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 312430 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
670 312430 : Self {
671 312430 : keyspace,
672 312430 : lsn,
673 312430 : path: Vec::new(),
674 312430 : }
675 312430 : }
676 :
677 446927 : pub fn record_layer_visit(
678 446927 : &mut self,
679 446927 : layer_to_read: &ReadableLayer,
680 446927 : keyspace_to_read: &KeySpace,
681 446927 : lsn_range: &Range<Lsn>,
682 446927 : ) {
683 446927 : let id = match layer_to_read {
684 139630 : ReadableLayer::PersistentLayer(layer) => {
685 139630 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
686 : }
687 307297 : ReadableLayer::InMemoryLayer(layer) => {
688 307297 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
689 : }
690 : };
691 446927 : self.path
692 446927 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
693 446927 : }
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 : /// Only for GC compaction.
943 : /// If set, the compaction will compact the metadata layers. Should be only set to true in unit tests
944 : /// because metadata compaction is not fully supported yet.
945 : pub gc_compaction_do_metadata_compaction: bool,
946 : }
947 :
948 : impl CompactOptions {
949 : #[cfg(test)]
950 28 : pub fn default_for_gc_compaction_unit_tests() -> Self {
951 28 : Self {
952 28 : gc_compaction_do_metadata_compaction: true,
953 28 : ..Default::default()
954 28 : }
955 28 : }
956 : }
957 :
958 : impl std::fmt::Debug for Timeline {
959 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
960 0 : write!(f, "Timeline<{}>", self.timeline_id)
961 0 : }
962 : }
963 :
964 : #[derive(thiserror::Error, Debug, Clone)]
965 : pub enum WaitLsnError {
966 : // Called on a timeline which is shutting down
967 : #[error("Shutdown")]
968 : Shutdown,
969 :
970 : // Called on an timeline not in active state or shutting down
971 : #[error("Bad timeline state: {0:?}")]
972 : BadState(TimelineState),
973 :
974 : // Timeout expired while waiting for LSN to catch up with goal.
975 : #[error("{0}")]
976 : Timeout(String),
977 : }
978 :
979 : impl WaitLsnError {
980 0 : pub(crate) fn is_cancel(&self) -> bool {
981 0 : match self {
982 0 : WaitLsnError::Shutdown => true,
983 0 : WaitLsnError::BadState(timeline_state) => match timeline_state {
984 0 : TimelineState::Loading => false,
985 0 : TimelineState::Active => false,
986 0 : TimelineState::Stopping => true,
987 0 : TimelineState::Broken { .. } => false,
988 : },
989 0 : WaitLsnError::Timeout(_) => false,
990 : }
991 0 : }
992 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
993 0 : match self {
994 0 : WaitLsnError::Shutdown => anyhow::Error::new(self),
995 0 : WaitLsnError::BadState(_) => anyhow::Error::new(self),
996 0 : WaitLsnError::Timeout(_) => anyhow::Error::new(self),
997 : }
998 0 : }
999 : }
1000 :
1001 : impl From<WaitLsnError> for tonic::Status {
1002 0 : fn from(err: WaitLsnError) -> Self {
1003 : use tonic::Code;
1004 0 : let code = if err.is_cancel() {
1005 0 : Code::Unavailable
1006 : } else {
1007 0 : Code::Internal
1008 : };
1009 0 : tonic::Status::new(code, err.to_string())
1010 0 : }
1011 : }
1012 :
1013 : // The impls below achieve cancellation mapping for errors.
1014 : // Perhaps there's a way of achieving this with less cruft.
1015 :
1016 : impl From<CreateImageLayersError> for CompactionError {
1017 0 : fn from(e: CreateImageLayersError) -> Self {
1018 0 : match e {
1019 0 : CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
1020 0 : CreateImageLayersError::Other(e) => {
1021 0 : CompactionError::Other(e.context("create image layers"))
1022 : }
1023 0 : _ => CompactionError::Other(e.into()),
1024 : }
1025 0 : }
1026 : }
1027 :
1028 : impl From<CreateImageLayersError> for FlushLayerError {
1029 0 : fn from(e: CreateImageLayersError) -> Self {
1030 0 : match e {
1031 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
1032 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
1033 : }
1034 0 : }
1035 : }
1036 :
1037 : impl From<PageReconstructError> for CreateImageLayersError {
1038 0 : fn from(e: PageReconstructError) -> Self {
1039 0 : match e {
1040 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
1041 0 : _ => CreateImageLayersError::PageReconstructError(e),
1042 : }
1043 0 : }
1044 : }
1045 :
1046 : impl From<super::storage_layer::errors::PutError> for CreateImageLayersError {
1047 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
1048 0 : if e.is_cancel() {
1049 0 : CreateImageLayersError::Cancelled
1050 : } else {
1051 0 : CreateImageLayersError::Other(e.into_anyhow())
1052 : }
1053 0 : }
1054 : }
1055 :
1056 : impl From<GetVectoredError> for CreateImageLayersError {
1057 0 : fn from(e: GetVectoredError) -> Self {
1058 0 : match e {
1059 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
1060 0 : _ => CreateImageLayersError::GetVectoredError(e),
1061 : }
1062 0 : }
1063 : }
1064 :
1065 : impl From<GetVectoredError> for PageReconstructError {
1066 3 : fn from(e: GetVectoredError) -> Self {
1067 3 : match e {
1068 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1069 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1070 0 : err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
1071 2 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1072 1 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1073 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1074 : }
1075 3 : }
1076 : }
1077 :
1078 : impl From<GetReadyAncestorError> for PageReconstructError {
1079 1 : fn from(e: GetReadyAncestorError) -> Self {
1080 : use GetReadyAncestorError::*;
1081 1 : match e {
1082 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1083 1 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1084 0 : Cancelled => PageReconstructError::Cancelled,
1085 : }
1086 1 : }
1087 : }
1088 :
1089 : pub(crate) enum WaitLsnTimeout {
1090 : Custom(Duration),
1091 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1092 : Default,
1093 : }
1094 :
1095 : pub(crate) enum WaitLsnWaiter<'a> {
1096 : Timeline(&'a Timeline),
1097 : Tenant,
1098 : PageService,
1099 : HttpEndpoint,
1100 : BaseBackupCache,
1101 : }
1102 :
1103 : /// Argument to [`Timeline::shutdown`].
1104 : #[derive(Debug, Clone, Copy)]
1105 : pub(crate) enum ShutdownMode {
1106 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
1107 : /// take multiple seconds for a busy timeline.
1108 : ///
1109 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1110 : /// the call to [`Timeline::shutdown`].
1111 : FreezeAndFlush,
1112 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1113 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1114 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1115 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1116 : Reload,
1117 : /// Shut down immediately, without waiting for any open layers to flush.
1118 : Hard,
1119 : }
1120 :
1121 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1122 : enum ImageLayerCreationOutcome {
1123 : /// We generated an image layer
1124 : Generated {
1125 : unfinished_image_layer: ImageLayerWriter,
1126 : },
1127 : /// The key range is empty
1128 : Empty,
1129 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1130 : /// the image layer creation.
1131 : Skip,
1132 : }
1133 :
1134 : enum RepartitionError {
1135 : Other(anyhow::Error),
1136 : CollectKeyspace(CollectKeySpaceError),
1137 : }
1138 :
1139 : impl RepartitionError {
1140 0 : fn is_cancel(&self) -> bool {
1141 0 : match self {
1142 0 : RepartitionError::Other(_) => false,
1143 0 : RepartitionError::CollectKeyspace(e) => e.is_cancel(),
1144 : }
1145 0 : }
1146 0 : fn into_anyhow(self) -> anyhow::Error {
1147 0 : match self {
1148 0 : RepartitionError::Other(e) => e,
1149 0 : RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
1150 : }
1151 0 : }
1152 : }
1153 :
1154 : /// Public interface functions
1155 : impl Timeline {
1156 : /// Get the LSN where this branch was created
1157 80 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1158 80 : self.ancestor_lsn
1159 80 : }
1160 :
1161 : /// Get the ancestor's timeline id
1162 416 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1163 416 : self.ancestor_timeline
1164 416 : .as_ref()
1165 416 : .map(|ancestor| ancestor.timeline_id)
1166 416 : }
1167 :
1168 : /// Get the ancestor timeline
1169 1 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1170 1 : self.ancestor_timeline.as_ref()
1171 1 : }
1172 :
1173 : /// Get the bytes written since the PITR cutoff on this branch, and
1174 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1175 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1176 : // TODO: for backwards compatibility, we return the full history back to 0 when the PITR
1177 : // cutoff has not yet been initialized. This should return None instead, but this is exposed
1178 : // in external HTTP APIs and callers may not handle a null value.
1179 0 : let gc_info = self.gc_info.read().unwrap();
1180 0 : let history = self
1181 0 : .get_last_record_lsn()
1182 0 : .checked_sub(gc_info.cutoffs.time.unwrap_or_default())
1183 0 : .unwrap_or_default()
1184 0 : .0;
1185 0 : (history, gc_info.within_ancestor_pitr)
1186 0 : }
1187 :
1188 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1189 427090 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1190 427090 : self.applied_gc_cutoff_lsn.read()
1191 427090 : }
1192 :
1193 : /// Read timeline's planned GC cutoff: this is the logical end of history that users are allowed
1194 : /// to read (based on configured PITR), even if physically we have more history. Returns None
1195 : /// if the PITR cutoff has not yet been initialized.
1196 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Option<Lsn> {
1197 0 : self.gc_info.read().unwrap().cutoffs.time
1198 0 : }
1199 :
1200 : /// Look up given page version.
1201 : ///
1202 : /// If a remote layer file is needed, it is downloaded as part of this
1203 : /// call.
1204 : ///
1205 : /// This method enforces [`Self::pagestream_throttle`] internally.
1206 : ///
1207 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1208 : /// abstraction above this needs to store suitable metadata to track what
1209 : /// data exists with what keys, in separate metadata entries. If a
1210 : /// non-existent key is requested, we may incorrectly return a value from
1211 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1212 : /// non-existing key.
1213 : ///
1214 : /// # Cancel-Safety
1215 : ///
1216 : /// This method is cancellation-safe.
1217 : #[inline(always)]
1218 301292 : pub(crate) async fn get(
1219 301292 : &self,
1220 301292 : key: Key,
1221 301292 : lsn: Lsn,
1222 301292 : ctx: &RequestContext,
1223 301292 : ) -> Result<Bytes, PageReconstructError> {
1224 301292 : if !lsn.is_valid() {
1225 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1226 301292 : }
1227 :
1228 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1229 : // already checked the key against the shard_identity when looking up the Timeline from
1230 : // page_service.
1231 301292 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1232 :
1233 301292 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1234 :
1235 301292 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1236 :
1237 301292 : let vectored_res = self
1238 301292 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1239 301292 : .await;
1240 :
1241 301292 : let key_value = vectored_res?.pop_first();
1242 301289 : match key_value {
1243 301283 : Some((got_key, value)) => {
1244 301283 : if got_key != key {
1245 0 : error!(
1246 0 : "Expected {}, but singular vectored get returned {}",
1247 : key, got_key
1248 : );
1249 0 : Err(PageReconstructError::Other(anyhow!(
1250 0 : "Singular vectored get returned wrong key"
1251 0 : )))
1252 : } else {
1253 301283 : value
1254 : }
1255 : }
1256 6 : None => Err(PageReconstructError::MissingKey(Box::new(
1257 6 : MissingKeyError {
1258 6 : keyspace: KeySpace::single(key..key.next()),
1259 6 : shard: self.shard_identity.get_shard_number(&key),
1260 6 : original_hwm_lsn: lsn,
1261 6 : ancestor_lsn: None,
1262 6 : backtrace: None,
1263 6 : read_path: None,
1264 6 : query: None,
1265 6 : },
1266 6 : ))),
1267 : }
1268 301292 : }
1269 :
1270 : #[inline(always)]
1271 0 : pub(crate) async fn debug_get(
1272 0 : &self,
1273 0 : key: Key,
1274 0 : lsn: Lsn,
1275 0 : ctx: &RequestContext,
1276 0 : reconstruct_state: &mut ValuesReconstructState,
1277 0 : ) -> Result<Bytes, PageReconstructError> {
1278 0 : if !lsn.is_valid() {
1279 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1280 0 : }
1281 :
1282 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1283 : // already checked the key against the shard_identity when looking up the Timeline from
1284 : // page_service.
1285 0 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1286 :
1287 0 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1288 0 : let vectored_res = self
1289 0 : .debug_get_vectored_impl(query, reconstruct_state, ctx)
1290 0 : .await;
1291 :
1292 0 : let key_value = vectored_res?.pop_first();
1293 0 : match key_value {
1294 0 : Some((got_key, value)) => {
1295 0 : if got_key != key {
1296 0 : error!(
1297 0 : "Expected {}, but singular vectored get returned {}",
1298 : key, got_key
1299 : );
1300 0 : Err(PageReconstructError::Other(anyhow!(
1301 0 : "Singular vectored get returned wrong key"
1302 0 : )))
1303 : } else {
1304 0 : value
1305 : }
1306 : }
1307 0 : None => Err(PageReconstructError::MissingKey(Box::new(
1308 0 : MissingKeyError {
1309 0 : keyspace: KeySpace::single(key..key.next()),
1310 0 : shard: self.shard_identity.get_shard_number(&key),
1311 0 : original_hwm_lsn: lsn,
1312 0 : ancestor_lsn: None,
1313 0 : backtrace: None,
1314 0 : read_path: None,
1315 0 : query: None,
1316 0 : },
1317 0 : ))),
1318 : }
1319 0 : }
1320 :
1321 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1322 :
1323 : /// Look up multiple page versions at a given LSN
1324 : ///
1325 : /// This naive implementation will be replaced with a more efficient one
1326 : /// which actually vectorizes the read path.
1327 10894 : pub(crate) async fn get_vectored(
1328 10894 : &self,
1329 10894 : query: VersionedKeySpaceQuery,
1330 10894 : io_concurrency: super::storage_layer::IoConcurrency,
1331 10894 : ctx: &RequestContext,
1332 10894 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1333 10894 : let total_keyspace = query.total_keyspace();
1334 :
1335 10894 : let key_count = total_keyspace.total_raw_size();
1336 10894 : if key_count > self.conf.max_get_vectored_keys.get() {
1337 0 : return Err(GetVectoredError::Oversized(
1338 0 : key_count as u64,
1339 0 : self.conf.max_get_vectored_keys.get() as u64,
1340 0 : ));
1341 10894 : }
1342 :
1343 34207 : for range in &total_keyspace.ranges {
1344 23313 : let mut key = range.start;
1345 65342 : while key != range.end {
1346 42029 : assert!(!self.shard_identity.is_key_disposable(&key));
1347 42029 : key = key.next();
1348 : }
1349 : }
1350 :
1351 10894 : trace!(
1352 0 : "get vectored query {} from task kind {:?}",
1353 : query,
1354 0 : ctx.task_kind(),
1355 : );
1356 :
1357 10894 : let start = crate::metrics::GET_VECTORED_LATENCY
1358 10894 : .for_task_kind(ctx.task_kind())
1359 10894 : .map(|metric| (metric, Instant::now()));
1360 :
1361 10894 : let res = self
1362 10894 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1363 10894 : .await;
1364 :
1365 10894 : if let Some((metric, start)) = start {
1366 0 : let elapsed = start.elapsed();
1367 0 : metric.observe(elapsed.as_secs_f64());
1368 10894 : }
1369 :
1370 10894 : res
1371 10894 : }
1372 :
1373 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1374 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1375 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1376 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1377 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1378 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1379 : /// the scan operation will not cause OOM in the future.
1380 8 : pub(crate) async fn scan(
1381 8 : &self,
1382 8 : keyspace: KeySpace,
1383 8 : lsn: Lsn,
1384 8 : ctx: &RequestContext,
1385 8 : io_concurrency: super::storage_layer::IoConcurrency,
1386 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1387 8 : if !lsn.is_valid() {
1388 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1389 8 : }
1390 :
1391 8 : trace!(
1392 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1393 : keyspace,
1394 : lsn,
1395 0 : ctx.task_kind()
1396 : );
1397 :
1398 : // We should generalize this into Keyspace::contains in the future.
1399 16 : for range in &keyspace.ranges {
1400 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1401 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1402 : {
1403 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1404 0 : "only metadata keyspace can be scanned"
1405 0 : )));
1406 8 : }
1407 : }
1408 :
1409 8 : let start = crate::metrics::SCAN_LATENCY
1410 8 : .for_task_kind(ctx.task_kind())
1411 8 : .map(ScanLatencyOngoingRecording::start_recording);
1412 :
1413 8 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1414 :
1415 8 : let vectored_res = self
1416 8 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1417 8 : .await;
1418 :
1419 8 : if let Some(recording) = start {
1420 0 : recording.observe();
1421 8 : }
1422 :
1423 8 : vectored_res
1424 8 : }
1425 :
1426 312430 : pub(super) async fn get_vectored_impl(
1427 312430 : &self,
1428 312430 : query: VersionedKeySpaceQuery,
1429 312430 : reconstruct_state: &mut ValuesReconstructState,
1430 312430 : ctx: &RequestContext,
1431 312430 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1432 312430 : if query.is_empty() {
1433 0 : return Ok(BTreeMap::default());
1434 312430 : }
1435 :
1436 312430 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1437 312430 : Some(ReadPath::new(
1438 312430 : query.total_keyspace(),
1439 312430 : query.high_watermark_lsn()?,
1440 : ))
1441 : } else {
1442 0 : None
1443 : };
1444 :
1445 312430 : reconstruct_state.read_path = read_path;
1446 :
1447 312430 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1448 0 : RedoAttemptType::LegacyCompaction
1449 : } else {
1450 312430 : RedoAttemptType::ReadPage
1451 : };
1452 :
1453 312430 : let traversal_res: Result<(), _> = {
1454 312430 : let ctx = RequestContextBuilder::from(ctx)
1455 312430 : .perf_span(|crnt_perf_span| {
1456 0 : info_span!(
1457 : target: PERF_TRACE_TARGET,
1458 0 : parent: crnt_perf_span,
1459 : "PLAN_IO",
1460 : )
1461 0 : })
1462 312430 : .attached_child();
1463 :
1464 312430 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1465 312430 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1466 312430 : .await
1467 : };
1468 :
1469 312430 : if let Err(err) = traversal_res {
1470 : // Wait for all the spawned IOs to complete.
1471 : // See comments on `spawn_io` inside `storage_layer` for more details.
1472 8 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1473 8 : .into_values()
1474 8 : .map(|state| state.collect_pending_ios())
1475 8 : .collect::<FuturesUnordered<_>>();
1476 8 : while collect_futs.next().await.is_some() {}
1477 :
1478 : // Enrich the missing key error with the original query.
1479 8 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1480 7 : missing_err.enrich(query.clone());
1481 7 : return Err(GetVectoredError::MissingKey(missing_err));
1482 1 : }
1483 :
1484 1 : return Err(err);
1485 312422 : };
1486 :
1487 312422 : let layers_visited = reconstruct_state.get_layers_visited();
1488 :
1489 312422 : let ctx = RequestContextBuilder::from(ctx)
1490 312422 : .perf_span(|crnt_perf_span| {
1491 0 : info_span!(
1492 : target: PERF_TRACE_TARGET,
1493 0 : parent: crnt_perf_span,
1494 : "RECONSTRUCT",
1495 : )
1496 0 : })
1497 312422 : .attached_child();
1498 :
1499 312422 : let futs = FuturesUnordered::new();
1500 363551 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1501 363551 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1502 :
1503 363551 : futs.push({
1504 363551 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1505 363551 : let ctx = RequestContextBuilder::from(&ctx)
1506 363551 : .perf_span(|crnt_perf_span| {
1507 0 : info_span!(
1508 : target: PERF_TRACE_TARGET,
1509 0 : parent: crnt_perf_span,
1510 : "RECONSTRUCT_KEY",
1511 : key = %key,
1512 : )
1513 0 : })
1514 363551 : .attached_child();
1515 :
1516 363551 : async move {
1517 363551 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1518 :
1519 363551 : let res = state
1520 363551 : .collect_pending_ios()
1521 363551 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1522 0 : info_span!(
1523 : target: PERF_TRACE_TARGET,
1524 0 : parent: crnt_perf_span,
1525 : "WAIT_FOR_IO_COMPLETIONS",
1526 : )
1527 0 : })
1528 363551 : .await;
1529 :
1530 363551 : let converted = match res {
1531 363551 : Ok(ok) => ok,
1532 0 : Err(err) => {
1533 0 : return (key, Err(err));
1534 : }
1535 : };
1536 363551 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1537 :
1538 : // The walredo module expects the records to be descending in terms of Lsn.
1539 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1540 363551 : debug_assert!(
1541 363551 : converted
1542 363551 : .records
1543 1403254 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1544 0 : "{converted:?}"
1545 : );
1546 :
1547 363551 : let walredo_deltas = converted.num_deltas();
1548 363551 : let walredo_res = walredo_self
1549 363551 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1550 363551 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1551 0 : info_span!(
1552 : target: PERF_TRACE_TARGET,
1553 0 : parent: crnt_perf_span,
1554 : "WALREDO",
1555 : deltas = %walredo_deltas,
1556 : )
1557 0 : })
1558 363551 : .await;
1559 :
1560 363551 : (key, walredo_res)
1561 363551 : }
1562 : });
1563 : }
1564 :
1565 312422 : let results = futs
1566 312422 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1567 312422 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1568 312422 : .await;
1569 :
1570 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1571 : // when they're missing. Instead they are omitted from the resulting btree
1572 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1573 : // to avoid infinite results.
1574 312422 : if !results.is_empty() {
1575 312243 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1576 0 : let total_keyspace = query.total_keyspace();
1577 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1578 :
1579 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1580 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1581 0 : LOG_PACER.lock().unwrap().call(|| {
1582 0 : let num_keys = total_keyspace.total_raw_size();
1583 0 : let num_pages = results.len();
1584 0 : tracing::info!(
1585 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1586 : lsn = %max_request_lsn,
1587 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1588 : );
1589 0 : });
1590 312243 : }
1591 :
1592 : // Records the number of layers visited in a few different ways:
1593 : //
1594 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1595 : // layer directly affects its observed latency.
1596 : //
1597 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1598 : // layer visits and access cost.
1599 : //
1600 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1601 : // read amplification after batching.
1602 312243 : let layers_visited = layers_visited as f64;
1603 312243 : let avg_layers_visited = layers_visited / results.len() as f64;
1604 312243 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1605 675794 : for _ in &results {
1606 363551 : self.metrics.layers_per_read.observe(layers_visited);
1607 363551 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1608 363551 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1609 363551 : }
1610 179 : }
1611 :
1612 312422 : Ok(results)
1613 312430 : }
1614 :
1615 : // A copy of the get_vectored_impl method except that we store the image and wal records into `reconstruct_state`.
1616 : // This is only used in the http getpage call for debugging purpose.
1617 0 : pub(super) async fn debug_get_vectored_impl(
1618 0 : &self,
1619 0 : query: VersionedKeySpaceQuery,
1620 0 : reconstruct_state: &mut ValuesReconstructState,
1621 0 : ctx: &RequestContext,
1622 0 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1623 0 : if query.is_empty() {
1624 0 : return Ok(BTreeMap::default());
1625 0 : }
1626 :
1627 0 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1628 0 : Some(ReadPath::new(
1629 0 : query.total_keyspace(),
1630 0 : query.high_watermark_lsn()?,
1631 : ))
1632 : } else {
1633 0 : None
1634 : };
1635 :
1636 0 : reconstruct_state.read_path = read_path;
1637 :
1638 0 : let traversal_res: Result<(), _> = self
1639 0 : .get_vectored_reconstruct_data(query.clone(), reconstruct_state, ctx)
1640 0 : .await;
1641 :
1642 0 : if let Err(err) = traversal_res {
1643 : // Wait for all the spawned IOs to complete.
1644 : // See comments on `spawn_io` inside `storage_layer` for more details.
1645 0 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1646 0 : .into_values()
1647 0 : .map(|state| state.collect_pending_ios())
1648 0 : .collect::<FuturesUnordered<_>>();
1649 0 : while collect_futs.next().await.is_some() {}
1650 0 : return Err(err);
1651 0 : };
1652 :
1653 0 : let reconstruct_state = Arc::new(Mutex::new(reconstruct_state));
1654 0 : let futs = FuturesUnordered::new();
1655 :
1656 0 : for (key, state) in std::mem::take(&mut reconstruct_state.lock().unwrap().keys) {
1657 0 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1658 0 : futs.push({
1659 0 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1660 0 : let rc_clone = Arc::clone(&reconstruct_state);
1661 :
1662 0 : async move {
1663 0 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1664 :
1665 0 : let converted = match state.collect_pending_ios().await {
1666 0 : Ok(ok) => ok,
1667 0 : Err(err) => {
1668 0 : return (key, Err(err));
1669 : }
1670 : };
1671 0 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1672 :
1673 : // The walredo module expects the records to be descending in terms of Lsn.
1674 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1675 0 : debug_assert!(
1676 0 : converted
1677 0 : .records
1678 0 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1679 0 : "{converted:?}"
1680 : );
1681 0 : {
1682 0 : let mut guard = rc_clone.lock().unwrap();
1683 0 : guard.set_debug_state(&converted);
1684 0 : }
1685 : (
1686 0 : key,
1687 0 : walredo_self
1688 0 : .reconstruct_value(
1689 0 : key,
1690 0 : req_lsn_for_key,
1691 0 : converted,
1692 0 : RedoAttemptType::ReadPage,
1693 0 : )
1694 0 : .await,
1695 : )
1696 0 : }
1697 : });
1698 : }
1699 :
1700 0 : let results = futs
1701 0 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1702 0 : .await;
1703 :
1704 0 : Ok(results)
1705 0 : }
1706 :
1707 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1708 138202 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1709 138202 : self.last_record_lsn.load().last
1710 138202 : }
1711 :
1712 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1713 0 : self.last_record_lsn.load().prev
1714 0 : }
1715 :
1716 : /// Atomically get both last and prev.
1717 117 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1718 117 : self.last_record_lsn.load()
1719 117 : }
1720 :
1721 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1722 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1723 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1724 0 : self.last_record_lsn.status_receiver()
1725 0 : }
1726 :
1727 610 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1728 610 : self.disk_consistent_lsn.load()
1729 610 : }
1730 :
1731 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1732 : /// not validated with control plane yet.
1733 : /// See [`Self::get_remote_consistent_lsn_visible`].
1734 2 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1735 2 : self.remote_client.remote_consistent_lsn_projected()
1736 2 : }
1737 :
1738 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1739 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1740 : /// generation validation in the deletion queue.
1741 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1742 0 : self.remote_client.remote_consistent_lsn_visible()
1743 0 : }
1744 :
1745 : /// The sum of the file size of all historic layers in the layer map.
1746 : /// This method makes no distinction between local and remote layers.
1747 : /// Hence, the result **does not represent local filesystem usage**.
1748 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1749 0 : let guard = self
1750 0 : .layers
1751 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1752 0 : .await;
1753 0 : guard.layer_size_sum()
1754 0 : }
1755 :
1756 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1757 0 : self.metrics.resident_physical_size_get()
1758 0 : }
1759 :
1760 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1761 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1762 0 : }
1763 :
1764 : ///
1765 : /// Wait until WAL has been received and processed up to this LSN.
1766 : ///
1767 : /// You should call this before any of the other get_* or list_* functions. Calling
1768 : /// those functions with an LSN that has been processed yet is an error.
1769 : ///
1770 114122 : pub(crate) async fn wait_lsn(
1771 114122 : &self,
1772 114122 : lsn: Lsn,
1773 114122 : who_is_waiting: WaitLsnWaiter<'_>,
1774 114122 : timeout: WaitLsnTimeout,
1775 114122 : ctx: &RequestContext, /* Prepare for use by cancellation */
1776 114122 : ) -> Result<(), WaitLsnError> {
1777 114122 : let state = self.current_state();
1778 114122 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1779 0 : return Err(WaitLsnError::Shutdown);
1780 114122 : } else if !matches!(state, TimelineState::Active) {
1781 0 : return Err(WaitLsnError::BadState(state));
1782 114122 : }
1783 :
1784 114122 : if cfg!(debug_assertions) {
1785 114122 : match ctx.task_kind() {
1786 : TaskKind::WalReceiverManager
1787 : | TaskKind::WalReceiverConnectionHandler
1788 : | TaskKind::WalReceiverConnectionPoller => {
1789 0 : let is_myself = match who_is_waiting {
1790 0 : WaitLsnWaiter::Timeline(waiter) => {
1791 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1792 : }
1793 : WaitLsnWaiter::Tenant
1794 : | WaitLsnWaiter::PageService
1795 : | WaitLsnWaiter::HttpEndpoint
1796 0 : | WaitLsnWaiter::BaseBackupCache => unreachable!(
1797 : "tenant or page_service context are not expected to have task kind {:?}",
1798 0 : ctx.task_kind()
1799 : ),
1800 : };
1801 0 : if is_myself {
1802 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1803 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1804 0 : panic!(
1805 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1806 : );
1807 0 : }
1808 0 : } else {
1809 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1810 0 : // our walreceiver task can make progress independent of theirs
1811 0 : }
1812 : }
1813 114122 : _ => {}
1814 : }
1815 0 : }
1816 :
1817 114122 : let timeout = match timeout {
1818 0 : WaitLsnTimeout::Custom(t) => t,
1819 114122 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1820 : };
1821 :
1822 114122 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1823 114122 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1824 :
1825 114122 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1826 114122 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1827 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1828 114122 : let log_slow_threshold = Duration::from_secs(1);
1829 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1830 114122 : let log_slow_period = Duration::from_secs(10);
1831 114122 : let mut logging_permit = None;
1832 114122 : let wait_for_timeout = monitor_slow_future(
1833 114122 : log_slow_threshold,
1834 114122 : log_slow_period,
1835 114122 : wait_for_timeout,
1836 : |MonitorSlowFutureCallback {
1837 : ready,
1838 : is_slow,
1839 : elapsed_total,
1840 : elapsed_since_last_callback,
1841 114122 : }| {
1842 114122 : self.metrics
1843 114122 : .wait_lsn_in_progress_micros
1844 114122 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1845 114122 : if !is_slow {
1846 114122 : return;
1847 0 : }
1848 : // It's slow, see if we should log it.
