Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : mod heatmap_layers_downloader;
8 : pub(crate) mod import_pgdata;
9 : mod init;
10 : pub mod layer_manager;
11 : pub(crate) mod logical_size;
12 : pub mod offload;
13 : pub mod span;
14 : pub mod uninit;
15 : mod walreceiver;
16 :
17 : use hashlink::LruCache;
18 : use std::array;
19 : use std::cmp::{max, min};
20 : use std::collections::btree_map::Entry;
21 : use std::collections::{BTreeMap, HashMap, HashSet};
22 : use std::ops::{ControlFlow, Deref, Range};
23 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
24 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
25 : use std::time::{Duration, Instant, SystemTime};
26 :
27 : use anyhow::{Context, Result, anyhow, bail, ensure};
28 : use arc_swap::{ArcSwap, ArcSwapOption};
29 : use bytes::Bytes;
30 : use camino::Utf8Path;
31 : use chrono::{DateTime, Utc};
32 : use compaction::{CompactionOutcome, GcCompactionCombinedSettings};
33 : use enumset::EnumSet;
34 : use fail::fail_point;
35 : use futures::stream::FuturesUnordered;
36 : use futures::{FutureExt, StreamExt};
37 : use handle::ShardTimelineId;
38 : use layer_manager::{
39 : LayerManagerLockHolder, LayerManagerReadGuard, LayerManagerWriteGuard, LockedLayerManager,
40 : Shutdown,
41 : };
42 :
43 : use once_cell::sync::Lazy;
44 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
45 : use pageserver_api::key::{
46 : KEY_SIZE, Key, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
47 : SPARSE_RANGE,
48 : };
49 : use pageserver_api::keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning};
50 : use pageserver_api::models::{
51 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
52 : DetachBehavior, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
53 : EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, LsnLease, PageTraceEvent, RelSizeMigration,
54 : TimelineState,
55 : };
56 : use pageserver_api::reltag::{BlockNumber, RelTag};
57 : use pageserver_api::shard::{ShardIdentity, ShardIndex, ShardNumber, TenantShardId};
58 : use postgres_connection::PgConnectionConfig;
59 : use postgres_ffi::v14::xlog_utils;
60 : use postgres_ffi::{PgMajorVersion, WAL_SEGMENT_SIZE, to_pg_timestamp};
61 : use rand::Rng;
62 : use remote_storage::DownloadError;
63 : use serde_with::serde_as;
64 : use storage_broker::BrokerClientChannel;
65 : use tokio::runtime::Handle;
66 : use tokio::sync::mpsc::Sender;
67 : use tokio::sync::{Notify, oneshot, watch};
68 : use tokio_util::sync::CancellationToken;
69 : use tracing::*;
70 : use utils::generation::Generation;
71 : use utils::guard_arc_swap::GuardArcSwap;
72 : use utils::id::TimelineId;
73 : use utils::logging::{MonitorSlowFutureCallback, monitor_slow_future};
74 : use utils::lsn::{AtomicLsn, Lsn, RecordLsn};
75 : use utils::postgres_client::PostgresClientProtocol;
76 : use utils::rate_limit::RateLimit;
77 : use utils::seqwait::SeqWait;
78 : use utils::simple_rcu::{Rcu, RcuReadGuard};
79 : use utils::sync::gate::{Gate, GateGuard};
80 : use utils::{completion, critical_timeline, fs_ext, pausable_failpoint};
81 : #[cfg(test)]
82 : use wal_decoder::models::value::Value;
83 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
84 :
85 : use self::delete::DeleteTimelineFlow;
86 : pub(super) use self::eviction_task::EvictionTaskTenantState;
87 : use self::eviction_task::EvictionTaskTimelineState;
88 : use self::logical_size::LogicalSize;
89 : use self::walreceiver::{WalReceiver, WalReceiverConf};
90 : use super::remote_timeline_client::RemoteTimelineClient;
91 : use super::remote_timeline_client::index::{GcCompactionState, IndexPart};
92 : use super::secondary::heatmap::HeatMapLayer;
93 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
94 : use super::tasks::log_compaction_error;
95 : use super::upload_queue::NotInitialized;
96 : use super::{
97 : AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
98 : debug_assert_current_span_has_tenant_and_timeline_id,
99 : };
100 : use crate::PERF_TRACE_TARGET;
101 : use crate::aux_file::AuxFileSizeEstimator;
102 : use crate::basebackup_cache::BasebackupCache;
103 : use crate::config::PageServerConf;
104 : use crate::context::{
105 : DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
106 : };
107 : use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
108 : use crate::feature_resolver::TenantFeatureResolver;
109 : use crate::keyspace::{KeyPartitioning, KeySpace};
110 : use crate::l0_flush::{self, L0FlushGlobalState};
111 : use crate::metrics::{
112 : DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_AMORTIZED_GLOBAL, LAYERS_PER_READ_BATCH_GLOBAL,
113 : LAYERS_PER_READ_GLOBAL, ScanLatencyOngoingRecording, TimelineMetrics,
114 : };
115 : use crate::page_service::TenantManagerTypes;
116 : use crate::pgdatadir_mapping::{
117 : CalculateLogicalSizeError, CollectKeySpaceError, DirectoryKind, LsnForTimestamp,
118 : MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
119 : };
120 : use crate::task_mgr::TaskKind;
121 : use crate::tenant::gc_result::GcResult;
122 : use crate::tenant::layer_map::LayerMap;
123 : use crate::tenant::metadata::TimelineMetadata;
124 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
125 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
126 : use crate::tenant::storage_layer::{
127 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
128 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
129 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
130 : ValueReconstructState, ValuesReconstructState,
131 : };
132 : use crate::tenant::tasks::BackgroundLoopKind;
133 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
134 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
135 : use crate::walingest::WalLagCooldown;
136 : use crate::walredo::RedoAttemptType;
137 : use crate::{ZERO_PAGE, task_mgr, walredo};
138 :
139 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
140 : pub(crate) enum FlushLoopState {
141 : NotStarted,
142 : Running {
143 : #[cfg(test)]
144 : expect_initdb_optimization: bool,
145 : #[cfg(test)]
146 : initdb_optimization_count: usize,
147 : },
148 : Exited,
149 : }
150 :
151 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
152 : pub enum ImageLayerCreationMode {
153 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
154 : Try,
155 : /// Force creating the image layers if possible. For now, no image layers will be created
156 : /// for metadata keys. Used in compaction code path with force flag enabled.
157 : Force,
158 : /// Initial ingestion of the data, and no data should be dropped in this function. This
159 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
160 : /// code path.
161 : Initial,
162 : }
163 :
164 : #[derive(Clone, Debug, Default)]
165 : pub enum LastImageLayerCreationStatus {
166 : Incomplete {
167 : /// The last key of the partition (exclusive) that was processed in the last
168 : /// image layer creation attempt. We will continue from this key in the next
169 : /// attempt.
170 : last_key: Key,
171 : },
172 : Complete,
173 : #[default]
174 : Initial,
175 : }
176 :
177 : impl std::fmt::Display for ImageLayerCreationMode {
178 191 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
179 191 : write!(f, "{self:?}")
180 191 : }
181 : }
182 :
183 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
184 : /// Can be removed after all refactors are done.
185 23 : fn drop_layer_manager_rlock(rlock: LayerManagerReadGuard<'_>) {
186 23 : drop(rlock)
187 23 : }
188 :
189 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
190 : /// Can be removed after all refactors are done.
191 214 : fn drop_layer_manager_wlock(rlock: LayerManagerWriteGuard<'_>) {
192 214 : drop(rlock)
193 214 : }
194 :
195 : /// The outward-facing resources required to build a Timeline
196 : pub struct TimelineResources {
197 : pub remote_client: RemoteTimelineClient,
198 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
199 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
200 : pub l0_compaction_trigger: Arc<Notify>,
201 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
202 : pub basebackup_cache: Arc<BasebackupCache>,
203 : pub feature_resolver: Arc<TenantFeatureResolver>,
204 : }
205 :
206 : pub struct Timeline {
207 : pub(crate) conf: &'static PageServerConf,
208 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
209 :
210 : myself: Weak<Self>,
211 :
212 : pub(crate) tenant_shard_id: TenantShardId,
213 : pub timeline_id: TimelineId,
214 :
215 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
216 : /// Never changes for the lifetime of this [`Timeline`] object.
217 : ///
218 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
219 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
220 : pub(crate) generation: Generation,
221 :
222 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
223 : /// to shards, and is constant through the lifetime of this Timeline.
224 : shard_identity: ShardIdentity,
225 :
226 : pub pg_version: PgMajorVersion,
227 :
228 : /// The tuple has two elements.
229 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
230 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
231 : ///
232 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
233 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
234 : ///
235 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
236 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
237 : /// `PersistentLayerDesc`'s.
238 : ///
239 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
240 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
241 : /// runtime, e.g., during page reconstruction.
242 : ///
243 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
244 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
245 : pub(crate) layers: LockedLayerManager,
246 :
247 : last_freeze_at: AtomicLsn,
248 : // Atomic would be more appropriate here.
249 : last_freeze_ts: RwLock<Instant>,
250 :
251 : pub(crate) standby_horizon: AtomicLsn,
252 :
253 : // WAL redo manager. `None` only for broken tenants.
254 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
255 :
256 : /// Remote storage client.
257 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
258 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
259 :
260 : // What page versions do we hold in the repository? If we get a
261 : // request > last_record_lsn, we need to wait until we receive all
262 : // the WAL up to the request. The SeqWait provides functions for
263 : // that. TODO: If we get a request for an old LSN, such that the
264 : // versions have already been garbage collected away, we should
265 : // throw an error, but we don't track that currently.
266 : //
267 : // last_record_lsn.load().last points to the end of last processed WAL record.
268 : //
269 : // We also remember the starting point of the previous record in
270 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
271 : // first WAL record when the node is started up. But here, we just
272 : // keep track of it.
273 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
274 :
275 : // All WAL records have been processed and stored durably on files on
276 : // local disk, up to this LSN. On crash and restart, we need to re-process
277 : // the WAL starting from this point.
278 : //
279 : // Some later WAL records might have been processed and also flushed to disk
280 : // already, so don't be surprised to see some, but there's no guarantee on
281 : // them yet.
282 : disk_consistent_lsn: AtomicLsn,
283 :
284 : // Parent timeline that this timeline was branched from, and the LSN
285 : // of the branch point.
286 : ancestor_timeline: Option<Arc<Timeline>>,
287 : ancestor_lsn: Lsn,
288 :
289 : // The LSN of gc-compaction that was last applied to this timeline.
290 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
291 :
292 : pub(crate) metrics: Arc<TimelineMetrics>,
293 :
294 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
295 : // in `crate::page_service` writes these metrics.
296 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
297 :
298 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
299 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
300 :
301 : /// Ensures layers aren't frozen by checkpointer between
302 : /// [`Timeline::get_layer_for_write`] and layer reads.
303 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
304 : /// Must always be acquired before the layer map/individual layer lock
305 : /// to avoid deadlock.
306 : ///
307 : /// The state is cleared upon freezing.
308 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
309 :
310 : /// Used to avoid multiple `flush_loop` tasks running
311 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
312 :
313 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
314 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
315 : /// The flush cycle counter is sent back on the layer_flush_done channel when
316 : /// the flush finishes. You can use that to wait for the flush to finish.
317 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
318 : /// read by whoever sends an update
319 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
320 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
321 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
322 :
323 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
324 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
325 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
326 : // the planned GC cutoff.
327 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
328 :
329 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
330 :
331 : // List of child timelines and their branch points. This is needed to avoid
332 : // garbage collecting data that is still needed by the child timelines.
333 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
334 :
335 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
336 :
337 : // It may change across major versions so for simplicity
338 : // keep it after running initdb for a timeline.
339 : // It is needed in checks when we want to error on some operations
340 : // when they are requested for pre-initdb lsn.
341 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
342 : // though let's keep them both for better error visibility.
343 : pub initdb_lsn: Lsn,
344 :
345 : /// The repartitioning result. Allows a single writer and multiple readers.
346 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
347 :
348 : /// Configuration: how often should the partitioning be recalculated.
349 : repartition_threshold: u64,
350 :
351 : last_image_layer_creation_check_at: AtomicLsn,
352 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
353 :
354 : /// Current logical size of the "datadir", at the last LSN.
355 : current_logical_size: LogicalSize,
356 :
357 : /// Information about the last processed message by the WAL receiver,
358 : /// or None if WAL receiver has not received anything for this timeline
359 : /// yet.
360 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
361 : pub walreceiver: Mutex<Option<WalReceiver>>,
362 :
363 : /// Relation size cache
364 : pub(crate) rel_size_latest_cache: RwLock<HashMap<RelTag, (Lsn, BlockNumber)>>,
365 : pub(crate) rel_size_snapshot_cache: Mutex<LruCache<(Lsn, RelTag), BlockNumber>>,
366 :
367 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
368 :
369 : state: watch::Sender<TimelineState>,
370 :
371 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
372 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
373 : pub delete_progress: TimelineDeleteProgress,
374 :
375 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
376 :
377 : /// Load or creation time information about the disk_consistent_lsn and when the loading
378 : /// happened. Used for consumption metrics.
379 : pub(crate) loaded_at: (Lsn, SystemTime),
380 :
381 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
382 : pub(crate) gate: Gate,
383 :
384 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
385 : /// to the timeline should drop out when this token fires.
386 : pub(crate) cancel: CancellationToken,
387 :
388 : /// Make sure we only have one running compaction at a time in tests.
389 : ///
390 : /// Must only be taken in two places:
391 : /// - [`Timeline::compact`] (this file)
392 : /// - [`delete::delete_local_timeline_directory`]
393 : ///
394 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
395 : compaction_lock: tokio::sync::Mutex<()>,
396 :
397 : /// If true, the last compaction failed.
398 : compaction_failed: AtomicBool,
399 :
400 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
401 : l0_compaction_trigger: Arc<Notify>,
402 :
403 : /// Make sure we only have one running gc at a time.
404 : ///
405 : /// Must only be taken in two places:
406 : /// - [`Timeline::gc`] (this file)
407 : /// - [`delete::delete_local_timeline_directory`]
408 : ///
409 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
410 : gc_lock: tokio::sync::Mutex<()>,
411 :
412 : /// Cloned from [`super::TenantShard::pagestream_throttle`] on construction.
413 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
414 :
415 : /// Size estimator for aux file v2
416 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
417 :
418 : /// Some test cases directly place keys into the timeline without actually modifying the directory
419 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
420 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
421 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
422 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
423 : #[cfg(test)]
424 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
425 :
426 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
427 :
428 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
429 :
430 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
431 :
432 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
433 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
434 :
435 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
436 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
437 :
438 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
439 :
440 : /// May host a background Tokio task which downloads all the layers from the current
441 : /// heatmap on demand.
442 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
443 :
444 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
445 :
446 : wait_lsn_log_slow: tokio::sync::Semaphore,
447 :
448 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
449 : basebackup_cache: Arc<BasebackupCache>,
450 :
451 : feature_resolver: Arc<TenantFeatureResolver>,
452 : }
453 :
454 : pub(crate) enum PreviousHeatmap {
455 : Active {
456 : heatmap: HeatMapTimeline,
457 : read_at: std::time::Instant,
458 : // End LSN covered by the heatmap if known
459 : end_lsn: Option<Lsn>,
460 : },
461 : Obsolete,
462 : }
463 :
464 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
465 :
466 : pub struct WalReceiverInfo {
467 : pub wal_source_connconf: PgConnectionConfig,
468 : pub last_received_msg_lsn: Lsn,
469 : pub last_received_msg_ts: u128,
470 : }
471 :
472 : /// Information about how much history needs to be retained, needed by
473 : /// Garbage Collection.
474 : #[derive(Default)]
475 : pub(crate) struct GcInfo {
476 : /// Specific LSNs that are needed.
477 : ///
478 : /// Currently, this includes all points where child branches have
479 : /// been forked off from. In the future, could also include
480 : /// explicit user-defined snapshot points.
481 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
482 :
483 : /// The cutoff coordinates, which are combined by selecting the minimum.
484 : pub(crate) cutoffs: GcCutoffs,
485 :
486 : /// Leases granted to particular LSNs.
487 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
488 :
489 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
490 : pub(crate) within_ancestor_pitr: bool,
491 : }
492 :
493 : impl GcInfo {
494 154 : pub(crate) fn min_cutoff(&self) -> Lsn {
495 154 : self.cutoffs.select_min()
496 154 : }
497 :
498 119 : pub(super) fn insert_child(
499 119 : &mut self,
500 119 : child_id: TimelineId,
501 119 : child_lsn: Lsn,
502 119 : is_offloaded: MaybeOffloaded,
503 119 : ) {
504 119 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
505 119 : self.retain_lsns.sort_by_key(|i| i.0);
506 119 : }
507 :
508 2 : pub(super) fn remove_child_maybe_offloaded(
509 2 : &mut self,
510 2 : child_id: TimelineId,
511 2 : maybe_offloaded: MaybeOffloaded,
512 2 : ) -> bool {
513 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
514 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
515 2 : let mut removed = false;
516 3 : self.retain_lsns.retain(|i| {
517 3 : if removed {
518 1 : return true;
519 2 : }
520 2 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
521 2 : removed |= remove;
522 2 : !remove
523 3 : });
524 2 : removed
525 2 : }
526 :
527 2 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
528 2 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
529 2 : }
530 :
531 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
532 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
533 0 : }
534 119 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
535 119 : self.leases.contains_key(&lsn)
536 119 : }
537 : }
538 :
539 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
540 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
541 : /// between time-based and space-based retention for observability and consumption metrics purposes.
542 : #[derive(Clone, Debug, Default)]
543 : pub(crate) struct GcCutoffs {
544 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
545 : /// history we must keep to retain a specified number of bytes of WAL.
546 : pub(crate) space: Lsn,
547 :
548 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates
549 : /// how much history we must keep to enable reading back at least the PITR interval duration.
550 : ///
551 : /// None indicates that the PITR cutoff has not been computed. A PITR interval of 0 will yield
552 : /// Some(last_record_lsn).
553 : pub(crate) time: Option<Lsn>,
554 : }
555 :
556 : impl GcCutoffs {
557 154 : fn select_min(&self) -> Lsn {
558 : // NB: if we haven't computed the PITR cutoff yet, we can't GC anything.
559 154 : self.space.min(self.time.unwrap_or_default())
560 154 : }
561 : }
562 :
563 : pub(crate) struct TimelineVisitOutcome {
564 : completed_keyspace: KeySpace,
565 : image_covered_keyspace: KeySpace,
566 : }
567 :
568 : /// An error happened in a get() operation.
569 : #[derive(thiserror::Error, Debug)]
570 : pub(crate) enum PageReconstructError {
571 : #[error(transparent)]
572 : Other(anyhow::Error),
573 :
574 : #[error("Ancestor LSN wait error: {0}")]
575 : AncestorLsnTimeout(WaitLsnError),
576 :
577 : #[error("timeline shutting down")]
578 : Cancelled,
579 :
580 : /// An error happened replaying WAL records
581 : #[error(transparent)]
582 : WalRedo(anyhow::Error),
583 :
584 : #[error("{0}")]
585 : MissingKey(Box<MissingKeyError>),
586 : }
587 :
588 : impl PageReconstructError {
589 0 : pub(crate) fn is_cancel(&self) -> bool {
590 0 : match self {
591 0 : PageReconstructError::Other(_) => false,
592 0 : PageReconstructError::AncestorLsnTimeout(e) => e.is_cancel(),
593 0 : PageReconstructError::Cancelled => true,
594 0 : PageReconstructError::WalRedo(_) => false,
595 0 : PageReconstructError::MissingKey(_) => false,
596 : }
597 0 : }
598 : #[allow(dead_code)] // we use the is_cancel + into_anyhow pattern in quite a few places, this one will follow soon enough
599 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
600 0 : match self {
601 0 : PageReconstructError::Other(e) => e,
602 0 : PageReconstructError::AncestorLsnTimeout(e) => e.into_anyhow(),
603 0 : PageReconstructError::Cancelled => anyhow::Error::new(self),
604 0 : PageReconstructError::WalRedo(e) => e,
605 0 : PageReconstructError::MissingKey(_) => anyhow::Error::new(self),
606 : }
607 0 : }
608 : }
609 :
610 : impl From<anyhow::Error> for PageReconstructError {
611 1 : fn from(value: anyhow::Error) -> Self {
612 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
613 1 : match value.downcast::<PageReconstructError>() {
614 0 : Ok(pre) => pre,
615 1 : Err(other) => PageReconstructError::Other(other),
616 : }
617 1 : }
618 : }
619 :
620 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
621 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
622 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
623 0 : }
624 : }
625 :
626 : impl From<layer_manager::Shutdown> for PageReconstructError {
627 0 : fn from(_: layer_manager::Shutdown) -> Self {
628 0 : PageReconstructError::Cancelled
629 0 : }
630 : }
631 :
632 : impl GetVectoredError {
633 : #[cfg(test)]
634 3 : pub(crate) fn is_missing_key_error(&self) -> bool {
635 3 : matches!(self, Self::MissingKey(_))
636 3 : }
637 : }
638 :
639 : impl From<layer_manager::Shutdown> for GetVectoredError {
640 0 : fn from(_: layer_manager::Shutdown) -> Self {
641 0 : GetVectoredError::Cancelled
642 0 : }
643 : }
644 :
645 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
646 : /// only and not used by the "real read path".
647 : pub enum ReadPathLayerId {
648 : PersistentLayer(PersistentLayerKey),
649 : InMemoryLayer(Range<Lsn>),
650 : }
651 :
652 : impl std::fmt::Display for ReadPathLayerId {
653 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
654 0 : match self {
655 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{key}"),
656 0 : ReadPathLayerId::InMemoryLayer(range) => {
657 0 : write!(f, "in-mem {}..{}", range.start, range.end)
658 : }
659 : }
660 0 : }
661 : }
662 : pub struct ReadPath {
663 : keyspace: KeySpace,
664 : lsn: Lsn,
665 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
666 : }
667 :
668 : impl ReadPath {
669 312414 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
670 312414 : Self {
671 312414 : keyspace,
672 312414 : lsn,
673 312414 : path: Vec::new(),
674 312414 : }
675 312414 : }
676 :
677 446162 : pub fn record_layer_visit(
678 446162 : &mut self,
679 446162 : layer_to_read: &ReadableLayer,
680 446162 : keyspace_to_read: &KeySpace,
681 446162 : lsn_range: &Range<Lsn>,
682 446162 : ) {
683 446162 : let id = match layer_to_read {
684 138845 : ReadableLayer::PersistentLayer(layer) => {
685 138845 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
686 : }
687 307317 : ReadableLayer::InMemoryLayer(layer) => {
688 307317 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
689 : }
690 : };
691 446162 : self.path
692 446162 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
693 446162 : }
694 : }
695 :
696 : impl std::fmt::Display for ReadPath {
697 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
698 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
699 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
700 0 : writeln!(
701 0 : f,
702 0 : "{}: {} {}..{} {}",
703 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
704 0 : )?;
705 : }
706 0 : Ok(())
707 0 : }
708 : }
709 :
710 : #[derive(thiserror::Error)]
711 : pub struct MissingKeyError {
712 : keyspace: KeySpace,
713 : shard: ShardNumber,
714 : query: Option<VersionedKeySpaceQuery>,
715 : // This is largest request LSN from the get page request batch
716 : original_hwm_lsn: Lsn,
717 : ancestor_lsn: Option<Lsn>,
718 : /// Debug information about the read path if there's an error
719 : read_path: Option<ReadPath>,
720 : backtrace: Option<std::backtrace::Backtrace>,
721 : }
722 :
723 : impl MissingKeyError {
724 7 : fn enrich(&mut self, query: VersionedKeySpaceQuery) {
725 7 : self.query = Some(query);
726 7 : }
727 : }
728 :
729 : impl std::fmt::Debug for MissingKeyError {
730 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
731 0 : write!(f, "{self}")
732 0 : }
733 : }
734 :
735 : impl std::fmt::Display for MissingKeyError {
736 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
737 0 : write!(
738 0 : f,
739 0 : "could not find data for key {} (shard {:?}), original HWM LSN {}",
740 : self.keyspace, self.shard, self.original_hwm_lsn
741 0 : )?;
742 :
743 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
744 0 : write!(f, ", ancestor {ancestor_lsn}")?;
745 0 : }
746 :
747 0 : if let Some(ref query) = self.query {
748 0 : write!(f, ", query {query}")?;
749 0 : }
750 :
751 0 : if let Some(ref read_path) = self.read_path {
752 0 : write!(f, "\n{read_path}")?;
753 0 : }
754 :
755 0 : if let Some(ref backtrace) = self.backtrace {
756 0 : write!(f, "\n{backtrace}")?;
757 0 : }
758 :
759 0 : Ok(())
760 0 : }
761 : }
762 :
763 : #[derive(thiserror::Error, Debug)]
764 : pub(crate) enum CreateImageLayersError {
765 : #[error("timeline shutting down")]
766 : Cancelled,
767 :
768 : #[error("read failed")]
769 : GetVectoredError(#[source] GetVectoredError),
770 :
771 : #[error("reconstruction failed")]
772 : PageReconstructError(#[source] PageReconstructError),
773 :
774 : #[error(transparent)]
775 : Other(anyhow::Error),
776 : }
777 :
778 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
779 0 : fn from(_: layer_manager::Shutdown) -> Self {
780 0 : CreateImageLayersError::Cancelled
781 0 : }
782 : }
783 :
784 : #[derive(thiserror::Error, Debug, Clone)]
785 : pub(crate) enum FlushLayerError {
786 : /// Timeline cancellation token was cancelled
787 : #[error("timeline shutting down")]
788 : Cancelled,
789 :
790 : /// We tried to flush a layer while the Timeline is in an unexpected state
791 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
792 : NotRunning(FlushLoopState),
793 :
794 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
795 : // loop via a watch channel, where we can only borrow it.
796 : #[error("create image layers (shared)")]
797 : CreateImageLayersError(Arc<CreateImageLayersError>),
798 :
799 : #[error("other (shared)")]
800 : Other(#[from] Arc<anyhow::Error>),
801 : }
802 :
803 : impl FlushLayerError {
804 : // When crossing from generic anyhow errors to this error type, we explicitly check
805 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
806 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
807 0 : let cancelled = timeline.cancel.is_cancelled()
808 : // The upload queue might have been shut down before the official cancellation of the timeline.
809 0 : || err
810 0 : .downcast_ref::<NotInitialized>()
811 0 : .map(NotInitialized::is_stopping)
812 0 : .unwrap_or_default();
813 0 : if cancelled {
814 0 : Self::Cancelled
815 : } else {
816 0 : Self::Other(Arc::new(err))
817 : }
818 0 : }
819 : }
820 :
821 : impl From<layer_manager::Shutdown> for FlushLayerError {
822 0 : fn from(_: layer_manager::Shutdown) -> Self {
823 0 : FlushLayerError::Cancelled
824 0 : }
825 : }
826 :
827 : #[derive(thiserror::Error, Debug)]
828 : pub enum GetVectoredError {
829 : #[error("timeline shutting down")]
830 : Cancelled,
831 :
832 : #[error("requested too many keys: {0} > {1}")]
833 : Oversized(u64, u64),
834 :
835 : #[error("requested at invalid LSN: {0}")]
836 : InvalidLsn(Lsn),
837 :
838 : #[error("requested key not found: {0}")]
839 : MissingKey(Box<MissingKeyError>),
840 :
841 : #[error("ancestry walk")]
842 : GetReadyAncestorError(#[source] GetReadyAncestorError),
843 :
844 : #[error(transparent)]
845 : Other(#[from] anyhow::Error),
846 : }
847 :
848 : impl From<GetReadyAncestorError> for GetVectoredError {
849 1 : fn from(value: GetReadyAncestorError) -> Self {
850 : use GetReadyAncestorError::*;
851 1 : match value {
852 0 : Cancelled => GetVectoredError::Cancelled,
853 : AncestorLsnTimeout(_) | BadState { .. } => {
854 1 : GetVectoredError::GetReadyAncestorError(value)
855 : }
856 : }
857 1 : }
858 : }
859 :
860 : #[derive(thiserror::Error, Debug)]
861 : pub enum GetReadyAncestorError {
862 : #[error("ancestor LSN wait error")]
863 : AncestorLsnTimeout(#[from] WaitLsnError),
864 :
865 : #[error("bad state on timeline {timeline_id}: {state:?}")]
866 : BadState {
867 : timeline_id: TimelineId,
868 : state: TimelineState,
869 : },
870 :
871 : #[error("cancelled")]
872 : Cancelled,
873 : }
874 :
875 : #[derive(Clone, Copy)]
876 : pub enum LogicalSizeCalculationCause {
877 : Initial,
878 : ConsumptionMetricsSyntheticSize,
879 : EvictionTaskImitation,
880 : TenantSizeHandler,
881 : }
882 :
883 : pub enum GetLogicalSizePriority {
884 : User,
885 : Background,
886 : }
887 :
888 : #[derive(Debug, enumset::EnumSetType)]
889 : pub(crate) enum CompactFlags {
890 : ForceRepartition,
891 : ForceImageLayerCreation,
892 : ForceL0Compaction,
893 : OnlyL0Compaction,
894 : EnhancedGcBottomMostCompaction,
895 : DryRun,
896 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
897 : /// the background compaction task, since we want to aggressively compact down L0 to bound
898 : /// read amplification.
899 : ///
900 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
901 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
902 : /// compaction).
903 : YieldForL0,
904 : }
905 :
906 : #[serde_with::serde_as]
907 0 : #[derive(Debug, Clone, serde::Deserialize)]
908 : pub(crate) struct CompactRequest {
909 : pub compact_key_range: Option<CompactKeyRange>,
910 : pub compact_lsn_range: Option<CompactLsnRange>,
911 : /// Whether the compaction job should be scheduled.
912 : #[serde(default)]
913 : pub scheduled: bool,
914 : /// Whether the compaction job should be split across key ranges.
915 : #[serde(default)]
916 : pub sub_compaction: bool,
917 : /// Max job size for each subcompaction job.
918 : pub sub_compaction_max_job_size_mb: Option<u64>,
919 : }
920 :
921 0 : #[derive(Debug, Clone, serde::Deserialize)]
922 : pub(crate) struct MarkInvisibleRequest {
923 : #[serde(default)]
924 : pub is_visible: Option<bool>,
925 : }
926 :
927 : #[derive(Debug, Clone, Default)]
928 : pub(crate) struct CompactOptions {
929 : pub flags: EnumSet<CompactFlags>,
930 : /// If set, the compaction will only compact the key range specified by this option.
931 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
932 : pub compact_key_range: Option<CompactKeyRange>,
933 : /// If set, the compaction will only compact the LSN within this value.
934 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
935 : pub compact_lsn_range: Option<CompactLsnRange>,
936 : /// Enable sub-compaction (split compaction job across key ranges).
937 : /// This option is only used by GC compaction.
938 : pub sub_compaction: bool,
939 : /// Set job size for the GC compaction.
940 : /// This option is only used by GC compaction.
941 : pub sub_compaction_max_job_size_mb: Option<u64>,
942 : }
943 :
944 : impl std::fmt::Debug for Timeline {
945 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
946 0 : write!(f, "Timeline<{}>", self.timeline_id)
947 0 : }
948 : }
949 :
950 : #[derive(thiserror::Error, Debug, Clone)]
951 : pub enum WaitLsnError {
952 : // Called on a timeline which is shutting down
953 : #[error("Shutdown")]
954 : Shutdown,
955 :
956 : // Called on an timeline not in active state or shutting down
957 : #[error("Bad timeline state: {0:?}")]
958 : BadState(TimelineState),
959 :
960 : // Timeout expired while waiting for LSN to catch up with goal.
961 : #[error("{0}")]
962 : Timeout(String),
963 : }
964 :
965 : impl WaitLsnError {
966 0 : pub(crate) fn is_cancel(&self) -> bool {
967 0 : match self {
968 0 : WaitLsnError::Shutdown => true,
969 0 : WaitLsnError::BadState(timeline_state) => match timeline_state {
970 0 : TimelineState::Loading => false,
971 0 : TimelineState::Active => false,
972 0 : TimelineState::Stopping => true,
973 0 : TimelineState::Broken { .. } => false,
974 : },
975 0 : WaitLsnError::Timeout(_) => false,
976 : }
977 0 : }
978 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
979 0 : match self {
980 0 : WaitLsnError::Shutdown => anyhow::Error::new(self),
981 0 : WaitLsnError::BadState(_) => anyhow::Error::new(self),
982 0 : WaitLsnError::Timeout(_) => anyhow::Error::new(self),
983 : }
984 0 : }
985 : }
986 :
987 : impl From<WaitLsnError> for tonic::Status {
988 0 : fn from(err: WaitLsnError) -> Self {
989 : use tonic::Code;
990 0 : let code = if err.is_cancel() {
991 0 : Code::Unavailable
992 : } else {
993 0 : Code::Internal
994 : };
995 0 : tonic::Status::new(code, err.to_string())
996 0 : }
997 : }
998 :
999 : // The impls below achieve cancellation mapping for errors.
1000 : // Perhaps there's a way of achieving this with less cruft.
