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::config::AttachmentMode;
122 : use crate::tenant::gc_result::GcResult;
123 : use crate::tenant::layer_map::LayerMap;
124 : use crate::tenant::metadata::TimelineMetadata;
125 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
126 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
127 : use crate::tenant::storage_layer::{
128 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
129 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
130 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
131 : ValueReconstructState, ValuesReconstructState,
132 : };
133 : use crate::tenant::tasks::BackgroundLoopKind;
134 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
135 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
136 : use crate::walingest::WalLagCooldown;
137 : use crate::walredo::RedoAttemptType;
138 : use crate::{ZERO_PAGE, task_mgr, walredo};
139 :
140 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
141 : pub(crate) enum FlushLoopState {
142 : NotStarted,
143 : Running {
144 : #[cfg(test)]
145 : expect_initdb_optimization: bool,
146 : #[cfg(test)]
147 : initdb_optimization_count: usize,
148 : },
149 : Exited,
150 : }
151 :
152 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
153 : pub enum ImageLayerCreationMode {
154 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
155 : Try,
156 : /// Force creating the image layers if possible. For now, no image layers will be created
157 : /// for metadata keys. Used in compaction code path with force flag enabled.
158 : Force,
159 : /// Initial ingestion of the data, and no data should be dropped in this function. This
160 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
161 : /// code path.
162 : Initial,
163 : }
164 :
165 : #[derive(Clone, Debug, Default)]
166 : pub enum LastImageLayerCreationStatus {
167 : Incomplete {
168 : /// The last key of the partition (exclusive) that was processed in the last
169 : /// image layer creation attempt. We will continue from this key in the next
170 : /// attempt.
171 : last_key: Key,
172 : },
173 : Complete,
174 : #[default]
175 : Initial,
176 : }
177 :
178 : impl std::fmt::Display for ImageLayerCreationMode {
179 190 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
180 190 : write!(f, "{self:?}")
181 190 : }
182 : }
183 :
184 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
185 : /// Can be removed after all refactors are done.
186 23 : fn drop_layer_manager_rlock(rlock: LayerManagerReadGuard<'_>) {
187 23 : drop(rlock)
188 23 : }
189 :
190 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
191 : /// Can be removed after all refactors are done.
192 213 : fn drop_layer_manager_wlock(rlock: LayerManagerWriteGuard<'_>) {
193 213 : drop(rlock)
194 213 : }
195 :
196 : /// The outward-facing resources required to build a Timeline
197 : pub struct TimelineResources {
198 : pub remote_client: RemoteTimelineClient,
199 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
200 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
201 : pub l0_compaction_trigger: Arc<Notify>,
202 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
203 : pub basebackup_cache: Arc<BasebackupCache>,
204 : pub feature_resolver: Arc<TenantFeatureResolver>,
205 : }
206 :
207 : pub struct Timeline {
208 : pub(crate) conf: &'static PageServerConf,
209 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
210 :
211 : myself: Weak<Self>,
212 :
213 : pub(crate) tenant_shard_id: TenantShardId,
214 : pub timeline_id: TimelineId,
215 :
216 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
217 : /// Never changes for the lifetime of this [`Timeline`] object.
218 : ///
219 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
220 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
221 : pub(crate) generation: Generation,
222 :
223 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
224 : /// to shards, and is constant through the lifetime of this Timeline.
225 : shard_identity: ShardIdentity,
226 :
227 : pub pg_version: PgMajorVersion,
228 :
229 : /// The tuple has two elements.
230 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
231 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
232 : ///
233 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
234 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
235 : ///
236 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
237 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
238 : /// `PersistentLayerDesc`'s.
239 : ///
240 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
241 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
242 : /// runtime, e.g., during page reconstruction.
243 : ///
244 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
245 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
246 : pub(crate) layers: LockedLayerManager,
247 :
248 : last_freeze_at: AtomicLsn,
249 : // Atomic would be more appropriate here.
250 : last_freeze_ts: RwLock<Instant>,
251 :
252 : pub(crate) standby_horizon: AtomicLsn,
253 :
254 : // WAL redo manager. `None` only for broken tenants.
255 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
256 :
257 : /// Remote storage client.
258 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
259 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
260 :
261 : // What page versions do we hold in the repository? If we get a
262 : // request > last_record_lsn, we need to wait until we receive all
263 : // the WAL up to the request. The SeqWait provides functions for
264 : // that. TODO: If we get a request for an old LSN, such that the
265 : // versions have already been garbage collected away, we should
266 : // throw an error, but we don't track that currently.
267 : //
268 : // last_record_lsn.load().last points to the end of last processed WAL record.
269 : //
270 : // We also remember the starting point of the previous record in
271 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
272 : // first WAL record when the node is started up. But here, we just
273 : // keep track of it.
274 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
275 :
276 : // All WAL records have been processed and stored durably on files on
277 : // local disk, up to this LSN. On crash and restart, we need to re-process
278 : // the WAL starting from this point.
279 : //
280 : // Some later WAL records might have been processed and also flushed to disk
281 : // already, so don't be surprised to see some, but there's no guarantee on
282 : // them yet.
283 : disk_consistent_lsn: AtomicLsn,
284 :
285 : // Parent timeline that this timeline was branched from, and the LSN
286 : // of the branch point.
287 : ancestor_timeline: Option<Arc<Timeline>>,
288 : ancestor_lsn: Lsn,
289 :
290 : // The LSN of gc-compaction that was last applied to this timeline.
291 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
292 :
293 : pub(crate) metrics: Arc<TimelineMetrics>,
294 :
295 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
296 : // in `crate::page_service` writes these metrics.
297 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
298 :
299 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
300 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
301 :
302 : /// Ensures layers aren't frozen by checkpointer between
303 : /// [`Timeline::get_layer_for_write`] and layer reads.
304 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
305 : /// Must always be acquired before the layer map/individual layer lock
306 : /// to avoid deadlock.
307 : ///
308 : /// The state is cleared upon freezing.
309 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
310 :
311 : /// Used to avoid multiple `flush_loop` tasks running
312 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
313 :
314 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
315 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
316 : /// The flush cycle counter is sent back on the layer_flush_done channel when
317 : /// the flush finishes. You can use that to wait for the flush to finish.
318 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
319 : /// read by whoever sends an update
320 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
321 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
322 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
323 :
324 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
325 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
326 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
327 : // the planned GC cutoff.
328 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
329 :
330 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
331 :
332 : // List of child timelines and their branch points. This is needed to avoid
333 : // garbage collecting data that is still needed by the child timelines.
334 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
335 :
336 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
337 :
338 : // It may change across major versions so for simplicity
339 : // keep it after running initdb for a timeline.
340 : // It is needed in checks when we want to error on some operations
341 : // when they are requested for pre-initdb lsn.
342 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
343 : // though let's keep them both for better error visibility.
344 : pub initdb_lsn: Lsn,
345 :
346 : /// The repartitioning result. Allows a single writer and multiple readers.
347 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
348 :
349 : /// Configuration: how often should the partitioning be recalculated.
350 : repartition_threshold: u64,
351 :
352 : last_image_layer_creation_check_at: AtomicLsn,
353 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
354 :
355 : /// Current logical size of the "datadir", at the last LSN.
356 : current_logical_size: LogicalSize,
357 :
358 : /// Information about the last processed message by the WAL receiver,
359 : /// or None if WAL receiver has not received anything for this timeline
360 : /// yet.
361 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
362 : pub walreceiver: Mutex<Option<WalReceiver>>,
363 :
364 : /// Relation size cache
365 : pub(crate) rel_size_latest_cache: RwLock<HashMap<RelTag, (Lsn, BlockNumber)>>,
366 : pub(crate) rel_size_snapshot_cache: Mutex<LruCache<(Lsn, RelTag), BlockNumber>>,
367 :
368 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
369 :
370 : state: watch::Sender<TimelineState>,
371 :
372 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
373 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
374 : pub delete_progress: TimelineDeleteProgress,
375 :
376 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
377 :
378 : /// Load or creation time information about the disk_consistent_lsn and when the loading
379 : /// happened. Used for consumption metrics.
380 : pub(crate) loaded_at: (Lsn, SystemTime),
381 :
382 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
383 : pub(crate) gate: Gate,
384 :
385 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
386 : /// to the timeline should drop out when this token fires.
387 : pub(crate) cancel: CancellationToken,
388 :
389 : /// Make sure we only have one running compaction at a time in tests.
390 : ///
391 : /// Must only be taken in two places:
392 : /// - [`Timeline::compact`] (this file)
393 : /// - [`delete::delete_local_timeline_directory`]
394 : ///
395 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
396 : compaction_lock: tokio::sync::Mutex<()>,
397 :
398 : /// If true, the last compaction failed.
399 : compaction_failed: AtomicBool,
400 :
401 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
402 : l0_compaction_trigger: Arc<Notify>,
403 :
404 : /// Make sure we only have one running gc at a time.
405 : ///
406 : /// Must only be taken in two places:
407 : /// - [`Timeline::gc`] (this file)
408 : /// - [`delete::delete_local_timeline_directory`]
409 : ///
410 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
411 : gc_lock: tokio::sync::Mutex<()>,
412 :
413 : /// Cloned from [`super::TenantShard::pagestream_throttle`] on construction.
414 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
415 :
416 : /// Size estimator for aux file v2
417 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
418 :
419 : /// Some test cases directly place keys into the timeline without actually modifying the directory
420 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
421 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
422 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
423 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
424 : #[cfg(test)]
425 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
426 :
427 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
428 :
429 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
430 :
431 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
432 :
433 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
434 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
435 :
436 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
437 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
438 :
439 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
440 :
441 : /// May host a background Tokio task which downloads all the layers from the current
442 : /// heatmap on demand.
443 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
444 :
445 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
446 :
447 : wait_lsn_log_slow: tokio::sync::Semaphore,
448 :
449 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
450 : basebackup_cache: Arc<BasebackupCache>,
451 :
452 : feature_resolver: Arc<TenantFeatureResolver>,
453 : }
454 :
455 : pub(crate) enum PreviousHeatmap {
456 : Active {
457 : heatmap: HeatMapTimeline,
458 : read_at: std::time::Instant,
459 : // End LSN covered by the heatmap if known
460 : end_lsn: Option<Lsn>,
461 : },
462 : Obsolete,
463 : }
464 :
465 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
466 :
467 : pub struct WalReceiverInfo {
468 : pub wal_source_connconf: PgConnectionConfig,
469 : pub last_received_msg_lsn: Lsn,
470 : pub last_received_msg_ts: u128,
471 : }
472 :
473 : /// Information about how much history needs to be retained, needed by
474 : /// Garbage Collection.
475 : #[derive(Default)]
476 : pub(crate) struct GcInfo {
477 : /// Specific LSNs that are needed.
478 : ///
479 : /// Currently, this includes all points where child branches have
480 : /// been forked off from. In the future, could also include
481 : /// explicit user-defined snapshot points.
482 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
483 :
484 : /// The cutoff coordinates, which are combined by selecting the minimum.
485 : pub(crate) cutoffs: GcCutoffs,
486 :
487 : /// Leases granted to particular LSNs.
488 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
489 :
490 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
491 : pub(crate) within_ancestor_pitr: bool,
492 : }
493 :
494 : impl GcInfo {
495 154 : pub(crate) fn min_cutoff(&self) -> Lsn {
496 154 : self.cutoffs.select_min()
497 154 : }
498 :
499 119 : pub(super) fn insert_child(
500 119 : &mut self,
501 119 : child_id: TimelineId,
502 119 : child_lsn: Lsn,
503 119 : is_offloaded: MaybeOffloaded,
504 119 : ) {
505 119 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
506 119 : self.retain_lsns.sort_by_key(|i| i.0);
507 119 : }
508 :
509 2 : pub(super) fn remove_child_maybe_offloaded(
510 2 : &mut self,
511 2 : child_id: TimelineId,
512 2 : maybe_offloaded: MaybeOffloaded,
513 2 : ) -> bool {
514 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
515 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
516 2 : let mut removed = false;
517 3 : self.retain_lsns.retain(|i| {
518 3 : if removed {
519 1 : return true;
520 2 : }
521 2 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
522 2 : removed |= remove;
523 2 : !remove
524 3 : });
525 2 : removed
526 2 : }
527 :
528 2 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
529 2 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
530 2 : }
531 :
532 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
533 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
534 0 : }
535 119 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
536 119 : self.leases.contains_key(&lsn)
537 119 : }
538 : }
539 :
540 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
541 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
542 : /// between time-based and space-based retention for observability and consumption metrics purposes.
543 : #[derive(Clone, Debug, Default)]
544 : pub(crate) struct GcCutoffs {
545 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
546 : /// history we must keep to retain a specified number of bytes of WAL.
547 : pub(crate) space: Lsn,
548 :
549 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates
550 : /// how much history we must keep to enable reading back at least the PITR interval duration.
551 : ///
552 : /// None indicates that the PITR cutoff has not been computed. A PITR interval of 0 will yield
553 : /// Some(last_record_lsn).
554 : pub(crate) time: Option<Lsn>,
555 : }
556 :
557 : impl GcCutoffs {
558 154 : fn select_min(&self) -> Lsn {
559 : // NB: if we haven't computed the PITR cutoff yet, we can't GC anything.
