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 312408 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
679 312408 : Self {
680 312408 : keyspace,
681 312408 : lsn,
682 312408 : path: Vec::new(),
683 312408 : }
684 312408 : }
685 :
686 447053 : pub fn record_layer_visit(
687 447053 : &mut self,
688 447053 : layer_to_read: &ReadableLayer,
689 447053 : keyspace_to_read: &KeySpace,
690 447053 : lsn_range: &Range<Lsn>,
691 447053 : ) {
692 447053 : let id = match layer_to_read {
693 139809 : ReadableLayer::PersistentLayer(layer) => {
694 139809 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
695 : }
696 307244 : ReadableLayer::InMemoryLayer(layer) => {
697 307244 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
698 : }
699 : };
700 447053 : self.path
701 447053 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
702 447053 : }
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 427207 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1199 427207 : self.applied_gc_cutoff_lsn.read()
1200 427207 : }
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 301270 : pub(crate) async fn get(
1228 301270 : &self,
1229 301270 : key: Key,
1230 301270 : lsn: Lsn,
1231 301270 : ctx: &RequestContext,
1232 301270 : ) -> Result<Bytes, PageReconstructError> {
1233 301270 : if !lsn.is_valid() {
1234 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1235 301270 : }
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 301270 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1241 :
1242 301270 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1243 :
1244 301270 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1245 :
1246 301270 : let vectored_res = self
1247 301270 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1248 301270 : .await;
1249 :
1250 301270 : let key_value = vectored_res?.pop_first();
1251 301267 : match key_value {
1252 301261 : Some((got_key, value)) => {
1253 301261 : 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 301261 : 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 301270 : }
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 312408 : pub(super) async fn get_vectored_impl(
1439 312408 : &self,
1440 312408 : query: VersionedKeySpaceQuery,
1441 312408 : reconstruct_state: &mut ValuesReconstructState,
1442 312408 : ctx: &RequestContext,
1443 312408 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1444 312408 : if query.is_empty() {
1445 0 : return Ok(BTreeMap::default());
1446 312408 : }
1447 :
1448 312408 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1449 312408 : Some(ReadPath::new(
1450 312408 : query.total_keyspace(),
1451 312408 : query.high_watermark_lsn()?,
1452 : ))
1453 : } else {
1454 0 : None
1455 : };
1456 :
1457 312408 : reconstruct_state.read_path = read_path;
1458 :
1459 312408 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1460 0 : RedoAttemptType::LegacyCompaction
1461 : } else {
1462 312408 : RedoAttemptType::ReadPage
1463 : };
1464 :
1465 312408 : let traversal_res: Result<(), _> = {
1466 312408 : let ctx = RequestContextBuilder::from(ctx)
1467 312408 : .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 312408 : .attached_child();
1475 :
1476 312408 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1477 312408 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1478 312408 : .await
1479 : };
1480 :
1481 312408 : 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 312400 : };
1498 :
1499 312400 : let layers_visited = reconstruct_state.get_layers_visited();
1500 :
1501 312400 : let ctx = RequestContextBuilder::from(ctx)
1502 312400 : .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 312400 : .attached_child();
1510 :
1511 312400 : let futs = FuturesUnordered::new();
1512 363529 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1513 363529 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1514 :
1515 363529 : futs.push({
1516 363529 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1517 363529 : let ctx = RequestContextBuilder::from(&ctx)
1518 363529 : .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 363529 : .attached_child();
1527 :
1528 363529 : async move {
1529 363529 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1530 :
1531 363529 : let res = state
1532 363529 : .collect_pending_ios()
1533 363529 : .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 363529 : .await;
1541 :
1542 363529 : let converted = match res {
1543 363529 : Ok(ok) => ok,
1544 0 : Err(err) => {
1545 0 : return (key, Err(err));
1546 : }
1547 : };
1548 363529 : 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 363529 : debug_assert!(
1553 363529 : converted
1554 363529 : .records
1555 1401588 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1556 0 : "{converted:?}"
1557 : );
1558 :
1559 363529 : let walredo_deltas = converted.num_deltas();
1560 363529 : let walredo_res = walredo_self
1561 363529 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1562 363529 : .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 363529 : .await;
1571 :
1572 363529 : (key, walredo_res)
1573 363529 : }
1574 : });
1575 : }
1576 :
1577 312400 : let results = futs
1578 312400 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1579 312400 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1580 312400 : .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 312400 : if !results.is_empty() {
1587 312221 : 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 312221 : }
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 312221 : let layers_visited = layers_visited as f64;
1615 312221 : let avg_layers_visited = layers_visited / results.len() as f64;
1616 312221 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1617 675750 : for _ in &results {
1618 363529 : self.metrics.layers_per_read.observe(layers_visited);
1619 363529 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1620 363529 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1621 363529 : }
1622 179 : }
1623 :
1624 312400 : Ok(results)
1625 312408 : }
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 609 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1740 609 : self.disk_consistent_lsn.load()
1741 609 : }
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 114262 : pub(crate) async fn wait_lsn(
1783 114262 : &self,
1784 114262 : lsn: Lsn,
1785 114262 : who_is_waiting: WaitLsnWaiter<'_>,
1786 114262 : timeout: WaitLsnTimeout,
1787 114262 : ctx: &RequestContext, /* Prepare for use by cancellation */
1788 114262 : ) -> Result<(), WaitLsnError> {
1789 114262 : let state = self.current_state();
1790 114262 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1791 0 : return Err(WaitLsnError::Shutdown);
1792 114262 : } else if !matches!(state, TimelineState::Active) {
1793 0 : return Err(WaitLsnError::BadState(state));
1794 114262 : }
1795 :
1796 114262 : if cfg!(debug_assertions) {
1797 114262 : 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 114262 : _ => {}
1826 : }
1827 0 : }
1828 :
1829 114262 : let timeout = match timeout {
1830 0 : WaitLsnTimeout::Custom(t) => t,
1831 114262 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1832 : };
1833 :
1834 114262 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1835 114262 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1836 :
1837 114262 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1838 114262 : 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 114262 : 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 114262 : let log_slow_period = Duration::from_secs(10);
1843 114262 : let mut logging_permit = None;
1844 114262 : let wait_for_timeout = monitor_slow_future(
1845 114262 : log_slow_threshold,
1846 114262 : log_slow_period,
1847 114262 : wait_for_timeout,
1848 : |MonitorSlowFutureCallback {
1849 : ready,
1850 : is_slow,
1851 : elapsed_total,
1852 : elapsed_since_last_callback,
1853 114262 : }| {
1854 114262 : self.metrics
1855 114262 : .wait_lsn_in_progress_micros
1856 114262 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1857 114262 : if !is_slow {
1858 114262 : 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 114262 : },
1882 : );
1883 114262 : 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 114262 : drop(logging_permit);
1886 114262 : drop(start_finish_counterpair_guard);
1887 114262 : drop(timer);
1888 114262 : match res {
1889 114262 : 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 114262 : }
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 115192 : pub(crate) fn current_state(&self) -> TimelineState {
2542 115192 : self.state.borrow().clone()
2543 115192 : }
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 568 : pub(crate) fn is_stopping(&self) -> bool {
2562 568 : self.current_state() == TimelineState::Stopping
2563 568 : }
2564 :
2565 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2566 0 : self.state.subscribe()
2567 0 : }
2568 :
2569 114263 : pub(crate) async fn wait_to_become_active(
2570 114263 : &self,
2571 114263 : _ctx: &RequestContext, // Prepare for use by cancellation
2572 114263 : ) -> Result<(), TimelineState> {
2573 114263 : let mut receiver = self.state.subscribe();
2574 : loop {
2575 114263 : let current_state = receiver.borrow().clone();
2576 114263 : 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 114262 : 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 114263 : }
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_enabled(&self) -> bool {
2895 973 : let tenant_conf = self.tenant_conf.load();
2896 973 : 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 : }
2901 :
2902 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> (RelSizeMigration, Option<Lsn>) {
2903 1099 : let (status, migrated_at) = self.rel_size_v2_status.load().as_ref().clone();
2904 1099 : (status.unwrap_or(RelSizeMigration::Legacy), migrated_at)
2905 1099 : }
2906 :
2907 23 : fn get_compaction_upper_limit(&self) -> usize {
2908 23 : let tenant_conf = self.tenant_conf.load();
2909 23 : tenant_conf
2910 23 : .tenant_conf
2911 23 : .compaction_upper_limit
2912 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2913 23 : }
2914 :
2915 0 : pub fn get_compaction_l0_first(&self) -> bool {
2916 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2917 0 : tenant_conf
2918 0 : .compaction_l0_first
2919 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2920 0 : }
2921 :
2922 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2923 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2924 0 : tenant_conf
2925 0 : .compaction_l0_semaphore
2926 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2927 0 : }
2928 :
2929 636 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2930 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2931 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2932 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2933 :
2934 : // If compaction is disabled, don't delay.
2935 636 : if self.get_compaction_period() == Duration::ZERO {
2936 632 : return None;
2937 4 : }
2938 :
2939 4 : let compaction_threshold = self.get_compaction_threshold();
2940 4 : let tenant_conf = self.tenant_conf.load();
2941 4 : let l0_flush_delay_threshold = tenant_conf
2942 4 : .tenant_conf
2943 4 : .l0_flush_delay_threshold
2944 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2945 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2946 :
2947 : // 0 disables backpressure.
2948 4 : if l0_flush_delay_threshold == 0 {
2949 0 : return None;
2950 4 : }
2951 :
2952 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2953 : // backpressure flushes below this.
2954 : // TODO: the tenant config should have validation to prevent this instead.
2955 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2956 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2957 636 : }
2958 :
2959 637 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2960 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2961 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2962 : // sibling timelines. We need more confidence before enabling this.
2963 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2964 :
2965 : // If compaction is disabled, don't stall.
2966 637 : if self.get_compaction_period() == Duration::ZERO {
2967 632 : return None;
2968 5 : }
2969 :
2970 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2971 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2972 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2973 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2974 5 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2975 0 : return None;
2976 5 : }
2977 :
2978 5 : let compaction_threshold = self.get_compaction_threshold();
2979 5 : let tenant_conf = self.tenant_conf.load();
2980 5 : let l0_flush_stall_threshold = tenant_conf
2981 5 : .tenant_conf
2982 5 : .l0_flush_stall_threshold
2983 5 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2984 :
2985 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2986 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2987 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2988 5 : if cfg!(feature = "testing")
2989 5 : && compaction_threshold == 1
2990 0 : && l0_flush_stall_threshold.is_none()
2991 : {
2992 0 : return None;
2993 5 : }
2994 :
2995 5 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2996 5 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2997 :
2998 : // 0 disables backpressure.
2999 5 : if l0_flush_stall_threshold == 0 {
3000 5 : return None;
3001 0 : }
3002 :
3003 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
3004 : // backpressure flushes below this.
3005 : // TODO: the tenant config should have validation to prevent this instead.
3006 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
3007 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
3008 637 : }
3009 :
3010 57 : fn get_image_creation_threshold(&self) -> usize {
3011 57 : let tenant_conf = self.tenant_conf.load();
3012 57 : tenant_conf
3013 57 : .tenant_conf
3014 57 : .image_creation_threshold
3015 57 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3016 57 : }
3017 :
3018 : // HADRON
3019 193 : fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
3020 193 : let tenant_conf = self.tenant_conf.load();
3021 193 : tenant_conf
3022 193 : .tenant_conf
3023 193 : .image_layer_force_creation_period
3024 193 : .or(self
3025 193 : .conf
3026 193 : .default_tenant_conf
3027 193 : .image_layer_force_creation_period)
3028 193 : }
3029 :
3030 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
3031 192 : let tenant_conf = &self.tenant_conf.load();
3032 192 : tenant_conf
3033 192 : .tenant_conf
3034 192 : .compaction_algorithm
3035 192 : .as_ref()
3036 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
3037 192 : .clone()
3038 192 : }
3039 :
3040 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
3041 192 : let tenant_conf = self.tenant_conf.load();
3042 192 : tenant_conf
3043 192 : .tenant_conf
3044 192 : .compaction_shard_ancestor
3045 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
3046 192 : }
3047 :
3048 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
3049 0 : let tenant_conf = self.tenant_conf.load();
3050 0 : tenant_conf
3051 0 : .tenant_conf
3052 0 : .eviction_policy
3053 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
3054 0 : }
3055 :
3056 235 : fn get_evictions_low_residence_duration_metric_threshold(
3057 235 : tenant_conf: &pageserver_api::models::TenantConfig,
3058 235 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
3059 235 : ) -> Duration {
3060 235 : tenant_conf
3061 235 : .evictions_low_residence_duration_metric_threshold
3062 235 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
3063 235 : }
3064 :
3065 191 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
3066 191 : let tenant_conf = self.tenant_conf.load();
3067 191 : tenant_conf
3068 191 : .tenant_conf
3069 191 : .image_layer_creation_check_threshold
3070 191 : .unwrap_or(
3071 191 : self.conf
3072 191 : .default_tenant_conf
3073 191 : .image_layer_creation_check_threshold,
3074 : )
3075 191 : }
3076 :
3077 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
3078 27 : let tenant_conf = &self.tenant_conf.load();
3079 27 : let gc_compaction_enabled = tenant_conf
3080 27 : .tenant_conf
3081 27 : .gc_compaction_enabled
3082 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
3083 27 : let gc_compaction_verification = tenant_conf
3084 27 : .tenant_conf
3085 27 : .gc_compaction_verification
3086 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
3087 27 : let gc_compaction_initial_threshold_kb = tenant_conf
3088 27 : .tenant_conf
3089 27 : .gc_compaction_initial_threshold_kb
3090 27 : .unwrap_or(
3091 27 : self.conf
3092 27 : .default_tenant_conf
3093 27 : .gc_compaction_initial_threshold_kb,
3094 : );
3095 27 : let gc_compaction_ratio_percent = tenant_conf
3096 27 : .tenant_conf
3097 27 : .gc_compaction_ratio_percent
3098 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
3099 27 : GcCompactionCombinedSettings {
3100 27 : gc_compaction_enabled,
3101 27 : gc_compaction_verification,
3102 27 : gc_compaction_initial_threshold_kb,
3103 27 : gc_compaction_ratio_percent,
3104 27 : }
3105 27 : }
3106 :
3107 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
3108 0 : let tenant_conf = self.tenant_conf.load();
3109 0 : tenant_conf
3110 0 : .tenant_conf
3111 0 : .image_creation_preempt_threshold
3112 0 : .unwrap_or(
3113 0 : self.conf
3114 0 : .default_tenant_conf
3115 0 : .image_creation_preempt_threshold,
3116 : )
3117 0 : }
3118 :
3119 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
3120 : // NB: Most tenant conf options are read by background loops, so,
3121 : // changes will automatically be picked up.
3122 :
3123 : // The threshold is embedded in the metric. So, we need to update it.
3124 : {
3125 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
3126 0 : &new_conf.tenant_conf,
3127 0 : &self.conf.default_tenant_conf,
3128 : );
3129 :
3130 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
3131 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
3132 :
3133 0 : let timeline_id_str = self.timeline_id.to_string();
3134 :
3135 0 : self.remote_client.update_config(&new_conf.location);
3136 :
3137 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
3138 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
3139 0 : if new_capacity != rel_size_cache.capacity() {
3140 0 : rel_size_cache.set_capacity(new_capacity);
3141 0 : }
3142 0 : }
3143 :
3144 0 : self.metrics
3145 0 : .evictions_with_low_residence_duration
3146 0 : .write()
3147 0 : .unwrap()
3148 0 : .change_threshold(
3149 0 : &tenant_id_str,
3150 0 : &shard_id_str,
3151 0 : &timeline_id_str,
3152 0 : new_threshold,
3153 0 : );
3154 : }
3155 0 : }
3156 :
3157 : /// Open a Timeline handle.
3158 : ///
3159 : /// Loads the metadata for the timeline into memory, but not the layer map.
3160 : #[allow(clippy::too_many_arguments)]
3161 235 : pub(super) fn new(
3162 235 : conf: &'static PageServerConf,
3163 235 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
3164 235 : metadata: &TimelineMetadata,
3165 235 : previous_heatmap: Option<PreviousHeatmap>,
3166 235 : ancestor: Option<Arc<Timeline>>,
3167 235 : timeline_id: TimelineId,
3168 235 : tenant_shard_id: TenantShardId,
3169 235 : generation: Generation,
3170 235 : shard_identity: ShardIdentity,
3171 235 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
3172 235 : resources: TimelineResources,
3173 235 : pg_version: PgMajorVersion,
3174 235 : state: TimelineState,
3175 235 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3176 235 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3177 235 : gc_compaction_state: Option<GcCompactionState>,
3178 235 : rel_size_v2_status: Option<RelSizeMigration>,
3179 235 : rel_size_migrated_at: Option<Lsn>,
3180 235 : cancel: CancellationToken,
3181 235 : ) -> Arc<Self> {
3182 235 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3183 235 : let (state, _) = watch::channel(state);
3184 :
3185 235 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3186 235 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3187 :
3188 235 : let evictions_low_residence_duration_metric_threshold = {
3189 235 : let loaded_tenant_conf = tenant_conf.load();
3190 235 : Self::get_evictions_low_residence_duration_metric_threshold(
3191 235 : &loaded_tenant_conf.tenant_conf,
3192 235 : &conf.default_tenant_conf,
3193 : )
3194 : };
3195 :
3196 235 : if let Some(ancestor) = &ancestor {
3197 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3198 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3199 118 : let is_offloaded = MaybeOffloaded::No;
3200 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3201 118 : }
3202 :
3203 235 : let relsize_snapshot_cache_capacity = {
3204 235 : let loaded_tenant_conf = tenant_conf.load();
3205 235 : loaded_tenant_conf
3206 235 : .tenant_conf
3207 235 : .relsize_snapshot_cache_capacity
3208 235 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3209 : };
3210 :
3211 235 : Arc::new_cyclic(|myself| {
3212 235 : let metrics = Arc::new(TimelineMetrics::new(
3213 235 : &tenant_shard_id,
3214 235 : &timeline_id,
3215 235 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3216 : "mtime",
3217 235 : evictions_low_residence_duration_metric_threshold,
3218 : ),
3219 : ));
3220 235 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3221 :
3222 235 : let mut result = Timeline {
3223 235 : conf,
3224 235 : tenant_conf,
3225 235 : myself: myself.clone(),
3226 235 : timeline_id,
3227 235 : tenant_shard_id,
3228 235 : generation,
3229 235 : shard_identity,
3230 235 : pg_version,
3231 235 : layers: Default::default(),
3232 235 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3233 :
3234 235 : walredo_mgr,
3235 235 : walreceiver: Mutex::new(None),
3236 :
3237 235 : remote_client: Arc::new(resources.remote_client),
3238 :
3239 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3240 235 : last_record_lsn: SeqWait::new(RecordLsn {
3241 235 : last: disk_consistent_lsn,
3242 235 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3243 235 : }),
3244 235 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3245 :
3246 235 : gc_compaction_state: ArcSwapOption::from_pointee(gc_compaction_state),
3247 :
3248 235 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3249 235 : last_freeze_ts: RwLock::new(Instant::now()),
3250 :
3251 235 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3252 :
3253 235 : ancestor_timeline: ancestor,
3254 235 : ancestor_lsn: metadata.ancestor_lsn(),
3255 :
3256 235 : metrics,
3257 :
3258 235 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3259 235 : &tenant_shard_id,
3260 235 : &timeline_id,
3261 235 : resources.pagestream_throttle_metrics,
3262 : ),
3263 :
3264 1880 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3265 1880 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3266 :
3267 235 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3268 :
3269 235 : layer_flush_start_tx,
3270 235 : layer_flush_done_tx,
3271 :
3272 235 : write_lock: tokio::sync::Mutex::new(None),
3273 :
3274 235 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3275 :
3276 235 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3277 235 : LastImageLayerCreationStatus::default(),
3278 : )),
3279 :
3280 235 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3281 235 : initdb_lsn: metadata.initdb_lsn(),
3282 :
3283 235 : current_logical_size: if disk_consistent_lsn.is_valid() {
3284 : // we're creating timeline data with some layer files existing locally,
3285 : // need to recalculate timeline's logical size based on data in the layers.
3286 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3287 : } else {
3288 : // we're creating timeline data without any layers existing locally,
3289 : // initial logical size is 0.
3290 115 : LogicalSize::empty_initial()
3291 : },
3292 :
3293 235 : partitioning: GuardArcSwap::new((
3294 235 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3295 235 : Lsn(0),
3296 235 : )),
3297 : repartition_threshold: 0,
3298 235 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3299 235 : last_image_layer_creation_check_instant: Mutex::new(None),
3300 235 : last_received_wal: Mutex::new(None),
3301 235 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3302 235 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3303 :
3304 235 : download_all_remote_layers_task_info: RwLock::new(None),
3305 :
3306 235 : state,
3307 :
3308 235 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3309 235 : EvictionTaskTimelineState::default(),
3310 : ),
3311 235 : delete_progress: TimelineDeleteProgress::default(),
3312 :
3313 235 : cancel,
3314 235 : gate: Gate::default(),
3315 :
3316 235 : compaction_lock: tokio::sync::Mutex::default(),
3317 235 : compaction_failed: AtomicBool::default(),
3318 235 : corruption_detected: AtomicBool::default(),
3319 235 : l0_compaction_trigger: resources.l0_compaction_trigger,
3320 235 : gc_lock: tokio::sync::Mutex::default(),
3321 :
3322 235 : standby_horizon: AtomicLsn::new(0),
3323 :
3324 235 : pagestream_throttle: resources.pagestream_throttle,
3325 :
3326 235 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3327 :
3328 : #[cfg(test)]
3329 235 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3330 :
3331 235 : l0_flush_global_state: resources.l0_flush_global_state,
3332 :
3333 235 : handles: Default::default(),
3334 :
3335 235 : attach_wal_lag_cooldown,
3336 :
3337 235 : create_idempotency,
3338 :
3339 235 : page_trace: Default::default(),
3340 :
3341 235 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3342 :
3343 235 : heatmap_layers_downloader: Mutex::new(None),
3344 :
3345 235 : rel_size_v2_status: ArcSwap::from_pointee((
3346 235 : rel_size_v2_status,
3347 235 : rel_size_migrated_at,
3348 235 : )),
3349 :
3350 235 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3351 :
3352 235 : basebackup_cache: resources.basebackup_cache,
3353 :
3354 235 : feature_resolver: resources.feature_resolver.clone(),
3355 :
3356 235 : db_rel_count: ArcSwapOption::from_pointee(None),
3357 : };
3358 :
3359 235 : result.repartition_threshold =
3360 235 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3361 :
3362 235 : result
3363 235 : .metrics
3364 235 : .last_record_lsn_gauge
3365 235 : .set(disk_consistent_lsn.0 as i64);
3366 235 : result
3367 235 : })
3368 235 : }
3369 :
3370 343 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3371 343 : let Ok(guard) = self.gate.enter() else {
3372 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3373 0 : return;
3374 : };
3375 343 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3376 343 : match *flush_loop_state {
3377 232 : FlushLoopState::NotStarted => (),
3378 : FlushLoopState::Running { .. } => {
3379 111 : info!(
3380 0 : "skipping attempt to start flush_loop twice {}/{}",
3381 0 : self.tenant_shard_id, self.timeline_id
3382 : );
3383 111 : return;
3384 : }
3385 : FlushLoopState::Exited => {
3386 0 : info!(
3387 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3388 0 : self.tenant_shard_id, self.timeline_id
3389 : );
3390 0 : return;
3391 : }
3392 : }
3393 :
3394 232 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3395 232 : let self_clone = Arc::clone(self);
3396 :
3397 232 : debug!("spawning flush loop");
3398 232 : *flush_loop_state = FlushLoopState::Running {
3399 232 : #[cfg(test)]
3400 232 : expect_initdb_optimization: false,
3401 232 : #[cfg(test)]
3402 232 : initdb_optimization_count: 0,
3403 232 : };
3404 232 : task_mgr::spawn(
3405 232 : task_mgr::BACKGROUND_RUNTIME.handle(),
3406 232 : task_mgr::TaskKind::LayerFlushTask,
3407 232 : self.tenant_shard_id,
3408 232 : Some(self.timeline_id),
3409 232 : "layer flush task",
3410 232 : async move {
3411 232 : let _guard = guard;
3412 232 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3413 232 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3414 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3415 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3416 5 : *flush_loop_state = FlushLoopState::Exited;
3417 5 : Ok(())
3418 5 : }
3419 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))
3420 : );
3421 343 : }
3422 :
3423 0 : pub(crate) fn update_gc_compaction_state(
3424 0 : &self,
3425 0 : gc_compaction_state: GcCompactionState,
3426 0 : ) -> anyhow::Result<()> {
3427 0 : self.gc_compaction_state
3428 0 : .store(Some(Arc::new(gc_compaction_state.clone())));
3429 0 : self.remote_client
3430 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3431 0 : }
3432 :
3433 0 : pub(crate) fn update_rel_size_v2_status(
3434 0 : &self,
3435 0 : rel_size_v2_status: RelSizeMigration,
3436 0 : rel_size_migrated_at: Option<Lsn>,
3437 0 : ) -> anyhow::Result<()> {
3438 0 : self.rel_size_v2_status.store(Arc::new((
3439 0 : Some(rel_size_v2_status.clone()),
3440 0 : rel_size_migrated_at,
3441 0 : )));
3442 0 : self.remote_client
3443 0 : .schedule_index_upload_for_rel_size_v2_status_update(
3444 0 : rel_size_v2_status,
3445 0 : rel_size_migrated_at,
3446 : )
3447 0 : }
3448 :
3449 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3450 0 : self.gc_compaction_state
3451 0 : .load()
3452 0 : .as_ref()
3453 0 : .map(|x| x.as_ref().clone())
3454 0 : }
3455 :
3456 : /// Creates and starts the wal receiver.