1001 :
1002 : impl From<CreateImageLayersError> for CompactionError {
1003 0 : fn from(e: CreateImageLayersError) -> Self {
1004 0 : match e {
1005 0 : CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
1006 0 : CreateImageLayersError::Other(e) => {
1007 0 : CompactionError::Other(e.context("create image layers"))
1008 : }
1009 0 : _ => CompactionError::Other(e.into()),
1010 : }
1011 0 : }
1012 : }
1013 :
1014 : impl From<CreateImageLayersError> for FlushLayerError {
1015 0 : fn from(e: CreateImageLayersError) -> Self {
1016 0 : match e {
1017 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
1018 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
1019 : }
1020 0 : }
1021 : }
1022 :
1023 : impl From<PageReconstructError> for CreateImageLayersError {
1024 0 : fn from(e: PageReconstructError) -> Self {
1025 0 : match e {
1026 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
1027 0 : _ => CreateImageLayersError::PageReconstructError(e),
1028 : }
1029 0 : }
1030 : }
1031 :
1032 : impl From<super::storage_layer::errors::PutError> for CreateImageLayersError {
1033 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
1034 0 : if e.is_cancel() {
1035 0 : CreateImageLayersError::Cancelled
1036 : } else {
1037 0 : CreateImageLayersError::Other(e.into_anyhow())
1038 : }
1039 0 : }
1040 : }
1041 :
1042 : impl From<GetVectoredError> for CreateImageLayersError {
1043 0 : fn from(e: GetVectoredError) -> Self {
1044 0 : match e {
1045 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
1046 0 : _ => CreateImageLayersError::GetVectoredError(e),
1047 : }
1048 0 : }
1049 : }
1050 :
1051 : impl From<GetVectoredError> for PageReconstructError {
1052 3 : fn from(e: GetVectoredError) -> Self {
1053 3 : match e {
1054 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1055 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1056 0 : err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
1057 2 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1058 1 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1059 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1060 : }
1061 3 : }
1062 : }
1063 :
1064 : impl From<GetReadyAncestorError> for PageReconstructError {
1065 1 : fn from(e: GetReadyAncestorError) -> Self {
1066 : use GetReadyAncestorError::*;
1067 1 : match e {
1068 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1069 1 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1070 0 : Cancelled => PageReconstructError::Cancelled,
1071 : }
1072 1 : }
1073 : }
1074 :
1075 : pub(crate) enum WaitLsnTimeout {
1076 : Custom(Duration),
1077 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1078 : Default,
1079 : }
1080 :
1081 : pub(crate) enum WaitLsnWaiter<'a> {
1082 : Timeline(&'a Timeline),
1083 : Tenant,
1084 : PageService,
1085 : HttpEndpoint,
1086 : BaseBackupCache,
1087 : }
1088 :
1089 : /// Argument to [`Timeline::shutdown`].
1090 : #[derive(Debug, Clone, Copy)]
1091 : pub(crate) enum ShutdownMode {
1092 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
1093 : /// take multiple seconds for a busy timeline.
1094 : ///
1095 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1096 : /// the call to [`Timeline::shutdown`].
1097 : FreezeAndFlush,
1098 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1099 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1100 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1101 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1102 : Reload,
1103 : /// Shut down immediately, without waiting for any open layers to flush.
1104 : Hard,
1105 : }
1106 :
1107 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1108 : enum ImageLayerCreationOutcome {
1109 : /// We generated an image layer
1110 : Generated {
1111 : unfinished_image_layer: ImageLayerWriter,
1112 : },
1113 : /// The key range is empty
1114 : Empty,
1115 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1116 : /// the image layer creation.
1117 : Skip,
1118 : }
1119 :
1120 : enum RepartitionError {
1121 : Other(anyhow::Error),
1122 : CollectKeyspace(CollectKeySpaceError),
1123 : }
1124 :
1125 : impl RepartitionError {
1126 0 : fn is_cancel(&self) -> bool {
1127 0 : match self {
1128 0 : RepartitionError::Other(_) => false,
1129 0 : RepartitionError::CollectKeyspace(e) => e.is_cancel(),
1130 : }
1131 0 : }
1132 0 : fn into_anyhow(self) -> anyhow::Error {
1133 0 : match self {
1134 0 : RepartitionError::Other(e) => e,
1135 0 : RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
1136 : }
1137 0 : }
1138 : }
1139 :
1140 : /// Public interface functions
1141 : impl Timeline {
1142 : /// Get the LSN where this branch was created
1143 80 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1144 80 : self.ancestor_lsn
1145 80 : }
1146 :
1147 : /// Get the ancestor's timeline id
1148 416 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1149 416 : self.ancestor_timeline
1150 416 : .as_ref()
1151 416 : .map(|ancestor| ancestor.timeline_id)
1152 416 : }
1153 :
1154 : /// Get the ancestor timeline
1155 1 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1156 1 : self.ancestor_timeline.as_ref()
1157 1 : }
1158 :
1159 : /// Get the bytes written since the PITR cutoff on this branch, and
1160 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1161 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1162 : // TODO: for backwards compatibility, we return the full history back to 0 when the PITR
1163 : // cutoff has not yet been initialized. This should return None instead, but this is exposed
1164 : // in external HTTP APIs and callers may not handle a null value.
1165 0 : let gc_info = self.gc_info.read().unwrap();
1166 0 : let history = self
1167 0 : .get_last_record_lsn()
1168 0 : .checked_sub(gc_info.cutoffs.time.unwrap_or_default())
1169 0 : .unwrap_or_default()
1170 0 : .0;
1171 0 : (history, gc_info.within_ancestor_pitr)
1172 0 : }
1173 :
1174 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1175 425931 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1176 425931 : self.applied_gc_cutoff_lsn.read()
1177 425931 : }
1178 :
1179 : /// Read timeline's planned GC cutoff: this is the logical end of history that users are allowed
1180 : /// to read (based on configured PITR), even if physically we have more history. Returns None
1181 : /// if the PITR cutoff has not yet been initialized.
1182 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Option<Lsn> {
1183 0 : self.gc_info.read().unwrap().cutoffs.time
1184 0 : }
1185 :
1186 : /// Look up given page version.
1187 : ///
1188 : /// If a remote layer file is needed, it is downloaded as part of this
1189 : /// call.
1190 : ///
1191 : /// This method enforces [`Self::pagestream_throttle`] internally.
1192 : ///
1193 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1194 : /// abstraction above this needs to store suitable metadata to track what
1195 : /// data exists with what keys, in separate metadata entries. If a
1196 : /// non-existent key is requested, we may incorrectly return a value from
1197 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1198 : /// non-existing key.
1199 : ///
1200 : /// # Cancel-Safety
1201 : ///
1202 : /// This method is cancellation-safe.
1203 : #[inline(always)]
1204 301276 : pub(crate) async fn get(
1205 301276 : &self,
1206 301276 : key: Key,
1207 301276 : lsn: Lsn,
1208 301276 : ctx: &RequestContext,
1209 301276 : ) -> Result<Bytes, PageReconstructError> {
1210 301276 : if !lsn.is_valid() {
1211 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1212 301276 : }
1213 :
1214 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1215 : // already checked the key against the shard_identity when looking up the Timeline from
1216 : // page_service.
1217 301276 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1218 :
1219 301276 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1220 :
1221 301276 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1222 :
1223 301276 : let vectored_res = self
1224 301276 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1225 301276 : .await;
1226 :
1227 301276 : let key_value = vectored_res?.pop_first();
1228 301273 : match key_value {
1229 301267 : Some((got_key, value)) => {
1230 301267 : if got_key != key {
1231 0 : error!(
1232 0 : "Expected {}, but singular vectored get returned {}",
1233 : key, got_key
1234 : );
1235 0 : Err(PageReconstructError::Other(anyhow!(
1236 0 : "Singular vectored get returned wrong key"
1237 0 : )))
1238 : } else {
1239 301267 : value
1240 : }
1241 : }
1242 6 : None => Err(PageReconstructError::MissingKey(Box::new(
1243 6 : MissingKeyError {
1244 6 : keyspace: KeySpace::single(key..key.next()),
1245 6 : shard: self.shard_identity.get_shard_number(&key),
1246 6 : original_hwm_lsn: lsn,
1247 6 : ancestor_lsn: None,
1248 6 : backtrace: None,
1249 6 : read_path: None,
1250 6 : query: None,
1251 6 : },
1252 6 : ))),
1253 : }
1254 301276 : }
1255 :
1256 : #[inline(always)]
1257 0 : pub(crate) async fn debug_get(
1258 0 : &self,
1259 0 : key: Key,
1260 0 : lsn: Lsn,
1261 0 : ctx: &RequestContext,
1262 0 : reconstruct_state: &mut ValuesReconstructState,
1263 0 : ) -> Result<Bytes, PageReconstructError> {
1264 0 : if !lsn.is_valid() {
1265 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1266 0 : }
1267 :
1268 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1269 : // already checked the key against the shard_identity when looking up the Timeline from
1270 : // page_service.
1271 0 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1272 :
1273 0 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1274 0 : let vectored_res = self
1275 0 : .debug_get_vectored_impl(query, reconstruct_state, ctx)
1276 0 : .await;
1277 :
1278 0 : let key_value = vectored_res?.pop_first();
1279 0 : match key_value {
1280 0 : Some((got_key, value)) => {
1281 0 : if got_key != key {
1282 0 : error!(
1283 0 : "Expected {}, but singular vectored get returned {}",
1284 : key, got_key
1285 : );
1286 0 : Err(PageReconstructError::Other(anyhow!(
1287 0 : "Singular vectored get returned wrong key"
1288 0 : )))
1289 : } else {
1290 0 : value
1291 : }
1292 : }
1293 0 : None => Err(PageReconstructError::MissingKey(Box::new(
1294 0 : MissingKeyError {
1295 0 : keyspace: KeySpace::single(key..key.next()),
1296 0 : shard: self.shard_identity.get_shard_number(&key),
1297 0 : original_hwm_lsn: lsn,
1298 0 : ancestor_lsn: None,
1299 0 : backtrace: None,
1300 0 : read_path: None,
1301 0 : query: None,
1302 0 : },
1303 0 : ))),
1304 : }
1305 0 : }
1306 :
1307 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1308 :
1309 : /// Look up multiple page versions at a given LSN
1310 : ///
1311 : /// This naive implementation will be replaced with a more efficient one
1312 : /// which actually vectorizes the read path.
1313 10894 : pub(crate) async fn get_vectored(
1314 10894 : &self,
1315 10894 : query: VersionedKeySpaceQuery,
1316 10894 : io_concurrency: super::storage_layer::IoConcurrency,
1317 10894 : ctx: &RequestContext,
1318 10894 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1319 10894 : let total_keyspace = query.total_keyspace();
1320 :
1321 10894 : let key_count = total_keyspace.total_raw_size();
1322 10894 : if key_count > self.conf.max_get_vectored_keys.get() {
1323 0 : return Err(GetVectoredError::Oversized(
1324 0 : key_count as u64,
1325 0 : self.conf.max_get_vectored_keys.get() as u64,
1326 0 : ));
1327 10894 : }
1328 :
1329 34207 : for range in &total_keyspace.ranges {
1330 23313 : let mut key = range.start;
1331 65342 : while key != range.end {
1332 42029 : assert!(!self.shard_identity.is_key_disposable(&key));
1333 42029 : key = key.next();
1334 : }
1335 : }
1336 :
1337 10894 : trace!(
1338 0 : "get vectored query {} from task kind {:?}",
1339 : query,
1340 0 : ctx.task_kind(),
1341 : );
1342 :
1343 10894 : let start = crate::metrics::GET_VECTORED_LATENCY
1344 10894 : .for_task_kind(ctx.task_kind())
1345 10894 : .map(|metric| (metric, Instant::now()));
1346 :
1347 10894 : let res = self
1348 10894 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1349 10894 : .await;
1350 :
1351 10894 : if let Some((metric, start)) = start {
1352 0 : let elapsed = start.elapsed();
1353 0 : metric.observe(elapsed.as_secs_f64());
1354 10894 : }
1355 :
1356 10894 : res
1357 10894 : }
1358 :
1359 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1360 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1361 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1362 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1363 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1364 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1365 : /// the scan operation will not cause OOM in the future.
1366 8 : pub(crate) async fn scan(
1367 8 : &self,
1368 8 : keyspace: KeySpace,
1369 8 : lsn: Lsn,
1370 8 : ctx: &RequestContext,
1371 8 : io_concurrency: super::storage_layer::IoConcurrency,
1372 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1373 8 : if !lsn.is_valid() {
1374 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1375 8 : }
1376 :
1377 8 : trace!(
1378 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1379 : keyspace,
1380 : lsn,
1381 0 : ctx.task_kind()
1382 : );
1383 :
1384 : // We should generalize this into Keyspace::contains in the future.
1385 16 : for range in &keyspace.ranges {
1386 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1387 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1388 : {
1389 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1390 0 : "only metadata keyspace can be scanned"
1391 0 : )));
1392 8 : }
1393 : }
1394 :
1395 8 : let start = crate::metrics::SCAN_LATENCY
1396 8 : .for_task_kind(ctx.task_kind())
1397 8 : .map(ScanLatencyOngoingRecording::start_recording);
1398 :
1399 8 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1400 :
1401 8 : let vectored_res = self
1402 8 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1403 8 : .await;
1404 :
1405 8 : if let Some(recording) = start {
1406 0 : recording.observe();
1407 8 : }
1408 :
1409 8 : vectored_res
1410 8 : }
1411 :
1412 312414 : pub(super) async fn get_vectored_impl(
1413 312414 : &self,
1414 312414 : query: VersionedKeySpaceQuery,
1415 312414 : reconstruct_state: &mut ValuesReconstructState,
1416 312414 : ctx: &RequestContext,
1417 312414 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1418 312414 : if query.is_empty() {
1419 0 : return Ok(BTreeMap::default());
1420 312414 : }
1421 :
1422 312414 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1423 312414 : Some(ReadPath::new(
1424 312414 : query.total_keyspace(),
1425 312414 : query.high_watermark_lsn()?,
1426 : ))
1427 : } else {
1428 0 : None
1429 : };
1430 :
1431 312414 : reconstruct_state.read_path = read_path;
1432 :
1433 312414 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1434 0 : RedoAttemptType::LegacyCompaction
1435 : } else {
1436 312414 : RedoAttemptType::ReadPage
1437 : };
1438 :
1439 312414 : let traversal_res: Result<(), _> = {
1440 312414 : let ctx = RequestContextBuilder::from(ctx)
1441 312414 : .perf_span(|crnt_perf_span| {
1442 0 : info_span!(
1443 : target: PERF_TRACE_TARGET,
1444 0 : parent: crnt_perf_span,
1445 : "PLAN_IO",
1446 : )
1447 0 : })
1448 312414 : .attached_child();
1449 :
1450 312414 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1451 312414 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1452 312414 : .await
1453 : };
1454 :
1455 312414 : if let Err(err) = traversal_res {
1456 : // Wait for all the spawned IOs to complete.
1457 : // See comments on `spawn_io` inside `storage_layer` for more details.
1458 8 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1459 8 : .into_values()
1460 8 : .map(|state| state.collect_pending_ios())
1461 8 : .collect::<FuturesUnordered<_>>();
1462 8 : while collect_futs.next().await.is_some() {}
1463 :
1464 : // Enrich the missing key error with the original query.
1465 8 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1466 7 : missing_err.enrich(query.clone());
1467 7 : return Err(GetVectoredError::MissingKey(missing_err));
1468 1 : }
1469 :
1470 1 : return Err(err);
1471 312406 : };
1472 :
1473 312406 : let layers_visited = reconstruct_state.get_layers_visited();
1474 :
1475 312406 : let ctx = RequestContextBuilder::from(ctx)
1476 312406 : .perf_span(|crnt_perf_span| {
1477 0 : info_span!(
1478 : target: PERF_TRACE_TARGET,
1479 0 : parent: crnt_perf_span,
1480 : "RECONSTRUCT",
1481 : )
1482 0 : })
1483 312406 : .attached_child();
1484 :
1485 312406 : let futs = FuturesUnordered::new();
1486 363535 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1487 363535 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1488 :
1489 363535 : futs.push({
1490 363535 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1491 363535 : let ctx = RequestContextBuilder::from(&ctx)
1492 363535 : .perf_span(|crnt_perf_span| {
1493 0 : info_span!(
1494 : target: PERF_TRACE_TARGET,
1495 0 : parent: crnt_perf_span,
1496 : "RECONSTRUCT_KEY",
1497 : key = %key,
1498 : )
1499 0 : })
1500 363535 : .attached_child();
1501 :
1502 363535 : async move {
1503 363535 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1504 :
1505 363535 : let res = state
1506 363535 : .collect_pending_ios()
1507 363535 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1508 0 : info_span!(
1509 : target: PERF_TRACE_TARGET,
1510 0 : parent: crnt_perf_span,
1511 : "WAIT_FOR_IO_COMPLETIONS",
1512 : )
1513 0 : })
1514 363535 : .await;
1515 :
1516 363535 : let converted = match res {
1517 363535 : Ok(ok) => ok,
1518 0 : Err(err) => {
1519 0 : return (key, Err(err));
1520 : }
1521 : };
1522 363535 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1523 :
1524 : // The walredo module expects the records to be descending in terms of Lsn.
1525 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1526 363535 : debug_assert!(
1527 363535 : converted
1528 363535 : .records
1529 1403254 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1530 0 : "{converted:?}"
1531 : );
1532 :
1533 363535 : let walredo_deltas = converted.num_deltas();
1534 363535 : let walredo_res = walredo_self
1535 363535 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1536 363535 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1537 0 : info_span!(
1538 : target: PERF_TRACE_TARGET,
1539 0 : parent: crnt_perf_span,
1540 : "WALREDO",
1541 : deltas = %walredo_deltas,
1542 : )
1543 0 : })
1544 363535 : .await;
1545 :
1546 363535 : (key, walredo_res)
1547 363535 : }
1548 : });
1549 : }
1550 :
1551 312406 : let results = futs
1552 312406 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1553 312406 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1554 312406 : .await;
1555 :
1556 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1557 : // when they're missing. Instead they are omitted from the resulting btree
1558 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1559 : // to avoid infinite results.
1560 312406 : if !results.is_empty() {
1561 312227 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1562 0 : let total_keyspace = query.total_keyspace();
1563 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1564 :
1565 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1566 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1567 0 : LOG_PACER.lock().unwrap().call(|| {
1568 0 : let num_keys = total_keyspace.total_raw_size();
1569 0 : let num_pages = results.len();
1570 0 : tracing::info!(
1571 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1572 : lsn = %max_request_lsn,
1573 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1574 : );
1575 0 : });
1576 312227 : }
1577 :
1578 : // Records the number of layers visited in a few different ways:
1579 : //
1580 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1581 : // layer directly affects its observed latency.
1582 : //
1583 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1584 : // layer visits and access cost.
1585 : //
1586 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1587 : // read amplification after batching.
1588 312227 : let layers_visited = layers_visited as f64;
1589 312227 : let avg_layers_visited = layers_visited / results.len() as f64;
1590 312227 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1591 675762 : for _ in &results {
1592 363535 : self.metrics.layers_per_read.observe(layers_visited);
1593 363535 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1594 363535 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1595 363535 : }
1596 179 : }
1597 :
1598 312406 : Ok(results)
1599 312414 : }
1600 :
1601 : // A copy of the get_vectored_impl method except that we store the image and wal records into `reconstruct_state`.
1602 : // This is only used in the http getpage call for debugging purpose.
1603 0 : pub(super) async fn debug_get_vectored_impl(
1604 0 : &self,
1605 0 : query: VersionedKeySpaceQuery,
1606 0 : reconstruct_state: &mut ValuesReconstructState,
1607 0 : ctx: &RequestContext,
1608 0 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1609 0 : if query.is_empty() {
1610 0 : return Ok(BTreeMap::default());
1611 0 : }
1612 :
1613 0 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1614 0 : Some(ReadPath::new(
1615 0 : query.total_keyspace(),
1616 0 : query.high_watermark_lsn()?,
1617 : ))
1618 : } else {
1619 0 : None
1620 : };
1621 :
1622 0 : reconstruct_state.read_path = read_path;
1623 :
1624 0 : let traversal_res: Result<(), _> = self
1625 0 : .get_vectored_reconstruct_data(query.clone(), reconstruct_state, ctx)
1626 0 : .await;
1627 :
1628 0 : if let Err(err) = traversal_res {
1629 : // Wait for all the spawned IOs to complete.
1630 : // See comments on `spawn_io` inside `storage_layer` for more details.
1631 0 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1632 0 : .into_values()
1633 0 : .map(|state| state.collect_pending_ios())
1634 0 : .collect::<FuturesUnordered<_>>();
1635 0 : while collect_futs.next().await.is_some() {}
1636 0 : return Err(err);
1637 0 : };
1638 :
1639 0 : let reconstruct_state = Arc::new(Mutex::new(reconstruct_state));
1640 0 : let futs = FuturesUnordered::new();
1641 :
1642 0 : for (key, state) in std::mem::take(&mut reconstruct_state.lock().unwrap().keys) {
1643 0 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1644 0 : futs.push({
1645 0 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1646 0 : let rc_clone = Arc::clone(&reconstruct_state);
1647 :
1648 0 : async move {
1649 0 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1650 :
1651 0 : let converted = match state.collect_pending_ios().await {
1652 0 : Ok(ok) => ok,
1653 0 : Err(err) => {
1654 0 : return (key, Err(err));
1655 : }
1656 : };
1657 0 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1658 :
1659 : // The walredo module expects the records to be descending in terms of Lsn.
1660 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1661 0 : debug_assert!(
1662 0 : converted
1663 0 : .records
1664 0 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1665 0 : "{converted:?}"
1666 : );
1667 0 : {
1668 0 : let mut guard = rc_clone.lock().unwrap();
1669 0 : guard.set_debug_state(&converted);
1670 0 : }
1671 : (
1672 0 : key,
1673 0 : walredo_self
1674 0 : .reconstruct_value(
1675 0 : key,
1676 0 : req_lsn_for_key,
1677 0 : converted,
1678 0 : RedoAttemptType::ReadPage,
1679 0 : )
1680 0 : .await,
1681 : )
1682 0 : }
1683 : });
1684 : }
1685 :
1686 0 : let results = futs
1687 0 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1688 0 : .await;
1689 :
1690 0 : Ok(results)
1691 0 : }
1692 :
1693 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1694 138202 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1695 138202 : self.last_record_lsn.load().last
1696 138202 : }
1697 :
1698 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1699 0 : self.last_record_lsn.load().prev
1700 0 : }
1701 :
1702 : /// Atomically get both last and prev.
1703 117 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1704 117 : self.last_record_lsn.load()
1705 117 : }
1706 :
1707 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1708 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1709 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1710 0 : self.last_record_lsn.status_receiver()
1711 0 : }
1712 :
1713 609 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1714 609 : self.disk_consistent_lsn.load()
1715 609 : }
1716 :
1717 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1718 : /// not validated with control plane yet.
1719 : /// See [`Self::get_remote_consistent_lsn_visible`].
1720 2 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1721 2 : self.remote_client.remote_consistent_lsn_projected()
1722 2 : }
1723 :
1724 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1725 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1726 : /// generation validation in the deletion queue.
1727 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1728 0 : self.remote_client.remote_consistent_lsn_visible()
1729 0 : }
1730 :
1731 : /// The sum of the file size of all historic layers in the layer map.
1732 : /// This method makes no distinction between local and remote layers.
1733 : /// Hence, the result **does not represent local filesystem usage**.
1734 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1735 0 : let guard = self
1736 0 : .layers
1737 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1738 0 : .await;
1739 0 : guard.layer_size_sum()
1740 0 : }
1741 :
1742 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1743 0 : self.metrics.resident_physical_size_get()
1744 0 : }
1745 :
1746 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1747 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1748 0 : }
1749 :
1750 : ///
1751 : /// Wait until WAL has been received and processed up to this LSN.
1752 : ///
1753 : /// You should call this before any of the other get_* or list_* functions. Calling
1754 : /// those functions with an LSN that has been processed yet is an error.
1755 : ///
1756 112980 : pub(crate) async fn wait_lsn(
1757 112980 : &self,
1758 112980 : lsn: Lsn,
1759 112980 : who_is_waiting: WaitLsnWaiter<'_>,
1760 112980 : timeout: WaitLsnTimeout,
1761 112980 : ctx: &RequestContext, /* Prepare for use by cancellation */
1762 112980 : ) -> Result<(), WaitLsnError> {
1763 112980 : let state = self.current_state();
1764 112980 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1765 0 : return Err(WaitLsnError::Shutdown);
1766 112980 : } else if !matches!(state, TimelineState::Active) {
1767 0 : return Err(WaitLsnError::BadState(state));
1768 112980 : }
1769 :
1770 112980 : if cfg!(debug_assertions) {
1771 112980 : match ctx.task_kind() {
1772 : TaskKind::WalReceiverManager
1773 : | TaskKind::WalReceiverConnectionHandler
1774 : | TaskKind::WalReceiverConnectionPoller => {
1775 0 : let is_myself = match who_is_waiting {
1776 0 : WaitLsnWaiter::Timeline(waiter) => {
1777 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1778 : }
1779 : WaitLsnWaiter::Tenant
1780 : | WaitLsnWaiter::PageService
1781 : | WaitLsnWaiter::HttpEndpoint
1782 0 : | WaitLsnWaiter::BaseBackupCache => unreachable!(
1783 : "tenant or page_service context are not expected to have task kind {:?}",
1784 0 : ctx.task_kind()
1785 : ),
1786 : };
1787 0 : if is_myself {
1788 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1789 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1790 0 : panic!(
1791 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1792 : );
1793 0 : }
1794 0 : } else {
1795 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1796 0 : // our walreceiver task can make progress independent of theirs
1797 0 : }
1798 : }
1799 112980 : _ => {}
1800 : }
1801 0 : }
1802 :
1803 112980 : let timeout = match timeout {
1804 0 : WaitLsnTimeout::Custom(t) => t,
1805 112980 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1806 : };
1807 :
1808 112980 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1809 112980 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1810 :
1811 112980 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1812 112980 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1813 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1814 112980 : let log_slow_threshold = Duration::from_secs(1);
1815 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1816 112980 : let log_slow_period = Duration::from_secs(10);
1817 112980 : let mut logging_permit = None;
1818 112980 : let wait_for_timeout = monitor_slow_future(
1819 112980 : log_slow_threshold,
1820 112980 : log_slow_period,
1821 112980 : wait_for_timeout,
1822 : |MonitorSlowFutureCallback {
1823 : ready,
1824 : is_slow,
1825 : elapsed_total,
1826 : elapsed_since_last_callback,
1827 112980 : }| {
1828 112980 : self.metrics
1829 112980 : .wait_lsn_in_progress_micros
1830 112980 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1831 112980 : if !is_slow {
1832 112980 : return;
1833 0 : }
1834 : // It's slow, see if we should log it.
1835 : // (We limit the logging to one per invocation per timeline to avoid excessive
1836 : // logging during an extended broker / networking outage that affects all timelines.)
1837 0 : if logging_permit.is_none() {
1838 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1839 0 : }
1840 0 : if logging_permit.is_none() {
1841 0 : return;
1842 0 : }
1843 : // We log it.
1844 0 : if ready {
1845 0 : info!(
1846 0 : "slow wait_lsn completed after {:.3}s",
1847 0 : elapsed_total.as_secs_f64()
1848 : );
1849 : } else {
1850 0 : info!(
1851 0 : "slow wait_lsn still running for {:.3}s",
1852 0 : elapsed_total.as_secs_f64()
1853 : );
1854 : }
1855 112980 : },
1856 : );
1857 112980 : let res = wait_for_timeout.await;
1858 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1859 112980 : drop(logging_permit);
1860 112980 : drop(start_finish_counterpair_guard);
1861 112980 : drop(timer);
1862 112980 : match res {
1863 112980 : Ok(()) => Ok(()),
1864 0 : Err(e) => {
1865 : use utils::seqwait::SeqWaitError::*;
1866 0 : match e {
1867 0 : Shutdown => Err(WaitLsnError::Shutdown),
1868 : Timeout => {
1869 0 : let walreceiver_status = self.walreceiver_status();
1870 0 : Err(WaitLsnError::Timeout(format!(
1871 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1872 0 : lsn,
1873 0 : self.get_last_record_lsn(),
1874 0 : self.get_disk_consistent_lsn(),
1875 0 : walreceiver_status,
1876 0 : )))
1877 : }
1878 : }
1879 : }
1880 : }
1881 112980 : }
1882 :
1883 0 : pub(crate) fn walreceiver_status(&self) -> String {
1884 0 : match &*self.walreceiver.lock().unwrap() {
1885 0 : None => "stopping or stopped".to_string(),
1886 0 : Some(walreceiver) => match walreceiver.status() {
1887 0 : Some(status) => status.to_human_readable_string(),
1888 0 : None => "Not active".to_string(),
1889 : },
1890 : }
1891 0 : }
1892 :
1893 : /// Check that it is valid to request operations with that lsn.
1894 119 : pub(crate) fn check_lsn_is_in_scope(
1895 119 : &self,
1896 119 : lsn: Lsn,
1897 119 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1898 119 : ) -> anyhow::Result<()> {
1899 119 : ensure!(
1900 119 : lsn >= **latest_gc_cutoff_lsn,
1901 2 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1902 : lsn,
1903 2 : **latest_gc_cutoff_lsn,
1904 : );
1905 117 : Ok(())
1906 119 : }
1907 :
1908 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1909 5 : pub(crate) fn init_lsn_lease(
1910 5 : &self,
1911 5 : lsn: Lsn,
1912 5 : length: Duration,
1913 5 : ctx: &RequestContext,
1914 5 : ) -> anyhow::Result<LsnLease> {
1915 5 : self.make_lsn_lease(lsn, length, true, ctx)
1916 5 : }
1917 :
1918 : /// 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.
1919 2 : pub(crate) fn renew_lsn_lease(
1920 2 : &self,
1921 2 : lsn: Lsn,
1922 2 : length: Duration,
1923 2 : ctx: &RequestContext,
1924 2 : ) -> anyhow::Result<LsnLease> {
1925 2 : self.make_lsn_lease(lsn, length, false, ctx)
1926 2 : }
1927 :
1928 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1929 : ///
1930 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1931 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1932 : ///
1933 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1934 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1935 7 : fn make_lsn_lease(
1936 7 : &self,
1937 7 : lsn: Lsn,
1938 7 : length: Duration,
1939 7 : init: bool,
1940 7 : _ctx: &RequestContext,
1941 7 : ) -> anyhow::Result<LsnLease> {
1942 6 : let lease = {
1943 : // Normalize the requested LSN to be aligned, and move to the first record
1944 : // if it points to the beginning of the page (header).
1945 7 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1946 :
1947 7 : let mut gc_info = self.gc_info.write().unwrap();
1948 7 : let planned_cutoff = gc_info.min_cutoff();
1949 :
1950 7 : let valid_until = SystemTime::now() + length;
1951 :
1952 7 : let entry = gc_info.leases.entry(lsn);
1953 :
1954 7 : match entry {
1955 3 : Entry::Occupied(mut occupied) => {
1956 3 : let existing_lease = occupied.get_mut();
1957 3 : if valid_until > existing_lease.valid_until {
1958 1 : existing_lease.valid_until = valid_until;
1959 1 : let dt: DateTime<Utc> = valid_until.into();
1960 1 : info!("lease extended to {}", dt);
1961 : } else {
1962 2 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1963 2 : info!("existing lease covers greater length, valid until {}", dt);
1964 : }
1965 :
1966 3 : existing_lease.clone()
1967 : }
1968 4 : Entry::Vacant(vacant) => {
1969 : // Never allow a lease to be requested for an LSN below the applied GC cutoff. The data could have been deleted.
1970 4 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1971 4 : if lsn < *latest_gc_cutoff_lsn {
1972 1 : bail!(
1973 1 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1974 : lsn,
1975 1 : *latest_gc_cutoff_lsn
1976 : );
1977 3 : }
1978 :
1979 : // We allow create lease for those below the planned gc cutoff if we are still within the grace period
1980 : // of GC blocking.
1981 3 : let validate = {
1982 3 : let conf = self.tenant_conf.load();
1983 3 : !conf.is_gc_blocked_by_lsn_lease_deadline()
1984 : };
1985 :
1986 : // Do not allow initial lease creation to be below the planned gc cutoff. The client (compute_ctl) determines
1987 : // whether it is a initial lease creation or a renewal.
1988 3 : if (init || validate) && lsn < planned_cutoff {
1989 0 : bail!(
1990 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1991 : lsn,
1992 : planned_cutoff
1993 : );
1994 3 : }
1995 :
1996 3 : let dt: DateTime<Utc> = valid_until.into();
1997 3 : info!("lease created, valid until {}", dt);
1998 3 : vacant.insert(LsnLease { valid_until }).clone()
1999 : }
2000 : }
2001 : };
2002 :
2003 6 : Ok(lease)
2004 7 : }
2005 :
2006 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2007 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2008 : #[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))]
2009 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
2010 : self.freeze0().await
2011 : }
2012 :
2013 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
2014 : #[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))]
2015 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
2016 : self.freeze_and_flush0().await
2017 : }
2018 :
2019 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2020 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2021 571 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
2022 571 : let mut g = self.write_lock.lock().await;
2023 571 : let to_lsn = self.get_last_record_lsn();
2024 571 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
2025 571 : }
2026 :
2027 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
2028 : // polluting the span hierarchy.
2029 571 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
2030 571 : let token = self.freeze0().await?;
2031 571 : self.wait_flush_completion(token).await
2032 571 : }
2033 :
2034 : // Check if an open ephemeral layer should be closed: this provides
2035 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
2036 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
2037 : // ephemeral layer bytes has been breached.
2038 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
2039 0 : debug_assert_current_span_has_tenant_and_timeline_id();
2040 :
2041 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
2042 : // If the write lock is held, there is an active wal receiver: rolling open layers
2043 : // is their responsibility while they hold this lock.
2044 0 : return;
2045 : };
2046 :
2047 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
2048 : // time, and this was missed.
2049 : // if write_guard.is_none() { return; }
2050 :
2051 0 : let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
2052 : // Don't block if the layer lock is busy
2053 0 : return;
2054 : };
2055 :
2056 0 : let Ok(lm) = layers_guard.layer_map() else {
2057 0 : return;
2058 : };
2059 :
2060 0 : let Some(open_layer) = &lm.open_layer else {
2061 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
2062 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
2063 : // that didn't result in writes to this shard.
2064 :
2065 : // Must not hold the layers lock while waiting for a flush.