560 154 : self.space.min(self.time.unwrap_or_default())
561 154 : }
562 : }
563 :
564 : pub(crate) struct TimelineVisitOutcome {
565 : completed_keyspace: KeySpace,
566 : image_covered_keyspace: KeySpace,
567 : }
568 :
569 : /// An error happened in a get() operation.
570 : #[derive(thiserror::Error, Debug)]
571 : pub(crate) enum PageReconstructError {
572 : #[error(transparent)]
573 : Other(anyhow::Error),
574 :
575 : #[error("Ancestor LSN wait error: {0}")]
576 : AncestorLsnTimeout(WaitLsnError),
577 :
578 : #[error("timeline shutting down")]
579 : Cancelled,
580 :
581 : /// An error happened replaying WAL records
582 : #[error(transparent)]
583 : WalRedo(anyhow::Error),
584 :
585 : #[error("{0}")]
586 : MissingKey(Box<MissingKeyError>),
587 : }
588 :
589 : impl PageReconstructError {
590 0 : pub(crate) fn is_cancel(&self) -> bool {
591 0 : match self {
592 0 : PageReconstructError::Other(_) => false,
593 0 : PageReconstructError::AncestorLsnTimeout(e) => e.is_cancel(),
594 0 : PageReconstructError::Cancelled => true,
595 0 : PageReconstructError::WalRedo(_) => false,
596 0 : PageReconstructError::MissingKey(_) => false,
597 : }
598 0 : }
599 : #[allow(dead_code)] // we use the is_cancel + into_anyhow pattern in quite a few places, this one will follow soon enough
600 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
601 0 : match self {
602 0 : PageReconstructError::Other(e) => e,
603 0 : PageReconstructError::AncestorLsnTimeout(e) => e.into_anyhow(),
604 0 : PageReconstructError::Cancelled => anyhow::Error::new(self),
605 0 : PageReconstructError::WalRedo(e) => e,
606 0 : PageReconstructError::MissingKey(_) => anyhow::Error::new(self),
607 : }
608 0 : }
609 : }
610 :
611 : impl From<anyhow::Error> for PageReconstructError {
612 1 : fn from(value: anyhow::Error) -> Self {
613 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
614 1 : match value.downcast::<PageReconstructError>() {
615 0 : Ok(pre) => pre,
616 1 : Err(other) => PageReconstructError::Other(other),
617 : }
618 1 : }
619 : }
620 :
621 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
622 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
623 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
624 0 : }
625 : }
626 :
627 : impl From<layer_manager::Shutdown> for PageReconstructError {
628 0 : fn from(_: layer_manager::Shutdown) -> Self {
629 0 : PageReconstructError::Cancelled
630 0 : }
631 : }
632 :
633 : impl GetVectoredError {
634 : #[cfg(test)]
635 3 : pub(crate) fn is_missing_key_error(&self) -> bool {
636 3 : matches!(self, Self::MissingKey(_))
637 3 : }
638 : }
639 :
640 : impl From<layer_manager::Shutdown> for GetVectoredError {
641 0 : fn from(_: layer_manager::Shutdown) -> Self {
642 0 : GetVectoredError::Cancelled
643 0 : }
644 : }
645 :
646 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
647 : /// only and not used by the "real read path".
648 : pub enum ReadPathLayerId {
649 : PersistentLayer(PersistentLayerKey),
650 : InMemoryLayer(Range<Lsn>),
651 : }
652 :
653 : impl std::fmt::Display for ReadPathLayerId {
654 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
655 0 : match self {
656 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{key}"),
657 0 : ReadPathLayerId::InMemoryLayer(range) => {
658 0 : write!(f, "in-mem {}..{}", range.start, range.end)
659 : }
660 : }
661 0 : }
662 : }
663 : pub struct ReadPath {
664 : keyspace: KeySpace,
665 : lsn: Lsn,
666 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
667 : }
668 :
669 : impl ReadPath {
670 312395 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
671 312395 : Self {
672 312395 : keyspace,
673 312395 : lsn,
674 312395 : path: Vec::new(),
675 312395 : }
676 312395 : }
677 :
678 445201 : pub fn record_layer_visit(
679 445201 : &mut self,
680 445201 : layer_to_read: &ReadableLayer,
681 445201 : keyspace_to_read: &KeySpace,
682 445201 : lsn_range: &Range<Lsn>,
683 445201 : ) {
684 445201 : let id = match layer_to_read {
685 137918 : ReadableLayer::PersistentLayer(layer) => {
686 137918 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
687 : }
688 307283 : ReadableLayer::InMemoryLayer(layer) => {
689 307283 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
690 : }
691 : };
692 445201 : self.path
693 445201 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
694 445201 : }
695 : }
696 :
697 : impl std::fmt::Display for ReadPath {
698 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
699 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
700 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
701 0 : writeln!(
702 0 : f,
703 0 : "{}: {} {}..{} {}",
704 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
705 0 : )?;
706 : }
707 0 : Ok(())
708 0 : }
709 : }
710 :
711 : #[derive(thiserror::Error)]
712 : pub struct MissingKeyError {
713 : keyspace: KeySpace,
714 : shard: ShardNumber,
715 : query: Option<VersionedKeySpaceQuery>,
716 : // This is largest request LSN from the get page request batch
717 : original_hwm_lsn: Lsn,
718 : ancestor_lsn: Option<Lsn>,
719 : /// Debug information about the read path if there's an error
720 : read_path: Option<ReadPath>,
721 : backtrace: Option<std::backtrace::Backtrace>,
722 : }
723 :
724 : impl MissingKeyError {
725 7 : fn enrich(&mut self, query: VersionedKeySpaceQuery) {
726 7 : self.query = Some(query);
727 7 : }
728 : }
729 :
730 : impl std::fmt::Debug for MissingKeyError {
731 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
732 0 : write!(f, "{self}")
733 0 : }
734 : }
735 :
736 : impl std::fmt::Display for MissingKeyError {
737 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
738 0 : write!(
739 0 : f,
740 0 : "could not find data for key {} (shard {:?}), original HWM LSN {}",
741 : self.keyspace, self.shard, self.original_hwm_lsn
742 0 : )?;
743 :
744 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
745 0 : write!(f, ", ancestor {ancestor_lsn}")?;
746 0 : }
747 :
748 0 : if let Some(ref query) = self.query {
749 0 : write!(f, ", query {query}")?;
750 0 : }
751 :
752 0 : if let Some(ref read_path) = self.read_path {
753 0 : write!(f, "\n{read_path}")?;
754 0 : }
755 :
756 0 : if let Some(ref backtrace) = self.backtrace {
757 0 : write!(f, "\n{backtrace}")?;
758 0 : }
759 :
760 0 : Ok(())
761 0 : }
762 : }
763 :
764 : #[derive(thiserror::Error, Debug)]
765 : pub(crate) enum CreateImageLayersError {
766 : #[error("timeline shutting down")]
767 : Cancelled,
768 :
769 : #[error("read failed")]
770 : GetVectoredError(#[source] GetVectoredError),
771 :
772 : #[error("reconstruction failed")]
773 : PageReconstructError(#[source] PageReconstructError),
774 :
775 : #[error(transparent)]
776 : Other(anyhow::Error),
777 : }
778 :
779 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
780 0 : fn from(_: layer_manager::Shutdown) -> Self {
781 0 : CreateImageLayersError::Cancelled
782 0 : }
783 : }
784 :
785 : #[derive(thiserror::Error, Debug, Clone)]
786 : pub(crate) enum FlushLayerError {
787 : /// Timeline cancellation token was cancelled
788 : #[error("timeline shutting down")]
789 : Cancelled,
790 :
791 : /// We tried to flush a layer while the Timeline is in an unexpected state
792 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
793 : NotRunning(FlushLoopState),
794 :
795 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
796 : // loop via a watch channel, where we can only borrow it.
797 : #[error("create image layers (shared)")]
798 : CreateImageLayersError(Arc<CreateImageLayersError>),
799 :
800 : #[error("other (shared)")]
801 : Other(#[from] Arc<anyhow::Error>),
802 : }
803 :
804 : impl FlushLayerError {
805 : // When crossing from generic anyhow errors to this error type, we explicitly check
806 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
807 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
808 0 : let cancelled = timeline.cancel.is_cancelled()
809 : // The upload queue might have been shut down before the official cancellation of the timeline.
810 0 : || err
811 0 : .downcast_ref::<NotInitialized>()
812 0 : .map(NotInitialized::is_stopping)
813 0 : .unwrap_or_default();
814 0 : if cancelled {
815 0 : Self::Cancelled
816 : } else {
817 0 : Self::Other(Arc::new(err))
818 : }
819 0 : }
820 : }
821 :
822 : impl From<layer_manager::Shutdown> for FlushLayerError {
823 0 : fn from(_: layer_manager::Shutdown) -> Self {
824 0 : FlushLayerError::Cancelled
825 0 : }
826 : }
827 :
828 : #[derive(thiserror::Error, Debug)]
829 : pub enum GetVectoredError {
830 : #[error("timeline shutting down")]
831 : Cancelled,
832 :
833 : #[error("requested too many keys: {0} > {1}")]
834 : Oversized(u64, u64),
835 :
836 : #[error("requested at invalid LSN: {0}")]
837 : InvalidLsn(Lsn),
838 :
839 : #[error("requested key not found: {0}")]
840 : MissingKey(Box<MissingKeyError>),
841 :
842 : #[error("ancestry walk")]
843 : GetReadyAncestorError(#[source] GetReadyAncestorError),
844 :
845 : #[error(transparent)]
846 : Other(#[from] anyhow::Error),
847 : }
848 :
849 : impl From<GetReadyAncestorError> for GetVectoredError {
850 1 : fn from(value: GetReadyAncestorError) -> Self {
851 : use GetReadyAncestorError::*;
852 1 : match value {
853 0 : Cancelled => GetVectoredError::Cancelled,
854 : AncestorLsnTimeout(_) | BadState { .. } => {
855 1 : GetVectoredError::GetReadyAncestorError(value)
856 : }
857 : }
858 1 : }
859 : }
860 :
861 : #[derive(thiserror::Error, Debug)]
862 : pub enum GetReadyAncestorError {
863 : #[error("ancestor LSN wait error")]
864 : AncestorLsnTimeout(#[from] WaitLsnError),
865 :
866 : #[error("bad state on timeline {timeline_id}: {state:?}")]
867 : BadState {
868 : timeline_id: TimelineId,
869 : state: TimelineState,
870 : },
871 :
872 : #[error("cancelled")]
873 : Cancelled,
874 : }
875 :
876 : #[derive(Clone, Copy)]
877 : pub enum LogicalSizeCalculationCause {
878 : Initial,
879 : ConsumptionMetricsSyntheticSize,
880 : EvictionTaskImitation,
881 : TenantSizeHandler,
882 : }
883 :
884 : pub enum GetLogicalSizePriority {
885 : User,
886 : Background,
887 : }
888 :
889 : #[derive(Debug, enumset::EnumSetType)]
890 : pub(crate) enum CompactFlags {
891 : ForceRepartition,
892 : ForceImageLayerCreation,
893 : ForceL0Compaction,
894 : OnlyL0Compaction,
895 : EnhancedGcBottomMostCompaction,
896 : DryRun,
897 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
898 : /// the background compaction task, since we want to aggressively compact down L0 to bound
899 : /// read amplification.
900 : ///
901 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
902 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
903 : /// compaction).
904 : YieldForL0,
905 : }
906 :
907 : #[serde_with::serde_as]
908 0 : #[derive(Debug, Clone, serde::Deserialize)]
909 : pub(crate) struct CompactRequest {
910 : pub compact_key_range: Option<CompactKeyRange>,
911 : pub compact_lsn_range: Option<CompactLsnRange>,
912 : /// Whether the compaction job should be scheduled.
913 : #[serde(default)]
914 : pub scheduled: bool,
915 : /// Whether the compaction job should be split across key ranges.
916 : #[serde(default)]
917 : pub sub_compaction: bool,
918 : /// Max job size for each subcompaction job.
919 : pub sub_compaction_max_job_size_mb: Option<u64>,
920 : }
921 :
922 0 : #[derive(Debug, Clone, serde::Deserialize)]
923 : pub(crate) struct MarkInvisibleRequest {
924 : #[serde(default)]
925 : pub is_visible: Option<bool>,
926 : }
927 :
928 : #[derive(Debug, Clone, Default)]
929 : pub(crate) struct CompactOptions {
930 : pub flags: EnumSet<CompactFlags>,
931 : /// If set, the compaction will only compact the key range specified by this option.
932 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
933 : pub compact_key_range: Option<CompactKeyRange>,
934 : /// If set, the compaction will only compact the LSN within this value.
935 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
936 : pub compact_lsn_range: Option<CompactLsnRange>,
937 : /// Enable sub-compaction (split compaction job across key ranges).
938 : /// This option is only used by GC compaction.
939 : pub sub_compaction: bool,
940 : /// Set job size for the GC compaction.
941 : /// This option is only used by GC compaction.