3457 : ///
3458 : /// This function is expected to be called at most once per Timeline's lifecycle
3459 : /// when the timeline is activated.
3460 0 : fn launch_wal_receiver(
3461 0 : self: &Arc<Self>,
3462 0 : ctx: &RequestContext,
3463 0 : broker_client: BrokerClientChannel,
3464 0 : ) {
3465 0 : info!(
3466 0 : "launching WAL receiver for timeline {} of tenant {}",
3467 0 : self.timeline_id, self.tenant_shard_id
3468 : );
3469 :
3470 0 : let tenant_conf = self.tenant_conf.load();
3471 0 : let wal_connect_timeout = tenant_conf
3472 0 : .tenant_conf
3473 0 : .walreceiver_connect_timeout
3474 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3475 0 : let lagging_wal_timeout = tenant_conf
3476 0 : .tenant_conf
3477 0 : .lagging_wal_timeout
3478 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3479 0 : let max_lsn_wal_lag = tenant_conf
3480 0 : .tenant_conf
3481 0 : .max_lsn_wal_lag
3482 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3483 :
3484 0 : let mut guard = self.walreceiver.lock().unwrap();
3485 0 : assert!(
3486 0 : guard.is_none(),
3487 0 : "multiple launches / re-launches of WAL receiver are not supported"
3488 : );
3489 :
3490 0 : let protocol = PostgresClientProtocol::Interpreted {
3491 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3492 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3493 0 : };
3494 :
3495 0 : *guard = Some(WalReceiver::start(
3496 0 : Arc::clone(self),
3497 0 : WalReceiverConf {
3498 0 : protocol,
3499 0 : wal_connect_timeout,
3500 0 : lagging_wal_timeout,
3501 0 : max_lsn_wal_lag,
3502 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3503 0 : availability_zone: self.conf.availability_zone.clone(),
3504 0 : ingest_batch_size: self.conf.ingest_batch_size,
3505 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3506 0 : },
3507 0 : broker_client,
3508 0 : ctx,
3509 0 : ));
3510 0 : }
3511 :
3512 : /// Initialize with an empty layer map. Used when creating a new timeline.
3513 232 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3514 232 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3515 232 : "in the context where we call this function, no other task has access to the object",
3516 : );
3517 232 : layers
3518 232 : .open_mut()
3519 232 : .expect("in this context the LayerManager must still be open")
3520 232 : .initialize_empty(Lsn(start_lsn.0));
3521 232 : }
3522 :
3523 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3524 : /// files.
3525 3 : pub(super) async fn load_layer_map(
3526 3 : &self,
3527 3 : disk_consistent_lsn: Lsn,
3528 3 : index_part: IndexPart,
3529 3 : ) -> anyhow::Result<()> {
3530 : use LayerName::*;
3531 : use init::Decision::*;
3532 : use init::{Discovered, DismissedLayer};
3533 :
3534 3 : let mut guard = self
3535 3 : .layers
3536 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3537 3 : .await;
3538 :
3539 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3540 :
3541 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3542 : // structs representing all files on disk
3543 3 : let timeline_path = self
3544 3 : .conf
3545 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3546 3 : let conf = self.conf;
3547 3 : let span = tracing::Span::current();
3548 :
3549 : // Copy to move into the task we're about to spawn
3550 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3551 :
3552 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3553 3 : move || {
3554 3 : let _g = span.entered();
3555 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3556 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3557 3 : let mut unrecognized_files = Vec::new();
3558 :
3559 3 : let mut path = timeline_path;
3560 :
3561 11 : for discovered in discovered {
3562 8 : let (name, kind) = match discovered {
3563 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3564 8 : discovered_layers.push((layer_file_name, local_metadata));
3565 8 : continue;
3566 : }
3567 0 : Discovered::IgnoredBackup(path) => {
3568 0 : std::fs::remove_file(path)
3569 0 : .or_else(fs_ext::ignore_not_found)
3570 0 : .fatal_err("Removing .old file");
3571 0 : continue;
3572 : }
3573 0 : Discovered::Unknown(file_name) => {
3574 : // we will later error if there are any
3575 0 : unrecognized_files.push(file_name);
3576 0 : continue;
3577 : }
3578 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3579 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3580 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3581 : };
3582 0 : path.push(Utf8Path::new(&name));
3583 0 : init::cleanup(&path, kind)?;
3584 0 : path.pop();
3585 : }
3586 :
3587 3 : if !unrecognized_files.is_empty() {
3588 : // assume that if there are any there are many many.
3589 0 : let n = unrecognized_files.len();
3590 0 : let first = &unrecognized_files[..n.min(10)];
3591 0 : anyhow::bail!(
3592 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3593 : );
3594 3 : }
3595 :
3596 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3597 :
3598 3 : let mut loaded_layers = Vec::new();
3599 3 : let mut needs_cleanup = Vec::new();
3600 3 : let mut total_physical_size = 0;
3601 :
3602 11 : for (name, decision) in decided {
3603 8 : let decision = match decision {
3604 8 : Ok(decision) => decision,
3605 0 : Err(DismissedLayer::Future { local }) => {
3606 0 : if let Some(local) = local {
3607 0 : init::cleanup_future_layer(
3608 0 : &local.local_path,
3609 0 : &name,
3610 0 : disk_consistent_lsn,
3611 0 : )?;
3612 0 : }
3613 0 : needs_cleanup.push(name);
3614 0 : continue;
3615 : }
3616 0 : Err(DismissedLayer::LocalOnly(local)) => {
3617 0 : init::cleanup_local_only_file(&name, &local)?;
3618 : // this file never existed remotely, we will have to do rework
3619 0 : continue;
3620 : }
3621 0 : Err(DismissedLayer::BadMetadata(local)) => {
3622 0 : init::cleanup_local_file_for_remote(&local)?;
3623 : // this file never existed remotely, we will have to do rework
3624 0 : continue;
3625 : }
3626 : };
3627 :
3628 8 : match &name {
3629 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3630 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3631 : }
3632 :
3633 8 : tracing::debug!(layer=%name, ?decision, "applied");
3634 :
3635 8 : let layer = match decision {
3636 8 : Resident { local, remote } => {
3637 8 : total_physical_size += local.file_size;
3638 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3639 8 : .drop_eviction_guard()
3640 : }
3641 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3642 : };
3643 :
3644 8 : loaded_layers.push(layer);
3645 : }
3646 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3647 3 : }
3648 : })
3649 3 : .await
3650 3 : .map_err(anyhow::Error::new)
3651 3 : .and_then(|x| x)?;
3652 :
3653 3 : let num_layers = loaded_layers.len();
3654 :
3655 3 : guard
3656 3 : .open_mut()
3657 3 : .expect("layermanager must be open during init")
3658 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3659 :
3660 3 : self.remote_client
3661 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3662 3 : self.remote_client
3663 3 : .schedule_index_upload_for_file_changes()?;
3664 : // This barrier orders above DELETEs before any later operations.
3665 : // This is critical because code executing after the barrier might
3666 : // create again objects with the same key that we just scheduled for deletion.
3667 : // For example, if we just scheduled deletion of an image layer "from the future",
3668 : // later compaction might run again and re-create the same image layer.
3669 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3670 : // "same" here means same key range and LSN.
3671 : //
3672 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3673 : // the upload queue may execute the PUT first, then the DELETE.
3674 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3675 : //
3676 : // 1. a future image layer is created and uploaded
3677 : // 2. ps restart
3678 : // 3. the future layer from (1) is deleted during load layer map
3679 : // 4. image layer is re-created and uploaded
3680 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3681 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3682 : //
3683 : // See https://github.com/neondatabase/neon/issues/5878
3684 : //
3685 : // NB: generation numbers naturally protect against this because they disambiguate
3686 : // (1) and (4)
3687 : // TODO: this is basically a no-op now, should we remove it?
3688 3 : self.remote_client.schedule_barrier()?;
3689 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3690 : // on retry.
3691 :
3692 3 : info!(
3693 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3694 : num_layers, disk_consistent_lsn, total_physical_size
3695 : );
3696 :
3697 3 : timer.stop_and_record();
3698 3 : Ok(())
3699 3 : }
3700 :
3701 : /// Retrieve current logical size of the timeline.
3702 : ///
3703 : /// The size could be lagging behind the actual number, in case
3704 : /// the initial size calculation has not been run (gets triggered on the first size access).
3705 : ///
3706 : /// return size and boolean flag that shows if the size is exact
3707 0 : pub(crate) fn get_current_logical_size(
3708 0 : self: &Arc<Self>,
3709 0 : priority: GetLogicalSizePriority,
3710 0 : ctx: &RequestContext,
3711 0 : ) -> logical_size::CurrentLogicalSize {
3712 0 : if !self.tenant_shard_id.is_shard_zero() {
3713 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3714 : // when HTTP API is serving a GET for timeline zero, return zero
3715 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3716 0 : }
3717 :
3718 0 : let current_size = self.current_logical_size.current_size();
3719 0 : debug!("Current size: {current_size:?}");
3720 :
3721 0 : match (current_size.accuracy(), priority) {
3722 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3723 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3724 0 : // background task will eventually deliver an exact value, we're in no rush
3725 0 : }
3726 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3727 : // background task is not ready, but user is asking for it now;
3728 : // => make the background task skip the line
3729 : // (The alternative would be to calculate the size here, but,
3730 : // it can actually take a long time if the user has a lot of rels.
3731 : // And we'll inevitable need it again; So, let the background task do the work.)
3732 0 : match self
3733 0 : .current_logical_size
3734 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3735 0 : .get()
3736 : {
3737 0 : Some(cancel) => cancel.cancel(),
3738 : None => {
3739 0 : match self.current_state() {
3740 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3741 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3742 0 : // Don't make noise.
3743 0 : }
3744 : TimelineState::Loading => {
3745 : // Import does not return an activated timeline.
3746 0 : info!(
3747 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3748 : );
3749 : }
3750 : TimelineState::Active => {
3751 : // activation should be setting the once cell
3752 0 : warn!(
3753 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3754 : );
3755 0 : debug_assert!(false);
3756 : }
3757 : }
3758 : }
3759 : }
3760 : }
3761 : }
3762 :
3763 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3764 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3765 0 : let first = self
3766 0 : .current_logical_size
3767 0 : .did_return_approximate_to_walreceiver
3768 0 : .compare_exchange(
3769 0 : false,
3770 0 : true,
3771 0 : AtomicOrdering::Relaxed,
3772 0 : AtomicOrdering::Relaxed,
3773 0 : )
3774 0 : .is_ok();
3775 0 : if first {
3776 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3777 0 : }
3778 0 : }
3779 0 : }
3780 :
3781 0 : current_size
3782 0 : }
3783 :
3784 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3785 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3786 : // nothing to do for freshly created timelines;
3787 0 : assert_eq!(
3788 0 : self.current_logical_size.current_size().accuracy(),
3789 : logical_size::Accuracy::Exact,
3790 : );
3791 0 : self.current_logical_size.initialized.add_permits(1);
3792 0 : return;
3793 : };
3794 :
3795 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3796 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3797 0 : self.current_logical_size
3798 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3799 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3800 :
3801 0 : let self_clone = Arc::clone(self);
3802 0 : let background_ctx = ctx.detached_child(
3803 0 : TaskKind::InitialLogicalSizeCalculation,
3804 0 : DownloadBehavior::Download,
3805 : );
3806 0 : task_mgr::spawn(
3807 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3808 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3809 0 : self.tenant_shard_id,
3810 0 : Some(self.timeline_id),
3811 0 : "initial size calculation",
3812 : // NB: don't log errors here, task_mgr will do that.
3813 0 : async move {
3814 0 : self_clone
3815 0 : .initial_logical_size_calculation_task(
3816 0 : initial_part_end,
3817 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3818 0 : background_ctx,
3819 0 : )
3820 0 : .await;
3821 0 : Ok(())
3822 0 : }
3823 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)),
3824 : );
3825 0 : }
3826 :
3827 : /// # Cancellation
3828 : ///
3829 : /// This method is sensitive to `Timeline::cancel`.
3830 : ///
3831 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3832 : ///
3833 : /// # Cancel-Safety
3834 : ///
3835 : /// It does Timeline IO, hence this should be polled to completion because
3836 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3837 : /// to reason about.
3838 0 : async fn initial_logical_size_calculation_task(
3839 0 : self: Arc<Self>,
3840 0 : initial_part_end: Lsn,
3841 0 : skip_concurrency_limiter: CancellationToken,
3842 0 : background_ctx: RequestContext,
3843 0 : ) {
3844 0 : scopeguard::defer! {
3845 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3846 : self.current_logical_size.initialized.add_permits(1);
3847 : }
3848 :
3849 0 : let try_once = |attempt: usize| {
3850 0 : let background_ctx = &background_ctx;
3851 0 : let self_ref = &self;
3852 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3853 0 : async move {
3854 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3855 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3856 0 : background_ctx,
3857 : );
3858 :
3859 : use crate::metrics::initial_logical_size::StartCircumstances;
3860 0 : let (_maybe_permit, circumstances) = tokio::select! {
3861 0 : permit = wait_for_permit => {
3862 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3863 : }
3864 0 : _ = self_ref.cancel.cancelled() => {
3865 0 : return Err(CalculateLogicalSizeError::Cancelled);
3866 : }
3867 0 : () = skip_concurrency_limiter.cancelled() => {
3868 : // Some action that is part of a end user interaction requested logical size
3869 : // => break out of the rate limit
3870 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3871 : // but then again what happens if they cancel; also, we should just be using
3872 : // one runtime across the entire process, so, let's leave this for now.
3873 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3874 : }
3875 : };
3876 :
3877 0 : let metrics_guard = if attempt == 1 {
3878 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3879 : } else {
3880 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3881 : };
3882 :
3883 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3884 0 : self_ref.conf.get_vectored_concurrent_io,
3885 0 : self_ref
3886 0 : .gate
3887 0 : .enter()
3888 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3889 : );
3890 :
3891 0 : let calculated_size = self_ref
3892 0 : .logical_size_calculation_task(
3893 0 : initial_part_end,
3894 0 : LogicalSizeCalculationCause::Initial,
3895 0 : background_ctx,
3896 0 : )
3897 0 : .await?;
3898 :
3899 0 : self_ref
3900 0 : .trigger_aux_file_size_computation(
3901 0 : initial_part_end,
3902 0 : background_ctx,
3903 0 : io_concurrency,
3904 0 : )
3905 0 : .await?;
3906 :
3907 : // TODO: add aux file size to logical size
3908 :
3909 0 : Ok((calculated_size, metrics_guard))
3910 0 : }
3911 0 : };
3912 :
3913 0 : let retrying = async {
3914 0 : let mut attempt = 0;
3915 : loop {
3916 0 : attempt += 1;
3917 :
3918 0 : match try_once(attempt).await {
3919 0 : Ok(res) => return ControlFlow::Continue(res),
3920 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3921 : Err(
3922 0 : e @ (CalculateLogicalSizeError::Decode(_)
3923 : | CalculateLogicalSizeError::PageRead(_)),
3924 : ) => {
3925 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3926 : // exponential back-off doesn't make sense at these long intervals;
3927 : // use fixed retry interval with generous jitter instead
3928 0 : let sleep_duration = Duration::from_secs(
3929 0 : u64::try_from(
3930 : // 1hour base
3931 0 : (60_i64 * 60_i64)
3932 0 : // 10min jitter
3933 0 : + rand::rng().random_range(-10 * 60..10 * 60),
3934 : )
3935 0 : .expect("10min < 1hour"),
3936 : );
3937 0 : tokio::select! {
3938 0 : _ = tokio::time::sleep(sleep_duration) => {}
3939 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3940 : }
3941 : }
3942 : }
3943 : }
3944 0 : };
3945 :
3946 0 : let (calculated_size, metrics_guard) = match retrying.await {
3947 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3948 0 : ControlFlow::Break(()) => return,
3949 : };
3950 :
3951 : // we cannot query current_logical_size.current_size() to know the current
3952 : // *negative* value, only truncated to u64.
3953 0 : let added = self
3954 0 : .current_logical_size
3955 0 : .size_added_after_initial
3956 0 : .load(AtomicOrdering::Relaxed);
3957 :
3958 0 : let sum = calculated_size.saturating_add_signed(added);
3959 :
3960 : // set the gauge value before it can be set in `update_current_logical_size`.
3961 0 : self.metrics.current_logical_size_gauge.set(sum);
3962 :
3963 0 : self.current_logical_size
3964 0 : .initial_logical_size
3965 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3966 0 : .ok()
3967 0 : .expect("only this task sets it");
3968 0 : }
3969 :
3970 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3971 7 : self: &Arc<Self>,
3972 7 : lsn: Lsn,
3973 7 : cause: LogicalSizeCalculationCause,
3974 7 : ctx: RequestContext,
3975 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3976 7 : let (sender, receiver) = oneshot::channel();
3977 7 : let self_clone = Arc::clone(self);
3978 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3979 : // we should stop the size calculation work and return an error.
3980 : // That would require restructuring this function's API to
3981 : // return the result directly, instead of a Receiver for the result.
3982 7 : let ctx = ctx.detached_child(
3983 7 : TaskKind::OndemandLogicalSizeCalculation,
3984 7 : DownloadBehavior::Download,
3985 : );
3986 7 : task_mgr::spawn(
3987 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3988 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3989 7 : self.tenant_shard_id,
3990 7 : Some(self.timeline_id),
3991 7 : "ondemand logical size calculation",
3992 7 : async move {
3993 7 : let res = self_clone
3994 7 : .logical_size_calculation_task(lsn, cause, &ctx)
3995 7 : .await;
3996 7 : let _ = sender.send(res).ok();
3997 7 : Ok(()) // Receiver is responsible for handling errors
3998 7 : }
3999 7 : .in_current_span(),
4000 : );
4001 7 : receiver
4002 7 : }
4003 :
4004 : #[instrument(skip_all)]
4005 : async fn logical_size_calculation_task(
4006 : self: &Arc<Self>,
4007 : lsn: Lsn,
4008 : cause: LogicalSizeCalculationCause,
4009 : ctx: &RequestContext,
4010 : ) -> Result<u64, CalculateLogicalSizeError> {
4011 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
4012 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
4013 : // accurate relation sizes, and they do not emit consumption metrics.
4014 : debug_assert!(self.tenant_shard_id.is_shard_zero());
4015 :
4016 : let guard = self
4017 : .gate
4018 : .enter()
4019 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
4020 :
4021 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
4022 : }
4023 :
4024 : /// Calculate the logical size of the database at the latest LSN.
4025 : ///
4026 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
4027 : /// especially if we need to download remote layers.
4028 7 : async fn calculate_logical_size(
4029 7 : &self,
4030 7 : up_to_lsn: Lsn,
4031 7 : cause: LogicalSizeCalculationCause,
4032 7 : _guard: &GateGuard,
4033 7 : ctx: &RequestContext,
4034 7 : ) -> Result<u64, CalculateLogicalSizeError> {
4035 7 : info!(
4036 0 : "Calculating logical size for timeline {} at {}",
4037 : self.timeline_id, up_to_lsn
4038 : );
4039 :
4040 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
4041 : {
4042 0 : return Err(CalculateLogicalSizeError::Cancelled);
4043 7 : }
4044 :
4045 : // See if we've already done the work for initial size calculation.
4046 : // This is a short-cut for timelines that are mostly unused.
4047 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
4048 0 : return Ok(size);
4049 7 : }
4050 7 : let storage_time_metrics = match cause {
4051 : LogicalSizeCalculationCause::Initial
4052 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
4053 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
4054 : LogicalSizeCalculationCause::EvictionTaskImitation => {
4055 7 : &self.metrics.imitate_logical_size_histo
4056 : }
4057 : };
4058 7 : let timer = storage_time_metrics.start_timer();
4059 7 : let logical_size = self
4060 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
4061 7 : .await?;
4062 7 : debug!("calculated logical size: {logical_size}");
4063 7 : timer.stop_and_record();
4064 7 : Ok(logical_size)
4065 7 : }
4066 :
4067 : /// Update current logical size, adding `delta' to the old value.
4068 135285 : fn update_current_logical_size(&self, delta: i64) {
4069 135285 : let logical_size = &self.current_logical_size;
4070 135285 : logical_size.increment_size(delta);
4071 :
4072 : // Also set the value in the prometheus gauge. Note that
4073 : // there is a race condition here: if this is is called by two
4074 : // threads concurrently, the prometheus gauge might be set to
4075 : // one value while current_logical_size is set to the
4076 : // other.
4077 135285 : match logical_size.current_size() {
4078 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
4079 135285 : .metrics
4080 135285 : .current_logical_size_gauge
4081 135285 : .set(new_current_size.into()),
4082 0 : CurrentLogicalSize::Approximate(_) => {
4083 0 : // don't update the gauge yet, this allows us not to update the gauge back and
4084 0 : // forth between the initial size calculation task.
4085 0 : }
4086 : }
4087 135285 : }
4088 :
4089 1527 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
4090 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
4091 : // for each of the database, but we only store one value, and therefore each pgdirmodification
4092 : // would overwrite the previous value if they modify different databases.