2066 0 : drop(layers_guard);
2067 :
2068 0 : let last_record_lsn = self.get_last_record_lsn();
2069 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
2070 0 : if last_record_lsn > disk_consistent_lsn {
2071 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
2072 : // we are a sharded tenant and have skipped some WAL
2073 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
2074 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
2075 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
2076 : // without any data ingested (yet) doesn't write a remote index as soon as it
2077 : // sees its LSN advance: we only do this if we've been layer-less
2078 : // for some time.
2079 0 : tracing::debug!(
2080 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
2081 : disk_consistent_lsn,
2082 : last_record_lsn
2083 : );
2084 :
2085 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
2086 : // We know there is no open layer, so we can request freezing without actually
2087 : // freezing anything. This is true even if we have dropped the layers_guard, we
2088 : // still hold the write_guard.
2089 0 : let _ = async {
2090 0 : let token = self
2091 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
2092 0 : .await?;
2093 0 : self.wait_flush_completion(token).await
2094 0 : }
2095 0 : .await;
2096 0 : }
2097 0 : }
2098 :
2099 0 : return;
2100 : };
2101 :
2102 0 : let current_size = open_layer.len();
2103 :
2104 0 : let current_lsn = self.get_last_record_lsn();
2105 :
2106 0 : let checkpoint_distance_override = open_layer.tick();
2107 :
2108 0 : if let Some(size_override) = checkpoint_distance_override {
2109 0 : if current_size > size_override {
2110 : // This is not harmful, but it only happens in relatively rare cases where
2111 : // time-based checkpoints are not happening fast enough to keep the amount of
2112 : // ephemeral data within configured limits. It's a sign of stress on the system.
2113 0 : tracing::info!(
2114 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
2115 : );
2116 0 : }
2117 0 : }
2118 :
2119 0 : let checkpoint_distance =
2120 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
2121 :
2122 0 : if self.should_roll(
2123 0 : current_size,
2124 0 : current_size,
2125 0 : checkpoint_distance,
2126 0 : self.get_last_record_lsn(),
2127 0 : self.last_freeze_at.load(),
2128 0 : open_layer.get_opened_at(),
2129 : ) {
2130 0 : match open_layer.info() {
2131 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
2132 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
2133 : // happens asynchronously in the background.
2134 0 : tracing::debug!(
2135 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
2136 : );
2137 : }
2138 : InMemoryLayerInfo::Open { .. } => {
2139 : // Upgrade to a write lock and freeze the layer
2140 0 : drop(layers_guard);
2141 0 : let res = self
2142 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
2143 0 : .await;
2144 :
2145 0 : if let Err(e) = res {
2146 0 : tracing::info!(
2147 0 : "failed to flush frozen layer after background freeze: {e:#}"
2148 : );
2149 0 : }
2150 : }
2151 : }
2152 0 : }
2153 0 : }
2154 :
2155 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
2156 : ///
2157 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
2158 : /// child timelines that are not offloaded yet.
2159 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
2160 0 : if self.remote_client.is_archived() != Some(true) {
2161 0 : return (false, "the timeline is not archived");
2162 0 : }
2163 0 : if !self.remote_client.no_pending_work() {
2164 : // if the remote client is still processing some work, we can't offload
2165 0 : return (false, "the upload queue is not drained yet");
2166 0 : }
2167 :
2168 0 : (true, "ok")
2169 0 : }
2170 :
2171 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
2172 : /// compaction tasks.
2173 192 : pub(crate) async fn compact(
2174 192 : self: &Arc<Self>,
2175 192 : cancel: &CancellationToken,
2176 192 : flags: EnumSet<CompactFlags>,
2177 192 : ctx: &RequestContext,
2178 192 : ) -> Result<CompactionOutcome, CompactionError> {
2179 192 : let res = self
2180 192 : .compact_with_options(
2181 192 : cancel,
2182 192 : CompactOptions {
2183 192 : flags,
2184 192 : compact_key_range: None,
2185 192 : compact_lsn_range: None,
2186 192 : sub_compaction: false,
2187 192 : sub_compaction_max_job_size_mb: None,
2188 192 : },
2189 192 : ctx,
2190 192 : )
2191 192 : .await;
2192 192 : if let Err(err) = &res {
2193 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
2194 192 : }
2195 192 : res
2196 192 : }
2197 :
2198 : /// Outermost timeline compaction operation; downloads needed layers.
2199 : ///
2200 : /// NB: the cancellation token is usually from a background task, but can also come from a
2201 : /// request task.
2202 192 : pub(crate) async fn compact_with_options(
2203 192 : self: &Arc<Self>,
2204 192 : cancel: &CancellationToken,
2205 192 : options: CompactOptions,
2206 192 : ctx: &RequestContext,
2207 192 : ) -> Result<CompactionOutcome, CompactionError> {
2208 : // Acquire the compaction lock and task semaphore.
2209 : //
2210 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
2211 : // out by other background tasks (including image compaction). We request this via
2212 : // `BackgroundLoopKind::L0Compaction`.
2213 : //
2214 : // Yield for pending L0 compaction while waiting for the semaphore.
2215 192 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
2216 192 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
2217 0 : true => BackgroundLoopKind::L0Compaction,
2218 192 : false => BackgroundLoopKind::Compaction,
2219 : };
2220 192 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2221 192 : if yield_for_l0 {
2222 : // If this is an L0 pass, it doesn't make sense to yield for L0.
2223 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2224 : // If `compaction_l0_first` is disabled, there's no point yielding.
2225 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2226 192 : }
2227 :
2228 192 : let acquire = async move {
2229 192 : let guard = self.compaction_lock.lock().await;
2230 192 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2231 192 : (guard, permit)
2232 192 : };
2233 :
2234 192 : let (_guard, _permit) = tokio::select! {
2235 192 : (guard, permit) = acquire => (guard, permit),
2236 192 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2237 0 : return Ok(CompactionOutcome::YieldForL0);
2238 : }
2239 192 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2240 192 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2241 : };
2242 :
2243 192 : let last_record_lsn = self.get_last_record_lsn();
2244 :
2245 : // Last record Lsn could be zero in case the timeline was just created
2246 192 : if !last_record_lsn.is_valid() {
2247 0 : warn!(
2248 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2249 : );
2250 0 : return Ok(CompactionOutcome::Skipped);
2251 192 : }
2252 :
2253 192 : let result = match self.get_compaction_algorithm_settings().kind {
2254 : CompactionAlgorithm::Tiered => {
2255 0 : self.compact_tiered(cancel, ctx).await?;
2256 0 : Ok(CompactionOutcome::Done)
2257 : }
2258 192 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2259 : };
2260 :
2261 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2262 0 : match &result {
2263 192 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2264 0 : Err(e) if e.is_cancel() => {}
2265 0 : Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
2266 : };
2267 :
2268 192 : result
2269 192 : }
2270 :
2271 : /// Mutate the timeline with a [`TimelineWriter`].
2272 2566609 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2273 : TimelineWriter {
2274 2566609 : tl: self,
2275 2566609 : write_guard: self.write_lock.lock().await,
2276 : }
2277 2566609 : }
2278 :
2279 0 : pub(crate) fn activate(
2280 0 : self: &Arc<Self>,
2281 0 : parent: Arc<crate::tenant::TenantShard>,
2282 0 : broker_client: BrokerClientChannel,
2283 0 : background_jobs_can_start: Option<&completion::Barrier>,
2284 0 : ctx: &RequestContext,
2285 0 : ) {
2286 0 : if self.tenant_shard_id.is_shard_zero() {
2287 0 : // Logical size is only maintained accurately on shard zero.
2288 0 : self.spawn_initial_logical_size_computation_task(ctx);
2289 0 : }
2290 0 : self.launch_wal_receiver(ctx, broker_client);
2291 0 : self.set_state(TimelineState::Active);
2292 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2293 0 : }
2294 :
2295 : /// After this function returns, there are no timeline-scoped tasks are left running.
2296 : ///
2297 : /// The preferred pattern for is:
2298 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2299 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2300 : /// go the extra mile and keep track of JoinHandles
2301 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2302 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2303 : ///
2304 : /// For legacy reasons, we still have multiple tasks spawned using
2305 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2306 : /// We refer to these as "timeline-scoped task_mgr tasks".
2307 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2308 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2309 : /// or [`task_mgr::shutdown_watcher`].
2310 : /// We want to gradually convert the code base away from these.
2311 : ///
2312 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2313 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2314 : /// ones that aren't mentioned here):
2315 : /// - [`TaskKind::TimelineDeletionWorker`]
2316 : /// - NB: also used for tenant deletion
2317 : /// - [`TaskKind::RemoteUploadTask`]`
2318 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2319 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2320 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2321 : /// - [`TaskKind::Eviction`]
2322 : /// - [`TaskKind::LayerFlushTask`]
2323 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2324 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2325 5 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2326 5 : debug_assert_current_span_has_tenant_and_timeline_id();
2327 :
2328 : // Regardless of whether we're going to try_freeze_and_flush
2329 : // cancel walreceiver to stop ingesting more data asap.
2330 : //
2331 : // Note that we're accepting a race condition here where we may
2332 : // do the final flush below, before walreceiver observes the
2333 : // cancellation and exits.
2334 : // This means we may open a new InMemoryLayer after the final flush below.
2335 : // Flush loop is also still running for a short while, so, in theory, it
2336 : // could also make its way into the upload queue.
2337 : //
2338 : // If we wait for the shutdown of the walreceiver before moving on to the
2339 : // flush, then that would be avoided. But we don't do it because the
2340 : // walreceiver entertains reads internally, which means that it possibly
2341 : // depends on the download of layers. Layer download is only sensitive to
2342 : // the cancellation of the entire timeline, so cancelling the walreceiver
2343 : // will have no effect on the individual get requests.
2344 : // This would cause problems when there is a lot of ongoing downloads or
2345 : // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
2346 : // and we can't deallocate resources of the timeline, etc.
2347 5 : let walreceiver = self.walreceiver.lock().unwrap().take();
2348 5 : tracing::debug!(
2349 0 : is_some = walreceiver.is_some(),
2350 0 : "Waiting for WalReceiverManager..."
2351 : );
2352 5 : if let Some(walreceiver) = walreceiver {
2353 0 : walreceiver.cancel().await;
2354 5 : }
2355 : // ... and inform any waiters for newer LSNs that there won't be any.
2356 5 : self.last_record_lsn.shutdown();
2357 :
2358 5 : if let ShutdownMode::FreezeAndFlush = mode {
2359 3 : let do_flush = if let Some((open, frozen)) = self
2360 3 : .layers
2361 3 : .read(LayerManagerLockHolder::Shutdown)
2362 3 : .await
2363 3 : .layer_map()
2364 3 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2365 3 : .ok()
2366 3 : .filter(|(open, frozen)| *open || *frozen > 0)
2367 : {
2368 0 : if self.remote_client.is_archived() == Some(true) {
2369 : // No point flushing on shutdown for an archived timeline: it is not important
2370 : // to have it nice and fresh after our restart, and trying to flush here might
2371 : // race with trying to offload it (which also stops the flush loop)
2372 0 : false
2373 : } else {
2374 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2375 0 : true
2376 : }
2377 : } else {
2378 : // this is double-shutdown, it'll be a no-op
2379 3 : true
2380 : };
2381 :
2382 : // we shut down walreceiver above, so, we won't add anything more
2383 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2384 : // to reach the disk & upload queue, then shut the upload queue and
2385 : // wait for it to drain.
2386 3 : if do_flush {
2387 3 : match self.freeze_and_flush().await {
2388 : Ok(_) => {
2389 : // drain the upload queue
2390 : // if we did not wait for completion here, it might be our shutdown process
2391 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2392 : // be spawned.
2393 : //
2394 : // what is problematic is the shutting down of RemoteTimelineClient, because
2395 : // obviously it does not make sense to stop while we wait for it, but what
2396 : // about corner cases like s3 suddenly hanging up?
2397 3 : self.remote_client.shutdown().await;
2398 : }
2399 : Err(FlushLayerError::Cancelled) => {
2400 : // this is likely the second shutdown, ignore silently.
2401 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2402 0 : debug_assert!(self.cancel.is_cancelled());
2403 : }
2404 0 : Err(e) => {
2405 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2406 : // we have some extra WAL replay to do next time the timeline starts.
2407 0 : warn!("failed to freeze and flush: {e:#}");
2408 : }
2409 : }
2410 :
2411 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2412 : // we also do a final check here to ensure that the queue is empty.
2413 3 : if !self.remote_client.no_pending_work() {
2414 0 : warn!(
2415 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2416 : );
2417 3 : }
2418 0 : }
2419 2 : }
2420 :
2421 5 : if let ShutdownMode::Reload = mode {
2422 : // drain the upload queue
2423 1 : self.remote_client.shutdown().await;
2424 1 : if !self.remote_client.no_pending_work() {
2425 0 : warn!(
2426 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2427 : );
2428 1 : }
2429 4 : }
2430 :
2431 : // Signal any subscribers to our cancellation token to drop out
2432 5 : tracing::debug!("Cancelling CancellationToken");
2433 5 : self.cancel.cancel();
2434 :
2435 : // If we have a background task downloading heatmap layers stop it.
2436 : // The background downloads are sensitive to timeline cancellation (done above),
2437 : // so the drain will be immediate.
2438 5 : self.stop_and_drain_heatmap_layers_download().await;
2439 :
2440 : // Ensure Prevent new page service requests from starting.
2441 5 : self.handles.shutdown();
2442 :
2443 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2444 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2445 5 : self.remote_client.stop();
2446 :
2447 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2448 : // to shut down the upload queue tasks.
2449 : // TODO: fix that, task management should be encapsulated inside remote_client.
2450 5 : task_mgr::shutdown_tasks(
2451 5 : Some(TaskKind::RemoteUploadTask),
2452 5 : Some(self.tenant_shard_id),
2453 5 : Some(self.timeline_id),
2454 5 : )
2455 5 : .await;
2456 :
2457 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2458 5 : tracing::debug!("Waiting for tasks...");
2459 5 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2460 :
2461 : {
2462 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2463 : // open.
2464 5 : let mut write_guard = self.write_lock.lock().await;
2465 5 : self.layers
2466 5 : .write(LayerManagerLockHolder::Shutdown)
2467 5 : .await
2468 5 : .shutdown(&mut write_guard);
2469 : }
2470 :
2471 : // Finally wait until any gate-holders are complete.
2472 : //
2473 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2474 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2475 5 : self.gate.close().await;
2476 :
2477 5 : self.metrics.shutdown();
2478 5 : }
2479 :
2480 236 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2481 236 : match (self.current_state(), new_state) {
2482 236 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2483 1 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2484 : }
2485 0 : (st, TimelineState::Loading) => {
2486 0 : error!("ignoring transition from {st:?} into Loading state");
2487 : }
2488 0 : (TimelineState::Broken { .. }, new_state) => {
2489 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2490 : }
2491 : (TimelineState::Stopping, TimelineState::Active) => {
2492 0 : error!("Not activating a Stopping timeline");
2493 : }
2494 235 : (_, new_state) => {
2495 235 : self.state.send_replace(new_state);
2496 235 : }
2497 : }
2498 236 : }
2499 :
2500 1 : pub(crate) fn set_broken(&self, reason: String) {
2501 1 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2502 1 : let broken_state = TimelineState::Broken {
2503 1 : reason,
2504 1 : backtrace: backtrace_str,
2505 1 : };
2506 1 : self.set_state(broken_state);
2507 :
2508 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2509 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2510 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2511 1 : self.cancel.cancel();
2512 1 : }
2513 :
2514 113910 : pub(crate) fn current_state(&self) -> TimelineState {
2515 113910 : self.state.borrow().clone()
2516 113910 : }
2517 :
2518 3 : pub(crate) fn is_broken(&self) -> bool {
2519 3 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2520 3 : }
2521 :
2522 126 : pub(crate) fn is_active(&self) -> bool {
2523 126 : self.current_state() == TimelineState::Active
2524 126 : }
2525 :
2526 8 : pub(crate) fn is_archived(&self) -> Option<bool> {
2527 8 : self.remote_client.is_archived()
2528 8 : }
2529 :
2530 8 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2531 8 : self.remote_client.is_invisible()
2532 8 : }
2533 :
2534 568 : pub(crate) fn is_stopping(&self) -> bool {
2535 568 : self.current_state() == TimelineState::Stopping
2536 568 : }
2537 :
2538 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2539 0 : self.state.subscribe()
2540 0 : }
2541 :
2542 112981 : pub(crate) async fn wait_to_become_active(
2543 112981 : &self,
2544 112981 : _ctx: &RequestContext, // Prepare for use by cancellation
2545 112981 : ) -> Result<(), TimelineState> {
2546 112981 : let mut receiver = self.state.subscribe();
2547 : loop {
2548 112981 : let current_state = receiver.borrow().clone();
2549 112981 : match current_state {
2550 : TimelineState::Loading => {
2551 0 : receiver
2552 0 : .changed()
2553 0 : .await
2554 0 : .expect("holding a reference to self");
2555 : }
2556 : TimelineState::Active => {
2557 112980 : return Ok(());
2558 : }
2559 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2560 : // There's no chance the timeline can transition back into ::Active
2561 1 : return Err(current_state);
2562 : }
2563 : }
2564 : }
2565 112981 : }
2566 :
2567 0 : pub(crate) async fn layer_map_info(
2568 0 : &self,
2569 0 : reset: LayerAccessStatsReset,
2570 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2571 0 : let guard = self
2572 0 : .layers
2573 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
2574 0 : .await;
2575 0 : let layer_map = guard.layer_map()?;
2576 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2577 0 : if let Some(open_layer) = &layer_map.open_layer {
2578 0 : in_memory_layers.push(open_layer.info());
2579 0 : }
2580 0 : for frozen_layer in &layer_map.frozen_layers {
2581 0 : in_memory_layers.push(frozen_layer.info());
2582 0 : }
2583 :
2584 0 : let historic_layers = layer_map
2585 0 : .iter_historic_layers()
2586 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2587 0 : .collect();
2588 :
2589 0 : Ok(LayerMapInfo {
2590 0 : in_memory_layers,
2591 0 : historic_layers,
2592 0 : })
2593 0 : }
2594 :
2595 : #[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))]
2596 : pub(crate) async fn download_layer(
2597 : &self,
2598 : layer_file_name: &LayerName,
2599 : ctx: &RequestContext,
2600 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2601 : let Some(layer) = self
2602 : .find_layer(layer_file_name)
2603 : .await
2604 0 : .map_err(|e| match e {
2605 : layer_manager::Shutdown => {
2606 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2607 : }
2608 0 : })?
2609 : else {
2610 : return Ok(None);
2611 : };
2612 :
2613 : layer.download(ctx).await?;
2614 :
2615 : Ok(Some(true))
2616 : }
2617 :
2618 : /// Evict just one layer.
2619 : ///
2620 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2621 0 : pub(crate) async fn evict_layer(
2622 0 : &self,
2623 0 : layer_file_name: &LayerName,
2624 0 : ) -> anyhow::Result<Option<bool>> {
2625 0 : let _gate = self
2626 0 : .gate
2627 0 : .enter()
2628 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2629 :
2630 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2631 0 : return Ok(None);
2632 : };
2633 :
2634 : // curl has this by default
2635 0 : let timeout = std::time::Duration::from_secs(120);
2636 :
2637 0 : match local_layer.evict_and_wait(timeout).await {
2638 0 : Ok(()) => Ok(Some(true)),
2639 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2640 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2641 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2642 : }
2643 0 : }
2644 :
2645 2401506 : fn should_roll(
2646 2401506 : &self,
2647 2401506 : layer_size: u64,
2648 2401506 : projected_layer_size: u64,
2649 2401506 : checkpoint_distance: u64,
2650 2401506 : projected_lsn: Lsn,
2651 2401506 : last_freeze_at: Lsn,
2652 2401506 : opened_at: Instant,
2653 2401506 : ) -> bool {
2654 2401506 : let distance = projected_lsn.widening_sub(last_freeze_at);
2655 :
2656 : // Rolling the open layer can be triggered by:
2657 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2658 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2659 : // account for how writes are distributed across shards: we expect each node to consume
2660 : // 1/count of the LSN on average.
2661 : // 2. The size of the currently open layer.
2662 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2663 : // up and suspend activity.
2664 2401506 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2665 0 : info!(
2666 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2667 : projected_lsn, layer_size, distance
2668 : );
2669 :
2670 0 : true
2671 2401506 : } else if projected_layer_size >= checkpoint_distance {
2672 : // NB: this check is relied upon by:
2673 40 : let _ = IndexEntry::validate_checkpoint_distance;
2674 40 : info!(
2675 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2676 : projected_lsn, layer_size, projected_layer_size
2677 : );
2678 :
2679 40 : true
2680 2401466 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2681 0 : info!(
2682 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2683 : projected_lsn,
2684 : layer_size,
2685 0 : opened_at.elapsed()
2686 : );
2687 :
2688 0 : true
2689 : } else {
2690 2401466 : false
2691 : }
2692 2401506 : }
2693 :
2694 1 : pub(crate) fn is_basebackup_cache_enabled(&self) -> bool {
2695 1 : let tenant_conf = self.tenant_conf.load();
2696 1 : tenant_conf
2697 1 : .tenant_conf
2698 1 : .basebackup_cache_enabled
2699 1 : .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
2700 1 : }
2701 :
2702 : /// Try to get a basebackup from the on-disk cache.
2703 0 : pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
2704 0 : self.basebackup_cache
2705 0 : .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
2706 0 : .await
2707 0 : }
2708 :
2709 : /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
2710 : /// the given request parameters.
2711 : ///
2712 : /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
2713 0 : pub async fn get_cached_basebackup_if_enabled(
2714 0 : &self,
2715 0 : lsn: Option<Lsn>,
2716 0 : prev_lsn: Option<Lsn>,
2717 0 : full: bool,
2718 0 : replica: bool,
2719 0 : gzip: bool,
2720 0 : ) -> Option<tokio::fs::File> {
2721 0 : if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
2722 0 : return None;
2723 0 : }
2724 : // We have to know which LSN to fetch the basebackup for.
2725 0 : let lsn = lsn?;
2726 : // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
2727 0 : if prev_lsn.is_some() || full || replica || !gzip {
2728 0 : return None;
2729 0 : }
2730 0 : self.get_cached_basebackup(lsn).await
2731 0 : }
2732 :
2733 : /// Prepare basebackup for the given LSN and store it in the basebackup cache.
2734 : /// The method is asynchronous and returns immediately.
2735 : /// The actual basebackup preparation is performed in the background
2736 : /// by the basebackup cache on a best-effort basis.
2737 1 : pub(crate) fn prepare_basebackup(&self, lsn: Lsn) {
2738 1 : if !self.is_basebackup_cache_enabled() {
2739 1 : return;
2740 0 : }
2741 0 : if !self.tenant_shard_id.is_shard_zero() {
2742 : // In theory we should never get here, but just in case check it.
2743 : // Preparing basebackup doesn't make sense for shards other than shard zero.
2744 0 : return;
2745 0 : }
2746 0 : if !self.is_active() {
2747 : // May happen during initial timeline creation.
2748 : // Such timeline is not in the global timeline map yet,
2749 : // so basebackup cache will not be able to find it.
2750 : // TODO(diko): We can prepare such timelines in finish_creation().
2751 0 : return;
2752 0 : }
2753 :
2754 0 : self.basebackup_cache
2755 0 : .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
2756 1 : }
2757 : }
2758 :
2759 : /// Number of times we will compute partition within a checkpoint distance.
2760 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2761 :
2762 : // Private functions
2763 : impl Timeline {
2764 6 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2765 6 : let tenant_conf = self.tenant_conf.load();
2766 6 : tenant_conf
2767 6 : .tenant_conf
2768 6 : .lsn_lease_length
2769 6 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2770 6 : }
2771 :
2772 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2773 0 : let tenant_conf = self.tenant_conf.load();
2774 0 : tenant_conf
2775 0 : .tenant_conf
2776 0 : .lsn_lease_length_for_ts
2777 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2778 0 : }
2779 :
2780 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2781 0 : let tenant_conf = self.tenant_conf.load();
2782 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2783 0 : }
2784 :
2785 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2786 0 : let tenant_conf = self.tenant_conf.load();
2787 0 : tenant_conf
2788 0 : .tenant_conf
2789 0 : .lazy_slru_download
2790 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2791 0 : }
2792 :
2793 : /// Checks if a get page request should get perf tracing
2794 : ///
2795 : /// The configuration priority is: tenant config override, default tenant config,
2796 : /// pageserver config.
2797 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2798 0 : let tenant_conf = self.tenant_conf.load();
2799 0 : let ratio = tenant_conf
2800 0 : .tenant_conf
2801 0 : .sampling_ratio
2802 0 : .flatten()
2803 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2804 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2805 :
2806 0 : match ratio {
2807 0 : Some(r) => {
2808 0 : if r.numerator == 0 {
2809 0 : false
2810 : } else {
2811 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2812 : }
2813 : }
2814 0 : None => false,
2815 : }
2816 0 : }
2817 :
2818 2402187 : fn get_checkpoint_distance(&self) -> u64 {
2819 2402187 : let tenant_conf = self.tenant_conf.load();
2820 2402187 : tenant_conf
2821 2402187 : .tenant_conf
2822 2402187 : .checkpoint_distance
2823 2402187 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2824 2402187 : }
2825 :
2826 2401522 : fn get_checkpoint_timeout(&self) -> Duration {
2827 2401522 : let tenant_conf = self.tenant_conf.load();
2828 2401522 : tenant_conf
2829 2401522 : .tenant_conf
2830 2401522 : .checkpoint_timeout
2831 2401522 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2832 2401522 : }
2833 :
2834 1 : pub(crate) fn get_pitr_interval(&self) -> Duration {
2835 1 : let tenant_conf = &self.tenant_conf.load().tenant_conf;
2836 1 : tenant_conf
2837 1 : .pitr_interval
2838 1 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2839 1 : }
2840 :
2841 1273 : fn get_compaction_period(&self) -> Duration {
2842 1273 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2843 1273 : tenant_conf
2844 1273 : .compaction_period
2845 1273 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2846 1273 : }
2847 :
2848 352 : fn get_compaction_target_size(&self) -> u64 {
2849 352 : let tenant_conf = self.tenant_conf.load();
2850 352 : tenant_conf
2851 352 : .tenant_conf
2852 352 : .compaction_target_size
2853 352 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2854 352 : }
2855 :
2856 820 : fn get_compaction_threshold(&self) -> usize {
2857 820 : let tenant_conf = self.tenant_conf.load();
2858 820 : tenant_conf
2859 820 : .tenant_conf
2860 820 : .compaction_threshold
2861 820 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2862 820 : }
2863 :
2864 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2865 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2866 : /// possible that the index part persists the state while the config doesn't get persisted.
2867 973 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2868 973 : let tenant_conf = self.tenant_conf.load();
2869 973 : tenant_conf
2870 973 : .tenant_conf
2871 973 : .rel_size_v2_enabled
2872 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2873 973 : }
2874 :
2875 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2876 1099 : self.rel_size_v2_status
2877 1099 : .load()
2878 1099 : .as_ref()
2879 1099 : .map(|s| s.as_ref().clone())
2880 1099 : .unwrap_or(RelSizeMigration::Legacy)
2881 1099 : }
2882 :
2883 23 : fn get_compaction_upper_limit(&self) -> usize {
2884 23 : let tenant_conf = self.tenant_conf.load();
2885 23 : tenant_conf
2886 23 : .tenant_conf
2887 23 : .compaction_upper_limit
2888 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2889 23 : }
2890 :
2891 0 : pub fn get_compaction_l0_first(&self) -> bool {
2892 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2893 0 : tenant_conf
2894 0 : .compaction_l0_first
2895 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2896 0 : }
2897 :
2898 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2899 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2900 0 : tenant_conf
2901 0 : .compaction_l0_semaphore
2902 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2903 0 : }
2904 :
2905 636 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2906 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2907 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2908 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2909 :
2910 : // If compaction is disabled, don't delay.
2911 636 : if self.get_compaction_period() == Duration::ZERO {
2912 632 : return None;
2913 4 : }
2914 :
2915 4 : let compaction_threshold = self.get_compaction_threshold();
2916 4 : let tenant_conf = self.tenant_conf.load();
2917 4 : let l0_flush_delay_threshold = tenant_conf
2918 4 : .tenant_conf
2919 4 : .l0_flush_delay_threshold
2920 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2921 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2922 :
2923 : // 0 disables backpressure.
2924 4 : if l0_flush_delay_threshold == 0 {
2925 0 : return None;
2926 4 : }
2927 :
2928 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2929 : // backpressure flushes below this.
2930 : // TODO: the tenant config should have validation to prevent this instead.
2931 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2932 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2933 636 : }
2934 :
2935 637 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2936 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2937 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2938 : // sibling timelines. We need more confidence before enabling this.
2939 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2940 :
2941 : // If compaction is disabled, don't stall.
2942 637 : if self.get_compaction_period() == Duration::ZERO {
2943 632 : return None;
2944 5 : }
2945 :
2946 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2947 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2948 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2949 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2950 5 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2951 0 : return None;
2952 5 : }
2953 :
2954 5 : let compaction_threshold = self.get_compaction_threshold();
2955 5 : let tenant_conf = self.tenant_conf.load();
2956 5 : let l0_flush_stall_threshold = tenant_conf
2957 5 : .tenant_conf
2958 5 : .l0_flush_stall_threshold
2959 5 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2960 :
2961 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2962 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2963 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2964 5 : if cfg!(feature = "testing")
2965 5 : && compaction_threshold == 1
2966 0 : && l0_flush_stall_threshold.is_none()
2967 : {
2968 0 : return None;
2969 5 : }
2970 :
2971 5 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2972 5 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2973 :
2974 : // 0 disables backpressure.
2975 5 : if l0_flush_stall_threshold == 0 {
2976 5 : return None;
2977 0 : }
2978 :
2979 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2980 : // backpressure flushes below this.
2981 : // TODO: the tenant config should have validation to prevent this instead.
2982 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2983 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2984 637 : }
2985 :
2986 57 : fn get_image_creation_threshold(&self) -> usize {
2987 57 : let tenant_conf = self.tenant_conf.load();
2988 57 : tenant_conf
2989 57 : .tenant_conf
2990 57 : .image_creation_threshold
2991 57 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2992 57 : }
2993 :
2994 : // HADRON
2995 193 : fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
2996 193 : let tenant_conf = self.tenant_conf.load();
2997 193 : tenant_conf
2998 193 : .tenant_conf
2999 193 : .image_layer_force_creation_period
3000 193 : .or(self
3001 193 : .conf
3002 193 : .default_tenant_conf
3003 193 : .image_layer_force_creation_period)
3004 193 : }
3005 :
3006 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
3007 192 : let tenant_conf = &self.tenant_conf.load();
3008 192 : tenant_conf
3009 192 : .tenant_conf
3010 192 : .compaction_algorithm
3011 192 : .as_ref()
3012 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
3013 192 : .clone()
3014 192 : }
3015 :
3016 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
3017 192 : let tenant_conf = self.tenant_conf.load();
3018 192 : tenant_conf
3019 192 : .tenant_conf
3020 192 : .compaction_shard_ancestor
3021 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
3022 192 : }
3023 :
3024 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
3025 0 : let tenant_conf = self.tenant_conf.load();
3026 0 : tenant_conf
3027 0 : .tenant_conf
3028 0 : .eviction_policy
3029 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
3030 0 : }
3031 :
3032 235 : fn get_evictions_low_residence_duration_metric_threshold(
3033 235 : tenant_conf: &pageserver_api::models::TenantConfig,
3034 235 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
3035 235 : ) -> Duration {
3036 235 : tenant_conf
3037 235 : .evictions_low_residence_duration_metric_threshold
3038 235 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
3039 235 : }
3040 :
3041 191 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
3042 191 : let tenant_conf = self.tenant_conf.load();
3043 191 : tenant_conf
3044 191 : .tenant_conf
3045 191 : .image_layer_creation_check_threshold
3046 191 : .unwrap_or(
3047 191 : self.conf
3048 191 : .default_tenant_conf
3049 191 : .image_layer_creation_check_threshold,
3050 : )
3051 191 : }
3052 :
3053 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
3054 27 : let tenant_conf = &self.tenant_conf.load();
3055 27 : let gc_compaction_enabled = tenant_conf
3056 27 : .tenant_conf
3057 27 : .gc_compaction_enabled
3058 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
3059 27 : let gc_compaction_verification = tenant_conf
3060 27 : .tenant_conf
3061 27 : .gc_compaction_verification
3062 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
3063 27 : let gc_compaction_initial_threshold_kb = tenant_conf
3064 27 : .tenant_conf
3065 27 : .gc_compaction_initial_threshold_kb
3066 27 : .unwrap_or(
3067 27 : self.conf
3068 27 : .default_tenant_conf
3069 27 : .gc_compaction_initial_threshold_kb,
3070 : );
3071 27 : let gc_compaction_ratio_percent = tenant_conf
3072 27 : .tenant_conf
3073 27 : .gc_compaction_ratio_percent
3074 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
3075 27 : GcCompactionCombinedSettings {
3076 27 : gc_compaction_enabled,
3077 27 : gc_compaction_verification,
3078 27 : gc_compaction_initial_threshold_kb,
3079 27 : gc_compaction_ratio_percent,
3080 27 : }
3081 27 : }
3082 :
3083 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
3084 0 : let tenant_conf = self.tenant_conf.load();
3085 0 : tenant_conf
3086 0 : .tenant_conf
3087 0 : .image_creation_preempt_threshold
3088 0 : .unwrap_or(
3089 0 : self.conf
3090 0 : .default_tenant_conf
3091 0 : .image_creation_preempt_threshold,
3092 : )
3093 0 : }
3094 :
3095 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
3096 : // NB: Most tenant conf options are read by background loops, so,
3097 : // changes will automatically be picked up.
3098 :
3099 : // The threshold is embedded in the metric. So, we need to update it.