1849 : // (We limit the logging to one per invocation per timeline to avoid excessive
1850 : // logging during an extended broker / networking outage that affects all timelines.)
1851 0 : if logging_permit.is_none() {
1852 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1853 0 : }
1854 0 : if logging_permit.is_none() {
1855 0 : return;
1856 0 : }
1857 : // We log it.
1858 0 : if ready {
1859 0 : info!(
1860 0 : "slow wait_lsn completed after {:.3}s",
1861 0 : elapsed_total.as_secs_f64()
1862 : );
1863 : } else {
1864 0 : info!(
1865 0 : "slow wait_lsn still running for {:.3}s",
1866 0 : elapsed_total.as_secs_f64()
1867 : );
1868 : }
1869 114122 : },
1870 : );
1871 114122 : let res = wait_for_timeout.await;
1872 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1873 114122 : drop(logging_permit);
1874 114122 : drop(start_finish_counterpair_guard);
1875 114122 : drop(timer);
1876 114122 : match res {
1877 114122 : Ok(()) => Ok(()),
1878 0 : Err(e) => {
1879 : use utils::seqwait::SeqWaitError::*;
1880 0 : match e {
1881 0 : Shutdown => Err(WaitLsnError::Shutdown),
1882 : Timeout => {
1883 0 : let walreceiver_status = self.walreceiver_status();
1884 0 : Err(WaitLsnError::Timeout(format!(
1885 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1886 0 : lsn,
1887 0 : self.get_last_record_lsn(),
1888 0 : self.get_disk_consistent_lsn(),
1889 0 : walreceiver_status,
1890 0 : )))
1891 : }
1892 : }
1893 : }
1894 : }
1895 114122 : }
1896 :
1897 0 : pub(crate) fn walreceiver_status(&self) -> String {
1898 0 : match &*self.walreceiver.lock().unwrap() {
1899 0 : None => "stopping or stopped".to_string(),
1900 0 : Some(walreceiver) => match walreceiver.status() {
1901 0 : Some(status) => status.to_human_readable_string(),
1902 0 : None => "Not active".to_string(),
1903 : },
1904 : }
1905 0 : }
1906 :
1907 : /// Check that it is valid to request operations with that lsn.
1908 119 : pub(crate) fn check_lsn_is_in_scope(
1909 119 : &self,
1910 119 : lsn: Lsn,
1911 119 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1912 119 : ) -> anyhow::Result<()> {
1913 119 : ensure!(
1914 119 : lsn >= **latest_gc_cutoff_lsn,
1915 2 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1916 : lsn,
1917 2 : **latest_gc_cutoff_lsn,
1918 : );
1919 117 : Ok(())
1920 119 : }
1921 :
1922 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1923 5 : pub(crate) fn init_lsn_lease(
1924 5 : &self,
1925 5 : lsn: Lsn,
1926 5 : length: Duration,
1927 5 : ctx: &RequestContext,
1928 5 : ) -> anyhow::Result<LsnLease> {
1929 5 : self.make_lsn_lease(lsn, length, true, ctx)
1930 5 : }
1931 :
1932 : /// 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.
1933 2 : pub(crate) fn renew_lsn_lease(
1934 2 : &self,
1935 2 : lsn: Lsn,
1936 2 : length: Duration,
1937 2 : ctx: &RequestContext,
1938 2 : ) -> anyhow::Result<LsnLease> {
1939 2 : self.make_lsn_lease(lsn, length, false, ctx)
1940 2 : }
1941 :
1942 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1943 : ///
1944 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1945 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1946 : ///
1947 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1948 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1949 7 : fn make_lsn_lease(
1950 7 : &self,
1951 7 : lsn: Lsn,
1952 7 : length: Duration,
1953 7 : init: bool,
1954 7 : _ctx: &RequestContext,
1955 7 : ) -> anyhow::Result<LsnLease> {
1956 6 : let lease = {
1957 : // Normalize the requested LSN to be aligned, and move to the first record
1958 : // if it points to the beginning of the page (header).
1959 7 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1960 :
1961 7 : let mut gc_info = self.gc_info.write().unwrap();
1962 7 : let planned_cutoff = gc_info.min_cutoff();
1963 :
1964 7 : let valid_until = SystemTime::now() + length;
1965 :
1966 7 : let entry = gc_info.leases.entry(lsn);
1967 :
1968 7 : match entry {
1969 3 : Entry::Occupied(mut occupied) => {
1970 3 : let existing_lease = occupied.get_mut();
1971 3 : if valid_until > existing_lease.valid_until {
1972 1 : existing_lease.valid_until = valid_until;
1973 1 : let dt: DateTime<Utc> = valid_until.into();
1974 1 : info!("lease extended to {}", dt);
1975 : } else {
1976 2 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1977 2 : info!("existing lease covers greater length, valid until {}", dt);
1978 : }
1979 :
1980 3 : existing_lease.clone()
1981 : }
1982 4 : Entry::Vacant(vacant) => {
1983 : // Never allow a lease to be requested for an LSN below the applied GC cutoff. The data could have been deleted.
1984 4 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1985 4 : if lsn < *latest_gc_cutoff_lsn {
1986 1 : bail!(
1987 1 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1988 : lsn,
1989 1 : *latest_gc_cutoff_lsn
1990 : );
1991 3 : }
1992 :
1993 : // We allow create lease for those below the planned gc cutoff if we are still within the grace period
1994 : // of GC blocking.
1995 3 : let validate = {
1996 3 : let conf = self.tenant_conf.load();
1997 3 : !conf.is_gc_blocked_by_lsn_lease_deadline()
1998 : };
1999 :
2000 : // Do not allow initial lease creation to be below the planned gc cutoff. The client (compute_ctl) determines
2001 : // whether it is a initial lease creation or a renewal.
2002 3 : if (init || validate) && lsn < planned_cutoff {
2003 0 : bail!(
2004 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
2005 : lsn,
2006 : planned_cutoff
2007 : );
2008 3 : }
2009 :
2010 3 : let dt: DateTime<Utc> = valid_until.into();
2011 3 : info!("lease created, valid until {}", dt);
2012 3 : vacant.insert(LsnLease { valid_until }).clone()
2013 : }
2014 : }
2015 : };
2016 :
2017 6 : Ok(lease)
2018 7 : }
2019 :
2020 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2021 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2022 : #[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))]
2023 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
2024 : self.freeze0().await
2025 : }
2026 :
2027 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
2028 : #[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))]
2029 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
2030 : self.freeze_and_flush0().await
2031 : }
2032 :
2033 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
2034 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
2035 571 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
2036 571 : let mut g = self.write_lock.lock().await;
2037 571 : let to_lsn = self.get_last_record_lsn();
2038 571 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
2039 571 : }
2040 :
2041 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
2042 : // polluting the span hierarchy.
2043 571 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
2044 571 : let token = self.freeze0().await?;
2045 571 : self.wait_flush_completion(token).await
2046 571 : }
2047 :
2048 : // Check if an open ephemeral layer should be closed: this provides
2049 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
2050 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
2051 : // ephemeral layer bytes has been breached.
2052 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
2053 0 : debug_assert_current_span_has_tenant_and_timeline_id();
2054 :
2055 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
2056 : // If the write lock is held, there is an active wal receiver: rolling open layers
2057 : // is their responsibility while they hold this lock.
2058 0 : return;
2059 : };
2060 :
2061 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
2062 : // time, and this was missed.
2063 : // if write_guard.is_none() { return; }
2064 :
2065 0 : let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
2066 : // Don't block if the layer lock is busy
2067 0 : return;
2068 : };
2069 :
2070 0 : let Ok(lm) = layers_guard.layer_map() else {
2071 0 : return;
2072 : };
2073 :
2074 0 : let Some(open_layer) = &lm.open_layer else {
2075 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
2076 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
2077 : // that didn't result in writes to this shard.
2078 :
2079 : // Must not hold the layers lock while waiting for a flush.
2080 0 : drop(layers_guard);
2081 :
2082 0 : let last_record_lsn = self.get_last_record_lsn();
2083 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
2084 0 : if last_record_lsn > disk_consistent_lsn {
2085 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
2086 : // we are a sharded tenant and have skipped some WAL
2087 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
2088 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
2089 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
2090 : // without any data ingested (yet) doesn't write a remote index as soon as it
2091 : // sees its LSN advance: we only do this if we've been layer-less
2092 : // for some time.
2093 0 : tracing::debug!(
2094 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
2095 : disk_consistent_lsn,
2096 : last_record_lsn
2097 : );
2098 :
2099 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
2100 : // We know there is no open layer, so we can request freezing without actually
2101 : // freezing anything. This is true even if we have dropped the layers_guard, we
2102 : // still hold the write_guard.
2103 0 : let _ = async {
2104 0 : let token = self
2105 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
2106 0 : .await?;
2107 0 : self.wait_flush_completion(token).await
2108 0 : }
2109 0 : .await;
2110 0 : }
2111 0 : }
2112 :
2113 0 : return;
2114 : };
2115 :
2116 0 : let current_size = open_layer.len();
2117 :
2118 0 : let current_lsn = self.get_last_record_lsn();
2119 :
2120 0 : let checkpoint_distance_override = open_layer.tick();
2121 :
2122 0 : if let Some(size_override) = checkpoint_distance_override {
2123 0 : if current_size > size_override {
2124 : // This is not harmful, but it only happens in relatively rare cases where
2125 : // time-based checkpoints are not happening fast enough to keep the amount of
2126 : // ephemeral data within configured limits. It's a sign of stress on the system.
2127 0 : tracing::info!(
2128 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
2129 : );
2130 0 : }
2131 0 : }
2132 :
2133 0 : let checkpoint_distance =
2134 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
2135 :
2136 0 : if self.should_roll(
2137 0 : current_size,
2138 0 : current_size,
2139 0 : checkpoint_distance,
2140 0 : self.get_last_record_lsn(),
2141 0 : self.last_freeze_at.load(),
2142 0 : open_layer.get_opened_at(),
2143 : ) {
2144 0 : match open_layer.info() {
2145 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
2146 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
2147 : // happens asynchronously in the background.
2148 0 : tracing::debug!(
2149 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
2150 : );
2151 : }
2152 : InMemoryLayerInfo::Open { .. } => {
2153 : // Upgrade to a write lock and freeze the layer
2154 0 : drop(layers_guard);
2155 0 : let res = self
2156 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
2157 0 : .await;
2158 :
2159 0 : if let Err(e) = res {
2160 0 : tracing::info!(
2161 0 : "failed to flush frozen layer after background freeze: {e:#}"
2162 : );
2163 0 : }
2164 : }
2165 : }
2166 0 : }
2167 0 : }
2168 :
2169 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
2170 : ///
2171 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
2172 : /// child timelines that are not offloaded yet.
2173 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
2174 0 : if self.remote_client.is_archived() != Some(true) {
2175 0 : return (false, "the timeline is not archived");
2176 0 : }
2177 0 : if !self.remote_client.no_pending_work() {
2178 : // if the remote client is still processing some work, we can't offload
2179 0 : return (false, "the upload queue is not drained yet");
2180 0 : }
2181 :
2182 0 : (true, "ok")
2183 0 : }
2184 :
2185 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
2186 : /// compaction tasks.
2187 192 : pub(crate) async fn compact(
2188 192 : self: &Arc<Self>,
2189 192 : cancel: &CancellationToken,
2190 192 : flags: EnumSet<CompactFlags>,
2191 192 : ctx: &RequestContext,
2192 192 : ) -> Result<CompactionOutcome, CompactionError> {
2193 192 : let res = self
2194 192 : .compact_with_options(
2195 192 : cancel,
2196 192 : CompactOptions {
2197 192 : flags,
2198 192 : compact_key_range: None,
2199 192 : compact_lsn_range: None,
2200 192 : sub_compaction: false,
2201 192 : sub_compaction_max_job_size_mb: None,
2202 192 : gc_compaction_do_metadata_compaction: false,
2203 192 : },
2204 192 : ctx,
2205 192 : )
2206 192 : .await;
2207 192 : if let Err(err) = &res {
2208 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
2209 192 : }
2210 192 : res
2211 192 : }
2212 :
2213 : /// Outermost timeline compaction operation; downloads needed layers.
2214 : ///
2215 : /// NB: the cancellation token is usually from a background task, but can also come from a
2216 : /// request task.
2217 192 : pub(crate) async fn compact_with_options(
2218 192 : self: &Arc<Self>,
2219 192 : cancel: &CancellationToken,
2220 192 : options: CompactOptions,
2221 192 : ctx: &RequestContext,
2222 192 : ) -> Result<CompactionOutcome, CompactionError> {
2223 : // Acquire the compaction lock and task semaphore.
2224 : //
2225 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
2226 : // out by other background tasks (including image compaction). We request this via
2227 : // `BackgroundLoopKind::L0Compaction`.
2228 : //
2229 : // Yield for pending L0 compaction while waiting for the semaphore.
2230 192 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
2231 192 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
2232 0 : true => BackgroundLoopKind::L0Compaction,
2233 192 : false => BackgroundLoopKind::Compaction,
2234 : };
2235 192 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2236 192 : if yield_for_l0 {
2237 : // If this is an L0 pass, it doesn't make sense to yield for L0.
2238 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2239 : // If `compaction_l0_first` is disabled, there's no point yielding.
2240 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2241 192 : }
2242 :
2243 192 : let acquire = async move {
2244 192 : let guard = self.compaction_lock.lock().await;
2245 192 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2246 192 : (guard, permit)
2247 192 : };
2248 :
2249 192 : let (_guard, _permit) = tokio::select! {
2250 192 : (guard, permit) = acquire => (guard, permit),
2251 192 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2252 0 : return Ok(CompactionOutcome::YieldForL0);
2253 : }
2254 192 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2255 192 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2256 : };
2257 :
2258 192 : let last_record_lsn = self.get_last_record_lsn();
2259 :
2260 : // Last record Lsn could be zero in case the timeline was just created
2261 192 : if !last_record_lsn.is_valid() {
2262 0 : warn!(
2263 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2264 : );
2265 0 : return Ok(CompactionOutcome::Skipped);
2266 192 : }
2267 :
2268 192 : let result = match self.get_compaction_algorithm_settings().kind {
2269 : CompactionAlgorithm::Tiered => {
2270 0 : self.compact_tiered(cancel, ctx).await?;
2271 0 : Ok(CompactionOutcome::Done)
2272 : }
2273 192 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2274 : };
2275 :
2276 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2277 0 : match &result {
2278 192 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2279 0 : Err(e) if e.is_cancel() => {}
2280 0 : Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
2281 : };
2282 :
2283 192 : result
2284 192 : }
2285 :
2286 : /// Mutate the timeline with a [`TimelineWriter`].
2287 2566609 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2288 : TimelineWriter {
2289 2566609 : tl: self,
2290 2566609 : write_guard: self.write_lock.lock().await,
2291 : }
2292 2566609 : }
2293 :
2294 0 : pub(crate) fn activate(
2295 0 : self: &Arc<Self>,
2296 0 : parent: Arc<crate::tenant::TenantShard>,
2297 0 : broker_client: BrokerClientChannel,
2298 0 : background_jobs_can_start: Option<&completion::Barrier>,
2299 0 : ctx: &RequestContext,
2300 0 : ) {
2301 0 : if self.tenant_shard_id.is_shard_zero() {
2302 0 : // Logical size is only maintained accurately on shard zero.
2303 0 : self.spawn_initial_logical_size_computation_task(ctx);
2304 0 : }
2305 0 : self.launch_wal_receiver(ctx, broker_client);
2306 0 : self.set_state(TimelineState::Active);
2307 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2308 0 : }
2309 :
2310 : /// After this function returns, there are no timeline-scoped tasks are left running.
2311 : ///
2312 : /// The preferred pattern for is:
2313 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2314 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2315 : /// go the extra mile and keep track of JoinHandles
2316 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2317 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2318 : ///
2319 : /// For legacy reasons, we still have multiple tasks spawned using
2320 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2321 : /// We refer to these as "timeline-scoped task_mgr tasks".
2322 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2323 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2324 : /// or [`task_mgr::shutdown_watcher`].
2325 : /// We want to gradually convert the code base away from these.
2326 : ///
2327 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2328 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2329 : /// ones that aren't mentioned here):
2330 : /// - [`TaskKind::TimelineDeletionWorker`]
2331 : /// - NB: also used for tenant deletion
2332 : /// - [`TaskKind::RemoteUploadTask`]`
2333 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2334 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2335 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2336 : /// - [`TaskKind::Eviction`]
2337 : /// - [`TaskKind::LayerFlushTask`]
2338 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2339 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2340 5 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2341 5 : debug_assert_current_span_has_tenant_and_timeline_id();
2342 :
2343 : // Regardless of whether we're going to try_freeze_and_flush
2344 : // cancel walreceiver to stop ingesting more data asap.
2345 : //
2346 : // Note that we're accepting a race condition here where we may
2347 : // do the final flush below, before walreceiver observes the
2348 : // cancellation and exits.
2349 : // This means we may open a new InMemoryLayer after the final flush below.
2350 : // Flush loop is also still running for a short while, so, in theory, it
2351 : // could also make its way into the upload queue.
2352 : //
2353 : // If we wait for the shutdown of the walreceiver before moving on to the
2354 : // flush, then that would be avoided. But we don't do it because the
2355 : // walreceiver entertains reads internally, which means that it possibly
2356 : // depends on the download of layers. Layer download is only sensitive to
2357 : // the cancellation of the entire timeline, so cancelling the walreceiver
2358 : // will have no effect on the individual get requests.
2359 : // This would cause problems when there is a lot of ongoing downloads or
2360 : // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
2361 : // and we can't deallocate resources of the timeline, etc.
2362 5 : let walreceiver = self.walreceiver.lock().unwrap().take();
2363 5 : tracing::debug!(
2364 0 : is_some = walreceiver.is_some(),
2365 0 : "Waiting for WalReceiverManager..."
2366 : );
2367 5 : if let Some(walreceiver) = walreceiver {
2368 0 : walreceiver.cancel().await;
2369 5 : }
2370 : // ... and inform any waiters for newer LSNs that there won't be any.
2371 5 : self.last_record_lsn.shutdown();
2372 :
2373 5 : if let ShutdownMode::FreezeAndFlush = mode {
2374 3 : let do_flush = if let Some((open, frozen)) = self
2375 3 : .layers
2376 3 : .read(LayerManagerLockHolder::Shutdown)
2377 3 : .await
2378 3 : .layer_map()
2379 3 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2380 3 : .ok()
2381 3 : .filter(|(open, frozen)| *open || *frozen > 0)
2382 : {
2383 0 : if self.remote_client.is_archived() == Some(true) {
2384 : // No point flushing on shutdown for an archived timeline: it is not important
2385 : // to have it nice and fresh after our restart, and trying to flush here might
2386 : // race with trying to offload it (which also stops the flush loop)
2387 0 : false
2388 : } else {
2389 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2390 0 : true
2391 : }
2392 : } else {
2393 : // this is double-shutdown, it'll be a no-op
2394 3 : true
2395 : };
2396 :
2397 : // we shut down walreceiver above, so, we won't add anything more
2398 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2399 : // to reach the disk & upload queue, then shut the upload queue and
2400 : // wait for it to drain.
2401 3 : if do_flush {
2402 3 : match self.freeze_and_flush().await {
2403 : Ok(_) => {
2404 : // drain the upload queue
2405 : // if we did not wait for completion here, it might be our shutdown process
2406 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2407 : // be spawned.
2408 : //
2409 : // what is problematic is the shutting down of RemoteTimelineClient, because
2410 : // obviously it does not make sense to stop while we wait for it, but what
2411 : // about corner cases like s3 suddenly hanging up?
2412 3 : self.remote_client.shutdown().await;
2413 : }
2414 : Err(FlushLayerError::Cancelled) => {
2415 : // this is likely the second shutdown, ignore silently.
2416 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2417 0 : debug_assert!(self.cancel.is_cancelled());
2418 : }
2419 0 : Err(e) => {
2420 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2421 : // we have some extra WAL replay to do next time the timeline starts.
2422 0 : warn!("failed to freeze and flush: {e:#}");
2423 : }
2424 : }
2425 :
2426 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2427 : // we also do a final check here to ensure that the queue is empty.
2428 3 : if !self.remote_client.no_pending_work() {
2429 0 : warn!(
2430 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2431 : );
2432 3 : }
2433 0 : }
2434 2 : }
2435 :
2436 5 : if let ShutdownMode::Reload = mode {
2437 : // drain the upload queue
2438 1 : self.remote_client.shutdown().await;
2439 1 : if !self.remote_client.no_pending_work() {
2440 0 : warn!(
2441 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2442 : );
2443 1 : }
2444 4 : }
2445 :
2446 : // Signal any subscribers to our cancellation token to drop out
2447 5 : tracing::debug!("Cancelling CancellationToken");
2448 5 : self.cancel.cancel();
2449 :
2450 : // If we have a background task downloading heatmap layers stop it.
2451 : // The background downloads are sensitive to timeline cancellation (done above),
2452 : // so the drain will be immediate.
2453 5 : self.stop_and_drain_heatmap_layers_download().await;
2454 :
2455 : // Ensure Prevent new page service requests from starting.
2456 5 : self.handles.shutdown();
2457 :
2458 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2459 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2460 5 : self.remote_client.stop();
2461 :
2462 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2463 : // to shut down the upload queue tasks.
2464 : // TODO: fix that, task management should be encapsulated inside remote_client.
2465 5 : task_mgr::shutdown_tasks(
2466 5 : Some(TaskKind::RemoteUploadTask),
2467 5 : Some(self.tenant_shard_id),
2468 5 : Some(self.timeline_id),
2469 5 : )
2470 5 : .await;
2471 :
2472 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2473 5 : tracing::debug!("Waiting for tasks...");
2474 5 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2475 :
2476 : {
2477 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2478 : // open.
2479 5 : let mut write_guard = self.write_lock.lock().await;
2480 5 : self.layers
2481 5 : .write(LayerManagerLockHolder::Shutdown)
2482 5 : .await
2483 5 : .shutdown(&mut write_guard);
2484 : }
2485 :
2486 : // Finally wait until any gate-holders are complete.
2487 : //
2488 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2489 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2490 5 : self.gate.close().await;
2491 :
2492 5 : self.metrics.shutdown();
2493 5 : }
2494 :
2495 236 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2496 236 : match (self.current_state(), new_state) {
2497 236 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2498 1 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2499 : }
2500 0 : (st, TimelineState::Loading) => {
2501 0 : error!("ignoring transition from {st:?} into Loading state");
2502 : }
2503 0 : (TimelineState::Broken { .. }, new_state) => {
2504 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2505 : }
2506 : (TimelineState::Stopping, TimelineState::Active) => {
2507 0 : error!("Not activating a Stopping timeline");
2508 : }
2509 235 : (_, new_state) => {
2510 235 : self.state.send_replace(new_state);
2511 235 : }
2512 : }
2513 236 : }
2514 :
2515 1 : pub(crate) fn set_broken(&self, reason: String) {
2516 1 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2517 1 : let broken_state = TimelineState::Broken {
2518 1 : reason,
2519 1 : backtrace: backtrace_str,
2520 1 : };
2521 1 : self.set_state(broken_state);
2522 :
2523 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2524 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2525 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2526 1 : self.cancel.cancel();
2527 1 : }
2528 :
2529 115053 : pub(crate) fn current_state(&self) -> TimelineState {
2530 115053 : self.state.borrow().clone()
2531 115053 : }
2532 :
2533 3 : pub(crate) fn is_broken(&self) -> bool {
2534 3 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2535 3 : }
2536 :
2537 126 : pub(crate) fn is_active(&self) -> bool {
2538 126 : self.current_state() == TimelineState::Active
2539 126 : }
2540 :
2541 8 : pub(crate) fn is_archived(&self) -> Option<bool> {
2542 8 : self.remote_client.is_archived()
2543 8 : }
2544 :
2545 8 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2546 8 : self.remote_client.is_invisible()
2547 8 : }
2548 :
2549 569 : pub(crate) fn is_stopping(&self) -> bool {
2550 569 : self.current_state() == TimelineState::Stopping
2551 569 : }
2552 :
2553 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2554 0 : self.state.subscribe()
2555 0 : }
2556 :
2557 114123 : pub(crate) async fn wait_to_become_active(
2558 114123 : &self,
2559 114123 : _ctx: &RequestContext, // Prepare for use by cancellation
2560 114123 : ) -> Result<(), TimelineState> {
2561 114123 : let mut receiver = self.state.subscribe();
2562 : loop {
2563 114123 : let current_state = receiver.borrow().clone();
2564 114123 : match current_state {
2565 : TimelineState::Loading => {
2566 0 : receiver
2567 0 : .changed()
2568 0 : .await
2569 0 : .expect("holding a reference to self");
2570 : }
2571 : TimelineState::Active => {
2572 114122 : return Ok(());
2573 : }
2574 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2575 : // There's no chance the timeline can transition back into ::Active
2576 1 : return Err(current_state);
2577 : }
2578 : }
2579 : }
2580 114123 : }
2581 :
2582 0 : pub(crate) async fn layer_map_info(
2583 0 : &self,
2584 0 : reset: LayerAccessStatsReset,
2585 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2586 0 : let guard = self
2587 0 : .layers
2588 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
2589 0 : .await;
2590 0 : let layer_map = guard.layer_map()?;
2591 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2592 0 : if let Some(open_layer) = &layer_map.open_layer {
2593 0 : in_memory_layers.push(open_layer.info());
2594 0 : }
2595 0 : for frozen_layer in &layer_map.frozen_layers {
2596 0 : in_memory_layers.push(frozen_layer.info());
2597 0 : }
2598 :
2599 0 : let historic_layers = layer_map
2600 0 : .iter_historic_layers()
2601 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2602 0 : .collect();
2603 :
2604 0 : Ok(LayerMapInfo {
2605 0 : in_memory_layers,
2606 0 : historic_layers,
2607 0 : })
2608 0 : }
2609 :
2610 : #[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))]
2611 : pub(crate) async fn download_layer(
2612 : &self,
2613 : layer_file_name: &LayerName,
2614 : ctx: &RequestContext,
2615 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2616 : let Some(layer) = self
2617 : .find_layer(layer_file_name)
2618 : .await
2619 0 : .map_err(|e| match e {
2620 : layer_manager::Shutdown => {
2621 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2622 : }
2623 0 : })?
2624 : else {
2625 : return Ok(None);
2626 : };
2627 :
2628 : layer.download(ctx).await?;
2629 :
2630 : Ok(Some(true))
2631 : }
2632 :
2633 : /// Evict just one layer.
2634 : ///
2635 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2636 0 : pub(crate) async fn evict_layer(
2637 0 : &self,
2638 0 : layer_file_name: &LayerName,
2639 0 : ) -> anyhow::Result<Option<bool>> {
2640 0 : let _gate = self
2641 0 : .gate
2642 0 : .enter()
2643 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2644 :
2645 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2646 0 : return Ok(None);
2647 : };
2648 :
2649 : // curl has this by default
2650 0 : let timeout = std::time::Duration::from_secs(120);
2651 :
2652 0 : match local_layer.evict_and_wait(timeout).await {
2653 0 : Ok(()) => Ok(Some(true)),
2654 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2655 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2656 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2657 : }
2658 0 : }
2659 :
2660 2401506 : fn should_roll(
2661 2401506 : &self,
2662 2401506 : layer_size: u64,
2663 2401506 : projected_layer_size: u64,
2664 2401506 : checkpoint_distance: u64,
2665 2401506 : projected_lsn: Lsn,
2666 2401506 : last_freeze_at: Lsn,
2667 2401506 : opened_at: Instant,
2668 2401506 : ) -> bool {
2669 2401506 : let distance = projected_lsn.widening_sub(last_freeze_at);
2670 :
2671 : // Rolling the open layer can be triggered by:
2672 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2673 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2674 : // account for how writes are distributed across shards: we expect each node to consume
2675 : // 1/count of the LSN on average.
2676 : // 2. The size of the currently open layer.
2677 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2678 : // up and suspend activity.
2679 2401506 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2680 0 : info!(
2681 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2682 : projected_lsn, layer_size, distance
2683 : );
2684 :
2685 0 : true
2686 2401506 : } else if projected_layer_size >= checkpoint_distance {
2687 : // NB: this check is relied upon by:
2688 40 : let _ = IndexEntry::validate_checkpoint_distance;
2689 40 : info!(
2690 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2691 : projected_lsn, layer_size, projected_layer_size
2692 : );
2693 :
2694 40 : true
2695 2401466 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2696 0 : info!(
2697 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2698 : projected_lsn,
2699 : layer_size,
2700 0 : opened_at.elapsed()
2701 : );
2702 :
2703 0 : true
2704 : } else {
2705 2401466 : false
2706 : }
2707 2401506 : }
2708 :
2709 1 : pub(crate) fn is_basebackup_cache_enabled(&self) -> bool {
2710 1 : let tenant_conf = self.tenant_conf.load();
2711 1 : tenant_conf
2712 1 : .tenant_conf
2713 1 : .basebackup_cache_enabled
2714 1 : .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
2715 1 : }
2716 :
2717 : /// Try to get a basebackup from the on-disk cache.
2718 0 : pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
2719 0 : self.basebackup_cache
2720 0 : .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
2721 0 : .await
2722 0 : }
2723 :
2724 : /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
2725 : /// the given request parameters.
2726 : ///
2727 : /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
2728 0 : pub async fn get_cached_basebackup_if_enabled(
2729 0 : &self,
2730 0 : lsn: Option<Lsn>,
2731 0 : prev_lsn: Option<Lsn>,
2732 0 : full: bool,
2733 0 : replica: bool,
2734 0 : gzip: bool,
2735 0 : ) -> Option<tokio::fs::File> {
2736 0 : if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
2737 0 : return None;
2738 0 : }
2739 : // We have to know which LSN to fetch the basebackup for.
2740 0 : let lsn = lsn?;
2741 : // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
2742 0 : if prev_lsn.is_some() || full || replica || !gzip {
2743 0 : return None;
2744 0 : }
2745 0 : self.get_cached_basebackup(lsn).await
2746 0 : }
2747 :
2748 : /// Prepare basebackup for the given LSN and store it in the basebackup cache.
2749 : /// The method is asynchronous and returns immediately.
2750 : /// The actual basebackup preparation is performed in the background
2751 : /// by the basebackup cache on a best-effort basis.