942 : pub sub_compaction_max_job_size_mb: Option<u64>,
943 : }
944 :
945 : impl std::fmt::Debug for Timeline {
946 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
947 0 : write!(f, "Timeline<{}>", self.timeline_id)
948 0 : }
949 : }
950 :
951 : #[derive(thiserror::Error, Debug, Clone)]
952 : pub enum WaitLsnError {
953 : // Called on a timeline which is shutting down
954 : #[error("Shutdown")]
955 : Shutdown,
956 :
957 : // Called on an timeline not in active state or shutting down
958 : #[error("Bad timeline state: {0:?}")]
959 : BadState(TimelineState),
960 :
961 : // Timeout expired while waiting for LSN to catch up with goal.
962 : #[error("{0}")]
963 : Timeout(String),
964 : }
965 :
966 : impl WaitLsnError {
967 0 : pub(crate) fn is_cancel(&self) -> bool {
968 0 : match self {
969 0 : WaitLsnError::Shutdown => true,
970 0 : WaitLsnError::BadState(timeline_state) => match timeline_state {
971 0 : TimelineState::Loading => false,
972 0 : TimelineState::Active => false,
973 0 : TimelineState::Stopping => true,
974 0 : TimelineState::Broken { .. } => false,
975 : },
976 0 : WaitLsnError::Timeout(_) => false,
977 : }
978 0 : }
979 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
980 0 : match self {
981 0 : WaitLsnError::Shutdown => anyhow::Error::new(self),
982 0 : WaitLsnError::BadState(_) => anyhow::Error::new(self),
983 0 : WaitLsnError::Timeout(_) => anyhow::Error::new(self),
984 : }
985 0 : }
986 : }
987 :
988 : impl From<WaitLsnError> for tonic::Status {
989 0 : fn from(err: WaitLsnError) -> Self {
990 : use tonic::Code;
991 0 : let code = if err.is_cancel() {
992 0 : Code::Unavailable
993 : } else {
994 0 : Code::Internal
995 : };
996 0 : tonic::Status::new(code, err.to_string())
997 0 : }
998 : }
999 :
1000 : // The impls below achieve cancellation mapping for errors.
1001 : // Perhaps there's a way of achieving this with less cruft.
1002 :
1003 : impl From<CreateImageLayersError> for CompactionError {
1004 0 : fn from(e: CreateImageLayersError) -> Self {
1005 0 : match e {
1006 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
1007 0 : CreateImageLayersError::Other(e) => {
1008 0 : CompactionError::Other(e.context("create image layers"))
1009 : }
1010 0 : _ => CompactionError::Other(e.into()),
1011 : }
1012 0 : }
1013 : }
1014 :
1015 : impl From<CreateImageLayersError> for FlushLayerError {
1016 0 : fn from(e: CreateImageLayersError) -> Self {
1017 0 : match e {
1018 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
1019 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
1020 : }
1021 0 : }
1022 : }
1023 :
1024 : impl From<PageReconstructError> for CreateImageLayersError {
1025 0 : fn from(e: PageReconstructError) -> Self {
1026 0 : match e {
1027 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
1028 0 : _ => CreateImageLayersError::PageReconstructError(e),
1029 : }
1030 0 : }
1031 : }
1032 :
1033 : impl From<super::storage_layer::errors::PutError> for CreateImageLayersError {
1034 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
1035 0 : if e.is_cancel() {
1036 0 : CreateImageLayersError::Cancelled
1037 : } else {
1038 0 : CreateImageLayersError::Other(e.into_anyhow())
1039 : }
1040 0 : }
1041 : }
1042 :
1043 : impl From<GetVectoredError> for CreateImageLayersError {
1044 0 : fn from(e: GetVectoredError) -> Self {
1045 0 : match e {
1046 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
1047 0 : _ => CreateImageLayersError::GetVectoredError(e),
1048 : }
1049 0 : }
1050 : }
1051 :
1052 : impl From<GetVectoredError> for PageReconstructError {
1053 3 : fn from(e: GetVectoredError) -> Self {
1054 3 : match e {
1055 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1056 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1057 0 : err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
1058 2 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1059 1 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1060 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1061 : }
1062 3 : }
1063 : }
1064 :
1065 : impl From<GetReadyAncestorError> for PageReconstructError {
1066 1 : fn from(e: GetReadyAncestorError) -> Self {
1067 : use GetReadyAncestorError::*;
1068 1 : match e {
1069 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1070 1 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1071 0 : Cancelled => PageReconstructError::Cancelled,
1072 : }
1073 1 : }
1074 : }
1075 :
1076 : pub(crate) enum WaitLsnTimeout {
1077 : Custom(Duration),
1078 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1079 : Default,
1080 : }
1081 :
1082 : pub(crate) enum WaitLsnWaiter<'a> {
1083 : Timeline(&'a Timeline),
1084 : Tenant,
1085 : PageService,
1086 : HttpEndpoint,
1087 : BaseBackupCache,
1088 : }
1089 :
1090 : /// Argument to [`Timeline::shutdown`].
1091 : #[derive(Debug, Clone, Copy)]
1092 : pub(crate) enum ShutdownMode {
1093 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
1094 : /// take multiple seconds for a busy timeline.
1095 : ///
1096 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1097 : /// the call to [`Timeline::shutdown`].
1098 : FreezeAndFlush,
1099 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1100 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1101 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1102 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1103 : Reload,
1104 : /// Shut down immediately, without waiting for any open layers to flush.
1105 : Hard,
1106 : }
1107 :
1108 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1109 : enum ImageLayerCreationOutcome {
1110 : /// We generated an image layer
1111 : Generated {
1112 : unfinished_image_layer: ImageLayerWriter,
1113 : },
1114 : /// The key range is empty
1115 : Empty,
1116 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1117 : /// the image layer creation.
1118 : Skip,
1119 : }
1120 :
1121 : enum RepartitionError {
1122 : Other(anyhow::Error),
1123 : CollectKeyspace(CollectKeySpaceError),
1124 : }
1125 :
1126 : impl RepartitionError {
1127 0 : fn is_cancel(&self) -> bool {
1128 0 : match self {
1129 0 : RepartitionError::Other(_) => false,
1130 0 : RepartitionError::CollectKeyspace(e) => e.is_cancel(),
1131 : }
1132 0 : }
1133 0 : fn into_anyhow(self) -> anyhow::Error {
1134 0 : match self {
1135 0 : RepartitionError::Other(e) => e,
1136 0 : RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
1137 : }
1138 0 : }
1139 : }
1140 :
1141 : /// Public interface functions
1142 : impl Timeline {
1143 : /// Get the LSN where this branch was created
1144 23 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1145 23 : self.ancestor_lsn
1146 23 : }
1147 :
1148 : /// Get the ancestor's timeline id
1149 416 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1150 416 : self.ancestor_timeline
1151 416 : .as_ref()
1152 416 : .map(|ancestor| ancestor.timeline_id)
1153 416 : }
1154 :
1155 : /// Get the ancestor timeline
1156 1 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1157 1 : self.ancestor_timeline.as_ref()
1158 1 : }
1159 :
1160 : /// Get the bytes written since the PITR cutoff on this branch, and
1161 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1162 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1163 : // TODO: for backwards compatibility, we return the full history back to 0 when the PITR
1164 : // cutoff has not yet been initialized. This should return None instead, but this is exposed
1165 : // in external HTTP APIs and callers may not handle a null value.
1166 0 : let gc_info = self.gc_info.read().unwrap();
1167 0 : let history = self
1168 0 : .get_last_record_lsn()
1169 0 : .checked_sub(gc_info.cutoffs.time.unwrap_or_default())
1170 0 : .unwrap_or_default()
1171 0 : .0;
1172 0 : (history, gc_info.within_ancestor_pitr)
1173 0 : }
1174 :
1175 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1176 426738 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1177 426738 : self.applied_gc_cutoff_lsn.read()
1178 426738 : }
1179 :
1180 : /// Read timeline's planned GC cutoff: this is the logical end of history that users are allowed
1181 : /// to read (based on configured PITR), even if physically we have more history. Returns None
1182 : /// if the PITR cutoff has not yet been initialized.
1183 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Option<Lsn> {
1184 0 : self.gc_info.read().unwrap().cutoffs.time
1185 0 : }
1186 :
1187 : /// Look up given page version.
1188 : ///
1189 : /// If a remote layer file is needed, it is downloaded as part of this
1190 : /// call.
1191 : ///
1192 : /// This method enforces [`Self::pagestream_throttle`] internally.
1193 : ///
1194 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1195 : /// abstraction above this needs to store suitable metadata to track what
1196 : /// data exists with what keys, in separate metadata entries. If a
1197 : /// non-existent key is requested, we may incorrectly return a value from
1198 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1199 : /// non-existing key.
1200 : ///
1201 : /// # Cancel-Safety
1202 : ///
1203 : /// This method is cancellation-safe.
1204 : #[inline(always)]
1205 301314 : pub(crate) async fn get(
1206 301314 : &self,
1207 301314 : key: Key,
1208 301314 : lsn: Lsn,
1209 301314 : ctx: &RequestContext,
1210 301314 : ) -> Result<Bytes, PageReconstructError> {
1211 301314 : if !lsn.is_valid() {
1212 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1213 301314 : }
1214 :
1215 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1216 : // already checked the key against the shard_identity when looking up the Timeline from
1217 : // page_service.
1218 301314 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1219 :
1220 301314 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1221 :
1222 301314 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1223 :
1224 301314 : let vectored_res = self
1225 301314 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1226 301314 : .await;
1227 :
1228 301314 : let key_value = vectored_res?.pop_first();
1229 301311 : match key_value {
1230 301305 : Some((got_key, value)) => {
1231 301305 : if got_key != key {
1232 0 : error!(
1233 0 : "Expected {}, but singular vectored get returned {}",
1234 : key, got_key
1235 : );
1236 0 : Err(PageReconstructError::Other(anyhow!(
1237 0 : "Singular vectored get returned wrong key"
1238 0 : )))
1239 : } else {
1240 301305 : value
1241 : }
1242 : }
1243 6 : None => Err(PageReconstructError::MissingKey(Box::new(
1244 6 : MissingKeyError {
1245 6 : keyspace: KeySpace::single(key..key.next()),
1246 6 : shard: self.shard_identity.get_shard_number(&key),
1247 6 : original_hwm_lsn: lsn,
1248 6 : ancestor_lsn: None,
1249 6 : backtrace: None,
1250 6 : read_path: None,
1251 6 : query: None,
1252 6 : },
1253 6 : ))),
1254 : }
1255 301314 : }
1256 :
1257 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1258 :
1259 : /// Look up multiple page versions at a given LSN
1260 : ///
1261 : /// This naive implementation will be replaced with a more efficient one
1262 : /// which actually vectorizes the read path.
1263 10888 : pub(crate) async fn get_vectored(
1264 10888 : &self,
1265 10888 : query: VersionedKeySpaceQuery,
1266 10888 : io_concurrency: super::storage_layer::IoConcurrency,
1267 10888 : ctx: &RequestContext,
1268 10888 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1269 10888 : let total_keyspace = query.total_keyspace();
1270 :
1271 10888 : let key_count = total_keyspace.total_raw_size();
1272 10888 : if key_count > self.conf.max_get_vectored_keys.get() {
1273 0 : return Err(GetVectoredError::Oversized(
1274 0 : key_count as u64,
1275 0 : self.conf.max_get_vectored_keys.get() as u64,
1276 0 : ));
1277 10888 : }
1278 :
1279 34195 : for range in &total_keyspace.ranges {
1280 23307 : let mut key = range.start;
1281 65329 : while key != range.end {
1282 42022 : assert!(!self.shard_identity.is_key_disposable(&key));
1283 42022 : key = key.next();
1284 : }
1285 : }
1286 :
1287 10888 : trace!(
1288 0 : "get vectored query {} from task kind {:?}",
1289 : query,
1290 0 : ctx.task_kind(),
1291 : );
1292 :
1293 10888 : let start = crate::metrics::GET_VECTORED_LATENCY
1294 10888 : .for_task_kind(ctx.task_kind())
1295 10888 : .map(|metric| (metric, Instant::now()));
1296 :
1297 10888 : let res = self
1298 10888 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1299 10888 : .await;
1300 :
1301 10888 : if let Some((metric, start)) = start {
1302 0 : let elapsed = start.elapsed();
1303 0 : metric.observe(elapsed.as_secs_f64());
1304 10888 : }
1305 :
1306 10888 : res
1307 10888 : }
1308 :
1309 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1310 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1311 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1312 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1313 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1314 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1315 : /// the scan operation will not cause OOM in the future.
1316 8 : pub(crate) async fn scan(
1317 8 : &self,
1318 8 : keyspace: KeySpace,
1319 8 : lsn: Lsn,
1320 8 : ctx: &RequestContext,
1321 8 : io_concurrency: super::storage_layer::IoConcurrency,
1322 8 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1323 8 : if !lsn.is_valid() {
1324 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1325 8 : }
1326 :
1327 8 : trace!(
1328 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1329 : keyspace,
1330 : lsn,
1331 0 : ctx.task_kind()
1332 : );
1333 :
1334 : // We should generalize this into Keyspace::contains in the future.