4093 :
4094 1527 : match count {
4095 566 : MetricsUpdate::Set(count) => {
4096 566 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
4097 566 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
4098 566 : }
4099 960 : MetricsUpdate::Add(count) => {
4100 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4101 : // it's fine.
4102 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4103 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
4104 960 : // the value reliably.
4105 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
4106 960 : }
4107 : // Otherwise, ignore this update
4108 : }
4109 1 : MetricsUpdate::Sub(count) => {
4110 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4111 : // it's fine.
4112 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4113 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
4114 1 : // The operation could overflow so we need to normalize the value.
4115 1 : let prev_val =
4116 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
4117 1 : let res = prev_val.saturating_sub(count);
4118 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
4119 1 : }
4120 : // Otherwise, ignore this update
4121 : }
4122 : };
4123 :
4124 : // TODO: remove this, there's no place in the code that updates this aux metrics.
4125 1527 : let aux_metric =
4126 1527 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
4127 :
4128 1527 : let sum_of_entries = self
4129 1527 : .directory_metrics
4130 1527 : .iter()
4131 12216 : .map(|v| v.load(AtomicOrdering::Relaxed))
4132 1527 : .sum();
4133 : // Set a high general threshold and a lower threshold for the auxiliary files,
4134 : // as we can have large numbers of relations in the db directory.
4135 : const SUM_THRESHOLD: u64 = 5000;
4136 : const AUX_THRESHOLD: u64 = 1000;
4137 1527 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
4138 0 : self.metrics
4139 0 : .directory_entries_count_gauge
4140 0 : .set(sum_of_entries);
4141 1527 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
4142 0 : metric.set(sum_of_entries);
4143 1527 : }
4144 1527 : }
4145 :
4146 0 : async fn find_layer(
4147 0 : &self,
4148 0 : layer_name: &LayerName,
4149 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
4150 0 : let guard = self
4151 0 : .layers
4152 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
4153 0 : .await;
4154 0 : let layer = guard
4155 0 : .layer_map()?
4156 0 : .iter_historic_layers()
4157 0 : .find(|l| &l.layer_name() == layer_name)
4158 0 : .map(|found| guard.get_from_desc(&found));
4159 0 : Ok(layer)
4160 0 : }
4161 :
4162 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
4163 0 : let crnt = self.previous_heatmap.load();
4164 0 : match crnt.as_deref() {
4165 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
4166 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
4167 0 : None => true,
4168 : },
4169 0 : Some(PreviousHeatmap::Obsolete) => false,
4170 0 : None => false,
4171 : }
4172 0 : }
4173 :
4174 : /// The timeline heatmap is a hint to secondary locations from the primary location,
4175 : /// indicating which layers are currently on-disk on the primary.
4176 : ///
4177 : /// None is returned if the Timeline is in a state where uploading a heatmap
4178 : /// doesn't make sense, such as shutting down or initializing. The caller
4179 : /// should treat this as a cue to simply skip doing any heatmap uploading
4180 : /// for this timeline.
4181 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
4182 8 : if !self.is_active() {
4183 0 : return None;
4184 8 : }
4185 :
4186 8 : let guard = self
4187 8 : .layers
4188 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
4189 8 : .await;
4190 :
4191 : // Firstly, if there's any heatmap left over from when this location
4192 : // was a secondary, take that into account. Keep layers that are:
4193 : // * present in the layer map
4194 : // * visible
4195 : // * non-resident
4196 : // * not evicted since we read the heatmap
4197 : //
4198 : // Without this, a new cold, attached location would clobber the previous
4199 : // heatamp.
4200 8 : let previous_heatmap = self.previous_heatmap.load();
4201 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4202 : Some(PreviousHeatmap::Active {
4203 6 : heatmap, read_at, ..
4204 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4205 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4206 23 : let layer = guard.try_get_from_key(&desc.key())?;
4207 :
4208 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4209 0 : return None;
4210 23 : }
4211 :
4212 23 : if layer.is_likely_resident() {
4213 10 : return None;
4214 13 : }
4215 :
4216 13 : if layer.last_evicted_at().happened_after(*read_at) {
4217 3 : return None;
4218 10 : }
4219 :
4220 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4221 23 : })),
4222 0 : Some(PreviousHeatmap::Obsolete) => None,
4223 2 : None => None,
4224 : };
4225 :
4226 : // Secondly, all currently visible, resident layers are included.
4227 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4228 18 : match layer.visibility() {
4229 : LayerVisibilityHint::Visible => {
4230 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4231 17 : let last_activity_ts = layer.latest_activity();
4232 17 : Some((
4233 17 : layer.layer_desc().clone(),
4234 17 : layer.metadata(),
4235 17 : last_activity_ts,
4236 17 : false, // these layers are not cold
4237 17 : ))
4238 : }
4239 : LayerVisibilityHint::Covered => {
4240 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4241 1 : None
4242 : }
4243 : }
4244 18 : });
4245 :
4246 8 : let mut layers = match visible_non_resident {
4247 6 : Some(non_resident) => {
4248 6 : let mut non_resident = non_resident.peekable();
4249 6 : if non_resident.peek().is_none() {
4250 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4251 2 : self.previous_heatmap
4252 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4253 4 : }
4254 :
4255 6 : non_resident.chain(resident).collect::<Vec<_>>()
4256 : }
4257 2 : None => resident.collect::<Vec<_>>(),
4258 : };
4259 :
4260 : // Sort layers in order of which to download first. For a large set of layers to download, we
4261 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4262 : // or hours later:
4263 : // - Cold layers go last for convenience when a human inspects the heatmap.
4264 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4265 : // only exist for a few minutes before being compacted into L1s.
4266 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4267 : // the layer is likely to be covered by an image layer during compaction.
4268 62 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4269 62 : std::cmp::Reverse((
4270 62 : *cold,
4271 62 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4272 62 : desc.lsn_range.end,
4273 62 : ))
4274 62 : });
4275 :
4276 8 : let layers = layers
4277 8 : .into_iter()
4278 27 : .map(|(desc, meta, atime, cold)| {
4279 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4280 27 : })
4281 8 : .collect();
4282 :
4283 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4284 8 : }
4285 :
4286 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4287 0 : let guard = self
4288 0 : .layers
4289 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4290 0 : .await;
4291 :
4292 0 : let now = SystemTime::now();
4293 0 : let mut heatmap_layers = Vec::default();
4294 0 : for vl in guard.visible_layers() {
4295 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4296 0 : continue;
4297 0 : }
4298 :
4299 0 : let hl = HeatMapLayer {
4300 0 : name: vl.layer_desc().layer_name(),
4301 0 : metadata: vl.metadata(),
4302 0 : access_time: now,
4303 0 : cold: true,
4304 0 : };
4305 0 : heatmap_layers.push(hl);
4306 : }
4307 :
4308 0 : tracing::info!(
4309 0 : "Generating unarchival heatmap with {} layers",
4310 0 : heatmap_layers.len()
4311 : );
4312 :
4313 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4314 0 : PreviousHeatmap::Active {
4315 0 : heatmap,
4316 0 : read_at: Instant::now(),
4317 0 : end_lsn: Some(end_lsn),
4318 0 : }
4319 0 : }
4320 :
4321 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4322 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4323 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4324 : // branchpoint in the value in IndexPart::lineage
4325 0 : self.ancestor_lsn == lsn
4326 0 : || (self.ancestor_lsn == Lsn::INVALID
4327 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4328 0 : }
4329 : }
4330 :
4331 : #[derive(Clone)]
4332 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4333 : /// In other words, a set of segments in a 2D space.
4334 : ///
4335 : /// This representation has the advatange of avoiding hash map
4336 : /// allocations for uniform queries.
4337 : pub(crate) enum VersionedKeySpaceQuery {
4338 : /// Variant for queries at a single [`Lsn`]
4339 : Uniform { keyspace: KeySpace, lsn: Lsn },
4340 : /// Variant for queries at multiple [`Lsn`]s
4341 : Scattered {
4342 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4343 : },
4344 : }
4345 :
4346 : impl VersionedKeySpaceQuery {
4347 302216 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4348 302216 : Self::Uniform { keyspace, lsn }
4349 302216 : }
4350 :
4351 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4352 10192 : Self::Scattered { keyspaces_at_lsn }
4353 10192 : }
4354 :
4355 : /// Returns the most recent (largest) LSN included in the query.
4356 : /// If any of the LSNs included in the query are invalid, returns
4357 : /// an error instead.
4358 624816 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4359 624816 : match self {
4360 604432 : Self::Uniform { lsn, .. } => {
4361 604432 : if !lsn.is_valid() {
4362 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4363 604432 : }
4364 :
4365 604432 : Ok(*lsn)
4366 : }
4367 20384 : Self::Scattered { keyspaces_at_lsn } => {
4368 20384 : let mut max_lsn = None;
4369 42216 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4370 42216 : if !lsn.is_valid() {
4371 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4372 42216 : }
4373 42216 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4374 : }
4375 :
4376 20384 : if let Some(computed) = max_lsn {
4377 20384 : Ok(*computed)
4378 : } else {
4379 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4380 : }
4381 : }
4382 : }
4383 624816 : }
4384 :
4385 : /// Returns the total keyspace being queried: the result of projecting
4386 : /// everything in the key dimensions onto the key axis.
4387 323485 : fn total_keyspace(&self) -> KeySpace {
4388 323485 : match self {
4389 303101 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4390 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4391 20384 : .iter()
4392 20384 : .map(|(_lsn, keyspace)| keyspace)
4393 42216 : .fold(KeySpace::default(), |mut acc, v| {
4394 42216 : acc.merge(v);
4395 42216 : acc
4396 42216 : }),
4397 : }
4398 323485 : }
4399 :
4400 : /// Returns LSN for a specific key.
4401 : ///
4402 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4403 395529 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4404 395529 : match self {
4405 322337 : Self::Uniform { lsn, .. } => *lsn,
4406 73192 : Self::Scattered { keyspaces_at_lsn } => {
4407 73192 : keyspaces_at_lsn
4408 73192 : .iter()
4409 420940 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4410 73192 : .expect("Returned key was requested")
4411 : .0
4412 : }
4413 : }
4414 395529 : }
4415 :
4416 : /// Remove any parts of the query (segments) which overlap with the provided
4417 : /// key space (also segments).
4418 967603 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4419 967603 : match self {
4420 947219 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4421 20384 : Self::Scattered { keyspaces_at_lsn } => {
4422 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4423 42216 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4424 42216 : let removed = keyspace.remove_overlapping_with(to_remove);
4425 42216 : removed_accum.add_keyspace(removed);
4426 42216 : });
4427 :
4428 20384 : removed_accum.to_keyspace()
4429 : }
4430 : }
4431 967603 : }
4432 :
4433 739078 : fn is_empty(&self) -> bool {
4434 739078 : match self {
4435 718694 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4436 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4437 20384 : .iter()
4438 31300 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4439 : }
4440 739078 : }
4441 :
4442 : /// "Lower" the query on the LSN dimension
4443 114263 : fn lower(&mut self, to: Lsn) {
4444 114263 : match self {
4445 114263 : Self::Uniform { lsn, .. } => {
4446 114263 : // If the originally requested LSN is smaller than the starting
4447 114263 : // LSN of the ancestor we are descending into, we need to respect that.
4448 114263 : // Hence the min.
4449 114263 : *lsn = std::cmp::min(*lsn, to);
4450 114263 : }
4451 0 : Self::Scattered { keyspaces_at_lsn } => {
4452 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4453 0 : *lsn = std::cmp::min(*lsn, to);
4454 0 : });
4455 : }
4456 : }
4457 114263 : }
4458 : }
4459 :
4460 : impl std::fmt::Display for VersionedKeySpaceQuery {
4461 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4462 0 : write!(f, "[")?;
4463 :
4464 0 : match self {
4465 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4466 0 : write!(f, "{keyspace} @ {lsn}")?;
4467 : }
4468 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4469 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4470 0 : write!(f, "{keyspace} @ {lsn},")?;
4471 : }
4472 : }
4473 : }
4474 :
4475 0 : write!(f, "]")
4476 0 : }
4477 : }
4478 :
4479 : impl Timeline {
4480 : #[allow(clippy::doc_lazy_continuation)]
4481 : /// Get the data needed to reconstruct all keys in the provided keyspace
4482 : ///
4483 : /// The algorithm is as follows:
4484 : /// 1. While some keys are still not done and there's a timeline to visit:
4485 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4486 : /// 2.1: Build the fringe for the current keyspace
4487 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4488 : /// intersects
4489 : /// 2.3. Pop the timeline from the fringe
4490 : /// 2.4. If the fringe is empty, go back to 1
4491 312408 : async fn get_vectored_reconstruct_data(
4492 312408 : &self,
4493 312408 : mut query: VersionedKeySpaceQuery,
4494 312408 : reconstruct_state: &mut ValuesReconstructState,
4495 312408 : ctx: &RequestContext,
4496 312408 : ) -> Result<(), GetVectoredError> {
4497 312408 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4498 :
4499 : let mut timeline_owned: Arc<Timeline>;
4500 312408 : let mut timeline = self;
4501 :
4502 312407 : let missing_keyspace = loop {
4503 426670 : if self.cancel.is_cancelled() {
4504 0 : return Err(GetVectoredError::Cancelled);
4505 426670 : }
4506 :
4507 : let TimelineVisitOutcome {
4508 426670 : completed_keyspace: completed,
4509 426670 : image_covered_keyspace,
4510 : } = {
4511 426670 : let ctx = RequestContextBuilder::from(ctx)
4512 426670 : .perf_span(|crnt_perf_span| {
4513 0 : info_span!(
4514 : target: PERF_TRACE_TARGET,
4515 0 : parent: crnt_perf_span,
4516 : "PLAN_IO_TIMELINE",
4517 : timeline = %timeline.timeline_id,
4518 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4519 : )
4520 0 : })
4521 426670 : .attached_child();
4522 :
4523 426670 : Self::get_vectored_reconstruct_data_timeline(
4524 426670 : timeline,
4525 426670 : &query,
4526 426670 : reconstruct_state,
4527 426670 : &self.cancel,
4528 426670 : &ctx,
4529 : )
4530 426670 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4531 426670 : .await?
4532 : };
4533 :
4534 426670 : query.remove_overlapping_with(&completed);
4535 :
4536 : // Do not descend into the ancestor timeline for aux files.
4537 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4538 : // stalling compaction.
4539 426670 : query.remove_overlapping_with(&KeySpace {
4540 426670 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4541 426670 : });
4542 :
4543 : // Keyspace is fully retrieved
4544 426670 : if query.is_empty() {
4545 312224 : break None;
4546 114446 : }
4547 :
4548 114446 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4549 : // Not fully retrieved but no ancestor timeline.
4550 183 : break Some(query.total_keyspace());
4551 : };
4552 :
4553 : // Now we see if there are keys covered by the image layer but does not exist in the
4554 : // image layer, which means that the key does not exist.
4555 :
4556 : // The block below will stop the vectored search if any of the keys encountered an image layer
4557 : // which did not contain a snapshot for said key. Since we have already removed all completed
4558 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4559 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4560 : // and stop the search as a result of that.
4561 114263 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4562 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4563 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4564 : // figuring out what is the inherited key range and do a fine-grained pruning.
4565 114263 : removed.remove_overlapping_with(&KeySpace {
4566 114263 : ranges: vec![SPARSE_RANGE],
4567 114263 : });
4568 114263 : if !removed.is_empty() {
4569 0 : break Some(removed);
4570 114263 : }
4571 :
4572 : // Each key range in the original query is at some point in the LSN space.
4573 : // When descending into the ancestor, lower all ranges in the LSN space
4574 : // such that new changes on the parent timeline are not visible.
4575 114263 : query.lower(timeline.ancestor_lsn);
4576 :
4577 114263 : let ctx = RequestContextBuilder::from(ctx)
4578 114263 : .perf_span(|crnt_perf_span| {
4579 0 : info_span!(
4580 : target: PERF_TRACE_TARGET,
4581 0 : parent: crnt_perf_span,
4582 : "GET_ANCESTOR",
4583 : timeline = %timeline.timeline_id,
4584 0 : ancestor = %ancestor_timeline.timeline_id,
4585 : ancestor_lsn = %timeline.ancestor_lsn
4586 : )
4587 0 : })
4588 114263 : .attached_child();
4589 :
4590 114263 : timeline_owned = timeline
4591 114263 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4592 114263 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4593 114263 : .await?;
4594 114262 : timeline = &*timeline_owned;
4595 : };
4596 :
4597 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4598 312407 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4599 183 : let mut missing_keyspace = missing_keyspace;
4600 183 : missing_keyspace.remove_overlapping_with(&KeySpace {
4601 183 : ranges: vec![SPARSE_RANGE],
4602 183 : });
4603 183 : if missing_keyspace.is_empty() {
4604 176 : None
4605 : } else {
4606 7 : Some(missing_keyspace)
4607 : }
4608 : } else {
4609 312224 : None
4610 : };
4611 :
4612 312407 : if let Some(missing_keyspace) = missing_keyspace {
4613 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4614 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4615 7 : shard: self.shard_identity.number,
4616 7 : original_hwm_lsn,
4617 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4618 7 : backtrace: None,
4619 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4620 7 : query: None,
4621 7 : })));
4622 312400 : }
4623 :
4624 312400 : Ok(())
4625 312408 : }
4626 :
4627 426670 : async fn get_vectored_init_fringe(
4628 426670 : &self,
4629 426670 : query: &VersionedKeySpaceQuery,
4630 426670 : ) -> Result<LayerFringe, GetVectoredError> {
4631 426670 : let mut fringe = LayerFringe::new();
4632 426670 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4633 :
4634 426670 : match query {
4635 416478 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4636 : // LSNs requested by the compute or determined by the pageserver
4637 : // are inclusive. Queries to the layer map use exclusive LSNs.
4638 : // Hence, bump the value before the query - same in the other
4639 : // match arm.
4640 416478 : let cont_lsn = Lsn(lsn.0 + 1);
4641 416478 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4642 : }
4643 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4644 21108 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4645 21108 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4646 21108 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4647 : }
4648 : }
4649 : }
4650 :
4651 426670 : Ok(fringe)
4652 426670 : }
4653 :
4654 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4655 : ///
4656 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4657 : /// the current keyspace. The current keyspace of the search at any given timeline
4658 : /// is the original keyspace minus all the keys that have been completed minus
4659 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4660 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4661 : ///
4662 : /// This is basically a depth-first search visitor implementation where a vertex
4663 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4664 : ///
4665 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4666 : /// and get all the required reconstruct data from the layer in one go.
4667 : ///
4668 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4669 : /// decides how to deal with these two keyspaces.
4670 426670 : async fn get_vectored_reconstruct_data_timeline(
4671 426670 : timeline: &Timeline,
4672 426670 : query: &VersionedKeySpaceQuery,
4673 426670 : reconstruct_state: &mut ValuesReconstructState,
4674 426670 : cancel: &CancellationToken,
4675 426670 : ctx: &RequestContext,
4676 426670 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4677 : // Prevent GC from progressing while visiting the current timeline.
4678 : // If we are GC-ing because a new image layer was added while traversing
4679 : // the timeline, then it will remove layers that are required for fulfilling
4680 : // the current get request (read-path cannot "look back" and notice the new
4681 : // image layer).
4682 426670 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4683 :
4684 : // See `compaction::compact_with_gc` for why we need this.
4685 426670 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4686 :
4687 : // Initialize the fringe
4688 426670 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4689 :
4690 426670 : let mut completed_keyspace = KeySpace::default();
4691 426670 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4692 :
4693 873723 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4694 447053 : if cancel.is_cancelled() {
4695 0 : return Err(GetVectoredError::Cancelled);
4696 447053 : }
4697 :
4698 447053 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4699 447053 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4700 447053 : }
4701 :
4702 : // Visit the layer and plan IOs for it
4703 447053 : let next_cont_lsn = lsn_range.start;
4704 447053 : layer_to_read
4705 447053 : .get_values_reconstruct_data(
4706 447053 : keyspace_to_read.clone(),
4707 447053 : lsn_range,
4708 447053 : reconstruct_state,
4709 447053 : ctx,
4710 447053 : )
4711 447053 : .await?;
4712 :
4713 447053 : let mut unmapped_keyspace = keyspace_to_read;
4714 447053 : let cont_lsn = next_cont_lsn;
4715 :
4716 447053 : reconstruct_state.on_layer_visited(&layer_to_read);
4717 :
4718 447053 : let (keys_done_last_step, keys_with_image_coverage) =
4719 447053 : reconstruct_state.consume_done_keys();
4720 447053 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4721 447053 : completed_keyspace.merge(&keys_done_last_step);
4722 447053 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4723 14233 : unmapped_keyspace
4724 14233 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4725 14233 : image_covered_keyspace.add_range(keys_with_image_coverage);
4726 432820 : }
4727 :
4728 : // Query the layer map for the next layers to read.
4729 : //
4730 : // Do not descent any further if the last layer we visited
4731 : // completed all keys in the keyspace it inspected. This is not
4732 : // required for correctness, but avoids visiting extra layers
4733 : // which turns out to be a perf bottleneck in some cases.
4734 447053 : if !unmapped_keyspace.is_empty() {
4735 130311 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4736 130311 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4737 :
4738 : // It's safe to drop the layer map lock after planning the next round of reads.
4739 : // The fringe keeps readable handles for the layers which are safe to read even
4740 : // if layers were compacted or flushed.
4741 : //
4742 : // The more interesting consideration is: "Why is the read algorithm still correct
4743 : // if the layer map changes while it is operating?". Doing a vectored read on a
4744 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4745 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4746 : // range at *a particular point in time*. It is fine for the answer to be different
4747 : // at two different time points.
4748 130311 : drop(guard);
4749 316742 : }
4750 : }
4751 :
4752 426670 : Ok(TimelineVisitOutcome {
4753 426670 : completed_keyspace,
4754 426670 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4755 426670 : })
4756 426670 : }
4757 :
4758 114263 : async fn get_ready_ancestor_timeline(
4759 114263 : &self,
4760 114263 : ancestor: &Arc<Timeline>,
4761 114263 : ctx: &RequestContext,
4762 114263 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4763 : // It's possible that the ancestor timeline isn't active yet, or
4764 : // is active but hasn't yet caught up to the branch point. Wait
4765 : // for it.
4766 : //
4767 : // This cannot happen while the pageserver is running normally,
4768 : // because you cannot create a branch from a point that isn't
4769 : // present in the pageserver yet. However, we don't wait for the
4770 : // branch point to be uploaded to cloud storage before creating
4771 : // a branch. I.e., the branch LSN need not be remote consistent
4772 : // for the branching operation to succeed.
4773 : //
4774 : // Hence, if we try to load a tenant in such a state where
4775 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4776 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4777 : // then we will need to wait for the ancestor timeline to
4778 : // re-stream WAL up to branch_lsn before we access it.
4779 : //
4780 : // How can a tenant get in such a state?
4781 : // - ungraceful pageserver process exit
4782 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4783 : //
4784 : // NB: this could be avoided by requiring
4785 : // branch_lsn >= remote_consistent_lsn
4786 : // during branch creation.