2752 1 : pub(crate) fn prepare_basebackup(&self, lsn: Lsn) {
2753 1 : if !self.is_basebackup_cache_enabled() {
2754 1 : return;
2755 0 : }
2756 0 : if !self.tenant_shard_id.is_shard_zero() {
2757 : // In theory we should never get here, but just in case check it.
2758 : // Preparing basebackup doesn't make sense for shards other than shard zero.
2759 0 : return;
2760 0 : }
2761 0 : if !self.is_active() {
2762 : // May happen during initial timeline creation.
2763 : // Such timeline is not in the global timeline map yet,
2764 : // so basebackup cache will not be able to find it.
2765 : // TODO(diko): We can prepare such timelines in finish_creation().
2766 0 : return;
2767 0 : }
2768 :
2769 0 : self.basebackup_cache
2770 0 : .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
2771 1 : }
2772 : }
2773 :
2774 : /// Number of times we will compute partition within a checkpoint distance.
2775 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2776 :
2777 : // Private functions
2778 : impl Timeline {
2779 6 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2780 6 : let tenant_conf = self.tenant_conf.load();
2781 6 : tenant_conf
2782 6 : .tenant_conf
2783 6 : .lsn_lease_length
2784 6 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2785 6 : }
2786 :
2787 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2788 0 : let tenant_conf = self.tenant_conf.load();
2789 0 : tenant_conf
2790 0 : .tenant_conf
2791 0 : .lsn_lease_length_for_ts
2792 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2793 0 : }
2794 :
2795 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2796 0 : let tenant_conf = self.tenant_conf.load();
2797 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2798 0 : }
2799 :
2800 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2801 0 : let tenant_conf = self.tenant_conf.load();
2802 0 : tenant_conf
2803 0 : .tenant_conf
2804 0 : .lazy_slru_download
2805 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2806 0 : }
2807 :
2808 : /// Checks if a get page request should get perf tracing
2809 : ///
2810 : /// The configuration priority is: tenant config override, default tenant config,
2811 : /// pageserver config.
2812 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2813 0 : let tenant_conf = self.tenant_conf.load();
2814 0 : let ratio = tenant_conf
2815 0 : .tenant_conf
2816 0 : .sampling_ratio
2817 0 : .flatten()
2818 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2819 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2820 :
2821 0 : match ratio {
2822 0 : Some(r) => {
2823 0 : if r.numerator == 0 {
2824 0 : false
2825 : } else {
2826 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2827 : }
2828 : }
2829 0 : None => false,
2830 : }
2831 0 : }
2832 :
2833 2402187 : fn get_checkpoint_distance(&self) -> u64 {
2834 2402187 : let tenant_conf = self.tenant_conf.load();
2835 2402187 : tenant_conf
2836 2402187 : .tenant_conf
2837 2402187 : .checkpoint_distance
2838 2402187 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2839 2402187 : }
2840 :
2841 2401522 : fn get_checkpoint_timeout(&self) -> Duration {
2842 2401522 : let tenant_conf = self.tenant_conf.load();
2843 2401522 : tenant_conf
2844 2401522 : .tenant_conf
2845 2401522 : .checkpoint_timeout
2846 2401522 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2847 2401522 : }
2848 :
2849 1 : pub(crate) fn get_pitr_interval(&self) -> Duration {
2850 1 : let tenant_conf = &self.tenant_conf.load().tenant_conf;
2851 1 : tenant_conf
2852 1 : .pitr_interval
2853 1 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2854 1 : }
2855 :
2856 1273 : fn get_compaction_period(&self) -> Duration {
2857 1273 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2858 1273 : tenant_conf
2859 1273 : .compaction_period
2860 1273 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2861 1273 : }
2862 :
2863 352 : fn get_compaction_target_size(&self) -> u64 {
2864 352 : let tenant_conf = self.tenant_conf.load();
2865 352 : tenant_conf
2866 352 : .tenant_conf
2867 352 : .compaction_target_size
2868 352 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2869 352 : }
2870 :
2871 820 : fn get_compaction_threshold(&self) -> usize {
2872 820 : let tenant_conf = self.tenant_conf.load();
2873 820 : tenant_conf
2874 820 : .tenant_conf
2875 820 : .compaction_threshold
2876 820 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2877 820 : }
2878 :
2879 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2880 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2881 : /// possible that the index part persists the state while the config doesn't get persisted.
2882 973 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2883 973 : let tenant_conf = self.tenant_conf.load();
2884 973 : tenant_conf
2885 973 : .tenant_conf
2886 973 : .rel_size_v2_enabled
2887 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2888 973 : }
2889 :
2890 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2891 1099 : self.rel_size_v2_status
2892 1099 : .load()
2893 1099 : .as_ref()
2894 1099 : .map(|s| s.as_ref().clone())
2895 1099 : .unwrap_or(RelSizeMigration::Legacy)
2896 1099 : }
2897 :
2898 23 : fn get_compaction_upper_limit(&self) -> usize {
2899 23 : let tenant_conf = self.tenant_conf.load();
2900 23 : tenant_conf
2901 23 : .tenant_conf
2902 23 : .compaction_upper_limit
2903 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2904 23 : }
2905 :
2906 0 : pub fn get_compaction_l0_first(&self) -> bool {
2907 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2908 0 : tenant_conf
2909 0 : .compaction_l0_first
2910 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2911 0 : }
2912 :
2913 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2914 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2915 0 : tenant_conf
2916 0 : .compaction_l0_semaphore
2917 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2918 0 : }
2919 :
2920 636 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2921 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2922 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2923 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2924 :
2925 : // If compaction is disabled, don't delay.
2926 636 : if self.get_compaction_period() == Duration::ZERO {
2927 632 : return None;
2928 4 : }
2929 :
2930 4 : let compaction_threshold = self.get_compaction_threshold();
2931 4 : let tenant_conf = self.tenant_conf.load();
2932 4 : let l0_flush_delay_threshold = tenant_conf
2933 4 : .tenant_conf
2934 4 : .l0_flush_delay_threshold
2935 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2936 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2937 :
2938 : // 0 disables backpressure.
2939 4 : if l0_flush_delay_threshold == 0 {
2940 0 : return None;
2941 4 : }
2942 :
2943 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2944 : // backpressure flushes below this.
2945 : // TODO: the tenant config should have validation to prevent this instead.
2946 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2947 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2948 636 : }
2949 :
2950 637 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2951 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2952 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2953 : // sibling timelines. We need more confidence before enabling this.
2954 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2955 :
2956 : // If compaction is disabled, don't stall.
2957 637 : if self.get_compaction_period() == Duration::ZERO {
2958 632 : return None;
2959 5 : }
2960 :
2961 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2962 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2963 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2964 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2965 5 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2966 0 : return None;
2967 5 : }
2968 :
2969 5 : let compaction_threshold = self.get_compaction_threshold();
2970 5 : let tenant_conf = self.tenant_conf.load();
2971 5 : let l0_flush_stall_threshold = tenant_conf
2972 5 : .tenant_conf
2973 5 : .l0_flush_stall_threshold
2974 5 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2975 :
2976 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2977 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2978 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2979 5 : if cfg!(feature = "testing")
2980 5 : && compaction_threshold == 1
2981 0 : && l0_flush_stall_threshold.is_none()
2982 : {
2983 0 : return None;
2984 5 : }
2985 :
2986 5 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2987 5 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2988 :
2989 : // 0 disables backpressure.
2990 5 : if l0_flush_stall_threshold == 0 {
2991 5 : return None;
2992 0 : }
2993 :
2994 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2995 : // backpressure flushes below this.
2996 : // TODO: the tenant config should have validation to prevent this instead.
2997 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2998 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2999 637 : }
3000 :
3001 57 : fn get_image_creation_threshold(&self) -> usize {
3002 57 : let tenant_conf = self.tenant_conf.load();
3003 57 : tenant_conf
3004 57 : .tenant_conf
3005 57 : .image_creation_threshold
3006 57 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
3007 57 : }
3008 :
3009 : // HADRON
3010 193 : fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
3011 193 : let tenant_conf = self.tenant_conf.load();
3012 193 : tenant_conf
3013 193 : .tenant_conf
3014 193 : .image_layer_force_creation_period
3015 193 : .or(self
3016 193 : .conf
3017 193 : .default_tenant_conf
3018 193 : .image_layer_force_creation_period)
3019 193 : }
3020 :
3021 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
3022 192 : let tenant_conf = &self.tenant_conf.load();
3023 192 : tenant_conf
3024 192 : .tenant_conf
3025 192 : .compaction_algorithm
3026 192 : .as_ref()
3027 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
3028 192 : .clone()
3029 192 : }
3030 :
3031 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
3032 192 : let tenant_conf = self.tenant_conf.load();
3033 192 : tenant_conf
3034 192 : .tenant_conf
3035 192 : .compaction_shard_ancestor
3036 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
3037 192 : }
3038 :
3039 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
3040 0 : let tenant_conf = self.tenant_conf.load();
3041 0 : tenant_conf
3042 0 : .tenant_conf
3043 0 : .eviction_policy
3044 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
3045 0 : }
3046 :
3047 235 : fn get_evictions_low_residence_duration_metric_threshold(
3048 235 : tenant_conf: &pageserver_api::models::TenantConfig,
3049 235 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
3050 235 : ) -> Duration {
3051 235 : tenant_conf
3052 235 : .evictions_low_residence_duration_metric_threshold
3053 235 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
3054 235 : }
3055 :
3056 191 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
3057 191 : let tenant_conf = self.tenant_conf.load();
3058 191 : tenant_conf
3059 191 : .tenant_conf
3060 191 : .image_layer_creation_check_threshold
3061 191 : .unwrap_or(
3062 191 : self.conf
3063 191 : .default_tenant_conf
3064 191 : .image_layer_creation_check_threshold,
3065 : )
3066 191 : }
3067 :
3068 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
3069 27 : let tenant_conf = &self.tenant_conf.load();
3070 27 : let gc_compaction_enabled = tenant_conf
3071 27 : .tenant_conf
3072 27 : .gc_compaction_enabled
3073 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
3074 27 : let gc_compaction_verification = tenant_conf
3075 27 : .tenant_conf
3076 27 : .gc_compaction_verification
3077 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
3078 27 : let gc_compaction_initial_threshold_kb = tenant_conf
3079 27 : .tenant_conf
3080 27 : .gc_compaction_initial_threshold_kb
3081 27 : .unwrap_or(
3082 27 : self.conf
3083 27 : .default_tenant_conf
3084 27 : .gc_compaction_initial_threshold_kb,
3085 : );
3086 27 : let gc_compaction_ratio_percent = tenant_conf
3087 27 : .tenant_conf
3088 27 : .gc_compaction_ratio_percent
3089 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
3090 27 : GcCompactionCombinedSettings {
3091 27 : gc_compaction_enabled,
3092 27 : gc_compaction_verification,
3093 27 : gc_compaction_initial_threshold_kb,
3094 27 : gc_compaction_ratio_percent,
3095 27 : }
3096 27 : }
3097 :
3098 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
3099 0 : let tenant_conf = self.tenant_conf.load();
3100 0 : tenant_conf
3101 0 : .tenant_conf
3102 0 : .image_creation_preempt_threshold
3103 0 : .unwrap_or(
3104 0 : self.conf
3105 0 : .default_tenant_conf
3106 0 : .image_creation_preempt_threshold,
3107 : )
3108 0 : }
3109 :
3110 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
3111 : // NB: Most tenant conf options are read by background loops, so,
3112 : // changes will automatically be picked up.
3113 :
3114 : // The threshold is embedded in the metric. So, we need to update it.
3115 : {
3116 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
3117 0 : &new_conf.tenant_conf,
3118 0 : &self.conf.default_tenant_conf,
3119 : );
3120 :
3121 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
3122 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
3123 :
3124 0 : let timeline_id_str = self.timeline_id.to_string();
3125 :
3126 0 : self.remote_client.update_config(&new_conf.location);
3127 :
3128 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
3129 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
3130 0 : if new_capacity != rel_size_cache.capacity() {
3131 0 : rel_size_cache.set_capacity(new_capacity);
3132 0 : }
3133 0 : }
3134 :
3135 0 : self.metrics
3136 0 : .evictions_with_low_residence_duration
3137 0 : .write()
3138 0 : .unwrap()
3139 0 : .change_threshold(
3140 0 : &tenant_id_str,
3141 0 : &shard_id_str,
3142 0 : &timeline_id_str,
3143 0 : new_threshold,
3144 0 : );
3145 : }
3146 0 : }
3147 :
3148 : /// Open a Timeline handle.
3149 : ///
3150 : /// Loads the metadata for the timeline into memory, but not the layer map.
3151 : #[allow(clippy::too_many_arguments)]
3152 235 : pub(super) fn new(
3153 235 : conf: &'static PageServerConf,
3154 235 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
3155 235 : metadata: &TimelineMetadata,
3156 235 : previous_heatmap: Option<PreviousHeatmap>,
3157 235 : ancestor: Option<Arc<Timeline>>,
3158 235 : timeline_id: TimelineId,
3159 235 : tenant_shard_id: TenantShardId,
3160 235 : generation: Generation,
3161 235 : shard_identity: ShardIdentity,
3162 235 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
3163 235 : resources: TimelineResources,
3164 235 : pg_version: PgMajorVersion,
3165 235 : state: TimelineState,
3166 235 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3167 235 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3168 235 : gc_compaction_state: Option<GcCompactionState>,
3169 235 : rel_size_v2_status: Option<RelSizeMigration>,
3170 235 : cancel: CancellationToken,
3171 235 : ) -> Arc<Self> {
3172 235 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3173 235 : let (state, _) = watch::channel(state);
3174 :
3175 235 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3176 235 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3177 :
3178 235 : let evictions_low_residence_duration_metric_threshold = {
3179 235 : let loaded_tenant_conf = tenant_conf.load();
3180 235 : Self::get_evictions_low_residence_duration_metric_threshold(
3181 235 : &loaded_tenant_conf.tenant_conf,
3182 235 : &conf.default_tenant_conf,
3183 : )
3184 : };
3185 :
3186 235 : if let Some(ancestor) = &ancestor {
3187 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3188 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3189 118 : let is_offloaded = MaybeOffloaded::No;
3190 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3191 118 : }
3192 :
3193 235 : let relsize_snapshot_cache_capacity = {
3194 235 : let loaded_tenant_conf = tenant_conf.load();
3195 235 : loaded_tenant_conf
3196 235 : .tenant_conf
3197 235 : .relsize_snapshot_cache_capacity
3198 235 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3199 : };
3200 :
3201 235 : Arc::new_cyclic(|myself| {
3202 235 : let metrics = Arc::new(TimelineMetrics::new(
3203 235 : &tenant_shard_id,
3204 235 : &timeline_id,
3205 235 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3206 : "mtime",
3207 235 : evictions_low_residence_duration_metric_threshold,
3208 : ),
3209 : ));
3210 235 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3211 :
3212 235 : let mut result = Timeline {
3213 235 : conf,
3214 235 : tenant_conf,
3215 235 : myself: myself.clone(),
3216 235 : timeline_id,
3217 235 : tenant_shard_id,
3218 235 : generation,
3219 235 : shard_identity,
3220 235 : pg_version,
3221 235 : layers: Default::default(),
3222 235 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3223 :
3224 235 : walredo_mgr,
3225 235 : walreceiver: Mutex::new(None),
3226 :
3227 235 : remote_client: Arc::new(resources.remote_client),
3228 :
3229 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3230 235 : last_record_lsn: SeqWait::new(RecordLsn {
3231 235 : last: disk_consistent_lsn,
3232 235 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3233 235 : }),
3234 235 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3235 :
3236 235 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
3237 :
3238 235 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3239 235 : last_freeze_ts: RwLock::new(Instant::now()),
3240 :
3241 235 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3242 :
3243 235 : ancestor_timeline: ancestor,
3244 235 : ancestor_lsn: metadata.ancestor_lsn(),
3245 :
3246 235 : metrics,
3247 :
3248 235 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3249 235 : &tenant_shard_id,
3250 235 : &timeline_id,
3251 235 : resources.pagestream_throttle_metrics,
3252 : ),
3253 :
3254 1880 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3255 1880 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3256 :
3257 235 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3258 :
3259 235 : layer_flush_start_tx,
3260 235 : layer_flush_done_tx,
3261 :
3262 235 : write_lock: tokio::sync::Mutex::new(None),
3263 :
3264 235 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3265 :
3266 235 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3267 235 : LastImageLayerCreationStatus::default(),
3268 : )),
3269 :
3270 235 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3271 235 : initdb_lsn: metadata.initdb_lsn(),
3272 :
3273 235 : current_logical_size: if disk_consistent_lsn.is_valid() {
3274 : // we're creating timeline data with some layer files existing locally,
3275 : // need to recalculate timeline's logical size based on data in the layers.
3276 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3277 : } else {
3278 : // we're creating timeline data without any layers existing locally,
3279 : // initial logical size is 0.
3280 115 : LogicalSize::empty_initial()
3281 : },
3282 :
3283 235 : partitioning: GuardArcSwap::new((
3284 235 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3285 235 : Lsn(0),
3286 235 : )),
3287 : repartition_threshold: 0,
3288 235 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3289 235 : last_image_layer_creation_check_instant: Mutex::new(None),
3290 235 : last_received_wal: Mutex::new(None),
3291 235 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3292 235 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3293 :
3294 235 : download_all_remote_layers_task_info: RwLock::new(None),
3295 :
3296 235 : state,
3297 :
3298 235 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3299 235 : EvictionTaskTimelineState::default(),
3300 : ),
3301 235 : delete_progress: TimelineDeleteProgress::default(),
3302 :
3303 235 : cancel,
3304 235 : gate: Gate::default(),
3305 :
3306 235 : compaction_lock: tokio::sync::Mutex::default(),
3307 235 : compaction_failed: AtomicBool::default(),
3308 235 : l0_compaction_trigger: resources.l0_compaction_trigger,
3309 235 : gc_lock: tokio::sync::Mutex::default(),
3310 :
3311 235 : standby_horizon: AtomicLsn::new(0),
3312 :
3313 235 : pagestream_throttle: resources.pagestream_throttle,
3314 :
3315 235 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3316 :
3317 : #[cfg(test)]
3318 235 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3319 :
3320 235 : l0_flush_global_state: resources.l0_flush_global_state,
3321 :
3322 235 : handles: Default::default(),
3323 :
3324 235 : attach_wal_lag_cooldown,
3325 :
3326 235 : create_idempotency,
3327 :
3328 235 : page_trace: Default::default(),
3329 :
3330 235 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3331 :
3332 235 : heatmap_layers_downloader: Mutex::new(None),
3333 :
3334 235 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
3335 :
3336 235 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3337 :
3338 235 : basebackup_cache: resources.basebackup_cache,
3339 :
3340 235 : feature_resolver: resources.feature_resolver.clone(),
3341 : };
3342 :
3343 235 : result.repartition_threshold =
3344 235 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3345 :
3346 235 : result
3347 235 : .metrics
3348 235 : .last_record_lsn_gauge
3349 235 : .set(disk_consistent_lsn.0 as i64);
3350 235 : result
3351 235 : })
3352 235 : }
3353 :
3354 343 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3355 343 : let Ok(guard) = self.gate.enter() else {
3356 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3357 0 : return;
3358 : };
3359 343 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3360 343 : match *flush_loop_state {
3361 232 : FlushLoopState::NotStarted => (),
3362 : FlushLoopState::Running { .. } => {
3363 111 : info!(
3364 0 : "skipping attempt to start flush_loop twice {}/{}",
3365 0 : self.tenant_shard_id, self.timeline_id
3366 : );
3367 111 : return;
3368 : }
3369 : FlushLoopState::Exited => {
3370 0 : info!(
3371 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3372 0 : self.tenant_shard_id, self.timeline_id
3373 : );
3374 0 : return;
3375 : }
3376 : }
3377 :
3378 232 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3379 232 : let self_clone = Arc::clone(self);
3380 :
3381 232 : debug!("spawning flush loop");
3382 232 : *flush_loop_state = FlushLoopState::Running {
3383 232 : #[cfg(test)]
3384 232 : expect_initdb_optimization: false,
3385 232 : #[cfg(test)]
3386 232 : initdb_optimization_count: 0,
3387 232 : };
3388 232 : task_mgr::spawn(
3389 232 : task_mgr::BACKGROUND_RUNTIME.handle(),
3390 232 : task_mgr::TaskKind::LayerFlushTask,
3391 232 : self.tenant_shard_id,
3392 232 : Some(self.timeline_id),
3393 232 : "layer flush task",
3394 232 : async move {
3395 232 : let _guard = guard;
3396 232 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3397 232 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3398 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3399 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3400 5 : *flush_loop_state = FlushLoopState::Exited;
3401 5 : Ok(())
3402 5 : }
3403 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))
3404 : );
3405 343 : }
3406 :
3407 0 : pub(crate) fn update_gc_compaction_state(
3408 0 : &self,
3409 0 : gc_compaction_state: GcCompactionState,
3410 0 : ) -> anyhow::Result<()> {
3411 0 : self.gc_compaction_state
3412 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3413 0 : self.remote_client
3414 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3415 0 : }
3416 :
3417 0 : pub(crate) fn update_rel_size_v2_status(
3418 0 : &self,
3419 0 : rel_size_v2_status: RelSizeMigration,
3420 0 : ) -> anyhow::Result<()> {
3421 0 : self.rel_size_v2_status
3422 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3423 0 : self.remote_client
3424 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3425 0 : }
3426 :
3427 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3428 0 : self.gc_compaction_state.load_full().as_ref().clone()
3429 0 : }
3430 :
3431 : /// Creates and starts the wal receiver.
3432 : ///
3433 : /// This function is expected to be called at most once per Timeline's lifecycle
3434 : /// when the timeline is activated.
3435 0 : fn launch_wal_receiver(
3436 0 : self: &Arc<Self>,
3437 0 : ctx: &RequestContext,
3438 0 : broker_client: BrokerClientChannel,
3439 0 : ) {
3440 0 : info!(
3441 0 : "launching WAL receiver for timeline {} of tenant {}",
3442 0 : self.timeline_id, self.tenant_shard_id
3443 : );
3444 :
3445 0 : let tenant_conf = self.tenant_conf.load();
3446 0 : let wal_connect_timeout = tenant_conf
3447 0 : .tenant_conf
3448 0 : .walreceiver_connect_timeout
3449 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3450 0 : let lagging_wal_timeout = tenant_conf
3451 0 : .tenant_conf
3452 0 : .lagging_wal_timeout
3453 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3454 0 : let max_lsn_wal_lag = tenant_conf
3455 0 : .tenant_conf
3456 0 : .max_lsn_wal_lag
3457 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3458 :
3459 0 : let mut guard = self.walreceiver.lock().unwrap();
3460 0 : assert!(
3461 0 : guard.is_none(),
3462 0 : "multiple launches / re-launches of WAL receiver are not supported"
3463 : );
3464 :
3465 0 : let protocol = PostgresClientProtocol::Interpreted {
3466 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3467 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3468 0 : };
3469 :
3470 0 : *guard = Some(WalReceiver::start(
3471 0 : Arc::clone(self),
3472 0 : WalReceiverConf {
3473 0 : protocol,
3474 0 : wal_connect_timeout,
3475 0 : lagging_wal_timeout,
3476 0 : max_lsn_wal_lag,
3477 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3478 0 : availability_zone: self.conf.availability_zone.clone(),
3479 0 : ingest_batch_size: self.conf.ingest_batch_size,
3480 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3481 0 : },
3482 0 : broker_client,
3483 0 : ctx,
3484 0 : ));
3485 0 : }
3486 :
3487 : /// Initialize with an empty layer map. Used when creating a new timeline.
3488 232 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3489 232 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3490 232 : "in the context where we call this function, no other task has access to the object",
3491 : );
3492 232 : layers
3493 232 : .open_mut()
3494 232 : .expect("in this context the LayerManager must still be open")
3495 232 : .initialize_empty(Lsn(start_lsn.0));
3496 232 : }
3497 :
3498 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3499 : /// files.
3500 3 : pub(super) async fn load_layer_map(
3501 3 : &self,
3502 3 : disk_consistent_lsn: Lsn,
3503 3 : index_part: IndexPart,
3504 3 : ) -> anyhow::Result<()> {
3505 : use LayerName::*;
3506 : use init::Decision::*;
3507 : use init::{Discovered, DismissedLayer};
3508 :
3509 3 : let mut guard = self
3510 3 : .layers
3511 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3512 3 : .await;
3513 :
3514 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3515 :
3516 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3517 : // structs representing all files on disk
3518 3 : let timeline_path = self
3519 3 : .conf
3520 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3521 3 : let conf = self.conf;
3522 3 : let span = tracing::Span::current();
3523 :
3524 : // Copy to move into the task we're about to spawn
3525 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3526 :
3527 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3528 3 : move || {
3529 3 : let _g = span.entered();
3530 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3531 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3532 3 : let mut unrecognized_files = Vec::new();
3533 :
3534 3 : let mut path = timeline_path;
3535 :
3536 11 : for discovered in discovered {
3537 8 : let (name, kind) = match discovered {
3538 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3539 8 : discovered_layers.push((layer_file_name, local_metadata));
3540 8 : continue;
3541 : }
3542 0 : Discovered::IgnoredBackup(path) => {
3543 0 : std::fs::remove_file(path)
3544 0 : .or_else(fs_ext::ignore_not_found)
3545 0 : .fatal_err("Removing .old file");
3546 0 : continue;
3547 : }
3548 0 : Discovered::Unknown(file_name) => {
3549 : // we will later error if there are any
3550 0 : unrecognized_files.push(file_name);
3551 0 : continue;
3552 : }
3553 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3554 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3555 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3556 : };
3557 0 : path.push(Utf8Path::new(&name));
3558 0 : init::cleanup(&path, kind)?;
3559 0 : path.pop();
3560 : }
3561 :
3562 3 : if !unrecognized_files.is_empty() {
3563 : // assume that if there are any there are many many.
3564 0 : let n = unrecognized_files.len();
3565 0 : let first = &unrecognized_files[..n.min(10)];
3566 0 : anyhow::bail!(
3567 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3568 : );
3569 3 : }
3570 :
3571 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3572 :
3573 3 : let mut loaded_layers = Vec::new();
3574 3 : let mut needs_cleanup = Vec::new();
3575 3 : let mut total_physical_size = 0;
3576 :
3577 11 : for (name, decision) in decided {
3578 8 : let decision = match decision {
3579 8 : Ok(decision) => decision,
3580 0 : Err(DismissedLayer::Future { local }) => {
3581 0 : if let Some(local) = local {
3582 0 : init::cleanup_future_layer(
3583 0 : &local.local_path,
3584 0 : &name,
3585 0 : disk_consistent_lsn,
3586 0 : )?;
3587 0 : }
3588 0 : needs_cleanup.push(name);
3589 0 : continue;
3590 : }
3591 0 : Err(DismissedLayer::LocalOnly(local)) => {
3592 0 : init::cleanup_local_only_file(&name, &local)?;
3593 : // this file never existed remotely, we will have to do rework
3594 0 : continue;
3595 : }
3596 0 : Err(DismissedLayer::BadMetadata(local)) => {
3597 0 : init::cleanup_local_file_for_remote(&local)?;
3598 : // this file never existed remotely, we will have to do rework
3599 0 : continue;
3600 : }
3601 : };
3602 :
3603 8 : match &name {
3604 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3605 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3606 : }
3607 :
3608 8 : tracing::debug!(layer=%name, ?decision, "applied");
3609 :
3610 8 : let layer = match decision {
3611 8 : Resident { local, remote } => {
3612 8 : total_physical_size += local.file_size;
3613 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3614 8 : .drop_eviction_guard()
3615 : }
3616 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3617 : };
3618 :
3619 8 : loaded_layers.push(layer);
3620 : }
3621 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3622 3 : }
3623 : })
3624 3 : .await
3625 3 : .map_err(anyhow::Error::new)
3626 3 : .and_then(|x| x)?;
3627 :
3628 3 : let num_layers = loaded_layers.len();
3629 :
3630 3 : guard
3631 3 : .open_mut()
3632 3 : .expect("layermanager must be open during init")
3633 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3634 :
3635 3 : self.remote_client
3636 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3637 3 : self.remote_client
3638 3 : .schedule_index_upload_for_file_changes()?;
3639 : // This barrier orders above DELETEs before any later operations.
3640 : // This is critical because code executing after the barrier might
3641 : // create again objects with the same key that we just scheduled for deletion.
3642 : // For example, if we just scheduled deletion of an image layer "from the future",
3643 : // later compaction might run again and re-create the same image layer.
3644 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3645 : // "same" here means same key range and LSN.
3646 : //
3647 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3648 : // the upload queue may execute the PUT first, then the DELETE.
3649 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3650 : //
3651 : // 1. a future image layer is created and uploaded
3652 : // 2. ps restart
3653 : // 3. the future layer from (1) is deleted during load layer map
3654 : // 4. image layer is re-created and uploaded
3655 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3656 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3657 : //
3658 : // See https://github.com/neondatabase/neon/issues/5878
3659 : //
3660 : // NB: generation numbers naturally protect against this because they disambiguate
3661 : // (1) and (4)
3662 : // TODO: this is basically a no-op now, should we remove it?
3663 3 : self.remote_client.schedule_barrier()?;
3664 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3665 : // on retry.
3666 :
3667 3 : info!(
3668 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3669 : num_layers, disk_consistent_lsn, total_physical_size
3670 : );
3671 :
3672 3 : timer.stop_and_record();
3673 3 : Ok(())
3674 3 : }
3675 :
3676 : /// Retrieve current logical size of the timeline.
3677 : ///
3678 : /// The size could be lagging behind the actual number, in case
3679 : /// the initial size calculation has not been run (gets triggered on the first size access).