1335 16 : for range in &keyspace.ranges {
1336 8 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1337 8 : || range.end.field1 > METADATA_KEY_END_PREFIX
1338 : {
1339 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1340 0 : "only metadata keyspace can be scanned"
1341 0 : )));
1342 8 : }
1343 : }
1344 :
1345 8 : let start = crate::metrics::SCAN_LATENCY
1346 8 : .for_task_kind(ctx.task_kind())
1347 8 : .map(ScanLatencyOngoingRecording::start_recording);
1348 :
1349 8 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1350 :
1351 8 : let vectored_res = self
1352 8 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1353 8 : .await;
1354 :
1355 8 : if let Some(recording) = start {
1356 0 : recording.observe();
1357 8 : }
1358 :
1359 8 : vectored_res
1360 8 : }
1361 :
1362 312395 : pub(super) async fn get_vectored_impl(
1363 312395 : &self,
1364 312395 : query: VersionedKeySpaceQuery,
1365 312395 : reconstruct_state: &mut ValuesReconstructState,
1366 312395 : ctx: &RequestContext,
1367 312395 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1368 312395 : if query.is_empty() {
1369 0 : return Ok(BTreeMap::default());
1370 312395 : }
1371 :
1372 312395 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1373 312395 : Some(ReadPath::new(
1374 312395 : query.total_keyspace(),
1375 312395 : query.high_watermark_lsn()?,
1376 : ))
1377 : } else {
1378 0 : None
1379 : };
1380 :
1381 312395 : reconstruct_state.read_path = read_path;
1382 :
1383 312395 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1384 0 : RedoAttemptType::LegacyCompaction
1385 : } else {
1386 312395 : RedoAttemptType::ReadPage
1387 : };
1388 :
1389 312395 : let traversal_res: Result<(), _> = {
1390 312395 : let ctx = RequestContextBuilder::from(ctx)
1391 312395 : .perf_span(|crnt_perf_span| {
1392 0 : info_span!(
1393 : target: PERF_TRACE_TARGET,
1394 0 : parent: crnt_perf_span,
1395 : "PLAN_IO",
1396 : )
1397 0 : })
1398 312395 : .attached_child();
1399 :
1400 312395 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1401 312395 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1402 312395 : .await
1403 : };
1404 :
1405 312395 : if let Err(err) = traversal_res {
1406 : // Wait for all the spawned IOs to complete.
1407 : // See comments on `spawn_io` inside `storage_layer` for more details.
1408 8 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1409 8 : .into_values()
1410 8 : .map(|state| state.collect_pending_ios())
1411 8 : .collect::<FuturesUnordered<_>>();
1412 8 : while collect_futs.next().await.is_some() {}
1413 :
1414 : // Enrich the missing key error with the original query.
1415 8 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1416 7 : missing_err.enrich(query.clone());
1417 7 : return Err(GetVectoredError::MissingKey(missing_err));
1418 1 : }
1419 :
1420 1 : return Err(err);
1421 312387 : };
1422 :
1423 312387 : let layers_visited = reconstruct_state.get_layers_visited();
1424 :
1425 312387 : let ctx = RequestContextBuilder::from(ctx)
1426 312387 : .perf_span(|crnt_perf_span| {
1427 0 : info_span!(
1428 : target: PERF_TRACE_TARGET,
1429 0 : parent: crnt_perf_span,
1430 : "RECONSTRUCT",
1431 : )
1432 0 : })
1433 312387 : .attached_child();
1434 :
1435 312387 : let futs = FuturesUnordered::new();
1436 363566 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1437 363566 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1438 :
1439 363566 : futs.push({
1440 363566 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1441 363566 : let ctx = RequestContextBuilder::from(&ctx)
1442 363566 : .perf_span(|crnt_perf_span| {
1443 0 : info_span!(
1444 : target: PERF_TRACE_TARGET,
1445 0 : parent: crnt_perf_span,
1446 : "RECONSTRUCT_KEY",
1447 : key = %key,
1448 : )
1449 0 : })
1450 363566 : .attached_child();
1451 :
1452 363566 : async move {
1453 363566 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1454 :
1455 363566 : let res = state
1456 363566 : .collect_pending_ios()
1457 363566 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1458 0 : info_span!(
1459 : target: PERF_TRACE_TARGET,
1460 0 : parent: crnt_perf_span,
1461 : "WAIT_FOR_IO_COMPLETIONS",
1462 : )
1463 0 : })
1464 363566 : .await;
1465 :
1466 363566 : let converted = match res {
1467 363566 : Ok(ok) => ok,
1468 0 : Err(err) => {
1469 0 : return (key, Err(err));
1470 : }
1471 : };
1472 363566 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1473 :
1474 : // The walredo module expects the records to be descending in terms of Lsn.
1475 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1476 363566 : debug_assert!(
1477 363566 : converted
1478 363566 : .records
1479 1403254 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1480 0 : "{converted:?}"
1481 : );
1482 :
1483 363566 : let walredo_deltas = converted.num_deltas();
1484 363566 : let walredo_res = walredo_self
1485 363566 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1486 363566 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1487 0 : info_span!(
1488 : target: PERF_TRACE_TARGET,
1489 0 : parent: crnt_perf_span,
1490 : "WALREDO",
1491 : deltas = %walredo_deltas,
1492 : )
1493 0 : })
1494 363566 : .await;
1495 :
1496 363566 : (key, walredo_res)
1497 363566 : }
1498 : });
1499 : }
1500 :
1501 312387 : let results = futs
1502 312387 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1503 312387 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1504 312387 : .await;
1505 :
1506 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1507 : // when they're missing. Instead they are omitted from the resulting btree
1508 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1509 : // to avoid infinite results.
1510 312387 : if !results.is_empty() {
1511 312259 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1512 0 : let total_keyspace = query.total_keyspace();
1513 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1514 :
1515 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1516 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1517 0 : LOG_PACER.lock().unwrap().call(|| {
1518 0 : let num_keys = total_keyspace.total_raw_size();
1519 0 : let num_pages = results.len();
1520 0 : tracing::info!(
1521 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1522 : lsn = %max_request_lsn,
1523 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1524 : );
1525 0 : });
1526 312259 : }
1527 :
1528 : // Records the number of layers visited in a few different ways:
1529 : //
1530 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1531 : // layer directly affects its observed latency.
1532 : //
1533 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1534 : // layer visits and access cost.
1535 : //
1536 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1537 : // read amplification after batching.
1538 312259 : let layers_visited = layers_visited as f64;
1539 312259 : let avg_layers_visited = layers_visited / results.len() as f64;
1540 312259 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1541 675825 : for _ in &results {
1542 363566 : self.metrics.layers_per_read.observe(layers_visited);
1543 363566 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1544 363566 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1545 363566 : }
1546 128 : }
1547 :
1548 312387 : Ok(results)
1549 312395 : }
1550 :
1551 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1552 138197 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1553 138197 : self.last_record_lsn.load().last
1554 138197 : }
1555 :
1556 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1557 0 : self.last_record_lsn.load().prev
1558 0 : }
1559 :
1560 : /// Atomically get both last and prev.
1561 117 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1562 117 : self.last_record_lsn.load()
1563 117 : }
1564 :
1565 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1566 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1567 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1568 0 : self.last_record_lsn.status_receiver()
1569 0 : }
1570 :
1571 609 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1572 609 : self.disk_consistent_lsn.load()
1573 609 : }
1574 :
1575 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1576 : /// not validated with control plane yet.
1577 : /// See [`Self::get_remote_consistent_lsn_visible`].
1578 2 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1579 2 : self.remote_client.remote_consistent_lsn_projected()
1580 2 : }
1581 :
1582 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1583 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1584 : /// generation validation in the deletion queue.
1585 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1586 0 : self.remote_client.remote_consistent_lsn_visible()
1587 0 : }
1588 :
1589 : /// The sum of the file size of all historic layers in the layer map.
1590 : /// This method makes no distinction between local and remote layers.
1591 : /// Hence, the result **does not represent local filesystem usage**.
1592 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1593 0 : let guard = self
1594 0 : .layers
1595 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
1596 0 : .await;
1597 0 : guard.layer_size_sum()
1598 0 : }
1599 :
1600 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1601 0 : self.metrics.resident_physical_size_get()
1602 0 : }
1603 :
1604 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1605 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1606 0 : }
1607 :
1608 : ///
1609 : /// Wait until WAL has been received and processed up to this LSN.
1610 : ///
1611 : /// You should call this before any of the other get_* or list_* functions. Calling
1612 : /// those functions with an LSN that has been processed yet is an error.
1613 : ///
1614 113805 : pub(crate) async fn wait_lsn(
1615 113805 : &self,
1616 113805 : lsn: Lsn,
1617 113805 : who_is_waiting: WaitLsnWaiter<'_>,
1618 113805 : timeout: WaitLsnTimeout,
1619 113805 : ctx: &RequestContext, /* Prepare for use by cancellation */
1620 113805 : ) -> Result<(), WaitLsnError> {
1621 113805 : let state = self.current_state();
1622 113805 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1623 0 : return Err(WaitLsnError::Shutdown);
1624 113805 : } else if !matches!(state, TimelineState::Active) {
1625 0 : return Err(WaitLsnError::BadState(state));
1626 113805 : }
1627 :
1628 113805 : if cfg!(debug_assertions) {
1629 113805 : match ctx.task_kind() {
1630 : TaskKind::WalReceiverManager
1631 : | TaskKind::WalReceiverConnectionHandler
1632 : | TaskKind::WalReceiverConnectionPoller => {
1633 0 : let is_myself = match who_is_waiting {
1634 0 : WaitLsnWaiter::Timeline(waiter) => {
1635 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1636 : }
1637 : WaitLsnWaiter::Tenant
1638 : | WaitLsnWaiter::PageService
1639 : | WaitLsnWaiter::HttpEndpoint
1640 0 : | WaitLsnWaiter::BaseBackupCache => unreachable!(
1641 : "tenant or page_service context are not expected to have task kind {:?}",
1642 0 : ctx.task_kind()
1643 : ),
1644 : };
1645 0 : if is_myself {
1646 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1647 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1648 0 : panic!(
1649 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1650 : );
1651 0 : }
1652 0 : } else {
1653 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1654 0 : // our walreceiver task can make progress independent of theirs
1655 0 : }
1656 : }
1657 113805 : _ => {}
1658 : }
1659 0 : }
1660 :
1661 113805 : let timeout = match timeout {
1662 0 : WaitLsnTimeout::Custom(t) => t,
1663 113805 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1664 : };
1665 :
1666 113805 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1667 113805 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1668 :
1669 113805 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1670 113805 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1671 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1672 113805 : let log_slow_threshold = Duration::from_secs(1);
1673 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1674 113805 : let log_slow_period = Duration::from_secs(10);
1675 113805 : let mut logging_permit = None;
1676 113805 : let wait_for_timeout = monitor_slow_future(
1677 113805 : log_slow_threshold,
1678 113805 : log_slow_period,
1679 113805 : wait_for_timeout,
1680 : |MonitorSlowFutureCallback {
1681 : ready,
1682 : is_slow,
1683 : elapsed_total,
1684 : elapsed_since_last_callback,
1685 113805 : }| {
1686 113805 : self.metrics
1687 113805 : .wait_lsn_in_progress_micros
1688 113805 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1689 113805 : if !is_slow {
1690 113805 : return;
1691 0 : }
1692 : // It's slow, see if we should log it.
1693 : // (We limit the logging to one per invocation per timeline to avoid excessive
1694 : // logging during an extended broker / networking outage that affects all timelines.)
1695 0 : if logging_permit.is_none() {
1696 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1697 0 : }
1698 0 : if logging_permit.is_none() {
1699 0 : return;
1700 0 : }
1701 : // We log it.
1702 0 : if ready {
1703 0 : info!(
1704 0 : "slow wait_lsn completed after {:.3}s",
1705 0 : elapsed_total.as_secs_f64()
1706 : );
1707 : } else {
1708 0 : info!(
1709 0 : "slow wait_lsn still running for {:.3}s",
1710 0 : elapsed_total.as_secs_f64()
1711 : );
1712 : }
1713 113805 : },
1714 : );
1715 113805 : let res = wait_for_timeout.await;
1716 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1717 113805 : drop(logging_permit);
1718 113805 : drop(start_finish_counterpair_guard);
1719 113805 : drop(timer);
1720 113805 : match res {
1721 113805 : Ok(()) => Ok(()),
1722 0 : Err(e) => {
1723 : use utils::seqwait::SeqWaitError::*;
1724 0 : match e {
1725 0 : Shutdown => Err(WaitLsnError::Shutdown),
1726 : Timeout => {
1727 0 : let walreceiver_status = self.walreceiver_status();
1728 0 : Err(WaitLsnError::Timeout(format!(
1729 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1730 0 : lsn,
1731 0 : self.get_last_record_lsn(),
1732 0 : self.get_disk_consistent_lsn(),
1733 0 : walreceiver_status,
1734 0 : )))
1735 : }
1736 : }
1737 : }
1738 : }
1739 113805 : }
1740 :
1741 0 : pub(crate) fn walreceiver_status(&self) -> String {
1742 0 : match &*self.walreceiver.lock().unwrap() {
1743 0 : None => "stopping or stopped".to_string(),
1744 0 : Some(walreceiver) => match walreceiver.status() {
1745 0 : Some(status) => status.to_human_readable_string(),
1746 0 : None => "Not active".to_string(),
1747 : },
1748 : }
1749 0 : }
1750 :
1751 : /// Check that it is valid to request operations with that lsn.