4787 114263 : match ancestor.wait_to_become_active(ctx).await {
4788 114262 : Ok(()) => {}
4789 : Err(TimelineState::Stopping) => {
4790 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4791 0 : return Err(GetReadyAncestorError::Cancelled);
4792 : }
4793 1 : Err(state) => {
4794 1 : return Err(GetReadyAncestorError::BadState {
4795 1 : timeline_id: ancestor.timeline_id,
4796 1 : state,
4797 1 : });
4798 : }
4799 : }
4800 114262 : ancestor
4801 114262 : .wait_lsn(
4802 114262 : self.ancestor_lsn,
4803 114262 : WaitLsnWaiter::Timeline(self),
4804 114262 : WaitLsnTimeout::Default,
4805 114262 : ctx,
4806 114262 : )
4807 114262 : .await
4808 114262 : .map_err(|e| match e {
4809 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4810 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4811 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4812 0 : timeline_id: ancestor.timeline_id,
4813 0 : state,
4814 0 : },
4815 0 : })?;
4816 :
4817 114262 : Ok(ancestor.clone())
4818 114263 : }
4819 :
4820 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4821 148592 : &self.shard_identity
4822 148592 : }
4823 :
4824 : #[inline(always)]
4825 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4826 0 : ShardTimelineId {
4827 0 : shard_index: ShardIndex {
4828 0 : shard_number: self.shard_identity.number,
4829 0 : shard_count: self.shard_identity.count,
4830 0 : },
4831 0 : timeline_id: self.timeline_id,
4832 0 : }
4833 0 : }
4834 :
4835 : /// Returns a non-frozen open in-memory layer for ingestion.
4836 : ///
4837 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4838 : /// this function without holding the mutex.
4839 660 : async fn get_layer_for_write(
4840 660 : &self,
4841 660 : lsn: Lsn,
4842 660 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4843 660 : ctx: &RequestContext,
4844 660 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4845 660 : let mut guard = self
4846 660 : .layers
4847 660 : .write(LayerManagerLockHolder::GetLayerForWrite)
4848 660 : .await;
4849 :
4850 660 : let last_record_lsn = self.get_last_record_lsn();
4851 660 : ensure!(
4852 660 : lsn > last_record_lsn,
4853 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4854 : lsn,
4855 : last_record_lsn,
4856 : );
4857 :
4858 660 : let layer = guard
4859 660 : .open_mut()?
4860 660 : .get_layer_for_write(
4861 660 : lsn,
4862 660 : self.conf,
4863 660 : self.timeline_id,
4864 660 : self.tenant_shard_id,
4865 660 : &self.gate,
4866 660 : &self.cancel,
4867 660 : ctx,
4868 : )
4869 660 : .await?;
4870 660 : Ok(layer)
4871 660 : }
4872 :
4873 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4874 2639560 : assert!(new_lsn.is_aligned());
4875 :
4876 2639560 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4877 2639560 : self.last_record_lsn.advance(new_lsn);
4878 2639560 : }
4879 :
4880 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4881 : ///
4882 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4883 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4884 611 : async fn freeze_inmem_layer_at(
4885 611 : &self,
4886 611 : at: Lsn,
4887 611 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4888 611 : ) -> Result<u64, FlushLayerError> {
4889 611 : let frozen = {
4890 611 : let mut guard = self
4891 611 : .layers
4892 611 : .write(LayerManagerLockHolder::TryFreezeLayer)
4893 611 : .await;
4894 611 : guard
4895 611 : .open_mut()?
4896 611 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4897 611 : .await
4898 : };
4899 :
4900 611 : if frozen {
4901 597 : let now = Instant::now();
4902 597 : *(self.last_freeze_ts.write().unwrap()) = now;
4903 597 : }
4904 :
4905 : // Increment the flush cycle counter and wake up the flush task.
4906 : // Remember the new value, so that when we listen for the flush
4907 : // to finish, we know when the flush that we initiated has
4908 : // finished, instead of some other flush that was started earlier.
4909 611 : let mut my_flush_request = 0;
4910 :
4911 611 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4912 611 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4913 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4914 611 : }
4915 :
4916 611 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4917 611 : my_flush_request = *counter + 1;
4918 611 : *counter = my_flush_request;
4919 611 : *lsn = std::cmp::max(at, *lsn);
4920 611 : });
4921 :
4922 611 : assert_ne!(my_flush_request, 0);
4923 :
4924 611 : Ok(my_flush_request)
4925 611 : }
4926 :
4927 : /// Layer flusher task's main loop.
4928 232 : async fn flush_loop(
4929 232 : self: &Arc<Self>,
4930 232 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4931 232 : ctx: &RequestContext,
4932 232 : ) {
4933 : // Always notify waiters about the flush loop exiting since the loop might stop
4934 : // when the timeline hasn't been cancelled.
4935 232 : let scopeguard_rx = layer_flush_start_rx.clone();
4936 232 : scopeguard::defer! {
4937 : let (flush_counter, _) = *scopeguard_rx.borrow();
4938 : let _ = self
4939 : .layer_flush_done_tx
4940 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4941 : }
4942 :
4943 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4944 232 : let mut watch_l0 = match self
4945 232 : .layers
4946 232 : .read(LayerManagerLockHolder::FlushLoop)
4947 232 : .await
4948 232 : .layer_map()
4949 : {
4950 232 : Ok(lm) => lm.watch_level0_deltas(),
4951 0 : Err(Shutdown) => return,
4952 : };
4953 :
4954 232 : info!("started flush loop");
4955 : loop {
4956 834 : tokio::select! {
4957 834 : _ = self.cancel.cancelled() => {
4958 5 : info!("shutting down layer flush task due to Timeline::cancel");
4959 5 : break;
4960 : },
4961 834 : _ = layer_flush_start_rx.changed() => {}
4962 : }
4963 602 : trace!("waking up");
4964 602 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4965 :
4966 : // The highest LSN to which we flushed in the loop over frozen layers
4967 602 : let mut flushed_to_lsn = Lsn(0);
4968 :
4969 602 : let result = loop {
4970 1198 : if self.cancel.is_cancelled() {
4971 0 : info!("dropping out of flush loop for timeline shutdown");
4972 0 : return;
4973 1198 : }
4974 :
4975 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4976 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4977 1198 : if flushed_to_lsn >= frozen_to_lsn {
4978 587 : break Ok(());
4979 611 : }
4980 :
4981 : // Fetch the next layer to flush, if any.
4982 611 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4983 611 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4984 611 : let Ok(lm) = layers.layer_map() else {
4985 0 : info!("dropping out of flush loop for timeline shutdown");
4986 0 : return;
4987 : };
4988 611 : let l0_count = lm.level0_deltas().len();
4989 611 : let frozen_count = lm.frozen_layers.len();
4990 611 : let frozen_size: u64 = lm
4991 611 : .frozen_layers
4992 611 : .iter()
4993 611 : .map(|l| l.estimated_in_mem_size())
4994 611 : .sum();
4995 611 : let open_layer_size: u64 = lm
4996 611 : .open_layer
4997 611 : .as_ref()
4998 611 : .map(|l| l.estimated_in_mem_size())
4999 611 : .unwrap_or(0);
5000 611 : let layer = lm.frozen_layers.front().cloned();
5001 611 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
5002 : // drop 'layers' lock
5003 : };
5004 611 : let Some(layer) = layer else {
5005 14 : break Ok(());
5006 : };
5007 :
5008 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
5009 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
5010 597 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
5011 0 : if l0_count >= stall_threshold {
5012 0 : warn!(
5013 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
5014 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5015 : );
5016 0 : let stall_timer = self
5017 0 : .metrics
5018 0 : .flush_delay_histo
5019 0 : .start_timer()
5020 0 : .record_on_drop();
5021 0 : tokio::select! {
5022 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
5023 0 : if let Ok(l0) = result.as_deref() {
5024 0 : let delay = stall_timer.elapsed().as_secs_f64();
5025 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
5026 0 : }
5027 : },
5028 0 : _ = self.cancel.cancelled() => {},
5029 : }
5030 0 : continue; // check again
5031 0 : }
5032 597 : }
5033 :
5034 : // Flush the layer.
5035 597 : let flush_timer = self.metrics.flush_time_histo.start_timer();
5036 597 : match self.flush_frozen_layer(layer, ctx).await {
5037 596 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
5038 : Err(FlushLayerError::Cancelled) => {
5039 0 : info!("dropping out of flush loop for timeline shutdown");
5040 0 : return;
5041 : }
5042 1 : err @ Err(
5043 : FlushLayerError::NotRunning(_)
5044 : | FlushLayerError::Other(_)
5045 : | FlushLayerError::CreateImageLayersError(_),
5046 : ) => {
5047 1 : error!("could not flush frozen layer: {err:?}");
5048 1 : break err.map(|_| ());
5049 : }
5050 : }
5051 596 : let flush_duration = flush_timer.stop_and_record();
5052 :
5053 : // Notify the tenant compaction loop if L0 compaction is needed.
5054 596 : let l0_count = *watch_l0.borrow();
5055 596 : if l0_count >= self.get_compaction_threshold() {
5056 238 : self.l0_compaction_trigger.notify_one();
5057 358 : }
5058 :
5059 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
5060 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
5061 : // ingestion by having ephemeral layer rolls wait for flushes.
5062 596 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
5063 4 : if l0_count >= delay_threshold {
5064 0 : let delay = flush_duration.as_secs_f64();
5065 0 : info!(
5066 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
5067 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5068 : );
5069 0 : let _delay_timer = self
5070 0 : .metrics
5071 0 : .flush_delay_histo
5072 0 : .start_timer()
5073 0 : .record_on_drop();
5074 0 : tokio::select! {
5075 0 : _ = tokio::time::sleep(flush_duration) => {},
5076 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
5077 0 : _ = self.cancel.cancelled() => {},
5078 : }
5079 4 : }
5080 592 : }
5081 : };
5082 :
5083 : // Unsharded tenants should never advance their LSN beyond the end of the
5084 : // highest layer they write: such gaps between layer data and the frozen LSN
5085 : // are only legal on sharded tenants.
5086 602 : debug_assert!(
5087 602 : self.shard_identity.count.count() > 1
5088 595 : || flushed_to_lsn >= frozen_to_lsn
5089 14 : || !flushed_to_lsn.is_valid()
5090 : );
5091 :
5092 602 : if flushed_to_lsn < frozen_to_lsn
5093 15 : && self.shard_identity.count.count() > 1
5094 1 : && result.is_ok()
5095 : {
5096 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
5097 : // to us via layer_flush_start_rx, then advance it here.
5098 : //
5099 : // This path is only taken for tenants with multiple shards: single sharded tenants should
5100 : // never encounter a gap in the wal.
5101 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
5102 0 : tracing::debug!(
5103 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
5104 : );
5105 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
5106 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
5107 0 : tracing::warn!(
5108 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
5109 : );
5110 0 : }
5111 0 : }
5112 602 : }
5113 :
5114 : // Notify any listeners that we're done
5115 602 : let _ = self
5116 602 : .layer_flush_done_tx
5117 602 : .send_replace((flush_counter, result));
5118 : }
5119 5 : }
5120 :
5121 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
5122 571 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
5123 571 : let mut rx = self.layer_flush_done_tx.subscribe();
5124 : loop {
5125 : {
5126 1162 : let (last_result_counter, last_result) = &*rx.borrow();
5127 1162 : if *last_result_counter >= request {
5128 571 : if let Err(err) = last_result {
5129 : // We already logged the original error in
5130 : // flush_loop. We cannot propagate it to the caller
5131 : // here, because it might not be Cloneable
5132 1 : return Err(err.clone());
5133 : } else {
5134 570 : return Ok(());
5135 : }
5136 591 : }
5137 : }
5138 591 : trace!("waiting for flush to complete");
5139 591 : tokio::select! {
5140 591 : rx_e = rx.changed() => {
5141 591 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
5142 : },
5143 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
5144 : // the notification from [`flush_loop`] that it completed.
5145 591 : _ = self.cancel.cancelled() => {
5146 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
5147 0 : return Ok(())
5148 : }
5149 : };
5150 591 : trace!("done")
5151 : }
5152 571 : }
5153 :
5154 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
5155 : ///
5156 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
5157 : #[instrument(skip_all, fields(layer=%frozen_layer))]
5158 : async fn flush_frozen_layer(
5159 : self: &Arc<Self>,
5160 : frozen_layer: Arc<InMemoryLayer>,
5161 : ctx: &RequestContext,
5162 : ) -> Result<Lsn, FlushLayerError> {
5163 : debug_assert_current_span_has_tenant_and_timeline_id();
5164 :
5165 : // As a special case, when we have just imported an image into the repository,
5166 : // instead of writing out a L0 delta layer, we directly write out image layer
5167 : // files instead. This is possible as long as *all* the data imported into the
5168 : // repository have the same LSN.
5169 : let lsn_range = frozen_layer.get_lsn_range();
5170 :
5171 : // Whether to directly create image layers for this flush, or flush them as delta layers
5172 : let create_image_layer =
5173 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
5174 :
5175 : #[cfg(test)]
5176 : {
5177 : match &mut *self.flush_loop_state.lock().unwrap() {
5178 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
5179 : panic!("flush loop not running")
5180 : }
5181 : FlushLoopState::Running {
5182 : expect_initdb_optimization,
5183 : initdb_optimization_count,
5184 : ..
5185 : } => {
5186 : if create_image_layer {
5187 : *initdb_optimization_count += 1;
5188 : } else {
5189 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
5190 : }
5191 : }
5192 : }
5193 : }
5194 :
5195 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5196 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5197 : // require downloading anything during initial import.
5198 : let ((rel_partition, metadata_partition), _lsn) = self
5199 : .repartition(
5200 : self.initdb_lsn,
5201 : self.get_compaction_target_size(),
5202 : EnumSet::empty(),
5203 : ctx,
5204 : )
5205 : .await
5206 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5207 :
5208 : if self.cancel.is_cancelled() {
5209 : return Err(FlushLayerError::Cancelled);
5210 : }
5211 :
5212 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5213 : // So that the key ranges don't overlap.
5214 : let mut partitions = KeyPartitioning::default();
5215 : partitions.parts.extend(rel_partition.parts);
5216 : if !metadata_partition.parts.is_empty() {
5217 : assert_eq!(
5218 : metadata_partition.parts.len(),
5219 : 1,
5220 : "currently sparse keyspace should only contain a single metadata keyspace"
5221 : );
5222 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5223 : // every single key within the keyspace, and therefore, it's safe to force converting it
5224 : // into a dense keyspace before calling this function.
5225 : partitions
5226 : .parts
5227 : .extend(metadata_partition.into_dense().parts);
5228 : }
5229 :
5230 : let mut layers_to_upload = Vec::new();
5231 : let (generated_image_layers, is_complete) = self
5232 : .create_image_layers(
5233 : &partitions,
5234 : self.initdb_lsn,
5235 : None,
5236 : ImageLayerCreationMode::Initial,
5237 : ctx,
5238 : LastImageLayerCreationStatus::Initial,
5239 : false, // don't yield for L0, we're flushing L0
5240 : )
5241 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5242 : .await?;
5243 : debug_assert!(
5244 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5245 : "init image generation mode must fully cover the keyspace"
5246 : );
5247 : layers_to_upload.extend(generated_image_layers);
5248 :
5249 : (layers_to_upload, None)
5250 : } else {
5251 : // Normal case, write out a L0 delta layer file.
5252 : // `create_delta_layer` will not modify the layer map.
5253 : // We will remove frozen layer and add delta layer in one atomic operation later.
5254 : let Some(layer) = self
5255 : .create_delta_layer(&frozen_layer, None, ctx)
5256 : .await
5257 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5258 : else {
5259 : panic!("delta layer cannot be empty if no filter is applied");
5260 : };
5261 : (
5262 : // FIXME: even though we have a single image and single delta layer assumption
5263 : // we push them to vec
5264 : vec![layer.clone()],
5265 : Some(layer),
5266 : )
5267 : };
5268 :
5269 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5270 :
5271 : if self.cancel.is_cancelled() {
5272 : return Err(FlushLayerError::Cancelled);
5273 : }
5274 :
5275 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5276 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5277 1 : });
5278 :
5279 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5280 :
5281 : // The new on-disk layers are now in the layer map. We can remove the
5282 : // in-memory layer from the map now. The flushed layer is stored in
5283 : // the mapping in `create_delta_layer`.
5284 : {
5285 : let mut guard = self
5286 : .layers
5287 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5288 : .await;
5289 :
5290 : guard.open_mut()?.finish_flush_l0_layer(
5291 : delta_layer_to_add.as_ref(),
5292 : &frozen_layer,
5293 : &self.metrics,
5294 : );
5295 :
5296 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5297 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5298 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5299 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5300 : }
5301 : // release lock on 'layers'
5302 : };
5303 :
5304 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5305 : // a compaction can delete the file and then it won't be available for uploads any more.
5306 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5307 : // race situation.
5308 : // See https://github.com/neondatabase/neon/issues/4526
5309 : pausable_failpoint!("flush-frozen-pausable");
5310 :
5311 : // This failpoint is used by another test case `test_pageserver_recovery`.
5312 : fail_point!("flush-frozen-exit");
5313 :
5314 : Ok(Lsn(lsn_range.end.0 - 1))
5315 : }
5316 :
5317 : /// Return true if the value changed
5318 : ///
5319 : /// This function must only be used from the layer flush task.
5320 596 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5321 596 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5322 596 : assert!(
5323 596 : new_value >= old_value,
5324 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5325 : );
5326 :
5327 596 : self.metrics
5328 596 : .disk_consistent_lsn_gauge
5329 596 : .set(new_value.0 as i64);
5330 596 : new_value != old_value
5331 596 : }
5332 :
5333 : /// Update metadata file
5334 623 : fn schedule_uploads(
5335 623 : &self,
5336 623 : disk_consistent_lsn: Lsn,
5337 623 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5338 623 : ) -> anyhow::Result<()> {
5339 : // We can only save a valid 'prev_record_lsn' value on disk if we
5340 : // flushed *all* in-memory changes to disk. We only track
5341 : // 'prev_record_lsn' in memory for the latest processed record, so we
5342 : // don't remember what the correct value that corresponds to some old
5343 : // LSN is. But if we flush everything, then the value corresponding
5344 : // current 'last_record_lsn' is correct and we can store it on disk.
5345 : let RecordLsn {
5346 623 : last: last_record_lsn,
5347 623 : prev: prev_record_lsn,
5348 623 : } = self.last_record_lsn.load();
5349 623 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5350 559 : Some(prev_record_lsn)
5351 : } else {
5352 64 : None
5353 : };
5354 :
5355 623 : let update = crate::tenant::metadata::MetadataUpdate::new(
5356 623 : disk_consistent_lsn,
5357 623 : ondisk_prev_record_lsn,
5358 623 : *self.applied_gc_cutoff_lsn.read(),
5359 : );
5360 :
5361 623 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5362 0 : "{}",
5363 0 : x.unwrap()
5364 : ));
5365 :
5366 1225 : for layer in layers_to_upload {
5367 602 : self.remote_client.schedule_layer_file_upload(layer)?;
5368 : }
5369 623 : self.remote_client
5370 623 : .schedule_index_upload_for_metadata_update(&update)?;
5371 :
5372 623 : Ok(())
5373 623 : }
5374 :
5375 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5376 0 : self.remote_client
5377 0 : .preserve_initdb_archive(
5378 0 : &self.tenant_shard_id.tenant_id,
5379 0 : &self.timeline_id,
5380 0 : &self.cancel,
5381 0 : )
5382 0 : .await
5383 0 : }
5384 :
5385 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5386 : // in layer map immediately. The caller is responsible to put it into the layer map.
5387 486 : async fn create_delta_layer(
5388 486 : self: &Arc<Self>,
5389 486 : frozen_layer: &Arc<InMemoryLayer>,
5390 486 : key_range: Option<Range<Key>>,
5391 486 : ctx: &RequestContext,
5392 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5393 486 : let self_clone = Arc::clone(self);
5394 486 : let frozen_layer = Arc::clone(frozen_layer);
5395 486 : let ctx = ctx.attached_child();
5396 486 : let work = async move {
5397 486 : let Some((desc, path)) = frozen_layer
5398 486 : .write_to_disk(
5399 486 : &ctx,
5400 486 : key_range,
5401 486 : self_clone.l0_flush_global_state.inner(),
5402 486 : &self_clone.gate,
5403 486 : self_clone.cancel.clone(),
5404 486 : )
5405 486 : .await?
5406 : else {
5407 0 : return Ok(None);
5408 : };
5409 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5410 :
5411 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5412 : // We just need to fsync the directory in which these inodes are linked,
5413 : // which we know to be the timeline directory.
5414 : //
5415 : // We use fatal_err() below because the after write_to_disk returns with success,
5416 : // the in-memory state of the filesystem already has the layer file in its final place,
5417 : // and subsequent pageserver code could think it's durable while it really isn't.
5418 486 : let timeline_dir = VirtualFile::open(
5419 486 : &self_clone
5420 486 : .conf
5421 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5422 486 : &ctx,
5423 486 : )
5424 486 : .await
5425 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5426 486 : timeline_dir
5427 486 : .sync_all()
5428 486 : .await
5429 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5430 486 : anyhow::Ok(Some(new_delta))
5431 486 : };
5432 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5433 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5434 : use crate::virtual_file::io_engine::IoEngine;
5435 486 : match crate::virtual_file::io_engine::get() {
5436 0 : IoEngine::NotSet => panic!("io engine not set"),
5437 : IoEngine::StdFs => {
5438 0 : let span = tracing::info_span!("blocking");
5439 0 : tokio::task::spawn_blocking({
5440 0 : move || Handle::current().block_on(work.instrument(span))
5441 : })
5442 0 : .await
5443 0 : .context("spawn_blocking")
5444 0 : .and_then(|x| x)
5445 : }
5446 : #[cfg(target_os = "linux")]
5447 486 : IoEngine::TokioEpollUring => work.await,
5448 : }
5449 486 : }
5450 :
5451 303 : async fn repartition(
5452 303 : &self,
5453 303 : lsn: Lsn,
5454 303 : partition_size: u64,
5455 303 : flags: EnumSet<CompactFlags>,
5456 303 : ctx: &RequestContext,
5457 303 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5458 303 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5459 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5460 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5461 : // and hence before the compaction task starts.