3680 : ///
3681 : /// return size and boolean flag that shows if the size is exact
3682 0 : pub(crate) fn get_current_logical_size(
3683 0 : self: &Arc<Self>,
3684 0 : priority: GetLogicalSizePriority,
3685 0 : ctx: &RequestContext,
3686 0 : ) -> logical_size::CurrentLogicalSize {
3687 0 : if !self.tenant_shard_id.is_shard_zero() {
3688 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3689 : // when HTTP API is serving a GET for timeline zero, return zero
3690 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3691 0 : }
3692 :
3693 0 : let current_size = self.current_logical_size.current_size();
3694 0 : debug!("Current size: {current_size:?}");
3695 :
3696 0 : match (current_size.accuracy(), priority) {
3697 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3698 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3699 0 : // background task will eventually deliver an exact value, we're in no rush
3700 0 : }
3701 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3702 : // background task is not ready, but user is asking for it now;
3703 : // => make the background task skip the line
3704 : // (The alternative would be to calculate the size here, but,
3705 : // it can actually take a long time if the user has a lot of rels.
3706 : // And we'll inevitable need it again; So, let the background task do the work.)
3707 0 : match self
3708 0 : .current_logical_size
3709 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3710 0 : .get()
3711 : {
3712 0 : Some(cancel) => cancel.cancel(),
3713 : None => {
3714 0 : match self.current_state() {
3715 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3716 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3717 0 : // Don't make noise.
3718 0 : }
3719 : TimelineState::Loading => {
3720 : // Import does not return an activated timeline.
3721 0 : info!(
3722 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3723 : );
3724 : }
3725 : TimelineState::Active => {
3726 : // activation should be setting the once cell
3727 0 : warn!(
3728 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3729 : );
3730 0 : debug_assert!(false);
3731 : }
3732 : }
3733 : }
3734 : }
3735 : }
3736 : }
3737 :
3738 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3739 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3740 0 : let first = self
3741 0 : .current_logical_size
3742 0 : .did_return_approximate_to_walreceiver
3743 0 : .compare_exchange(
3744 0 : false,
3745 0 : true,
3746 0 : AtomicOrdering::Relaxed,
3747 0 : AtomicOrdering::Relaxed,
3748 0 : )
3749 0 : .is_ok();
3750 0 : if first {
3751 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3752 0 : }
3753 0 : }
3754 0 : }
3755 :
3756 0 : current_size
3757 0 : }
3758 :
3759 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3760 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3761 : // nothing to do for freshly created timelines;
3762 0 : assert_eq!(
3763 0 : self.current_logical_size.current_size().accuracy(),
3764 : logical_size::Accuracy::Exact,
3765 : );
3766 0 : self.current_logical_size.initialized.add_permits(1);
3767 0 : return;
3768 : };
3769 :
3770 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3771 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3772 0 : self.current_logical_size
3773 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3774 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3775 :
3776 0 : let self_clone = Arc::clone(self);
3777 0 : let background_ctx = ctx.detached_child(
3778 0 : TaskKind::InitialLogicalSizeCalculation,
3779 0 : DownloadBehavior::Download,
3780 : );
3781 0 : task_mgr::spawn(
3782 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3783 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3784 0 : self.tenant_shard_id,
3785 0 : Some(self.timeline_id),
3786 0 : "initial size calculation",
3787 : // NB: don't log errors here, task_mgr will do that.
3788 0 : async move {
3789 0 : self_clone
3790 0 : .initial_logical_size_calculation_task(
3791 0 : initial_part_end,
3792 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3793 0 : background_ctx,
3794 0 : )
3795 0 : .await;
3796 0 : Ok(())
3797 0 : }
3798 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)),
3799 : );
3800 0 : }
3801 :
3802 : /// # Cancellation
3803 : ///
3804 : /// This method is sensitive to `Timeline::cancel`.
3805 : ///
3806 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3807 : ///
3808 : /// # Cancel-Safety
3809 : ///
3810 : /// It does Timeline IO, hence this should be polled to completion because
3811 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3812 : /// to reason about.
3813 0 : async fn initial_logical_size_calculation_task(
3814 0 : self: Arc<Self>,
3815 0 : initial_part_end: Lsn,
3816 0 : skip_concurrency_limiter: CancellationToken,
3817 0 : background_ctx: RequestContext,
3818 0 : ) {
3819 0 : scopeguard::defer! {
3820 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3821 : self.current_logical_size.initialized.add_permits(1);
3822 : }
3823 :
3824 0 : let try_once = |attempt: usize| {
3825 0 : let background_ctx = &background_ctx;
3826 0 : let self_ref = &self;
3827 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3828 0 : async move {
3829 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3830 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3831 0 : background_ctx,
3832 : );
3833 :
3834 : use crate::metrics::initial_logical_size::StartCircumstances;
3835 0 : let (_maybe_permit, circumstances) = tokio::select! {
3836 0 : permit = wait_for_permit => {
3837 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3838 : }
3839 0 : _ = self_ref.cancel.cancelled() => {
3840 0 : return Err(CalculateLogicalSizeError::Cancelled);
3841 : }
3842 0 : () = skip_concurrency_limiter.cancelled() => {
3843 : // Some action that is part of a end user interaction requested logical size
3844 : // => break out of the rate limit
3845 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3846 : // but then again what happens if they cancel; also, we should just be using
3847 : // one runtime across the entire process, so, let's leave this for now.
3848 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3849 : }
3850 : };
3851 :
3852 0 : let metrics_guard = if attempt == 1 {
3853 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3854 : } else {
3855 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3856 : };
3857 :
3858 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3859 0 : self_ref.conf.get_vectored_concurrent_io,
3860 0 : self_ref
3861 0 : .gate
3862 0 : .enter()
3863 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3864 : );
3865 :
3866 0 : let calculated_size = self_ref
3867 0 : .logical_size_calculation_task(
3868 0 : initial_part_end,
3869 0 : LogicalSizeCalculationCause::Initial,
3870 0 : background_ctx,
3871 0 : )
3872 0 : .await?;
3873 :
3874 0 : self_ref
3875 0 : .trigger_aux_file_size_computation(
3876 0 : initial_part_end,
3877 0 : background_ctx,
3878 0 : io_concurrency,
3879 0 : )
3880 0 : .await?;
3881 :
3882 : // TODO: add aux file size to logical size
3883 :
3884 0 : Ok((calculated_size, metrics_guard))
3885 0 : }
3886 0 : };
3887 :
3888 0 : let retrying = async {
3889 0 : let mut attempt = 0;
3890 : loop {
3891 0 : attempt += 1;
3892 :
3893 0 : match try_once(attempt).await {
3894 0 : Ok(res) => return ControlFlow::Continue(res),
3895 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3896 : Err(
3897 0 : e @ (CalculateLogicalSizeError::Decode(_)
3898 : | CalculateLogicalSizeError::PageRead(_)),
3899 : ) => {
3900 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3901 : // exponential back-off doesn't make sense at these long intervals;
3902 : // use fixed retry interval with generous jitter instead
3903 0 : let sleep_duration = Duration::from_secs(
3904 0 : u64::try_from(
3905 : // 1hour base
3906 0 : (60_i64 * 60_i64)
3907 0 : // 10min jitter
3908 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3909 : )
3910 0 : .expect("10min < 1hour"),
3911 : );
3912 0 : tokio::select! {
3913 0 : _ = tokio::time::sleep(sleep_duration) => {}
3914 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3915 : }
3916 : }
3917 : }
3918 : }
3919 0 : };
3920 :
3921 0 : let (calculated_size, metrics_guard) = match retrying.await {
3922 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3923 0 : ControlFlow::Break(()) => return,
3924 : };
3925 :
3926 : // we cannot query current_logical_size.current_size() to know the current
3927 : // *negative* value, only truncated to u64.
3928 0 : let added = self
3929 0 : .current_logical_size
3930 0 : .size_added_after_initial
3931 0 : .load(AtomicOrdering::Relaxed);
3932 :
3933 0 : let sum = calculated_size.saturating_add_signed(added);
3934 :
3935 : // set the gauge value before it can be set in `update_current_logical_size`.
3936 0 : self.metrics.current_logical_size_gauge.set(sum);
3937 :
3938 0 : self.current_logical_size
3939 0 : .initial_logical_size
3940 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3941 0 : .ok()
3942 0 : .expect("only this task sets it");
3943 0 : }
3944 :
3945 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3946 7 : self: &Arc<Self>,
3947 7 : lsn: Lsn,
3948 7 : cause: LogicalSizeCalculationCause,
3949 7 : ctx: RequestContext,
3950 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3951 7 : let (sender, receiver) = oneshot::channel();
3952 7 : let self_clone = Arc::clone(self);
3953 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3954 : // we should stop the size calculation work and return an error.
3955 : // That would require restructuring this function's API to
3956 : // return the result directly, instead of a Receiver for the result.
3957 7 : let ctx = ctx.detached_child(
3958 7 : TaskKind::OndemandLogicalSizeCalculation,
3959 7 : DownloadBehavior::Download,
3960 : );
3961 7 : task_mgr::spawn(
3962 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3963 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3964 7 : self.tenant_shard_id,
3965 7 : Some(self.timeline_id),
3966 7 : "ondemand logical size calculation",
3967 7 : async move {
3968 7 : let res = self_clone
3969 7 : .logical_size_calculation_task(lsn, cause, &ctx)
3970 7 : .await;
3971 7 : let _ = sender.send(res).ok();
3972 7 : Ok(()) // Receiver is responsible for handling errors
3973 7 : }
3974 7 : .in_current_span(),
3975 : );
3976 7 : receiver
3977 7 : }
3978 :
3979 : #[instrument(skip_all)]
3980 : async fn logical_size_calculation_task(
3981 : self: &Arc<Self>,
3982 : lsn: Lsn,
3983 : cause: LogicalSizeCalculationCause,
3984 : ctx: &RequestContext,
3985 : ) -> Result<u64, CalculateLogicalSizeError> {
3986 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3987 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3988 : // accurate relation sizes, and they do not emit consumption metrics.
3989 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3990 :
3991 : let guard = self
3992 : .gate
3993 : .enter()
3994 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3995 :
3996 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3997 : }
3998 :
3999 : /// Calculate the logical size of the database at the latest LSN.
4000 : ///
4001 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
4002 : /// especially if we need to download remote layers.
4003 7 : async fn calculate_logical_size(
4004 7 : &self,
4005 7 : up_to_lsn: Lsn,
4006 7 : cause: LogicalSizeCalculationCause,
4007 7 : _guard: &GateGuard,
4008 7 : ctx: &RequestContext,
4009 7 : ) -> Result<u64, CalculateLogicalSizeError> {
4010 7 : info!(
4011 0 : "Calculating logical size for timeline {} at {}",
4012 : self.timeline_id, up_to_lsn
4013 : );
4014 :
4015 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
4016 : {
4017 0 : return Err(CalculateLogicalSizeError::Cancelled);
4018 7 : }
4019 :
4020 : // See if we've already done the work for initial size calculation.
4021 : // This is a short-cut for timelines that are mostly unused.
4022 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
4023 0 : return Ok(size);
4024 7 : }
4025 7 : let storage_time_metrics = match cause {
4026 : LogicalSizeCalculationCause::Initial
4027 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
4028 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
4029 : LogicalSizeCalculationCause::EvictionTaskImitation => {
4030 7 : &self.metrics.imitate_logical_size_histo
4031 : }
4032 : };
4033 7 : let timer = storage_time_metrics.start_timer();
4034 7 : let logical_size = self
4035 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
4036 7 : .await?;
4037 7 : debug!("calculated logical size: {logical_size}");
4038 7 : timer.stop_and_record();
4039 7 : Ok(logical_size)
4040 7 : }
4041 :
4042 : /// Update current logical size, adding `delta' to the old value.
4043 135285 : fn update_current_logical_size(&self, delta: i64) {
4044 135285 : let logical_size = &self.current_logical_size;
4045 135285 : logical_size.increment_size(delta);
4046 :
4047 : // Also set the value in the prometheus gauge. Note that
4048 : // there is a race condition here: if this is is called by two
4049 : // threads concurrently, the prometheus gauge might be set to
4050 : // one value while current_logical_size is set to the
4051 : // other.
4052 135285 : match logical_size.current_size() {
4053 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
4054 135285 : .metrics
4055 135285 : .current_logical_size_gauge
4056 135285 : .set(new_current_size.into()),
4057 0 : CurrentLogicalSize::Approximate(_) => {
4058 0 : // don't update the gauge yet, this allows us not to update the gauge back and
4059 0 : // forth between the initial size calculation task.
4060 0 : }
4061 : }
4062 135285 : }
4063 :
4064 1527 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
4065 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
4066 : // for each of the database, but we only store one value, and therefore each pgdirmodification
4067 : // would overwrite the previous value if they modify different databases.
4068 :
4069 1527 : match count {
4070 566 : MetricsUpdate::Set(count) => {
4071 566 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
4072 566 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
4073 566 : }
4074 960 : MetricsUpdate::Add(count) => {
4075 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4076 : // it's fine.
4077 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4078 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
4079 960 : // the value reliably.
4080 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
4081 960 : }
4082 : // Otherwise, ignore this update
4083 : }
4084 1 : MetricsUpdate::Sub(count) => {
4085 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
4086 : // it's fine.
4087 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
4088 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
4089 1 : // The operation could overflow so we need to normalize the value.
4090 1 : let prev_val =
4091 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
4092 1 : let res = prev_val.saturating_sub(count);
4093 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
4094 1 : }
4095 : // Otherwise, ignore this update
4096 : }
4097 : };
4098 :
4099 : // TODO: remove this, there's no place in the code that updates this aux metrics.
4100 1527 : let aux_metric =
4101 1527 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
4102 :
4103 1527 : let sum_of_entries = self
4104 1527 : .directory_metrics
4105 1527 : .iter()
4106 12216 : .map(|v| v.load(AtomicOrdering::Relaxed))
4107 1527 : .sum();
4108 : // Set a high general threshold and a lower threshold for the auxiliary files,
4109 : // as we can have large numbers of relations in the db directory.
4110 : const SUM_THRESHOLD: u64 = 5000;
4111 : const AUX_THRESHOLD: u64 = 1000;
4112 1527 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
4113 0 : self.metrics
4114 0 : .directory_entries_count_gauge
4115 0 : .set(sum_of_entries);
4116 1527 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
4117 0 : metric.set(sum_of_entries);
4118 1527 : }
4119 1527 : }
4120 :
4121 0 : async fn find_layer(
4122 0 : &self,
4123 0 : layer_name: &LayerName,
4124 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
4125 0 : let guard = self
4126 0 : .layers
4127 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
4128 0 : .await;
4129 0 : let layer = guard
4130 0 : .layer_map()?
4131 0 : .iter_historic_layers()
4132 0 : .find(|l| &l.layer_name() == layer_name)
4133 0 : .map(|found| guard.get_from_desc(&found));
4134 0 : Ok(layer)
4135 0 : }
4136 :
4137 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
4138 0 : let crnt = self.previous_heatmap.load();
4139 0 : match crnt.as_deref() {
4140 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
4141 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
4142 0 : None => true,
4143 : },
4144 0 : Some(PreviousHeatmap::Obsolete) => false,
4145 0 : None => false,
4146 : }
4147 0 : }
4148 :
4149 : /// The timeline heatmap is a hint to secondary locations from the primary location,
4150 : /// indicating which layers are currently on-disk on the primary.
4151 : ///
4152 : /// None is returned if the Timeline is in a state where uploading a heatmap
4153 : /// doesn't make sense, such as shutting down or initializing. The caller
4154 : /// should treat this as a cue to simply skip doing any heatmap uploading
4155 : /// for this timeline.
4156 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
4157 8 : if !self.is_active() {
4158 0 : return None;
4159 8 : }
4160 :
4161 8 : let guard = self
4162 8 : .layers
4163 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
4164 8 : .await;
4165 :
4166 : // Firstly, if there's any heatmap left over from when this location
4167 : // was a secondary, take that into account. Keep layers that are:
4168 : // * present in the layer map
4169 : // * visible
4170 : // * non-resident
4171 : // * not evicted since we read the heatmap
4172 : //
4173 : // Without this, a new cold, attached location would clobber the previous
4174 : // heatamp.
4175 8 : let previous_heatmap = self.previous_heatmap.load();
4176 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4177 : Some(PreviousHeatmap::Active {
4178 6 : heatmap, read_at, ..
4179 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4180 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4181 23 : let layer = guard.try_get_from_key(&desc.key())?;
4182 :
4183 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4184 0 : return None;
4185 23 : }
4186 :
4187 23 : if layer.is_likely_resident() {
4188 10 : return None;
4189 13 : }
4190 :
4191 13 : if layer.last_evicted_at().happened_after(*read_at) {
4192 3 : return None;
4193 10 : }
4194 :
4195 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4196 23 : })),
4197 0 : Some(PreviousHeatmap::Obsolete) => None,
4198 2 : None => None,
4199 : };
4200 :
4201 : // Secondly, all currently visible, resident layers are included.
4202 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4203 18 : match layer.visibility() {
4204 : LayerVisibilityHint::Visible => {
4205 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4206 17 : let last_activity_ts = layer.latest_activity();
4207 17 : Some((
4208 17 : layer.layer_desc().clone(),
4209 17 : layer.metadata(),
4210 17 : last_activity_ts,
4211 17 : false, // these layers are not cold
4212 17 : ))
4213 : }
4214 : LayerVisibilityHint::Covered => {
4215 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4216 1 : None
4217 : }
4218 : }
4219 18 : });
4220 :
4221 8 : let mut layers = match visible_non_resident {
4222 6 : Some(non_resident) => {
4223 6 : let mut non_resident = non_resident.peekable();
4224 6 : if non_resident.peek().is_none() {
4225 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4226 2 : self.previous_heatmap
4227 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4228 4 : }
4229 :
4230 6 : non_resident.chain(resident).collect::<Vec<_>>()
4231 : }
4232 2 : None => resident.collect::<Vec<_>>(),
4233 : };
4234 :
4235 : // Sort layers in order of which to download first. For a large set of layers to download, we
4236 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4237 : // or hours later:
4238 : // - Cold layers go last for convenience when a human inspects the heatmap.
4239 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4240 : // only exist for a few minutes before being compacted into L1s.
4241 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4242 : // the layer is likely to be covered by an image layer during compaction.
4243 60 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4244 60 : std::cmp::Reverse((
4245 60 : *cold,
4246 60 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4247 60 : desc.lsn_range.end,
4248 60 : ))
4249 60 : });
4250 :
4251 8 : let layers = layers
4252 8 : .into_iter()
4253 27 : .map(|(desc, meta, atime, cold)| {
4254 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4255 27 : })
4256 8 : .collect();
4257 :
4258 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4259 8 : }
4260 :
4261 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4262 0 : let guard = self
4263 0 : .layers
4264 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4265 0 : .await;
4266 :
4267 0 : let now = SystemTime::now();
4268 0 : let mut heatmap_layers = Vec::default();
4269 0 : for vl in guard.visible_layers() {
4270 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4271 0 : continue;
4272 0 : }
4273 :
4274 0 : let hl = HeatMapLayer {
4275 0 : name: vl.layer_desc().layer_name(),
4276 0 : metadata: vl.metadata(),
4277 0 : access_time: now,
4278 0 : cold: true,
4279 0 : };
4280 0 : heatmap_layers.push(hl);
4281 : }
4282 :
4283 0 : tracing::info!(
4284 0 : "Generating unarchival heatmap with {} layers",
4285 0 : heatmap_layers.len()
4286 : );
4287 :
4288 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4289 0 : PreviousHeatmap::Active {
4290 0 : heatmap,
4291 0 : read_at: Instant::now(),
4292 0 : end_lsn: Some(end_lsn),
4293 0 : }
4294 0 : }
4295 :
4296 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4297 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4298 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4299 : // branchpoint in the value in IndexPart::lineage
4300 0 : self.ancestor_lsn == lsn
4301 0 : || (self.ancestor_lsn == Lsn::INVALID
4302 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4303 0 : }
4304 : }
4305 :
4306 : #[derive(Clone)]
4307 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4308 : /// In other words, a set of segments in a 2D space.
4309 : ///
4310 : /// This representation has the advatange of avoiding hash map
4311 : /// allocations for uniform queries.
4312 : pub(crate) enum VersionedKeySpaceQuery {
4313 : /// Variant for queries at a single [`Lsn`]
4314 : Uniform { keyspace: KeySpace, lsn: Lsn },
4315 : /// Variant for queries at multiple [`Lsn`]s
4316 : Scattered {
4317 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4318 : },
4319 : }
4320 :
4321 : impl VersionedKeySpaceQuery {
4322 302238 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4323 302238 : Self::Uniform { keyspace, lsn }
4324 302238 : }
4325 :
4326 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4327 10192 : Self::Scattered { keyspaces_at_lsn }
4328 10192 : }
4329 :
4330 : /// Returns the most recent (largest) LSN included in the query.
4331 : /// If any of the LSNs included in the query are invalid, returns
4332 : /// an error instead.
4333 624860 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4334 624860 : match self {
4335 604476 : Self::Uniform { lsn, .. } => {
4336 604476 : if !lsn.is_valid() {
4337 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4338 604476 : }
4339 :
4340 604476 : Ok(*lsn)
4341 : }
4342 20384 : Self::Scattered { keyspaces_at_lsn } => {
4343 20384 : let mut max_lsn = None;
4344 42218 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4345 42218 : if !lsn.is_valid() {
4346 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4347 42218 : }
4348 42218 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4349 : }
4350 :
4351 20384 : if let Some(computed) = max_lsn {
4352 20384 : Ok(*computed)
4353 : } else {
4354 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4355 : }
4356 : }
4357 : }
4358 624860 : }
4359 :
4360 : /// Returns the total keyspace being queried: the result of projecting
4361 : /// everything in the key dimensions onto the key axis.
4362 323507 : fn total_keyspace(&self) -> KeySpace {
4363 323507 : match self {
4364 303123 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4365 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4366 20384 : .iter()
4367 20384 : .map(|(_lsn, keyspace)| keyspace)
4368 42218 : .fold(KeySpace::default(), |mut acc, v| {
4369 42218 : acc.merge(v);
4370 42218 : acc
4371 42218 : }),
4372 : }
4373 323507 : }
4374 :
4375 : /// Returns LSN for a specific key.
4376 : ///
4377 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4378 395551 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4379 395551 : match self {
4380 322359 : Self::Uniform { lsn, .. } => *lsn,
4381 73192 : Self::Scattered { keyspaces_at_lsn } => {
4382 73192 : keyspaces_at_lsn
4383 73192 : .iter()
4384 416770 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4385 73192 : .expect("Returned key was requested")
4386 : .0
4387 : }
4388 : }
4389 395551 : }
4390 :
4391 : /// Remove any parts of the query (segments) which overlap with the provided
4392 : /// key space (also segments).
4393 967227 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4394 967227 : match self {
4395 946843 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4396 20384 : Self::Scattered { keyspaces_at_lsn } => {
4397 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4398 42218 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4399 42218 : let removed = keyspace.remove_overlapping_with(to_remove);
4400 42218 : removed_accum.add_keyspace(removed);
4401 42218 : });
4402 :
4403 20384 : removed_accum.to_keyspace()
4404 : }
4405 : }
4406 967227 : }
4407 :
4408 738982 : fn is_empty(&self) -> bool {
4409 738982 : match self {
4410 718598 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4411 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4412 20384 : .iter()
4413 31301 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4414 : }
4415 738982 : }
4416 :
4417 : /// "Lower" the query on the LSN dimension
4418 114123 : fn lower(&mut self, to: Lsn) {
4419 114123 : match self {
4420 114123 : Self::Uniform { lsn, .. } => {
4421 114123 : // If the originally requested LSN is smaller than the starting
4422 114123 : // LSN of the ancestor we are descending into, we need to respect that.
4423 114123 : // Hence the min.
4424 114123 : *lsn = std::cmp::min(*lsn, to);
4425 114123 : }
4426 0 : Self::Scattered { keyspaces_at_lsn } => {
4427 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4428 0 : *lsn = std::cmp::min(*lsn, to);
4429 0 : });
4430 : }
4431 : }
4432 114123 : }
4433 : }
4434 :
4435 : impl std::fmt::Display for VersionedKeySpaceQuery {
4436 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4437 0 : write!(f, "[")?;
4438 :
4439 0 : match self {
4440 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4441 0 : write!(f, "{keyspace} @ {lsn}")?;
4442 : }
4443 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4444 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4445 0 : write!(f, "{keyspace} @ {lsn},")?;
4446 : }
4447 : }
4448 : }
4449 :
4450 0 : write!(f, "]")
4451 0 : }
4452 : }
4453 :
4454 : impl Timeline {
4455 : #[allow(clippy::doc_lazy_continuation)]
4456 : /// Get the data needed to reconstruct all keys in the provided keyspace
4457 : ///
4458 : /// The algorithm is as follows:
4459 : /// 1. While some keys are still not done and there's a timeline to visit:
4460 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4461 : /// 2.1: Build the fringe for the current keyspace
4462 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4463 : /// intersects
4464 : /// 2.3. Pop the timeline from the fringe
4465 : /// 2.4. If the fringe is empty, go back to 1
4466 312430 : async fn get_vectored_reconstruct_data(
4467 312430 : &self,
4468 312430 : mut query: VersionedKeySpaceQuery,
4469 312430 : reconstruct_state: &mut ValuesReconstructState,
4470 312430 : ctx: &RequestContext,
4471 312430 : ) -> Result<(), GetVectoredError> {
4472 312430 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4473 :
4474 : let mut timeline_owned: Arc<Timeline>;
4475 312430 : let mut timeline = self;
4476 :
4477 312429 : let missing_keyspace = loop {
4478 426552 : if self.cancel.is_cancelled() {
4479 0 : return Err(GetVectoredError::Cancelled);
4480 426552 : }
4481 :
4482 : let TimelineVisitOutcome {
4483 426552 : completed_keyspace: completed,
4484 426552 : image_covered_keyspace,
4485 : } = {
4486 426552 : let ctx = RequestContextBuilder::from(ctx)
4487 426552 : .perf_span(|crnt_perf_span| {
4488 0 : info_span!(
4489 : target: PERF_TRACE_TARGET,
4490 0 : parent: crnt_perf_span,
4491 : "PLAN_IO_TIMELINE",
4492 : timeline = %timeline.timeline_id,
4493 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4494 : )
4495 0 : })
4496 426552 : .attached_child();
4497 :
4498 426552 : Self::get_vectored_reconstruct_data_timeline(
4499 426552 : timeline,
4500 426552 : &query,
4501 426552 : reconstruct_state,
4502 426552 : &self.cancel,
4503 426552 : &ctx,
4504 : )
4505 426552 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4506 426552 : .await?
4507 : };
4508 :
4509 426552 : query.remove_overlapping_with(&completed);
4510 :
4511 : // Do not descend into the ancestor timeline for aux files.
4512 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4513 : // stalling compaction.
4514 426552 : query.remove_overlapping_with(&KeySpace {
4515 426552 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4516 426552 : });
4517 :
4518 : // Keyspace is fully retrieved
4519 426552 : if query.is_empty() {
4520 312246 : break None;
4521 114306 : }
4522 :
4523 114306 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4524 : // Not fully retrieved but no ancestor timeline.
4525 183 : break Some(query.total_keyspace());
4526 : };
4527 :
4528 : // Now we see if there are keys covered by the image layer but does not exist in the
4529 : // image layer, which means that the key does not exist.
4530 :
4531 : // The block below will stop the vectored search if any of the keys encountered an image layer
4532 : // which did not contain a snapshot for said key. Since we have already removed all completed
4533 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4534 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4535 : // and stop the search as a result of that.
4536 114123 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4537 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4538 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4539 : // figuring out what is the inherited key range and do a fine-grained pruning.
4540 114123 : removed.remove_overlapping_with(&KeySpace {
4541 114123 : ranges: vec![SPARSE_RANGE],
4542 114123 : });
4543 114123 : if !removed.is_empty() {
4544 0 : break Some(removed);
4545 114123 : }
4546 :
4547 : // Each key range in the original query is at some point in the LSN space.
4548 : // When descending into the ancestor, lower all ranges in the LSN space
4549 : // such that new changes on the parent timeline are not visible.
4550 114123 : query.lower(timeline.ancestor_lsn);
4551 :
4552 114123 : let ctx = RequestContextBuilder::from(ctx)
4553 114123 : .perf_span(|crnt_perf_span| {
4554 0 : info_span!(
4555 : target: PERF_TRACE_TARGET,
4556 0 : parent: crnt_perf_span,
4557 : "GET_ANCESTOR",
4558 : timeline = %timeline.timeline_id,
4559 0 : ancestor = %ancestor_timeline.timeline_id,
4560 : ancestor_lsn = %timeline.ancestor_lsn
4561 : )
4562 0 : })
4563 114123 : .attached_child();
4564 :
4565 114123 : timeline_owned = timeline
4566 114123 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4567 114123 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4568 114123 : .await?;
4569 114122 : timeline = &*timeline_owned;
4570 : };
4571 :
4572 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4573 312429 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4574 183 : let mut missing_keyspace = missing_keyspace;
4575 183 : missing_keyspace.remove_overlapping_with(&KeySpace {
4576 183 : ranges: vec![SPARSE_RANGE],
4577 183 : });
4578 183 : if missing_keyspace.is_empty() {
4579 176 : None
4580 : } else {
4581 7 : Some(missing_keyspace)
4582 : }
4583 : } else {
4584 312246 : None
4585 : };
4586 :
4587 312429 : if let Some(missing_keyspace) = missing_keyspace {
4588 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4589 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4590 7 : shard: self.shard_identity.number,
4591 7 : original_hwm_lsn,
4592 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4593 7 : backtrace: None,
4594 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4595 7 : query: None,
4596 7 : })));
4597 312422 : }
4598 :
4599 312422 : Ok(())
4600 312430 : }
4601 :
4602 426552 : async fn get_vectored_init_fringe(
4603 426552 : &self,
4604 426552 : query: &VersionedKeySpaceQuery,
4605 426552 : ) -> Result<LayerFringe, GetVectoredError> {
4606 426552 : let mut fringe = LayerFringe::new();
4607 426552 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4608 :
4609 426552 : match query {
4610 416360 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4611 : // LSNs requested by the compute or determined by the pageserver
4612 : // are inclusive. Queries to the layer map use exclusive LSNs.
4613 : // Hence, bump the value before the query - same in the other
4614 : // match arm.
4615 416360 : let cont_lsn = Lsn(lsn.0 + 1);
4616 416360 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4617 : }
4618 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4619 21109 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4620 21109 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4621 21109 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4622 : }
4623 : }
4624 : }
4625 :
4626 426552 : Ok(fringe)
4627 426552 : }
4628 :
4629 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4630 : ///
4631 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4632 : /// the current keyspace. The current keyspace of the search at any given timeline
4633 : /// is the original keyspace minus all the keys that have been completed minus
4634 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4635 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4636 : ///
4637 : /// This is basically a depth-first search visitor implementation where a vertex
4638 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4639 : ///
4640 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4641 : /// and get all the required reconstruct data from the layer in one go.