1752 119 : pub(crate) fn check_lsn_is_in_scope(
1753 119 : &self,
1754 119 : lsn: Lsn,
1755 119 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1756 119 : ) -> anyhow::Result<()> {
1757 119 : ensure!(
1758 119 : lsn >= **latest_gc_cutoff_lsn,
1759 2 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1760 : lsn,
1761 2 : **latest_gc_cutoff_lsn,
1762 : );
1763 117 : Ok(())
1764 119 : }
1765 :
1766 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1767 5 : pub(crate) fn init_lsn_lease(
1768 5 : &self,
1769 5 : lsn: Lsn,
1770 5 : length: Duration,
1771 5 : ctx: &RequestContext,
1772 5 : ) -> anyhow::Result<LsnLease> {
1773 5 : self.make_lsn_lease(lsn, length, true, ctx)
1774 5 : }
1775 :
1776 : /// 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.
1777 2 : pub(crate) fn renew_lsn_lease(
1778 2 : &self,
1779 2 : lsn: Lsn,
1780 2 : length: Duration,
1781 2 : ctx: &RequestContext,
1782 2 : ) -> anyhow::Result<LsnLease> {
1783 2 : self.make_lsn_lease(lsn, length, false, ctx)
1784 2 : }
1785 :
1786 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1787 : ///
1788 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1789 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1790 : ///
1791 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1792 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1793 7 : fn make_lsn_lease(
1794 7 : &self,
1795 7 : lsn: Lsn,
1796 7 : length: Duration,
1797 7 : init: bool,
1798 7 : _ctx: &RequestContext,
1799 7 : ) -> anyhow::Result<LsnLease> {
1800 6 : let lease = {
1801 : // Normalize the requested LSN to be aligned, and move to the first record
1802 : // if it points to the beginning of the page (header).
1803 7 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1804 :
1805 7 : let mut gc_info = self.gc_info.write().unwrap();
1806 7 : let planned_cutoff = gc_info.min_cutoff();
1807 :
1808 7 : let valid_until = SystemTime::now() + length;
1809 :
1810 7 : let entry = gc_info.leases.entry(lsn);
1811 :
1812 7 : match entry {
1813 3 : Entry::Occupied(mut occupied) => {
1814 3 : let existing_lease = occupied.get_mut();
1815 3 : if valid_until > existing_lease.valid_until {
1816 1 : existing_lease.valid_until = valid_until;
1817 1 : let dt: DateTime<Utc> = valid_until.into();
1818 1 : info!("lease extended to {}", dt);
1819 : } else {
1820 2 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1821 2 : info!("existing lease covers greater length, valid until {}", dt);
1822 : }
1823 :
1824 3 : existing_lease.clone()
1825 : }
1826 4 : Entry::Vacant(vacant) => {
1827 : // Reject already GC-ed LSN if we are in AttachedSingle and
1828 : // not blocked by the lsn lease deadline.
1829 4 : let validate = {
1830 4 : let conf = self.tenant_conf.load();
1831 4 : conf.location.attach_mode == AttachmentMode::Single
1832 4 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1833 : };
1834 :
1835 4 : if init || validate {
1836 4 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1837 4 : if lsn < *latest_gc_cutoff_lsn {
1838 1 : bail!(
1839 1 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1840 : lsn,
1841 1 : *latest_gc_cutoff_lsn
1842 : );
1843 3 : }
1844 3 : if lsn < planned_cutoff {
1845 0 : bail!(
1846 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1847 : lsn,
1848 : planned_cutoff
1849 : );
1850 3 : }
1851 0 : }
1852 :
1853 3 : let dt: DateTime<Utc> = valid_until.into();
1854 3 : info!("lease created, valid until {}", dt);
1855 3 : vacant.insert(LsnLease { valid_until }).clone()
1856 : }
1857 : }
1858 : };
1859 :
1860 6 : Ok(lease)
1861 7 : }
1862 :
1863 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1864 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1865 : #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
1866 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1867 : self.freeze0().await
1868 : }
1869 :
1870 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1871 : #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
1872 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1873 : self.freeze_and_flush0().await
1874 : }
1875 :
1876 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1877 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1878 570 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1879 570 : let mut g = self.write_lock.lock().await;
1880 570 : let to_lsn = self.get_last_record_lsn();
1881 570 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1882 570 : }
1883 :
1884 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1885 : // polluting the span hierarchy.
1886 570 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1887 570 : let token = self.freeze0().await?;
1888 570 : self.wait_flush_completion(token).await
1889 570 : }
1890 :
1891 : // Check if an open ephemeral layer should be closed: this provides
1892 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1893 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1894 : // ephemeral layer bytes has been breached.
1895 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1896 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1897 : // If the write lock is held, there is an active wal receiver: rolling open layers
1898 : // is their responsibility while they hold this lock.
1899 0 : return;
1900 : };
1901 :
1902 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1903 : // time, and this was missed.
1904 : // if write_guard.is_none() { return; }
1905 :
1906 0 : let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
1907 : // Don't block if the layer lock is busy
1908 0 : return;
1909 : };
1910 :
1911 0 : let Ok(lm) = layers_guard.layer_map() else {
1912 0 : return;
1913 : };
1914 :
1915 0 : let Some(open_layer) = &lm.open_layer else {
1916 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1917 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1918 : // that didn't result in writes to this shard.
1919 :
1920 : // Must not hold the layers lock while waiting for a flush.
1921 0 : drop(layers_guard);
1922 :
1923 0 : let last_record_lsn = self.get_last_record_lsn();
1924 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1925 0 : if last_record_lsn > disk_consistent_lsn {
1926 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1927 : // we are a sharded tenant and have skipped some WAL
1928 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1929 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1930 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1931 : // without any data ingested (yet) doesn't write a remote index as soon as it
1932 : // sees its LSN advance: we only do this if we've been layer-less
1933 : // for some time.
1934 0 : tracing::debug!(
1935 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1936 : disk_consistent_lsn,
1937 : last_record_lsn
1938 : );
1939 :
1940 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1941 : // We know there is no open layer, so we can request freezing without actually
1942 : // freezing anything. This is true even if we have dropped the layers_guard, we
1943 : // still hold the write_guard.
1944 0 : let _ = async {
1945 0 : let token = self
1946 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1947 0 : .await?;
1948 0 : self.wait_flush_completion(token).await
1949 0 : }
1950 0 : .await;
1951 0 : }
1952 0 : }
1953 :
1954 0 : return;
1955 : };
1956 :
1957 0 : let current_size = open_layer.len();
1958 :
1959 0 : let current_lsn = self.get_last_record_lsn();
1960 :
1961 0 : let checkpoint_distance_override = open_layer.tick();
1962 :
1963 0 : if let Some(size_override) = checkpoint_distance_override {
1964 0 : if current_size > size_override {
1965 : // This is not harmful, but it only happens in relatively rare cases where
1966 : // time-based checkpoints are not happening fast enough to keep the amount of
1967 : // ephemeral data within configured limits. It's a sign of stress on the system.
1968 0 : tracing::info!(
1969 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
1970 : );
1971 0 : }
1972 0 : }
1973 :
1974 0 : let checkpoint_distance =
1975 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1976 :
1977 0 : if self.should_roll(
1978 0 : current_size,
1979 0 : current_size,
1980 0 : checkpoint_distance,
1981 0 : self.get_last_record_lsn(),
1982 0 : self.last_freeze_at.load(),
1983 0 : open_layer.get_opened_at(),
1984 : ) {
1985 0 : match open_layer.info() {
1986 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1987 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1988 : // happens asynchronously in the background.
1989 0 : tracing::debug!(
1990 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1991 : );
1992 : }
1993 : InMemoryLayerInfo::Open { .. } => {
1994 : // Upgrade to a write lock and freeze the layer
1995 0 : drop(layers_guard);
1996 0 : let res = self
1997 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1998 0 : .await;
1999 :
2000 0 : if let Err(e) = res {
2001 0 : tracing::info!(
2002 0 : "failed to flush frozen layer after background freeze: {e:#}"
2003 : );
2004 0 : }
2005 : }
2006 : }
2007 0 : }
2008 0 : }
2009 :
2010 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
2011 : ///
2012 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
2013 : /// child timelines that are not offloaded yet.
2014 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
2015 0 : if self.remote_client.is_archived() != Some(true) {
2016 0 : return (false, "the timeline is not archived");
2017 0 : }
2018 0 : if !self.remote_client.no_pending_work() {
2019 : // if the remote client is still processing some work, we can't offload
2020 0 : return (false, "the upload queue is not drained yet");
2021 0 : }
2022 :
2023 0 : (true, "ok")
2024 0 : }
2025 :
2026 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
2027 : /// compaction tasks.
2028 192 : pub(crate) async fn compact(
2029 192 : self: &Arc<Self>,
2030 192 : cancel: &CancellationToken,
2031 192 : flags: EnumSet<CompactFlags>,
2032 192 : ctx: &RequestContext,
2033 192 : ) -> Result<CompactionOutcome, CompactionError> {
2034 192 : let res = self
2035 192 : .compact_with_options(
2036 192 : cancel,
2037 192 : CompactOptions {
2038 192 : flags,
2039 192 : compact_key_range: None,
2040 192 : compact_lsn_range: None,
2041 192 : sub_compaction: false,
2042 192 : sub_compaction_max_job_size_mb: None,
2043 192 : },
2044 192 : ctx,
2045 192 : )
2046 192 : .await;
2047 192 : if let Err(err) = &res {
2048 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
2049 192 : }
2050 192 : res
2051 192 : }
2052 :
2053 : /// Outermost timeline compaction operation; downloads needed layers.
2054 : ///
2055 : /// NB: the cancellation token is usually from a background task, but can also come from a
2056 : /// request task.
2057 192 : pub(crate) async fn compact_with_options(
2058 192 : self: &Arc<Self>,
2059 192 : cancel: &CancellationToken,
2060 192 : options: CompactOptions,
2061 192 : ctx: &RequestContext,
2062 192 : ) -> Result<CompactionOutcome, CompactionError> {
2063 : // Acquire the compaction lock and task semaphore.
2064 : //
2065 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
2066 : // out by other background tasks (including image compaction). We request this via
2067 : // `BackgroundLoopKind::L0Compaction`.
2068 : //
2069 : // Yield for pending L0 compaction while waiting for the semaphore.
2070 192 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
2071 192 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
2072 0 : true => BackgroundLoopKind::L0Compaction,
2073 192 : false => BackgroundLoopKind::Compaction,
2074 : };
2075 192 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2076 192 : if yield_for_l0 {
2077 : // If this is an L0 pass, it doesn't make sense to yield for L0.
2078 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2079 : // If `compaction_l0_first` is disabled, there's no point yielding.
2080 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2081 192 : }
2082 :
2083 192 : let acquire = async move {
2084 192 : let guard = self.compaction_lock.lock().await;
2085 192 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2086 192 : (guard, permit)
2087 192 : };
2088 :
2089 192 : let (_guard, _permit) = tokio::select! {
2090 192 : (guard, permit) = acquire => (guard, permit),
2091 192 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2092 0 : return Ok(CompactionOutcome::YieldForL0);
2093 : }
2094 192 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2095 192 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2096 : };
2097 :
2098 192 : let last_record_lsn = self.get_last_record_lsn();
2099 :
2100 : // Last record Lsn could be zero in case the timeline was just created
2101 192 : if !last_record_lsn.is_valid() {
2102 0 : warn!(
2103 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2104 : );
2105 0 : return Ok(CompactionOutcome::Skipped);
2106 192 : }
2107 :
2108 192 : let result = match self.get_compaction_algorithm_settings().kind {
2109 : CompactionAlgorithm::Tiered => {
2110 0 : self.compact_tiered(cancel, ctx).await?;
2111 0 : Ok(CompactionOutcome::Done)
2112 : }
2113 192 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2114 : };
2115 :
2116 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2117 0 : match &result {
2118 192 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2119 0 : Err(e) if e.is_cancel() => {}
2120 0 : Err(CompactionError::ShuttingDown) => {
2121 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
2122 0 : }
2123 : Err(CompactionError::Other(_)) => {
2124 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
2125 : }
2126 : };
2127 :
2128 192 : result
2129 192 : }
2130 :
2131 : /// Mutate the timeline with a [`TimelineWriter`].
2132 2566607 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2133 : TimelineWriter {
2134 2566607 : tl: self,
2135 2566607 : write_guard: self.write_lock.lock().await,
2136 : }
2137 2566607 : }
2138 :
2139 0 : pub(crate) fn activate(
2140 0 : self: &Arc<Self>,
2141 0 : parent: Arc<crate::tenant::TenantShard>,
2142 0 : broker_client: BrokerClientChannel,
2143 0 : background_jobs_can_start: Option<&completion::Barrier>,
2144 0 : ctx: &RequestContext,
2145 0 : ) {
2146 0 : if self.tenant_shard_id.is_shard_zero() {
2147 0 : // Logical size is only maintained accurately on shard zero.