5462 0 : return Err(RepartitionError::Other(anyhow!(
5463 0 : "repartition() called concurrently"
5464 0 : )));
5465 : };
5466 303 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5467 303 : if lsn < *partition_lsn {
5468 0 : return Err(RepartitionError::Other(anyhow!(
5469 0 : "repartition() called with LSN going backwards, this should not happen"
5470 0 : )));
5471 303 : }
5472 :
5473 303 : let distance = lsn.0 - partition_lsn.0;
5474 303 : if *partition_lsn != Lsn(0)
5475 141 : && distance <= self.repartition_threshold
5476 141 : && !flags.contains(CompactFlags::ForceRepartition)
5477 : {
5478 134 : debug!(
5479 : distance,
5480 : threshold = self.repartition_threshold,
5481 0 : "no repartitioning needed"
5482 : );
5483 134 : return Ok((
5484 134 : (dense_partition.clone(), sparse_partition.clone()),
5485 134 : *partition_lsn,
5486 134 : ));
5487 169 : }
5488 :
5489 169 : let (dense_ks, sparse_ks) = self
5490 169 : .collect_keyspace(lsn, ctx)
5491 169 : .await
5492 169 : .map_err(RepartitionError::CollectKeyspace)?;
5493 169 : let dense_partitioning = dense_ks.partition(
5494 169 : &self.shard_identity,
5495 169 : partition_size,
5496 169 : postgres_ffi::BLCKSZ as u64,
5497 : );
5498 169 : let sparse_partitioning = SparseKeyPartitioning {
5499 169 : parts: vec![sparse_ks],
5500 169 : }; // no partitioning for metadata keys for now
5501 169 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5502 169 : guard.write(result.clone());
5503 169 : Ok(result)
5504 303 : }
5505 :
5506 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5507 57 : async fn time_for_new_image_layer(
5508 57 : &self,
5509 57 : partition: &KeySpace,
5510 57 : lsn: Lsn,
5511 57 : force_image_creation_lsn: Option<Lsn>,
5512 57 : ) -> bool {
5513 57 : let threshold = self.get_image_creation_threshold();
5514 :
5515 57 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5516 57 : let Ok(layers) = guard.layer_map() else {
5517 0 : return false;
5518 : };
5519 57 : let mut min_image_lsn: Lsn = Lsn::MAX;
5520 57 : let mut max_deltas = 0;
5521 364 : for part_range in &partition.ranges {
5522 307 : let image_coverage = layers.image_coverage(part_range, lsn);
5523 614 : for (img_range, last_img) in image_coverage {
5524 307 : let img_lsn = if let Some(last_img) = last_img {
5525 0 : last_img.get_lsn_range().end
5526 : } else {
5527 307 : Lsn(0)
5528 : };
5529 : // Let's consider an example:
5530 : //
5531 : // delta layer with LSN range 71-81
5532 : // delta layer with LSN range 81-91
5533 : // delta layer with LSN range 91-101
5534 : // image layer at LSN 100
5535 : //
5536 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5537 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5538 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5539 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5540 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5541 307 : if img_lsn < lsn {
5542 307 : let num_deltas =
5543 307 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5544 :
5545 307 : max_deltas = max_deltas.max(num_deltas);
5546 307 : if num_deltas >= threshold {
5547 0 : debug!(
5548 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5549 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5550 : );
5551 0 : return true;
5552 307 : }
5553 0 : }
5554 307 : min_image_lsn = min(min_image_lsn, img_lsn);
5555 : }
5556 : }
5557 :
5558 : // HADRON
5559 : // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
5560 57 : min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
5561 57 : if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
5562 0 : info!(
5563 0 : "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
5564 0 : partition.ranges[0].start,
5565 0 : partition.ranges[0].end,
5566 : min_image_lsn,
5567 0 : force_image_creation_lsn.unwrap(),
5568 : max_deltas
5569 : );
5570 0 : return true;
5571 57 : }
5572 :
5573 57 : debug!(
5574 : max_deltas,
5575 0 : "none of the partitioned ranges had >= {threshold} deltas"
5576 : );
5577 57 : false
5578 57 : }
5579 :
5580 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5581 : /// so that at most one image layer will be produced from this function.
5582 : #[allow(clippy::too_many_arguments)]
5583 124 : async fn create_image_layer_for_rel_blocks(
5584 124 : self: &Arc<Self>,
5585 124 : partition: &KeySpace,
5586 124 : mut image_layer_writer: ImageLayerWriter,
5587 124 : lsn: Lsn,
5588 124 : ctx: &RequestContext,
5589 124 : img_range: Range<Key>,
5590 124 : io_concurrency: IoConcurrency,
5591 124 : progress: Option<(usize, usize)>,
5592 124 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5593 124 : let mut wrote_keys = false;
5594 :
5595 124 : let mut key_request_accum = KeySpaceAccum::new();
5596 824 : for range in &partition.ranges {
5597 700 : let mut key = range.start;
5598 1517 : while key < range.end {
5599 : // Decide whether to retain this key: usually we do, but sharded tenants may
5600 : // need to drop keys that don't belong to them. If we retain the key, add it
5601 : // to `key_request_accum` for later issuing a vectored get
5602 817 : if self.shard_identity.is_key_disposable(&key) {
5603 0 : debug!(
5604 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5605 : key,
5606 0 : self.shard_identity.get_shard_number(&key)
5607 : );
5608 817 : } else {
5609 817 : key_request_accum.add_key(key);
5610 817 : }
5611 :
5612 817 : let last_key_in_range = key.next() == range.end;
5613 817 : key = key.next();
5614 :
5615 : // Maybe flush `key_rest_accum`
5616 817 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5617 817 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5618 : {
5619 700 : let query =
5620 700 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5621 :
5622 700 : let results = self
5623 700 : .get_vectored(query, io_concurrency.clone(), ctx)
5624 700 : .await?;
5625 :
5626 700 : if self.cancel.is_cancelled() {
5627 0 : return Err(CreateImageLayersError::Cancelled);
5628 700 : }
5629 :
5630 1517 : for (img_key, img) in results {
5631 817 : let img = match img {
5632 817 : Ok(img) => img,
5633 0 : Err(err) => {
5634 : // If we fail to reconstruct a VM or FSM page, we can zero the
5635 : // page without losing any actual user data. That seems better
5636 : // than failing repeatedly and getting stuck.
5637 : //
5638 : // We had a bug at one point, where we truncated the FSM and VM
5639 : // in the pageserver, but the Postgres didn't know about that
5640 : // and continued to generate incremental WAL records for pages
5641 : // that didn't exist in the pageserver. Trying to replay those
5642 : // WAL records failed to find the previous image of the page.
5643 : // This special case allows us to recover from that situation.
5644 : // See https://github.com/neondatabase/neon/issues/2601.
5645 : //
5646 : // Unfortunately we cannot do this for the main fork, or for
5647 : // any metadata keys, keys, as that would lead to actual data
5648 : // loss.
5649 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5650 0 : warn!(
5651 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5652 : );
5653 0 : ZERO_PAGE.clone()
5654 : } else {
5655 0 : return Err(CreateImageLayersError::from(err));
5656 : }
5657 : }
5658 : };
5659 :
5660 : // Write all the keys we just read into our new image layer.
5661 817 : image_layer_writer.put_image(img_key, img, ctx).await?;
5662 817 : wrote_keys = true;
5663 : }
5664 117 : }
5665 : }
5666 : }
5667 :
5668 124 : let progress_report = progress
5669 124 : .map(|(idx, total)| format!("({idx}/{total}) "))
5670 124 : .unwrap_or_default();
5671 124 : if wrote_keys {
5672 : // Normal path: we have written some data into the new image layer for this
5673 : // partition, so flush it to disk.
5674 124 : info!(
5675 0 : "{} produced image layer for rel {}",
5676 : progress_report,
5677 0 : ImageLayerName {
5678 0 : key_range: img_range.clone(),
5679 0 : lsn
5680 0 : },
5681 : );
5682 124 : Ok(ImageLayerCreationOutcome::Generated {
5683 124 : unfinished_image_layer: image_layer_writer,
5684 124 : })
5685 : } else {
5686 0 : tracing::debug!(
5687 0 : "{} no data in range {}-{}",
5688 : progress_report,
5689 : img_range.start,
5690 : img_range.end
5691 : );
5692 0 : Ok(ImageLayerCreationOutcome::Empty)
5693 : }
5694 124 : }
5695 :
5696 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5697 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5698 : /// would not be too large to fit in a single image layer.
5699 : ///
5700 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5701 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5702 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5703 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5704 : #[allow(clippy::too_many_arguments)]
5705 169 : async fn create_image_layer_for_metadata_keys(
5706 169 : self: &Arc<Self>,
5707 169 : partition: &KeySpace,
5708 169 : mut image_layer_writer: ImageLayerWriter,
5709 169 : lsn: Lsn,
5710 169 : ctx: &RequestContext,
5711 169 : img_range: Range<Key>,
5712 169 : mode: ImageLayerCreationMode,
5713 169 : io_concurrency: IoConcurrency,
5714 169 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5715 : // Metadata keys image layer creation.
5716 169 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5717 169 : let begin = Instant::now();
5718 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5719 : // not contain too many keys, otherwise this takes a lot of memory.
5720 169 : let data = self
5721 169 : .get_vectored_impl(
5722 169 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5723 169 : &mut reconstruct_state,
5724 169 : ctx,
5725 169 : )
5726 169 : .await?;
5727 169 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5728 169 : let mut new_data = BTreeMap::new();
5729 169 : let mut total_kb_retrieved = 0;
5730 169 : let mut total_keys_retrieved = 0;
5731 5175 : for (k, v) in data {
5732 5006 : let v = v?;
5733 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5734 5006 : total_keys_retrieved += 1;
5735 5006 : new_data.insert(k, v);
5736 : }
5737 169 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5738 : };
5739 169 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5740 169 : let elapsed = begin.elapsed();
5741 :
5742 169 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5743 169 : info!(
5744 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",
5745 0 : elapsed.as_secs_f64()
5746 : );
5747 :
5748 169 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5749 16 : return Ok(ImageLayerCreationOutcome::Skip);
5750 153 : }
5751 153 : if self.cancel.is_cancelled() {
5752 0 : return Err(CreateImageLayersError::Cancelled);
5753 153 : }
5754 153 : let mut wrote_any_image = false;
5755 5159 : for (k, v) in data {
5756 5006 : if v.is_empty() {
5757 : // the key has been deleted, it does not need an image
5758 : // in metadata keyspace, an empty image == tombstone
5759 4 : continue;
5760 5002 : }
5761 5002 : wrote_any_image = true;
5762 :
5763 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5764 :
5765 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5766 : // on the normal data path either.
5767 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5768 : }
5769 :
5770 153 : if wrote_any_image {
5771 : // Normal path: we have written some data into the new image layer for this
5772 : // partition, so flush it to disk.
5773 6 : info!(
5774 0 : "created image layer for metadata {}",
5775 0 : ImageLayerName {
5776 0 : key_range: img_range.clone(),
5777 0 : lsn
5778 0 : }
5779 : );
5780 6 : Ok(ImageLayerCreationOutcome::Generated {
5781 6 : unfinished_image_layer: image_layer_writer,
5782 6 : })
5783 : } else {
5784 147 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5785 147 : Ok(ImageLayerCreationOutcome::Empty)
5786 : }
5787 169 : }
5788 :
5789 : /// Predicate function which indicates whether we should check if new image layers
5790 : /// are required. Since checking if new image layers are required is expensive in
5791 : /// terms of CPU, we only do it in the following cases:
5792 : /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
5793 : /// 2. If enough time has passed since the last check:
5794 : /// 1. For large tenants, we wish to perform the check more often since they
5795 : /// suffer from the lack of image layers. Note that we assume sharded tenants
5796 : /// to be large since non-zero shards do not track the logical size.
5797 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5798 191 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5799 191 : let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
5800 :
5801 191 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5802 191 : let distance = lsn
5803 191 : .checked_sub(last_checks_at)
5804 191 : .expect("Attempt to compact with LSN going backwards");
5805 191 : let min_distance =
5806 191 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5807 :
5808 191 : let distance_based_decision = distance.0 >= min_distance;
5809 :
5810 191 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5811 191 : let check_required_after = (|| {
5812 191 : if self.shard_identity.is_unsharded() {
5813 135 : if let CurrentLogicalSize::Exact(logical_size) =
5814 186 : self.current_logical_size.current_size()
5815 : {
5816 135 : if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
5817 135 : return Duration::from_secs(3600 * 48);
5818 0 : }
5819 51 : }
5820 5 : }
5821 :
5822 56 : self.get_checkpoint_timeout()
5823 : })();
5824 :
5825 191 : let time_based_decision = match *last_check_instant {
5826 29 : Some(last_check) => {
5827 29 : let elapsed = last_check.elapsed();
5828 29 : elapsed >= check_required_after
5829 : }
5830 162 : None => true,
5831 : };
5832 :
5833 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5834 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5835 : // check.
5836 191 : let decision = distance_based_decision || time_based_decision;
5837 191 : tracing::info!(
5838 0 : "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
5839 : decision,
5840 : distance_based_decision,
5841 : time_based_decision
5842 : );
5843 191 : if decision {
5844 162 : self.last_image_layer_creation_check_at.store(lsn);
5845 162 : *last_check_instant = Some(Instant::now());
5846 162 : }
5847 :
5848 191 : decision
5849 191 : }
5850 :
5851 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5852 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5853 : ///
5854 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5855 : #[allow(clippy::too_many_arguments)]
5856 191 : async fn create_image_layers(
5857 191 : self: &Arc<Timeline>,
5858 191 : partitioning: &KeyPartitioning,
5859 191 : lsn: Lsn,
5860 191 : force_image_creation_lsn: Option<Lsn>,
5861 191 : mode: ImageLayerCreationMode,
5862 191 : ctx: &RequestContext,
5863 191 : last_status: LastImageLayerCreationStatus,
5864 191 : yield_for_l0: bool,
5865 191 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5866 191 : let timer = self.metrics.create_images_time_histo.start_timer();
5867 :
5868 191 : if partitioning.parts.is_empty() {
5869 0 : warn!("no partitions to create image layers for");
5870 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5871 191 : }
5872 :
5873 : // We need to avoid holes between generated image layers.
5874 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5875 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5876 : //
5877 : // How such hole between partitions can appear?
5878 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5879 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5880 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5881 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5882 191 : let mut start = Key::MIN;
5883 :
5884 191 : let check_for_image_layers =
5885 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5886 0 : info!(
5887 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5888 : last_key
5889 : );
5890 0 : true
5891 : } else {
5892 191 : self.should_check_if_image_layers_required(lsn)
5893 : };
5894 :
5895 191 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5896 :
5897 191 : let mut all_generated = true;
5898 :
5899 191 : let mut partition_processed = 0;
5900 191 : let mut total_partitions = partitioning.parts.len();
5901 191 : let mut last_partition_processed = None;
5902 191 : let mut partition_parts = partitioning.parts.clone();
5903 :
5904 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5905 : // We need to skip the partitions that have already been processed.
5906 0 : let mut found = false;
5907 0 : for (i, partition) in partition_parts.iter().enumerate() {
5908 0 : if last_key <= partition.end().unwrap() {
5909 : // ```plain
5910 : // |------|--------|----------|------|
5911 : // ^last_key
5912 : // ^start from this partition
5913 : // ```
5914 : // Why `i+1` instead of `i`?
5915 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5916 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5917 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5918 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5919 0 : if i + 1 >= total_partitions {
5920 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5921 : // iteration of image layer creation should return a complete status.
5922 0 : break; // with found=false
5923 0 : }
5924 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5925 0 : total_partitions = partition_parts.len();
5926 : // Update the start key to the partition start.
5927 0 : start = partition_parts[0].start().unwrap();
5928 0 : found = true;
5929 0 : break;
5930 0 : }
5931 : }
5932 0 : if !found {
5933 : // Last key is within the last partition, or larger than all partitions.
5934 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5935 0 : }
5936 191 : }
5937 :
5938 191 : let total = partition_parts.len();
5939 394 : for (idx, partition) in partition_parts.iter().enumerate() {
5940 394 : if self.cancel.is_cancelled() {
5941 0 : return Err(CreateImageLayersError::Cancelled);
5942 394 : }
5943 394 : partition_processed += 1;
5944 394 : let img_range = start..partition.ranges.last().unwrap().end;
5945 394 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5946 394 : if compact_metadata {
5947 764 : for range in &partition.ranges {
5948 573 : assert!(
5949 573 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5950 573 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5951 0 : "metadata keys must be partitioned separately"
5952 : );
5953 : }
5954 191 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5955 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5956 : // might mess up with evictions.
5957 22 : start = img_range.end;
5958 22 : continue;
5959 169 : }
5960 : // For initial and force modes, we always generate image layers for metadata keys.
5961 203 : } else if let ImageLayerCreationMode::Try = mode {
5962 : // check_for_image_layers = false -> skip
5963 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5964 79 : if !check_for_image_layers
5965 57 : || !self
5966 57 : .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
5967 57 : .await
5968 : {
5969 79 : start = img_range.end;
5970 79 : continue;
5971 0 : }
5972 124 : }
5973 293 : if let ImageLayerCreationMode::Force = mode {
5974 : // When forced to create image layers, we might try and create them where they already
5975 : // exist. This mode is only used in tests/debug.
5976 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5977 14 : if layers.contains_key(&PersistentLayerKey {
5978 14 : key_range: img_range.clone(),
5979 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5980 14 : is_delta: false,
5981 14 : }) {
5982 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5983 : // in the future.
5984 0 : tracing::info!(
5985 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5986 : img_range.start,
5987 : img_range.end
5988 : );
5989 0 : start = img_range.end;
5990 0 : continue;
5991 14 : }
5992 279 : }
5993 :
5994 293 : let image_layer_writer = ImageLayerWriter::new(
5995 293 : self.conf,
5996 293 : self.timeline_id,
5997 293 : self.tenant_shard_id,
5998 293 : &img_range,
5999 293 : lsn,
6000 293 : &self.gate,
6001 293 : self.cancel.clone(),
6002 293 : ctx,
6003 293 : )
6004 293 : .await
6005 293 : .map_err(CreateImageLayersError::Other)?;
6006 :
6007 293 : fail_point!("image-layer-writer-fail-before-finish", |_| {
6008 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
6009 0 : "failpoint image-layer-writer-fail-before-finish"
6010 0 : )))
6011 0 : });
6012 :
6013 : // Begin Hadron
6014 : //
6015 293 : fail_point!("create-image-layer-fail-simulated-corruption", |_| {
6016 0 : self.corruption_detected
6017 0 : .store(true, std::sync::atomic::Ordering::Relaxed);
6018 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
6019 0 : "failpoint create-image-layer-fail-simulated-corruption"
6020 0 : )))
6021 0 : });
6022 : // End Hadron
6023 :
6024 293 : let io_concurrency = IoConcurrency::spawn_from_conf(
6025 293 : self.conf.get_vectored_concurrent_io,
6026 293 : self.gate
6027 293 : .enter()
6028 293 : .map_err(|_| CreateImageLayersError::Cancelled)?,
6029 : );
6030 :
6031 293 : let outcome = if !compact_metadata {
6032 124 : self.create_image_layer_for_rel_blocks(
6033 124 : partition,
6034 124 : image_layer_writer,
6035 124 : lsn,
6036 124 : ctx,
6037 124 : img_range.clone(),
6038 124 : io_concurrency,
6039 124 : Some((idx, total)),
6040 124 : )
6041 124 : .await?
6042 : } else {
6043 169 : self.create_image_layer_for_metadata_keys(
6044 169 : partition,
6045 169 : image_layer_writer,
6046 169 : lsn,
6047 169 : ctx,
6048 169 : img_range.clone(),
6049 169 : mode,
6050 169 : io_concurrency,
6051 169 : )
6052 169 : .await?
6053 : };
6054 293 : match outcome {
6055 147 : ImageLayerCreationOutcome::Empty => {
6056 147 : // No data in this partition, so we don't need to create an image layer (for now).
6057 147 : // The next image layer should cover this key range, so we don't advance the `start`
6058 147 : // key.
6059 147 : }
6060 : ImageLayerCreationOutcome::Generated {
6061 130 : unfinished_image_layer,
6062 130 : } => {
6063 130 : batch_image_writer.add_unfinished_image_writer(
6064 130 : unfinished_image_layer,
6065 130 : img_range.clone(),
6066 130 : lsn,
6067 130 : );
6068 130 : // The next image layer should be generated right after this one.
6069 130 : start = img_range.end;
6070 130 : }
6071 16 : ImageLayerCreationOutcome::Skip => {
6072 16 : // We don't need to create an image layer for this partition.
6073 16 : // The next image layer should NOT cover this range, otherwise
6074 16 : // the keyspace becomes empty (reads don't go past image layers).
6075 16 : start = img_range.end;
6076 16 : }
6077 : }
6078 :
6079 293 : if let ImageLayerCreationMode::Try = mode {
6080 : // We have at least made some progress
6081 51 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
6082 : // The `Try` mode is currently only used on the compaction path. We want to avoid
6083 : // image layer generation taking too long time and blocking L0 compaction. So in this
6084 : // mode, we also inspect the current number of L0 layers and skip image layer generation
6085 : // if there are too many of them.
6086 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
6087 0 : * self.get_compaction_threshold();
6088 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
6089 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
6090 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
6091 0 : if image_preempt_threshold != 0 {
6092 0 : let should_yield = self
6093 0 : .l0_compaction_trigger
6094 0 : .notified()
6095 0 : .now_or_never()
6096 0 : .is_some();
6097 0 : if should_yield {
6098 0 : tracing::info!(
6099 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
6100 0 : partition.start().unwrap(),
6101 0 : partition.end().unwrap()
6102 : );
6103 0 : last_partition_processed = Some(partition.clone());
6104 0 : all_generated = false;
6105 0 : break;
6106 0 : }
6107 0 : }
6108 51 : }
6109 242 : }
6110 : }
6111 :
6112 191 : let image_layers = batch_image_writer
6113 191 : .finish(self, ctx)
6114 191 : .await
6115 191 : .map_err(CreateImageLayersError::Other)?;
6116 :
6117 191 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6118 :
6119 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
6120 : // now they are being scheduled outside of write lock; current way is inconsistent with
6121 : // compaction lock order.
6122 191 : guard
6123 191 : .open_mut()?
6124 191 : .track_new_image_layers(&image_layers, &self.metrics);
6125 191 : drop_layer_manager_wlock(guard);
6126 191 : let duration = timer.stop_and_record();
6127 :
6128 : // Creating image layers may have caused some previously visible layers to be covered
6129 191 : if !image_layers.is_empty() {
6130 118 : self.update_layer_visibility().await?;
6131 73 : }
6132 :
6133 191 : let total_layer_size = image_layers
6134 191 : .iter()
6135 191 : .map(|l| l.metadata().file_size)
6136 191 : .sum::<u64>();
6137 :
6138 191 : if !image_layers.is_empty() {
6139 118 : info!(
6140 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
6141 0 : image_layers.len(),
6142 : total_layer_size,
6143 0 : duration.as_secs_f64(),
6144 : partition_processed,
6145 : total_partitions
6146 : );
6147 73 : }
6148 :
6149 : Ok((
6150 191 : image_layers,
6151 191 : if all_generated {
6152 191 : LastImageLayerCreationStatus::Complete
6153 : } else {
6154 : LastImageLayerCreationStatus::Incomplete {
6155 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
6156 0 : last_partition_processed.end().unwrap_or(Key::MIN)
6157 : } else {
6158 : // This branch should be unreachable, but in case it happens, we can just return the start key.
6159 0 : Key::MIN
6160 : },
6161 : }
6162 : },
6163 : ))
6164 191 : }
6165 :
6166 : /// Wait until the background initial logical size calculation is complete, or
6167 : /// this Timeline is shut down. Calling this function will cause the initial
6168 : /// logical size calculation to skip waiting for the background jobs barrier.
6169 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
6170 0 : if !self.shard_identity.is_shard_zero() {
6171 : // We don't populate logical size on shard >0: skip waiting for it.
6172 0 : return;
6173 0 : }
6174 :
6175 0 : if self.remote_client.is_deleting() {
6176 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
6177 0 : return;
6178 0 : }
6179 :
6180 0 : if self.current_logical_size.current_size().is_exact() {
6181 : // root timelines are initialized with exact count, but never start the background
6182 : // calculation
6183 0 : return;
6184 0 : }
6185 :
6186 0 : if self.cancel.is_cancelled() {
6187 : // We already requested stopping the tenant, so we cannot wait for the logical size
6188 : // calculation to complete given the task might have been already cancelled.