4642 : ///
4643 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4644 : /// decides how to deal with these two keyspaces.
4645 426552 : async fn get_vectored_reconstruct_data_timeline(
4646 426552 : timeline: &Timeline,
4647 426552 : query: &VersionedKeySpaceQuery,
4648 426552 : reconstruct_state: &mut ValuesReconstructState,
4649 426552 : cancel: &CancellationToken,
4650 426552 : ctx: &RequestContext,
4651 426552 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4652 : // Prevent GC from progressing while visiting the current timeline.
4653 : // If we are GC-ing because a new image layer was added while traversing
4654 : // the timeline, then it will remove layers that are required for fulfilling
4655 : // the current get request (read-path cannot "look back" and notice the new
4656 : // image layer).
4657 426552 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4658 :
4659 : // See `compaction::compact_with_gc` for why we need this.
4660 426552 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4661 :
4662 : // Initialize the fringe
4663 426552 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4664 :
4665 426552 : let mut completed_keyspace = KeySpace::default();
4666 426552 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4667 :
4668 873479 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4669 446927 : if cancel.is_cancelled() {
4670 0 : return Err(GetVectoredError::Cancelled);
4671 446927 : }
4672 :
4673 446927 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4674 446927 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4675 446927 : }
4676 :
4677 : // Visit the layer and plan IOs for it
4678 446927 : let next_cont_lsn = lsn_range.start;
4679 446927 : layer_to_read
4680 446927 : .get_values_reconstruct_data(
4681 446927 : keyspace_to_read.clone(),
4682 446927 : lsn_range,
4683 446927 : reconstruct_state,
4684 446927 : ctx,
4685 446927 : )
4686 446927 : .await?;
4687 :
4688 446927 : let mut unmapped_keyspace = keyspace_to_read;
4689 446927 : let cont_lsn = next_cont_lsn;
4690 :
4691 446927 : reconstruct_state.on_layer_visited(&layer_to_read);
4692 :
4693 446927 : let (keys_done_last_step, keys_with_image_coverage) =
4694 446927 : reconstruct_state.consume_done_keys();
4695 446927 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4696 446927 : completed_keyspace.merge(&keys_done_last_step);
4697 446927 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4698 14120 : unmapped_keyspace
4699 14120 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4700 14120 : image_covered_keyspace.add_range(keys_with_image_coverage);
4701 432807 : }
4702 :
4703 : // Query the layer map for the next layers to read.
4704 : //
4705 : // Do not descent any further if the last layer we visited
4706 : // completed all keys in the keyspace it inspected. This is not
4707 : // required for correctness, but avoids visiting extra layers
4708 : // which turns out to be a perf bottleneck in some cases.
4709 446927 : if !unmapped_keyspace.is_empty() {
4710 130229 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4711 130229 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4712 :
4713 : // It's safe to drop the layer map lock after planning the next round of reads.
4714 : // The fringe keeps readable handles for the layers which are safe to read even
4715 : // if layers were compacted or flushed.
4716 : //
4717 : // The more interesting consideration is: "Why is the read algorithm still correct
4718 : // if the layer map changes while it is operating?". Doing a vectored read on a
4719 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4720 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4721 : // range at *a particular point in time*. It is fine for the answer to be different
4722 : // at two different time points.
4723 130229 : drop(guard);
4724 316698 : }
4725 : }
4726 :
4727 426552 : Ok(TimelineVisitOutcome {
4728 426552 : completed_keyspace,
4729 426552 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4730 426552 : })
4731 426552 : }
4732 :
4733 114123 : async fn get_ready_ancestor_timeline(
4734 114123 : &self,
4735 114123 : ancestor: &Arc<Timeline>,
4736 114123 : ctx: &RequestContext,
4737 114123 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4738 : // It's possible that the ancestor timeline isn't active yet, or
4739 : // is active but hasn't yet caught up to the branch point. Wait
4740 : // for it.
4741 : //
4742 : // This cannot happen while the pageserver is running normally,
4743 : // because you cannot create a branch from a point that isn't
4744 : // present in the pageserver yet. However, we don't wait for the
4745 : // branch point to be uploaded to cloud storage before creating
4746 : // a branch. I.e., the branch LSN need not be remote consistent
4747 : // for the branching operation to succeed.
4748 : //
4749 : // Hence, if we try to load a tenant in such a state where
4750 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4751 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4752 : // then we will need to wait for the ancestor timeline to
4753 : // re-stream WAL up to branch_lsn before we access it.
4754 : //
4755 : // How can a tenant get in such a state?
4756 : // - ungraceful pageserver process exit
4757 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4758 : //
4759 : // NB: this could be avoided by requiring
4760 : // branch_lsn >= remote_consistent_lsn
4761 : // during branch creation.
4762 114123 : match ancestor.wait_to_become_active(ctx).await {
4763 114122 : Ok(()) => {}
4764 : Err(TimelineState::Stopping) => {
4765 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4766 0 : return Err(GetReadyAncestorError::Cancelled);
4767 : }
4768 1 : Err(state) => {
4769 1 : return Err(GetReadyAncestorError::BadState {
4770 1 : timeline_id: ancestor.timeline_id,
4771 1 : state,
4772 1 : });
4773 : }
4774 : }
4775 114122 : ancestor
4776 114122 : .wait_lsn(
4777 114122 : self.ancestor_lsn,
4778 114122 : WaitLsnWaiter::Timeline(self),
4779 114122 : WaitLsnTimeout::Default,
4780 114122 : ctx,
4781 114122 : )
4782 114122 : .await
4783 114122 : .map_err(|e| match e {
4784 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4785 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4786 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4787 0 : timeline_id: ancestor.timeline_id,
4788 0 : state,
4789 0 : },
4790 0 : })?;
4791 :
4792 114122 : Ok(ancestor.clone())
4793 114123 : }
4794 :
4795 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4796 148592 : &self.shard_identity
4797 148592 : }
4798 :
4799 : #[inline(always)]
4800 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4801 0 : ShardTimelineId {
4802 0 : shard_index: ShardIndex {
4803 0 : shard_number: self.shard_identity.number,
4804 0 : shard_count: self.shard_identity.count,
4805 0 : },
4806 0 : timeline_id: self.timeline_id,
4807 0 : }
4808 0 : }
4809 :
4810 : /// Returns a non-frozen open in-memory layer for ingestion.
4811 : ///
4812 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4813 : /// this function without holding the mutex.
4814 660 : async fn get_layer_for_write(
4815 660 : &self,
4816 660 : lsn: Lsn,
4817 660 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4818 660 : ctx: &RequestContext,
4819 660 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4820 660 : let mut guard = self
4821 660 : .layers
4822 660 : .write(LayerManagerLockHolder::GetLayerForWrite)
4823 660 : .await;
4824 :
4825 660 : let last_record_lsn = self.get_last_record_lsn();
4826 660 : ensure!(
4827 660 : lsn > last_record_lsn,
4828 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4829 : lsn,
4830 : last_record_lsn,
4831 : );
4832 :
4833 660 : let layer = guard
4834 660 : .open_mut()?
4835 660 : .get_layer_for_write(
4836 660 : lsn,
4837 660 : self.conf,
4838 660 : self.timeline_id,
4839 660 : self.tenant_shard_id,
4840 660 : &self.gate,
4841 660 : &self.cancel,
4842 660 : ctx,
4843 : )
4844 660 : .await?;
4845 660 : Ok(layer)
4846 660 : }
4847 :
4848 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4849 2639560 : assert!(new_lsn.is_aligned());
4850 :
4851 2639560 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4852 2639560 : self.last_record_lsn.advance(new_lsn);
4853 2639560 : }
4854 :
4855 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4856 : ///
4857 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4858 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4859 611 : async fn freeze_inmem_layer_at(
4860 611 : &self,
4861 611 : at: Lsn,
4862 611 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4863 611 : ) -> Result<u64, FlushLayerError> {
4864 611 : let frozen = {
4865 611 : let mut guard = self
4866 611 : .layers
4867 611 : .write(LayerManagerLockHolder::TryFreezeLayer)
4868 611 : .await;
4869 611 : guard
4870 611 : .open_mut()?
4871 611 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4872 611 : .await
4873 : };
4874 :
4875 611 : if frozen {
4876 597 : let now = Instant::now();
4877 597 : *(self.last_freeze_ts.write().unwrap()) = now;
4878 597 : }
4879 :
4880 : // Increment the flush cycle counter and wake up the flush task.
4881 : // Remember the new value, so that when we listen for the flush
4882 : // to finish, we know when the flush that we initiated has
4883 : // finished, instead of some other flush that was started earlier.
4884 611 : let mut my_flush_request = 0;
4885 :
4886 611 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4887 611 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4888 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4889 611 : }
4890 :
4891 611 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4892 611 : my_flush_request = *counter + 1;
4893 611 : *counter = my_flush_request;
4894 611 : *lsn = std::cmp::max(at, *lsn);
4895 611 : });
4896 :
4897 611 : assert_ne!(my_flush_request, 0);
4898 :
4899 611 : Ok(my_flush_request)
4900 611 : }
4901 :
4902 : /// Layer flusher task's main loop.
4903 232 : async fn flush_loop(
4904 232 : self: &Arc<Self>,
4905 232 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4906 232 : ctx: &RequestContext,
4907 232 : ) {
4908 : // Always notify waiters about the flush loop exiting since the loop might stop
4909 : // when the timeline hasn't been cancelled.
4910 232 : let scopeguard_rx = layer_flush_start_rx.clone();
4911 232 : scopeguard::defer! {
4912 : let (flush_counter, _) = *scopeguard_rx.borrow();
4913 : let _ = self
4914 : .layer_flush_done_tx
4915 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4916 : }
4917 :
4918 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4919 232 : let mut watch_l0 = match self
4920 232 : .layers
4921 232 : .read(LayerManagerLockHolder::FlushLoop)
4922 232 : .await
4923 232 : .layer_map()
4924 : {
4925 232 : Ok(lm) => lm.watch_level0_deltas(),
4926 0 : Err(Shutdown) => return,
4927 : };
4928 :
4929 232 : info!("started flush loop");
4930 : loop {
4931 833 : tokio::select! {
4932 833 : _ = self.cancel.cancelled() => {
4933 5 : info!("shutting down layer flush task due to Timeline::cancel");
4934 5 : break;
4935 : },
4936 833 : _ = layer_flush_start_rx.changed() => {}
4937 : }
4938 601 : trace!("waking up");
4939 601 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4940 :
4941 : // The highest LSN to which we flushed in the loop over frozen layers
4942 601 : let mut flushed_to_lsn = Lsn(0);
4943 :
4944 601 : let result = loop {
4945 1197 : if self.cancel.is_cancelled() {
4946 0 : info!("dropping out of flush loop for timeline shutdown");
4947 0 : return;
4948 1197 : }
4949 :
4950 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4951 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4952 1197 : if flushed_to_lsn >= frozen_to_lsn {
4953 586 : break Ok(());
4954 611 : }
4955 :
4956 : // Fetch the next layer to flush, if any.
4957 611 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4958 611 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4959 611 : let Ok(lm) = layers.layer_map() else {
4960 0 : info!("dropping out of flush loop for timeline shutdown");
4961 0 : return;
4962 : };
4963 611 : let l0_count = lm.level0_deltas().len();
4964 611 : let frozen_count = lm.frozen_layers.len();
4965 611 : let frozen_size: u64 = lm
4966 611 : .frozen_layers
4967 611 : .iter()
4968 611 : .map(|l| l.estimated_in_mem_size())
4969 611 : .sum();
4970 611 : let open_layer_size: u64 = lm
4971 611 : .open_layer
4972 611 : .as_ref()
4973 611 : .map(|l| l.estimated_in_mem_size())
4974 611 : .unwrap_or(0);
4975 611 : let layer = lm.frozen_layers.front().cloned();
4976 611 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
4977 : // drop 'layers' lock
4978 : };
4979 611 : let Some(layer) = layer else {
4980 14 : break Ok(());
4981 : };
4982 :
4983 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4984 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4985 597 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4986 0 : if l0_count >= stall_threshold {
4987 0 : warn!(
4988 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4989 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4990 : );
4991 0 : let stall_timer = self
4992 0 : .metrics
4993 0 : .flush_delay_histo
4994 0 : .start_timer()
4995 0 : .record_on_drop();
4996 0 : tokio::select! {
4997 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4998 0 : if let Ok(l0) = result.as_deref() {
4999 0 : let delay = stall_timer.elapsed().as_secs_f64();
5000 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
5001 0 : }
5002 : },
5003 0 : _ = self.cancel.cancelled() => {},
5004 : }
5005 0 : continue; // check again
5006 0 : }
5007 597 : }
5008 :
5009 : // Flush the layer.
5010 597 : let flush_timer = self.metrics.flush_time_histo.start_timer();
5011 597 : match self.flush_frozen_layer(layer, ctx).await {
5012 596 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
5013 : Err(FlushLayerError::Cancelled) => {
5014 0 : info!("dropping out of flush loop for timeline shutdown");
5015 0 : return;
5016 : }
5017 1 : err @ Err(
5018 : FlushLayerError::NotRunning(_)
5019 : | FlushLayerError::Other(_)
5020 : | FlushLayerError::CreateImageLayersError(_),
5021 : ) => {
5022 1 : error!("could not flush frozen layer: {err:?}");
5023 1 : break err.map(|_| ());
5024 : }
5025 : }
5026 596 : let flush_duration = flush_timer.stop_and_record();
5027 :
5028 : // Notify the tenant compaction loop if L0 compaction is needed.
5029 596 : let l0_count = *watch_l0.borrow();
5030 596 : if l0_count >= self.get_compaction_threshold() {
5031 238 : self.l0_compaction_trigger.notify_one();
5032 358 : }
5033 :
5034 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
5035 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
5036 : // ingestion by having ephemeral layer rolls wait for flushes.
5037 596 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
5038 4 : if l0_count >= delay_threshold {
5039 0 : let delay = flush_duration.as_secs_f64();
5040 0 : info!(
5041 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
5042 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
5043 : );
5044 0 : let _delay_timer = self
5045 0 : .metrics
5046 0 : .flush_delay_histo
5047 0 : .start_timer()
5048 0 : .record_on_drop();
5049 0 : tokio::select! {
5050 0 : _ = tokio::time::sleep(flush_duration) => {},
5051 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
5052 0 : _ = self.cancel.cancelled() => {},
5053 : }
5054 4 : }
5055 592 : }
5056 : };
5057 :
5058 : // Unsharded tenants should never advance their LSN beyond the end of the
5059 : // highest layer they write: such gaps between layer data and the frozen LSN
5060 : // are only legal on sharded tenants.
5061 601 : debug_assert!(
5062 601 : self.shard_identity.count.count() > 1
5063 594 : || flushed_to_lsn >= frozen_to_lsn
5064 14 : || !flushed_to_lsn.is_valid()
5065 : );
5066 :
5067 601 : if flushed_to_lsn < frozen_to_lsn
5068 15 : && self.shard_identity.count.count() > 1
5069 1 : && result.is_ok()
5070 : {
5071 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
5072 : // to us via layer_flush_start_rx, then advance it here.
5073 : //
5074 : // This path is only taken for tenants with multiple shards: single sharded tenants should
5075 : // never encounter a gap in the wal.
5076 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
5077 0 : tracing::debug!(
5078 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
5079 : );
5080 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
5081 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
5082 0 : tracing::warn!(
5083 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
5084 : );
5085 0 : }
5086 0 : }
5087 601 : }
5088 :
5089 : // Notify any listeners that we're done
5090 601 : let _ = self
5091 601 : .layer_flush_done_tx
5092 601 : .send_replace((flush_counter, result));
5093 : }
5094 5 : }
5095 :
5096 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
5097 571 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
5098 571 : let mut rx = self.layer_flush_done_tx.subscribe();
5099 : loop {
5100 : {
5101 1159 : let (last_result_counter, last_result) = &*rx.borrow();
5102 1159 : if *last_result_counter >= request {
5103 571 : if let Err(err) = last_result {
5104 : // We already logged the original error in
5105 : // flush_loop. We cannot propagate it to the caller
5106 : // here, because it might not be Cloneable
5107 1 : return Err(err.clone());
5108 : } else {
5109 570 : return Ok(());
5110 : }
5111 588 : }
5112 : }
5113 588 : trace!("waiting for flush to complete");
5114 588 : tokio::select! {
5115 588 : rx_e = rx.changed() => {
5116 588 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
5117 : },
5118 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
5119 : // the notification from [`flush_loop`] that it completed.
5120 588 : _ = self.cancel.cancelled() => {
5121 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
5122 0 : return Ok(())
5123 : }
5124 : };
5125 588 : trace!("done")
5126 : }
5127 571 : }
5128 :
5129 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
5130 : ///
5131 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
5132 : #[instrument(skip_all, fields(layer=%frozen_layer))]
5133 : async fn flush_frozen_layer(
5134 : self: &Arc<Self>,
5135 : frozen_layer: Arc<InMemoryLayer>,
5136 : ctx: &RequestContext,
5137 : ) -> Result<Lsn, FlushLayerError> {
5138 : debug_assert_current_span_has_tenant_and_timeline_id();
5139 :
5140 : // As a special case, when we have just imported an image into the repository,
5141 : // instead of writing out a L0 delta layer, we directly write out image layer
5142 : // files instead. This is possible as long as *all* the data imported into the
5143 : // repository have the same LSN.
5144 : let lsn_range = frozen_layer.get_lsn_range();
5145 :
5146 : // Whether to directly create image layers for this flush, or flush them as delta layers
5147 : let create_image_layer =
5148 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
5149 :
5150 : #[cfg(test)]
5151 : {
5152 : match &mut *self.flush_loop_state.lock().unwrap() {
5153 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
5154 : panic!("flush loop not running")
5155 : }
5156 : FlushLoopState::Running {
5157 : expect_initdb_optimization,
5158 : initdb_optimization_count,
5159 : ..
5160 : } => {
5161 : if create_image_layer {
5162 : *initdb_optimization_count += 1;
5163 : } else {
5164 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
5165 : }
5166 : }
5167 : }
5168 : }
5169 :
5170 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5171 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5172 : // require downloading anything during initial import.
5173 : let ((rel_partition, metadata_partition), _lsn) = self
5174 : .repartition(
5175 : self.initdb_lsn,
5176 : self.get_compaction_target_size(),
5177 : EnumSet::empty(),
5178 : ctx,
5179 : )
5180 : .await
5181 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5182 :
5183 : if self.cancel.is_cancelled() {
5184 : return Err(FlushLayerError::Cancelled);
5185 : }
5186 :
5187 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5188 : // So that the key ranges don't overlap.
5189 : let mut partitions = KeyPartitioning::default();
5190 : partitions.parts.extend(rel_partition.parts);
5191 : if !metadata_partition.parts.is_empty() {
5192 : assert_eq!(
5193 : metadata_partition.parts.len(),
5194 : 1,
5195 : "currently sparse keyspace should only contain a single metadata keyspace"
5196 : );
5197 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5198 : // every single key within the keyspace, and therefore, it's safe to force converting it
5199 : // into a dense keyspace before calling this function.
5200 : partitions
5201 : .parts
5202 : .extend(metadata_partition.into_dense().parts);
5203 : }
5204 :
5205 : let mut layers_to_upload = Vec::new();
5206 : let (generated_image_layers, is_complete) = self
5207 : .create_image_layers(
5208 : &partitions,
5209 : self.initdb_lsn,
5210 : None,
5211 : ImageLayerCreationMode::Initial,
5212 : ctx,
5213 : LastImageLayerCreationStatus::Initial,
5214 : false, // don't yield for L0, we're flushing L0
5215 : )
5216 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5217 : .await?;
5218 : debug_assert!(
5219 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5220 : "init image generation mode must fully cover the keyspace"
5221 : );
5222 : layers_to_upload.extend(generated_image_layers);
5223 :
5224 : (layers_to_upload, None)
5225 : } else {
5226 : // Normal case, write out a L0 delta layer file.
5227 : // `create_delta_layer` will not modify the layer map.
5228 : // We will remove frozen layer and add delta layer in one atomic operation later.
5229 : let Some(layer) = self
5230 : .create_delta_layer(&frozen_layer, None, ctx)
5231 : .await
5232 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5233 : else {
5234 : panic!("delta layer cannot be empty if no filter is applied");
5235 : };
5236 : (
5237 : // FIXME: even though we have a single image and single delta layer assumption
5238 : // we push them to vec
5239 : vec![layer.clone()],
5240 : Some(layer),
5241 : )
5242 : };
5243 :
5244 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5245 :
5246 : if self.cancel.is_cancelled() {
5247 : return Err(FlushLayerError::Cancelled);
5248 : }
5249 :
5250 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5251 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5252 1 : });
5253 :
5254 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5255 :
5256 : // The new on-disk layers are now in the layer map. We can remove the
5257 : // in-memory layer from the map now. The flushed layer is stored in
5258 : // the mapping in `create_delta_layer`.
5259 : {
5260 : let mut guard = self
5261 : .layers
5262 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5263 : .await;
5264 :
5265 : guard.open_mut()?.finish_flush_l0_layer(
5266 : delta_layer_to_add.as_ref(),
5267 : &frozen_layer,
5268 : &self.metrics,
5269 : );
5270 :
5271 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5272 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5273 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5274 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5275 : }
5276 : // release lock on 'layers'
5277 : };
5278 :
5279 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5280 : // a compaction can delete the file and then it won't be available for uploads any more.
5281 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5282 : // race situation.
5283 : // See https://github.com/neondatabase/neon/issues/4526
5284 : pausable_failpoint!("flush-frozen-pausable");
5285 :
5286 : // This failpoint is used by another test case `test_pageserver_recovery`.
5287 : fail_point!("flush-frozen-exit");
5288 :
5289 : Ok(Lsn(lsn_range.end.0 - 1))
5290 : }
5291 :
5292 : /// Return true if the value changed
5293 : ///
5294 : /// This function must only be used from the layer flush task.
5295 596 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5296 596 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5297 596 : assert!(
5298 596 : new_value >= old_value,
5299 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5300 : );
5301 :
5302 596 : self.metrics
5303 596 : .disk_consistent_lsn_gauge
5304 596 : .set(new_value.0 as i64);
5305 596 : new_value != old_value
5306 596 : }
5307 :
5308 : /// Update metadata file
5309 623 : fn schedule_uploads(
5310 623 : &self,
5311 623 : disk_consistent_lsn: Lsn,
5312 623 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5313 623 : ) -> anyhow::Result<()> {
5314 : // We can only save a valid 'prev_record_lsn' value on disk if we
5315 : // flushed *all* in-memory changes to disk. We only track
5316 : // 'prev_record_lsn' in memory for the latest processed record, so we
5317 : // don't remember what the correct value that corresponds to some old
5318 : // LSN is. But if we flush everything, then the value corresponding
5319 : // current 'last_record_lsn' is correct and we can store it on disk.
5320 : let RecordLsn {
5321 623 : last: last_record_lsn,
5322 623 : prev: prev_record_lsn,
5323 623 : } = self.last_record_lsn.load();
5324 623 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5325 563 : Some(prev_record_lsn)
5326 : } else {
5327 60 : None
5328 : };
5329 :
5330 623 : let update = crate::tenant::metadata::MetadataUpdate::new(
5331 623 : disk_consistent_lsn,
5332 623 : ondisk_prev_record_lsn,
5333 623 : *self.applied_gc_cutoff_lsn.read(),
5334 : );
5335 :
5336 623 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5337 0 : "{}",
5338 0 : x.unwrap()
5339 : ));
5340 :
5341 1225 : for layer in layers_to_upload {
5342 602 : self.remote_client.schedule_layer_file_upload(layer)?;
5343 : }
5344 623 : self.remote_client
5345 623 : .schedule_index_upload_for_metadata_update(&update)?;
5346 :
5347 623 : Ok(())
5348 623 : }
5349 :
5350 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5351 0 : self.remote_client
5352 0 : .preserve_initdb_archive(
5353 0 : &self.tenant_shard_id.tenant_id,
5354 0 : &self.timeline_id,
5355 0 : &self.cancel,
5356 0 : )
5357 0 : .await
5358 0 : }
5359 :
5360 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5361 : // in layer map immediately. The caller is responsible to put it into the layer map.
5362 486 : async fn create_delta_layer(
5363 486 : self: &Arc<Self>,
5364 486 : frozen_layer: &Arc<InMemoryLayer>,
5365 486 : key_range: Option<Range<Key>>,
5366 486 : ctx: &RequestContext,
5367 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5368 486 : let self_clone = Arc::clone(self);
5369 486 : let frozen_layer = Arc::clone(frozen_layer);
5370 486 : let ctx = ctx.attached_child();
5371 486 : let work = async move {
5372 486 : let Some((desc, path)) = frozen_layer
5373 486 : .write_to_disk(
5374 486 : &ctx,
5375 486 : key_range,
5376 486 : self_clone.l0_flush_global_state.inner(),
5377 486 : &self_clone.gate,
5378 486 : self_clone.cancel.clone(),
5379 486 : )
5380 486 : .await?
5381 : else {
5382 0 : return Ok(None);
5383 : };
5384 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5385 :
5386 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5387 : // We just need to fsync the directory in which these inodes are linked,
5388 : // which we know to be the timeline directory.
5389 : //
5390 : // We use fatal_err() below because the after write_to_disk returns with success,
5391 : // the in-memory state of the filesystem already has the layer file in its final place,
5392 : // and subsequent pageserver code could think it's durable while it really isn't.
5393 486 : let timeline_dir = VirtualFile::open(
5394 486 : &self_clone
5395 486 : .conf
5396 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5397 486 : &ctx,
5398 486 : )
5399 486 : .await
5400 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5401 486 : timeline_dir
5402 486 : .sync_all()
5403 486 : .await
5404 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5405 486 : anyhow::Ok(Some(new_delta))
5406 486 : };
5407 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5408 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5409 : use crate::virtual_file::io_engine::IoEngine;
5410 486 : match crate::virtual_file::io_engine::get() {
5411 0 : IoEngine::NotSet => panic!("io engine not set"),
5412 : IoEngine::StdFs => {
5413 0 : let span = tracing::info_span!("blocking");
5414 0 : tokio::task::spawn_blocking({
5415 0 : move || Handle::current().block_on(work.instrument(span))
5416 : })
5417 0 : .await
5418 0 : .context("spawn_blocking")
5419 0 : .and_then(|x| x)
5420 : }
5421 : #[cfg(target_os = "linux")]
5422 486 : IoEngine::TokioEpollUring => work.await,
5423 : }
5424 486 : }
5425 :
5426 303 : async fn repartition(
5427 303 : &self,
5428 303 : lsn: Lsn,
5429 303 : partition_size: u64,
5430 303 : flags: EnumSet<CompactFlags>,
5431 303 : ctx: &RequestContext,
5432 303 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5433 303 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5434 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5435 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5436 : // and hence before the compaction task starts.
5437 0 : return Err(RepartitionError::Other(anyhow!(
5438 0 : "repartition() called concurrently"
5439 0 : )));
5440 : };
5441 303 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5442 303 : if lsn < *partition_lsn {
5443 0 : return Err(RepartitionError::Other(anyhow!(
5444 0 : "repartition() called with LSN going backwards, this should not happen"
5445 0 : )));
5446 303 : }
5447 :
5448 303 : let distance = lsn.0 - partition_lsn.0;
5449 303 : if *partition_lsn != Lsn(0)
5450 141 : && distance <= self.repartition_threshold
5451 141 : && !flags.contains(CompactFlags::ForceRepartition)
5452 : {
5453 134 : debug!(
5454 : distance,
5455 : threshold = self.repartition_threshold,
5456 0 : "no repartitioning needed"
5457 : );
5458 134 : return Ok((
5459 134 : (dense_partition.clone(), sparse_partition.clone()),
5460 134 : *partition_lsn,
5461 134 : ));
5462 169 : }
5463 :
5464 169 : let (dense_ks, sparse_ks) = self
5465 169 : .collect_keyspace(lsn, ctx)
5466 169 : .await
5467 169 : .map_err(RepartitionError::CollectKeyspace)?;
5468 169 : let dense_partitioning = dense_ks.partition(
5469 169 : &self.shard_identity,
5470 169 : partition_size,
5471 169 : postgres_ffi::BLCKSZ as u64,
5472 : );
5473 169 : let sparse_partitioning = SparseKeyPartitioning {
5474 169 : parts: vec![sparse_ks],
5475 169 : }; // no partitioning for metadata keys for now
5476 169 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5477 169 : guard.write(result.clone());
5478 169 : Ok(result)
5479 303 : }
5480 :
5481 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5482 57 : async fn time_for_new_image_layer(
5483 57 : &self,
5484 57 : partition: &KeySpace,
5485 57 : lsn: Lsn,
5486 57 : force_image_creation_lsn: Option<Lsn>,
5487 57 : ) -> bool {
5488 57 : let threshold = self.get_image_creation_threshold();
5489 :
5490 57 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5491 57 : let Ok(layers) = guard.layer_map() else {
5492 0 : return false;
5493 : };
5494 57 : let mut min_image_lsn: Lsn = Lsn::MAX;
5495 57 : let mut max_deltas = 0;
5496 364 : for part_range in &partition.ranges {
5497 307 : let image_coverage = layers.image_coverage(part_range, lsn);
5498 614 : for (img_range, last_img) in image_coverage {
5499 307 : let img_lsn = if let Some(last_img) = last_img {
5500 0 : last_img.get_lsn_range().end
5501 : } else {
5502 307 : Lsn(0)
5503 : };
5504 : // Let's consider an example:
5505 : //
5506 : // delta layer with LSN range 71-81
5507 : // delta layer with LSN range 81-91
5508 : // delta layer with LSN range 91-101
5509 : // image layer at LSN 100
5510 : //
5511 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5512 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5513 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5514 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5515 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5516 307 : if img_lsn < lsn {
5517 307 : let num_deltas =
5518 307 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5519 :
5520 307 : max_deltas = max_deltas.max(num_deltas);
5521 307 : if num_deltas >= threshold {
5522 0 : debug!(
5523 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5524 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5525 : );
5526 0 : return true;
5527 307 : }
5528 0 : }
5529 307 : min_image_lsn = min(min_image_lsn, img_lsn);
5530 : }
5531 : }
5532 :
5533 : // HADRON
5534 : // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
5535 57 : min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
5536 57 : if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
5537 0 : info!(
5538 0 : "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
5539 0 : partition.ranges[0].start,
5540 0 : partition.ranges[0].end,
5541 : min_image_lsn,
5542 0 : force_image_creation_lsn.unwrap(),
5543 : max_deltas
5544 : );
5545 0 : return true;
5546 57 : }
5547 :
5548 57 : debug!(
5549 : max_deltas,
5550 0 : "none of the partitioned ranges had >= {threshold} deltas"
5551 : );
5552 57 : false
5553 57 : }
5554 :
5555 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5556 : /// so that at most one image layer will be produced from this function.