2148 0 : self.spawn_initial_logical_size_computation_task(ctx);
2149 0 : }
2150 0 : self.launch_wal_receiver(ctx, broker_client);
2151 0 : self.set_state(TimelineState::Active);
2152 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2153 0 : }
2154 :
2155 : /// After this function returns, there are no timeline-scoped tasks are left running.
2156 : ///
2157 : /// The preferred pattern for is:
2158 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2159 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2160 : /// go the extra mile and keep track of JoinHandles
2161 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2162 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2163 : ///
2164 : /// For legacy reasons, we still have multiple tasks spawned using
2165 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2166 : /// We refer to these as "timeline-scoped task_mgr tasks".
2167 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2168 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2169 : /// or [`task_mgr::shutdown_watcher`].
2170 : /// We want to gradually convert the code base away from these.
2171 : ///
2172 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2173 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2174 : /// ones that aren't mentioned here):
2175 : /// - [`TaskKind::TimelineDeletionWorker`]
2176 : /// - NB: also used for tenant deletion
2177 : /// - [`TaskKind::RemoteUploadTask`]`
2178 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2179 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2180 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2181 : /// - [`TaskKind::Eviction`]
2182 : /// - [`TaskKind::LayerFlushTask`]
2183 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2184 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2185 5 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2186 5 : debug_assert_current_span_has_tenant_and_timeline_id();
2187 :
2188 : // Regardless of whether we're going to try_freeze_and_flush
2189 : // cancel walreceiver to stop ingesting more data asap.
2190 : //
2191 : // Note that we're accepting a race condition here where we may
2192 : // do the final flush below, before walreceiver observes the
2193 : // cancellation and exits.
2194 : // This means we may open a new InMemoryLayer after the final flush below.
2195 : // Flush loop is also still running for a short while, so, in theory, it
2196 : // could also make its way into the upload queue.
2197 : //
2198 : // If we wait for the shutdown of the walreceiver before moving on to the
2199 : // flush, then that would be avoided. But we don't do it because the
2200 : // walreceiver entertains reads internally, which means that it possibly
2201 : // depends on the download of layers. Layer download is only sensitive to
2202 : // the cancellation of the entire timeline, so cancelling the walreceiver
2203 : // will have no effect on the individual get requests.
2204 : // This would cause problems when there is a lot of ongoing downloads or
2205 : // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
2206 : // and we can't deallocate resources of the timeline, etc.
2207 5 : let walreceiver = self.walreceiver.lock().unwrap().take();
2208 5 : tracing::debug!(
2209 0 : is_some = walreceiver.is_some(),
2210 0 : "Waiting for WalReceiverManager..."
2211 : );
2212 5 : if let Some(walreceiver) = walreceiver {
2213 0 : walreceiver.cancel().await;
2214 5 : }
2215 : // ... and inform any waiters for newer LSNs that there won't be any.
2216 5 : self.last_record_lsn.shutdown();
2217 :
2218 5 : if let ShutdownMode::FreezeAndFlush = mode {
2219 3 : let do_flush = if let Some((open, frozen)) = self
2220 3 : .layers
2221 3 : .read(LayerManagerLockHolder::Shutdown)
2222 3 : .await
2223 3 : .layer_map()
2224 3 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2225 3 : .ok()
2226 3 : .filter(|(open, frozen)| *open || *frozen > 0)
2227 : {
2228 0 : if self.remote_client.is_archived() == Some(true) {
2229 : // No point flushing on shutdown for an archived timeline: it is not important
2230 : // to have it nice and fresh after our restart, and trying to flush here might
2231 : // race with trying to offload it (which also stops the flush loop)
2232 0 : false
2233 : } else {
2234 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2235 0 : true
2236 : }
2237 : } else {
2238 : // this is double-shutdown, it'll be a no-op
2239 3 : true
2240 : };
2241 :
2242 : // we shut down walreceiver above, so, we won't add anything more
2243 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2244 : // to reach the disk & upload queue, then shut the upload queue and
2245 : // wait for it to drain.
2246 3 : if do_flush {
2247 3 : match self.freeze_and_flush().await {
2248 : Ok(_) => {
2249 : // drain the upload queue
2250 : // if we did not wait for completion here, it might be our shutdown process
2251 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2252 : // be spawned.
2253 : //
2254 : // what is problematic is the shutting down of RemoteTimelineClient, because
2255 : // obviously it does not make sense to stop while we wait for it, but what
2256 : // about corner cases like s3 suddenly hanging up?
2257 3 : self.remote_client.shutdown().await;
2258 : }
2259 : Err(FlushLayerError::Cancelled) => {
2260 : // this is likely the second shutdown, ignore silently.
2261 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2262 0 : debug_assert!(self.cancel.is_cancelled());
2263 : }
2264 0 : Err(e) => {
2265 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2266 : // we have some extra WAL replay to do next time the timeline starts.
2267 0 : warn!("failed to freeze and flush: {e:#}");
2268 : }
2269 : }
2270 :
2271 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2272 : // we also do a final check here to ensure that the queue is empty.
2273 3 : if !self.remote_client.no_pending_work() {
2274 0 : warn!(
2275 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2276 : );
2277 3 : }
2278 0 : }
2279 2 : }
2280 :
2281 5 : if let ShutdownMode::Reload = mode {
2282 : // drain the upload queue
2283 1 : self.remote_client.shutdown().await;
2284 1 : if !self.remote_client.no_pending_work() {
2285 0 : warn!(
2286 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2287 : );
2288 1 : }
2289 4 : }
2290 :
2291 : // Signal any subscribers to our cancellation token to drop out
2292 5 : tracing::debug!("Cancelling CancellationToken");
2293 5 : self.cancel.cancel();
2294 :
2295 : // If we have a background task downloading heatmap layers stop it.
2296 : // The background downloads are sensitive to timeline cancellation (done above),
2297 : // so the drain will be immediate.
2298 5 : self.stop_and_drain_heatmap_layers_download().await;
2299 :
2300 : // Ensure Prevent new page service requests from starting.
2301 5 : self.handles.shutdown();
2302 :
2303 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2304 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2305 5 : self.remote_client.stop();
2306 :
2307 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2308 : // to shut down the upload queue tasks.
2309 : // TODO: fix that, task management should be encapsulated inside remote_client.
2310 5 : task_mgr::shutdown_tasks(
2311 5 : Some(TaskKind::RemoteUploadTask),
2312 5 : Some(self.tenant_shard_id),
2313 5 : Some(self.timeline_id),
2314 5 : )
2315 5 : .await;
2316 :
2317 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2318 5 : tracing::debug!("Waiting for tasks...");
2319 5 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2320 :
2321 : {
2322 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2323 : // open.
2324 5 : let mut write_guard = self.write_lock.lock().await;
2325 5 : self.layers
2326 5 : .write(LayerManagerLockHolder::Shutdown)
2327 5 : .await
2328 5 : .shutdown(&mut write_guard);
2329 : }
2330 :
2331 : // Finally wait until any gate-holders are complete.
2332 : //
2333 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2334 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2335 5 : self.gate.close().await;
2336 :
2337 5 : self.metrics.shutdown();
2338 5 : }
2339 :
2340 235 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2341 235 : match (self.current_state(), new_state) {
2342 235 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2343 1 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2344 : }
2345 0 : (st, TimelineState::Loading) => {
2346 0 : error!("ignoring transition from {st:?} into Loading state");
2347 : }
2348 0 : (TimelineState::Broken { .. }, new_state) => {
2349 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2350 : }
2351 : (TimelineState::Stopping, TimelineState::Active) => {
2352 0 : error!("Not activating a Stopping timeline");
2353 : }
2354 234 : (_, new_state) => {
2355 234 : self.state.send_replace(new_state);
2356 234 : }
2357 : }
2358 235 : }
2359 :
2360 1 : pub(crate) fn set_broken(&self, reason: String) {
2361 1 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2362 1 : let broken_state = TimelineState::Broken {
2363 1 : reason,
2364 1 : backtrace: backtrace_str,
2365 1 : };
2366 1 : self.set_state(broken_state);
2367 :
2368 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2369 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2370 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2371 1 : self.cancel.cancel();
2372 1 : }
2373 :
2374 114735 : pub(crate) fn current_state(&self) -> TimelineState {
2375 114735 : self.state.borrow().clone()
2376 114735 : }
2377 :
2378 3 : pub(crate) fn is_broken(&self) -> bool {
2379 3 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2380 3 : }
2381 :
2382 126 : pub(crate) fn is_active(&self) -> bool {
2383 126 : self.current_state() == TimelineState::Active
2384 126 : }
2385 :
2386 8 : pub(crate) fn is_archived(&self) -> Option<bool> {
2387 8 : self.remote_client.is_archived()
2388 8 : }
2389 :
2390 8 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2391 8 : self.remote_client.is_invisible()
2392 8 : }
2393 :
2394 569 : pub(crate) fn is_stopping(&self) -> bool {
2395 569 : self.current_state() == TimelineState::Stopping
2396 569 : }
2397 :
2398 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2399 0 : self.state.subscribe()
2400 0 : }
2401 :
2402 113806 : pub(crate) async fn wait_to_become_active(
2403 113806 : &self,
2404 113806 : _ctx: &RequestContext, // Prepare for use by cancellation
2405 113806 : ) -> Result<(), TimelineState> {
2406 113806 : let mut receiver = self.state.subscribe();
2407 : loop {
2408 113806 : let current_state = receiver.borrow().clone();
2409 113806 : match current_state {
2410 : TimelineState::Loading => {
2411 0 : receiver
2412 0 : .changed()
2413 0 : .await
2414 0 : .expect("holding a reference to self");
2415 : }
2416 : TimelineState::Active => {
2417 113805 : return Ok(());
2418 : }
2419 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2420 : // There's no chance the timeline can transition back into ::Active
2421 1 : return Err(current_state);
2422 : }
2423 : }
2424 : }
2425 113806 : }
2426 :
2427 0 : pub(crate) async fn layer_map_info(
2428 0 : &self,
2429 0 : reset: LayerAccessStatsReset,
2430 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2431 0 : let guard = self
2432 0 : .layers
2433 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
2434 0 : .await;
2435 0 : let layer_map = guard.layer_map()?;
2436 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2437 0 : if let Some(open_layer) = &layer_map.open_layer {
2438 0 : in_memory_layers.push(open_layer.info());
2439 0 : }
2440 0 : for frozen_layer in &layer_map.frozen_layers {
2441 0 : in_memory_layers.push(frozen_layer.info());
2442 0 : }
2443 :
2444 0 : let historic_layers = layer_map
2445 0 : .iter_historic_layers()
2446 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2447 0 : .collect();
2448 :
2449 0 : Ok(LayerMapInfo {
2450 0 : in_memory_layers,
2451 0 : historic_layers,
2452 0 : })
2453 0 : }
2454 :
2455 : #[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))]
2456 : pub(crate) async fn download_layer(
2457 : &self,
2458 : layer_file_name: &LayerName,
2459 : ctx: &RequestContext,
2460 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2461 : let Some(layer) = self
2462 : .find_layer(layer_file_name)
2463 : .await
2464 0 : .map_err(|e| match e {
2465 : layer_manager::Shutdown => {
2466 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2467 : }
2468 0 : })?
2469 : else {
2470 : return Ok(None);
2471 : };
2472 :
2473 : layer.download(ctx).await?;
2474 :
2475 : Ok(Some(true))
2476 : }
2477 :
2478 : /// Evict just one layer.
2479 : ///
2480 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2481 0 : pub(crate) async fn evict_layer(
2482 0 : &self,
2483 0 : layer_file_name: &LayerName,
2484 0 : ) -> anyhow::Result<Option<bool>> {
2485 0 : let _gate = self
2486 0 : .gate
2487 0 : .enter()
2488 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2489 :
2490 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2491 0 : return Ok(None);
2492 : };
2493 :
2494 : // curl has this by default
2495 0 : let timeout = std::time::Duration::from_secs(120);
2496 :
2497 0 : match local_layer.evict_and_wait(timeout).await {
2498 0 : Ok(()) => Ok(Some(true)),
2499 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2500 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2501 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2502 : }
2503 0 : }
2504 :
2505 2401506 : fn should_roll(
2506 2401506 : &self,
2507 2401506 : layer_size: u64,
2508 2401506 : projected_layer_size: u64,
2509 2401506 : checkpoint_distance: u64,
2510 2401506 : projected_lsn: Lsn,
2511 2401506 : last_freeze_at: Lsn,
2512 2401506 : opened_at: Instant,
2513 2401506 : ) -> bool {
2514 2401506 : let distance = projected_lsn.widening_sub(last_freeze_at);
2515 :
2516 : // Rolling the open layer can be triggered by:
2517 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2518 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2519 : // account for how writes are distributed across shards: we expect each node to consume
2520 : // 1/count of the LSN on average.
2521 : // 2. The size of the currently open layer.
2522 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2523 : // up and suspend activity.