3100 : {
3101 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
3102 0 : &new_conf.tenant_conf,
3103 0 : &self.conf.default_tenant_conf,
3104 : );
3105 :
3106 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
3107 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
3108 :
3109 0 : let timeline_id_str = self.timeline_id.to_string();
3110 :
3111 0 : self.remote_client.update_config(&new_conf.location);
3112 :
3113 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
3114 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
3115 0 : if new_capacity != rel_size_cache.capacity() {
3116 0 : rel_size_cache.set_capacity(new_capacity);
3117 0 : }
3118 0 : }
3119 :
3120 0 : self.metrics
3121 0 : .evictions_with_low_residence_duration
3122 0 : .write()
3123 0 : .unwrap()
3124 0 : .change_threshold(
3125 0 : &tenant_id_str,
3126 0 : &shard_id_str,
3127 0 : &timeline_id_str,
3128 0 : new_threshold,
3129 0 : );
3130 : }
3131 0 : }
3132 :
3133 : /// Open a Timeline handle.
3134 : ///
3135 : /// Loads the metadata for the timeline into memory, but not the layer map.
3136 : #[allow(clippy::too_many_arguments)]
3137 235 : pub(super) fn new(
3138 235 : conf: &'static PageServerConf,
3139 235 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
3140 235 : metadata: &TimelineMetadata,
3141 235 : previous_heatmap: Option<PreviousHeatmap>,
3142 235 : ancestor: Option<Arc<Timeline>>,
3143 235 : timeline_id: TimelineId,
3144 235 : tenant_shard_id: TenantShardId,
3145 235 : generation: Generation,
3146 235 : shard_identity: ShardIdentity,
3147 235 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
3148 235 : resources: TimelineResources,
3149 235 : pg_version: PgMajorVersion,
3150 235 : state: TimelineState,
3151 235 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3152 235 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3153 235 : gc_compaction_state: Option<GcCompactionState>,
3154 235 : rel_size_v2_status: Option<RelSizeMigration>,
3155 235 : cancel: CancellationToken,
3156 235 : ) -> Arc<Self> {
3157 235 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3158 235 : let (state, _) = watch::channel(state);
3159 :
3160 235 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3161 235 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3162 :
3163 235 : let evictions_low_residence_duration_metric_threshold = {
3164 235 : let loaded_tenant_conf = tenant_conf.load();
3165 235 : Self::get_evictions_low_residence_duration_metric_threshold(
3166 235 : &loaded_tenant_conf.tenant_conf,
3167 235 : &conf.default_tenant_conf,
3168 : )
3169 : };
3170 :
3171 235 : if let Some(ancestor) = &ancestor {
3172 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3173 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3174 118 : let is_offloaded = MaybeOffloaded::No;
3175 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3176 118 : }
3177 :
3178 235 : let relsize_snapshot_cache_capacity = {
3179 235 : let loaded_tenant_conf = tenant_conf.load();
3180 235 : loaded_tenant_conf
3181 235 : .tenant_conf
3182 235 : .relsize_snapshot_cache_capacity
3183 235 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3184 : };
3185 :
3186 235 : Arc::new_cyclic(|myself| {
3187 235 : let metrics = Arc::new(TimelineMetrics::new(
3188 235 : &tenant_shard_id,
3189 235 : &timeline_id,
3190 235 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3191 : "mtime",
3192 235 : evictions_low_residence_duration_metric_threshold,
3193 : ),
3194 : ));
3195 235 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3196 :
3197 235 : let mut result = Timeline {
3198 235 : conf,
3199 235 : tenant_conf,
3200 235 : myself: myself.clone(),
3201 235 : timeline_id,
3202 235 : tenant_shard_id,
3203 235 : generation,
3204 235 : shard_identity,
3205 235 : pg_version,
3206 235 : layers: Default::default(),
3207 235 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3208 :
3209 235 : walredo_mgr,
3210 235 : walreceiver: Mutex::new(None),
3211 :
3212 235 : remote_client: Arc::new(resources.remote_client),
3213 :
3214 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3215 235 : last_record_lsn: SeqWait::new(RecordLsn {
3216 235 : last: disk_consistent_lsn,
3217 235 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3218 235 : }),
3219 235 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3220 :
3221 235 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
3222 :
3223 235 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3224 235 : last_freeze_ts: RwLock::new(Instant::now()),
3225 :
3226 235 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3227 :
3228 235 : ancestor_timeline: ancestor,
3229 235 : ancestor_lsn: metadata.ancestor_lsn(),
3230 :
3231 235 : metrics,
3232 :
3233 235 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3234 235 : &tenant_shard_id,
3235 235 : &timeline_id,
3236 235 : resources.pagestream_throttle_metrics,
3237 : ),
3238 :
3239 1880 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3240 1880 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3241 :
3242 235 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3243 :
3244 235 : layer_flush_start_tx,
3245 235 : layer_flush_done_tx,
3246 :
3247 235 : write_lock: tokio::sync::Mutex::new(None),
3248 :
3249 235 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3250 :
3251 235 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3252 235 : LastImageLayerCreationStatus::default(),
3253 : )),
3254 :
3255 235 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3256 235 : initdb_lsn: metadata.initdb_lsn(),
3257 :
3258 235 : current_logical_size: if disk_consistent_lsn.is_valid() {
3259 : // we're creating timeline data with some layer files existing locally,
3260 : // need to recalculate timeline's logical size based on data in the layers.
3261 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3262 : } else {
3263 : // we're creating timeline data without any layers existing locally,
3264 : // initial logical size is 0.
3265 115 : LogicalSize::empty_initial()
3266 : },
3267 :
3268 235 : partitioning: GuardArcSwap::new((
3269 235 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3270 235 : Lsn(0),
3271 235 : )),
3272 : repartition_threshold: 0,
3273 235 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3274 235 : last_image_layer_creation_check_instant: Mutex::new(None),
3275 235 : last_received_wal: Mutex::new(None),
3276 235 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3277 235 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3278 :
3279 235 : download_all_remote_layers_task_info: RwLock::new(None),
3280 :
3281 235 : state,
3282 :
3283 235 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3284 235 : EvictionTaskTimelineState::default(),
3285 : ),
3286 235 : delete_progress: TimelineDeleteProgress::default(),
3287 :
3288 235 : cancel,
3289 235 : gate: Gate::default(),
3290 :
3291 235 : compaction_lock: tokio::sync::Mutex::default(),
3292 235 : compaction_failed: AtomicBool::default(),
3293 235 : l0_compaction_trigger: resources.l0_compaction_trigger,
3294 235 : gc_lock: tokio::sync::Mutex::default(),
3295 :
3296 235 : standby_horizon: AtomicLsn::new(0),
3297 :
3298 235 : pagestream_throttle: resources.pagestream_throttle,
3299 :
3300 235 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3301 :
3302 : #[cfg(test)]
3303 235 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3304 :
3305 235 : l0_flush_global_state: resources.l0_flush_global_state,
3306 :
3307 235 : handles: Default::default(),
3308 :
3309 235 : attach_wal_lag_cooldown,
3310 :
3311 235 : create_idempotency,
3312 :
3313 235 : page_trace: Default::default(),
3314 :
3315 235 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3316 :
3317 235 : heatmap_layers_downloader: Mutex::new(None),
3318 :
3319 235 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
3320 :
3321 235 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3322 :
3323 235 : basebackup_cache: resources.basebackup_cache,
3324 :
3325 235 : feature_resolver: resources.feature_resolver.clone(),
3326 : };
3327 :
3328 235 : result.repartition_threshold =
3329 235 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3330 :
3331 235 : result
3332 235 : .metrics
3333 235 : .last_record_lsn_gauge
3334 235 : .set(disk_consistent_lsn.0 as i64);
3335 235 : result
3336 235 : })
3337 235 : }
3338 :
3339 343 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3340 343 : let Ok(guard) = self.gate.enter() else {
3341 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3342 0 : return;
3343 : };
3344 343 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3345 343 : match *flush_loop_state {
3346 232 : FlushLoopState::NotStarted => (),
3347 : FlushLoopState::Running { .. } => {
3348 111 : info!(
3349 0 : "skipping attempt to start flush_loop twice {}/{}",
3350 0 : self.tenant_shard_id, self.timeline_id
3351 : );
3352 111 : return;
3353 : }
3354 : FlushLoopState::Exited => {
3355 0 : info!(
3356 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3357 0 : self.tenant_shard_id, self.timeline_id
3358 : );
3359 0 : return;
3360 : }
3361 : }
3362 :
3363 232 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3364 232 : let self_clone = Arc::clone(self);
3365 :
3366 232 : debug!("spawning flush loop");
3367 232 : *flush_loop_state = FlushLoopState::Running {
3368 232 : #[cfg(test)]
3369 232 : expect_initdb_optimization: false,
3370 232 : #[cfg(test)]
3371 232 : initdb_optimization_count: 0,
3372 232 : };
3373 232 : task_mgr::spawn(
3374 232 : task_mgr::BACKGROUND_RUNTIME.handle(),
3375 232 : task_mgr::TaskKind::LayerFlushTask,
3376 232 : self.tenant_shard_id,
3377 232 : Some(self.timeline_id),
3378 232 : "layer flush task",
3379 232 : async move {
3380 232 : let _guard = guard;
3381 232 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3382 232 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3383 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3384 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3385 5 : *flush_loop_state = FlushLoopState::Exited;
3386 5 : Ok(())
3387 5 : }
3388 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))
3389 : );
3390 343 : }
3391 :
3392 0 : pub(crate) fn update_gc_compaction_state(
3393 0 : &self,
3394 0 : gc_compaction_state: GcCompactionState,
3395 0 : ) -> anyhow::Result<()> {
3396 0 : self.gc_compaction_state
3397 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3398 0 : self.remote_client
3399 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3400 0 : }
3401 :
3402 0 : pub(crate) fn update_rel_size_v2_status(
3403 0 : &self,
3404 0 : rel_size_v2_status: RelSizeMigration,
3405 0 : ) -> anyhow::Result<()> {
3406 0 : self.rel_size_v2_status
3407 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3408 0 : self.remote_client
3409 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3410 0 : }
3411 :
3412 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3413 0 : self.gc_compaction_state.load_full().as_ref().clone()
3414 0 : }
3415 :
3416 : /// Creates and starts the wal receiver.
3417 : ///
3418 : /// This function is expected to be called at most once per Timeline's lifecycle
3419 : /// when the timeline is activated.
3420 0 : fn launch_wal_receiver(
3421 0 : self: &Arc<Self>,
3422 0 : ctx: &RequestContext,
3423 0 : broker_client: BrokerClientChannel,
3424 0 : ) {
3425 0 : info!(
3426 0 : "launching WAL receiver for timeline {} of tenant {}",
3427 0 : self.timeline_id, self.tenant_shard_id
3428 : );
3429 :
3430 0 : let tenant_conf = self.tenant_conf.load();
3431 0 : let wal_connect_timeout = tenant_conf
3432 0 : .tenant_conf
3433 0 : .walreceiver_connect_timeout
3434 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3435 0 : let lagging_wal_timeout = tenant_conf
3436 0 : .tenant_conf
3437 0 : .lagging_wal_timeout
3438 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3439 0 : let max_lsn_wal_lag = tenant_conf
3440 0 : .tenant_conf
3441 0 : .max_lsn_wal_lag
3442 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3443 :
3444 0 : let mut guard = self.walreceiver.lock().unwrap();
3445 0 : assert!(
3446 0 : guard.is_none(),
3447 0 : "multiple launches / re-launches of WAL receiver are not supported"
3448 : );
3449 :
3450 0 : let protocol = PostgresClientProtocol::Interpreted {
3451 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3452 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3453 0 : };
3454 :
3455 0 : *guard = Some(WalReceiver::start(
3456 0 : Arc::clone(self),
3457 0 : WalReceiverConf {
3458 0 : protocol,
3459 0 : wal_connect_timeout,
3460 0 : lagging_wal_timeout,
3461 0 : max_lsn_wal_lag,
3462 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3463 0 : availability_zone: self.conf.availability_zone.clone(),
3464 0 : ingest_batch_size: self.conf.ingest_batch_size,
3465 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3466 0 : },
3467 0 : broker_client,
3468 0 : ctx,
3469 0 : ));
3470 0 : }
3471 :
3472 : /// Initialize with an empty layer map. Used when creating a new timeline.
3473 232 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3474 232 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3475 232 : "in the context where we call this function, no other task has access to the object",
3476 : );
3477 232 : layers
3478 232 : .open_mut()
3479 232 : .expect("in this context the LayerManager must still be open")
3480 232 : .initialize_empty(Lsn(start_lsn.0));
3481 232 : }
3482 :
3483 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3484 : /// files.
3485 3 : pub(super) async fn load_layer_map(
3486 3 : &self,
3487 3 : disk_consistent_lsn: Lsn,
3488 3 : index_part: IndexPart,
3489 3 : ) -> anyhow::Result<()> {
3490 : use LayerName::*;
3491 : use init::Decision::*;
3492 : use init::{Discovered, DismissedLayer};
3493 :
3494 3 : let mut guard = self
3495 3 : .layers
3496 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3497 3 : .await;
3498 :
3499 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3500 :
3501 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3502 : // structs representing all files on disk
3503 3 : let timeline_path = self
3504 3 : .conf
3505 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3506 3 : let conf = self.conf;
3507 3 : let span = tracing::Span::current();
3508 :
3509 : // Copy to move into the task we're about to spawn
3510 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3511 :
3512 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3513 3 : move || {
3514 3 : let _g = span.entered();
3515 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3516 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3517 3 : let mut unrecognized_files = Vec::new();
3518 :
3519 3 : let mut path = timeline_path;
3520 :
3521 11 : for discovered in discovered {
3522 8 : let (name, kind) = match discovered {
3523 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3524 8 : discovered_layers.push((layer_file_name, local_metadata));
3525 8 : continue;
3526 : }
3527 0 : Discovered::IgnoredBackup(path) => {
3528 0 : std::fs::remove_file(path)
3529 0 : .or_else(fs_ext::ignore_not_found)
3530 0 : .fatal_err("Removing .old file");
3531 0 : continue;
3532 : }
3533 0 : Discovered::Unknown(file_name) => {
3534 : // we will later error if there are any
3535 0 : unrecognized_files.push(file_name);
3536 0 : continue;
3537 : }
3538 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3539 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3540 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3541 : };
3542 0 : path.push(Utf8Path::new(&name));
3543 0 : init::cleanup(&path, kind)?;
3544 0 : path.pop();
3545 : }
3546 :
3547 3 : if !unrecognized_files.is_empty() {
3548 : // assume that if there are any there are many many.
3549 0 : let n = unrecognized_files.len();
3550 0 : let first = &unrecognized_files[..n.min(10)];
3551 0 : anyhow::bail!(
3552 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3553 : );
3554 3 : }
3555 :
3556 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3557 :
3558 3 : let mut loaded_layers = Vec::new();
3559 3 : let mut needs_cleanup = Vec::new();
3560 3 : let mut total_physical_size = 0;
3561 :
3562 11 : for (name, decision) in decided {
3563 8 : let decision = match decision {
3564 8 : Ok(decision) => decision,
3565 0 : Err(DismissedLayer::Future { local }) => {
3566 0 : if let Some(local) = local {
3567 0 : init::cleanup_future_layer(
3568 0 : &local.local_path,
3569 0 : &name,
3570 0 : disk_consistent_lsn,
3571 0 : )?;
3572 0 : }
3573 0 : needs_cleanup.push(name);
3574 0 : continue;
3575 : }
3576 0 : Err(DismissedLayer::LocalOnly(local)) => {
3577 0 : init::cleanup_local_only_file(&name, &local)?;
3578 : // this file never existed remotely, we will have to do rework
3579 0 : continue;
3580 : }
3581 0 : Err(DismissedLayer::BadMetadata(local)) => {
3582 0 : init::cleanup_local_file_for_remote(&local)?;
3583 : // this file never existed remotely, we will have to do rework
3584 0 : continue;
3585 : }
3586 : };
3587 :
3588 8 : match &name {
3589 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3590 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3591 : }
3592 :
3593 8 : tracing::debug!(layer=%name, ?decision, "applied");
3594 :
3595 8 : let layer = match decision {
3596 8 : Resident { local, remote } => {
3597 8 : total_physical_size += local.file_size;
3598 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3599 8 : .drop_eviction_guard()
3600 : }
3601 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3602 : };
3603 :
3604 8 : loaded_layers.push(layer);
3605 : }
3606 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3607 3 : }
3608 : })
3609 3 : .await
3610 3 : .map_err(anyhow::Error::new)
3611 3 : .and_then(|x| x)?;
3612 :
3613 3 : let num_layers = loaded_layers.len();
3614 :
3615 3 : guard
3616 3 : .open_mut()
3617 3 : .expect("layermanager must be open during init")
3618 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3619 :
3620 3 : self.remote_client
3621 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3622 3 : self.remote_client
3623 3 : .schedule_index_upload_for_file_changes()?;
3624 : // This barrier orders above DELETEs before any later operations.
3625 : // This is critical because code executing after the barrier might
3626 : // create again objects with the same key that we just scheduled for deletion.
3627 : // For example, if we just scheduled deletion of an image layer "from the future",
3628 : // later compaction might run again and re-create the same image layer.
3629 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3630 : // "same" here means same key range and LSN.
3631 : //
3632 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3633 : // the upload queue may execute the PUT first, then the DELETE.
3634 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3635 : //
3636 : // 1. a future image layer is created and uploaded
3637 : // 2. ps restart
3638 : // 3. the future layer from (1) is deleted during load layer map
3639 : // 4. image layer is re-created and uploaded
3640 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3641 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3642 : //
3643 : // See https://github.com/neondatabase/neon/issues/5878
3644 : //
3645 : // NB: generation numbers naturally protect against this because they disambiguate
3646 : // (1) and (4)
3647 : // TODO: this is basically a no-op now, should we remove it?
3648 3 : self.remote_client.schedule_barrier()?;
3649 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3650 : // on retry.
3651 :
3652 3 : info!(
3653 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3654 : num_layers, disk_consistent_lsn, total_physical_size
3655 : );
3656 :
3657 3 : timer.stop_and_record();
3658 3 : Ok(())
3659 3 : }
3660 :
3661 : /// Retrieve current logical size of the timeline.
3662 : ///
3663 : /// The size could be lagging behind the actual number, in case
3664 : /// the initial size calculation has not been run (gets triggered on the first size access).
3665 : ///
3666 : /// return size and boolean flag that shows if the size is exact
3667 0 : pub(crate) fn get_current_logical_size(
3668 0 : self: &Arc<Self>,
3669 0 : priority: GetLogicalSizePriority,
3670 0 : ctx: &RequestContext,
3671 0 : ) -> logical_size::CurrentLogicalSize {
3672 0 : if !self.tenant_shard_id.is_shard_zero() {
3673 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3674 : // when HTTP API is serving a GET for timeline zero, return zero
3675 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3676 0 : }
3677 :
3678 0 : let current_size = self.current_logical_size.current_size();
3679 0 : debug!("Current size: {current_size:?}");
3680 :
3681 0 : match (current_size.accuracy(), priority) {
3682 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3683 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3684 0 : // background task will eventually deliver an exact value, we're in no rush
3685 0 : }
3686 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3687 : // background task is not ready, but user is asking for it now;
3688 : // => make the background task skip the line
3689 : // (The alternative would be to calculate the size here, but,
3690 : // it can actually take a long time if the user has a lot of rels.
3691 : // And we'll inevitable need it again; So, let the background task do the work.)
3692 0 : match self
3693 0 : .current_logical_size
3694 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3695 0 : .get()
3696 : {
3697 0 : Some(cancel) => cancel.cancel(),
3698 : None => {
3699 0 : match self.current_state() {
3700 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3701 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3702 0 : // Don't make noise.
3703 0 : }
3704 : TimelineState::Loading => {
3705 : // Import does not return an activated timeline.
3706 0 : info!(
3707 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3708 : );
3709 : }
3710 : TimelineState::Active => {
3711 : // activation should be setting the once cell
3712 0 : warn!(
3713 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3714 : );
3715 0 : debug_assert!(false);
3716 : }
3717 : }
3718 : }
3719 : }
3720 : }
3721 : }
3722 :
3723 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3724 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3725 0 : let first = self
3726 0 : .current_logical_size
3727 0 : .did_return_approximate_to_walreceiver
3728 0 : .compare_exchange(
3729 0 : false,
3730 0 : true,
3731 0 : AtomicOrdering::Relaxed,
3732 0 : AtomicOrdering::Relaxed,
3733 0 : )
3734 0 : .is_ok();
3735 0 : if first {
3736 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3737 0 : }
3738 0 : }
3739 0 : }
3740 :
3741 0 : current_size
3742 0 : }
3743 :
3744 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3745 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3746 : // nothing to do for freshly created timelines;
3747 0 : assert_eq!(
3748 0 : self.current_logical_size.current_size().accuracy(),
3749 : logical_size::Accuracy::Exact,
3750 : );
3751 0 : self.current_logical_size.initialized.add_permits(1);
3752 0 : return;
3753 : };
3754 :
3755 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3756 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3757 0 : self.current_logical_size
3758 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3759 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3760 :
3761 0 : let self_clone = Arc::clone(self);
3762 0 : let background_ctx = ctx.detached_child(
3763 0 : TaskKind::InitialLogicalSizeCalculation,
3764 0 : DownloadBehavior::Download,
3765 : );
3766 0 : task_mgr::spawn(
3767 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3768 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3769 0 : self.tenant_shard_id,
3770 0 : Some(self.timeline_id),
3771 0 : "initial size calculation",
3772 : // NB: don't log errors here, task_mgr will do that.
3773 0 : async move {
3774 0 : self_clone
3775 0 : .initial_logical_size_calculation_task(
3776 0 : initial_part_end,
3777 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3778 0 : background_ctx,
3779 0 : )
3780 0 : .await;
3781 0 : Ok(())
3782 0 : }
3783 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)),
3784 : );
3785 0 : }
3786 :
3787 : /// # Cancellation
3788 : ///
3789 : /// This method is sensitive to `Timeline::cancel`.
3790 : ///
3791 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3792 : ///
3793 : /// # Cancel-Safety
3794 : ///
3795 : /// It does Timeline IO, hence this should be polled to completion because
3796 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3797 : /// to reason about.
3798 0 : async fn initial_logical_size_calculation_task(
3799 0 : self: Arc<Self>,
3800 0 : initial_part_end: Lsn,
3801 0 : skip_concurrency_limiter: CancellationToken,
3802 0 : background_ctx: RequestContext,
3803 0 : ) {
3804 0 : scopeguard::defer! {
3805 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3806 : self.current_logical_size.initialized.add_permits(1);
3807 : }
3808 :
3809 0 : let try_once = |attempt: usize| {
3810 0 : let background_ctx = &background_ctx;
3811 0 : let self_ref = &self;
3812 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3813 0 : async move {
3814 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3815 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3816 0 : background_ctx,
3817 : );
3818 :
3819 : use crate::metrics::initial_logical_size::StartCircumstances;
3820 0 : let (_maybe_permit, circumstances) = tokio::select! {
3821 0 : permit = wait_for_permit => {
3822 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3823 : }
3824 0 : _ = self_ref.cancel.cancelled() => {
3825 0 : return Err(CalculateLogicalSizeError::Cancelled);
3826 : }
3827 0 : () = skip_concurrency_limiter.cancelled() => {
3828 : // Some action that is part of a end user interaction requested logical size
3829 : // => break out of the rate limit
3830 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3831 : // but then again what happens if they cancel; also, we should just be using
3832 : // one runtime across the entire process, so, let's leave this for now.
3833 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3834 : }
3835 : };
3836 :
3837 0 : let metrics_guard = if attempt == 1 {
3838 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3839 : } else {
3840 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3841 : };
3842 :
3843 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3844 0 : self_ref.conf.get_vectored_concurrent_io,
3845 0 : self_ref
3846 0 : .gate
3847 0 : .enter()
3848 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3849 : );
3850 :
3851 0 : let calculated_size = self_ref
3852 0 : .logical_size_calculation_task(
3853 0 : initial_part_end,
3854 0 : LogicalSizeCalculationCause::Initial,
3855 0 : background_ctx,
3856 0 : )
3857 0 : .await?;
3858 :
3859 0 : self_ref
3860 0 : .trigger_aux_file_size_computation(
3861 0 : initial_part_end,
3862 0 : background_ctx,
3863 0 : io_concurrency,
3864 0 : )
3865 0 : .await?;
3866 :
3867 : // TODO: add aux file size to logical size
3868 :
3869 0 : Ok((calculated_size, metrics_guard))
3870 0 : }
3871 0 : };
3872 :
3873 0 : let retrying = async {
3874 0 : let mut attempt = 0;
3875 : loop {
3876 0 : attempt += 1;
3877 :
3878 0 : match try_once(attempt).await {
3879 0 : Ok(res) => return ControlFlow::Continue(res),
3880 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3881 : Err(
3882 0 : e @ (CalculateLogicalSizeError::Decode(_)
3883 : | CalculateLogicalSizeError::PageRead(_)),
3884 : ) => {
3885 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3886 : // exponential back-off doesn't make sense at these long intervals;
3887 : // use fixed retry interval with generous jitter instead
3888 0 : let sleep_duration = Duration::from_secs(
3889 0 : u64::try_from(
3890 : // 1hour base
3891 0 : (60_i64 * 60_i64)
3892 0 : // 10min jitter
3893 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3894 : )
3895 0 : .expect("10min < 1hour"),
3896 : );
3897 0 : tokio::select! {
3898 0 : _ = tokio::time::sleep(sleep_duration) => {}
3899 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3900 : }
3901 : }
3902 : }
3903 : }
3904 0 : };
3905 :
3906 0 : let (calculated_size, metrics_guard) = match retrying.await {
3907 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3908 0 : ControlFlow::Break(()) => return,
3909 : };
3910 :
3911 : // we cannot query current_logical_size.current_size() to know the current
3912 : // *negative* value, only truncated to u64.
3913 0 : let added = self
3914 0 : .current_logical_size
3915 0 : .size_added_after_initial
3916 0 : .load(AtomicOrdering::Relaxed);
3917 :
3918 0 : let sum = calculated_size.saturating_add_signed(added);
3919 :
3920 : // set the gauge value before it can be set in `update_current_logical_size`.
3921 0 : self.metrics.current_logical_size_gauge.set(sum);
3922 :
3923 0 : self.current_logical_size
3924 0 : .initial_logical_size
3925 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3926 0 : .ok()
3927 0 : .expect("only this task sets it");
3928 0 : }
3929 :
3930 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3931 7 : self: &Arc<Self>,
3932 7 : lsn: Lsn,
3933 7 : cause: LogicalSizeCalculationCause,
3934 7 : ctx: RequestContext,
3935 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3936 7 : let (sender, receiver) = oneshot::channel();
3937 7 : let self_clone = Arc::clone(self);
3938 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3939 : // we should stop the size calculation work and return an error.
3940 : // That would require restructuring this function's API to
3941 : // return the result directly, instead of a Receiver for the result.
3942 7 : let ctx = ctx.detached_child(
3943 7 : TaskKind::OndemandLogicalSizeCalculation,
3944 7 : DownloadBehavior::Download,
3945 : );
3946 7 : task_mgr::spawn(
3947 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3948 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3949 7 : self.tenant_shard_id,
3950 7 : Some(self.timeline_id),
3951 7 : "ondemand logical size calculation",
3952 7 : async move {
3953 7 : let res = self_clone
3954 7 : .logical_size_calculation_task(lsn, cause, &ctx)
3955 7 : .await;
3956 7 : let _ = sender.send(res).ok();
3957 7 : Ok(()) // Receiver is responsible for handling errors
3958 7 : }
3959 7 : .in_current_span(),
3960 : );
3961 7 : receiver
3962 7 : }
3963 :
3964 : #[instrument(skip_all)]
3965 : async fn logical_size_calculation_task(
3966 : self: &Arc<Self>,
3967 : lsn: Lsn,
3968 : cause: LogicalSizeCalculationCause,
3969 : ctx: &RequestContext,
3970 : ) -> Result<u64, CalculateLogicalSizeError> {
3971 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3972 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3973 : // accurate relation sizes, and they do not emit consumption metrics.
3974 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3975 :
3976 : let guard = self
3977 : .gate
3978 : .enter()
3979 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3980 :
3981 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3982 : }
3983 :
3984 : /// Calculate the logical size of the database at the latest LSN.
3985 : ///
3986 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3987 : /// especially if we need to download remote layers.
3988 7 : async fn calculate_logical_size(
3989 7 : &self,
3990 7 : up_to_lsn: Lsn,
3991 7 : cause: LogicalSizeCalculationCause,
3992 7 : _guard: &GateGuard,
3993 7 : ctx: &RequestContext,
3994 7 : ) -> Result<u64, CalculateLogicalSizeError> {
3995 7 : info!(
3996 0 : "Calculating logical size for timeline {} at {}",
3997 : self.timeline_id, up_to_lsn
3998 : );
3999 :
4000 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
4001 : {
4002 0 : return Err(CalculateLogicalSizeError::Cancelled);
4003 7 : }
4004 :
4005 : // See if we've already done the work for initial size calculation.
4006 : // This is a short-cut for timelines that are mostly unused.
4007 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
4008 0 : return Ok(size);
4009 7 : }
4010 7 : let storage_time_metrics = match cause {
4011 : LogicalSizeCalculationCause::Initial
4012 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
4013 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
4014 : LogicalSizeCalculationCause::EvictionTaskImitation => {
4015 7 : &self.metrics.imitate_logical_size_histo
4016 : }
4017 : };
4018 7 : let timer = storage_time_metrics.start_timer();
4019 7 : let logical_size = self
4020 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
4021 7 : .await?;
4022 7 : debug!("calculated logical size: {logical_size}");
4023 7 : timer.stop_and_record();
4024 7 : Ok(logical_size)
4025 7 : }
4026 :
4027 : /// Update current logical size, adding `delta' to the old value.
4028 135285 : fn update_current_logical_size(&self, delta: i64) {
4029 135285 : let logical_size = &self.current_logical_size;
4030 135285 : logical_size.increment_size(delta);
4031 :
4032 : // Also set the value in the prometheus gauge. Note that
4033 : // there is a race condition here: if this is is called by two
4034 : // threads concurrently, the prometheus gauge might be set to
4035 : // one value while current_logical_size is set to the
4036 : // other.
4037 135285 : match logical_size.current_size() {
4038 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
4039 135285 : .metrics
4040 135285 : .current_logical_size_gauge
4041 135285 : .set(new_current_size.into()),
4042 0 : CurrentLogicalSize::Approximate(_) => {
4043 0 : // don't update the gauge yet, this allows us not to update the gauge back and
4044 0 : // forth between the initial size calculation task.
4045 0 : }
4046 : }
4047 135285 : }
4048 :
4049 1527 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
4050 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
4051 : // for each of the database, but we only store one value, and therefore each pgdirmodification
4052 : // would overwrite the previous value if they modify different databases.
4053 :
4054 1527 : match count {
4055 566 : MetricsUpdate::Set(count) => {
4056 566 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
4057 566 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
4058 566 : }
4059 960 : MetricsUpdate::Add(count) => {
4060 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4061 : // it's fine.
4062 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4063 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
4064 960 : // the value reliably.
4065 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
4066 960 : }
4067 : // Otherwise, ignore this update
4068 : }
4069 1 : MetricsUpdate::Sub(count) => {
4070 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4071 : // it's fine.
4072 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4073 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
4074 1 : // The operation could overflow so we need to normalize the value.
4075 1 : let prev_val =
4076 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
4077 1 : let res = prev_val.saturating_sub(count);
4078 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
4079 1 : }
4080 : // Otherwise, ignore this update
4081 : }
4082 : };
4083 :
4084 : // TODO: remove this, there's no place in the code that updates this aux metrics.
4085 1527 : let aux_metric =
4086 1527 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
4087 :
4088 1527 : let sum_of_entries = self
4089 1527 : .directory_metrics
4090 1527 : .iter()
4091 12216 : .map(|v| v.load(AtomicOrdering::Relaxed))
4092 1527 : .sum();
4093 : // Set a high general threshold and a lower threshold for the auxiliary files,
4094 : // as we can have large numbers of relations in the db directory.
4095 : const SUM_THRESHOLD: u64 = 5000;
4096 : const AUX_THRESHOLD: u64 = 1000;
4097 1527 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
4098 0 : self.metrics
4099 0 : .directory_entries_count_gauge
4100 0 : .set(sum_of_entries);
4101 1527 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
4102 0 : metric.set(sum_of_entries);
4103 1527 : }
4104 1527 : }
4105 :
4106 0 : async fn find_layer(
4107 0 : &self,
4108 0 : layer_name: &LayerName,
4109 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
4110 0 : let guard = self
4111 0 : .layers
4112 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
4113 0 : .await;
4114 0 : let layer = guard
4115 0 : .layer_map()?
4116 0 : .iter_historic_layers()
4117 0 : .find(|l| &l.layer_name() == layer_name)
4118 0 : .map(|found| guard.get_from_desc(&found));
4119 0 : Ok(layer)
4120 0 : }
4121 :
4122 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
4123 0 : let crnt = self.previous_heatmap.load();
4124 0 : match crnt.as_deref() {
4125 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
4126 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
4127 0 : None => true,
4128 : },
4129 0 : Some(PreviousHeatmap::Obsolete) => false,
4130 0 : None => false,
4131 : }
4132 0 : }
4133 :
4134 : /// The timeline heatmap is a hint to secondary locations from the primary location,
4135 : /// indicating which layers are currently on-disk on the primary.
4136 : ///
4137 : /// None is returned if the Timeline is in a state where uploading a heatmap
4138 : /// doesn't make sense, such as shutting down or initializing. The caller
4139 : /// should treat this as a cue to simply skip doing any heatmap uploading
4140 : /// for this timeline.
4141 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
4142 8 : if !self.is_active() {
4143 0 : return None;
4144 8 : }
4145 :
4146 8 : let guard = self
4147 8 : .layers
4148 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
4149 8 : .await;
4150 :
4151 : // Firstly, if there's any heatmap left over from when this location
4152 : // was a secondary, take that into account. Keep layers that are:
4153 : // * present in the layer map
4154 : // * visible
4155 : // * non-resident
4156 : // * not evicted since we read the heatmap
4157 : //
4158 : // Without this, a new cold, attached location would clobber the previous
4159 : // heatamp.