5557 : #[allow(clippy::too_many_arguments)]
5558 124 : async fn create_image_layer_for_rel_blocks(
5559 124 : self: &Arc<Self>,
5560 124 : partition: &KeySpace,
5561 124 : mut image_layer_writer: ImageLayerWriter,
5562 124 : lsn: Lsn,
5563 124 : ctx: &RequestContext,
5564 124 : img_range: Range<Key>,
5565 124 : io_concurrency: IoConcurrency,
5566 124 : progress: Option<(usize, usize)>,
5567 124 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5568 124 : let mut wrote_keys = false;
5569 :
5570 124 : let mut key_request_accum = KeySpaceAccum::new();
5571 824 : for range in &partition.ranges {
5572 700 : let mut key = range.start;
5573 1517 : while key < range.end {
5574 : // Decide whether to retain this key: usually we do, but sharded tenants may
5575 : // need to drop keys that don't belong to them. If we retain the key, add it
5576 : // to `key_request_accum` for later issuing a vectored get
5577 817 : if self.shard_identity.is_key_disposable(&key) {
5578 0 : debug!(
5579 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5580 : key,
5581 0 : self.shard_identity.get_shard_number(&key)
5582 : );
5583 817 : } else {
5584 817 : key_request_accum.add_key(key);
5585 817 : }
5586 :
5587 817 : let last_key_in_range = key.next() == range.end;
5588 817 : key = key.next();
5589 :
5590 : // Maybe flush `key_rest_accum`
5591 817 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5592 817 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5593 : {
5594 700 : let query =
5595 700 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5596 :
5597 700 : let results = self
5598 700 : .get_vectored(query, io_concurrency.clone(), ctx)
5599 700 : .await?;
5600 :
5601 700 : if self.cancel.is_cancelled() {
5602 0 : return Err(CreateImageLayersError::Cancelled);
5603 700 : }
5604 :
5605 1517 : for (img_key, img) in results {
5606 817 : let img = match img {
5607 817 : Ok(img) => img,
5608 0 : Err(err) => {
5609 : // If we fail to reconstruct a VM or FSM page, we can zero the
5610 : // page without losing any actual user data. That seems better
5611 : // than failing repeatedly and getting stuck.
5612 : //
5613 : // We had a bug at one point, where we truncated the FSM and VM
5614 : // in the pageserver, but the Postgres didn't know about that
5615 : // and continued to generate incremental WAL records for pages
5616 : // that didn't exist in the pageserver. Trying to replay those
5617 : // WAL records failed to find the previous image of the page.
5618 : // This special case allows us to recover from that situation.
5619 : // See https://github.com/neondatabase/neon/issues/2601.
5620 : //
5621 : // Unfortunately we cannot do this for the main fork, or for
5622 : // any metadata keys, keys, as that would lead to actual data
5623 : // loss.
5624 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5625 0 : warn!(
5626 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5627 : );
5628 0 : ZERO_PAGE.clone()
5629 : } else {
5630 0 : return Err(CreateImageLayersError::from(err));
5631 : }
5632 : }
5633 : };
5634 :
5635 : // Write all the keys we just read into our new image layer.
5636 817 : image_layer_writer.put_image(img_key, img, ctx).await?;
5637 817 : wrote_keys = true;
5638 : }
5639 117 : }
5640 : }
5641 : }
5642 :
5643 124 : let progress_report = progress
5644 124 : .map(|(idx, total)| format!("({idx}/{total}) "))
5645 124 : .unwrap_or_default();
5646 124 : if wrote_keys {
5647 : // Normal path: we have written some data into the new image layer for this
5648 : // partition, so flush it to disk.
5649 124 : info!(
5650 0 : "{} produced image layer for rel {}",
5651 : progress_report,
5652 0 : ImageLayerName {
5653 0 : key_range: img_range.clone(),
5654 0 : lsn
5655 0 : },
5656 : );
5657 124 : Ok(ImageLayerCreationOutcome::Generated {
5658 124 : unfinished_image_layer: image_layer_writer,
5659 124 : })
5660 : } else {
5661 0 : tracing::debug!(
5662 0 : "{} no data in range {}-{}",
5663 : progress_report,
5664 : img_range.start,
5665 : img_range.end
5666 : );
5667 0 : Ok(ImageLayerCreationOutcome::Empty)
5668 : }
5669 124 : }
5670 :
5671 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5672 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5673 : /// would not be too large to fit in a single image layer.
5674 : ///
5675 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5676 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5677 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5678 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5679 : #[allow(clippy::too_many_arguments)]
5680 169 : async fn create_image_layer_for_metadata_keys(
5681 169 : self: &Arc<Self>,
5682 169 : partition: &KeySpace,
5683 169 : mut image_layer_writer: ImageLayerWriter,
5684 169 : lsn: Lsn,
5685 169 : ctx: &RequestContext,
5686 169 : img_range: Range<Key>,
5687 169 : mode: ImageLayerCreationMode,
5688 169 : io_concurrency: IoConcurrency,
5689 169 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5690 : // Metadata keys image layer creation.
5691 169 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5692 169 : let begin = Instant::now();
5693 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5694 : // not contain too many keys, otherwise this takes a lot of memory.
5695 169 : let data = self
5696 169 : .get_vectored_impl(
5697 169 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5698 169 : &mut reconstruct_state,
5699 169 : ctx,
5700 169 : )
5701 169 : .await?;
5702 169 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5703 169 : let mut new_data = BTreeMap::new();
5704 169 : let mut total_kb_retrieved = 0;
5705 169 : let mut total_keys_retrieved = 0;
5706 5175 : for (k, v) in data {
5707 5006 : let v = v?;
5708 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5709 5006 : total_keys_retrieved += 1;
5710 5006 : new_data.insert(k, v);
5711 : }
5712 169 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5713 : };
5714 169 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5715 169 : let elapsed = begin.elapsed();
5716 :
5717 169 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5718 169 : info!(
5719 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",
5720 0 : elapsed.as_secs_f64()
5721 : );
5722 :
5723 169 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5724 16 : return Ok(ImageLayerCreationOutcome::Skip);
5725 153 : }
5726 153 : if self.cancel.is_cancelled() {
5727 0 : return Err(CreateImageLayersError::Cancelled);
5728 153 : }
5729 153 : let mut wrote_any_image = false;
5730 5159 : for (k, v) in data {
5731 5006 : if v.is_empty() {
5732 : // the key has been deleted, it does not need an image
5733 : // in metadata keyspace, an empty image == tombstone
5734 4 : continue;
5735 5002 : }
5736 5002 : wrote_any_image = true;
5737 :
5738 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5739 :
5740 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5741 : // on the normal data path either.
5742 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5743 : }
5744 :
5745 153 : if wrote_any_image {
5746 : // Normal path: we have written some data into the new image layer for this
5747 : // partition, so flush it to disk.
5748 6 : info!(
5749 0 : "created image layer for metadata {}",
5750 0 : ImageLayerName {
5751 0 : key_range: img_range.clone(),
5752 0 : lsn
5753 0 : }
5754 : );
5755 6 : Ok(ImageLayerCreationOutcome::Generated {
5756 6 : unfinished_image_layer: image_layer_writer,
5757 6 : })
5758 : } else {
5759 147 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5760 147 : Ok(ImageLayerCreationOutcome::Empty)
5761 : }
5762 169 : }
5763 :
5764 : /// Predicate function which indicates whether we should check if new image layers
5765 : /// are required. Since checking if new image layers are required is expensive in
5766 : /// terms of CPU, we only do it in the following cases:
5767 : /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
5768 : /// 2. If enough time has passed since the last check:
5769 : /// 1. For large tenants, we wish to perform the check more often since they
5770 : /// suffer from the lack of image layers. Note that we assume sharded tenants
5771 : /// to be large since non-zero shards do not track the logical size.
5772 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5773 191 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5774 191 : let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
5775 :
5776 191 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5777 191 : let distance = lsn
5778 191 : .checked_sub(last_checks_at)
5779 191 : .expect("Attempt to compact with LSN going backwards");
5780 191 : let min_distance =
5781 191 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5782 :
5783 191 : let distance_based_decision = distance.0 >= min_distance;
5784 :
5785 191 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5786 191 : let check_required_after = (|| {
5787 191 : if self.shard_identity.is_unsharded() {
5788 135 : if let CurrentLogicalSize::Exact(logical_size) =
5789 186 : self.current_logical_size.current_size()
5790 : {
5791 135 : if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
5792 135 : return Duration::from_secs(3600 * 48);
5793 0 : }
5794 51 : }
5795 5 : }
5796 :
5797 56 : self.get_checkpoint_timeout()
5798 : })();
5799 :
5800 191 : let time_based_decision = match *last_check_instant {
5801 29 : Some(last_check) => {
5802 29 : let elapsed = last_check.elapsed();
5803 29 : elapsed >= check_required_after
5804 : }
5805 162 : None => true,
5806 : };
5807 :
5808 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5809 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5810 : // check.
5811 191 : let decision = distance_based_decision || time_based_decision;
5812 191 : tracing::info!(
5813 0 : "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
5814 : decision,
5815 : distance_based_decision,
5816 : time_based_decision
5817 : );
5818 191 : if decision {
5819 162 : self.last_image_layer_creation_check_at.store(lsn);
5820 162 : *last_check_instant = Some(Instant::now());
5821 162 : }
5822 :
5823 191 : decision
5824 191 : }
5825 :
5826 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5827 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5828 : ///
5829 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5830 : #[allow(clippy::too_many_arguments)]
5831 191 : async fn create_image_layers(
5832 191 : self: &Arc<Timeline>,
5833 191 : partitioning: &KeyPartitioning,
5834 191 : lsn: Lsn,
5835 191 : force_image_creation_lsn: Option<Lsn>,
5836 191 : mode: ImageLayerCreationMode,
5837 191 : ctx: &RequestContext,
5838 191 : last_status: LastImageLayerCreationStatus,
5839 191 : yield_for_l0: bool,
5840 191 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5841 191 : let timer = self.metrics.create_images_time_histo.start_timer();
5842 :
5843 191 : if partitioning.parts.is_empty() {
5844 0 : warn!("no partitions to create image layers for");
5845 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5846 191 : }
5847 :
5848 : // We need to avoid holes between generated image layers.
5849 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5850 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5851 : //
5852 : // How such hole between partitions can appear?
5853 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5854 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5855 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5856 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5857 191 : let mut start = Key::MIN;
5858 :
5859 191 : let check_for_image_layers =
5860 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5861 0 : info!(
5862 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5863 : last_key
5864 : );
5865 0 : true
5866 : } else {
5867 191 : self.should_check_if_image_layers_required(lsn)
5868 : };
5869 :
5870 191 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5871 :
5872 191 : let mut all_generated = true;
5873 :
5874 191 : let mut partition_processed = 0;
5875 191 : let mut total_partitions = partitioning.parts.len();
5876 191 : let mut last_partition_processed = None;
5877 191 : let mut partition_parts = partitioning.parts.clone();
5878 :
5879 191 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5880 : // We need to skip the partitions that have already been processed.
5881 0 : let mut found = false;
5882 0 : for (i, partition) in partition_parts.iter().enumerate() {
5883 0 : if last_key <= partition.end().unwrap() {
5884 : // ```plain
5885 : // |------|--------|----------|------|
5886 : // ^last_key
5887 : // ^start from this partition
5888 : // ```
5889 : // Why `i+1` instead of `i`?
5890 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5891 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5892 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5893 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5894 0 : if i + 1 >= total_partitions {
5895 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5896 : // iteration of image layer creation should return a complete status.
5897 0 : break; // with found=false
5898 0 : }
5899 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5900 0 : total_partitions = partition_parts.len();
5901 : // Update the start key to the partition start.
5902 0 : start = partition_parts[0].start().unwrap();
5903 0 : found = true;
5904 0 : break;
5905 0 : }
5906 : }
5907 0 : if !found {
5908 : // Last key is within the last partition, or larger than all partitions.
5909 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5910 0 : }
5911 191 : }
5912 :
5913 191 : let total = partition_parts.len();
5914 394 : for (idx, partition) in partition_parts.iter().enumerate() {
5915 394 : if self.cancel.is_cancelled() {
5916 0 : return Err(CreateImageLayersError::Cancelled);
5917 394 : }
5918 394 : partition_processed += 1;
5919 394 : let img_range = start..partition.ranges.last().unwrap().end;
5920 394 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5921 394 : if compact_metadata {
5922 764 : for range in &partition.ranges {
5923 573 : assert!(
5924 573 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5925 573 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5926 0 : "metadata keys must be partitioned separately"
5927 : );
5928 : }
5929 191 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5930 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5931 : // might mess up with evictions.
5932 22 : start = img_range.end;
5933 22 : continue;
5934 169 : }
5935 : // For initial and force modes, we always generate image layers for metadata keys.
5936 203 : } else if let ImageLayerCreationMode::Try = mode {
5937 : // check_for_image_layers = false -> skip
5938 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5939 79 : if !check_for_image_layers
5940 57 : || !self
5941 57 : .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
5942 57 : .await
5943 : {
5944 79 : start = img_range.end;
5945 79 : continue;
5946 0 : }
5947 124 : }
5948 293 : if let ImageLayerCreationMode::Force = mode {
5949 : // When forced to create image layers, we might try and create them where they already
5950 : // exist. This mode is only used in tests/debug.
5951 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5952 14 : if layers.contains_key(&PersistentLayerKey {
5953 14 : key_range: img_range.clone(),
5954 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5955 14 : is_delta: false,
5956 14 : }) {
5957 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5958 : // in the future.
5959 0 : tracing::info!(
5960 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5961 : img_range.start,
5962 : img_range.end
5963 : );
5964 0 : start = img_range.end;
5965 0 : continue;
5966 14 : }
5967 279 : }
5968 :
5969 293 : let image_layer_writer = ImageLayerWriter::new(
5970 293 : self.conf,
5971 293 : self.timeline_id,
5972 293 : self.tenant_shard_id,
5973 293 : &img_range,
5974 293 : lsn,
5975 293 : &self.gate,
5976 293 : self.cancel.clone(),
5977 293 : ctx,
5978 293 : )
5979 293 : .await
5980 293 : .map_err(CreateImageLayersError::Other)?;
5981 :
5982 293 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5983 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5984 0 : "failpoint image-layer-writer-fail-before-finish"
5985 0 : )))
5986 0 : });
5987 :
5988 293 : let io_concurrency = IoConcurrency::spawn_from_conf(
5989 293 : self.conf.get_vectored_concurrent_io,
5990 293 : self.gate
5991 293 : .enter()
5992 293 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5993 : );
5994 :
5995 293 : let outcome = if !compact_metadata {
5996 124 : self.create_image_layer_for_rel_blocks(
5997 124 : partition,
5998 124 : image_layer_writer,
5999 124 : lsn,
6000 124 : ctx,
6001 124 : img_range.clone(),
6002 124 : io_concurrency,
6003 124 : Some((idx, total)),
6004 124 : )
6005 124 : .await?
6006 : } else {
6007 169 : self.create_image_layer_for_metadata_keys(
6008 169 : partition,
6009 169 : image_layer_writer,
6010 169 : lsn,
6011 169 : ctx,
6012 169 : img_range.clone(),
6013 169 : mode,
6014 169 : io_concurrency,
6015 169 : )
6016 169 : .await?
6017 : };
6018 293 : match outcome {
6019 147 : ImageLayerCreationOutcome::Empty => {
6020 147 : // No data in this partition, so we don't need to create an image layer (for now).
6021 147 : // The next image layer should cover this key range, so we don't advance the `start`
6022 147 : // key.
6023 147 : }
6024 : ImageLayerCreationOutcome::Generated {
6025 130 : unfinished_image_layer,
6026 130 : } => {
6027 130 : batch_image_writer.add_unfinished_image_writer(
6028 130 : unfinished_image_layer,
6029 130 : img_range.clone(),
6030 130 : lsn,
6031 130 : );
6032 130 : // The next image layer should be generated right after this one.
6033 130 : start = img_range.end;
6034 130 : }
6035 16 : ImageLayerCreationOutcome::Skip => {
6036 16 : // We don't need to create an image layer for this partition.
6037 16 : // The next image layer should NOT cover this range, otherwise
6038 16 : // the keyspace becomes empty (reads don't go past image layers).
6039 16 : start = img_range.end;
6040 16 : }
6041 : }
6042 :
6043 293 : if let ImageLayerCreationMode::Try = mode {
6044 : // We have at least made some progress
6045 51 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
6046 : // The `Try` mode is currently only used on the compaction path. We want to avoid
6047 : // image layer generation taking too long time and blocking L0 compaction. So in this
6048 : // mode, we also inspect the current number of L0 layers and skip image layer generation
6049 : // if there are too many of them.
6050 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
6051 0 : * self.get_compaction_threshold();
6052 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
6053 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
6054 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
6055 0 : if image_preempt_threshold != 0 {
6056 0 : let should_yield = self
6057 0 : .l0_compaction_trigger
6058 0 : .notified()
6059 0 : .now_or_never()
6060 0 : .is_some();
6061 0 : if should_yield {
6062 0 : tracing::info!(
6063 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
6064 0 : partition.start().unwrap(),
6065 0 : partition.end().unwrap()
6066 : );
6067 0 : last_partition_processed = Some(partition.clone());
6068 0 : all_generated = false;
6069 0 : break;
6070 0 : }
6071 0 : }
6072 51 : }
6073 242 : }
6074 : }
6075 :
6076 191 : let image_layers = batch_image_writer
6077 191 : .finish(self, ctx)
6078 191 : .await
6079 191 : .map_err(CreateImageLayersError::Other)?;
6080 :
6081 191 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6082 :
6083 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
6084 : // now they are being scheduled outside of write lock; current way is inconsistent with
6085 : // compaction lock order.
6086 191 : guard
6087 191 : .open_mut()?
6088 191 : .track_new_image_layers(&image_layers, &self.metrics);
6089 191 : drop_layer_manager_wlock(guard);
6090 191 : let duration = timer.stop_and_record();
6091 :
6092 : // Creating image layers may have caused some previously visible layers to be covered
6093 191 : if !image_layers.is_empty() {
6094 118 : self.update_layer_visibility().await?;
6095 73 : }
6096 :
6097 191 : let total_layer_size = image_layers
6098 191 : .iter()
6099 191 : .map(|l| l.metadata().file_size)
6100 191 : .sum::<u64>();
6101 :
6102 191 : if !image_layers.is_empty() {
6103 118 : info!(
6104 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
6105 0 : image_layers.len(),
6106 : total_layer_size,
6107 0 : duration.as_secs_f64(),
6108 : partition_processed,
6109 : total_partitions
6110 : );
6111 73 : }
6112 :
6113 : Ok((
6114 191 : image_layers,
6115 191 : if all_generated {
6116 191 : LastImageLayerCreationStatus::Complete
6117 : } else {
6118 : LastImageLayerCreationStatus::Incomplete {
6119 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
6120 0 : last_partition_processed.end().unwrap_or(Key::MIN)
6121 : } else {
6122 : // This branch should be unreachable, but in case it happens, we can just return the start key.
6123 0 : Key::MIN
6124 : },
6125 : }
6126 : },
6127 : ))
6128 191 : }
6129 :
6130 : /// Wait until the background initial logical size calculation is complete, or
6131 : /// this Timeline is shut down. Calling this function will cause the initial
6132 : /// logical size calculation to skip waiting for the background jobs barrier.
6133 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
6134 0 : if !self.shard_identity.is_shard_zero() {
6135 : // We don't populate logical size on shard >0: skip waiting for it.
6136 0 : return;
6137 0 : }
6138 :
6139 0 : if self.remote_client.is_deleting() {
6140 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
6141 0 : return;
6142 0 : }
6143 :
6144 0 : if self.current_logical_size.current_size().is_exact() {
6145 : // root timelines are initialized with exact count, but never start the background
6146 : // calculation
6147 0 : return;
6148 0 : }
6149 :
6150 0 : if self.cancel.is_cancelled() {
6151 : // We already requested stopping the tenant, so we cannot wait for the logical size
6152 : // calculation to complete given the task might have been already cancelled.
6153 0 : return;
6154 0 : }
6155 :
6156 0 : if let Some(await_bg_cancel) = self
6157 0 : .current_logical_size
6158 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
6159 0 : .get()
6160 0 : {
6161 0 : await_bg_cancel.cancel();
6162 0 : } else {
6163 : // We should not wait if we were not able to explicitly instruct
6164 : // the logical size cancellation to skip the concurrency limit semaphore.
6165 : // TODO: this is an unexpected case. We should restructure so that it
6166 : // can't happen.
6167 0 : tracing::warn!(
6168 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
6169 : );
6170 0 : debug_assert!(false);
6171 : }
6172 :
6173 0 : tokio::select!(
6174 0 : _ = self.current_logical_size.initialized.acquire() => {},
6175 0 : _ = self.cancel.cancelled() => {}
6176 : )
6177 0 : }
6178 :
6179 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
6180 : /// Timelines layers up to the ancestor_lsn.
6181 : ///
6182 : /// Requires a timeline that:
6183 : /// - has an ancestor to detach from
6184 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
6185 : /// a technical requirement
6186 : ///
6187 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
6188 : /// polled again until completion.
6189 : ///
6190 : /// During the operation all timelines sharing the data with this timeline will be reparented
6191 : /// from our ancestor to be branches of this timeline.
6192 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
6193 0 : self: &Arc<Timeline>,
6194 0 : tenant: &crate::tenant::TenantShard,
6195 0 : options: detach_ancestor::Options,
6196 0 : behavior: DetachBehavior,
6197 0 : ctx: &RequestContext,
6198 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
6199 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
6200 0 : }
6201 :
6202 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
6203 : /// reparents any reparentable children of previous ancestor.
6204 : ///
6205 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6206 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6207 : /// successfully, tenant must be reloaded.
6208 : ///
6209 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6210 : /// resetting the tenant.
6211 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6212 0 : self: &Arc<Timeline>,
6213 0 : tenant: &crate::tenant::TenantShard,
6214 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6215 0 : ancestor_timeline_id: TimelineId,
6216 0 : ancestor_lsn: Lsn,
6217 0 : behavior: DetachBehavior,
6218 0 : ctx: &RequestContext,
6219 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6220 0 : detach_ancestor::detach_and_reparent(
6221 0 : self,
6222 0 : tenant,
6223 0 : prepared,
6224 0 : ancestor_timeline_id,
6225 0 : ancestor_lsn,
6226 0 : behavior,
6227 0 : ctx,
6228 0 : )
6229 0 : .await
6230 0 : }
6231 :
6232 : /// Final step which unblocks the GC.
6233 : ///
6234 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6235 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6236 0 : self: &Arc<Timeline>,
6237 0 : tenant: &crate::tenant::TenantShard,
6238 0 : attempt: detach_ancestor::Attempt,
6239 0 : ctx: &RequestContext,
6240 0 : ) -> Result<(), detach_ancestor::Error> {
6241 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6242 0 : }
6243 : }
6244 :
6245 : impl Drop for Timeline {
6246 5 : fn drop(&mut self) {
6247 5 : if let Some(ancestor) = &self.ancestor_timeline {
6248 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6249 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6250 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6251 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6252 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6253 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6254 2 : }
6255 0 : }
6256 3 : }
6257 5 : info!(
6258 0 : "Timeline {} for tenant {} is being dropped",
6259 : self.timeline_id, self.tenant_shard_id.tenant_id
6260 : );
6261 5 : }
6262 : }
6263 :
6264 : pub(crate) use compaction_error::CompactionError;
6265 : /// In a private mod to enforce that [`CompactionError::is_cancel`] is used
6266 : /// instead of `match`ing on [`CompactionError::ShuttingDown`].
6267 : mod compaction_error {
6268 : use utils::sync::gate::GateError;
6269 :
6270 : use crate::{
6271 : pgdatadir_mapping::CollectKeySpaceError,
6272 : tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
6273 : virtual_file::owned_buffers_io::write::FlushTaskError,
6274 : };
6275 :
6276 : /// Top-level failure to compact. Use [`Self::is_cancel`].
6277 : #[derive(Debug, thiserror::Error)]
6278 : pub(crate) enum CompactionError {
6279 : /// Use [`Self::is_cancel`] instead of checking for this variant.
6280 : #[error("The timeline or pageserver is shutting down")]
6281 : #[allow(private_interfaces)]
6282 : ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
6283 : #[error(transparent)]
6284 : Other(anyhow::Error),
6285 : }
6286 :
6287 : #[derive(Debug)]
6288 : struct ForbidMatching;
6289 :
6290 : impl CompactionError {
6291 0 : pub fn new_cancelled() -> Self {
6292 0 : Self::ShuttingDown(ForbidMatching)
6293 0 : }
6294 : /// Errors that can be ignored, i.e., cancel and shutdown.
6295 0 : pub fn is_cancel(&self) -> bool {
6296 0 : let other = match self {
6297 0 : CompactionError::ShuttingDown(_) => return true,
6298 0 : CompactionError::Other(other) => other,
6299 : };
6300 :
6301 : // The write path of compaction in particular often lacks differentiated
6302 : // handling errors stemming from cancellation from other errors.
6303 : // So, if requested, we also check the ::Other variant by downcasting.
6304 : // The list below has been found empirically from flaky tests and production logs.
6305 : // The process is simple: on ::Other(), compaction will print the enclosed
6306 : // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
6307 : // line where the write path / compaction code does undifferentiated error handling
6308 : // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
6309 : // below, following the same is_cancel() pattern.
6310 :
6311 0 : let root_cause = other.root_cause();
6312 :
6313 0 : let upload_queue = root_cause
6314 0 : .downcast_ref::<NotInitialized>()
6315 0 : .is_some_and(|e| e.is_stopping());
6316 0 : let timeline = root_cause
6317 0 : .downcast_ref::<PageReconstructError>()
6318 0 : .is_some_and(|e| e.is_cancel());
6319 0 : let buffered_writer_flush_task_canelled = root_cause
6320 0 : .downcast_ref::<FlushTaskError>()
6321 0 : .is_some_and(|e| e.is_cancel());
6322 0 : let write_blob_cancelled = root_cause
6323 0 : .downcast_ref::<WriteBlobError>()
6324 0 : .is_some_and(|e| e.is_cancel());
6325 0 : let gate_closed = root_cause
6326 0 : .downcast_ref::<GateError>()
6327 0 : .is_some_and(|e| e.is_cancel());
6328 0 : upload_queue
6329 0 : || timeline
6330 0 : || buffered_writer_flush_task_canelled
6331 0 : || write_blob_cancelled
6332 0 : || gate_closed
6333 0 : }
6334 0 : pub fn into_anyhow(self) -> anyhow::Error {
6335 0 : match self {
6336 0 : CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
6337 0 : CompactionError::Other(e) => e,
6338 : }
6339 0 : }
6340 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6341 0 : if err.is_cancel() {
6342 0 : Self::new_cancelled()
6343 : } else {
6344 0 : Self::Other(err.into_anyhow())
6345 : }
6346 0 : }
6347 : }
6348 : }
6349 :
6350 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6351 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6352 0 : match value {
6353 : super::upload_queue::NotInitialized::Uninitialized => {
6354 0 : CompactionError::Other(anyhow::anyhow!(value))
6355 : }
6356 : super::upload_queue::NotInitialized::ShuttingDown
6357 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
6358 : }
6359 0 : }
6360 : }
6361 :
6362 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6363 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6364 0 : match e {
6365 : super::storage_layer::layer::DownloadError::TimelineShutdown
6366 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6367 0 : CompactionError::new_cancelled()
6368 : }
6369 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6370 : | super::storage_layer::layer::DownloadError::DownloadRequired
6371 : | super::storage_layer::layer::DownloadError::NotFile(_)
6372 : | super::storage_layer::layer::DownloadError::DownloadFailed
6373 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6374 0 : CompactionError::Other(anyhow::anyhow!(e))
6375 : }
6376 : #[cfg(test)]
6377 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6378 0 : CompactionError::Other(anyhow::anyhow!(e))
6379 : }
6380 : }
6381 0 : }
6382 : }
6383 :
6384 : impl From<layer_manager::Shutdown> for CompactionError {
6385 0 : fn from(_: layer_manager::Shutdown) -> Self {
6386 0 : CompactionError::new_cancelled()
6387 0 : }
6388 : }
6389 :
6390 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6391 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6392 0 : if e.is_cancel() {
6393 0 : CompactionError::new_cancelled()
6394 : } else {
6395 0 : CompactionError::Other(e.into_anyhow())
6396 : }
6397 0 : }
6398 : }
6399 :
6400 : #[serde_as]
6401 : #[derive(serde::Serialize)]
6402 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6403 :
6404 : #[derive(Default)]
6405 : enum DurationRecorder {
6406 : #[default]
6407 : NotStarted,
6408 : Recorded(RecordedDuration, tokio::time::Instant),
6409 : }
6410 :
6411 : impl DurationRecorder {
6412 115 : fn till_now(&self) -> DurationRecorder {
6413 115 : match self {
6414 : DurationRecorder::NotStarted => {
6415 0 : panic!("must only call on recorded measurements")
6416 : }
6417 115 : DurationRecorder::Recorded(_, ended) => {
6418 115 : let now = tokio::time::Instant::now();
6419 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6420 : }
6421 : }
6422 115 : }
6423 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6424 138 : match self {
6425 0 : DurationRecorder::NotStarted => None,
6426 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6427 : }
6428 138 : }
6429 : }
6430 :
6431 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6432 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6433 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6434 : /// keys specified in the `data` field.