2524 2401506 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2525 0 : info!(
2526 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2527 : projected_lsn, layer_size, distance
2528 : );
2529 :
2530 0 : true
2531 2401506 : } else if projected_layer_size >= checkpoint_distance {
2532 : // NB: this check is relied upon by:
2533 40 : let _ = IndexEntry::validate_checkpoint_distance;
2534 40 : info!(
2535 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2536 : projected_lsn, layer_size, projected_layer_size
2537 : );
2538 :
2539 40 : true
2540 2401466 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2541 0 : info!(
2542 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2543 : projected_lsn,
2544 : layer_size,
2545 0 : opened_at.elapsed()
2546 : );
2547 :
2548 0 : true
2549 : } else {
2550 2401466 : false
2551 : }
2552 2401506 : }
2553 :
2554 1 : pub(crate) fn is_basebackup_cache_enabled(&self) -> bool {
2555 1 : let tenant_conf = self.tenant_conf.load();
2556 1 : tenant_conf
2557 1 : .tenant_conf
2558 1 : .basebackup_cache_enabled
2559 1 : .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
2560 1 : }
2561 :
2562 : /// Try to get a basebackup from the on-disk cache.
2563 0 : pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
2564 0 : self.basebackup_cache
2565 0 : .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
2566 0 : .await
2567 0 : }
2568 :
2569 : /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
2570 : /// the given request parameters.
2571 : ///
2572 : /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
2573 0 : pub async fn get_cached_basebackup_if_enabled(
2574 0 : &self,
2575 0 : lsn: Option<Lsn>,
2576 0 : prev_lsn: Option<Lsn>,
2577 0 : full: bool,
2578 0 : replica: bool,
2579 0 : gzip: bool,
2580 0 : ) -> Option<tokio::fs::File> {
2581 0 : if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
2582 0 : return None;
2583 0 : }
2584 : // We have to know which LSN to fetch the basebackup for.
2585 0 : let lsn = lsn?;
2586 : // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
2587 0 : if prev_lsn.is_some() || full || replica || !gzip {
2588 0 : return None;
2589 0 : }
2590 0 : self.get_cached_basebackup(lsn).await
2591 0 : }
2592 :
2593 : /// Prepare basebackup for the given LSN and store it in the basebackup cache.
2594 : /// The method is asynchronous and returns immediately.
2595 : /// The actual basebackup preparation is performed in the background
2596 : /// by the basebackup cache on a best-effort basis.
2597 1 : pub(crate) fn prepare_basebackup(&self, lsn: Lsn) {
2598 1 : if !self.is_basebackup_cache_enabled() {
2599 1 : return;
2600 0 : }
2601 0 : if !self.tenant_shard_id.is_shard_zero() {
2602 : // In theory we should never get here, but just in case check it.
2603 : // Preparing basebackup doesn't make sense for shards other than shard zero.
2604 0 : return;
2605 0 : }
2606 0 : if !self.is_active() {
2607 : // May happen during initial timeline creation.
2608 : // Such timeline is not in the global timeline map yet,
2609 : // so basebackup cache will not be able to find it.
2610 : // TODO(diko): We can prepare such timelines in finish_creation().
2611 0 : return;
2612 0 : }
2613 :
2614 0 : self.basebackup_cache
2615 0 : .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
2616 1 : }
2617 : }
2618 :
2619 : /// Number of times we will compute partition within a checkpoint distance.
2620 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2621 :
2622 : // Private functions
2623 : impl Timeline {
2624 6 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2625 6 : let tenant_conf = self.tenant_conf.load();
2626 6 : tenant_conf
2627 6 : .tenant_conf
2628 6 : .lsn_lease_length
2629 6 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2630 6 : }
2631 :
2632 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2633 0 : let tenant_conf = self.tenant_conf.load();
2634 0 : tenant_conf
2635 0 : .tenant_conf
2636 0 : .lsn_lease_length_for_ts
2637 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2638 0 : }
2639 :
2640 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2641 0 : let tenant_conf = self.tenant_conf.load();
2642 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2643 0 : }
2644 :
2645 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2646 0 : let tenant_conf = self.tenant_conf.load();
2647 0 : tenant_conf
2648 0 : .tenant_conf
2649 0 : .lazy_slru_download
2650 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2651 0 : }
2652 :
2653 : /// Checks if a get page request should get perf tracing
2654 : ///
2655 : /// The configuration priority is: tenant config override, default tenant config,
2656 : /// pageserver config.
2657 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2658 0 : let tenant_conf = self.tenant_conf.load();
2659 0 : let ratio = tenant_conf
2660 0 : .tenant_conf
2661 0 : .sampling_ratio
2662 0 : .flatten()
2663 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2664 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2665 :
2666 0 : match ratio {
2667 0 : Some(r) => {
2668 0 : if r.numerator == 0 {
2669 0 : false
2670 : } else {
2671 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2672 : }
2673 : }
2674 0 : None => false,
2675 : }
2676 0 : }
2677 :
2678 2402185 : fn get_checkpoint_distance(&self) -> u64 {
2679 2402185 : let tenant_conf = self.tenant_conf.load();
2680 2402185 : tenant_conf
2681 2402185 : .tenant_conf
2682 2402185 : .checkpoint_distance
2683 2402185 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2684 2402185 : }
2685 :
2686 2401466 : fn get_checkpoint_timeout(&self) -> Duration {
2687 2401466 : let tenant_conf = self.tenant_conf.load();
2688 2401466 : tenant_conf
2689 2401466 : .tenant_conf
2690 2401466 : .checkpoint_timeout
2691 2401466 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2692 2401466 : }
2693 :
2694 0 : pub(crate) fn get_pitr_interval(&self) -> Duration {
2695 0 : let tenant_conf = &self.tenant_conf.load().tenant_conf;
2696 0 : tenant_conf
2697 0 : .pitr_interval
2698 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2699 0 : }
2700 :
2701 1271 : fn get_compaction_period(&self) -> Duration {
2702 1271 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2703 1271 : tenant_conf
2704 1271 : .compaction_period
2705 1271 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2706 1271 : }
2707 :
2708 351 : fn get_compaction_target_size(&self) -> u64 {
2709 351 : let tenant_conf = self.tenant_conf.load();
2710 351 : tenant_conf
2711 351 : .tenant_conf
2712 351 : .compaction_target_size
2713 351 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2714 351 : }
2715 :
2716 817 : fn get_compaction_threshold(&self) -> usize {
2717 817 : let tenant_conf = self.tenant_conf.load();
2718 817 : tenant_conf
2719 817 : .tenant_conf
2720 817 : .compaction_threshold
2721 817 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2722 817 : }
2723 :
2724 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2725 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2726 : /// possible that the index part persists the state while the config doesn't get persisted.
2727 973 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2728 973 : let tenant_conf = self.tenant_conf.load();
2729 973 : tenant_conf
2730 973 : .tenant_conf
2731 973 : .rel_size_v2_enabled
2732 973 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2733 973 : }
2734 :
2735 1099 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2736 1099 : self.rel_size_v2_status
2737 1099 : .load()
2738 1099 : .as_ref()
2739 1099 : .map(|s| s.as_ref().clone())
2740 1099 : .unwrap_or(RelSizeMigration::Legacy)
2741 1099 : }
2742 :
2743 23 : fn get_compaction_upper_limit(&self) -> usize {
2744 23 : let tenant_conf = self.tenant_conf.load();
2745 23 : tenant_conf
2746 23 : .tenant_conf
2747 23 : .compaction_upper_limit
2748 23 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2749 23 : }
2750 :
2751 0 : pub fn get_compaction_l0_first(&self) -> bool {
2752 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2753 0 : tenant_conf
2754 0 : .compaction_l0_first
2755 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2756 0 : }
2757 :
2758 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2759 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2760 0 : tenant_conf
2761 0 : .compaction_l0_semaphore
2762 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2763 0 : }
2764 :
2765 635 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2766 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2767 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2768 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2769 :
2770 : // If compaction is disabled, don't delay.
2771 635 : if self.get_compaction_period() == Duration::ZERO {
2772 632 : return None;
2773 3 : }
2774 :
2775 3 : let compaction_threshold = self.get_compaction_threshold();
2776 3 : let tenant_conf = self.tenant_conf.load();
2777 3 : let l0_flush_delay_threshold = tenant_conf
2778 3 : .tenant_conf
2779 3 : .l0_flush_delay_threshold
2780 3 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2781 3 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2782 :
2783 : // 0 disables backpressure.
2784 3 : if l0_flush_delay_threshold == 0 {
2785 0 : return None;
2786 3 : }
2787 :
2788 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2789 : // backpressure flushes below this.
2790 : // TODO: the tenant config should have validation to prevent this instead.
2791 3 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2792 3 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2793 635 : }
2794 :
2795 636 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2796 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2797 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2798 : // sibling timelines. We need more confidence before enabling this.
2799 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2800 :
2801 : // If compaction is disabled, don't stall.
2802 636 : if self.get_compaction_period() == Duration::ZERO {
2803 632 : return None;
2804 4 : }
2805 :
2806 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2807 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2808 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2809 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2810 4 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2811 0 : return None;
2812 4 : }
2813 :
2814 4 : let compaction_threshold = self.get_compaction_threshold();
2815 4 : let tenant_conf = self.tenant_conf.load();
2816 4 : let l0_flush_stall_threshold = tenant_conf
2817 4 : .tenant_conf
2818 4 : .l0_flush_stall_threshold
2819 4 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2820 :
2821 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2822 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2823 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2824 4 : if cfg!(feature = "testing")
2825 4 : && compaction_threshold == 1
2826 0 : && l0_flush_stall_threshold.is_none()
2827 : {
2828 0 : return None;
2829 4 : }
2830 :
2831 4 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2832 4 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2833 :
2834 : // 0 disables backpressure.
2835 4 : if l0_flush_stall_threshold == 0 {
2836 4 : return None;
2837 0 : }
2838 :
2839 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2840 : // backpressure flushes below this.
2841 : // TODO: the tenant config should have validation to prevent this instead.
2842 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2843 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2844 636 : }
2845 :
2846 7 : fn get_image_creation_threshold(&self) -> usize {
2847 7 : let tenant_conf = self.tenant_conf.load();
2848 7 : tenant_conf
2849 7 : .tenant_conf
2850 7 : .image_creation_threshold
2851 7 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2852 7 : }
2853 :
2854 192 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2855 192 : let tenant_conf = &self.tenant_conf.load();
2856 192 : tenant_conf
2857 192 : .tenant_conf
2858 192 : .compaction_algorithm
2859 192 : .as_ref()
2860 192 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2861 192 : .clone()
2862 192 : }
2863 :
2864 192 : pub fn get_compaction_shard_ancestor(&self) -> bool {
2865 192 : let tenant_conf = self.tenant_conf.load();
2866 192 : tenant_conf
2867 192 : .tenant_conf
2868 192 : .compaction_shard_ancestor
2869 192 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
2870 192 : }
2871 :
2872 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2873 0 : let tenant_conf = self.tenant_conf.load();
2874 0 : tenant_conf
2875 0 : .tenant_conf
2876 0 : .eviction_policy
2877 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2878 0 : }
2879 :
2880 234 : fn get_evictions_low_residence_duration_metric_threshold(
2881 234 : tenant_conf: &pageserver_api::models::TenantConfig,
2882 234 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
2883 234 : ) -> Duration {
2884 234 : tenant_conf
2885 234 : .evictions_low_residence_duration_metric_threshold
2886 234 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2887 234 : }
2888 :
2889 190 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2890 190 : let tenant_conf = self.tenant_conf.load();
2891 190 : tenant_conf
2892 190 : .tenant_conf
2893 190 : .image_layer_creation_check_threshold
2894 190 : .unwrap_or(
2895 190 : self.conf
2896 190 : .default_tenant_conf
2897 190 : .image_layer_creation_check_threshold,
2898 : )
2899 190 : }
2900 :
2901 27 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
2902 27 : let tenant_conf = &self.tenant_conf.load();
2903 27 : let gc_compaction_enabled = tenant_conf
2904 27 : .tenant_conf
2905 27 : .gc_compaction_enabled
2906 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
2907 27 : let gc_compaction_verification = tenant_conf
2908 27 : .tenant_conf
2909 27 : .gc_compaction_verification
2910 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
2911 27 : let gc_compaction_initial_threshold_kb = tenant_conf
2912 27 : .tenant_conf
2913 27 : .gc_compaction_initial_threshold_kb
2914 27 : .unwrap_or(
2915 27 : self.conf
2916 27 : .default_tenant_conf
2917 27 : .gc_compaction_initial_threshold_kb,
2918 : );
2919 27 : let gc_compaction_ratio_percent = tenant_conf
2920 27 : .tenant_conf
2921 27 : .gc_compaction_ratio_percent
2922 27 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
2923 27 : GcCompactionCombinedSettings {
2924 27 : gc_compaction_enabled,
2925 27 : gc_compaction_verification,
2926 27 : gc_compaction_initial_threshold_kb,
2927 27 : gc_compaction_ratio_percent,
2928 27 : }
2929 27 : }
2930 :
2931 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2932 0 : let tenant_conf = self.tenant_conf.load();
2933 0 : tenant_conf
2934 0 : .tenant_conf
2935 0 : .image_creation_preempt_threshold
2936 0 : .unwrap_or(
2937 0 : self.conf
2938 0 : .default_tenant_conf
2939 0 : .image_creation_preempt_threshold,
2940 : )
2941 0 : }
2942 :
2943 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2944 : // NB: Most tenant conf options are read by background loops, so,
2945 : // changes will automatically be picked up.