6189 0 : return;
6190 0 : }
6191 :
6192 0 : if let Some(await_bg_cancel) = self
6193 0 : .current_logical_size
6194 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
6195 0 : .get()
6196 0 : {
6197 0 : await_bg_cancel.cancel();
6198 0 : } else {
6199 : // We should not wait if we were not able to explicitly instruct
6200 : // the logical size cancellation to skip the concurrency limit semaphore.
6201 : // TODO: this is an unexpected case. We should restructure so that it
6202 : // can't happen.
6203 0 : tracing::warn!(
6204 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
6205 : );
6206 0 : debug_assert!(false);
6207 : }
6208 :
6209 0 : tokio::select!(
6210 0 : _ = self.current_logical_size.initialized.acquire() => {},
6211 0 : _ = self.cancel.cancelled() => {}
6212 : )
6213 0 : }
6214 :
6215 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
6216 : /// Timelines layers up to the ancestor_lsn.
6217 : ///
6218 : /// Requires a timeline that:
6219 : /// - has an ancestor to detach from
6220 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
6221 : /// a technical requirement
6222 : ///
6223 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
6224 : /// polled again until completion.
6225 : ///
6226 : /// During the operation all timelines sharing the data with this timeline will be reparented
6227 : /// from our ancestor to be branches of this timeline.
6228 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
6229 0 : self: &Arc<Timeline>,
6230 0 : tenant: &crate::tenant::TenantShard,
6231 0 : options: detach_ancestor::Options,
6232 0 : behavior: DetachBehavior,
6233 0 : ctx: &RequestContext,
6234 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
6235 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
6236 0 : }
6237 :
6238 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
6239 : /// reparents any reparentable children of previous ancestor.
6240 : ///
6241 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6242 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6243 : /// successfully, tenant must be reloaded.
6244 : ///
6245 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6246 : /// resetting the tenant.
6247 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6248 0 : self: &Arc<Timeline>,
6249 0 : tenant: &crate::tenant::TenantShard,
6250 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6251 0 : ancestor_timeline_id: TimelineId,
6252 0 : ancestor_lsn: Lsn,
6253 0 : behavior: DetachBehavior,
6254 0 : ctx: &RequestContext,
6255 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6256 0 : detach_ancestor::detach_and_reparent(
6257 0 : self,
6258 0 : tenant,
6259 0 : prepared,
6260 0 : ancestor_timeline_id,
6261 0 : ancestor_lsn,
6262 0 : behavior,
6263 0 : ctx,
6264 0 : )
6265 0 : .await
6266 0 : }
6267 :
6268 : /// Final step which unblocks the GC.
6269 : ///
6270 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6271 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6272 0 : self: &Arc<Timeline>,
6273 0 : tenant: &crate::tenant::TenantShard,
6274 0 : attempt: detach_ancestor::Attempt,
6275 0 : ctx: &RequestContext,
6276 0 : ) -> Result<(), detach_ancestor::Error> {
6277 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6278 0 : }
6279 : }
6280 :
6281 : impl Drop for Timeline {
6282 5 : fn drop(&mut self) {
6283 5 : if let Some(ancestor) = &self.ancestor_timeline {
6284 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6285 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6286 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6287 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6288 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6289 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6290 2 : }
6291 0 : }
6292 3 : }
6293 5 : info!(
6294 0 : "Timeline {} for tenant {} is being dropped",
6295 : self.timeline_id, self.tenant_shard_id.tenant_id
6296 : );
6297 5 : }
6298 : }
6299 :
6300 : pub(crate) use compaction_error::CompactionError;
6301 : /// In a private mod to enforce that [`CompactionError::is_cancel`] is used
6302 : /// instead of `match`ing on [`CompactionError::ShuttingDown`].
6303 : mod compaction_error {
6304 : use utils::sync::gate::GateError;
6305 :
6306 : use crate::{
6307 : pgdatadir_mapping::CollectKeySpaceError,
6308 : tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
6309 : virtual_file::owned_buffers_io::write::FlushTaskError,
6310 : };
6311 :
6312 : /// Top-level failure to compact. Use [`Self::is_cancel`].
6313 : #[derive(Debug, thiserror::Error)]
6314 : pub(crate) enum CompactionError {
6315 : /// Use [`Self::is_cancel`] instead of checking for this variant.
6316 : #[error("The timeline or pageserver is shutting down")]
6317 : #[allow(private_interfaces)]
6318 : ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
6319 : #[error(transparent)]
6320 : Other(anyhow::Error),
6321 : }
6322 :
6323 : #[derive(Debug)]
6324 : struct ForbidMatching;
6325 :
6326 : impl CompactionError {
6327 0 : pub fn new_cancelled() -> Self {
6328 0 : Self::ShuttingDown(ForbidMatching)
6329 0 : }
6330 : /// Errors that can be ignored, i.e., cancel and shutdown.
6331 0 : pub fn is_cancel(&self) -> bool {
6332 0 : let other = match self {
6333 0 : CompactionError::ShuttingDown(_) => return true,
6334 0 : CompactionError::Other(other) => other,
6335 : };
6336 :
6337 : // The write path of compaction in particular often lacks differentiated
6338 : // handling errors stemming from cancellation from other errors.
6339 : // So, if requested, we also check the ::Other variant by downcasting.
6340 : // The list below has been found empirically from flaky tests and production logs.
6341 : // The process is simple: on ::Other(), compaction will print the enclosed
6342 : // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
6343 : // line where the write path / compaction code does undifferentiated error handling
6344 : // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
6345 : // below, following the same is_cancel() pattern.
6346 :
6347 0 : let root_cause = other.root_cause();
6348 :
6349 0 : let upload_queue = root_cause
6350 0 : .downcast_ref::<NotInitialized>()
6351 0 : .is_some_and(|e| e.is_stopping());
6352 0 : let timeline = root_cause
6353 0 : .downcast_ref::<PageReconstructError>()
6354 0 : .is_some_and(|e| e.is_cancel());
6355 0 : let buffered_writer_flush_task_canelled = root_cause
6356 0 : .downcast_ref::<FlushTaskError>()
6357 0 : .is_some_and(|e| e.is_cancel());
6358 0 : let write_blob_cancelled = root_cause
6359 0 : .downcast_ref::<WriteBlobError>()
6360 0 : .is_some_and(|e| e.is_cancel());
6361 0 : let gate_closed = root_cause
6362 0 : .downcast_ref::<GateError>()
6363 0 : .is_some_and(|e| e.is_cancel());
6364 0 : upload_queue
6365 0 : || timeline
6366 0 : || buffered_writer_flush_task_canelled
6367 0 : || write_blob_cancelled
6368 0 : || gate_closed
6369 0 : }
6370 0 : pub fn into_anyhow(self) -> anyhow::Error {
6371 0 : match self {
6372 0 : CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
6373 0 : CompactionError::Other(e) => e,
6374 : }
6375 0 : }
6376 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6377 0 : if err.is_cancel() {
6378 0 : Self::new_cancelled()
6379 : } else {
6380 0 : Self::Other(err.into_anyhow())
6381 : }
6382 0 : }
6383 : }
6384 : }
6385 :
6386 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6387 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6388 0 : match value {
6389 : super::upload_queue::NotInitialized::Uninitialized => {
6390 0 : CompactionError::Other(anyhow::anyhow!(value))
6391 : }
6392 : super::upload_queue::NotInitialized::ShuttingDown
6393 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
6394 : }
6395 0 : }
6396 : }
6397 :
6398 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6399 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6400 0 : match e {
6401 : super::storage_layer::layer::DownloadError::TimelineShutdown
6402 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6403 0 : CompactionError::new_cancelled()
6404 : }
6405 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6406 : | super::storage_layer::layer::DownloadError::DownloadRequired
6407 : | super::storage_layer::layer::DownloadError::NotFile(_)
6408 : | super::storage_layer::layer::DownloadError::DownloadFailed
6409 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6410 0 : CompactionError::Other(anyhow::anyhow!(e))
6411 : }
6412 : #[cfg(test)]
6413 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6414 0 : CompactionError::Other(anyhow::anyhow!(e))
6415 : }
6416 : }
6417 0 : }
6418 : }
6419 :
6420 : impl From<layer_manager::Shutdown> for CompactionError {
6421 0 : fn from(_: layer_manager::Shutdown) -> Self {
6422 0 : CompactionError::new_cancelled()
6423 0 : }
6424 : }
6425 :
6426 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6427 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6428 0 : if e.is_cancel() {
6429 0 : CompactionError::new_cancelled()
6430 : } else {
6431 0 : CompactionError::Other(e.into_anyhow())
6432 : }
6433 0 : }
6434 : }
6435 :
6436 : #[serde_as]
6437 : #[derive(serde::Serialize)]
6438 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6439 :
6440 : #[derive(Default)]
6441 : enum DurationRecorder {
6442 : #[default]
6443 : NotStarted,
6444 : Recorded(RecordedDuration, tokio::time::Instant),
6445 : }
6446 :
6447 : impl DurationRecorder {
6448 115 : fn till_now(&self) -> DurationRecorder {
6449 115 : match self {
6450 : DurationRecorder::NotStarted => {
6451 0 : panic!("must only call on recorded measurements")
6452 : }
6453 115 : DurationRecorder::Recorded(_, ended) => {
6454 115 : let now = tokio::time::Instant::now();
6455 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6456 : }
6457 : }
6458 115 : }
6459 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6460 138 : match self {
6461 0 : DurationRecorder::NotStarted => None,
6462 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6463 : }
6464 138 : }
6465 : }
6466 :
6467 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6468 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6469 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6470 : /// keys specified in the `data` field.
6471 : #[cfg(test)]
6472 : #[derive(Clone)]
6473 : pub struct DeltaLayerTestDesc {
6474 : pub lsn_range: Range<Lsn>,
6475 : pub key_range: Range<Key>,
6476 : pub data: Vec<(Key, Lsn, Value)>,
6477 : }
6478 :
6479 : #[cfg(test)]
6480 : #[derive(Clone)]
6481 : pub struct InMemoryLayerTestDesc {
6482 : pub lsn_range: Range<Lsn>,
6483 : pub data: Vec<(Key, Lsn, Value)>,
6484 : pub is_open: bool,
6485 : }
6486 :
6487 : #[cfg(test)]
6488 : impl DeltaLayerTestDesc {
6489 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6490 2 : Self {
6491 2 : lsn_range,
6492 2 : key_range,
6493 2 : data,
6494 2 : }
6495 2 : }
6496 :
6497 45 : pub fn new_with_inferred_key_range(
6498 45 : lsn_range: Range<Lsn>,
6499 45 : data: Vec<(Key, Lsn, Value)>,
6500 45 : ) -> Self {
6501 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6502 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6503 45 : Self {
6504 45 : key_range: (*key_min)..(key_max.next()),
6505 45 : lsn_range,
6506 45 : data,
6507 45 : }
6508 45 : }
6509 :
6510 5 : pub(crate) fn layer_name(&self) -> LayerName {
6511 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6512 5 : key_range: self.key_range.clone(),
6513 5 : lsn_range: self.lsn_range.clone(),
6514 5 : })
6515 5 : }
6516 : }
6517 :
6518 : impl Timeline {
6519 23 : async fn finish_compact_batch(
6520 23 : self: &Arc<Self>,
6521 23 : new_deltas: &[ResidentLayer],
6522 23 : new_images: &[ResidentLayer],
6523 23 : layers_to_remove: &[Layer],
6524 23 : ) -> Result<(), CompactionError> {
6525 23 : let mut guard = tokio::select! {
6526 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6527 23 : _ = self.cancel.cancelled() => {
6528 0 : return Err(CompactionError::new_cancelled());
6529 : }
6530 : };
6531 :
6532 23 : let mut duplicated_layers = HashSet::new();
6533 :
6534 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6535 :
6536 186 : for l in new_deltas {
6537 163 : if guard.contains(l.as_ref()) {
6538 : // expected in tests
6539 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6540 :
6541 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6542 : // `guard` on self.layers. as of writing this, there are no error returns except
6543 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6544 : // because we have not implemented L0 => L0 compaction.
6545 0 : duplicated_layers.insert(l.layer_desc().key());
6546 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6547 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6548 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6549 0 : )));
6550 163 : } else {
6551 163 : insert_layers.push(l.clone());
6552 163 : }
6553 : }
6554 :
6555 : // only remove those inputs which were not outputs
6556 23 : let remove_layers: Vec<Layer> = layers_to_remove
6557 23 : .iter()
6558 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6559 23 : .cloned()
6560 23 : .collect();
6561 :
6562 23 : if !new_images.is_empty() {
6563 0 : guard
6564 0 : .open_mut()?
6565 0 : .track_new_image_layers(new_images, &self.metrics);
6566 23 : }
6567 :
6568 23 : guard
6569 23 : .open_mut()?
6570 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6571 :
6572 23 : self.remote_client
6573 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6574 :
6575 23 : drop_layer_manager_wlock(guard);
6576 :
6577 23 : Ok(())
6578 23 : }
6579 :
6580 0 : async fn rewrite_layers(
6581 0 : self: &Arc<Self>,
6582 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6583 0 : mut drop_layers: Vec<Layer>,
6584 0 : ) -> Result<(), CompactionError> {
6585 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6586 :
6587 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6588 : // to avoid double-removing, and avoid rewriting something that was removed.
6589 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6590 0 : drop_layers.retain(|l| guard.contains(l));
6591 :
6592 0 : guard
6593 0 : .open_mut()?
6594 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6595 :
6596 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6597 :
6598 0 : self.remote_client
6599 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6600 :
6601 0 : Ok(())
6602 0 : }
6603 :
6604 : /// Schedules the uploads of the given image layers
6605 80 : fn upload_new_image_layers(
6606 80 : self: &Arc<Self>,
6607 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6608 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6609 93 : for layer in new_images {
6610 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6611 : }
6612 : // should any new image layer been created, not uploading index_part will
6613 : // result in a mismatch between remote_physical_size and layermap calculated
6614 : // size, which will fail some tests, but should not be an issue otherwise.
6615 80 : self.remote_client
6616 80 : .schedule_index_upload_for_file_changes()?;
6617 80 : Ok(())
6618 80 : }
6619 :
6620 0 : async fn find_gc_time_cutoff(
6621 0 : &self,
6622 0 : now: SystemTime,
6623 0 : pitr: Duration,
6624 0 : cancel: &CancellationToken,
6625 0 : ctx: &RequestContext,
6626 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6627 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6628 0 : if self.shard_identity.is_shard_zero() {
6629 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6630 0 : let time_range = if pitr == Duration::ZERO {
6631 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6632 : } else {
6633 0 : pitr
6634 : };
6635 :
6636 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6637 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6638 0 : let timestamp = to_pg_timestamp(time_cutoff);
6639 :
6640 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6641 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6642 0 : LsnForTimestamp::Future(lsn) => {
6643 : // The timestamp is in the future. That sounds impossible,
6644 : // but what it really means is that there hasn't been
6645 : // any commits since the cutoff timestamp.
6646 : //
6647 : // In this case we should use the LSN of the most recent commit,
6648 : // which is implicitly the last LSN in the log.
6649 0 : debug!("future({})", lsn);
6650 0 : Some(self.get_last_record_lsn())
6651 : }
6652 0 : LsnForTimestamp::Past(lsn) => {
6653 0 : debug!("past({})", lsn);
6654 0 : None
6655 : }
6656 0 : LsnForTimestamp::NoData(lsn) => {
6657 0 : debug!("nodata({})", lsn);
6658 0 : None
6659 : }
6660 : };
6661 0 : Ok(time_cutoff)
6662 : } else {
6663 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6664 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6665 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6666 : // space cutoff that we would also have respected ourselves.
6667 0 : match self
6668 0 : .remote_client
6669 0 : .download_foreign_index(ShardNumber(0), cancel)
6670 0 : .await
6671 : {
6672 0 : Ok((index_part, index_generation, _index_mtime)) => {
6673 0 : tracing::info!(
6674 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6675 0 : index_part.metadata.latest_gc_cutoff_lsn()
6676 : );
6677 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6678 : }
6679 : Err(DownloadError::NotFound) => {
6680 : // This is unexpected, because during timeline creations shard zero persists to remote
6681 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6682 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6683 : // state, then shard zero should _eventually_ write an index when it GCs.
6684 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6685 0 : Ok(None)
6686 : }
6687 0 : Err(e) => {
6688 : // TODO: this function should return a different error type than page reconstruct error
6689 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6690 : }
6691 : }
6692 :
6693 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6694 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6695 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6696 : // new location. This is legal in principle, but problematic in practice because it might result
6697 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6698 : // because they have GC'd past the branch point.
6699 : }
6700 0 : }
6701 :
6702 : /// Find the Lsns above which layer files need to be retained on
6703 : /// garbage collection.
6704 : ///
6705 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6706 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6707 : /// the space-based retention.
6708 : ///
6709 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6710 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6711 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6712 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6713 : /// in the response as the GC cutoff point for the timeline.
6714 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6715 : pub(super) async fn find_gc_cutoffs(
6716 : &self,
6717 : now: SystemTime,
6718 : space_cutoff: Lsn,
6719 : pitr: Duration,
6720 : cancel: &CancellationToken,
6721 : ctx: &RequestContext,
6722 : ) -> Result<GcCutoffs, PageReconstructError> {
6723 : let _timer = self
6724 : .metrics
6725 : .find_gc_cutoffs_histo
6726 : .start_timer()
6727 : .record_on_drop();
6728 :
6729 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6730 :
6731 : if cfg!(test) && pitr == Duration::ZERO {
6732 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6733 : return Ok(GcCutoffs {
6734 : time: Some(self.get_last_record_lsn()),
6735 : space: space_cutoff,
6736 : });
6737 : }
6738 :
6739 : // Calculate a time-based limit on how much to retain:
6740 : // - if PITR interval is set, then this is our cutoff.
6741 : // - if PITR interval is not set, then we do a lookup
6742 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6743 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6744 :
6745 : Ok(match (pitr, time_cutoff) {
6746 : (Duration::ZERO, Some(time_cutoff)) => {
6747 : // PITR is not set. Retain the size-based limit, or the default time retention,
6748 : // whichever requires less data.
6749 : GcCutoffs {
6750 : time: Some(self.get_last_record_lsn()),
6751 : space: std::cmp::max(time_cutoff, space_cutoff),
6752 : }
6753 : }
6754 : (Duration::ZERO, None) => {
6755 : // PITR is not set, and time lookup failed
6756 : GcCutoffs {
6757 : time: Some(self.get_last_record_lsn()),
6758 : space: space_cutoff,
6759 : }
6760 : }
6761 : (_, None) => {
6762 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6763 : // cannot advance beyond what was already GC'd, and respect space-based retention
6764 : GcCutoffs {
6765 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6766 : space: space_cutoff,
6767 : }
6768 : }
6769 : (_, Some(time_cutoff)) => {
6770 : // PITR interval is set and we looked up timestamp successfully. Ignore
6771 : // size based retention and make time cutoff authoritative
6772 : GcCutoffs {
6773 : time: Some(time_cutoff),
6774 : space: time_cutoff,
6775 : }
6776 : }
6777 : })
6778 : }
6779 :
6780 : /// Garbage collect layer files on a timeline that are no longer needed.
6781 : ///
6782 : /// Currently, we don't make any attempt at removing unneeded page versions
6783 : /// within a layer file. We can only remove the whole file if it's fully
6784 : /// obsolete.
6785 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6786 : // this is most likely the background tasks, but it might be the spawned task from
6787 : // immediate_gc
6788 377 : let _g = tokio::select! {
6789 377 : guard = self.gc_lock.lock() => guard,
6790 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6791 : };
6792 376 : let timer = self.metrics.garbage_collect_histo.start_timer();
6793 :
6794 376 : fail_point!("before-timeline-gc");
6795 :
6796 : // Is the timeline being deleted?
6797 376 : if self.is_stopping() {
6798 0 : return Err(GcError::TimelineCancelled);
6799 376 : }
6800 :
6801 376 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6802 376 : let gc_info = self.gc_info.read().unwrap();
6803 :
6804 376 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6805 376 : let time_cutoff = gc_info.cutoffs.time;
6806 376 : let retain_lsns = gc_info
6807 376 : .retain_lsns
6808 376 : .iter()
6809 376 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6810 376 : .collect();
6811 :
6812 : // Gets the maximum LSN that holds the valid lease.
6813 : //
6814 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6815 : // Here, we do not check for stale leases again.
6816 376 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6817 :
6818 376 : (
6819 376 : space_cutoff,
6820 376 : time_cutoff,
6821 376 : retain_lsns,
6822 376 : max_lsn_with_valid_lease,
6823 376 : )
6824 : };
6825 :
6826 376 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6827 376 : let standby_horizon = self.standby_horizon.load();
6828 : // Hold GC for the standby, but as a safety guard do it only within some
6829 : // reasonable lag.
6830 376 : if standby_horizon != Lsn::INVALID {
6831 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6832 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6833 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6834 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6835 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6836 : } else {
6837 0 : warn!(
6838 0 : "standby is lagging for more than {}MB, not holding gc for it",
6839 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6840 : )
6841 : }
6842 0 : }
6843 376 : }
6844 :
6845 : // Reset standby horizon to ignore it if it is not updated till next GC.
6846 : // It is an easy way to unset it when standby disappears without adding
6847 : // more conf options.
6848 376 : self.standby_horizon.store(Lsn::INVALID);
6849 376 : self.metrics
6850 376 : .standby_horizon_gauge
6851 376 : .set(Lsn::INVALID.0 as i64);
6852 :
6853 376 : let res = self
6854 376 : .gc_timeline(
6855 376 : space_cutoff,
6856 376 : time_cutoff,
6857 376 : retain_lsns,
6858 376 : max_lsn_with_valid_lease,
6859 376 : new_gc_cutoff,
6860 : )
6861 376 : .instrument(
6862 376 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6863 : )
6864 376 : .await?;
6865 :
6866 : // only record successes
6867 376 : timer.stop_and_record();
6868 :
6869 376 : Ok(res)
6870 377 : }
6871 :
6872 376 : async fn gc_timeline(
6873 376 : &self,
6874 376 : space_cutoff: Lsn,
6875 376 : time_cutoff: Option<Lsn>, // None if uninitialized
6876 376 : retain_lsns: Vec<Lsn>,
6877 376 : max_lsn_with_valid_lease: Option<Lsn>,
6878 376 : new_gc_cutoff: Lsn,
6879 376 : ) -> Result<GcResult, GcError> {
6880 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6881 :
6882 376 : let now = SystemTime::now();
6883 376 : let mut result: GcResult = GcResult::default();
6884 :
6885 : // Nothing to GC. Return early.
6886 376 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6887 376 : if latest_gc_cutoff >= new_gc_cutoff {
6888 11 : info!(
6889 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6890 : );
6891 11 : return Ok(result);
6892 365 : }
6893 :
6894 365 : let Some(time_cutoff) = time_cutoff else {
6895 : // The GC cutoff should have been computed by now, but let's be defensive.
6896 0 : info!("Nothing to GC: time_cutoff not yet computed");
6897 0 : return Ok(result);
6898 : };
6899 :
6900 : // We need to ensure that no one tries to read page versions or create
6901 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6902 : // for details. This will block until the old value is no longer in use.
6903 : //
6904 : // The GC cutoff should only ever move forwards.