6435 : #[cfg(test)]
6436 : #[derive(Clone)]
6437 : pub struct DeltaLayerTestDesc {
6438 : pub lsn_range: Range<Lsn>,
6439 : pub key_range: Range<Key>,
6440 : pub data: Vec<(Key, Lsn, Value)>,
6441 : }
6442 :
6443 : #[cfg(test)]
6444 : #[derive(Clone)]
6445 : pub struct InMemoryLayerTestDesc {
6446 : pub lsn_range: Range<Lsn>,
6447 : pub data: Vec<(Key, Lsn, Value)>,
6448 : pub is_open: bool,
6449 : }
6450 :
6451 : #[cfg(test)]
6452 : impl DeltaLayerTestDesc {
6453 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6454 2 : Self {
6455 2 : lsn_range,
6456 2 : key_range,
6457 2 : data,
6458 2 : }
6459 2 : }
6460 :
6461 45 : pub fn new_with_inferred_key_range(
6462 45 : lsn_range: Range<Lsn>,
6463 45 : data: Vec<(Key, Lsn, Value)>,
6464 45 : ) -> Self {
6465 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6466 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6467 45 : Self {
6468 45 : key_range: (*key_min)..(key_max.next()),
6469 45 : lsn_range,
6470 45 : data,
6471 45 : }
6472 45 : }
6473 :
6474 5 : pub(crate) fn layer_name(&self) -> LayerName {
6475 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6476 5 : key_range: self.key_range.clone(),
6477 5 : lsn_range: self.lsn_range.clone(),
6478 5 : })
6479 5 : }
6480 : }
6481 :
6482 : impl Timeline {
6483 23 : async fn finish_compact_batch(
6484 23 : self: &Arc<Self>,
6485 23 : new_deltas: &[ResidentLayer],
6486 23 : new_images: &[ResidentLayer],
6487 23 : layers_to_remove: &[Layer],
6488 23 : ) -> Result<(), CompactionError> {
6489 23 : let mut guard = tokio::select! {
6490 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6491 23 : _ = self.cancel.cancelled() => {
6492 0 : return Err(CompactionError::new_cancelled());
6493 : }
6494 : };
6495 :
6496 23 : let mut duplicated_layers = HashSet::new();
6497 :
6498 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6499 :
6500 186 : for l in new_deltas {
6501 163 : if guard.contains(l.as_ref()) {
6502 : // expected in tests
6503 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6504 :
6505 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6506 : // `guard` on self.layers. as of writing this, there are no error returns except
6507 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6508 : // because we have not implemented L0 => L0 compaction.
6509 0 : duplicated_layers.insert(l.layer_desc().key());
6510 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6511 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6512 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6513 0 : )));
6514 163 : } else {
6515 163 : insert_layers.push(l.clone());
6516 163 : }
6517 : }
6518 :
6519 : // only remove those inputs which were not outputs
6520 23 : let remove_layers: Vec<Layer> = layers_to_remove
6521 23 : .iter()
6522 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6523 23 : .cloned()
6524 23 : .collect();
6525 :
6526 23 : if !new_images.is_empty() {
6527 0 : guard
6528 0 : .open_mut()?
6529 0 : .track_new_image_layers(new_images, &self.metrics);
6530 23 : }
6531 :
6532 23 : guard
6533 23 : .open_mut()?
6534 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6535 :
6536 23 : self.remote_client
6537 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6538 :
6539 23 : drop_layer_manager_wlock(guard);
6540 :
6541 23 : Ok(())
6542 23 : }
6543 :
6544 0 : async fn rewrite_layers(
6545 0 : self: &Arc<Self>,
6546 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6547 0 : mut drop_layers: Vec<Layer>,
6548 0 : ) -> Result<(), CompactionError> {
6549 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6550 :
6551 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6552 : // to avoid double-removing, and avoid rewriting something that was removed.
6553 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6554 0 : drop_layers.retain(|l| guard.contains(l));
6555 :
6556 0 : guard
6557 0 : .open_mut()?
6558 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6559 :
6560 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6561 :
6562 0 : self.remote_client
6563 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6564 :
6565 0 : Ok(())
6566 0 : }
6567 :
6568 : /// Schedules the uploads of the given image layers
6569 80 : fn upload_new_image_layers(
6570 80 : self: &Arc<Self>,
6571 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6572 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6573 93 : for layer in new_images {
6574 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6575 : }
6576 : // should any new image layer been created, not uploading index_part will
6577 : // result in a mismatch between remote_physical_size and layermap calculated
6578 : // size, which will fail some tests, but should not be an issue otherwise.
6579 80 : self.remote_client
6580 80 : .schedule_index_upload_for_file_changes()?;
6581 80 : Ok(())
6582 80 : }
6583 :
6584 0 : async fn find_gc_time_cutoff(
6585 0 : &self,
6586 0 : now: SystemTime,
6587 0 : pitr: Duration,
6588 0 : cancel: &CancellationToken,
6589 0 : ctx: &RequestContext,
6590 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6591 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6592 0 : if self.shard_identity.is_shard_zero() {
6593 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6594 0 : let time_range = if pitr == Duration::ZERO {
6595 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6596 : } else {
6597 0 : pitr
6598 : };
6599 :
6600 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6601 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6602 0 : let timestamp = to_pg_timestamp(time_cutoff);
6603 :
6604 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6605 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6606 0 : LsnForTimestamp::Future(lsn) => {
6607 : // The timestamp is in the future. That sounds impossible,
6608 : // but what it really means is that there hasn't been
6609 : // any commits since the cutoff timestamp.
6610 : //
6611 : // In this case we should use the LSN of the most recent commit,
6612 : // which is implicitly the last LSN in the log.
6613 0 : debug!("future({})", lsn);
6614 0 : Some(self.get_last_record_lsn())
6615 : }
6616 0 : LsnForTimestamp::Past(lsn) => {
6617 0 : debug!("past({})", lsn);
6618 0 : None
6619 : }
6620 0 : LsnForTimestamp::NoData(lsn) => {
6621 0 : debug!("nodata({})", lsn);
6622 0 : None
6623 : }
6624 : };
6625 0 : Ok(time_cutoff)
6626 : } else {
6627 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6628 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6629 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6630 : // space cutoff that we would also have respected ourselves.
6631 0 : match self
6632 0 : .remote_client
6633 0 : .download_foreign_index(ShardNumber(0), cancel)
6634 0 : .await
6635 : {
6636 0 : Ok((index_part, index_generation, _index_mtime)) => {
6637 0 : tracing::info!(
6638 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6639 0 : index_part.metadata.latest_gc_cutoff_lsn()
6640 : );
6641 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6642 : }
6643 : Err(DownloadError::NotFound) => {
6644 : // This is unexpected, because during timeline creations shard zero persists to remote
6645 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6646 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6647 : // state, then shard zero should _eventually_ write an index when it GCs.
6648 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6649 0 : Ok(None)
6650 : }
6651 0 : Err(e) => {
6652 : // TODO: this function should return a different error type than page reconstruct error
6653 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6654 : }
6655 : }
6656 :
6657 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6658 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6659 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6660 : // new location. This is legal in principle, but problematic in practice because it might result
6661 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6662 : // because they have GC'd past the branch point.
6663 : }
6664 0 : }
6665 :
6666 : /// Find the Lsns above which layer files need to be retained on
6667 : /// garbage collection.
6668 : ///
6669 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6670 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6671 : /// the space-based retention.
6672 : ///
6673 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6674 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6675 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6676 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6677 : /// in the response as the GC cutoff point for the timeline.
6678 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6679 : pub(super) async fn find_gc_cutoffs(
6680 : &self,
6681 : now: SystemTime,
6682 : space_cutoff: Lsn,
6683 : pitr: Duration,
6684 : cancel: &CancellationToken,
6685 : ctx: &RequestContext,
6686 : ) -> Result<GcCutoffs, PageReconstructError> {
6687 : let _timer = self
6688 : .metrics
6689 : .find_gc_cutoffs_histo
6690 : .start_timer()
6691 : .record_on_drop();
6692 :
6693 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6694 :
6695 : if cfg!(test) && pitr == Duration::ZERO {
6696 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6697 : return Ok(GcCutoffs {
6698 : time: Some(self.get_last_record_lsn()),
6699 : space: space_cutoff,
6700 : });
6701 : }
6702 :
6703 : // Calculate a time-based limit on how much to retain:
6704 : // - if PITR interval is set, then this is our cutoff.
6705 : // - if PITR interval is not set, then we do a lookup
6706 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6707 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6708 :
6709 : Ok(match (pitr, time_cutoff) {
6710 : (Duration::ZERO, Some(time_cutoff)) => {
6711 : // PITR is not set. Retain the size-based limit, or the default time retention,
6712 : // whichever requires less data.
6713 : GcCutoffs {
6714 : time: Some(self.get_last_record_lsn()),
6715 : space: std::cmp::max(time_cutoff, space_cutoff),
6716 : }
6717 : }
6718 : (Duration::ZERO, None) => {
6719 : // PITR is not set, and time lookup failed
6720 : GcCutoffs {
6721 : time: Some(self.get_last_record_lsn()),
6722 : space: space_cutoff,
6723 : }
6724 : }
6725 : (_, None) => {
6726 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6727 : // cannot advance beyond what was already GC'd, and respect space-based retention
6728 : GcCutoffs {
6729 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6730 : space: space_cutoff,
6731 : }
6732 : }
6733 : (_, Some(time_cutoff)) => {
6734 : // PITR interval is set and we looked up timestamp successfully. Ignore
6735 : // size based retention and make time cutoff authoritative
6736 : GcCutoffs {
6737 : time: Some(time_cutoff),
6738 : space: time_cutoff,
6739 : }
6740 : }
6741 : })
6742 : }
6743 :
6744 : /// Garbage collect layer files on a timeline that are no longer needed.
6745 : ///
6746 : /// Currently, we don't make any attempt at removing unneeded page versions
6747 : /// within a layer file. We can only remove the whole file if it's fully
6748 : /// obsolete.
6749 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6750 : // this is most likely the background tasks, but it might be the spawned task from
6751 : // immediate_gc
6752 377 : let _g = tokio::select! {
6753 377 : guard = self.gc_lock.lock() => guard,
6754 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6755 : };
6756 377 : let timer = self.metrics.garbage_collect_histo.start_timer();
6757 :
6758 377 : fail_point!("before-timeline-gc");
6759 :
6760 : // Is the timeline being deleted?
6761 377 : if self.is_stopping() {
6762 0 : return Err(GcError::TimelineCancelled);
6763 377 : }
6764 :
6765 377 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6766 377 : let gc_info = self.gc_info.read().unwrap();
6767 :
6768 377 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6769 377 : let time_cutoff = gc_info.cutoffs.time;
6770 377 : let retain_lsns = gc_info
6771 377 : .retain_lsns
6772 377 : .iter()
6773 377 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6774 377 : .collect();
6775 :
6776 : // Gets the maximum LSN that holds the valid lease.
6777 : //
6778 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6779 : // Here, we do not check for stale leases again.
6780 377 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6781 :
6782 377 : (
6783 377 : space_cutoff,
6784 377 : time_cutoff,
6785 377 : retain_lsns,
6786 377 : max_lsn_with_valid_lease,
6787 377 : )
6788 : };
6789 :
6790 377 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6791 377 : let standby_horizon = self.standby_horizon.load();
6792 : // Hold GC for the standby, but as a safety guard do it only within some
6793 : // reasonable lag.
6794 377 : if standby_horizon != Lsn::INVALID {
6795 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6796 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6797 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6798 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6799 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6800 : } else {
6801 0 : warn!(
6802 0 : "standby is lagging for more than {}MB, not holding gc for it",
6803 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6804 : )
6805 : }
6806 0 : }
6807 377 : }
6808 :
6809 : // Reset standby horizon to ignore it if it is not updated till next GC.
6810 : // It is an easy way to unset it when standby disappears without adding
6811 : // more conf options.
6812 377 : self.standby_horizon.store(Lsn::INVALID);
6813 377 : self.metrics
6814 377 : .standby_horizon_gauge
6815 377 : .set(Lsn::INVALID.0 as i64);
6816 :
6817 377 : let res = self
6818 377 : .gc_timeline(
6819 377 : space_cutoff,
6820 377 : time_cutoff,
6821 377 : retain_lsns,
6822 377 : max_lsn_with_valid_lease,
6823 377 : new_gc_cutoff,
6824 : )
6825 377 : .instrument(
6826 377 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6827 : )
6828 377 : .await?;
6829 :
6830 : // only record successes
6831 377 : timer.stop_and_record();
6832 :
6833 377 : Ok(res)
6834 377 : }
6835 :
6836 377 : async fn gc_timeline(
6837 377 : &self,
6838 377 : space_cutoff: Lsn,
6839 377 : time_cutoff: Option<Lsn>, // None if uninitialized
6840 377 : retain_lsns: Vec<Lsn>,
6841 377 : max_lsn_with_valid_lease: Option<Lsn>,
6842 377 : new_gc_cutoff: Lsn,
6843 377 : ) -> Result<GcResult, GcError> {
6844 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6845 :
6846 377 : let now = SystemTime::now();
6847 377 : let mut result: GcResult = GcResult::default();
6848 :
6849 : // Nothing to GC. Return early.
6850 377 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6851 377 : if latest_gc_cutoff >= new_gc_cutoff {
6852 11 : info!(
6853 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6854 : );
6855 11 : return Ok(result);
6856 366 : }
6857 :
6858 366 : let Some(time_cutoff) = time_cutoff else {
6859 : // The GC cutoff should have been computed by now, but let's be defensive.
6860 0 : info!("Nothing to GC: time_cutoff not yet computed");
6861 0 : return Ok(result);
6862 : };
6863 :
6864 : // We need to ensure that no one tries to read page versions or create
6865 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6866 : // for details. This will block until the old value is no longer in use.
6867 : //
6868 : // The GC cutoff should only ever move forwards.
6869 366 : let waitlist = {
6870 366 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6871 366 : if *write_guard > new_gc_cutoff {
6872 0 : return Err(GcError::BadLsn {
6873 0 : why: format!(
6874 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6875 0 : *write_guard, new_gc_cutoff
6876 0 : ),
6877 0 : });
6878 366 : }
6879 :
6880 366 : write_guard.store_and_unlock(new_gc_cutoff)
6881 : };
6882 366 : waitlist.wait().await;
6883 :
6884 366 : info!("GC starting");
6885 :
6886 366 : debug!("retain_lsns: {:?}", retain_lsns);
6887 :
6888 366 : let max_retain_lsn = retain_lsns.iter().max();
6889 :
6890 : // Scan all layers in the timeline (remote or on-disk).
6891 : //
6892 : // Garbage collect the layer if all conditions are satisfied:
6893 : // 1. it is older than cutoff LSN;
6894 : // 2. it is older than PITR interval;
6895 : // 3. it doesn't need to be retained for 'retain_lsns';
6896 : // 4. it does not need to be kept for LSNs holding valid leases.
6897 : // 5. newer on-disk image layers cover the layer's whole key range
6898 366 : let layers_to_remove = {
6899 366 : let mut layers_to_remove = Vec::new();
6900 :
6901 366 : let guard = self
6902 366 : .layers
6903 366 : .read(LayerManagerLockHolder::GarbageCollection)
6904 366 : .await;
6905 366 : let layers = guard.layer_map()?;
6906 6209 : 'outer: for l in layers.iter_historic_layers() {
6907 6209 : result.layers_total += 1;
6908 :
6909 : // 1. Is it newer than GC horizon cutoff point?
6910 6209 : if l.get_lsn_range().end > space_cutoff {
6911 371 : debug!(
6912 0 : "keeping {} because it's newer than space_cutoff {}",
6913 0 : l.layer_name(),
6914 : space_cutoff,
6915 : );
6916 371 : result.layers_needed_by_cutoff += 1;
6917 371 : continue 'outer;
6918 5838 : }
6919 :
6920 : // 2. It is newer than PiTR cutoff point?
6921 5838 : if l.get_lsn_range().end > time_cutoff {
6922 0 : debug!(
6923 0 : "keeping {} because it's newer than time_cutoff {}",
6924 0 : l.layer_name(),
6925 : time_cutoff,
6926 : );
6927 0 : result.layers_needed_by_pitr += 1;
6928 0 : continue 'outer;
6929 5838 : }
6930 :
6931 : // 3. Is it needed by a child branch?
6932 : // NOTE With that we would keep data that
6933 : // might be referenced by child branches forever.
6934 : // We can track this in child timeline GC and delete parent layers when
6935 : // they are no longer needed. This might be complicated with long inheritance chains.
6936 5838 : if let Some(retain_lsn) = max_retain_lsn {
6937 : // start_lsn is inclusive
6938 6 : if &l.get_lsn_range().start <= retain_lsn {
6939 6 : debug!(
6940 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6941 0 : l.layer_name(),
6942 : retain_lsn,
6943 0 : l.is_incremental(),
6944 : );
6945 6 : result.layers_needed_by_branches += 1;
6946 6 : continue 'outer;
6947 0 : }
6948 5832 : }
6949 :
6950 : // 4. Is there a valid lease that requires us to keep this layer?
6951 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
6952 : // keep if layer start <= any of the lease
6953 9 : if &l.get_lsn_range().start <= lsn {
6954 7 : debug!(
6955 0 : "keeping {} because there is a valid lease preventing GC at {}",
6956 0 : l.layer_name(),
6957 : lsn,
6958 : );
6959 7 : result.layers_needed_by_leases += 1;
6960 7 : continue 'outer;
6961 2 : }
6962 5823 : }
6963 :
6964 : // 5. Is there a later on-disk layer for this relation?
6965 : //
6966 : // The end-LSN is exclusive, while disk_consistent_lsn is
6967 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6968 : // OK for a delta layer to have end LSN 101, but if the end LSN
6969 : // is 102, then it might not have been fully flushed to disk
6970 : // before crash.
6971 : //
6972 : // For example, imagine that the following layers exist:
6973 : //
6974 : // 1000 - image (A)
6975 : // 1000-2000 - delta (B)
6976 : // 2000 - image (C)
6977 : // 2000-3000 - delta (D)
6978 : // 3000 - image (E)
6979 : //
6980 : // If GC horizon is at 2500, we can remove layers A and B, but
6981 : // we cannot remove C, even though it's older than 2500, because
6982 : // the delta layer 2000-3000 depends on it.
6983 5825 : if !layers
6984 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6985 : {
6986 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
6987 5821 : result.layers_not_updated += 1;
6988 5821 : continue 'outer;
6989 4 : }
6990 :
6991 : // We didn't find any reason to keep this file, so remove it.
6992 4 : info!(
6993 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6994 0 : l.layer_name(),
6995 0 : l.is_incremental(),
6996 : );
6997 4 : layers_to_remove.push(l);
6998 : }
6999 :
7000 366 : layers_to_remove
7001 : };
7002 :
7003 366 : if !layers_to_remove.is_empty() {
7004 : // Persist the new GC cutoff value before we actually remove anything.
7005 : // This unconditionally schedules also an index_part.json update, even though, we will
7006 : // be doing one a bit later with the unlinked gc'd layers.
7007 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
7008 3 : self.schedule_uploads(disk_consistent_lsn, None)
7009 3 : .map_err(|e| {
7010 0 : if self.cancel.is_cancelled() {
7011 0 : GcError::TimelineCancelled
7012 : } else {
7013 0 : GcError::Remote(e)
7014 : }
7015 0 : })?;
7016 :
7017 3 : let mut guard = self
7018 3 : .layers
7019 3 : .write(LayerManagerLockHolder::GarbageCollection)
7020 3 : .await;
7021 :
7022 3 : let gc_layers = layers_to_remove
7023 3 : .iter()
7024 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
7025 3 : .collect::<Vec<Layer>>();
7026 :
7027 3 : result.layers_removed = gc_layers.len() as u64;
7028 :
7029 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
7030 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
7031 :
7032 : #[cfg(feature = "testing")]
7033 3 : {
7034 3 : result.doomed_layers = gc_layers;
7035 3 : }
7036 363 : }
7037 :
7038 366 : info!(
7039 0 : "GC completed removing {} layers, cutoff {}",
7040 : result.layers_removed, new_gc_cutoff
7041 : );
7042 :
7043 366 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
7044 366 : Ok(result)
7045 377 : }
7046 :
7047 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
7048 364884 : pub(crate) async fn reconstruct_value(
7049 364884 : &self,
7050 364884 : key: Key,
7051 364884 : request_lsn: Lsn,
7052 364884 : mut data: ValueReconstructState,
7053 364884 : redo_attempt_type: RedoAttemptType,
7054 364884 : ) -> Result<Bytes, PageReconstructError> {
7055 : // Perform WAL redo if needed
7056 364884 : data.records.reverse();
7057 :
7058 364884 : let fire_critical_error = match redo_attempt_type {
7059 363551 : RedoAttemptType::ReadPage => true,
7060 0 : RedoAttemptType::LegacyCompaction => true,
7061 1333 : RedoAttemptType::GcCompaction => false,
7062 : };
7063 :
7064 : // If we have a page image, and no WAL, we're all set
7065 364884 : if data.records.is_empty() {
7066 338110 : if let Some((img_lsn, img)) = &data.img {
7067 338110 : trace!(
7068 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
7069 : key, img_lsn, request_lsn,
7070 : );
7071 338110 : Ok(img.clone())
7072 : } else {
7073 0 : Err(PageReconstructError::from(anyhow!(
7074 0 : "base image for {key} at {request_lsn} not found"
7075 0 : )))
7076 : }
7077 : } else {
7078 : // We need to do WAL redo.
7079 : //
7080 : // If we don't have a base image, then the oldest WAL record better initialize
7081 : // the page
7082 26774 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
7083 0 : Err(PageReconstructError::from(anyhow!(
7084 0 : "Base image for {} at {} not found, but got {} WAL records",
7085 0 : key,
7086 0 : request_lsn,
7087 0 : data.records.len()
7088 0 : )))
7089 : } else {
7090 26774 : if data.img.is_some() {
7091 13029 : trace!(
7092 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
7093 0 : data.records.len(),
7094 : key,
7095 : request_lsn
7096 : );
7097 : } else {
7098 13745 : trace!(
7099 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
7100 0 : data.records.len(),
7101 : key,
7102 : request_lsn
7103 : );
7104 : };
7105 26774 : let res = self
7106 26774 : .walredo_mgr
7107 26774 : .as_ref()
7108 26774 : .context("timeline has no walredo manager")
7109 26774 : .map_err(PageReconstructError::WalRedo)?
7110 26774 : .request_redo(
7111 26774 : key,
7112 26774 : request_lsn,
7113 26774 : data.img,
7114 26774 : data.records,
7115 26774 : self.pg_version,
7116 26774 : redo_attempt_type,
7117 : )
7118 26774 : .await;
7119 26773 : let img = match res {
7120 26773 : Ok(img) => img,
7121 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
7122 1 : Err(walredo::Error::Other(err)) => {
7123 1 : if fire_critical_error {
7124 0 : critical_timeline!(
7125 0 : self.tenant_shard_id,
7126 0 : self.timeline_id,
7127 0 : "walredo failure during page reconstruction: {err:?}"
7128 : );
7129 1 : }
7130 1 : return Err(PageReconstructError::WalRedo(
7131 1 : err.context("reconstruct a page image"),
7132 1 : ));
7133 : }
7134 : };
7135 26773 : Ok(img)
7136 : }
7137 : }
7138 364884 : }
7139 :
7140 0 : pub(crate) async fn spawn_download_all_remote_layers(
7141 0 : self: Arc<Self>,
7142 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7143 0 : ctx: &RequestContext,
7144 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
7145 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7146 :
7147 : // this is not really needed anymore; it has tests which really check the return value from
7148 : // http api. it would be better not to maintain this anymore.
7149 :
7150 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
7151 0 : if let Some(st) = &*status_guard {
7152 0 : match &st.state {
7153 : DownloadRemoteLayersTaskState::Running => {
7154 0 : return Err(st.clone());
7155 : }
7156 : DownloadRemoteLayersTaskState::ShutDown
7157 0 : | DownloadRemoteLayersTaskState::Completed => {
7158 0 : *status_guard = None;
7159 0 : }
7160 : }
7161 0 : }
7162 :
7163 0 : let self_clone = Arc::clone(&self);
7164 0 : let task_ctx = ctx.detached_child(
7165 0 : TaskKind::DownloadAllRemoteLayers,
7166 0 : DownloadBehavior::Download,
7167 : );
7168 0 : let task_id = task_mgr::spawn(
7169 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
7170 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
7171 0 : self.tenant_shard_id,
7172 0 : Some(self.timeline_id),
7173 0 : "download all remote layers task",
7174 0 : async move {
7175 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
7176 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
7177 0 : match &mut *status_guard {
7178 : None => {
7179 0 : warn!("tasks status is supposed to be Some(), since we are running");
7180 : }
7181 0 : Some(st) => {
7182 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
7183 0 : if st.task_id != exp_task_id {
7184 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
7185 0 : } else {
7186 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7187 0 : }
7188 : }
7189 : };
7190 0 : Ok(())
7191 0 : }
7192 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))
7193 : );
7194 :
7195 0 : let initial_info = DownloadRemoteLayersTaskInfo {
7196 0 : task_id: format!("{task_id}"),
7197 0 : state: DownloadRemoteLayersTaskState::Running,
7198 0 : total_layer_count: 0,
7199 0 : successful_download_count: 0,
7200 0 : failed_download_count: 0,
7201 0 : };
7202 0 : *status_guard = Some(initial_info.clone());
7203 :
7204 0 : Ok(initial_info)
7205 0 : }
7206 :
7207 0 : async fn download_all_remote_layers(
7208 0 : self: &Arc<Self>,
7209 0 : request: DownloadRemoteLayersTaskSpawnRequest,
7210 0 : ctx: &RequestContext,
7211 0 : ) {
7212 : use pageserver_api::models::DownloadRemoteLayersTaskState;
7213 :
7214 0 : let remaining = {
7215 0 : let guard = self
7216 0 : .layers
7217 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7218 0 : .await;
7219 0 : let Ok(lm) = guard.layer_map() else {
7220 : // technically here we could look into iterating accessible layers, but downloading
7221 : // all layers of a shutdown timeline makes no sense regardless.
7222 0 : tracing::info!("attempted to download all layers of shutdown timeline");
7223 0 : return;
7224 : };
7225 0 : lm.iter_historic_layers()
7226 0 : .map(|desc| guard.get_from_desc(&desc))
7227 0 : .collect::<Vec<_>>()
7228 : };
7229 0 : let total_layer_count = remaining.len();
7230 :
7231 : macro_rules! lock_status {
7232 : ($st:ident) => {
7233 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
7234 : let st = st
7235 : .as_mut()
7236 : .expect("this function is only called after the task has been spawned");
7237 : assert_eq!(
7238 : st.task_id,
7239 : format!(
7240 : "{}",
7241 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
7242 : )
7243 : );
7244 : let $st = st;
7245 : };
7246 : }
7247 :
7248 : {
7249 0 : lock_status!(st);
7250 0 : st.total_layer_count = total_layer_count as u64;
7251 : }
7252 :
7253 0 : let mut remaining = remaining.into_iter();
7254 0 : let mut have_remaining = true;
7255 0 : let mut js = tokio::task::JoinSet::new();
7256 :
7257 0 : let cancel = task_mgr::shutdown_token();
7258 :
7259 0 : let limit = request.max_concurrent_downloads;
7260 :
7261 : loop {
7262 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
7263 0 : let Some(next) = remaining.next() else {
7264 0 : have_remaining = false;
7265 0 : break;
7266 : };
7267 :
7268 0 : let span = tracing::info_span!("download", layer = %next);
7269 :
7270 0 : let ctx = ctx.attached_child();
7271 0 : js.spawn(
7272 0 : async move {
7273 0 : let res = next.download(&ctx).await;
7274 0 : (next, res)
7275 0 : }
7276 0 : .instrument(span),
7277 : );
7278 : }
7279 :
7280 0 : while let Some(res) = js.join_next().await {
7281 0 : match res {
7282 : Ok((_, Ok(_))) => {
7283 0 : lock_status!(st);
7284 0 : st.successful_download_count += 1;
7285 : }
7286 0 : Ok((layer, Err(e))) => {
7287 0 : tracing::error!(%layer, "download failed: {e:#}");
7288 0 : lock_status!(st);
7289 0 : st.failed_download_count += 1;
7290 : }
7291 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7292 0 : Err(je) if je.is_panic() => {
7293 0 : lock_status!(st);
7294 0 : st.failed_download_count += 1;
7295 : }
7296 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7297 : }
7298 : }
7299 :
7300 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7301 0 : break;
7302 0 : }
7303 : }
7304 :
7305 : {
7306 0 : lock_status!(st);
7307 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7308 : }
7309 0 : }
7310 :
7311 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7312 0 : &self,
7313 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7314 0 : self.download_all_remote_layers_task_info
7315 0 : .read()
7316 0 : .unwrap()
7317 0 : .clone()
7318 0 : }
7319 :
7320 : /* BEGIN_HADRON */
7321 0 : pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
7322 0 : let guard = self
7323 0 : .layers
7324 0 : .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
7325 0 : .await;
7326 0 : let layer_map = guard.layer_map()?;
7327 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
7328 :
7329 0 : Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
7330 0 : }
7331 : /* END_HADRON */
7332 : }
7333 :
7334 : impl Timeline {
7335 : /// Returns non-remote layers for eviction.
7336 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7337 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7338 0 : let mut max_layer_size: Option<u64> = None;
7339 :
7340 0 : let resident_layers = guard
7341 0 : .likely_resident_layers()
7342 0 : .map(|layer| {
7343 0 : let file_size = layer.layer_desc().file_size;
7344 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7345 :
7346 0 : let last_activity_ts = layer.latest_activity();
7347 :
7348 0 : EvictionCandidate {
7349 0 : layer: layer.to_owned().into(),
7350 0 : last_activity_ts,
7351 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7352 0 : visibility: layer.visibility(),
7353 0 : }
7354 0 : })
7355 0 : .collect();
7356 :
7357 0 : DiskUsageEvictionInfo {
7358 0 : max_layer_size,
7359 0 : resident_layers,
7360 0 : }
7361 0 : }
7362 :
7363 962 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7364 962 : ShardIndex {
7365 962 : shard_number: self.tenant_shard_id.shard_number,
7366 962 : shard_count: self.tenant_shard_id.shard_count,
7367 962 : }
7368 962 : }
7369 :
7370 : /// Persistently blocks gc for `Manual` reason.