4160 8 : let previous_heatmap = self.previous_heatmap.load();
4161 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4162 : Some(PreviousHeatmap::Active {
4163 6 : heatmap, read_at, ..
4164 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4165 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4166 23 : let layer = guard.try_get_from_key(&desc.key())?;
4167 :
4168 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4169 0 : return None;
4170 23 : }
4171 :
4172 23 : if layer.is_likely_resident() {
4173 10 : return None;
4174 13 : }
4175 :
4176 13 : if layer.last_evicted_at().happened_after(*read_at) {
4177 3 : return None;
4178 10 : }
4179 :
4180 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4181 23 : })),
4182 0 : Some(PreviousHeatmap::Obsolete) => None,
4183 2 : None => None,
4184 : };
4185 :
4186 : // Secondly, all currently visible, resident layers are included.
4187 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4188 18 : match layer.visibility() {
4189 : LayerVisibilityHint::Visible => {
4190 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4191 17 : let last_activity_ts = layer.latest_activity();
4192 17 : Some((
4193 17 : layer.layer_desc().clone(),
4194 17 : layer.metadata(),
4195 17 : last_activity_ts,
4196 17 : false, // these layers are not cold
4197 17 : ))
4198 : }
4199 : LayerVisibilityHint::Covered => {
4200 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4201 1 : None
4202 : }
4203 : }
4204 18 : });
4205 :
4206 8 : let mut layers = match visible_non_resident {
4207 6 : Some(non_resident) => {
4208 6 : let mut non_resident = non_resident.peekable();
4209 6 : if non_resident.peek().is_none() {
4210 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4211 2 : self.previous_heatmap
4212 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4213 4 : }
4214 :
4215 6 : non_resident.chain(resident).collect::<Vec<_>>()
4216 : }
4217 2 : None => resident.collect::<Vec<_>>(),
4218 : };
4219 :
4220 : // Sort layers in order of which to download first. For a large set of layers to download, we
4221 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4222 : // or hours later:
4223 : // - Cold layers go last for convenience when a human inspects the heatmap.
4224 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4225 : // only exist for a few minutes before being compacted into L1s.
4226 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4227 : // the layer is likely to be covered by an image layer during compaction.
4228 60 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4229 60 : std::cmp::Reverse((
4230 60 : *cold,
4231 60 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4232 60 : desc.lsn_range.end,
4233 60 : ))
4234 60 : });
4235 :
4236 8 : let layers = layers
4237 8 : .into_iter()
4238 27 : .map(|(desc, meta, atime, cold)| {
4239 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4240 27 : })
4241 8 : .collect();
4242 :
4243 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4244 8 : }
4245 :
4246 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4247 0 : let guard = self
4248 0 : .layers
4249 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4250 0 : .await;
4251 :
4252 0 : let now = SystemTime::now();
4253 0 : let mut heatmap_layers = Vec::default();
4254 0 : for vl in guard.visible_layers() {
4255 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4256 0 : continue;
4257 0 : }
4258 :
4259 0 : let hl = HeatMapLayer {
4260 0 : name: vl.layer_desc().layer_name(),
4261 0 : metadata: vl.metadata(),
4262 0 : access_time: now,
4263 0 : cold: true,
4264 0 : };
4265 0 : heatmap_layers.push(hl);
4266 : }
4267 :
4268 0 : tracing::info!(
4269 0 : "Generating unarchival heatmap with {} layers",
4270 0 : heatmap_layers.len()
4271 : );
4272 :
4273 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4274 0 : PreviousHeatmap::Active {
4275 0 : heatmap,
4276 0 : read_at: Instant::now(),
4277 0 : end_lsn: Some(end_lsn),
4278 0 : }
4279 0 : }
4280 :
4281 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4282 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4283 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4284 : // branchpoint in the value in IndexPart::lineage
4285 0 : self.ancestor_lsn == lsn
4286 0 : || (self.ancestor_lsn == Lsn::INVALID
4287 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4288 0 : }
4289 : }
4290 :
4291 : #[derive(Clone)]
4292 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4293 : /// In other words, a set of segments in a 2D space.
4294 : ///
4295 : /// This representation has the advatange of avoiding hash map
4296 : /// allocations for uniform queries.
4297 : pub(crate) enum VersionedKeySpaceQuery {
4298 : /// Variant for queries at a single [`Lsn`]
4299 : Uniform { keyspace: KeySpace, lsn: Lsn },
4300 : /// Variant for queries at multiple [`Lsn`]s
4301 : Scattered {
4302 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4303 : },
4304 : }
4305 :
4306 : impl VersionedKeySpaceQuery {
4307 302222 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4308 302222 : Self::Uniform { keyspace, lsn }
4309 302222 : }
4310 :
4311 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4312 10192 : Self::Scattered { keyspaces_at_lsn }
4313 10192 : }
4314 :
4315 : /// Returns the most recent (largest) LSN included in the query.
4316 : /// If any of the LSNs included in the query are invalid, returns
4317 : /// an error instead.
4318 624828 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4319 624828 : match self {
4320 604444 : Self::Uniform { lsn, .. } => {
4321 604444 : if !lsn.is_valid() {
4322 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4323 604444 : }
4324 :
4325 604444 : Ok(*lsn)
4326 : }
4327 20384 : Self::Scattered { keyspaces_at_lsn } => {
4328 20384 : let mut max_lsn = None;
4329 42218 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4330 42218 : if !lsn.is_valid() {
4331 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4332 42218 : }
4333 42218 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4334 : }
4335 :
4336 20384 : if let Some(computed) = max_lsn {
4337 20384 : Ok(*computed)
4338 : } else {
4339 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4340 : }
4341 : }
4342 : }
4343 624828 : }
4344 :
4345 : /// Returns the total keyspace being queried: the result of projecting
4346 : /// everything in the key dimensions onto the key axis.
4347 323491 : fn total_keyspace(&self) -> KeySpace {
4348 323491 : match self {
4349 303107 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4350 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4351 20384 : .iter()
4352 20384 : .map(|(_lsn, keyspace)| keyspace)
4353 42218 : .fold(KeySpace::default(), |mut acc, v| {
4354 42218 : acc.merge(v);
4355 42218 : acc
4356 42218 : }),
4357 : }
4358 323491 : }
4359 :
4360 : /// Returns LSN for a specific key.
4361 : ///
4362 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4363 395535 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4364 395535 : match self {
4365 322343 : Self::Uniform { lsn, .. } => *lsn,
4366 73192 : Self::Scattered { keyspaces_at_lsn } => {
4367 73192 : keyspaces_at_lsn
4368 73192 : .iter()
4369 418280 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4370 73192 : .expect("Returned key was requested")
4371 : .0
4372 : }
4373 : }
4374 395535 : }
4375 :
4376 : /// Remove any parts of the query (segments) which overlap with the provided
4377 : /// key space (also segments).
4378 963769 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4379 963769 : match self {
4380 943385 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4381 20384 : Self::Scattered { keyspaces_at_lsn } => {
4382 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4383 42218 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4384 42218 : let removed = keyspace.remove_overlapping_with(to_remove);
4385 42218 : removed_accum.add_keyspace(removed);
4386 42218 : });
4387 :
4388 20384 : removed_accum.to_keyspace()
4389 : }
4390 : }
4391 963769 : }
4392 :
4393 737808 : fn is_empty(&self) -> bool {
4394 737808 : match self {
4395 717424 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4396 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4397 20384 : .iter()
4398 31301 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4399 : }
4400 737808 : }
4401 :
4402 : /// "Lower" the query on the LSN dimension
4403 112981 : fn lower(&mut self, to: Lsn) {
4404 112981 : match self {
4405 112981 : Self::Uniform { lsn, .. } => {
4406 112981 : // If the originally requested LSN is smaller than the starting
4407 112981 : // LSN of the ancestor we are descending into, we need to respect that.
4408 112981 : // Hence the min.
4409 112981 : *lsn = std::cmp::min(*lsn, to);
4410 112981 : }
4411 0 : Self::Scattered { keyspaces_at_lsn } => {
4412 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4413 0 : *lsn = std::cmp::min(*lsn, to);
4414 0 : });
4415 : }
4416 : }
4417 112981 : }
4418 : }
4419 :
4420 : impl std::fmt::Display for VersionedKeySpaceQuery {
4421 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4422 0 : write!(f, "[")?;
4423 :
4424 0 : match self {
4425 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4426 0 : write!(f, "{keyspace} @ {lsn}")?;
4427 : }
4428 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4429 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4430 0 : write!(f, "{keyspace} @ {lsn},")?;
4431 : }
4432 : }
4433 : }
4434 :
4435 0 : write!(f, "]")
4436 0 : }
4437 : }
4438 :
4439 : impl Timeline {
4440 : #[allow(clippy::doc_lazy_continuation)]
4441 : /// Get the data needed to reconstruct all keys in the provided keyspace
4442 : ///
4443 : /// The algorithm is as follows:
4444 : /// 1. While some keys are still not done and there's a timeline to visit:
4445 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4446 : /// 2.1: Build the fringe for the current keyspace
4447 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4448 : /// intersects
4449 : /// 2.3. Pop the timeline from the fringe
4450 : /// 2.4. If the fringe is empty, go back to 1
4451 312414 : async fn get_vectored_reconstruct_data(
4452 312414 : &self,
4453 312414 : mut query: VersionedKeySpaceQuery,
4454 312414 : reconstruct_state: &mut ValuesReconstructState,
4455 312414 : ctx: &RequestContext,
4456 312414 : ) -> Result<(), GetVectoredError> {
4457 312414 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4458 :
4459 : let mut timeline_owned: Arc<Timeline>;
4460 312414 : let mut timeline = self;
4461 :
4462 312413 : let missing_keyspace = loop {
4463 425394 : if self.cancel.is_cancelled() {
4464 0 : return Err(GetVectoredError::Cancelled);
4465 425394 : }
4466 :
4467 : let TimelineVisitOutcome {
4468 425394 : completed_keyspace: completed,
4469 425394 : image_covered_keyspace,
4470 : } = {
4471 425394 : let ctx = RequestContextBuilder::from(ctx)
4472 425394 : .perf_span(|crnt_perf_span| {
4473 0 : info_span!(
4474 : target: PERF_TRACE_TARGET,
4475 0 : parent: crnt_perf_span,
4476 : "PLAN_IO_TIMELINE",
4477 : timeline = %timeline.timeline_id,
4478 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4479 : )
4480 0 : })
4481 425394 : .attached_child();
4482 :
4483 425394 : Self::get_vectored_reconstruct_data_timeline(
4484 425394 : timeline,
4485 425394 : &query,
4486 425394 : reconstruct_state,
4487 425394 : &self.cancel,
4488 425394 : &ctx,
4489 : )
4490 425394 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4491 425394 : .await?
4492 : };
4493 :
4494 425394 : query.remove_overlapping_with(&completed);
4495 :
4496 : // Do not descend into the ancestor timeline for aux files.
4497 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4498 : // stalling compaction.
4499 425394 : query.remove_overlapping_with(&KeySpace {
4500 425394 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4501 425394 : });
4502 :
4503 : // Keyspace is fully retrieved
4504 425394 : if query.is_empty() {
4505 312230 : break None;
4506 113164 : }
4507 :
4508 113164 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4509 : // Not fully retrieved but no ancestor timeline.
4510 183 : break Some(query.total_keyspace());
4511 : };
4512 :
4513 : // Now we see if there are keys covered by the image layer but does not exist in the
4514 : // image layer, which means that the key does not exist.
4515 :
4516 : // The block below will stop the vectored search if any of the keys encountered an image layer
4517 : // which did not contain a snapshot for said key. Since we have already removed all completed
4518 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4519 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4520 : // and stop the search as a result of that.
4521 112981 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4522 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4523 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4524 : // figuring out what is the inherited key range and do a fine-grained pruning.
4525 112981 : removed.remove_overlapping_with(&KeySpace {
4526 112981 : ranges: vec![SPARSE_RANGE],
4527 112981 : });
4528 112981 : if !removed.is_empty() {
4529 0 : break Some(removed);
4530 112981 : }
4531 :
4532 : // Each key range in the original query is at some point in the LSN space.
4533 : // When descending into the ancestor, lower all ranges in the LSN space
4534 : // such that new changes on the parent timeline are not visible.
4535 112981 : query.lower(timeline.ancestor_lsn);
4536 :
4537 112981 : let ctx = RequestContextBuilder::from(ctx)
4538 112981 : .perf_span(|crnt_perf_span| {
4539 0 : info_span!(
4540 : target: PERF_TRACE_TARGET,
4541 0 : parent: crnt_perf_span,
4542 : "GET_ANCESTOR",
4543 : timeline = %timeline.timeline_id,
4544 0 : ancestor = %ancestor_timeline.timeline_id,
4545 : ancestor_lsn = %timeline.ancestor_lsn
4546 : )
4547 0 : })
4548 112981 : .attached_child();
4549 :
4550 112981 : timeline_owned = timeline
4551 112981 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4552 112981 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4553 112981 : .await?;
4554 112980 : timeline = &*timeline_owned;
4555 : };
4556 :
4557 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4558 312413 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4559 183 : let mut missing_keyspace = missing_keyspace;
4560 183 : missing_keyspace.remove_overlapping_with(&KeySpace {
4561 183 : ranges: vec![SPARSE_RANGE],
4562 183 : });
4563 183 : if missing_keyspace.is_empty() {
4564 176 : None
4565 : } else {
4566 7 : Some(missing_keyspace)
4567 : }
4568 : } else {
4569 312230 : None
4570 : };
4571 :
4572 312413 : if let Some(missing_keyspace) = missing_keyspace {
4573 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4574 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4575 7 : shard: self.shard_identity.number,
4576 7 : original_hwm_lsn,
4577 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4578 7 : backtrace: None,
4579 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4580 7 : query: None,
4581 7 : })));
4582 312406 : }
4583 :
4584 312406 : Ok(())
4585 312414 : }
4586 :
4587 425394 : async fn get_vectored_init_fringe(
4588 425394 : &self,
4589 425394 : query: &VersionedKeySpaceQuery,
4590 425394 : ) -> Result<LayerFringe, GetVectoredError> {
4591 425394 : let mut fringe = LayerFringe::new();
4592 425394 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4593 :
4594 425394 : match query {
4595 415202 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4596 : // LSNs requested by the compute or determined by the pageserver
4597 : // are inclusive. Queries to the layer map use exclusive LSNs.
4598 : // Hence, bump the value before the query - same in the other
4599 : // match arm.
4600 415202 : let cont_lsn = Lsn(lsn.0 + 1);
4601 415202 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4602 : }
4603 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4604 21109 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4605 21109 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4606 21109 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4607 : }
4608 : }
4609 : }
4610 :
4611 425394 : Ok(fringe)
4612 425394 : }
4613 :
4614 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4615 : ///
4616 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4617 : /// the current keyspace. The current keyspace of the search at any given timeline
4618 : /// is the original keyspace minus all the keys that have been completed minus
4619 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4620 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4621 : ///
4622 : /// This is basically a depth-first search visitor implementation where a vertex
4623 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4624 : ///
4625 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4626 : /// and get all the required reconstruct data from the layer in one go.
4627 : ///
4628 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4629 : /// decides how to deal with these two keyspaces.
4630 425394 : async fn get_vectored_reconstruct_data_timeline(
4631 425394 : timeline: &Timeline,
4632 425394 : query: &VersionedKeySpaceQuery,
4633 425394 : reconstruct_state: &mut ValuesReconstructState,
4634 425394 : cancel: &CancellationToken,
4635 425394 : ctx: &RequestContext,
4636 425394 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4637 : // Prevent GC from progressing while visiting the current timeline.
4638 : // If we are GC-ing because a new image layer was added while traversing
4639 : // the timeline, then it will remove layers that are required for fulfilling
4640 : // the current get request (read-path cannot "look back" and notice the new
4641 : // image layer).
4642 425394 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4643 :
4644 : // See `compaction::compact_with_gc` for why we need this.
4645 425394 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4646 :
4647 : // Initialize the fringe
4648 425394 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4649 :
4650 425394 : let mut completed_keyspace = KeySpace::default();
4651 425394 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4652 :
4653 871556 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4654 446162 : if cancel.is_cancelled() {
4655 0 : return Err(GetVectoredError::Cancelled);
4656 446162 : }
4657 :
4658 446162 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4659 446162 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4660 446162 : }
4661 :
4662 : // Visit the layer and plan IOs for it
4663 446162 : let next_cont_lsn = lsn_range.start;
4664 446162 : layer_to_read
4665 446162 : .get_values_reconstruct_data(
4666 446162 : keyspace_to_read.clone(),
4667 446162 : lsn_range,
4668 446162 : reconstruct_state,
4669 446162 : ctx,
4670 446162 : )
4671 446162 : .await?;
4672 :
4673 446162 : let mut unmapped_keyspace = keyspace_to_read;
4674 446162 : let cont_lsn = next_cont_lsn;
4675 :
4676 446162 : reconstruct_state.on_layer_visited(&layer_to_read);
4677 :
4678 446162 : let (keys_done_last_step, keys_with_image_coverage) =
4679 446162 : reconstruct_state.consume_done_keys();
4680 446162 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4681 446162 : completed_keyspace.merge(&keys_done_last_step);
4682 446162 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4683 14122 : unmapped_keyspace
4684 14122 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4685 14122 : image_covered_keyspace.add_range(keys_with_image_coverage);
4686 432040 : }
4687 :
4688 : // Query the layer map for the next layers to read.
4689 : //
4690 : // Do not descent any further if the last layer we visited
4691 : // completed all keys in the keyspace it inspected. This is not
4692 : // required for correctness, but avoids visiting extra layers
4693 : // which turns out to be a perf bottleneck in some cases.
4694 446162 : if !unmapped_keyspace.is_empty() {
4695 129478 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4696 129478 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4697 :
4698 : // It's safe to drop the layer map lock after planning the next round of reads.
4699 : // The fringe keeps readable handles for the layers which are safe to read even
4700 : // if layers were compacted or flushed.
4701 : //
4702 : // The more interesting consideration is: "Why is the read algorithm still correct
4703 : // if the layer map changes while it is operating?". Doing a vectored read on a
4704 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4705 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4706 : // range at *a particular point in time*. It is fine for the answer to be different
4707 : // at two different time points.
4708 129478 : drop(guard);
4709 316684 : }
4710 : }
4711 :
4712 425394 : Ok(TimelineVisitOutcome {
4713 425394 : completed_keyspace,
4714 425394 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4715 425394 : })
4716 425394 : }
4717 :
4718 112981 : async fn get_ready_ancestor_timeline(
4719 112981 : &self,
4720 112981 : ancestor: &Arc<Timeline>,
4721 112981 : ctx: &RequestContext,
4722 112981 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4723 : // It's possible that the ancestor timeline isn't active yet, or
4724 : // is active but hasn't yet caught up to the branch point. Wait
4725 : // for it.
4726 : //
4727 : // This cannot happen while the pageserver is running normally,
4728 : // because you cannot create a branch from a point that isn't
4729 : // present in the pageserver yet. However, we don't wait for the
4730 : // branch point to be uploaded to cloud storage before creating
4731 : // a branch. I.e., the branch LSN need not be remote consistent
4732 : // for the branching operation to succeed.
4733 : //
4734 : // Hence, if we try to load a tenant in such a state where
4735 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4736 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4737 : // then we will need to wait for the ancestor timeline to
4738 : // re-stream WAL up to branch_lsn before we access it.
4739 : //
4740 : // How can a tenant get in such a state?
4741 : // - ungraceful pageserver process exit
4742 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4743 : //
4744 : // NB: this could be avoided by requiring
4745 : // branch_lsn >= remote_consistent_lsn
4746 : // during branch creation.
4747 112981 : match ancestor.wait_to_become_active(ctx).await {
4748 112980 : Ok(()) => {}
4749 : Err(TimelineState::Stopping) => {
4750 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4751 0 : return Err(GetReadyAncestorError::Cancelled);
4752 : }
4753 1 : Err(state) => {
4754 1 : return Err(GetReadyAncestorError::BadState {
4755 1 : timeline_id: ancestor.timeline_id,
4756 1 : state,
4757 1 : });
4758 : }
4759 : }
4760 112980 : ancestor
4761 112980 : .wait_lsn(
4762 112980 : self.ancestor_lsn,
4763 112980 : WaitLsnWaiter::Timeline(self),
4764 112980 : WaitLsnTimeout::Default,
4765 112980 : ctx,
4766 112980 : )
4767 112980 : .await
4768 112980 : .map_err(|e| match e {
4769 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4770 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4771 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4772 0 : timeline_id: ancestor.timeline_id,
4773 0 : state,
4774 0 : },
4775 0 : })?;
4776 :
4777 112980 : Ok(ancestor.clone())
4778 112981 : }
4779 :
4780 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4781 148592 : &self.shard_identity
4782 148592 : }
4783 :
4784 : #[inline(always)]
4785 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4786 0 : ShardTimelineId {
4787 0 : shard_index: ShardIndex {
4788 0 : shard_number: self.shard_identity.number,
4789 0 : shard_count: self.shard_identity.count,
4790 0 : },
4791 0 : timeline_id: self.timeline_id,
4792 0 : }
4793 0 : }
4794 :
4795 : /// Returns a non-frozen open in-memory layer for ingestion.
4796 : ///
4797 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4798 : /// this function without holding the mutex.
4799 660 : async fn get_layer_for_write(
4800 660 : &self,
4801 660 : lsn: Lsn,
4802 660 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4803 660 : ctx: &RequestContext,
4804 660 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4805 660 : let mut guard = self
4806 660 : .layers
4807 660 : .write(LayerManagerLockHolder::GetLayerForWrite)
4808 660 : .await;
4809 :
4810 660 : let last_record_lsn = self.get_last_record_lsn();
4811 660 : ensure!(
4812 660 : lsn > last_record_lsn,
4813 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4814 : lsn,
4815 : last_record_lsn,
4816 : );
4817 :
4818 660 : let layer = guard
4819 660 : .open_mut()?
4820 660 : .get_layer_for_write(
4821 660 : lsn,
4822 660 : self.conf,
4823 660 : self.timeline_id,
4824 660 : self.tenant_shard_id,
4825 660 : &self.gate,
4826 660 : &self.cancel,
4827 660 : ctx,
4828 : )
4829 660 : .await?;
4830 660 : Ok(layer)
4831 660 : }
4832 :
4833 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4834 2639560 : assert!(new_lsn.is_aligned());
4835 :
4836 2639560 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4837 2639560 : self.last_record_lsn.advance(new_lsn);
4838 2639560 : }
4839 :
4840 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4841 : ///
4842 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4843 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4844 611 : async fn freeze_inmem_layer_at(
4845 611 : &self,
4846 611 : at: Lsn,
4847 611 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4848 611 : ) -> Result<u64, FlushLayerError> {
4849 611 : let frozen = {
4850 611 : let mut guard = self
4851 611 : .layers
4852 611 : .write(LayerManagerLockHolder::TryFreezeLayer)
4853 611 : .await;
4854 611 : guard
4855 611 : .open_mut()?
4856 611 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4857 611 : .await
4858 : };
4859 :
4860 611 : if frozen {
4861 597 : let now = Instant::now();
4862 597 : *(self.last_freeze_ts.write().unwrap()) = now;
4863 597 : }
4864 :
4865 : // Increment the flush cycle counter and wake up the flush task.
4866 : // Remember the new value, so that when we listen for the flush
4867 : // to finish, we know when the flush that we initiated has
4868 : // finished, instead of some other flush that was started earlier.
4869 611 : let mut my_flush_request = 0;
4870 :
4871 611 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4872 611 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4873 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4874 611 : }
4875 :
4876 611 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4877 611 : my_flush_request = *counter + 1;
4878 611 : *counter = my_flush_request;
4879 611 : *lsn = std::cmp::max(at, *lsn);
4880 611 : });
4881 :
4882 611 : assert_ne!(my_flush_request, 0);
4883 :
4884 611 : Ok(my_flush_request)
4885 611 : }
4886 :
4887 : /// Layer flusher task's main loop.
4888 232 : async fn flush_loop(
4889 232 : self: &Arc<Self>,
4890 232 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4891 232 : ctx: &RequestContext,
4892 232 : ) {
4893 : // Always notify waiters about the flush loop exiting since the loop might stop
4894 : // when the timeline hasn't been cancelled.
4895 232 : let scopeguard_rx = layer_flush_start_rx.clone();
4896 232 : scopeguard::defer! {
4897 : let (flush_counter, _) = *scopeguard_rx.borrow();
4898 : let _ = self
4899 : .layer_flush_done_tx
4900 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4901 : }
4902 :
4903 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4904 232 : let mut watch_l0 = match self
4905 232 : .layers
4906 232 : .read(LayerManagerLockHolder::FlushLoop)
4907 232 : .await
4908 232 : .layer_map()
4909 : {
4910 232 : Ok(lm) => lm.watch_level0_deltas(),
4911 0 : Err(Shutdown) => return,
4912 : };
4913 :
4914 232 : info!("started flush loop");
4915 : loop {
4916 830 : tokio::select! {
4917 830 : _ = self.cancel.cancelled() => {
4918 5 : info!("shutting down layer flush task due to Timeline::cancel");
4919 5 : break;
4920 : },
4921 830 : _ = layer_flush_start_rx.changed() => {}
4922 : }
4923 598 : trace!("waking up");
4924 598 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4925 :
4926 : // The highest LSN to which we flushed in the loop over frozen layers
4927 598 : let mut flushed_to_lsn = Lsn(0);
4928 :
4929 598 : let result = loop {
4930 1194 : if self.cancel.is_cancelled() {
4931 0 : info!("dropping out of flush loop for timeline shutdown");
4932 0 : return;
4933 1194 : }
4934 :
4935 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4936 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4937 1194 : if flushed_to_lsn >= frozen_to_lsn {
4938 583 : break Ok(());
4939 611 : }
4940 :
4941 : // Fetch the next layer to flush, if any.
4942 611 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4943 611 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4944 611 : let Ok(lm) = layers.layer_map() else {
4945 0 : info!("dropping out of flush loop for timeline shutdown");
4946 0 : return;
4947 : };
4948 611 : let l0_count = lm.level0_deltas().len();
4949 611 : let frozen_count = lm.frozen_layers.len();
4950 611 : let frozen_size: u64 = lm
4951 611 : .frozen_layers
4952 611 : .iter()
4953 611 : .map(|l| l.estimated_in_mem_size())
4954 611 : .sum();
4955 611 : let open_layer_size: u64 = lm
4956 611 : .open_layer
4957 611 : .as_ref()
4958 611 : .map(|l| l.estimated_in_mem_size())
4959 611 : .unwrap_or(0);
4960 611 : let layer = lm.frozen_layers.front().cloned();
4961 611 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
4962 : // drop 'layers' lock
4963 : };
4964 611 : let Some(layer) = layer else {
4965 14 : break Ok(());
4966 : };
4967 :
4968 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4969 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4970 597 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4971 0 : if l0_count >= stall_threshold {
4972 0 : warn!(
4973 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4974 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4975 : );
4976 0 : let stall_timer = self
4977 0 : .metrics
4978 0 : .flush_delay_histo
4979 0 : .start_timer()
4980 0 : .record_on_drop();
4981 0 : tokio::select! {
4982 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4983 0 : if let Ok(l0) = result.as_deref() {
4984 0 : let delay = stall_timer.elapsed().as_secs_f64();
4985 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4986 0 : }
4987 : },
4988 0 : _ = self.cancel.cancelled() => {},
4989 : }
4990 0 : continue; // check again
4991 0 : }
4992 597 : }
4993 :
4994 : // Flush the layer.
4995 597 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4996 597 : match self.flush_frozen_layer(layer, ctx).await {
4997 596 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4998 : Err(FlushLayerError::Cancelled) => {
4999 0 : info!("dropping out of flush loop for timeline shutdown");
5000 0 : return;
5001 : }
5002 1 : err @ Err(
5003 : FlushLayerError::NotRunning(_)
5004 : | FlushLayerError::Other(_)
5005 : | FlushLayerError::CreateImageLayersError(_),
5006 : ) => {
5007 1 : error!("could not flush frozen layer: {err:?}");
5008 1 : break err.map(|_| ());
5009 : }
5010 : }
5011 596 : let flush_duration = flush_timer.stop_and_record();
5012 :
5013 : // Notify the tenant compaction loop if L0 compaction is needed.
5014 596 : let l0_count = *watch_l0.borrow();
5015 596 : if l0_count >= self.get_compaction_threshold() {
5016 238 : self.l0_compaction_trigger.notify_one();
5017 358 : }
5018 :
5019 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
5020 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
5021 : // ingestion by having ephemeral layer rolls wait for flushes.
5022 596 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
5023 4 : if l0_count >= delay_threshold {
5024 0 : let delay = flush_duration.as_secs_f64();
5025 0 : info!(
5026 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
5027 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5028 : );
5029 0 : let _delay_timer = self
5030 0 : .metrics
5031 0 : .flush_delay_histo
5032 0 : .start_timer()
5033 0 : .record_on_drop();
5034 0 : tokio::select! {
5035 0 : _ = tokio::time::sleep(flush_duration) => {},
5036 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
5037 0 : _ = self.cancel.cancelled() => {},
5038 : }
5039 4 : }
5040 592 : }
5041 : };
5042 :
5043 : // Unsharded tenants should never advance their LSN beyond the end of the
5044 : // highest layer they write: such gaps between layer data and the frozen LSN
5045 : // are only legal on sharded tenants.
5046 598 : debug_assert!(
5047 598 : self.shard_identity.count.count() > 1
5048 591 : || flushed_to_lsn >= frozen_to_lsn
5049 14 : || !flushed_to_lsn.is_valid()
5050 : );
5051 :
5052 598 : if flushed_to_lsn < frozen_to_lsn
5053 15 : && self.shard_identity.count.count() > 1
5054 1 : && result.is_ok()
5055 : {
5056 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
5057 : // to us via layer_flush_start_rx, then advance it here.
5058 : //
5059 : // This path is only taken for tenants with multiple shards: single sharded tenants should
5060 : // never encounter a gap in the wal.
5061 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
5062 0 : tracing::debug!(
5063 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
5064 : );
5065 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
5066 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
5067 0 : tracing::warn!(
5068 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
5069 : );
5070 0 : }
5071 0 : }
5072 598 : }
5073 :
5074 : // Notify any listeners that we're done
5075 598 : let _ = self
5076 598 : .layer_flush_done_tx
5077 598 : .send_replace((flush_counter, result));
5078 : }
5079 5 : }
5080 :
5081 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
5082 571 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
5083 571 : let mut rx = self.layer_flush_done_tx.subscribe();
5084 : loop {
5085 : {
5086 1161 : let (last_result_counter, last_result) = &*rx.borrow();
5087 1161 : if *last_result_counter >= request {
5088 571 : if let Err(err) = last_result {
5089 : // We already logged the original error in
5090 : // flush_loop. We cannot propagate it to the caller
5091 : // here, because it might not be Cloneable
5092 1 : return Err(err.clone());
5093 : } else {
5094 570 : return Ok(());
5095 : }
5096 590 : }
5097 : }
5098 590 : trace!("waiting for flush to complete");
5099 590 : tokio::select! {
5100 590 : rx_e = rx.changed() => {
5101 590 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
5102 : },
5103 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
5104 : // the notification from [`flush_loop`] that it completed.
5105 590 : _ = self.cancel.cancelled() => {
5106 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
5107 0 : return Ok(())
5108 : }
5109 : };
5110 590 : trace!("done")
5111 : }
5112 571 : }
5113 :
5114 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
5115 : ///
5116 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
5117 : #[instrument(skip_all, fields(layer=%frozen_layer))]
5118 : async fn flush_frozen_layer(
5119 : self: &Arc<Self>,
5120 : frozen_layer: Arc<InMemoryLayer>,
5121 : ctx: &RequestContext,
5122 : ) -> Result<Lsn, FlushLayerError> {
5123 : debug_assert_current_span_has_tenant_and_timeline_id();
5124 :
5125 : // As a special case, when we have just imported an image into the repository,
5126 : // instead of writing out a L0 delta layer, we directly write out image layer
5127 : // files instead. This is possible as long as *all* the data imported into the
5128 : // repository have the same LSN.
5129 : let lsn_range = frozen_layer.get_lsn_range();
5130 :
5131 : // Whether to directly create image layers for this flush, or flush them as delta layers
5132 : let create_image_layer =
5133 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
5134 :
5135 : #[cfg(test)]
5136 : {
5137 : match &mut *self.flush_loop_state.lock().unwrap() {
5138 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
5139 : panic!("flush loop not running")
5140 : }
5141 : FlushLoopState::Running {
5142 : expect_initdb_optimization,
5143 : initdb_optimization_count,
5144 : ..
5145 : } => {
5146 : if create_image_layer {
5147 : *initdb_optimization_count += 1;
5148 : } else {
5149 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
5150 : }
5151 : }
5152 : }
5153 : }
5154 :
5155 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5156 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5157 : // require downloading anything during initial import.
5158 : let ((rel_partition, metadata_partition), _lsn) = self
5159 : .repartition(
5160 : self.initdb_lsn,
5161 : self.get_compaction_target_size(),
5162 : EnumSet::empty(),
5163 : ctx,
5164 : )
5165 : .await
5166 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5167 :
5168 : if self.cancel.is_cancelled() {
5169 : return Err(FlushLayerError::Cancelled);
5170 : }
5171 :
5172 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5173 : // So that the key ranges don't overlap.
5174 : let mut partitions = KeyPartitioning::default();
5175 : partitions.parts.extend(rel_partition.parts);
5176 : if !metadata_partition.parts.is_empty() {
5177 : assert_eq!(
5178 : metadata_partition.parts.len(),
5179 : 1,
5180 : "currently sparse keyspace should only contain a single metadata keyspace"
5181 : );
5182 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5183 : // every single key within the keyspace, and therefore, it's safe to force converting it
5184 : // into a dense keyspace before calling this function.