6905 365 : let waitlist = {
6906 365 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6907 365 : if *write_guard > new_gc_cutoff {
6908 0 : return Err(GcError::BadLsn {
6909 0 : why: format!(
6910 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6911 0 : *write_guard, new_gc_cutoff
6912 0 : ),
6913 0 : });
6914 365 : }
6915 :
6916 365 : write_guard.store_and_unlock(new_gc_cutoff)
6917 : };
6918 365 : let waitlist_wait_fut = std::pin::pin!(waitlist.wait());
6919 365 : log_slow(
6920 365 : "applied_gc_cutoff waitlist wait",
6921 365 : Duration::from_secs(30),
6922 365 : waitlist_wait_fut,
6923 365 : )
6924 365 : .await;
6925 :
6926 365 : info!("GC starting");
6927 :
6928 365 : debug!("retain_lsns: {:?}", retain_lsns);
6929 :
6930 365 : let max_retain_lsn = retain_lsns.iter().max();
6931 :
6932 : // Scan all layers in the timeline (remote or on-disk).
6933 : //
6934 : // Garbage collect the layer if all conditions are satisfied:
6935 : // 1. it is older than cutoff LSN;
6936 : // 2. it is older than PITR interval;
6937 : // 3. it doesn't need to be retained for 'retain_lsns';
6938 : // 4. it does not need to be kept for LSNs holding valid leases.
6939 : // 5. newer on-disk image layers cover the layer's whole key range
6940 365 : let layers_to_remove = {
6941 365 : let mut layers_to_remove = Vec::new();
6942 :
6943 365 : let guard = self
6944 365 : .layers
6945 365 : .read(LayerManagerLockHolder::GarbageCollection)
6946 365 : .await;
6947 365 : let layers = guard.layer_map()?;
6948 6206 : 'outer: for l in layers.iter_historic_layers() {
6949 6206 : result.layers_total += 1;
6950 :
6951 : // 1. Is it newer than GC horizon cutoff point?
6952 6206 : if l.get_lsn_range().end > space_cutoff {
6953 370 : debug!(
6954 0 : "keeping {} because it's newer than space_cutoff {}",
6955 0 : l.layer_name(),
6956 : space_cutoff,
6957 : );
6958 370 : result.layers_needed_by_cutoff += 1;
6959 370 : continue 'outer;
6960 5836 : }
6961 :
6962 : // 2. It is newer than PiTR cutoff point?
6963 5836 : if l.get_lsn_range().end > time_cutoff {
6964 0 : debug!(
6965 0 : "keeping {} because it's newer than time_cutoff {}",
6966 0 : l.layer_name(),
6967 : time_cutoff,
6968 : );
6969 0 : result.layers_needed_by_pitr += 1;
6970 0 : continue 'outer;
6971 5836 : }
6972 :
6973 : // 3. Is it needed by a child branch?
6974 : // NOTE With that we would keep data that
6975 : // might be referenced by child branches forever.
6976 : // We can track this in child timeline GC and delete parent layers when
6977 : // they are no longer needed. This might be complicated with long inheritance chains.
6978 5836 : if let Some(retain_lsn) = max_retain_lsn {
6979 : // start_lsn is inclusive
6980 4 : if &l.get_lsn_range().start <= retain_lsn {
6981 4 : debug!(
6982 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6983 0 : l.layer_name(),
6984 : retain_lsn,
6985 0 : l.is_incremental(),
6986 : );
6987 4 : result.layers_needed_by_branches += 1;
6988 4 : continue 'outer;
6989 0 : }
6990 5832 : }
6991 :
6992 : // 4. Is there a valid lease that requires us to keep this layer?
6993 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
6994 : // keep if layer start <= any of the lease
6995 9 : if &l.get_lsn_range().start <= lsn {
6996 7 : debug!(
6997 0 : "keeping {} because there is a valid lease preventing GC at {}",
6998 0 : l.layer_name(),
6999 : lsn,
7000 : );
7001 7 : result.layers_needed_by_leases += 1;
7002 7 : continue 'outer;
7003 2 : }
7004 5823 : }
7005 :
7006 : // 5. Is there a later on-disk layer for this relation?
7007 : //
7008 : // The end-LSN is exclusive, while disk_consistent_lsn is
7009 : // inclusive. For example, if disk_consistent_lsn is 100, it is
7010 : // OK for a delta layer to have end LSN 101, but if the end LSN
7011 : // is 102, then it might not have been fully flushed to disk
7012 : // before crash.
7013 : //
7014 : // For example, imagine that the following layers exist:
7015 : //
7016 : // 1000 - image (A)
7017 : // 1000-2000 - delta (B)
7018 : // 2000 - image (C)
7019 : // 2000-3000 - delta (D)
7020 : // 3000 - image (E)
7021 : //
7022 : // If GC horizon is at 2500, we can remove layers A and B, but
7023 : // we cannot remove C, even though it's older than 2500, because
7024 : // the delta layer 2000-3000 depends on it.
7025 5825 : if !layers
7026 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
7027 : {
7028 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
7029 5821 : result.layers_not_updated += 1;
7030 5821 : continue 'outer;
7031 4 : }
7032 :
7033 : // We didn't find any reason to keep this file, so remove it.
7034 4 : info!(
7035 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
7036 0 : l.layer_name(),
7037 0 : l.is_incremental(),
7038 : );
7039 4 : layers_to_remove.push(l);
7040 : }
7041 :
7042 365 : layers_to_remove
7043 : };
7044 :
7045 365 : if !layers_to_remove.is_empty() {
7046 : // Persist the new GC cutoff value before we actually remove anything.
7047 : // This unconditionally schedules also an index_part.json update, even though, we will
7048 : // be doing one a bit later with the unlinked gc'd layers.
7049 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
7050 3 : self.schedule_uploads(disk_consistent_lsn, None)
7051 3 : .map_err(|e| {
7052 0 : if self.cancel.is_cancelled() {
7053 0 : GcError::TimelineCancelled
7054 : } else {
7055 0 : GcError::Remote(e)
7056 : }
7057 0 : })?;
7058 :
7059 3 : let mut guard = self
7060 3 : .layers
7061 3 : .write(LayerManagerLockHolder::GarbageCollection)
7062 3 : .await;
7063 :
7064 3 : let gc_layers = layers_to_remove
7065 3 : .iter()
7066 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
7067 3 : .collect::<Vec<Layer>>();
7068 :
7069 3 : result.layers_removed = gc_layers.len() as u64;
7070 :
7071 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
7072 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
7073 :
7074 : #[cfg(feature = "testing")]
7075 3 : {
7076 3 : result.doomed_layers = gc_layers;
7077 3 : }
7078 362 : }
7079 :
7080 365 : info!(
7081 0 : "GC completed removing {} layers, cutoff {}",
7082 : result.layers_removed, new_gc_cutoff
7083 : );
7084 :
7085 365 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
7086 365 : Ok(result)
7087 376 : }
7088 :
7089 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
7090 364862 : pub(crate) async fn reconstruct_value(
7091 364862 : &self,
7092 364862 : key: Key,
7093 364862 : request_lsn: Lsn,
7094 364862 : mut data: ValueReconstructState,
7095 364862 : redo_attempt_type: RedoAttemptType,
7096 364862 : ) -> Result<Bytes, PageReconstructError> {
7097 : // Perform WAL redo if needed
7098 364862 : data.records.reverse();
7099 :
7100 364862 : let fire_critical_error = match redo_attempt_type {
7101 363529 : RedoAttemptType::ReadPage => true,
7102 0 : RedoAttemptType::LegacyCompaction => true,
7103 1333 : RedoAttemptType::GcCompaction => false,
7104 : };
7105 :
7106 : // If we have a page image, and no WAL, we're all set
7107 364862 : if data.records.is_empty() {
7108 338213 : if let Some((img_lsn, img)) = &data.img {
7109 338213 : trace!(
7110 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
7111 : key, img_lsn, request_lsn,
7112 : );
7113 338213 : Ok(img.clone())
7114 : } else {
7115 0 : Err(PageReconstructError::from(anyhow!(
7116 0 : "base image for {key} at {request_lsn} not found"
7117 0 : )))
7118 : }
7119 : } else {
7120 : // We need to do WAL redo.
7121 : //
7122 : // If we don't have a base image, then the oldest WAL record better initialize
7123 : // the page
7124 26649 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
7125 0 : Err(PageReconstructError::from(anyhow!(
7126 0 : "Base image for {} at {} not found, but got {} WAL records",
7127 0 : key,
7128 0 : request_lsn,
7129 0 : data.records.len()
7130 0 : )))
7131 : } else {
7132 26649 : if data.img.is_some() {
7133 12781 : trace!(
7134 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
7135 0 : data.records.len(),
7136 : key,
7137 : request_lsn
7138 : );
7139 : } else {
7140 13868 : trace!(
7141 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
7142 0 : data.records.len(),
7143 : key,
7144 : request_lsn
7145 : );
7146 : };
7147 26649 : let res = self
7148 26649 : .walredo_mgr
7149 26649 : .as_ref()
7150 26649 : .context("timeline has no walredo manager")
7151 26649 : .map_err(PageReconstructError::WalRedo)?
7152 26649 : .request_redo(
7153 26649 : key,
7154 26649 : request_lsn,
7155 26649 : data.img,
7156 26649 : data.records,
7157 26649 : self.pg_version,
7158 26649 : redo_attempt_type,
7159 : )
7160 26649 : .await;
7161 26648 : let img = match res {
7162 26648 : Ok(img) => img,
7163 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
7164 1 : Err(walredo::Error::Other(err)) => {
7165 1 : if fire_critical_error {
7166 0 : critical_timeline!(
7167 0 : self.tenant_shard_id,
7168 0 : self.timeline_id,
7169 0 : Some(&self.corruption_detected),
7170 0 : "walredo failure during page reconstruction: {err:?}"
7171 : );
7172 1 : }
7173 1 : return Err(PageReconstructError::WalRedo(
7174 1 : err.context("reconstruct a page image"),
7175 1 : ));
7176 : }
7177 : };
7178 26648 : Ok(img)
7179 : }
7180 : }
7181 364862 : }
7182 :
7183 0 : pub(crate) async fn spawn_download_all_remote_layers(
7184 0 : self: Arc<Self>,
7185 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7186 0 : ctx: &RequestContext,
7187 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
7188 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7189 :
7190 : // this is not really needed anymore; it has tests which really check the return value from
7191 : // http api. it would be better not to maintain this anymore.
7192 :
7193 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
7194 0 : if let Some(st) = &*status_guard {
7195 0 : match &st.state {
7196 : DownloadRemoteLayersTaskState::Running => {
7197 0 : return Err(st.clone());
7198 : }
7199 : DownloadRemoteLayersTaskState::ShutDown
7200 0 : | DownloadRemoteLayersTaskState::Completed => {
7201 0 : *status_guard = None;
7202 0 : }
7203 : }
7204 0 : }
7205 :
7206 0 : let self_clone = Arc::clone(&self);
7207 0 : let task_ctx = ctx.detached_child(
7208 0 : TaskKind::DownloadAllRemoteLayers,
7209 0 : DownloadBehavior::Download,
7210 : );
7211 0 : let task_id = task_mgr::spawn(
7212 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
7213 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
7214 0 : self.tenant_shard_id,
7215 0 : Some(self.timeline_id),
7216 0 : "download all remote layers task",
7217 0 : async move {
7218 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
7219 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
7220 0 : match &mut *status_guard {
7221 : None => {
7222 0 : warn!("tasks status is supposed to be Some(), since we are running");
7223 : }
7224 0 : Some(st) => {
7225 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
7226 0 : if st.task_id != exp_task_id {
7227 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
7228 0 : } else {
7229 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7230 0 : }
7231 : }
7232 : };
7233 0 : Ok(())
7234 0 : }
7235 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))
7236 : );
7237 :
7238 0 : let initial_info = DownloadRemoteLayersTaskInfo {
7239 0 : task_id: format!("{task_id}"),
7240 0 : state: DownloadRemoteLayersTaskState::Running,
7241 0 : total_layer_count: 0,
7242 0 : successful_download_count: 0,
7243 0 : failed_download_count: 0,
7244 0 : };
7245 0 : *status_guard = Some(initial_info.clone());
7246 :
7247 0 : Ok(initial_info)
7248 0 : }
7249 :
7250 0 : async fn download_all_remote_layers(
7251 0 : self: &Arc<Self>,
7252 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7253 0 : ctx: &RequestContext,
7254 0 : ) {
7255 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7256 :
7257 0 : let remaining = {
7258 0 : let guard = self
7259 0 : .layers
7260 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7261 0 : .await;
7262 0 : let Ok(lm) = guard.layer_map() else {
7263 : // technically here we could look into iterating accessible layers, but downloading
7264 : // all layers of a shutdown timeline makes no sense regardless.
7265 0 : tracing::info!("attempted to download all layers of shutdown timeline");
7266 0 : return;
7267 : };
7268 0 : lm.iter_historic_layers()
7269 0 : .map(|desc| guard.get_from_desc(&desc))
7270 0 : .collect::<Vec<_>>()
7271 : };
7272 0 : let total_layer_count = remaining.len();
7273 :
7274 : macro_rules! lock_status {
7275 : ($st:ident) => {
7276 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
7277 : let st = st
7278 : .as_mut()
7279 : .expect("this function is only called after the task has been spawned");
7280 : assert_eq!(
7281 : st.task_id,
7282 : format!(
7283 : "{}",
7284 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
7285 : )
7286 : );
7287 : let $st = st;
7288 : };
7289 : }
7290 :
7291 : {
7292 0 : lock_status!(st);
7293 0 : st.total_layer_count = total_layer_count as u64;
7294 : }
7295 :
7296 0 : let mut remaining = remaining.into_iter();
7297 0 : let mut have_remaining = true;
7298 0 : let mut js = tokio::task::JoinSet::new();
7299 :
7300 0 : let cancel = task_mgr::shutdown_token();
7301 :
7302 0 : let limit = request.max_concurrent_downloads;
7303 :
7304 : loop {
7305 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
7306 0 : let Some(next) = remaining.next() else {
7307 0 : have_remaining = false;
7308 0 : break;
7309 : };
7310 :
7311 0 : let span = tracing::info_span!("download", layer = %next);
7312 :
7313 0 : let ctx = ctx.attached_child();
7314 0 : js.spawn(
7315 0 : async move {
7316 0 : let res = next.download(&ctx).await;
7317 0 : (next, res)
7318 0 : }
7319 0 : .instrument(span),
7320 : );
7321 : }
7322 :
7323 0 : while let Some(res) = js.join_next().await {
7324 0 : match res {
7325 : Ok((_, Ok(_))) => {
7326 0 : lock_status!(st);
7327 0 : st.successful_download_count += 1;
7328 : }
7329 0 : Ok((layer, Err(e))) => {
7330 0 : tracing::error!(%layer, "download failed: {e:#}");
7331 0 : lock_status!(st);
7332 0 : st.failed_download_count += 1;
7333 : }
7334 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7335 0 : Err(je) if je.is_panic() => {
7336 0 : lock_status!(st);
7337 0 : st.failed_download_count += 1;
7338 : }
7339 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7340 : }
7341 : }
7342 :
7343 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7344 0 : break;
7345 0 : }
7346 : }
7347 :
7348 : {
7349 0 : lock_status!(st);
7350 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7351 : }
7352 0 : }
7353 :
7354 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7355 0 : &self,
7356 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7357 0 : self.download_all_remote_layers_task_info
7358 0 : .read()
7359 0 : .unwrap()
7360 0 : .clone()
7361 0 : }
7362 :
7363 : /* BEGIN_HADRON */
7364 0 : pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
7365 0 : let guard = self
7366 0 : .layers
7367 0 : .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
7368 0 : .await;
7369 0 : let layer_map = guard.layer_map()?;
7370 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
7371 :
7372 0 : Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
7373 0 : }
7374 : /* END_HADRON */
7375 : }
7376 :
7377 : impl Timeline {
7378 : /// Returns non-remote layers for eviction.
7379 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7380 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7381 0 : let mut max_layer_size: Option<u64> = None;
7382 :
7383 0 : let resident_layers = guard
7384 0 : .likely_resident_layers()
7385 0 : .map(|layer| {
7386 0 : let file_size = layer.layer_desc().file_size;
7387 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7388 :
7389 0 : let last_activity_ts = layer.latest_activity();
7390 :
7391 0 : EvictionCandidate {
7392 0 : layer: layer.to_owned().into(),
7393 0 : last_activity_ts,
7394 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7395 0 : visibility: layer.visibility(),
7396 0 : }
7397 0 : })
7398 0 : .collect();
7399 :
7400 0 : DiskUsageEvictionInfo {
7401 0 : max_layer_size,
7402 0 : resident_layers,
7403 0 : }
7404 0 : }
7405 :
7406 962 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7407 962 : ShardIndex {
7408 962 : shard_number: self.tenant_shard_id.shard_number,
7409 962 : shard_count: self.tenant_shard_id.shard_count,
7410 962 : }
7411 962 : }
7412 :
7413 : /// Persistently blocks gc for `Manual` reason.
7414 : ///
7415 : /// Returns true if no such block existed before, false otherwise.
7416 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7417 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7418 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7419 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7420 0 : }
7421 :
7422 : /// Persistently unblocks gc for `Manual` reason.
7423 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7424 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7425 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7426 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7427 0 : }
7428 :
7429 : #[cfg(test)]
7430 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7431 31 : self.last_record_lsn.advance(new_lsn);
7432 31 : }
7433 :
7434 : #[cfg(test)]
7435 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7436 2 : self.disk_consistent_lsn.store(new_value);
7437 2 : }
7438 :
7439 : /// Force create an image layer and place it into the layer map.
7440 : ///
7441 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7442 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7443 : /// placed into the layer map in one run AND be validated.
7444 : #[cfg(test)]
7445 36 : pub(super) async fn force_create_image_layer(
7446 36 : self: &Arc<Timeline>,
7447 36 : lsn: Lsn,
7448 36 : mut images: Vec<(Key, Bytes)>,
7449 36 : check_start_lsn: Option<Lsn>,
7450 36 : ctx: &RequestContext,
7451 36 : ) -> anyhow::Result<()> {
7452 36 : let last_record_lsn = self.get_last_record_lsn();
7453 36 : assert!(
7454 36 : lsn <= last_record_lsn,
7455 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7456 : );
7457 36 : if let Some(check_start_lsn) = check_start_lsn {
7458 36 : assert!(lsn >= check_start_lsn);
7459 0 : }
7460 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7461 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7462 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7463 36 : let mut image_layer_writer = ImageLayerWriter::new(
7464 36 : self.conf,
7465 36 : self.timeline_id,
7466 36 : self.tenant_shard_id,
7467 36 : &(min_key..end_key),
7468 36 : lsn,
7469 36 : &self.gate,
7470 36 : self.cancel.clone(),
7471 36 : ctx,
7472 36 : )
7473 36 : .await?;
7474 312 : for (key, img) in images {
7475 276 : image_layer_writer.put_image(key, img, ctx).await?;
7476 : }
7477 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7478 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7479 36 : info!("force created image layer {}", image_layer.local_path());
7480 : {
7481 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7482 36 : guard
7483 36 : .open_mut()
7484 36 : .unwrap()
7485 36 : .force_insert_layer(image_layer.clone());
7486 : }
7487 :
7488 : // Update remote_timeline_client state to reflect existence of this layer
7489 36 : self.remote_client
7490 36 : .schedule_layer_file_upload(image_layer)
7491 36 : .unwrap();
7492 :
7493 36 : Ok(())
7494 36 : }
7495 :
7496 : /// Force create a delta layer and place it into the layer map.
7497 : ///
7498 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7499 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7500 : /// placed into the layer map in one run AND be validated.
7501 : #[cfg(test)]
7502 50 : pub(super) async fn force_create_delta_layer(
7503 50 : self: &Arc<Timeline>,
7504 50 : mut deltas: DeltaLayerTestDesc,
7505 50 : check_start_lsn: Option<Lsn>,
7506 50 : ctx: &RequestContext,
7507 50 : ) -> anyhow::Result<()> {
7508 50 : let last_record_lsn = self.get_last_record_lsn();
7509 50 : deltas
7510 50 : .data
7511 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7512 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7513 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7514 10464 : for (_, lsn, _) in &deltas.data {
7515 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7516 : }
7517 50 : assert!(
7518 50 : deltas.lsn_range.end <= last_record_lsn,
7519 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7520 : deltas.lsn_range.end,
7521 : last_record_lsn
7522 : );
7523 50 : if let Some(check_start_lsn) = check_start_lsn {
7524 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7525 0 : }
7526 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7527 50 : self.conf,
7528 50 : self.timeline_id,
7529 50 : self.tenant_shard_id,
7530 50 : deltas.key_range.start,
7531 50 : deltas.lsn_range,
7532 50 : &self.gate,
7533 50 : self.cancel.clone(),
7534 50 : ctx,
7535 50 : )
7536 50 : .await?;
7537 10464 : for (key, lsn, val) in deltas.data {
7538 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7539 : }
7540 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7541 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7542 50 : info!("force created delta layer {}", delta_layer.local_path());
7543 : {
7544 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7545 50 : guard
7546 50 : .open_mut()
7547 50 : .unwrap()
7548 50 : .force_insert_layer(delta_layer.clone());
7549 : }
7550 :
7551 : // Update remote_timeline_client state to reflect existence of this layer
7552 50 : self.remote_client
7553 50 : .schedule_layer_file_upload(delta_layer)
7554 50 : .unwrap();
7555 :
7556 50 : Ok(())
7557 50 : }
7558 :
7559 : /// Force create an in-memory layer and place them into the layer map.