7371 : ///
7372 : /// Returns true if no such block existed before, false otherwise.
7373 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7374 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7375 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7376 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7377 0 : }
7378 :
7379 : /// Persistently unblocks gc for `Manual` reason.
7380 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7381 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7382 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7383 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7384 0 : }
7385 :
7386 : #[cfg(test)]
7387 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7388 31 : self.last_record_lsn.advance(new_lsn);
7389 31 : }
7390 :
7391 : #[cfg(test)]
7392 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7393 2 : self.disk_consistent_lsn.store(new_value);
7394 2 : }
7395 :
7396 : /// Force create an image layer and place it into the layer map.
7397 : ///
7398 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7399 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7400 : /// placed into the layer map in one run AND be validated.
7401 : #[cfg(test)]
7402 36 : pub(super) async fn force_create_image_layer(
7403 36 : self: &Arc<Timeline>,
7404 36 : lsn: Lsn,
7405 36 : mut images: Vec<(Key, Bytes)>,
7406 36 : check_start_lsn: Option<Lsn>,
7407 36 : ctx: &RequestContext,
7408 36 : ) -> anyhow::Result<()> {
7409 36 : let last_record_lsn = self.get_last_record_lsn();
7410 36 : assert!(
7411 36 : lsn <= last_record_lsn,
7412 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7413 : );
7414 36 : if let Some(check_start_lsn) = check_start_lsn {
7415 36 : assert!(lsn >= check_start_lsn);
7416 0 : }
7417 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7418 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7419 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7420 36 : let mut image_layer_writer = ImageLayerWriter::new(
7421 36 : self.conf,
7422 36 : self.timeline_id,
7423 36 : self.tenant_shard_id,
7424 36 : &(min_key..end_key),
7425 36 : lsn,
7426 36 : &self.gate,
7427 36 : self.cancel.clone(),
7428 36 : ctx,
7429 36 : )
7430 36 : .await?;
7431 312 : for (key, img) in images {
7432 276 : image_layer_writer.put_image(key, img, ctx).await?;
7433 : }
7434 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7435 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7436 36 : info!("force created image layer {}", image_layer.local_path());
7437 : {
7438 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7439 36 : guard
7440 36 : .open_mut()
7441 36 : .unwrap()
7442 36 : .force_insert_layer(image_layer.clone());
7443 : }
7444 :
7445 : // Update remote_timeline_client state to reflect existence of this layer
7446 36 : self.remote_client
7447 36 : .schedule_layer_file_upload(image_layer)
7448 36 : .unwrap();
7449 :
7450 36 : Ok(())
7451 36 : }
7452 :
7453 : /// Force create a delta layer and place it into the layer map.
7454 : ///
7455 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7456 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7457 : /// placed into the layer map in one run AND be validated.
7458 : #[cfg(test)]
7459 50 : pub(super) async fn force_create_delta_layer(
7460 50 : self: &Arc<Timeline>,
7461 50 : mut deltas: DeltaLayerTestDesc,
7462 50 : check_start_lsn: Option<Lsn>,
7463 50 : ctx: &RequestContext,
7464 50 : ) -> anyhow::Result<()> {
7465 50 : let last_record_lsn = self.get_last_record_lsn();
7466 50 : deltas
7467 50 : .data
7468 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7469 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7470 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7471 10464 : for (_, lsn, _) in &deltas.data {
7472 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7473 : }
7474 50 : assert!(
7475 50 : deltas.lsn_range.end <= last_record_lsn,
7476 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7477 : deltas.lsn_range.end,
7478 : last_record_lsn
7479 : );
7480 50 : if let Some(check_start_lsn) = check_start_lsn {
7481 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7482 0 : }
7483 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7484 50 : self.conf,
7485 50 : self.timeline_id,
7486 50 : self.tenant_shard_id,
7487 50 : deltas.key_range.start,
7488 50 : deltas.lsn_range,
7489 50 : &self.gate,
7490 50 : self.cancel.clone(),
7491 50 : ctx,
7492 50 : )
7493 50 : .await?;
7494 10464 : for (key, lsn, val) in deltas.data {
7495 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7496 : }
7497 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7498 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7499 50 : info!("force created delta layer {}", delta_layer.local_path());
7500 : {
7501 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7502 50 : guard
7503 50 : .open_mut()
7504 50 : .unwrap()
7505 50 : .force_insert_layer(delta_layer.clone());
7506 : }
7507 :
7508 : // Update remote_timeline_client state to reflect existence of this layer
7509 50 : self.remote_client
7510 50 : .schedule_layer_file_upload(delta_layer)
7511 50 : .unwrap();
7512 :
7513 50 : Ok(())
7514 50 : }
7515 :
7516 : /// Force create an in-memory layer and place them into the layer map.
7517 : #[cfg(test)]
7518 4 : pub(super) async fn force_create_in_memory_layer(
7519 4 : self: &Arc<Timeline>,
7520 4 : mut in_memory: InMemoryLayerTestDesc,
7521 4 : check_start_lsn: Option<Lsn>,
7522 4 : ctx: &RequestContext,
7523 4 : ) -> anyhow::Result<()> {
7524 : use utils::bin_ser::BeSer;
7525 :
7526 : // Validate LSNs
7527 4 : if let Some(check_start_lsn) = check_start_lsn {
7528 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7529 0 : }
7530 :
7531 4 : let last_record_lsn = self.get_last_record_lsn();
7532 4 : let layer_end_lsn = if in_memory.is_open {
7533 1 : in_memory
7534 1 : .data
7535 1 : .iter()
7536 1 : .map(|(_key, lsn, _value)| lsn)
7537 1 : .max()
7538 1 : .cloned()
7539 : } else {
7540 3 : Some(in_memory.lsn_range.end)
7541 : };
7542 :
7543 4 : if let Some(end) = layer_end_lsn {
7544 4 : assert!(
7545 4 : end <= last_record_lsn,
7546 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7547 : );
7548 0 : }
7549 :
7550 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7551 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7552 19820 : assert!(*lsn < in_memory.lsn_range.end);
7553 19820 : });
7554 :
7555 : // Build the batch
7556 4 : in_memory
7557 4 : .data
7558 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7559 :
7560 4 : let data = in_memory
7561 4 : .data
7562 4 : .into_iter()
7563 19820 : .map(|(key, lsn, value)| {
7564 19820 : let value_size = value.serialized_size().unwrap() as usize;
7565 19820 : (key.to_compact(), lsn, value_size, value)
7566 19820 : })
7567 4 : .collect::<Vec<_>>();
7568 :
7569 4 : let batch = SerializedValueBatch::from_values(data);
7570 :
7571 : // Create the in-memory layer and write the batch into it
7572 4 : let layer = InMemoryLayer::create(
7573 4 : self.conf,
7574 4 : self.timeline_id,
7575 4 : self.tenant_shard_id,
7576 4 : in_memory.lsn_range.start,
7577 4 : &self.gate,
7578 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7579 4 : &CancellationToken::new(),
7580 4 : ctx,
7581 4 : )
7582 4 : .await
7583 4 : .unwrap();
7584 :
7585 4 : layer.put_batch(batch, ctx).await.unwrap();
7586 4 : if !in_memory.is_open {
7587 3 : layer.freeze(in_memory.lsn_range.end).await;
7588 1 : }
7589 :
7590 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7591 :
7592 : // Link the layer to the layer map
7593 : {
7594 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7595 4 : let layer_map = guard.open_mut().unwrap();
7596 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7597 : }
7598 :
7599 4 : Ok(())
7600 4 : }
7601 :
7602 : /// Return all keys at the LSN in the image layers
7603 : #[cfg(test)]
7604 3 : pub(crate) async fn inspect_image_layers(
7605 3 : self: &Arc<Timeline>,
7606 3 : lsn: Lsn,
7607 3 : ctx: &RequestContext,
7608 3 : io_concurrency: IoConcurrency,
7609 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7610 3 : let mut all_data = Vec::new();
7611 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7612 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7613 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7614 4 : let layer = guard.get_from_desc(&layer);
7615 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7616 4 : layer
7617 4 : .get_values_reconstruct_data(
7618 4 : KeySpace::single(Key::MIN..Key::MAX),
7619 4 : lsn..Lsn(lsn.0 + 1),
7620 4 : &mut reconstruct_data,
7621 4 : ctx,
7622 4 : )
7623 4 : .await?;
7624 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7625 33 : let v = v.collect_pending_ios().await?;
7626 33 : all_data.push((k, v.img.unwrap().1));
7627 : }
7628 13 : }
7629 : }
7630 3 : all_data.sort();
7631 3 : Ok(all_data)
7632 3 : }
7633 :
7634 : /// Get all historic layer descriptors in the layer map
7635 : #[cfg(test)]
7636 12 : pub(crate) async fn inspect_historic_layers(
7637 12 : self: &Arc<Timeline>,
7638 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7639 12 : let mut layers = Vec::new();
7640 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7641 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7642 57 : layers.push(layer.key());
7643 57 : }
7644 12 : Ok(layers)
7645 12 : }
7646 :
7647 : #[cfg(test)]
7648 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7649 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7650 5 : keyspace.merge(&ks);
7651 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7652 5 : }
7653 : }
7654 :
7655 : /// Tracking writes ingestion does to a particular in-memory layer.
7656 : ///
7657 : /// Cleared upon freezing a layer.
7658 : pub(crate) struct TimelineWriterState {
7659 : open_layer: Arc<InMemoryLayer>,
7660 : current_size: u64,
7661 : // Previous Lsn which passed through
7662 : prev_lsn: Option<Lsn>,
7663 : // Largest Lsn which passed through the current writer
7664 : max_lsn: Option<Lsn>,
7665 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7666 : cached_last_freeze_at: Lsn,
7667 : }
7668 :
7669 : impl TimelineWriterState {
7670 660 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7671 660 : Self {
7672 660 : open_layer,
7673 660 : current_size,
7674 660 : prev_lsn: None,
7675 660 : max_lsn: None,
7676 660 : cached_last_freeze_at: last_freeze_at,
7677 660 : }
7678 660 : }
7679 : }
7680 :
7681 : /// Various functions to mutate the timeline.
7682 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7683 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7684 : // but will cause large code changes.
7685 : pub(crate) struct TimelineWriter<'a> {
7686 : tl: &'a Timeline,
7687 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7688 : }
7689 :
7690 : impl Deref for TimelineWriter<'_> {
7691 : type Target = Timeline;
7692 :
7693 4949226 : fn deref(&self) -> &Self::Target {
7694 4949226 : self.tl
7695 4949226 : }
7696 : }
7697 :
7698 : #[derive(PartialEq)]
7699 : enum OpenLayerAction {
7700 : Roll,
7701 : Open,
7702 : None,
7703 : }
7704 :
7705 : impl TimelineWriter<'_> {
7706 2402129 : async fn handle_open_layer_action(
7707 2402129 : &mut self,
7708 2402129 : at: Lsn,
7709 2402129 : action: OpenLayerAction,
7710 2402129 : ctx: &RequestContext,
7711 2402129 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7712 2402129 : match action {
7713 : OpenLayerAction::Roll => {
7714 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7715 40 : self.roll_layer(freeze_at).await?;
7716 40 : self.open_layer(at, ctx).await?;
7717 : }
7718 620 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7719 : OpenLayerAction::None => {
7720 2401469 : assert!(self.write_guard.is_some());
7721 : }
7722 : }
7723 :
7724 2402129 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7725 2402129 : }
7726 :
7727 660 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7728 660 : let layer = self
7729 660 : .tl
7730 660 : .get_layer_for_write(at, &self.write_guard, ctx)
7731 660 : .await?;
7732 660 : let initial_size = layer.len();
7733 :
7734 660 : let last_freeze_at = self.last_freeze_at.load();
7735 660 : self.write_guard.replace(TimelineWriterState::new(
7736 660 : layer,
7737 660 : initial_size,
7738 660 : last_freeze_at,
7739 660 : ));
7740 :
7741 660 : Ok(())
7742 660 : }
7743 :
7744 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7745 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7746 :
7747 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7748 : // to propagate the backpressure up into WAL ingestion.
7749 40 : let l0_count = self
7750 40 : .tl
7751 40 : .layers
7752 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7753 40 : .await
7754 40 : .layer_map()?
7755 40 : .level0_deltas()
7756 40 : .len();
7757 40 : let wait_thresholds = [
7758 40 : self.get_l0_flush_delay_threshold(),
7759 40 : self.get_l0_flush_stall_threshold(),
7760 40 : ];
7761 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7762 :
7763 : // self.write_guard will be taken by the freezing
7764 40 : let flush_id = self
7765 40 : .tl
7766 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7767 40 : .await?;
7768 :
7769 40 : assert!(self.write_guard.is_none());
7770 :
7771 40 : if let Some(wait_threshold) = wait_threshold {
7772 0 : if l0_count >= wait_threshold {
7773 0 : debug!(
7774 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7775 : );
7776 0 : self.tl.wait_flush_completion(flush_id).await?;
7777 0 : }
7778 40 : }
7779 :
7780 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7781 0 : warn!("Flushed oversized open layer with size {}", current_size)
7782 40 : }
7783 :
7784 40 : Ok(())
7785 40 : }
7786 :
7787 2402129 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7788 2402129 : let state = &*self.write_guard;
7789 2402129 : let Some(state) = &state else {
7790 620 : return OpenLayerAction::Open;
7791 : };
7792 :
7793 : #[cfg(feature = "testing")]
7794 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7795 : // this check and assertion are not really needed because
7796 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7797 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7798 : // is no TimelineWriterState.
7799 0 : assert!(
7800 0 : state.open_layer.end_lsn.get().is_some(),
7801 0 : "our open_layer must be outdated"
7802 : );
7803 :
7804 : // this would be a memory leak waiting to happen because the in-memory layer always has
7805 : // an index
7806 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7807 2401509 : }
7808 :
7809 2401509 : if state.prev_lsn == Some(lsn) {
7810 : // Rolling mid LSN is not supported by [downstream code].
7811 : // Hence, only roll at LSN boundaries.
7812 : //
7813 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7814 3 : return OpenLayerAction::None;
7815 2401506 : }
7816 :
7817 2401506 : if state.current_size == 0 {
7818 : // Don't roll empty layers
7819 0 : return OpenLayerAction::None;
7820 2401506 : }
7821 :
7822 2401506 : if self.tl.should_roll(
7823 2401506 : state.current_size,
7824 2401506 : state.current_size + new_value_size,
7825 2401506 : self.get_checkpoint_distance(),
7826 2401506 : lsn,
7827 2401506 : state.cached_last_freeze_at,
7828 2401506 : state.open_layer.get_opened_at(),
7829 : ) {
7830 40 : OpenLayerAction::Roll
7831 : } else {
7832 2401466 : OpenLayerAction::None
7833 : }
7834 2402129 : }
7835 :
7836 : /// Put a batch of keys at the specified Lsns.
7837 2402128 : pub(crate) async fn put_batch(
7838 2402128 : &mut self,
7839 2402128 : batch: SerializedValueBatch,
7840 2402128 : ctx: &RequestContext,
7841 2402128 : ) -> anyhow::Result<()> {
7842 2402128 : if !batch.has_data() {
7843 0 : return Ok(());
7844 2402128 : }
7845 :
7846 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7847 : // We don't assert this in release builds, since key ownership policies may change over
7848 : // time. Stray keys will be removed during compaction.
7849 2402128 : if cfg!(debug_assertions) {
7850 4947529 : for metadata in &batch.metadata {
7851 2545401 : if let ValueMeta::Serialized(metadata) = metadata {
7852 2545401 : let key = Key::from_compact(metadata.key);
7853 2545401 : assert!(
7854 2545401 : self.shard_identity.is_key_local(&key)
7855 12 : || self.shard_identity.is_key_global(&key),
7856 0 : "key {key} does not belong on shard {}",
7857 0 : self.shard_identity.shard_index()
7858 : );
7859 0 : }
7860 : }
7861 0 : }
7862 :
7863 2402128 : let batch_max_lsn = batch.max_lsn;
7864 2402128 : let buf_size: u64 = batch.buffer_size() as u64;
7865 :
7866 2402128 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7867 2402128 : let layer = self
7868 2402128 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7869 2402128 : .await?;
7870 :
7871 2402128 : let res = layer.put_batch(batch, ctx).await;
7872 :
7873 2402128 : if res.is_ok() {
7874 2402128 : // Update the current size only when the entire write was ok.
7875 2402128 : // In case of failures, we may have had partial writes which
7876 2402128 : // render the size tracking out of sync. That's ok because
7877 2402128 : // the checkpoint distance should be significantly smaller
7878 2402128 : // than the S3 single shot upload limit of 5GiB.
7879 2402128 : let state = self.write_guard.as_mut().unwrap();
7880 2402128 :
7881 2402128 : state.current_size += buf_size;
7882 2402128 : state.prev_lsn = Some(batch_max_lsn);
7883 2402128 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7884 2402128 : }
7885 :
7886 2402128 : res
7887 2402128 : }
7888 :
7889 : #[cfg(test)]
7890 : /// Test helper, for tests that would like to poke individual values without composing a batch
7891 2195079 : pub(crate) async fn put(
7892 2195079 : &mut self,
7893 2195079 : key: Key,
7894 2195079 : lsn: Lsn,
7895 2195079 : value: &Value,
7896 2195079 : ctx: &RequestContext,
7897 2195079 : ) -> anyhow::Result<()> {
7898 : use utils::bin_ser::BeSer;
7899 2195079 : if !key.is_valid_key_on_write_path() {
7900 0 : bail!(
7901 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7902 : key
7903 : );
7904 2195079 : }
7905 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7906 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7907 2195079 : key.to_compact(),
7908 2195079 : lsn,
7909 2195079 : val_ser_size,
7910 2195079 : value.clone(),
7911 2195079 : )]);
7912 :
7913 2195079 : self.put_batch(batch, ctx).await
7914 2195079 : }
7915 :
7916 1 : pub(crate) async fn delete_batch(
7917 1 : &mut self,
7918 1 : batch: &[(Range<Key>, Lsn)],
7919 1 : ctx: &RequestContext,
7920 1 : ) -> anyhow::Result<()> {
7921 1 : if let Some((_, lsn)) = batch.first() {
7922 1 : let action = self.get_open_layer_action(*lsn, 0);
7923 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7924 1 : layer.put_tombstones(batch).await?;
7925 0 : }
7926 :
7927 1 : Ok(())
7928 1 : }
7929 :
7930 : /// Track the end of the latest digested WAL record.
7931 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7932 : ///
7933 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7934 : ///
7935 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7936 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7937 : /// the latest and can be read.
7938 2639560 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7939 2639560 : self.tl.finish_write(new_lsn);
7940 2639560 : }
7941 :
7942 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7943 135285 : self.tl.update_current_logical_size(delta)
7944 135285 : }
7945 : }
7946 :
7947 : // We need TimelineWriter to be send in upcoming conversion of
7948 : // Timeline::layers to tokio::sync::RwLock.
7949 : #[test]
7950 1 : fn is_send() {
7951 1 : fn _assert_send<T: Send>() {}
7952 1 : _assert_send::<TimelineWriter<'_>>();
7953 1 : }
7954 :
7955 : #[cfg(test)]
7956 : mod tests {
7957 : use std::sync::Arc;
7958 :
7959 : use pageserver_api::key::Key;
7960 : use postgres_ffi::PgMajorVersion;
7961 : use std::iter::Iterator;
7962 : use tracing::Instrument;
7963 : use utils::id::TimelineId;
7964 : use utils::lsn::Lsn;
7965 : use wal_decoder::models::value::Value;
7966 :
7967 : use super::HeatMapTimeline;
7968 : use crate::context::RequestContextBuilder;
7969 : use crate::tenant::harness::{TenantHarness, test_img};
7970 : use crate::tenant::layer_map::LayerMap;
7971 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7972 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
7973 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7974 : use crate::tenant::{PreviousHeatmap, Timeline};
7975 :
7976 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7977 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7978 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7979 25 : for (l, r) in lhs_rhs {
7980 20 : assert_eq!(l.name, r.name);
7981 20 : assert_eq!(l.metadata, r.metadata);
7982 : }
7983 5 : }
7984 :
7985 : #[tokio::test]
7986 1 : async fn test_heatmap_generation() {
7987 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7988 :
7989 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7990 1 : Lsn(0x10)..Lsn(0x20),
7991 1 : vec![(
7992 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7993 1 : Lsn(0x11),
7994 1 : Value::Image(test_img("foo")),
7995 1 : )],
7996 : );
7997 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7998 1 : Lsn(0x10)..Lsn(0x20),
7999 1 : vec![(
8000 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8001 1 : Lsn(0x11),
8002 1 : Value::Image(test_img("foo")),
8003 1 : )],
8004 : );
8005 1 : let l0_delta = DeltaLayerTestDesc::new(
8006 1 : Lsn(0x20)..Lsn(0x30),
8007 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8008 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8009 1 : vec![(
8010 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8011 1 : Lsn(0x25),
8012 1 : Value::Image(test_img("foo")),
8013 1 : )],
8014 : );
8015 1 : let delta_layers = vec![
8016 1 : covered_delta.clone(),
8017 1 : visible_delta.clone(),
8018 1 : l0_delta.clone(),
8019 : ];
8020 :
8021 1 : let image_layer = (
8022 1 : Lsn(0x40),
8023 1 : vec![(
8024 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8025 1 : test_img("bar"),
8026 1 : )],
8027 1 : );
8028 1 : let image_layers = vec![image_layer];
8029 :
8030 1 : let (tenant, ctx) = harness.load().await;
8031 1 : let timeline = tenant
8032 1 : .create_test_timeline_with_layers(
8033 1 : TimelineId::generate(),
8034 1 : Lsn(0x10),
8035 1 : PgMajorVersion::PG14,
8036 1 : &ctx,
8037 1 : Vec::new(), // in-memory layers
8038 1 : delta_layers,
8039 1 : image_layers,
8040 1 : Lsn(0x100),
8041 1 : )
8042 1 : .await
8043 1 : .unwrap();
8044 1 : let ctx = &ctx.with_scope_timeline(&timeline);
8045 :
8046 : // Layer visibility is an input to heatmap generation, so refresh it first
8047 1 : timeline.update_layer_visibility().await.unwrap();
8048 :
8049 1 : let heatmap = timeline
8050 1 : .generate_heatmap()
8051 1 : .await
8052 1 : .expect("Infallible while timeline is not shut down");
8053 :
8054 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
8055 :
8056 : // L0 should come last
8057 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
8058 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
8059 :
8060 1 : let mut last_lsn = Lsn::MAX;
8061 5 : for layer in heatmap_layers {
8062 : // Covered layer should be omitted
8063 4 : assert!(layer.name != covered_delta.layer_name());
8064 :
8065 4 : let layer_lsn = match &layer.name {
8066 2 : LayerName::Delta(d) => d.lsn_range.end,
8067 2 : LayerName::Image(i) => i.lsn,
8068 : };
8069 :
8070 : // Apart from L0s, newest Layers should come first
8071 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
8072 3 : assert!(layer_lsn <= last_lsn);
8073 3 : last_lsn = layer_lsn;
8074 1 : }
8075 : }
8076 :
8077 : // Evict all the layers and stash the old heatmap in the timeline.
8078 : // This simulates a migration to a cold secondary location.
8079 :
8080 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8081 1 : let mut all_layers = Vec::new();
8082 1 : let forever = std::time::Duration::from_secs(120);
8083 5 : for layer in guard.likely_resident_layers() {
8084 5 : all_layers.push(layer.clone());
8085 5 : layer.evict_and_wait(forever).await.unwrap();
8086 : }
8087 1 : drop(guard);
8088 :
8089 1 : timeline
8090 1 : .previous_heatmap
8091 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8092 1 : heatmap: heatmap.clone(),
8093 1 : read_at: std::time::Instant::now(),
8094 1 : end_lsn: None,
8095 1 : })));
8096 :
8097 : // Generate a new heatmap and assert that it contains the same layers as the old one.
8098 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
8099 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
8100 :
8101 : // Download each layer one by one. Generate the heatmap at each step and check
8102 : // that it's stable.
8103 6 : for layer in all_layers {
8104 5 : if layer.visibility() == LayerVisibilityHint::Covered {
8105 1 : continue;
8106 4 : }
8107 1 :
8108 4 : eprintln!("Downloading {layer} and re-generating heatmap");
8109 1 :
8110 4 : let ctx = &RequestContextBuilder::from(ctx)
8111 4 : .download_behavior(crate::context::DownloadBehavior::Download)
8112 4 : .attached_child();
8113 1 :
8114 4 : let _resident = layer
8115 4 : .download_and_keep_resident(ctx)
8116 4 : .instrument(tracing::info_span!(
8117 4 : parent: None,
8118 1 : "download_layer",
8119 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
8120 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
8121 1 : timeline_id = %timeline.timeline_id
8122 1 : ))
8123 4 : .await
8124 4 : .unwrap();
8125 1 :
8126 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
8127 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
8128 1 : }
8129 1 :
8130 1 : // Everything from the post-migration heatmap is now resident.
8131 1 : // Check that we drop it from memory.
8132 1 : assert!(matches!(
8133 1 : timeline.previous_heatmap.load().as_deref(),
8134 1 : Some(PreviousHeatmap::Obsolete)
8135 1 : ));
8136 1 : }
8137 :
8138 : #[tokio::test]
8139 1 : async fn test_previous_heatmap_obsoletion() {
8140 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
8141 1 : .await
8142 1 : .unwrap();
8143 :
8144 1 : let l0_delta = DeltaLayerTestDesc::new(
8145 1 : Lsn(0x20)..Lsn(0x30),
8146 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
8147 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
8148 1 : vec![(
8149 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
8150 1 : Lsn(0x25),
8151 1 : Value::Image(test_img("foo")),
8152 1 : )],
8153 : );
8154 :
8155 1 : let image_layer = (
8156 1 : Lsn(0x40),
8157 1 : vec![(
8158 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
8159 1 : test_img("bar"),
8160 1 : )],
8161 1 : );
8162 :
8163 1 : let delta_layers = vec![l0_delta];
8164 1 : let image_layers = vec![image_layer];
8165 :
8166 1 : let (tenant, ctx) = harness.load().await;
8167 1 : let timeline = tenant
8168 1 : .create_test_timeline_with_layers(
8169 1 : TimelineId::generate(),
8170 1 : Lsn(0x10),
8171 1 : PgMajorVersion::PG14,
8172 1 : &ctx,
8173 1 : Vec::new(), // in-memory layers
8174 1 : delta_layers,
8175 1 : image_layers,
8176 1 : Lsn(0x100),
8177 1 : )
8178 1 : .await
8179 1 : .unwrap();
8180 :
8181 : // Layer visibility is an input to heatmap generation, so refresh it first
8182 1 : timeline.update_layer_visibility().await.unwrap();
8183 :
8184 1 : let heatmap = timeline
8185 1 : .generate_heatmap()
8186 1 : .await
8187 1 : .expect("Infallible while timeline is not shut down");
8188 :
8189 : // Both layers should be in the heatmap
8190 1 : assert!(heatmap.all_layers().count() > 0);
8191 :
8192 : // Now simulate a migration.
8193 1 : timeline
8194 1 : .previous_heatmap
8195 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
8196 1 : heatmap: heatmap.clone(),
8197 1 : read_at: std::time::Instant::now(),
8198 1 : end_lsn: None,
8199 1 : })));
8200 :
8201 : // Evict all the layers in the previous heatmap
8202 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
8203 1 : let forever = std::time::Duration::from_secs(120);
8204 3 : for layer in guard.likely_resident_layers() {
8205 3 : layer.evict_and_wait(forever).await.unwrap();
8206 : }
8207 1 : drop(guard);
8208 :
8209 : // Generate a new heatmap and check that the previous heatmap
8210 : // has been marked obsolete.
8211 1 : let post_eviction_heatmap = timeline
8212 1 : .generate_heatmap()
8213 1 : .await
8214 1 : .expect("Infallible while timeline is not shut down");
8215 :
8216 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
8217 1 : assert!(matches!(
8218 1 : timeline.previous_heatmap.load().as_deref(),
8219 1 : Some(PreviousHeatmap::Obsolete)
8220 1 : ));
8221 1 : }
8222 :
8223 : #[tokio::test]
8224 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
8225 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
8226 1 : .await
8227 1 : .unwrap();
8228 :
8229 1 : let (tenant, ctx) = harness.load().await;
8230 1 : let timeline = tenant
8231 1 : .create_test_timeline(
8232 1 : TimelineId::generate(),
8233 1 : Lsn(0x10),
8234 1 : PgMajorVersion::PG14,
8235 1 : &ctx,
8236 1 : )
8237 1 : .await
8238 1 : .unwrap();
8239 :
8240 1 : let layer = find_some_layer(&timeline).await;
8241 1 : let layer = layer
8242 1 : .keep_resident()
8243 1 : .await
8244 1 : .expect("no download => no downloading errors")
8245 1 : .drop_eviction_guard();
8246 :
8247 1 : let forever = std::time::Duration::from_secs(120);
8248 :
8249 1 : let first = layer.evict_and_wait(forever);
8250 1 : let second = layer.evict_and_wait(forever);
8251 :
8252 1 : let (first, second) = tokio::join!(first, second);
8253 :
8254 1 : let res = layer.keep_resident().await;
8255 1 : assert!(res.is_none(), "{res:?}");
8256 :
8257 1 : match (first, second) {
8258 1 : (Ok(()), Ok(())) => {
8259 1 : // because there are no more timeline locks being taken on eviction path, we can
8260 1 : // witness all three outcomes here.
8261 1 : }
8262 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
8263 0 : // if one completes before the other, this is fine just as well.
8264 0 : }
8265 1 : other => unreachable!("unexpected {:?}", other),
8266 1 : }
8267 1 : }
8268 :
8269 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
8270 1 : let layers = timeline
8271 1 : .layers
8272 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
8273 1 : .await;
8274 1 : let desc = layers
8275 1 : .layer_map()
8276 1 : .unwrap()
8277 1 : .iter_historic_layers()
8278 1 : .next()
8279 1 : .expect("must find one layer to evict");
8280 :
8281 1 : layers.get_from_desc(&desc)
8282 1 : }
8283 : }
|