5185 : partitions
5186 : .parts
5187 : .extend(metadata_partition.into_dense().parts);
5188 : }
5189 :
5190 : let mut layers_to_upload = Vec::new();
5191 : let (generated_image_layers, is_complete) = self
5192 : .create_image_layers(
5193 : &partitions,
5194 : self.initdb_lsn,
5195 : None,
5196 : ImageLayerCreationMode::Initial,
5197 : ctx,
5198 : LastImageLayerCreationStatus::Initial,
5199 : false, // don't yield for L0, we're flushing L0
5200 : )
5201 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5202 : .await?;
5203 : debug_assert!(
5204 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5205 : "init image generation mode must fully cover the keyspace"
5206 : );
5207 : layers_to_upload.extend(generated_image_layers);
5208 :
5209 : (layers_to_upload, None)
5210 : } else {
5211 : // Normal case, write out a L0 delta layer file.
5212 : // `create_delta_layer` will not modify the layer map.
5213 : // We will remove frozen layer and add delta layer in one atomic operation later.
5214 : let Some(layer) = self
5215 : .create_delta_layer(&frozen_layer, None, ctx)
5216 : .await
5217 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5218 : else {
5219 : panic!("delta layer cannot be empty if no filter is applied");
5220 : };
5221 : (
5222 : // FIXME: even though we have a single image and single delta layer assumption
5223 : // we push them to vec
5224 : vec![layer.clone()],
5225 : Some(layer),
5226 : )
5227 : };
5228 :
5229 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5230 :
5231 : if self.cancel.is_cancelled() {
5232 : return Err(FlushLayerError::Cancelled);
5233 : }
5234 :
5235 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5236 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5237 1 : });
5238 :
5239 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5240 :
5241 : // The new on-disk layers are now in the layer map. We can remove the
5242 : // in-memory layer from the map now. The flushed layer is stored in
5243 : // the mapping in `create_delta_layer`.
5244 : {
5245 : let mut guard = self
5246 : .layers
5247 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5248 : .await;
5249 :
5250 : guard.open_mut()?.finish_flush_l0_layer(
5251 : delta_layer_to_add.as_ref(),
5252 : &frozen_layer,
5253 : &self.metrics,
5254 : );
5255 :
5256 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5257 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5258 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5259 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5260 : }
5261 : // release lock on 'layers'
5262 : };
5263 :
5264 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5265 : // a compaction can delete the file and then it won't be available for uploads any more.
5266 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5267 : // race situation.
5268 : // See https://github.com/neondatabase/neon/issues/4526
5269 : pausable_failpoint!("flush-frozen-pausable");
5270 :
5271 : // This failpoint is used by another test case `test_pageserver_recovery`.
5272 : fail_point!("flush-frozen-exit");
5273 :
5274 : Ok(Lsn(lsn_range.end.0 - 1))
5275 : }
5276 :
5277 : /// Return true if the value changed
5278 : ///
5279 : /// This function must only be used from the layer flush task.
5280 596 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5281 596 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5282 596 : assert!(
5283 596 : new_value >= old_value,
5284 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5285 : );
5286 :
5287 596 : self.metrics
5288 596 : .disk_consistent_lsn_gauge
5289 596 : .set(new_value.0 as i64);
5290 596 : new_value != old_value
5291 596 : }
5292 :
5293 : /// Update metadata file
5294 623 : fn schedule_uploads(
5295 623 : &self,
5296 623 : disk_consistent_lsn: Lsn,
5297 623 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5298 623 : ) -> anyhow::Result<()> {
5299 : // We can only save a valid 'prev_record_lsn' value on disk if we
5300 : // flushed *all* in-memory changes to disk. We only track
5301 : // 'prev_record_lsn' in memory for the latest processed record, so we
5302 : // don't remember what the correct value that corresponds to some old
5303 : // LSN is. But if we flush everything, then the value corresponding
5304 : // current 'last_record_lsn' is correct and we can store it on disk.
5305 : let RecordLsn {
5306 623 : last: last_record_lsn,
5307 623 : prev: prev_record_lsn,
5308 623 : } = self.last_record_lsn.load();
5309 623 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5310 560 : Some(prev_record_lsn)
5311 : } else {
5312 63 : None
5313 : };
5314 :
5315 623 : let update = crate::tenant::metadata::MetadataUpdate::new(
5316 623 : disk_consistent_lsn,
5317 623 : ondisk_prev_record_lsn,
5318 623 : *self.applied_gc_cutoff_lsn.read(),
5319 : );
5320 :
5321 623 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5322 0 : "{}",
5323 0 : x.unwrap()
5324 : ));
5325 :
5326 1225 : for layer in layers_to_upload {
5327 602 : self.remote_client.schedule_layer_file_upload(layer)?;
5328 : }
5329 623 : self.remote_client
5330 623 : .schedule_index_upload_for_metadata_update(&update)?;
5331 :
5332 623 : Ok(())
5333 623 : }
5334 :
5335 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5336 0 : self.remote_client
5337 0 : .preserve_initdb_archive(
5338 0 : &self.tenant_shard_id.tenant_id,
5339 0 : &self.timeline_id,
5340 0 : &self.cancel,
5341 0 : )
5342 0 : .await
5343 0 : }
5344 :
5345 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5346 : // in layer map immediately. The caller is responsible to put it into the layer map.
5347 486 : async fn create_delta_layer(
5348 486 : self: &Arc<Self>,
5349 486 : frozen_layer: &Arc<InMemoryLayer>,
5350 486 : key_range: Option<Range<Key>>,
5351 486 : ctx: &RequestContext,
5352 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5353 486 : let self_clone = Arc::clone(self);
5354 486 : let frozen_layer = Arc::clone(frozen_layer);
5355 486 : let ctx = ctx.attached_child();
5356 486 : let work = async move {
5357 486 : let Some((desc, path)) = frozen_layer
5358 486 : .write_to_disk(
5359 486 : &ctx,
5360 486 : key_range,
5361 486 : self_clone.l0_flush_global_state.inner(),
5362 486 : &self_clone.gate,
5363 486 : self_clone.cancel.clone(),
5364 486 : )
5365 486 : .await?
5366 : else {
5367 0 : return Ok(None);
5368 : };
5369 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5370 :
5371 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5372 : // We just need to fsync the directory in which these inodes are linked,
5373 : // which we know to be the timeline directory.
5374 : //
5375 : // We use fatal_err() below because the after write_to_disk returns with success,
5376 : // the in-memory state of the filesystem already has the layer file in its final place,
5377 : // and subsequent pageserver code could think it's durable while it really isn't.
5378 486 : let timeline_dir = VirtualFile::open(
5379 486 : &self_clone
5380 486 : .conf
5381 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5382 486 : &ctx,
5383 486 : )
5384 486 : .await
5385 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5386 486 : timeline_dir
5387 486 : .sync_all()
5388 486 : .await
5389 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5390 486 : anyhow::Ok(Some(new_delta))
5391 486 : };
5392 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5393 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5394 : use crate::virtual_file::io_engine::IoEngine;
5395 486 : match crate::virtual_file::io_engine::get() {
5396 0 : IoEngine::NotSet => panic!("io engine not set"),
5397 : IoEngine::StdFs => {
5398 0 : let span = tracing::info_span!("blocking");
5399 0 : tokio::task::spawn_blocking({
5400 0 : move || Handle::current().block_on(work.instrument(span))
5401 : })
5402 0 : .await
5403 0 : .context("spawn_blocking")
5404 0 : .and_then(|x| x)
5405 : }
5406 : #[cfg(target_os = "linux")]
5407 486 : IoEngine::TokioEpollUring => work.await,
5408 : }
5409 486 : }
5410 :
5411 303 : async fn repartition(
5412 303 : &self,
5413 303 : lsn: Lsn,
5414 303 : partition_size: u64,
5415 303 : flags: EnumSet<CompactFlags>,
5416 303 : ctx: &RequestContext,
5417 303 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5418 303 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5419 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5420 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5421 : // and hence before the compaction task starts.
5422 0 : return Err(RepartitionError::Other(anyhow!(
5423 0 : "repartition() called concurrently"
5424 0 : )));
5425 : };
5426 303 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5427 303 : if lsn < *partition_lsn {
5428 0 : return Err(RepartitionError::Other(anyhow!(
5429 0 : "repartition() called with LSN going backwards, this should not happen"
5430 0 : )));
5431 303 : }
5432 :
5433 303 : let distance = lsn.0 - partition_lsn.0;
5434 303 : if *partition_lsn != Lsn(0)
5435 141 : && distance <= self.repartition_threshold
5436 141 : && !flags.contains(CompactFlags::ForceRepartition)
5437 : {
5438 134 : debug!(
5439 : distance,
5440 : threshold = self.repartition_threshold,
5441 0 : "no repartitioning needed"
5442 : );
5443 134 : return Ok((
5444 134 : (dense_partition.clone(), sparse_partition.clone()),
5445 134 : *partition_lsn,
5446 134 : ));
5447 169 : }
5448 :
5449 169 : let (dense_ks, sparse_ks) = self
5450 169 : .collect_keyspace(lsn, ctx)
5451 169 : .await
5452 169 : .map_err(RepartitionError::CollectKeyspace)?;
5453 169 : let dense_partitioning = dense_ks.partition(
5454 169 : &self.shard_identity,
5455 169 : partition_size,
5456 169 : postgres_ffi::BLCKSZ as u64,
5457 : );
5458 169 : let sparse_partitioning = SparseKeyPartitioning {
5459 169 : parts: vec![sparse_ks],
5460 169 : }; // no partitioning for metadata keys for now
5461 169 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5462 169 : guard.write(result.clone());
5463 169 : Ok(result)
5464 303 : }
5465 :
5466 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5467 57 : async fn time_for_new_image_layer(
5468 57 : &self,
5469 57 : partition: &KeySpace,
5470 57 : lsn: Lsn,
5471 57 : force_image_creation_lsn: Option<Lsn>,
5472 57 : ) -> bool {
5473 57 : let threshold = self.get_image_creation_threshold();
5474 :
5475 57 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5476 57 : let Ok(layers) = guard.layer_map() else {
5477 0 : return false;
5478 : };
5479 57 : let mut min_image_lsn: Lsn = Lsn::MAX;
5480 57 : let mut max_deltas = 0;
5481 364 : for part_range in &partition.ranges {
5482 307 : let image_coverage = layers.image_coverage(part_range, lsn);
5483 614 : for (img_range, last_img) in image_coverage {
5484 307 : let img_lsn = if let Some(last_img) = last_img {
5485 0 : last_img.get_lsn_range().end
5486 : } else {
5487 307 : Lsn(0)
5488 : };
5489 : // Let's consider an example:
5490 : //
5491 : // delta layer with LSN range 71-81
5492 : // delta layer with LSN range 81-91
5493 : // delta layer with LSN range 91-101
5494 : // image layer at LSN 100
5495 : //
5496 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5497 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5498 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5499 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5500 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5501 307 : if img_lsn < lsn {
5502 307 : let num_deltas =
5503 307 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5504 :
5505 307 : max_deltas = max_deltas.max(num_deltas);
5506 307 : if num_deltas >= threshold {
5507 0 : debug!(
5508 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5509 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5510 : );
5511 0 : return true;
5512 307 : }
5513 0 : }
5514 307 : min_image_lsn = min(min_image_lsn, img_lsn);
5515 : }
5516 : }
5517 :
5518 : // HADRON
5519 : // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
5520 57 : min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
5521 57 : if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
5522 0 : info!(
5523 0 : "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
5524 0 : partition.ranges[0].start,
5525 0 : partition.ranges[0].end,
5526 : min_image_lsn,
5527 0 : force_image_creation_lsn.unwrap(),
5528 : max_deltas
5529 : );
5530 0 : return true;
5531 57 : }
5532 :
5533 57 : debug!(
5534 : max_deltas,
5535 0 : "none of the partitioned ranges had >= {threshold} deltas"
5536 : );
5537 57 : false
5538 57 : }
5539 :
5540 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5541 : /// so that at most one image layer will be produced from this function.
5542 : #[allow(clippy::too_many_arguments)]
5543 124 : async fn create_image_layer_for_rel_blocks(
5544 124 : self: &Arc<Self>,
5545 124 : partition: &KeySpace,
5546 124 : mut image_layer_writer: ImageLayerWriter,
5547 124 : lsn: Lsn,
5548 124 : ctx: &RequestContext,
5549 124 : img_range: Range<Key>,
5550 124 : io_concurrency: IoConcurrency,
5551 124 : progress: Option<(usize, usize)>,
5552 124 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5553 124 : let mut wrote_keys = false;
5554 :
5555 124 : let mut key_request_accum = KeySpaceAccum::new();
5556 824 : for range in &partition.ranges {
5557 700 : let mut key = range.start;
5558 1517 : while key < range.end {
5559 : // Decide whether to retain this key: usually we do, but sharded tenants may
5560 : // need to drop keys that don't belong to them. If we retain the key, add it
5561 : // to `key_request_accum` for later issuing a vectored get
5562 817 : if self.shard_identity.is_key_disposable(&key) {
5563 0 : debug!(
5564 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5565 : key,
5566 0 : self.shard_identity.get_shard_number(&key)
5567 : );
5568 817 : } else {
5569 817 : key_request_accum.add_key(key);
5570 817 : }
5571 :
5572 817 : let last_key_in_range = key.next() == range.end;
5573 817 : key = key.next();
5574 :
5575 : // Maybe flush `key_rest_accum`
5576 817 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5577 817 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5578 : {
5579 700 : let query =
5580 700 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5581 :
5582 700 : let results = self
5583 700 : .get_vectored(query, io_concurrency.clone(), ctx)
5584 700 : .await?;
5585 :
5586 700 : if self.cancel.is_cancelled() {
5587 0 : return Err(CreateImageLayersError::Cancelled);
5588 700 : }
5589 :
5590 1517 : for (img_key, img) in results {
5591 817 : let img = match img {
5592 817 : Ok(img) => img,
5593 0 : Err(err) => {
5594 : // If we fail to reconstruct a VM or FSM page, we can zero the
5595 : // page without losing any actual user data. That seems better
5596 : // than failing repeatedly and getting stuck.
5597 : //
5598 : // We had a bug at one point, where we truncated the FSM and VM
5599 : // in the pageserver, but the Postgres didn't know about that
5600 : // and continued to generate incremental WAL records for pages
5601 : // that didn't exist in the pageserver. Trying to replay those
5602 : // WAL records failed to find the previous image of the page.
5603 : // This special case allows us to recover from that situation.
5604 : // See https://github.com/neondatabase/neon/issues/2601.
5605 : //
5606 : // Unfortunately we cannot do this for the main fork, or for
5607 : // any metadata keys, keys, as that would lead to actual data
5608 : // loss.
5609 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5610 0 : warn!(
5611 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5612 : );
5613 0 : ZERO_PAGE.clone()
5614 : } else {
5615 0 : return Err(CreateImageLayersError::from(err));
5616 : }
5617 : }
5618 : };
5619 :
5620 : // Write all the keys we just read into our new image layer.
5621 817 : image_layer_writer.put_image(img_key, img, ctx).await?;
5622 817 : wrote_keys = true;
5623 : }
5624 117 : }
5625 : }
5626 : }
5627 :
5628 124 : let progress_report = progress
5629 124 : .map(|(idx, total)| format!("({idx}/{total}) "))
5630 124 : .unwrap_or_default();
5631 124 : if wrote_keys {
5632 : // Normal path: we have written some data into the new image layer for this
5633 : // partition, so flush it to disk.
5634 124 : info!(
5635 0 : "{} produced image layer for rel {}",
5636 : progress_report,
5637 0 : ImageLayerName {
5638 0 : key_range: img_range.clone(),
5639 0 : lsn
5640 0 : },
5641 : );
5642 124 : Ok(ImageLayerCreationOutcome::Generated {
5643 124 : unfinished_image_layer: image_layer_writer,
5644 124 : })
5645 : } else {
5646 0 : tracing::debug!(
5647 0 : "{} no data in range {}-{}",
5648 : progress_report,
5649 : img_range.start,
5650 : img_range.end
5651 : );
5652 0 : Ok(ImageLayerCreationOutcome::Empty)
5653 : }
5654 124 : }
5655 :
5656 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5657 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5658 : /// would not be too large to fit in a single image layer.
5659 : ///
5660 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5661 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5662 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5663 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5664 : #[allow(clippy::too_many_arguments)]
5665 169 : async fn create_image_layer_for_metadata_keys(
5666 169 : self: &Arc<Self>,
5667 169 : partition: &KeySpace,
5668 169 : mut image_layer_writer: ImageLayerWriter,
5669 169 : lsn: Lsn,
5670 169 : ctx: &RequestContext,
5671 169 : img_range: Range<Key>,
5672 169 : mode: ImageLayerCreationMode,
5673 169 : io_concurrency: IoConcurrency,
5674 169 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5675 : // Metadata keys image layer creation.
5676 169 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5677 169 : let begin = Instant::now();
5678 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5679 : // not contain too many keys, otherwise this takes a lot of memory.
5680 169 : let data = self
5681 169 : .get_vectored_impl(
5682 169 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5683 169 : &mut reconstruct_state,
5684 169 : ctx,
5685 169 : )
5686 169 : .await?;
5687 169 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5688 169 : let mut new_data = BTreeMap::new();
5689 169 : let mut total_kb_retrieved = 0;
5690 169 : let mut total_keys_retrieved = 0;
5691 5175 : for (k, v) in data {
5692 5006 : let v = v?;
5693 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5694 5006 : total_keys_retrieved += 1;
5695 5006 : new_data.insert(k, v);
5696 : }
5697 169 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5698 : };
5699 169 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5700 169 : let elapsed = begin.elapsed();
5701 :
5702 169 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5703 169 : info!(
5704 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",
5705 0 : elapsed.as_secs_f64()
5706 : );
5707 :
5708 169 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5709 16 : return Ok(ImageLayerCreationOutcome::Skip);
5710 153 : }
5711 153 : if self.cancel.is_cancelled() {
5712 0 : return Err(CreateImageLayersError::Cancelled);
5713 153 : }
5714 153 : let mut wrote_any_image = false;
5715 5159 : for (k, v) in data {
5716 5006 : if v.is_empty() {
5717 : // the key has been deleted, it does not need an image
5718 : // in metadata keyspace, an empty image == tombstone
5719 4 : continue;
5720 5002 : }
5721 5002 : wrote_any_image = true;
5722 :
5723 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5724 :
5725 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5726 : // on the normal data path either.
5727 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5728 : }
5729 :
5730 153 : if wrote_any_image {
5731 : // Normal path: we have written some data into the new image layer for this
5732 : // partition, so flush it to disk.
5733 6 : info!(
5734 0 : "created image layer for metadata {}",
5735 0 : ImageLayerName {
5736 0 : key_range: img_range.clone(),
5737 0 : lsn
5738 0 : }
5739 : );
5740 6 : Ok(ImageLayerCreationOutcome::Generated {
5741 6 : unfinished_image_layer: image_layer_writer,
5742 6 : })
5743 : } else {
5744 147 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5745 147 : Ok(ImageLayerCreationOutcome::Empty)
5746 : }
5747 169 : }
5748 :
5749 : /// Predicate function which indicates whether we should check if new image layers
5750 : /// are required. Since checking if new image layers are required is expensive in
5751 : /// terms of CPU, we only do it in the following cases:
5752 : /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
5753 : /// 2. If enough time has passed since the last check:
5754 : /// 1. For large tenants, we wish to perform the check more often since they
5755 : /// suffer from the lack of image layers. Note that we assume sharded tenants
5756 : /// to be large since non-zero shards do not track the logical size.
5757 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5758 191 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5759 191 : let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
5760 :
5761 191 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5762 191 : let distance = lsn
5763 191 : .checked_sub(last_checks_at)
5764 191 : .expect("Attempt to compact with LSN going backwards");
5765 191 : let min_distance =
5766 191 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5767 :
5768 191 : let distance_based_decision = distance.0 >= min_distance;
5769 :
5770 191 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5771 191 : let check_required_after = (|| {
5772 191 : if self.shard_identity.is_unsharded() {
5773 135 : if let CurrentLogicalSize::Exact(logical_size) =
5774 186 : self.current_logical_size.current_size()
5775 : {
5776 135 : if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
5777 135 : return Duration::from_secs(3600 * 48);
5778 0 : }
5779 51 : }
5780 5 : }
5781 :
5782 56 : self.get_checkpoint_timeout()
5783 : })();
5784 :
5785 191 : let time_based_decision = match *last_check_instant {
5786 29 : Some(last_check) => {
5787 29 : let elapsed = last_check.elapsed();
5788 29 : elapsed >= check_required_after
5789 : }
5790 162 : None => true,
5791 : };
5792 :
5793 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5794 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5795 : // check.
5796 191 : let decision = distance_based_decision || time_based_decision;
5797 191 : tracing::info!(
5798 0 : "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
5799 : decision,
5800 : distance_based_decision,
5801 : time_based_decision
5802 : );
5803 191 : if decision {
5804 162 : self.last_image_layer_creation_check_at.store(lsn);
5805 162 : *last_check_instant = Some(Instant::now());
5806 162 : }
5807 :
5808 191 : decision
5809 191 : }
5810 :
5811 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5812 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5813 : ///
5814 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5815 : #[allow(clippy::too_many_arguments)]
5816 191 : async fn create_image_layers(
5817 191 : self: &Arc<Timeline>,
5818 191 : partitioning: &KeyPartitioning,
5819 191 : lsn: Lsn,
5820 191 : force_image_creation_lsn: Option<Lsn>,
5821 191 : mode: ImageLayerCreationMode,
5822 191 : ctx: &RequestContext,
5823 191 : last_status: LastImageLayerCreationStatus,
5824 191 : yield_for_l0: bool,
5825 191 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5826 191 : let timer = self.metrics.create_images_time_histo.start_timer();
5827 :
5828 191 : if partitioning.parts.is_empty() {
5829 0 : warn!("no partitions to create image layers for");
5830 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5831 191 : }
5832 :
5833 : // We need to avoid holes between generated image layers.
5834 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5835 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5836 : //
5837 : // How such hole between partitions can appear?
5838 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5839 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5840 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5841 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5842 191 : let mut start = Key::MIN;
5843 :
5844 191 : let check_for_image_layers =
5845 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5846 0 : info!(
5847 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5848 : last_key
5849 : );
5850 0 : true
5851 : } else {
5852 191 : self.should_check_if_image_layers_required(lsn)
5853 : };
5854 :
5855 191 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5856 :
5857 191 : let mut all_generated = true;
5858 :
5859 191 : let mut partition_processed = 0;
5860 191 : let mut total_partitions = partitioning.parts.len();
5861 191 : let mut last_partition_processed = None;
5862 191 : let mut partition_parts = partitioning.parts.clone();
5863 :
5864 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5865 : // We need to skip the partitions that have already been processed.
5866 0 : let mut found = false;
5867 0 : for (i, partition) in partition_parts.iter().enumerate() {
5868 0 : if last_key <= partition.end().unwrap() {
5869 : // ```plain
5870 : // |------|--------|----------|------|
5871 : // ^last_key
5872 : // ^start from this partition
5873 : // ```
5874 : // Why `i+1` instead of `i`?
5875 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5876 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5877 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5878 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5879 0 : if i + 1 >= total_partitions {
5880 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5881 : // iteration of image layer creation should return a complete status.
5882 0 : break; // with found=false
5883 0 : }
5884 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5885 0 : total_partitions = partition_parts.len();
5886 : // Update the start key to the partition start.
5887 0 : start = partition_parts[0].start().unwrap();
5888 0 : found = true;
5889 0 : break;
5890 0 : }
5891 : }
5892 0 : if !found {
5893 : // Last key is within the last partition, or larger than all partitions.
5894 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5895 0 : }
5896 191 : }
5897 :
5898 191 : let total = partition_parts.len();
5899 394 : for (idx, partition) in partition_parts.iter().enumerate() {
5900 394 : if self.cancel.is_cancelled() {
5901 0 : return Err(CreateImageLayersError::Cancelled);
5902 394 : }
5903 394 : partition_processed += 1;
5904 394 : let img_range = start..partition.ranges.last().unwrap().end;
5905 394 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5906 394 : if compact_metadata {
5907 764 : for range in &partition.ranges {
5908 573 : assert!(
5909 573 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5910 573 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5911 0 : "metadata keys must be partitioned separately"
5912 : );
5913 : }
5914 191 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5915 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5916 : // might mess up with evictions.
5917 22 : start = img_range.end;
5918 22 : continue;
5919 169 : }
5920 : // For initial and force modes, we always generate image layers for metadata keys.
5921 203 : } else if let ImageLayerCreationMode::Try = mode {
5922 : // check_for_image_layers = false -> skip
5923 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5924 79 : if !check_for_image_layers
5925 57 : || !self
5926 57 : .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
5927 57 : .await
5928 : {
5929 79 : start = img_range.end;
5930 79 : continue;
5931 0 : }
5932 124 : }
5933 293 : if let ImageLayerCreationMode::Force = mode {
5934 : // When forced to create image layers, we might try and create them where they already
5935 : // exist. This mode is only used in tests/debug.
5936 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5937 14 : if layers.contains_key(&PersistentLayerKey {
5938 14 : key_range: img_range.clone(),
5939 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5940 14 : is_delta: false,
5941 14 : }) {
5942 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5943 : // in the future.
5944 0 : tracing::info!(
5945 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5946 : img_range.start,
5947 : img_range.end
5948 : );
5949 0 : start = img_range.end;
5950 0 : continue;
5951 14 : }
5952 279 : }
5953 :
5954 293 : let image_layer_writer = ImageLayerWriter::new(
5955 293 : self.conf,
5956 293 : self.timeline_id,
5957 293 : self.tenant_shard_id,
5958 293 : &img_range,
5959 293 : lsn,
5960 293 : &self.gate,
5961 293 : self.cancel.clone(),
5962 293 : ctx,
5963 293 : )
5964 293 : .await
5965 293 : .map_err(CreateImageLayersError::Other)?;
5966 :
5967 293 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5968 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5969 0 : "failpoint image-layer-writer-fail-before-finish"
5970 0 : )))
5971 0 : });
5972 :
5973 293 : let io_concurrency = IoConcurrency::spawn_from_conf(
5974 293 : self.conf.get_vectored_concurrent_io,
5975 293 : self.gate
5976 293 : .enter()
5977 293 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5978 : );
5979 :
5980 293 : let outcome = if !compact_metadata {
5981 124 : self.create_image_layer_for_rel_blocks(
5982 124 : partition,
5983 124 : image_layer_writer,
5984 124 : lsn,
5985 124 : ctx,
5986 124 : img_range.clone(),
5987 124 : io_concurrency,
5988 124 : Some((idx, total)),
5989 124 : )
5990 124 : .await?
5991 : } else {
5992 169 : self.create_image_layer_for_metadata_keys(
5993 169 : partition,
5994 169 : image_layer_writer,
5995 169 : lsn,
5996 169 : ctx,
5997 169 : img_range.clone(),
5998 169 : mode,
5999 169 : io_concurrency,
6000 169 : )
6001 169 : .await?
6002 : };
6003 293 : match outcome {
6004 147 : ImageLayerCreationOutcome::Empty => {
6005 147 : // No data in this partition, so we don't need to create an image layer (for now).
6006 147 : // The next image layer should cover this key range, so we don't advance the `start`
6007 147 : // key.
6008 147 : }
6009 : ImageLayerCreationOutcome::Generated {
6010 130 : unfinished_image_layer,
6011 130 : } => {
6012 130 : batch_image_writer.add_unfinished_image_writer(
6013 130 : unfinished_image_layer,
6014 130 : img_range.clone(),
6015 130 : lsn,
6016 130 : );
6017 130 : // The next image layer should be generated right after this one.
6018 130 : start = img_range.end;
6019 130 : }
6020 16 : ImageLayerCreationOutcome::Skip => {
6021 16 : // We don't need to create an image layer for this partition.
6022 16 : // The next image layer should NOT cover this range, otherwise
6023 16 : // the keyspace becomes empty (reads don't go past image layers).
6024 16 : start = img_range.end;
6025 16 : }
6026 : }
6027 :
6028 293 : if let ImageLayerCreationMode::Try = mode {
6029 : // We have at least made some progress
6030 51 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
6031 : // The `Try` mode is currently only used on the compaction path. We want to avoid
6032 : // image layer generation taking too long time and blocking L0 compaction. So in this
6033 : // mode, we also inspect the current number of L0 layers and skip image layer generation
6034 : // if there are too many of them.
6035 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
6036 0 : * self.get_compaction_threshold();
6037 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
6038 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
6039 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
6040 0 : if image_preempt_threshold != 0 {
6041 0 : let should_yield = self
6042 0 : .l0_compaction_trigger
6043 0 : .notified()
6044 0 : .now_or_never()
6045 0 : .is_some();
6046 0 : if should_yield {
6047 0 : tracing::info!(
6048 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
6049 0 : partition.start().unwrap(),
6050 0 : partition.end().unwrap()
6051 : );
6052 0 : last_partition_processed = Some(partition.clone());
6053 0 : all_generated = false;
6054 0 : break;
6055 0 : }
6056 0 : }
6057 51 : }
6058 242 : }
6059 : }
6060 :
6061 191 : let image_layers = batch_image_writer
6062 191 : .finish(self, ctx)
6063 191 : .await
6064 191 : .map_err(CreateImageLayersError::Other)?;
6065 :
6066 191 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6067 :
6068 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
6069 : // now they are being scheduled outside of write lock; current way is inconsistent with
6070 : // compaction lock order.
6071 191 : guard
6072 191 : .open_mut()?
6073 191 : .track_new_image_layers(&image_layers, &self.metrics);
6074 191 : drop_layer_manager_wlock(guard);
6075 191 : let duration = timer.stop_and_record();
6076 :
6077 : // Creating image layers may have caused some previously visible layers to be covered
6078 191 : if !image_layers.is_empty() {
6079 118 : self.update_layer_visibility().await?;
6080 73 : }
6081 :
6082 191 : let total_layer_size = image_layers
6083 191 : .iter()
6084 191 : .map(|l| l.metadata().file_size)
6085 191 : .sum::<u64>();
6086 :
6087 191 : if !image_layers.is_empty() {
6088 118 : info!(
6089 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
6090 0 : image_layers.len(),
6091 : total_layer_size,
6092 0 : duration.as_secs_f64(),
6093 : partition_processed,
6094 : total_partitions
6095 : );
6096 73 : }
6097 :
6098 : Ok((
6099 191 : image_layers,
6100 191 : if all_generated {
6101 191 : LastImageLayerCreationStatus::Complete
6102 : } else {
6103 : LastImageLayerCreationStatus::Incomplete {
6104 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
6105 0 : last_partition_processed.end().unwrap_or(Key::MIN)
6106 : } else {
6107 : // This branch should be unreachable, but in case it happens, we can just return the start key.
6108 0 : Key::MIN
6109 : },
6110 : }
6111 : },
6112 : ))
6113 191 : }
6114 :
6115 : /// Wait until the background initial logical size calculation is complete, or
6116 : /// this Timeline is shut down. Calling this function will cause the initial
6117 : /// logical size calculation to skip waiting for the background jobs barrier.
6118 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
6119 0 : if !self.shard_identity.is_shard_zero() {
6120 : // We don't populate logical size on shard >0: skip waiting for it.
6121 0 : return;
6122 0 : }
6123 :
6124 0 : if self.remote_client.is_deleting() {
6125 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
6126 0 : return;
6127 0 : }
6128 :
6129 0 : if self.current_logical_size.current_size().is_exact() {
6130 : // root timelines are initialized with exact count, but never start the background
6131 : // calculation
6132 0 : return;
6133 0 : }
6134 :
6135 0 : if self.cancel.is_cancelled() {
6136 : // We already requested stopping the tenant, so we cannot wait for the logical size
6137 : // calculation to complete given the task might have been already cancelled.
6138 0 : return;
6139 0 : }
6140 :
6141 0 : if let Some(await_bg_cancel) = self
6142 0 : .current_logical_size
6143 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
6144 0 : .get()
6145 0 : {
6146 0 : await_bg_cancel.cancel();
6147 0 : } else {
6148 : // We should not wait if we were not able to explicitly instruct
6149 : // the logical size cancellation to skip the concurrency limit semaphore.
6150 : // TODO: this is an unexpected case. We should restructure so that it
6151 : // can't happen.
6152 0 : tracing::warn!(
6153 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
6154 : );
6155 0 : debug_assert!(false);
6156 : }
6157 :
6158 0 : tokio::select!(
6159 0 : _ = self.current_logical_size.initialized.acquire() => {},
6160 0 : _ = self.cancel.cancelled() => {}
6161 : )
6162 0 : }
6163 :
6164 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
6165 : /// Timelines layers up to the ancestor_lsn.
6166 : ///
6167 : /// Requires a timeline that:
6168 : /// - has an ancestor to detach from
6169 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
6170 : /// a technical requirement
6171 : ///
6172 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
6173 : /// polled again until completion.
6174 : ///
6175 : /// During the operation all timelines sharing the data with this timeline will be reparented
6176 : /// from our ancestor to be branches of this timeline.
6177 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
6178 0 : self: &Arc<Timeline>,
6179 0 : tenant: &crate::tenant::TenantShard,
6180 0 : options: detach_ancestor::Options,
6181 0 : behavior: DetachBehavior,
6182 0 : ctx: &RequestContext,
6183 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
6184 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
6185 0 : }
6186 :
6187 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
6188 : /// reparents any reparentable children of previous ancestor.
6189 : ///
6190 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6191 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6192 : /// successfully, tenant must be reloaded.
6193 : ///
6194 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6195 : /// resetting the tenant.