7560 : #[cfg(test)]
7561 4 : pub(super) async fn force_create_in_memory_layer(
7562 4 : self: &Arc<Timeline>,
7563 4 : mut in_memory: InMemoryLayerTestDesc,
7564 4 : check_start_lsn: Option<Lsn>,
7565 4 : ctx: &RequestContext,
7566 4 : ) -> anyhow::Result<()> {
7567 : use utils::bin_ser::BeSer;
7568 :
7569 : // Validate LSNs
7570 4 : if let Some(check_start_lsn) = check_start_lsn {
7571 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7572 0 : }
7573 :
7574 4 : let last_record_lsn = self.get_last_record_lsn();
7575 4 : let layer_end_lsn = if in_memory.is_open {
7576 1 : in_memory
7577 1 : .data
7578 1 : .iter()
7579 1 : .map(|(_key, lsn, _value)| lsn)
7580 1 : .max()
7581 1 : .cloned()
7582 : } else {
7583 3 : Some(in_memory.lsn_range.end)
7584 : };
7585 :
7586 4 : if let Some(end) = layer_end_lsn {
7587 4 : assert!(
7588 4 : end <= last_record_lsn,
7589 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7590 : );
7591 0 : }
7592 :
7593 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7594 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7595 19820 : assert!(*lsn < in_memory.lsn_range.end);
7596 19820 : });
7597 :
7598 : // Build the batch
7599 4 : in_memory
7600 4 : .data
7601 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7602 :
7603 4 : let data = in_memory
7604 4 : .data
7605 4 : .into_iter()
7606 19820 : .map(|(key, lsn, value)| {
7607 19820 : let value_size = value.serialized_size().unwrap() as usize;
7608 19820 : (key.to_compact(), lsn, value_size, value)
7609 19820 : })
7610 4 : .collect::<Vec<_>>();
7611 :
7612 4 : let batch = SerializedValueBatch::from_values(data);
7613 :
7614 : // Create the in-memory layer and write the batch into it
7615 4 : let layer = InMemoryLayer::create(
7616 4 : self.conf,
7617 4 : self.timeline_id,
7618 4 : self.tenant_shard_id,
7619 4 : in_memory.lsn_range.start,
7620 4 : &self.gate,
7621 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7622 4 : &CancellationToken::new(),
7623 4 : ctx,
7624 4 : )
7625 4 : .await
7626 4 : .unwrap();
7627 :
7628 4 : layer.put_batch(batch, ctx).await.unwrap();
7629 4 : if !in_memory.is_open {
7630 3 : layer.freeze(in_memory.lsn_range.end).await;
7631 1 : }
7632 :
7633 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7634 :
7635 : // Link the layer to the layer map
7636 : {
7637 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7638 4 : let layer_map = guard.open_mut().unwrap();
7639 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7640 : }
7641 :
7642 4 : Ok(())
7643 4 : }
7644 :
7645 : /// Return all keys at the LSN in the image layers
7646 : #[cfg(test)]
7647 3 : pub(crate) async fn inspect_image_layers(
7648 3 : self: &Arc<Timeline>,
7649 3 : lsn: Lsn,
7650 3 : ctx: &RequestContext,
7651 3 : io_concurrency: IoConcurrency,
7652 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7653 3 : let mut all_data = Vec::new();
7654 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7655 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7656 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7657 4 : let layer = guard.get_from_desc(&layer);
7658 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7659 4 : layer
7660 4 : .get_values_reconstruct_data(
7661 4 : KeySpace::single(Key::MIN..Key::MAX),
7662 4 : lsn..Lsn(lsn.0 + 1),
7663 4 : &mut reconstruct_data,
7664 4 : ctx,
7665 4 : )
7666 4 : .await?;
7667 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7668 33 : let v = v.collect_pending_ios().await?;
7669 33 : all_data.push((k, v.img.unwrap().1));
7670 : }
7671 13 : }
7672 : }
7673 3 : all_data.sort();
7674 3 : Ok(all_data)
7675 3 : }
7676 :
7677 : /// Get all historic layer descriptors in the layer map
7678 : #[cfg(test)]
7679 12 : pub(crate) async fn inspect_historic_layers(
7680 12 : self: &Arc<Timeline>,
7681 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7682 12 : let mut layers = Vec::new();
7683 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7684 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7685 57 : layers.push(layer.key());
7686 57 : }
7687 12 : Ok(layers)
7688 12 : }
7689 :
7690 : #[cfg(test)]
7691 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7692 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7693 5 : keyspace.merge(&ks);
7694 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7695 5 : }
7696 : }
7697 :
7698 : /// Tracking writes ingestion does to a particular in-memory layer.
7699 : ///
7700 : /// Cleared upon freezing a layer.
7701 : pub(crate) struct TimelineWriterState {
7702 : open_layer: Arc<InMemoryLayer>,
7703 : current_size: u64,
7704 : // Previous Lsn which passed through
7705 : prev_lsn: Option<Lsn>,
7706 : // Largest Lsn which passed through the current writer
7707 : max_lsn: Option<Lsn>,
7708 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7709 : cached_last_freeze_at: Lsn,
7710 : }
7711 :
7712 : impl TimelineWriterState {
7713 660 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7714 660 : Self {
7715 660 : open_layer,
7716 660 : current_size,
7717 660 : prev_lsn: None,
7718 660 : max_lsn: None,
7719 660 : cached_last_freeze_at: last_freeze_at,
7720 660 : }
7721 660 : }
7722 : }
7723 :
7724 : /// Various functions to mutate the timeline.
7725 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7726 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7727 : // but will cause large code changes.
7728 : pub(crate) struct TimelineWriter<'a> {
7729 : tl: &'a Timeline,
7730 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7731 : }
7732 :
7733 : impl Deref for TimelineWriter<'_> {
7734 : type Target = Timeline;
7735 :
7736 4949226 : fn deref(&self) -> &Self::Target {
7737 4949226 : self.tl
7738 4949226 : }
7739 : }
7740 :
7741 : #[derive(PartialEq)]
7742 : enum OpenLayerAction {
7743 : Roll,
7744 : Open,
7745 : None,
7746 : }
7747 :
7748 : impl TimelineWriter<'_> {
7749 2402129 : async fn handle_open_layer_action(
7750 2402129 : &mut self,
7751 2402129 : at: Lsn,
7752 2402129 : action: OpenLayerAction,
7753 2402129 : ctx: &RequestContext,
7754 2402129 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7755 2402129 : match action {
7756 : OpenLayerAction::Roll => {
7757 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7758 40 : self.roll_layer(freeze_at).await?;
7759 40 : self.open_layer(at, ctx).await?;
7760 : }
7761 620 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7762 : OpenLayerAction::None => {
7763 2401469 : assert!(self.write_guard.is_some());
7764 : }
7765 : }
7766 :
7767 2402129 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7768 2402129 : }
7769 :
7770 660 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7771 660 : let layer = self
7772 660 : .tl
7773 660 : .get_layer_for_write(at, &self.write_guard, ctx)
7774 660 : .await?;
7775 660 : let initial_size = layer.len();
7776 :
7777 660 : let last_freeze_at = self.last_freeze_at.load();
7778 660 : self.write_guard.replace(TimelineWriterState::new(
7779 660 : layer,
7780 660 : initial_size,
7781 660 : last_freeze_at,
7782 660 : ));
7783 :
7784 660 : Ok(())
7785 660 : }
7786 :
7787 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7788 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7789 :
7790 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7791 : // to propagate the backpressure up into WAL ingestion.
7792 40 : let l0_count = self
7793 40 : .tl
7794 40 : .layers
7795 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7796 40 : .await
7797 40 : .layer_map()?
7798 40 : .level0_deltas()
7799 40 : .len();
7800 40 : let wait_thresholds = [
7801 40 : self.get_l0_flush_delay_threshold(),
7802 40 : self.get_l0_flush_stall_threshold(),
7803 40 : ];
7804 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7805 :
7806 : // self.write_guard will be taken by the freezing
7807 40 : let flush_id = self
7808 40 : .tl
7809 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7810 40 : .await?;
7811 :
7812 40 : assert!(self.write_guard.is_none());
7813 :
7814 40 : if let Some(wait_threshold) = wait_threshold {
7815 0 : if l0_count >= wait_threshold {
7816 0 : debug!(
7817 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7818 : );
7819 0 : self.tl.wait_flush_completion(flush_id).await?;
7820 0 : }
7821 40 : }
7822 :
7823 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7824 0 : warn!("Flushed oversized open layer with size {}", current_size)
7825 40 : }
7826 :
7827 40 : Ok(())
7828 40 : }
7829 :
7830 2402129 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7831 2402129 : let state = &*self.write_guard;
7832 2402129 : let Some(state) = &state else {
7833 620 : return OpenLayerAction::Open;
7834 : };
7835 :
7836 : #[cfg(feature = "testing")]
7837 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7838 : // this check and assertion are not really needed because
7839 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7840 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7841 : // is no TimelineWriterState.
7842 0 : assert!(
7843 0 : state.open_layer.end_lsn.get().is_some(),
7844 0 : "our open_layer must be outdated"
7845 : );
7846 :
7847 : // this would be a memory leak waiting to happen because the in-memory layer always has
7848 : // an index
7849 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7850 2401509 : }
7851 :
7852 2401509 : if state.prev_lsn == Some(lsn) {
7853 : // Rolling mid LSN is not supported by [downstream code].
7854 : // Hence, only roll at LSN boundaries.
7855 : //
7856 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7857 3 : return OpenLayerAction::None;
7858 2401506 : }
7859 :
7860 2401506 : if state.current_size == 0 {
7861 : // Don't roll empty layers
7862 0 : return OpenLayerAction::None;
7863 2401506 : }
7864 :
7865 2401506 : if self.tl.should_roll(
7866 2401506 : state.current_size,
7867 2401506 : state.current_size + new_value_size,
7868 2401506 : self.get_checkpoint_distance(),
7869 2401506 : lsn,
7870 2401506 : state.cached_last_freeze_at,
7871 2401506 : state.open_layer.get_opened_at(),
7872 : ) {
7873 40 : OpenLayerAction::Roll
7874 : } else {
7875 2401466 : OpenLayerAction::None
7876 : }
7877 2402129 : }
7878 :
7879 : /// Put a batch of keys at the specified Lsns.
7880 2402128 : pub(crate) async fn put_batch(
7881 2402128 : &mut self,
7882 2402128 : batch: SerializedValueBatch,
7883 2402128 : ctx: &RequestContext,
7884 2402128 : ) -> anyhow::Result<()> {
7885 2402128 : if !batch.has_data() {
7886 0 : return Ok(());
7887 2402128 : }
7888 :
7889 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7890 : // We don't assert this in release builds, since key ownership policies may change over
7891 : // time. Stray keys will be removed during compaction.
7892 2402128 : if cfg!(debug_assertions) {
7893 4947529 : for metadata in &batch.metadata {
7894 2545401 : if let ValueMeta::Serialized(metadata) = metadata {
7895 2545401 : let key = Key::from_compact(metadata.key);
7896 2545401 : assert!(
7897 2545401 : self.shard_identity.is_key_local(&key)
7898 12 : || self.shard_identity.is_key_global(&key),
7899 0 : "key {key} does not belong on shard {}",
7900 0 : self.shard_identity.shard_index()
7901 : );
7902 0 : }
7903 : }
7904 0 : }
7905 :
7906 2402128 : let batch_max_lsn = batch.max_lsn;
7907 2402128 : let buf_size: u64 = batch.buffer_size() as u64;
7908 :
7909 2402128 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7910 2402128 : let layer = self
7911 2402128 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7912 2402128 : .await?;
7913 :
7914 2402128 : let res = layer.put_batch(batch, ctx).await;
7915 :
7916 2402128 : if res.is_ok() {
7917 2402128 : // Update the current size only when the entire write was ok.
7918 2402128 : // In case of failures, we may have had partial writes which
7919 2402128 : // render the size tracking out of sync. That's ok because
7920 2402128 : // the checkpoint distance should be significantly smaller
7921 2402128 : // than the S3 single shot upload limit of 5GiB.
7922 2402128 : let state = self.write_guard.as_mut().unwrap();
7923 2402128 :
7924 2402128 : state.current_size += buf_size;
7925 2402128 : state.prev_lsn = Some(batch_max_lsn);
7926 2402128 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7927 2402128 : }
7928 :
7929 2402128 : res
7930 2402128 : }
7931 :
7932 : #[cfg(test)]
7933 : /// Test helper, for tests that would like to poke individual values without composing a batch
7934 2195079 : pub(crate) async fn put(
7935 2195079 : &mut self,
7936 2195079 : key: Key,
7937 2195079 : lsn: Lsn,
7938 2195079 : value: &Value,
7939 2195079 : ctx: &RequestContext,
7940 2195079 : ) -> anyhow::Result<()> {
7941 : use utils::bin_ser::BeSer;
7942 2195079 : if !key.is_valid_key_on_write_path() {
7943 0 : bail!(
7944 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7945 : key
7946 : );
7947 2195079 : }
7948 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7949 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7950 2195079 : key.to_compact(),
7951 2195079 : lsn,
7952 2195079 : val_ser_size,
7953 2195079 : value.clone(),
7954 2195079 : )]);
7955 :
7956 2195079 : self.put_batch(batch, ctx).await
7957 2195079 : }
7958 :
7959 1 : pub(crate) async fn delete_batch(
7960 1 : &mut self,
7961 1 : batch: &[(Range<Key>, Lsn)],
7962 1 : ctx: &RequestContext,
7963 1 : ) -> anyhow::Result<()> {
7964 1 : if let Some((_, lsn)) = batch.first() {
7965 1 : let action = self.get_open_layer_action(*lsn, 0);
7966 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7967 1 : layer.put_tombstones(batch).await?;
7968 0 : }
7969 :
7970 1 : Ok(())
7971 1 : }
7972 :
7973 : /// Track the end of the latest digested WAL record.
7974 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7975 : ///
7976 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7977 : ///
7978 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7979 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7980 : /// the latest and can be read.
7981 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7982 2639560 : self.tl.finish_write(new_lsn);
7983 2639560 : }
7984 :
7985 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7986 135285 : self.tl.update_current_logical_size(delta)
7987 135285 : }
7988 : }
7989 :
7990 : // We need TimelineWriter to be send in upcoming conversion of
7991 : // Timeline::layers to tokio::sync::RwLock.
7992 : #[test]
7993 1 : fn is_send() {
7994 1 : fn _assert_send<T: Send>() {}
7995 1 : _assert_send::<TimelineWriter<'_>>();
7996 1 : }
7997 :
7998 : #[cfg(test)]
7999 : mod tests {
8000 : use std::sync::Arc;
8001 :
8002 : use pageserver_api::key::Key;
8003 : use postgres_ffi::PgMajorVersion;
8004 : use std::iter::Iterator;
8005 : use tracing::Instrument;
8006 : use utils::id::TimelineId;
8007 : use utils::lsn::Lsn;
8008 : use wal_decoder::models::value::Value;
8009 :
8010 : use super::HeatMapTimeline;
8011 : use crate::context::RequestContextBuilder;
8012 : use crate::tenant::harness::{TenantHarness, test_img};
8013 : use crate::tenant::layer_map::LayerMap;
8014 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
8015 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
8016 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
8017 : use crate::tenant::{PreviousHeatmap, Timeline};
8018 :
8019 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
8020 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
8021 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
8022 25 : for (l, r) in lhs_rhs {
8023 20 : assert_eq!(l.name, r.name);
8024 20 : assert_eq!(l.metadata, r.metadata);
8025 : }
8026 5 : }
8027 :
8028 : #[tokio::test]
8029 1 : async fn test_heatmap_generation() {
8030 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
8031 :
8032 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
8033 1 : Lsn(0x10)..Lsn(0x20),
8034 1 : vec![(
8035 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8036 1 : Lsn(0x11),
8037 1 : Value::Image(test_img("foo")),
8038 1 : )],
8039 : );
8040 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
8041 1 : Lsn(0x10)..Lsn(0x20),
8042 1 : vec![(
8043 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8044 1 : Lsn(0x11),
8045 1 : Value::Image(test_img("foo")),
8046 1 : )],
8047 : );
8048 1 : let l0_delta = DeltaLayerTestDesc::new(
8049 1 : Lsn(0x20)..Lsn(0x30),
8050 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8051 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8052 1 : vec![(
8053 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8054 1 : Lsn(0x25),
8055 1 : Value::Image(test_img("foo")),
8056 1 : )],
8057 : );
8058 1 : let delta_layers = vec![
8059 1 : covered_delta.clone(),
8060 1 : visible_delta.clone(),
8061 1 : l0_delta.clone(),
8062 : ];
8063 :
8064 1 : let image_layer = (
8065 1 : Lsn(0x40),
8066 1 : vec![(
8067 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8068 1 : test_img("bar"),
8069 1 : )],
8070 1 : );
8071 1 : let image_layers = vec![image_layer];
8072 :
8073 1 : let (tenant, ctx) = harness.load().await;
8074 1 : let timeline = tenant
8075 1 : .create_test_timeline_with_layers(
8076 1 : TimelineId::generate(),
8077 1 : Lsn(0x10),
8078 1 : PgMajorVersion::PG14,
8079 1 : &ctx,
8080 1 : Vec::new(), // in-memory layers
8081 1 : delta_layers,
8082 1 : image_layers,
8083 1 : Lsn(0x100),
8084 1 : )
8085 1 : .await
8086 1 : .unwrap();
8087 1 : let ctx = &ctx.with_scope_timeline(&timeline);
8088 :
8089 : // Layer visibility is an input to heatmap generation, so refresh it first
8090 1 : timeline.update_layer_visibility().await.unwrap();
8091 :
8092 1 : let heatmap = timeline
8093 1 : .generate_heatmap()
8094 1 : .await
8095 1 : .expect("Infallible while timeline is not shut down");
8096 :
8097 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
8098 :
8099 : // L0 should come last
8100 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
8101 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
8102 :
8103 1 : let mut last_lsn = Lsn::MAX;
8104 5 : for layer in heatmap_layers {
8105 : // Covered layer should be omitted
8106 4 : assert!(layer.name != covered_delta.layer_name());
8107 :
8108 4 : let layer_lsn = match &layer.name {
8109 2 : LayerName::Delta(d) => d.lsn_range.end,
8110 2 : LayerName::Image(i) => i.lsn,
8111 : };
8112 :
8113 : // Apart from L0s, newest Layers should come first
8114 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
8115 3 : assert!(layer_lsn <= last_lsn);
8116 3 : last_lsn = layer_lsn;
8117 1 : }
8118 : }
8119 :
8120 : // Evict all the layers and stash the old heatmap in the timeline.
8121 : // This simulates a migration to a cold secondary location.
8122 :
8123 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8124 1 : let mut all_layers = Vec::new();
8125 1 : let forever = std::time::Duration::from_secs(120);
8126 5 : for layer in guard.likely_resident_layers() {
8127 5 : all_layers.push(layer.clone());
8128 5 : layer.evict_and_wait(forever).await.unwrap();
8129 : }
8130 1 : drop(guard);
8131 :
8132 1 : timeline
8133 1 : .previous_heatmap
8134 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8135 1 : heatmap: heatmap.clone(),
8136 1 : read_at: std::time::Instant::now(),
8137 1 : end_lsn: None,
8138 1 : })));
8139 :
8140 : // Generate a new heatmap and assert that it contains the same layers as the old one.
8141 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
8142 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
8143 :
8144 : // Download each layer one by one. Generate the heatmap at each step and check
8145 : // that it's stable.
8146 6 : for layer in all_layers {
8147 5 : if layer.visibility() == LayerVisibilityHint::Covered {
8148 1 : continue;
8149 4 : }
8150 1 :
8151 4 : eprintln!("Downloading {layer} and re-generating heatmap");
8152 1 :
8153 4 : let ctx = &RequestContextBuilder::from(ctx)
8154 4 : .download_behavior(crate::context::DownloadBehavior::Download)
8155 4 : .attached_child();
8156 1 :
8157 4 : let _resident = layer
8158 4 : .download_and_keep_resident(ctx)
8159 4 : .instrument(tracing::info_span!(
8160 4 : parent: None,
8161 1 : "download_layer",
8162 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
8163 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
8164 1 : timeline_id = %timeline.timeline_id
8165 1 : ))
8166 4 : .await
8167 4 : .unwrap();
8168 1 :
8169 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
8170 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
8171 1 : }
8172 1 :
8173 1 : // Everything from the post-migration heatmap is now resident.
8174 1 : // Check that we drop it from memory.
8175 1 : assert!(matches!(
8176 1 : timeline.previous_heatmap.load().as_deref(),
8177 1 : Some(PreviousHeatmap::Obsolete)
8178 1 : ));
8179 1 : }
8180 :
8181 : #[tokio::test]
8182 1 : async fn test_previous_heatmap_obsoletion() {
8183 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
8184 1 : .await
8185 1 : .unwrap();
8186 :
8187 1 : let l0_delta = DeltaLayerTestDesc::new(
8188 1 : Lsn(0x20)..Lsn(0x30),
8189 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8190 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8191 1 : vec![(
8192 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8193 1 : Lsn(0x25),
8194 1 : Value::Image(test_img("foo")),
8195 1 : )],
8196 : );
8197 :
8198 1 : let image_layer = (
8199 1 : Lsn(0x40),
8200 1 : vec![(
8201 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8202 1 : test_img("bar"),
8203 1 : )],
8204 1 : );
8205 :
8206 1 : let delta_layers = vec![l0_delta];
8207 1 : let image_layers = vec![image_layer];
8208 :
8209 1 : let (tenant, ctx) = harness.load().await;
8210 1 : let timeline = tenant
8211 1 : .create_test_timeline_with_layers(
8212 1 : TimelineId::generate(),
8213 1 : Lsn(0x10),
8214 1 : PgMajorVersion::PG14,
8215 1 : &ctx,
8216 1 : Vec::new(), // in-memory layers
8217 1 : delta_layers,
8218 1 : image_layers,
8219 1 : Lsn(0x100),
8220 1 : )
8221 1 : .await
8222 1 : .unwrap();
8223 :
8224 : // Layer visibility is an input to heatmap generation, so refresh it first
8225 1 : timeline.update_layer_visibility().await.unwrap();
8226 :
8227 1 : let heatmap = timeline
8228 1 : .generate_heatmap()
8229 1 : .await
8230 1 : .expect("Infallible while timeline is not shut down");
8231 :
8232 : // Both layers should be in the heatmap
8233 1 : assert!(heatmap.all_layers().count() > 0);
8234 :
8235 : // Now simulate a migration.
8236 1 : timeline
8237 1 : .previous_heatmap
8238 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8239 1 : heatmap: heatmap.clone(),
8240 1 : read_at: std::time::Instant::now(),
8241 1 : end_lsn: None,
8242 1 : })));
8243 :
8244 : // Evict all the layers in the previous heatmap
8245 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8246 1 : let forever = std::time::Duration::from_secs(120);
8247 3 : for layer in guard.likely_resident_layers() {
8248 3 : layer.evict_and_wait(forever).await.unwrap();
8249 : }
8250 1 : drop(guard);
8251 :
8252 : // Generate a new heatmap and check that the previous heatmap
8253 : // has been marked obsolete.
8254 1 : let post_eviction_heatmap = timeline
8255 1 : .generate_heatmap()
8256 1 : .await
8257 1 : .expect("Infallible while timeline is not shut down");
8258 :
8259 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
8260 1 : assert!(matches!(
8261 1 : timeline.previous_heatmap.load().as_deref(),
8262 1 : Some(PreviousHeatmap::Obsolete)
8263 1 : ));
8264 1 : }
8265 :
8266 : #[tokio::test]
8267 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
8268 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
8269 1 : .await
8270 1 : .unwrap();
8271 :
8272 1 : let (tenant, ctx) = harness.load().await;
8273 1 : let timeline = tenant
8274 1 : .create_test_timeline(
8275 1 : TimelineId::generate(),
8276 1 : Lsn(0x10),
8277 1 : PgMajorVersion::PG14,
8278 1 : &ctx,
8279 1 : )
8280 1 : .await
8281 1 : .unwrap();
8282 :
8283 1 : let layer = find_some_layer(&timeline).await;
8284 1 : let layer = layer
8285 1 : .keep_resident()
8286 1 : .await
8287 1 : .expect("no download => no downloading errors")
8288 1 : .drop_eviction_guard();
8289 :
8290 1 : let forever = std::time::Duration::from_secs(120);
8291 :
8292 1 : let first = layer.evict_and_wait(forever);
8293 1 : let second = layer.evict_and_wait(forever);
8294 :
8295 1 : let (first, second) = tokio::join!(first, second);
8296 :
8297 1 : let res = layer.keep_resident().await;
8298 1 : assert!(res.is_none(), "{res:?}");
8299 :
8300 1 : match (first, second) {
8301 1 : (Ok(()), Ok(())) => {
8302 1 : // because there are no more timeline locks being taken on eviction path, we can
8303 1 : // witness all three outcomes here.
8304 1 : }
8305 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
8306 0 : // if one completes before the other, this is fine just as well.
8307 0 : }
8308 1 : other => unreachable!("unexpected {:?}", other),
8309 1 : }
8310 1 : }
8311 :
8312 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
8313 1 : let layers = timeline
8314 1 : .layers
8315 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
8316 1 : .await;
8317 1 : let desc = layers
8318 1 : .layer_map()
8319 1 : .unwrap()
8320 1 : .iter_historic_layers()
8321 1 : .next()
8322 1 : .expect("must find one layer to evict");
8323 :
8324 1 : layers.get_from_desc(&desc)
8325 1 : }
8326 : }
|