6196 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6197 0 : self: &Arc<Timeline>,
6198 0 : tenant: &crate::tenant::TenantShard,
6199 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6200 0 : ancestor_timeline_id: TimelineId,
6201 0 : ancestor_lsn: Lsn,
6202 0 : behavior: DetachBehavior,
6203 0 : ctx: &RequestContext,
6204 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6205 0 : detach_ancestor::detach_and_reparent(
6206 0 : self,
6207 0 : tenant,
6208 0 : prepared,
6209 0 : ancestor_timeline_id,
6210 0 : ancestor_lsn,
6211 0 : behavior,
6212 0 : ctx,
6213 0 : )
6214 0 : .await
6215 0 : }
6216 :
6217 : /// Final step which unblocks the GC.
6218 : ///
6219 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6220 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6221 0 : self: &Arc<Timeline>,
6222 0 : tenant: &crate::tenant::TenantShard,
6223 0 : attempt: detach_ancestor::Attempt,
6224 0 : ctx: &RequestContext,
6225 0 : ) -> Result<(), detach_ancestor::Error> {
6226 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6227 0 : }
6228 : }
6229 :
6230 : impl Drop for Timeline {
6231 5 : fn drop(&mut self) {
6232 5 : if let Some(ancestor) = &self.ancestor_timeline {
6233 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6234 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6235 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6236 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6237 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6238 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6239 2 : }
6240 0 : }
6241 3 : }
6242 5 : info!(
6243 0 : "Timeline {} for tenant {} is being dropped",
6244 : self.timeline_id, self.tenant_shard_id.tenant_id
6245 : );
6246 5 : }
6247 : }
6248 :
6249 : pub(crate) use compaction_error::CompactionError;
6250 : /// In a private mod to enforce that [`CompactionError::is_cancel`] is used
6251 : /// instead of `match`ing on [`CompactionError::ShuttingDown`].
6252 : mod compaction_error {
6253 : use utils::sync::gate::GateError;
6254 :
6255 : use crate::{
6256 : pgdatadir_mapping::CollectKeySpaceError,
6257 : tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
6258 : virtual_file::owned_buffers_io::write::FlushTaskError,
6259 : };
6260 :
6261 : /// Top-level failure to compact. Use [`Self::is_cancel`].
6262 : #[derive(Debug, thiserror::Error)]
6263 : pub(crate) enum CompactionError {
6264 : /// Use [`Self::is_cancel`] instead of checking for this variant.
6265 : #[error("The timeline or pageserver is shutting down")]
6266 : #[allow(private_interfaces)]
6267 : ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
6268 : #[error(transparent)]
6269 : Other(anyhow::Error),
6270 : }
6271 :
6272 : #[derive(Debug)]
6273 : struct ForbidMatching;
6274 :
6275 : impl CompactionError {
6276 0 : pub fn new_cancelled() -> Self {
6277 0 : Self::ShuttingDown(ForbidMatching)
6278 0 : }
6279 : /// Errors that can be ignored, i.e., cancel and shutdown.
6280 0 : pub fn is_cancel(&self) -> bool {
6281 0 : let other = match self {
6282 0 : CompactionError::ShuttingDown(_) => return true,
6283 0 : CompactionError::Other(other) => other,
6284 : };
6285 :
6286 : // The write path of compaction in particular often lacks differentiated
6287 : // handling errors stemming from cancellation from other errors.
6288 : // So, if requested, we also check the ::Other variant by downcasting.
6289 : // The list below has been found empirically from flaky tests and production logs.
6290 : // The process is simple: on ::Other(), compaction will print the enclosed
6291 : // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
6292 : // line where the write path / compaction code does undifferentiated error handling
6293 : // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
6294 : // below, following the same is_cancel() pattern.
6295 :
6296 0 : let root_cause = other.root_cause();
6297 :
6298 0 : let upload_queue = root_cause
6299 0 : .downcast_ref::<NotInitialized>()
6300 0 : .is_some_and(|e| e.is_stopping());
6301 0 : let timeline = root_cause
6302 0 : .downcast_ref::<PageReconstructError>()
6303 0 : .is_some_and(|e| e.is_cancel());
6304 0 : let buffered_writer_flush_task_canelled = root_cause
6305 0 : .downcast_ref::<FlushTaskError>()
6306 0 : .is_some_and(|e| e.is_cancel());
6307 0 : let write_blob_cancelled = root_cause
6308 0 : .downcast_ref::<WriteBlobError>()
6309 0 : .is_some_and(|e| e.is_cancel());
6310 0 : let gate_closed = root_cause
6311 0 : .downcast_ref::<GateError>()
6312 0 : .is_some_and(|e| e.is_cancel());
6313 0 : upload_queue
6314 0 : || timeline
6315 0 : || buffered_writer_flush_task_canelled
6316 0 : || write_blob_cancelled
6317 0 : || gate_closed
6318 0 : }
6319 0 : pub fn into_anyhow(self) -> anyhow::Error {
6320 0 : match self {
6321 0 : CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
6322 0 : CompactionError::Other(e) => e,
6323 : }
6324 0 : }
6325 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6326 0 : if err.is_cancel() {
6327 0 : Self::new_cancelled()
6328 : } else {
6329 0 : Self::Other(err.into_anyhow())
6330 : }
6331 0 : }
6332 : }
6333 : }
6334 :
6335 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6336 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6337 0 : match value {
6338 : super::upload_queue::NotInitialized::Uninitialized => {
6339 0 : CompactionError::Other(anyhow::anyhow!(value))
6340 : }
6341 : super::upload_queue::NotInitialized::ShuttingDown
6342 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
6343 : }
6344 0 : }
6345 : }
6346 :
6347 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6348 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6349 0 : match e {
6350 : super::storage_layer::layer::DownloadError::TimelineShutdown
6351 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6352 0 : CompactionError::new_cancelled()
6353 : }
6354 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6355 : | super::storage_layer::layer::DownloadError::DownloadRequired
6356 : | super::storage_layer::layer::DownloadError::NotFile(_)
6357 : | super::storage_layer::layer::DownloadError::DownloadFailed
6358 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6359 0 : CompactionError::Other(anyhow::anyhow!(e))
6360 : }
6361 : #[cfg(test)]
6362 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6363 0 : CompactionError::Other(anyhow::anyhow!(e))
6364 : }
6365 : }
6366 0 : }
6367 : }
6368 :
6369 : impl From<layer_manager::Shutdown> for CompactionError {
6370 0 : fn from(_: layer_manager::Shutdown) -> Self {
6371 0 : CompactionError::new_cancelled()
6372 0 : }
6373 : }
6374 :
6375 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6376 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6377 0 : if e.is_cancel() {
6378 0 : CompactionError::new_cancelled()
6379 : } else {
6380 0 : CompactionError::Other(e.into_anyhow())
6381 : }
6382 0 : }
6383 : }
6384 :
6385 : #[serde_as]
6386 : #[derive(serde::Serialize)]
6387 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6388 :
6389 : #[derive(Default)]
6390 : enum DurationRecorder {
6391 : #[default]
6392 : NotStarted,
6393 : Recorded(RecordedDuration, tokio::time::Instant),
6394 : }
6395 :
6396 : impl DurationRecorder {
6397 115 : fn till_now(&self) -> DurationRecorder {
6398 115 : match self {
6399 : DurationRecorder::NotStarted => {
6400 0 : panic!("must only call on recorded measurements")
6401 : }
6402 115 : DurationRecorder::Recorded(_, ended) => {
6403 115 : let now = tokio::time::Instant::now();
6404 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6405 : }
6406 : }
6407 115 : }
6408 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6409 138 : match self {
6410 0 : DurationRecorder::NotStarted => None,
6411 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6412 : }
6413 138 : }
6414 : }
6415 :
6416 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6417 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6418 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6419 : /// keys specified in the `data` field.
6420 : #[cfg(test)]
6421 : #[derive(Clone)]
6422 : pub struct DeltaLayerTestDesc {
6423 : pub lsn_range: Range<Lsn>,
6424 : pub key_range: Range<Key>,
6425 : pub data: Vec<(Key, Lsn, Value)>,
6426 : }
6427 :
6428 : #[cfg(test)]
6429 : #[derive(Clone)]
6430 : pub struct InMemoryLayerTestDesc {
6431 : pub lsn_range: Range<Lsn>,
6432 : pub data: Vec<(Key, Lsn, Value)>,
6433 : pub is_open: bool,
6434 : }
6435 :
6436 : #[cfg(test)]
6437 : impl DeltaLayerTestDesc {
6438 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6439 2 : Self {
6440 2 : lsn_range,
6441 2 : key_range,
6442 2 : data,
6443 2 : }
6444 2 : }
6445 :
6446 45 : pub fn new_with_inferred_key_range(
6447 45 : lsn_range: Range<Lsn>,
6448 45 : data: Vec<(Key, Lsn, Value)>,
6449 45 : ) -> Self {
6450 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6451 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6452 45 : Self {
6453 45 : key_range: (*key_min)..(key_max.next()),
6454 45 : lsn_range,
6455 45 : data,
6456 45 : }
6457 45 : }
6458 :
6459 5 : pub(crate) fn layer_name(&self) -> LayerName {
6460 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6461 5 : key_range: self.key_range.clone(),
6462 5 : lsn_range: self.lsn_range.clone(),
6463 5 : })
6464 5 : }
6465 : }
6466 :
6467 : impl Timeline {
6468 23 : async fn finish_compact_batch(
6469 23 : self: &Arc<Self>,
6470 23 : new_deltas: &[ResidentLayer],
6471 23 : new_images: &[ResidentLayer],
6472 23 : layers_to_remove: &[Layer],
6473 23 : ) -> Result<(), CompactionError> {
6474 23 : let mut guard = tokio::select! {
6475 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6476 23 : _ = self.cancel.cancelled() => {
6477 0 : return Err(CompactionError::new_cancelled());
6478 : }
6479 : };
6480 :
6481 23 : let mut duplicated_layers = HashSet::new();
6482 :
6483 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6484 :
6485 186 : for l in new_deltas {
6486 163 : if guard.contains(l.as_ref()) {
6487 : // expected in tests
6488 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6489 :
6490 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6491 : // `guard` on self.layers. as of writing this, there are no error returns except
6492 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6493 : // because we have not implemented L0 => L0 compaction.
6494 0 : duplicated_layers.insert(l.layer_desc().key());
6495 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6496 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6497 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6498 0 : )));
6499 163 : } else {
6500 163 : insert_layers.push(l.clone());
6501 163 : }
6502 : }
6503 :
6504 : // only remove those inputs which were not outputs
6505 23 : let remove_layers: Vec<Layer> = layers_to_remove
6506 23 : .iter()
6507 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6508 23 : .cloned()
6509 23 : .collect();
6510 :
6511 23 : if !new_images.is_empty() {
6512 0 : guard
6513 0 : .open_mut()?
6514 0 : .track_new_image_layers(new_images, &self.metrics);
6515 23 : }
6516 :
6517 23 : guard
6518 23 : .open_mut()?
6519 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6520 :
6521 23 : self.remote_client
6522 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6523 :
6524 23 : drop_layer_manager_wlock(guard);
6525 :
6526 23 : Ok(())
6527 23 : }
6528 :
6529 0 : async fn rewrite_layers(
6530 0 : self: &Arc<Self>,
6531 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6532 0 : mut drop_layers: Vec<Layer>,
6533 0 : ) -> Result<(), CompactionError> {
6534 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6535 :
6536 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6537 : // to avoid double-removing, and avoid rewriting something that was removed.
6538 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6539 0 : drop_layers.retain(|l| guard.contains(l));
6540 :
6541 0 : guard
6542 0 : .open_mut()?
6543 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6544 :
6545 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6546 :
6547 0 : self.remote_client
6548 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6549 :
6550 0 : Ok(())
6551 0 : }
6552 :
6553 : /// Schedules the uploads of the given image layers
6554 80 : fn upload_new_image_layers(
6555 80 : self: &Arc<Self>,
6556 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6557 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6558 93 : for layer in new_images {
6559 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6560 : }
6561 : // should any new image layer been created, not uploading index_part will
6562 : // result in a mismatch between remote_physical_size and layermap calculated
6563 : // size, which will fail some tests, but should not be an issue otherwise.
6564 80 : self.remote_client
6565 80 : .schedule_index_upload_for_file_changes()?;
6566 80 : Ok(())
6567 80 : }
6568 :
6569 0 : async fn find_gc_time_cutoff(
6570 0 : &self,
6571 0 : now: SystemTime,
6572 0 : pitr: Duration,
6573 0 : cancel: &CancellationToken,
6574 0 : ctx: &RequestContext,
6575 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6576 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6577 0 : if self.shard_identity.is_shard_zero() {
6578 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6579 0 : let time_range = if pitr == Duration::ZERO {
6580 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6581 : } else {
6582 0 : pitr
6583 : };
6584 :
6585 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6586 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6587 0 : let timestamp = to_pg_timestamp(time_cutoff);
6588 :
6589 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6590 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6591 0 : LsnForTimestamp::Future(lsn) => {
6592 : // The timestamp is in the future. That sounds impossible,
6593 : // but what it really means is that there hasn't been
6594 : // any commits since the cutoff timestamp.
6595 : //
6596 : // In this case we should use the LSN of the most recent commit,
6597 : // which is implicitly the last LSN in the log.
6598 0 : debug!("future({})", lsn);
6599 0 : Some(self.get_last_record_lsn())
6600 : }
6601 0 : LsnForTimestamp::Past(lsn) => {
6602 0 : debug!("past({})", lsn);
6603 0 : None
6604 : }
6605 0 : LsnForTimestamp::NoData(lsn) => {
6606 0 : debug!("nodata({})", lsn);
6607 0 : None
6608 : }
6609 : };
6610 0 : Ok(time_cutoff)
6611 : } else {
6612 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6613 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6614 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6615 : // space cutoff that we would also have respected ourselves.
6616 0 : match self
6617 0 : .remote_client
6618 0 : .download_foreign_index(ShardNumber(0), cancel)
6619 0 : .await
6620 : {
6621 0 : Ok((index_part, index_generation, _index_mtime)) => {
6622 0 : tracing::info!(
6623 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6624 0 : index_part.metadata.latest_gc_cutoff_lsn()
6625 : );
6626 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6627 : }
6628 : Err(DownloadError::NotFound) => {
6629 : // This is unexpected, because during timeline creations shard zero persists to remote
6630 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6631 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6632 : // state, then shard zero should _eventually_ write an index when it GCs.
6633 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6634 0 : Ok(None)
6635 : }
6636 0 : Err(e) => {
6637 : // TODO: this function should return a different error type than page reconstruct error
6638 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6639 : }
6640 : }
6641 :
6642 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6643 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6644 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6645 : // new location. This is legal in principle, but problematic in practice because it might result
6646 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6647 : // because they have GC'd past the branch point.
6648 : }
6649 0 : }
6650 :
6651 : /// Find the Lsns above which layer files need to be retained on
6652 : /// garbage collection.
6653 : ///
6654 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6655 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6656 : /// the space-based retention.
6657 : ///
6658 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6659 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6660 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6661 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6662 : /// in the response as the GC cutoff point for the timeline.
6663 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6664 : pub(super) async fn find_gc_cutoffs(
6665 : &self,
6666 : now: SystemTime,
6667 : space_cutoff: Lsn,
6668 : pitr: Duration,
6669 : cancel: &CancellationToken,
6670 : ctx: &RequestContext,
6671 : ) -> Result<GcCutoffs, PageReconstructError> {
6672 : let _timer = self
6673 : .metrics
6674 : .find_gc_cutoffs_histo
6675 : .start_timer()
6676 : .record_on_drop();
6677 :
6678 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6679 :
6680 : if cfg!(test) && pitr == Duration::ZERO {
6681 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6682 : return Ok(GcCutoffs {
6683 : time: Some(self.get_last_record_lsn()),
6684 : space: space_cutoff,
6685 : });
6686 : }
6687 :
6688 : // Calculate a time-based limit on how much to retain:
6689 : // - if PITR interval is set, then this is our cutoff.
6690 : // - if PITR interval is not set, then we do a lookup
6691 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6692 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6693 :
6694 : Ok(match (pitr, time_cutoff) {
6695 : (Duration::ZERO, Some(time_cutoff)) => {
6696 : // PITR is not set. Retain the size-based limit, or the default time retention,
6697 : // whichever requires less data.
6698 : GcCutoffs {
6699 : time: Some(self.get_last_record_lsn()),
6700 : space: std::cmp::max(time_cutoff, space_cutoff),
6701 : }
6702 : }
6703 : (Duration::ZERO, None) => {
6704 : // PITR is not set, and time lookup failed
6705 : GcCutoffs {
6706 : time: Some(self.get_last_record_lsn()),
6707 : space: space_cutoff,
6708 : }
6709 : }
6710 : (_, None) => {
6711 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6712 : // cannot advance beyond what was already GC'd, and respect space-based retention
6713 : GcCutoffs {
6714 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6715 : space: space_cutoff,
6716 : }
6717 : }
6718 : (_, Some(time_cutoff)) => {
6719 : // PITR interval is set and we looked up timestamp successfully. Ignore
6720 : // size based retention and make time cutoff authoritative
6721 : GcCutoffs {
6722 : time: Some(time_cutoff),
6723 : space: time_cutoff,
6724 : }
6725 : }
6726 : })
6727 : }
6728 :
6729 : /// Garbage collect layer files on a timeline that are no longer needed.
6730 : ///
6731 : /// Currently, we don't make any attempt at removing unneeded page versions
6732 : /// within a layer file. We can only remove the whole file if it's fully
6733 : /// obsolete.
6734 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6735 : // this is most likely the background tasks, but it might be the spawned task from
6736 : // immediate_gc
6737 377 : let _g = tokio::select! {
6738 377 : guard = self.gc_lock.lock() => guard,
6739 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6740 : };
6741 376 : let timer = self.metrics.garbage_collect_histo.start_timer();
6742 :
6743 376 : fail_point!("before-timeline-gc");
6744 :
6745 : // Is the timeline being deleted?
6746 376 : if self.is_stopping() {
6747 0 : return Err(GcError::TimelineCancelled);
6748 376 : }
6749 :
6750 376 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6751 376 : let gc_info = self.gc_info.read().unwrap();
6752 :
6753 376 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6754 376 : let time_cutoff = gc_info.cutoffs.time;
6755 376 : let retain_lsns = gc_info
6756 376 : .retain_lsns
6757 376 : .iter()
6758 376 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6759 376 : .collect();
6760 :
6761 : // Gets the maximum LSN that holds the valid lease.
6762 : //
6763 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6764 : // Here, we do not check for stale leases again.
6765 376 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6766 :
6767 376 : (
6768 376 : space_cutoff,
6769 376 : time_cutoff,
6770 376 : retain_lsns,
6771 376 : max_lsn_with_valid_lease,
6772 376 : )
6773 : };
6774 :
6775 376 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6776 376 : let standby_horizon = self.standby_horizon.load();
6777 : // Hold GC for the standby, but as a safety guard do it only within some
6778 : // reasonable lag.
6779 376 : if standby_horizon != Lsn::INVALID {
6780 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6781 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6782 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6783 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6784 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6785 : } else {
6786 0 : warn!(
6787 0 : "standby is lagging for more than {}MB, not holding gc for it",
6788 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6789 : )
6790 : }
6791 0 : }
6792 376 : }
6793 :
6794 : // Reset standby horizon to ignore it if it is not updated till next GC.
6795 : // It is an easy way to unset it when standby disappears without adding
6796 : // more conf options.
6797 376 : self.standby_horizon.store(Lsn::INVALID);
6798 376 : self.metrics
6799 376 : .standby_horizon_gauge
6800 376 : .set(Lsn::INVALID.0 as i64);
6801 :
6802 376 : let res = self
6803 376 : .gc_timeline(
6804 376 : space_cutoff,
6805 376 : time_cutoff,
6806 376 : retain_lsns,
6807 376 : max_lsn_with_valid_lease,
6808 376 : new_gc_cutoff,
6809 : )
6810 376 : .instrument(
6811 376 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6812 : )
6813 376 : .await?;
6814 :
6815 : // only record successes
6816 376 : timer.stop_and_record();
6817 :
6818 376 : Ok(res)
6819 377 : }
6820 :
6821 376 : async fn gc_timeline(
6822 376 : &self,
6823 376 : space_cutoff: Lsn,
6824 376 : time_cutoff: Option<Lsn>, // None if uninitialized
6825 376 : retain_lsns: Vec<Lsn>,
6826 376 : max_lsn_with_valid_lease: Option<Lsn>,
6827 376 : new_gc_cutoff: Lsn,
6828 376 : ) -> Result<GcResult, GcError> {
6829 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6830 :
6831 376 : let now = SystemTime::now();
6832 376 : let mut result: GcResult = GcResult::default();
6833 :
6834 : // Nothing to GC. Return early.
6835 376 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6836 376 : if latest_gc_cutoff >= new_gc_cutoff {
6837 11 : info!(
6838 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6839 : );
6840 11 : return Ok(result);
6841 365 : }
6842 :
6843 365 : let Some(time_cutoff) = time_cutoff else {
6844 : // The GC cutoff should have been computed by now, but let's be defensive.
6845 0 : info!("Nothing to GC: time_cutoff not yet computed");
6846 0 : return Ok(result);
6847 : };
6848 :
6849 : // We need to ensure that no one tries to read page versions or create
6850 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6851 : // for details. This will block until the old value is no longer in use.
6852 : //
6853 : // The GC cutoff should only ever move forwards.
6854 365 : let waitlist = {
6855 365 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6856 365 : if *write_guard > new_gc_cutoff {
6857 0 : return Err(GcError::BadLsn {
6858 0 : why: format!(
6859 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6860 0 : *write_guard, new_gc_cutoff
6861 0 : ),
6862 0 : });
6863 365 : }
6864 :
6865 365 : write_guard.store_and_unlock(new_gc_cutoff)
6866 : };
6867 365 : waitlist.wait().await;
6868 :
6869 365 : info!("GC starting");
6870 :
6871 365 : debug!("retain_lsns: {:?}", retain_lsns);
6872 :
6873 365 : let max_retain_lsn = retain_lsns.iter().max();
6874 :
6875 : // Scan all layers in the timeline (remote or on-disk).
6876 : //
6877 : // Garbage collect the layer if all conditions are satisfied:
6878 : // 1. it is older than cutoff LSN;
6879 : // 2. it is older than PITR interval;
6880 : // 3. it doesn't need to be retained for 'retain_lsns';
6881 : // 4. it does not need to be kept for LSNs holding valid leases.
6882 : // 5. newer on-disk image layers cover the layer's whole key range
6883 365 : let layers_to_remove = {
6884 365 : let mut layers_to_remove = Vec::new();
6885 :
6886 365 : let guard = self
6887 365 : .layers
6888 365 : .read(LayerManagerLockHolder::GarbageCollection)
6889 365 : .await;
6890 365 : let layers = guard.layer_map()?;
6891 6206 : 'outer: for l in layers.iter_historic_layers() {
6892 6206 : result.layers_total += 1;
6893 :
6894 : // 1. Is it newer than GC horizon cutoff point?
6895 6206 : if l.get_lsn_range().end > space_cutoff {
6896 370 : debug!(
6897 0 : "keeping {} because it's newer than space_cutoff {}",
6898 0 : l.layer_name(),
6899 : space_cutoff,
6900 : );
6901 370 : result.layers_needed_by_cutoff += 1;
6902 370 : continue 'outer;
6903 5836 : }
6904 :
6905 : // 2. It is newer than PiTR cutoff point?
6906 5836 : if l.get_lsn_range().end > time_cutoff {
6907 0 : debug!(
6908 0 : "keeping {} because it's newer than time_cutoff {}",
6909 0 : l.layer_name(),
6910 : time_cutoff,
6911 : );
6912 0 : result.layers_needed_by_pitr += 1;
6913 0 : continue 'outer;
6914 5836 : }
6915 :
6916 : // 3. Is it needed by a child branch?
6917 : // NOTE With that we would keep data that
6918 : // might be referenced by child branches forever.
6919 : // We can track this in child timeline GC and delete parent layers when
6920 : // they are no longer needed. This might be complicated with long inheritance chains.
6921 5836 : if let Some(retain_lsn) = max_retain_lsn {
6922 : // start_lsn is inclusive
6923 4 : if &l.get_lsn_range().start <= retain_lsn {
6924 4 : debug!(
6925 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6926 0 : l.layer_name(),
6927 : retain_lsn,
6928 0 : l.is_incremental(),
6929 : );
6930 4 : result.layers_needed_by_branches += 1;
6931 4 : continue 'outer;
6932 0 : }
6933 5832 : }
6934 :
6935 : // 4. Is there a valid lease that requires us to keep this layer?
6936 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
6937 : // keep if layer start <= any of the lease
6938 9 : if &l.get_lsn_range().start <= lsn {
6939 7 : debug!(
6940 0 : "keeping {} because there is a valid lease preventing GC at {}",
6941 0 : l.layer_name(),
6942 : lsn,
6943 : );
6944 7 : result.layers_needed_by_leases += 1;
6945 7 : continue 'outer;
6946 2 : }
6947 5823 : }
6948 :
6949 : // 5. Is there a later on-disk layer for this relation?
6950 : //
6951 : // The end-LSN is exclusive, while disk_consistent_lsn is
6952 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6953 : // OK for a delta layer to have end LSN 101, but if the end LSN
6954 : // is 102, then it might not have been fully flushed to disk
6955 : // before crash.
6956 : //
6957 : // For example, imagine that the following layers exist:
6958 : //
6959 : // 1000 - image (A)
6960 : // 1000-2000 - delta (B)
6961 : // 2000 - image (C)
6962 : // 2000-3000 - delta (D)
6963 : // 3000 - image (E)
6964 : //
6965 : // If GC horizon is at 2500, we can remove layers A and B, but
6966 : // we cannot remove C, even though it's older than 2500, because
6967 : // the delta layer 2000-3000 depends on it.
6968 5825 : if !layers
6969 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6970 : {
6971 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
6972 5821 : result.layers_not_updated += 1;
6973 5821 : continue 'outer;
6974 4 : }
6975 :
6976 : // We didn't find any reason to keep this file, so remove it.
6977 4 : info!(
6978 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6979 0 : l.layer_name(),
6980 0 : l.is_incremental(),
6981 : );
6982 4 : layers_to_remove.push(l);
6983 : }
6984 :
6985 365 : layers_to_remove
6986 : };
6987 :
6988 365 : if !layers_to_remove.is_empty() {
6989 : // Persist the new GC cutoff value before we actually remove anything.
6990 : // This unconditionally schedules also an index_part.json update, even though, we will
6991 : // be doing one a bit later with the unlinked gc'd layers.
6992 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6993 3 : self.schedule_uploads(disk_consistent_lsn, None)
6994 3 : .map_err(|e| {
6995 0 : if self.cancel.is_cancelled() {
6996 0 : GcError::TimelineCancelled
6997 : } else {
6998 0 : GcError::Remote(e)
6999 : }
7000 0 : })?;
7001 :
7002 3 : let mut guard = self
7003 3 : .layers
7004 3 : .write(LayerManagerLockHolder::GarbageCollection)
7005 3 : .await;
7006 :
7007 3 : let gc_layers = layers_to_remove
7008 3 : .iter()
7009 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
7010 3 : .collect::<Vec<Layer>>();
7011 :
7012 3 : result.layers_removed = gc_layers.len() as u64;
7013 :
7014 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
7015 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
7016 :
7017 : #[cfg(feature = "testing")]
7018 3 : {
7019 3 : result.doomed_layers = gc_layers;
7020 3 : }
7021 362 : }
7022 :
7023 365 : info!(
7024 0 : "GC completed removing {} layers, cutoff {}",
7025 : result.layers_removed, new_gc_cutoff
7026 : );
7027 :
7028 365 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
7029 365 : Ok(result)
7030 376 : }
7031 :
7032 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
7033 364868 : pub(crate) async fn reconstruct_value(
7034 364868 : &self,
7035 364868 : key: Key,
7036 364868 : request_lsn: Lsn,
7037 364868 : mut data: ValueReconstructState,
7038 364868 : redo_attempt_type: RedoAttemptType,
7039 364868 : ) -> Result<Bytes, PageReconstructError> {
7040 : // Perform WAL redo if needed
7041 364868 : data.records.reverse();
7042 :
7043 364868 : let fire_critical_error = match redo_attempt_type {
7044 363535 : RedoAttemptType::ReadPage => true,
7045 0 : RedoAttemptType::LegacyCompaction => true,
7046 1333 : RedoAttemptType::GcCompaction => false,
7047 : };
7048 :
7049 : // If we have a page image, and no WAL, we're all set
7050 364868 : if data.records.is_empty() {
7051 338094 : if let Some((img_lsn, img)) = &data.img {
7052 338094 : trace!(
7053 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
7054 : key, img_lsn, request_lsn,
7055 : );
7056 338094 : Ok(img.clone())
7057 : } else {
7058 0 : Err(PageReconstructError::from(anyhow!(
7059 0 : "base image for {key} at {request_lsn} not found"
7060 0 : )))
7061 : }
7062 : } else {
7063 : // We need to do WAL redo.
7064 : //
7065 : // If we don't have a base image, then the oldest WAL record better initialize
7066 : // the page
7067 26774 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
7068 0 : Err(PageReconstructError::from(anyhow!(
7069 0 : "Base image for {} at {} not found, but got {} WAL records",
7070 0 : key,
7071 0 : request_lsn,
7072 0 : data.records.len()
7073 0 : )))
7074 : } else {
7075 26774 : if data.img.is_some() {
7076 13029 : trace!(
7077 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
7078 0 : data.records.len(),
7079 : key,
7080 : request_lsn
7081 : );
7082 : } else {
7083 13745 : trace!(
7084 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
7085 0 : data.records.len(),
7086 : key,
7087 : request_lsn
7088 : );
7089 : };
7090 26774 : let res = self
7091 26774 : .walredo_mgr
7092 26774 : .as_ref()
7093 26774 : .context("timeline has no walredo manager")
7094 26774 : .map_err(PageReconstructError::WalRedo)?
7095 26774 : .request_redo(
7096 26774 : key,
7097 26774 : request_lsn,
7098 26774 : data.img,
7099 26774 : data.records,
7100 26774 : self.pg_version,
7101 26774 : redo_attempt_type,
7102 : )
7103 26774 : .await;
7104 26773 : let img = match res {
7105 26773 : Ok(img) => img,
7106 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
7107 1 : Err(walredo::Error::Other(err)) => {
7108 1 : if fire_critical_error {
7109 0 : critical_timeline!(
7110 0 : self.tenant_shard_id,
7111 0 : self.timeline_id,
7112 0 : "walredo failure during page reconstruction: {err:?}"
7113 : );
7114 1 : }
7115 1 : return Err(PageReconstructError::WalRedo(
7116 1 : err.context("reconstruct a page image"),
7117 1 : ));
7118 : }
7119 : };
7120 26773 : Ok(img)
7121 : }
7122 : }
7123 364868 : }
7124 :
7125 0 : pub(crate) async fn spawn_download_all_remote_layers(
7126 0 : self: Arc<Self>,
7127 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7128 0 : ctx: &RequestContext,
7129 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
7130 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7131 :
7132 : // this is not really needed anymore; it has tests which really check the return value from
7133 : // http api. it would be better not to maintain this anymore.
7134 :
7135 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
7136 0 : if let Some(st) = &*status_guard {
7137 0 : match &st.state {
7138 : DownloadRemoteLayersTaskState::Running => {
7139 0 : return Err(st.clone());
7140 : }
7141 : DownloadRemoteLayersTaskState::ShutDown
7142 0 : | DownloadRemoteLayersTaskState::Completed => {
7143 0 : *status_guard = None;
7144 0 : }
7145 : }
7146 0 : }
7147 :
7148 0 : let self_clone = Arc::clone(&self);
7149 0 : let task_ctx = ctx.detached_child(
7150 0 : TaskKind::DownloadAllRemoteLayers,
7151 0 : DownloadBehavior::Download,
7152 : );
7153 0 : let task_id = task_mgr::spawn(
7154 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
7155 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
7156 0 : self.tenant_shard_id,
7157 0 : Some(self.timeline_id),
7158 0 : "download all remote layers task",
7159 0 : async move {
7160 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
7161 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
7162 0 : match &mut *status_guard {
7163 : None => {
7164 0 : warn!("tasks status is supposed to be Some(), since we are running");
7165 : }
7166 0 : Some(st) => {
7167 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
7168 0 : if st.task_id != exp_task_id {
7169 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
7170 0 : } else {
7171 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7172 0 : }
7173 : }
7174 : };
7175 0 : Ok(())
7176 0 : }
7177 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))
7178 : );
7179 :
7180 0 : let initial_info = DownloadRemoteLayersTaskInfo {
7181 0 : task_id: format!("{task_id}"),
7182 0 : state: DownloadRemoteLayersTaskState::Running,
7183 0 : total_layer_count: 0,
7184 0 : successful_download_count: 0,
7185 0 : failed_download_count: 0,
7186 0 : };
7187 0 : *status_guard = Some(initial_info.clone());
7188 :
7189 0 : Ok(initial_info)
7190 0 : }
7191 :
7192 0 : async fn download_all_remote_layers(
7193 0 : self: &Arc<Self>,
7194 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7195 0 : ctx: &RequestContext,
7196 0 : ) {
7197 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7198 :
7199 0 : let remaining = {
7200 0 : let guard = self
7201 0 : .layers
7202 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7203 0 : .await;
7204 0 : let Ok(lm) = guard.layer_map() else {
7205 : // technically here we could look into iterating accessible layers, but downloading
7206 : // all layers of a shutdown timeline makes no sense regardless.
7207 0 : tracing::info!("attempted to download all layers of shutdown timeline");
7208 0 : return;
7209 : };
7210 0 : lm.iter_historic_layers()
7211 0 : .map(|desc| guard.get_from_desc(&desc))
7212 0 : .collect::<Vec<_>>()
7213 : };
7214 0 : let total_layer_count = remaining.len();
7215 :
7216 : macro_rules! lock_status {
7217 : ($st:ident) => {
7218 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
7219 : let st = st
7220 : .as_mut()
7221 : .expect("this function is only called after the task has been spawned");
7222 : assert_eq!(
7223 : st.task_id,
7224 : format!(
7225 : "{}",
7226 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
7227 : )
7228 : );
7229 : let $st = st;
7230 : };
7231 : }
7232 :
7233 : {
7234 0 : lock_status!(st);
7235 0 : st.total_layer_count = total_layer_count as u64;
7236 : }
7237 :
7238 0 : let mut remaining = remaining.into_iter();
7239 0 : let mut have_remaining = true;
7240 0 : let mut js = tokio::task::JoinSet::new();
7241 :
7242 0 : let cancel = task_mgr::shutdown_token();
7243 :
7244 0 : let limit = request.max_concurrent_downloads;
7245 :
7246 : loop {
7247 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
7248 0 : let Some(next) = remaining.next() else {
7249 0 : have_remaining = false;
7250 0 : break;
7251 : };
7252 :
7253 0 : let span = tracing::info_span!("download", layer = %next);
7254 :
7255 0 : let ctx = ctx.attached_child();
7256 0 : js.spawn(
7257 0 : async move {
7258 0 : let res = next.download(&ctx).await;
7259 0 : (next, res)
7260 0 : }
7261 0 : .instrument(span),
7262 : );
7263 : }
7264 :
7265 0 : while let Some(res) = js.join_next().await {
7266 0 : match res {
7267 : Ok((_, Ok(_))) => {
7268 0 : lock_status!(st);
7269 0 : st.successful_download_count += 1;
7270 : }
7271 0 : Ok((layer, Err(e))) => {
7272 0 : tracing::error!(%layer, "download failed: {e:#}");
7273 0 : lock_status!(st);
7274 0 : st.failed_download_count += 1;
7275 : }
7276 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7277 0 : Err(je) if je.is_panic() => {
7278 0 : lock_status!(st);
7279 0 : st.failed_download_count += 1;
7280 : }
7281 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7282 : }
7283 : }
7284 :
7285 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7286 0 : break;
7287 0 : }
7288 : }
7289 :
7290 : {
7291 0 : lock_status!(st);
7292 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7293 : }
7294 0 : }
7295 :
7296 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7297 0 : &self,
7298 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7299 0 : self.download_all_remote_layers_task_info
7300 0 : .read()
7301 0 : .unwrap()
7302 0 : .clone()
7303 0 : }
7304 :
7305 : /* BEGIN_HADRON */
7306 0 : pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
7307 0 : let guard = self
7308 0 : .layers
7309 0 : .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
7310 0 : .await;
7311 0 : let layer_map = guard.layer_map()?;
7312 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
7313 :
7314 0 : Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
7315 0 : }
7316 : /* END_HADRON */
7317 : }
7318 :
7319 : impl Timeline {
7320 : /// Returns non-remote layers for eviction.
7321 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7322 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7323 0 : let mut max_layer_size: Option<u64> = None;
7324 :
7325 0 : let resident_layers = guard
7326 0 : .likely_resident_layers()
7327 0 : .map(|layer| {
7328 0 : let file_size = layer.layer_desc().file_size;
7329 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7330 :
7331 0 : let last_activity_ts = layer.latest_activity();
7332 :
7333 0 : EvictionCandidate {
7334 0 : layer: layer.to_owned().into(),
7335 0 : last_activity_ts,
7336 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7337 0 : visibility: layer.visibility(),
7338 0 : }
7339 0 : })
7340 0 : .collect();
7341 :
7342 0 : DiskUsageEvictionInfo {
7343 0 : max_layer_size,
7344 0 : resident_layers,
7345 0 : }
7346 0 : }
7347 :
7348 962 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7349 962 : ShardIndex {
7350 962 : shard_number: self.tenant_shard_id.shard_number,
7351 962 : shard_count: self.tenant_shard_id.shard_count,
7352 962 : }
7353 962 : }
7354 :
7355 : /// Persistently blocks gc for `Manual` reason.
7356 : ///
7357 : /// Returns true if no such block existed before, false otherwise.
7358 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7359 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7360 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7361 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7362 0 : }
7363 :
7364 : /// Persistently unblocks gc for `Manual` reason.
7365 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7366 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7367 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7368 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7369 0 : }
7370 :
7371 : #[cfg(test)]
7372 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7373 31 : self.last_record_lsn.advance(new_lsn);
7374 31 : }
7375 :
7376 : #[cfg(test)]
7377 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7378 2 : self.disk_consistent_lsn.store(new_value);
7379 2 : }
7380 :
7381 : /// Force create an image layer and place it into the layer map.
7382 : ///
7383 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7384 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7385 : /// placed into the layer map in one run AND be validated.
7386 : #[cfg(test)]
7387 36 : pub(super) async fn force_create_image_layer(
7388 36 : self: &Arc<Timeline>,
7389 36 : lsn: Lsn,
7390 36 : mut images: Vec<(Key, Bytes)>,
7391 36 : check_start_lsn: Option<Lsn>,
7392 36 : ctx: &RequestContext,
7393 36 : ) -> anyhow::Result<()> {
7394 36 : let last_record_lsn = self.get_last_record_lsn();
7395 36 : assert!(
7396 36 : lsn <= last_record_lsn,
7397 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7398 : );
7399 36 : if let Some(check_start_lsn) = check_start_lsn {
7400 36 : assert!(lsn >= check_start_lsn);
7401 0 : }
7402 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7403 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7404 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7405 36 : let mut image_layer_writer = ImageLayerWriter::new(
7406 36 : self.conf,
7407 36 : self.timeline_id,
7408 36 : self.tenant_shard_id,
7409 36 : &(min_key..end_key),
7410 36 : lsn,
7411 36 : &self.gate,
7412 36 : self.cancel.clone(),
7413 36 : ctx,
7414 36 : )
7415 36 : .await?;
7416 312 : for (key, img) in images {
7417 276 : image_layer_writer.put_image(key, img, ctx).await?;
7418 : }
7419 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7420 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7421 36 : info!("force created image layer {}", image_layer.local_path());
7422 : {
7423 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7424 36 : guard
7425 36 : .open_mut()
7426 36 : .unwrap()
7427 36 : .force_insert_layer(image_layer.clone());
7428 : }
7429 :
7430 : // Update remote_timeline_client state to reflect existence of this layer
7431 36 : self.remote_client
7432 36 : .schedule_layer_file_upload(image_layer)
7433 36 : .unwrap();
7434 :
7435 36 : Ok(())
7436 36 : }
7437 :
7438 : /// Force create a delta layer and place it into the layer map.
7439 : ///
7440 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7441 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7442 : /// placed into the layer map in one run AND be validated.
7443 : #[cfg(test)]
7444 50 : pub(super) async fn force_create_delta_layer(
7445 50 : self: &Arc<Timeline>,
7446 50 : mut deltas: DeltaLayerTestDesc,
7447 50 : check_start_lsn: Option<Lsn>,
7448 50 : ctx: &RequestContext,
7449 50 : ) -> anyhow::Result<()> {
7450 50 : let last_record_lsn = self.get_last_record_lsn();
7451 50 : deltas
7452 50 : .data
7453 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7454 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7455 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7456 10464 : for (_, lsn, _) in &deltas.data {
7457 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7458 : }
7459 50 : assert!(
7460 50 : deltas.lsn_range.end <= last_record_lsn,
7461 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7462 : deltas.lsn_range.end,
7463 : last_record_lsn
7464 : );
7465 50 : if let Some(check_start_lsn) = check_start_lsn {
7466 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7467 0 : }
7468 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7469 50 : self.conf,
7470 50 : self.timeline_id,
7471 50 : self.tenant_shard_id,
7472 50 : deltas.key_range.start,
7473 50 : deltas.lsn_range,
7474 50 : &self.gate,
7475 50 : self.cancel.clone(),
7476 50 : ctx,
7477 50 : )
7478 50 : .await?;
7479 10464 : for (key, lsn, val) in deltas.data {
7480 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7481 : }
7482 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7483 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7484 50 : info!("force created delta layer {}", delta_layer.local_path());
7485 : {
7486 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7487 50 : guard
7488 50 : .open_mut()
7489 50 : .unwrap()
7490 50 : .force_insert_layer(delta_layer.clone());
7491 : }
7492 :
7493 : // Update remote_timeline_client state to reflect existence of this layer
7494 50 : self.remote_client
7495 50 : .schedule_layer_file_upload(delta_layer)
7496 50 : .unwrap();
7497 :
7498 50 : Ok(())
7499 50 : }
7500 :
7501 : /// Force create an in-memory layer and place them into the layer map.
7502 : #[cfg(test)]
7503 4 : pub(super) async fn force_create_in_memory_layer(
7504 4 : self: &Arc<Timeline>,
7505 4 : mut in_memory: InMemoryLayerTestDesc,
7506 4 : check_start_lsn: Option<Lsn>,
7507 4 : ctx: &RequestContext,
7508 4 : ) -> anyhow::Result<()> {
7509 : use utils::bin_ser::BeSer;
7510 :
7511 : // Validate LSNs
7512 4 : if let Some(check_start_lsn) = check_start_lsn {
7513 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7514 0 : }
7515 :
7516 4 : let last_record_lsn = self.get_last_record_lsn();
7517 4 : let layer_end_lsn = if in_memory.is_open {
7518 1 : in_memory
7519 1 : .data
7520 1 : .iter()
7521 1 : .map(|(_key, lsn, _value)| lsn)
7522 1 : .max()
7523 1 : .cloned()
7524 : } else {
7525 3 : Some(in_memory.lsn_range.end)
7526 : };
7527 :
7528 4 : if let Some(end) = layer_end_lsn {
7529 4 : assert!(
7530 4 : end <= last_record_lsn,
7531 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7532 : );
7533 0 : }
7534 :
7535 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7536 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7537 19820 : assert!(*lsn < in_memory.lsn_range.end);
7538 19820 : });
7539 :
7540 : // Build the batch
7541 4 : in_memory
7542 4 : .data
7543 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7544 :
7545 4 : let data = in_memory
7546 4 : .data
7547 4 : .into_iter()
7548 19820 : .map(|(key, lsn, value)| {
7549 19820 : let value_size = value.serialized_size().unwrap() as usize;
7550 19820 : (key.to_compact(), lsn, value_size, value)
7551 19820 : })
7552 4 : .collect::<Vec<_>>();
7553 :
7554 4 : let batch = SerializedValueBatch::from_values(data);
7555 :
7556 : // Create the in-memory layer and write the batch into it
7557 4 : let layer = InMemoryLayer::create(
7558 4 : self.conf,
7559 4 : self.timeline_id,
7560 4 : self.tenant_shard_id,
7561 4 : in_memory.lsn_range.start,
7562 4 : &self.gate,
7563 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7564 4 : &CancellationToken::new(),
7565 4 : ctx,
7566 4 : )
7567 4 : .await
7568 4 : .unwrap();
7569 :
7570 4 : layer.put_batch(batch, ctx).await.unwrap();
7571 4 : if !in_memory.is_open {
7572 3 : layer.freeze(in_memory.lsn_range.end).await;
7573 1 : }
7574 :
7575 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7576 :
7577 : // Link the layer to the layer map
7578 : {
7579 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7580 4 : let layer_map = guard.open_mut().unwrap();
7581 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7582 : }
7583 :
7584 4 : Ok(())
7585 4 : }
7586 :
7587 : /// Return all keys at the LSN in the image layers
7588 : #[cfg(test)]
7589 3 : pub(crate) async fn inspect_image_layers(
7590 3 : self: &Arc<Timeline>,
7591 3 : lsn: Lsn,
7592 3 : ctx: &RequestContext,
7593 3 : io_concurrency: IoConcurrency,
7594 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7595 3 : let mut all_data = Vec::new();
7596 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7597 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7598 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7599 4 : let layer = guard.get_from_desc(&layer);
7600 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7601 4 : layer
7602 4 : .get_values_reconstruct_data(
7603 4 : KeySpace::single(Key::MIN..Key::MAX),
7604 4 : lsn..Lsn(lsn.0 + 1),
7605 4 : &mut reconstruct_data,
7606 4 : ctx,
7607 4 : )
7608 4 : .await?;
7609 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7610 33 : let v = v.collect_pending_ios().await?;
7611 33 : all_data.push((k, v.img.unwrap().1));
7612 : }
7613 13 : }
7614 : }
7615 3 : all_data.sort();
7616 3 : Ok(all_data)
7617 3 : }
7618 :
7619 : /// Get all historic layer descriptors in the layer map
7620 : #[cfg(test)]
7621 12 : pub(crate) async fn inspect_historic_layers(
7622 12 : self: &Arc<Timeline>,
7623 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7624 12 : let mut layers = Vec::new();
7625 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7626 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7627 57 : layers.push(layer.key());
7628 57 : }
7629 12 : Ok(layers)
7630 12 : }
7631 :
7632 : #[cfg(test)]
7633 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7634 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7635 5 : keyspace.merge(&ks);
7636 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7637 5 : }
7638 : }
7639 :
7640 : /// Tracking writes ingestion does to a particular in-memory layer.
7641 : ///
7642 : /// Cleared upon freezing a layer.
7643 : pub(crate) struct TimelineWriterState {
7644 : open_layer: Arc<InMemoryLayer>,
7645 : current_size: u64,
7646 : // Previous Lsn which passed through
7647 : prev_lsn: Option<Lsn>,
7648 : // Largest Lsn which passed through the current writer
7649 : max_lsn: Option<Lsn>,
7650 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7651 : cached_last_freeze_at: Lsn,
7652 : }
7653 :
7654 : impl TimelineWriterState {
7655 660 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7656 660 : Self {
7657 660 : open_layer,
7658 660 : current_size,
7659 660 : prev_lsn: None,
7660 660 : max_lsn: None,
7661 660 : cached_last_freeze_at: last_freeze_at,
7662 660 : }
7663 660 : }
7664 : }
7665 :
7666 : /// Various functions to mutate the timeline.
7667 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7668 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7669 : // but will cause large code changes.
7670 : pub(crate) struct TimelineWriter<'a> {
7671 : tl: &'a Timeline,
7672 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7673 : }
7674 :
7675 : impl Deref for TimelineWriter<'_> {
7676 : type Target = Timeline;
7677 :
7678 4949226 : fn deref(&self) -> &Self::Target {
7679 4949226 : self.tl
7680 4949226 : }
7681 : }
7682 :
7683 : #[derive(PartialEq)]
7684 : enum OpenLayerAction {
7685 : Roll,
7686 : Open,
7687 : None,
7688 : }
7689 :
7690 : impl TimelineWriter<'_> {
7691 2402129 : async fn handle_open_layer_action(
7692 2402129 : &mut self,
7693 2402129 : at: Lsn,
7694 2402129 : action: OpenLayerAction,
7695 2402129 : ctx: &RequestContext,
7696 2402129 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7697 2402129 : match action {
7698 : OpenLayerAction::Roll => {
7699 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7700 40 : self.roll_layer(freeze_at).await?;
7701 40 : self.open_layer(at, ctx).await?;
7702 : }
7703 620 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7704 : OpenLayerAction::None => {
7705 2401469 : assert!(self.write_guard.is_some());
7706 : }
7707 : }
7708 :
7709 2402129 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7710 2402129 : }
7711 :
7712 660 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7713 660 : let layer = self
7714 660 : .tl
7715 660 : .get_layer_for_write(at, &self.write_guard, ctx)
7716 660 : .await?;
7717 660 : let initial_size = layer.len();
7718 :
7719 660 : let last_freeze_at = self.last_freeze_at.load();
7720 660 : self.write_guard.replace(TimelineWriterState::new(
7721 660 : layer,
7722 660 : initial_size,
7723 660 : last_freeze_at,
7724 660 : ));
7725 :
7726 660 : Ok(())
7727 660 : }
7728 :
7729 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7730 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7731 :
7732 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7733 : // to propagate the backpressure up into WAL ingestion.
7734 40 : let l0_count = self
7735 40 : .tl
7736 40 : .layers
7737 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7738 40 : .await
7739 40 : .layer_map()?
7740 40 : .level0_deltas()
7741 40 : .len();
7742 40 : let wait_thresholds = [
7743 40 : self.get_l0_flush_delay_threshold(),
7744 40 : self.get_l0_flush_stall_threshold(),
7745 40 : ];
7746 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7747 :
7748 : // self.write_guard will be taken by the freezing
7749 40 : let flush_id = self
7750 40 : .tl
7751 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7752 40 : .await?;
7753 :
7754 40 : assert!(self.write_guard.is_none());
7755 :
7756 40 : if let Some(wait_threshold) = wait_threshold {
7757 0 : if l0_count >= wait_threshold {
7758 0 : debug!(
7759 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7760 : );
7761 0 : self.tl.wait_flush_completion(flush_id).await?;
7762 0 : }
7763 40 : }
7764 :
7765 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7766 0 : warn!("Flushed oversized open layer with size {}", current_size)
7767 40 : }
7768 :
7769 40 : Ok(())
7770 40 : }
7771 :
7772 2402129 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7773 2402129 : let state = &*self.write_guard;
7774 2402129 : let Some(state) = &state else {
7775 620 : return OpenLayerAction::Open;
7776 : };
7777 :
7778 : #[cfg(feature = "testing")]
7779 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7780 : // this check and assertion are not really needed because
7781 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7782 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7783 : // is no TimelineWriterState.
7784 0 : assert!(
7785 0 : state.open_layer.end_lsn.get().is_some(),
7786 0 : "our open_layer must be outdated"
7787 : );
7788 :
7789 : // this would be a memory leak waiting to happen because the in-memory layer always has
7790 : // an index
7791 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7792 2401509 : }
7793 :
7794 2401509 : if state.prev_lsn == Some(lsn) {
7795 : // Rolling mid LSN is not supported by [downstream code].
7796 : // Hence, only roll at LSN boundaries.
7797 : //
7798 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7799 3 : return OpenLayerAction::None;
7800 2401506 : }
7801 :
7802 2401506 : if state.current_size == 0 {
7803 : // Don't roll empty layers
7804 0 : return OpenLayerAction::None;
7805 2401506 : }
7806 :
7807 2401506 : if self.tl.should_roll(
7808 2401506 : state.current_size,
7809 2401506 : state.current_size + new_value_size,
7810 2401506 : self.get_checkpoint_distance(),
7811 2401506 : lsn,
7812 2401506 : state.cached_last_freeze_at,
7813 2401506 : state.open_layer.get_opened_at(),
7814 : ) {
7815 40 : OpenLayerAction::Roll
7816 : } else {
7817 2401466 : OpenLayerAction::None
7818 : }
7819 2402129 : }
7820 :
7821 : /// Put a batch of keys at the specified Lsns.
7822 2402128 : pub(crate) async fn put_batch(
7823 2402128 : &mut self,
7824 2402128 : batch: SerializedValueBatch,
7825 2402128 : ctx: &RequestContext,
7826 2402128 : ) -> anyhow::Result<()> {
7827 2402128 : if !batch.has_data() {
7828 0 : return Ok(());
7829 2402128 : }
7830 :
7831 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7832 : // We don't assert this in release builds, since key ownership policies may change over
7833 : // time. Stray keys will be removed during compaction.
7834 2402128 : if cfg!(debug_assertions) {
7835 4947529 : for metadata in &batch.metadata {
7836 2545401 : if let ValueMeta::Serialized(metadata) = metadata {
7837 2545401 : let key = Key::from_compact(metadata.key);
7838 2545401 : assert!(
7839 2545401 : self.shard_identity.is_key_local(&key)
7840 12 : || self.shard_identity.is_key_global(&key),
7841 0 : "key {key} does not belong on shard {}",
7842 0 : self.shard_identity.shard_index()
7843 : );
7844 0 : }
7845 : }
7846 0 : }
7847 :
7848 2402128 : let batch_max_lsn = batch.max_lsn;
7849 2402128 : let buf_size: u64 = batch.buffer_size() as u64;
7850 :
7851 2402128 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7852 2402128 : let layer = self
7853 2402128 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7854 2402128 : .await?;
7855 :
7856 2402128 : let res = layer.put_batch(batch, ctx).await;
7857 :
7858 2402128 : if res.is_ok() {
7859 2402128 : // Update the current size only when the entire write was ok.
7860 2402128 : // In case of failures, we may have had partial writes which
7861 2402128 : // render the size tracking out of sync. That's ok because
7862 2402128 : // the checkpoint distance should be significantly smaller
7863 2402128 : // than the S3 single shot upload limit of 5GiB.
7864 2402128 : let state = self.write_guard.as_mut().unwrap();
7865 2402128 :
7866 2402128 : state.current_size += buf_size;
7867 2402128 : state.prev_lsn = Some(batch_max_lsn);
7868 2402128 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7869 2402128 : }
7870 :
7871 2402128 : res
7872 2402128 : }
7873 :
7874 : #[cfg(test)]
7875 : /// Test helper, for tests that would like to poke individual values without composing a batch
7876 2195079 : pub(crate) async fn put(
7877 2195079 : &mut self,
7878 2195079 : key: Key,
7879 2195079 : lsn: Lsn,
7880 2195079 : value: &Value,
7881 2195079 : ctx: &RequestContext,
7882 2195079 : ) -> anyhow::Result<()> {
7883 : use utils::bin_ser::BeSer;
7884 2195079 : if !key.is_valid_key_on_write_path() {
7885 0 : bail!(
7886 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7887 : key
7888 : );
7889 2195079 : }
7890 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7891 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7892 2195079 : key.to_compact(),
7893 2195079 : lsn,
7894 2195079 : val_ser_size,
7895 2195079 : value.clone(),
7896 2195079 : )]);
7897 :
7898 2195079 : self.put_batch(batch, ctx).await
7899 2195079 : }
7900 :
7901 1 : pub(crate) async fn delete_batch(
7902 1 : &mut self,
7903 1 : batch: &[(Range<Key>, Lsn)],
7904 1 : ctx: &RequestContext,
7905 1 : ) -> anyhow::Result<()> {
7906 1 : if let Some((_, lsn)) = batch.first() {
7907 1 : let action = self.get_open_layer_action(*lsn, 0);
7908 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7909 1 : layer.put_tombstones(batch).await?;
7910 0 : }
7911 :
7912 1 : Ok(())
7913 1 : }
7914 :
7915 : /// Track the end of the latest digested WAL record.
7916 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7917 : ///
7918 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7919 : ///
7920 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7921 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7922 : /// the latest and can be read.
7923 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7924 2639560 : self.tl.finish_write(new_lsn);
7925 2639560 : }
7926 :
7927 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7928 135285 : self.tl.update_current_logical_size(delta)
7929 135285 : }
7930 : }
7931 :
7932 : // We need TimelineWriter to be send in upcoming conversion of
7933 : // Timeline::layers to tokio::sync::RwLock.
7934 : #[test]
7935 1 : fn is_send() {
7936 1 : fn _assert_send<T: Send>() {}
7937 1 : _assert_send::<TimelineWriter<'_>>();
7938 1 : }
7939 :
7940 : #[cfg(test)]
7941 : mod tests {
7942 : use std::sync::Arc;
7943 :
7944 : use pageserver_api::key::Key;
7945 : use postgres_ffi::PgMajorVersion;
7946 : use std::iter::Iterator;
7947 : use tracing::Instrument;
7948 : use utils::id::TimelineId;
7949 : use utils::lsn::Lsn;
7950 : use wal_decoder::models::value::Value;
7951 :
7952 : use super::HeatMapTimeline;
7953 : use crate::context::RequestContextBuilder;
7954 : use crate::tenant::harness::{TenantHarness, test_img};
7955 : use crate::tenant::layer_map::LayerMap;
7956 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7957 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
7958 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7959 : use crate::tenant::{PreviousHeatmap, Timeline};
7960 :
7961 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7962 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7963 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7964 25 : for (l, r) in lhs_rhs {
7965 20 : assert_eq!(l.name, r.name);
7966 20 : assert_eq!(l.metadata, r.metadata);
7967 : }
7968 5 : }
7969 :
7970 : #[tokio::test]
7971 1 : async fn test_heatmap_generation() {
7972 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7973 :
7974 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7975 1 : Lsn(0x10)..Lsn(0x20),
7976 1 : vec![(
7977 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7978 1 : Lsn(0x11),
7979 1 : Value::Image(test_img("foo")),
7980 1 : )],
7981 : );
7982 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7983 1 : Lsn(0x10)..Lsn(0x20),
7984 1 : vec![(
7985 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7986 1 : Lsn(0x11),
7987 1 : Value::Image(test_img("foo")),
7988 1 : )],
7989 : );
7990 1 : let l0_delta = DeltaLayerTestDesc::new(
7991 1 : Lsn(0x20)..Lsn(0x30),
7992 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7993 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7994 1 : vec![(
7995 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7996 1 : Lsn(0x25),
7997 1 : Value::Image(test_img("foo")),
7998 1 : )],
7999 : );
8000 1 : let delta_layers = vec![
8001 1 : covered_delta.clone(),
8002 1 : visible_delta.clone(),
8003 1 : l0_delta.clone(),
8004 : ];
8005 :
8006 1 : let image_layer = (
8007 1 : Lsn(0x40),
8008 1 : vec![(
8009 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8010 1 : test_img("bar"),
8011 1 : )],
8012 1 : );
8013 1 : let image_layers = vec![image_layer];
8014 :
8015 1 : let (tenant, ctx) = harness.load().await;
8016 1 : let timeline = tenant
8017 1 : .create_test_timeline_with_layers(
8018 1 : TimelineId::generate(),
8019 1 : Lsn(0x10),
8020 1 : PgMajorVersion::PG14,
8021 1 : &ctx,
8022 1 : Vec::new(), // in-memory layers
8023 1 : delta_layers,
8024 1 : image_layers,
8025 1 : Lsn(0x100),
8026 1 : )
8027 1 : .await
8028 1 : .unwrap();
8029 1 : let ctx = &ctx.with_scope_timeline(&timeline);
8030 :
8031 : // Layer visibility is an input to heatmap generation, so refresh it first
8032 1 : timeline.update_layer_visibility().await.unwrap();
8033 :
8034 1 : let heatmap = timeline
8035 1 : .generate_heatmap()
8036 1 : .await
8037 1 : .expect("Infallible while timeline is not shut down");
8038 :
8039 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
8040 :
8041 : // L0 should come last
8042 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
8043 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
8044 :
8045 1 : let mut last_lsn = Lsn::MAX;
8046 5 : for layer in heatmap_layers {
8047 : // Covered layer should be omitted
8048 4 : assert!(layer.name != covered_delta.layer_name());
8049 :
8050 4 : let layer_lsn = match &layer.name {
8051 2 : LayerName::Delta(d) => d.lsn_range.end,
8052 2 : LayerName::Image(i) => i.lsn,
8053 : };
8054 :
8055 : // Apart from L0s, newest Layers should come first
8056 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
8057 3 : assert!(layer_lsn <= last_lsn);
8058 3 : last_lsn = layer_lsn;
8059 1 : }
8060 : }
8061 :
8062 : // Evict all the layers and stash the old heatmap in the timeline.
8063 : // This simulates a migration to a cold secondary location.
8064 :
8065 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8066 1 : let mut all_layers = Vec::new();
8067 1 : let forever = std::time::Duration::from_secs(120);
8068 5 : for layer in guard.likely_resident_layers() {
8069 5 : all_layers.push(layer.clone());
8070 5 : layer.evict_and_wait(forever).await.unwrap();
8071 : }
8072 1 : drop(guard);
8073 :
8074 1 : timeline
8075 1 : .previous_heatmap
8076 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8077 1 : heatmap: heatmap.clone(),
8078 1 : read_at: std::time::Instant::now(),
8079 1 : end_lsn: None,
8080 1 : })));
8081 :
8082 : // Generate a new heatmap and assert that it contains the same layers as the old one.
8083 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
8084 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
8085 :
8086 : // Download each layer one by one. Generate the heatmap at each step and check
8087 : // that it's stable.
8088 6 : for layer in all_layers {
8089 5 : if layer.visibility() == LayerVisibilityHint::Covered {
8090 1 : continue;
8091 4 : }
8092 1 :
8093 4 : eprintln!("Downloading {layer} and re-generating heatmap");
8094 1 :
8095 4 : let ctx = &RequestContextBuilder::from(ctx)
8096 4 : .download_behavior(crate::context::DownloadBehavior::Download)
8097 4 : .attached_child();
8098 1 :
8099 4 : let _resident = layer
8100 4 : .download_and_keep_resident(ctx)
8101 4 : .instrument(tracing::info_span!(
8102 4 : parent: None,
8103 1 : "download_layer",
8104 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
8105 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
8106 1 : timeline_id = %timeline.timeline_id
8107 1 : ))
8108 4 : .await
8109 4 : .unwrap();
8110 1 :
8111 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
8112 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
8113 1 : }
8114 1 :
8115 1 : // Everything from the post-migration heatmap is now resident.
8116 1 : // Check that we drop it from memory.
8117 1 : assert!(matches!(
8118 1 : timeline.previous_heatmap.load().as_deref(),
8119 1 : Some(PreviousHeatmap::Obsolete)
8120 1 : ));
8121 1 : }
8122 :
8123 : #[tokio::test]
8124 1 : async fn test_previous_heatmap_obsoletion() {
8125 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
8126 1 : .await
8127 1 : .unwrap();
8128 :
8129 1 : let l0_delta = DeltaLayerTestDesc::new(
8130 1 : Lsn(0x20)..Lsn(0x30),
8131 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8132 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8133 1 : vec![(
8134 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8135 1 : Lsn(0x25),
8136 1 : Value::Image(test_img("foo")),
8137 1 : )],
8138 : );
8139 :
8140 1 : let image_layer = (
8141 1 : Lsn(0x40),
8142 1 : vec![(
8143 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8144 1 : test_img("bar"),
8145 1 : )],
8146 1 : );
8147 :
8148 1 : let delta_layers = vec![l0_delta];
8149 1 : let image_layers = vec![image_layer];
8150 :
8151 1 : let (tenant, ctx) = harness.load().await;
8152 1 : let timeline = tenant
8153 1 : .create_test_timeline_with_layers(
8154 1 : TimelineId::generate(),
8155 1 : Lsn(0x10),
8156 1 : PgMajorVersion::PG14,
8157 1 : &ctx,
8158 1 : Vec::new(), // in-memory layers
8159 1 : delta_layers,
8160 1 : image_layers,
8161 1 : Lsn(0x100),
8162 1 : )
8163 1 : .await
8164 1 : .unwrap();
8165 :
8166 : // Layer visibility is an input to heatmap generation, so refresh it first
8167 1 : timeline.update_layer_visibility().await.unwrap();
8168 :
8169 1 : let heatmap = timeline
8170 1 : .generate_heatmap()
8171 1 : .await
8172 1 : .expect("Infallible while timeline is not shut down");
8173 :
8174 : // Both layers should be in the heatmap
8175 1 : assert!(heatmap.all_layers().count() > 0);
8176 :
8177 : // Now simulate a migration.
8178 1 : timeline
8179 1 : .previous_heatmap
8180 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8181 1 : heatmap: heatmap.clone(),
8182 1 : read_at: std::time::Instant::now(),
8183 1 : end_lsn: None,
8184 1 : })));
8185 :
8186 : // Evict all the layers in the previous heatmap
8187 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8188 1 : let forever = std::time::Duration::from_secs(120);
8189 3 : for layer in guard.likely_resident_layers() {
8190 3 : layer.evict_and_wait(forever).await.unwrap();
8191 : }
8192 1 : drop(guard);
8193 :
8194 : // Generate a new heatmap and check that the previous heatmap
8195 : // has been marked obsolete.
8196 1 : let post_eviction_heatmap = timeline
8197 1 : .generate_heatmap()
8198 1 : .await
8199 1 : .expect("Infallible while timeline is not shut down");
8200 :
8201 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
8202 1 : assert!(matches!(
8203 1 : timeline.previous_heatmap.load().as_deref(),
8204 1 : Some(PreviousHeatmap::Obsolete)
8205 1 : ));
8206 1 : }
8207 :
8208 : #[tokio::test]
8209 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
8210 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
8211 1 : .await
8212 1 : .unwrap();
8213 :
8214 1 : let (tenant, ctx) = harness.load().await;
8215 1 : let timeline = tenant
8216 1 : .create_test_timeline(
8217 1 : TimelineId::generate(),
8218 1 : Lsn(0x10),
8219 1 : PgMajorVersion::PG14,
8220 1 : &ctx,
8221 1 : )
8222 1 : .await
8223 1 : .unwrap();
8224 :
8225 1 : let layer = find_some_layer(&timeline).await;
8226 1 : let layer = layer
8227 1 : .keep_resident()
8228 1 : .await
8229 1 : .expect("no download => no downloading errors")
8230 1 : .drop_eviction_guard();
8231 :
8232 1 : let forever = std::time::Duration::from_secs(120);
8233 :
8234 1 : let first = layer.evict_and_wait(forever);
8235 1 : let second = layer.evict_and_wait(forever);
8236 :
8237 1 : let (first, second) = tokio::join!(first, second);
8238 :
8239 1 : let res = layer.keep_resident().await;
8240 1 : assert!(res.is_none(), "{res:?}");
8241 :
8242 1 : match (first, second) {
8243 1 : (Ok(()), Ok(())) => {
8244 1 : // because there are no more timeline locks being taken on eviction path, we can
8245 1 : // witness all three outcomes here.
8246 1 : }
8247 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
8248 0 : // if one completes before the other, this is fine just as well.
8249 0 : }
8250 1 : other => unreachable!("unexpected {:?}", other),
8251 1 : }
8252 1 : }
8253 :
8254 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
8255 1 : let layers = timeline
8256 1 : .layers
8257 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
8258 1 : .await;
8259 1 : let desc = layers
8260 1 : .layer_map()
8261 1 : .unwrap()
8262 1 : .iter_historic_layers()
8263 1 : .next()
8264 1 : .expect("must find one layer to evict");
8265 :
8266 1 : layers.get_from_desc(&desc)
8267 1 : }
8268 : }
|