2946 :
2947 : // The threshold is embedded in the metric. So, we need to update it.
2948 : {
2949 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2950 0 : &new_conf.tenant_conf,
2951 0 : &self.conf.default_tenant_conf,
2952 : );
2953 :
2954 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2955 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2956 :
2957 0 : let timeline_id_str = self.timeline_id.to_string();
2958 :
2959 0 : self.remote_client.update_config(&new_conf.location);
2960 :
2961 0 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
2962 0 : if let Some(new_capacity) = new_conf.tenant_conf.relsize_snapshot_cache_capacity {
2963 0 : if new_capacity != rel_size_cache.capacity() {
2964 0 : rel_size_cache.set_capacity(new_capacity);
2965 0 : }
2966 0 : }
2967 :
2968 0 : self.metrics
2969 0 : .evictions_with_low_residence_duration
2970 0 : .write()
2971 0 : .unwrap()
2972 0 : .change_threshold(
2973 0 : &tenant_id_str,
2974 0 : &shard_id_str,
2975 0 : &timeline_id_str,
2976 0 : new_threshold,
2977 0 : );
2978 : }
2979 0 : }
2980 :
2981 : /// Open a Timeline handle.
2982 : ///
2983 : /// Loads the metadata for the timeline into memory, but not the layer map.
2984 : #[allow(clippy::too_many_arguments)]
2985 234 : pub(super) fn new(
2986 234 : conf: &'static PageServerConf,
2987 234 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2988 234 : metadata: &TimelineMetadata,
2989 234 : previous_heatmap: Option<PreviousHeatmap>,
2990 234 : ancestor: Option<Arc<Timeline>>,
2991 234 : timeline_id: TimelineId,
2992 234 : tenant_shard_id: TenantShardId,
2993 234 : generation: Generation,
2994 234 : shard_identity: ShardIdentity,
2995 234 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2996 234 : resources: TimelineResources,
2997 234 : pg_version: PgMajorVersion,
2998 234 : state: TimelineState,
2999 234 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
3000 234 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
3001 234 : gc_compaction_state: Option<GcCompactionState>,
3002 234 : rel_size_v2_status: Option<RelSizeMigration>,
3003 234 : cancel: CancellationToken,
3004 234 : ) -> Arc<Self> {
3005 234 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
3006 234 : let (state, _) = watch::channel(state);
3007 :
3008 234 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
3009 234 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
3010 :
3011 234 : let evictions_low_residence_duration_metric_threshold = {
3012 234 : let loaded_tenant_conf = tenant_conf.load();
3013 234 : Self::get_evictions_low_residence_duration_metric_threshold(
3014 234 : &loaded_tenant_conf.tenant_conf,
3015 234 : &conf.default_tenant_conf,
3016 : )
3017 : };
3018 :
3019 234 : if let Some(ancestor) = &ancestor {
3020 118 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
3021 118 : // If we construct an explicit timeline object, it's obviously not offloaded
3022 118 : let is_offloaded = MaybeOffloaded::No;
3023 118 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
3024 118 : }
3025 :
3026 234 : let relsize_snapshot_cache_capacity = {
3027 234 : let loaded_tenant_conf = tenant_conf.load();
3028 234 : loaded_tenant_conf
3029 234 : .tenant_conf
3030 234 : .relsize_snapshot_cache_capacity
3031 234 : .unwrap_or(conf.default_tenant_conf.relsize_snapshot_cache_capacity)
3032 : };
3033 :
3034 234 : Arc::new_cyclic(|myself| {
3035 234 : let metrics = Arc::new(TimelineMetrics::new(
3036 234 : &tenant_shard_id,
3037 234 : &timeline_id,
3038 234 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
3039 : "mtime",
3040 234 : evictions_low_residence_duration_metric_threshold,
3041 : ),
3042 : ));
3043 234 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
3044 :
3045 234 : let mut result = Timeline {
3046 234 : conf,
3047 234 : tenant_conf,
3048 234 : myself: myself.clone(),
3049 234 : timeline_id,
3050 234 : tenant_shard_id,
3051 234 : generation,
3052 234 : shard_identity,
3053 234 : pg_version,
3054 234 : layers: Default::default(),
3055 234 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
3056 :
3057 234 : walredo_mgr,
3058 234 : walreceiver: Mutex::new(None),
3059 :
3060 234 : remote_client: Arc::new(resources.remote_client),
3061 :
3062 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
3063 234 : last_record_lsn: SeqWait::new(RecordLsn {
3064 234 : last: disk_consistent_lsn,
3065 234 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
3066 234 : }),
3067 234 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
3068 :
3069 234 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
3070 :
3071 234 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
3072 234 : last_freeze_ts: RwLock::new(Instant::now()),
3073 :
3074 234 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
3075 :
3076 234 : ancestor_timeline: ancestor,
3077 234 : ancestor_lsn: metadata.ancestor_lsn(),
3078 :
3079 234 : metrics,
3080 :
3081 234 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
3082 234 : &tenant_shard_id,
3083 234 : &timeline_id,
3084 234 : resources.pagestream_throttle_metrics,
3085 : ),
3086 :
3087 1872 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
3088 1872 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
3089 :
3090 234 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
3091 :
3092 234 : layer_flush_start_tx,
3093 234 : layer_flush_done_tx,
3094 :
3095 234 : write_lock: tokio::sync::Mutex::new(None),
3096 :
3097 234 : gc_info: std::sync::RwLock::new(GcInfo::default()),
3098 :
3099 234 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
3100 234 : LastImageLayerCreationStatus::default(),
3101 : )),
3102 :
3103 234 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
3104 234 : initdb_lsn: metadata.initdb_lsn(),
3105 :
3106 234 : current_logical_size: if disk_consistent_lsn.is_valid() {
3107 : // we're creating timeline data with some layer files existing locally,
3108 : // need to recalculate timeline's logical size based on data in the layers.
3109 120 : LogicalSize::deferred_initial(disk_consistent_lsn)
3110 : } else {
3111 : // we're creating timeline data without any layers existing locally,
3112 : // initial logical size is 0.
3113 114 : LogicalSize::empty_initial()
3114 : },
3115 :
3116 234 : partitioning: GuardArcSwap::new((
3117 234 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
3118 234 : Lsn(0),
3119 234 : )),
3120 : repartition_threshold: 0,
3121 234 : last_image_layer_creation_check_at: AtomicLsn::new(0),
3122 234 : last_image_layer_creation_check_instant: Mutex::new(None),
3123 :
3124 234 : last_received_wal: Mutex::new(None),
3125 234 : rel_size_latest_cache: RwLock::new(HashMap::new()),
3126 234 : rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
3127 :
3128 234 : download_all_remote_layers_task_info: RwLock::new(None),
3129 :
3130 234 : state,
3131 :
3132 234 : eviction_task_timeline_state: tokio::sync::Mutex::new(
3133 234 : EvictionTaskTimelineState::default(),
3134 : ),
3135 234 : delete_progress: TimelineDeleteProgress::default(),
3136 :
3137 234 : cancel,
3138 234 : gate: Gate::default(),
3139 :
3140 234 : compaction_lock: tokio::sync::Mutex::default(),
3141 234 : compaction_failed: AtomicBool::default(),
3142 234 : l0_compaction_trigger: resources.l0_compaction_trigger,
3143 234 : gc_lock: tokio::sync::Mutex::default(),
3144 :
3145 234 : standby_horizon: AtomicLsn::new(0),
3146 :
3147 234 : pagestream_throttle: resources.pagestream_throttle,
3148 :
3149 234 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
3150 :
3151 : #[cfg(test)]
3152 234 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3153 :
3154 234 : l0_flush_global_state: resources.l0_flush_global_state,
3155 :
3156 234 : handles: Default::default(),
3157 :
3158 234 : attach_wal_lag_cooldown,
3159 :
3160 234 : create_idempotency,
3161 :
3162 234 : page_trace: Default::default(),
3163 :
3164 234 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3165 :
3166 234 : heatmap_layers_downloader: Mutex::new(None),
3167 :
3168 234 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
3169 :
3170 234 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3171 :
3172 234 : basebackup_cache: resources.basebackup_cache,
3173 :
3174 234 : feature_resolver: resources.feature_resolver.clone(),
3175 : };
3176 :
3177 234 : result.repartition_threshold =
3178 234 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3179 :
3180 234 : result
3181 234 : .metrics
3182 234 : .last_record_lsn_gauge
3183 234 : .set(disk_consistent_lsn.0 as i64);
3184 234 : result
3185 234 : })
3186 234 : }
3187 :
3188 341 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3189 341 : let Ok(guard) = self.gate.enter() else {
3190 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3191 0 : return;
3192 : };
3193 341 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3194 341 : match *flush_loop_state {
3195 231 : FlushLoopState::NotStarted => (),
3196 : FlushLoopState::Running { .. } => {
3197 110 : info!(
3198 0 : "skipping attempt to start flush_loop twice {}/{}",
3199 0 : self.tenant_shard_id, self.timeline_id
3200 : );
3201 110 : return;
3202 : }
3203 : FlushLoopState::Exited => {
3204 0 : info!(
3205 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3206 0 : self.tenant_shard_id, self.timeline_id
3207 : );
3208 0 : return;
3209 : }
3210 : }
3211 :
3212 231 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3213 231 : let self_clone = Arc::clone(self);
3214 :
3215 231 : debug!("spawning flush loop");
3216 231 : *flush_loop_state = FlushLoopState::Running {
3217 231 : #[cfg(test)]
3218 231 : expect_initdb_optimization: false,
3219 231 : #[cfg(test)]
3220 231 : initdb_optimization_count: 0,
3221 231 : };
3222 231 : task_mgr::spawn(
3223 231 : task_mgr::BACKGROUND_RUNTIME.handle(),
3224 231 : task_mgr::TaskKind::LayerFlushTask,
3225 231 : self.tenant_shard_id,
3226 231 : Some(self.timeline_id),
3227 231 : "layer flush task",
3228 231 : async move {
3229 231 : let _guard = guard;
3230 231 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3231 231 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3232 5 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3233 5 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3234 5 : *flush_loop_state = FlushLoopState::Exited;
3235 5 : Ok(())
3236 5 : }
3237 231 : .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))
3238 : );
3239 341 : }
3240 :
3241 0 : pub(crate) fn update_gc_compaction_state(
3242 0 : &self,
3243 0 : gc_compaction_state: GcCompactionState,
3244 0 : ) -> anyhow::Result<()> {
3245 0 : self.gc_compaction_state
3246 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3247 0 : self.remote_client
3248 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3249 0 : }
3250 :
3251 0 : pub(crate) fn update_rel_size_v2_status(
3252 0 : &self,
3253 0 : rel_size_v2_status: RelSizeMigration,
3254 0 : ) -> anyhow::Result<()> {
3255 0 : self.rel_size_v2_status
3256 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3257 0 : self.remote_client
3258 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3259 0 : }
3260 :
3261 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3262 0 : self.gc_compaction_state.load_full().as_ref().clone()
3263 0 : }
3264 :
3265 : /// Creates and starts the wal receiver.
3266 : ///
3267 : /// This function is expected to be called at most once per Timeline's lifecycle
3268 : /// when the timeline is activated.
3269 0 : fn launch_wal_receiver(
3270 0 : self: &Arc<Self>,
3271 0 : ctx: &RequestContext,
3272 0 : broker_client: BrokerClientChannel,
3273 0 : ) {
3274 0 : info!(
3275 0 : "launching WAL receiver for timeline {} of tenant {}",
3276 0 : self.timeline_id, self.tenant_shard_id
3277 : );
3278 :
3279 0 : let tenant_conf = self.tenant_conf.load();
3280 0 : let wal_connect_timeout = tenant_conf
3281 0 : .tenant_conf
3282 0 : .walreceiver_connect_timeout
3283 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3284 0 : let lagging_wal_timeout = tenant_conf
3285 0 : .tenant_conf
3286 0 : .lagging_wal_timeout
3287 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3288 0 : let max_lsn_wal_lag = tenant_conf
3289 0 : .tenant_conf
3290 0 : .max_lsn_wal_lag
3291 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3292 :
3293 0 : let mut guard = self.walreceiver.lock().unwrap();
3294 0 : assert!(
3295 0 : guard.is_none(),
3296 0 : "multiple launches / re-launches of WAL receiver are not supported"
3297 : );
3298 :
3299 0 : let protocol = PostgresClientProtocol::Interpreted {
3300 0 : format: utils::postgres_client::InterpretedFormat::Protobuf,
3301 0 : compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
3302 0 : };
3303 :
3304 0 : *guard = Some(WalReceiver::start(
3305 0 : Arc::clone(self),
3306 0 : WalReceiverConf {
3307 0 : protocol,
3308 0 : wal_connect_timeout,
3309 0 : lagging_wal_timeout,
3310 0 : max_lsn_wal_lag,
3311 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3312 0 : availability_zone: self.conf.availability_zone.clone(),
3313 0 : ingest_batch_size: self.conf.ingest_batch_size,
3314 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3315 0 : },
3316 0 : broker_client,
3317 0 : ctx,
3318 0 : ));
3319 0 : }
3320 :
3321 : /// Initialize with an empty layer map. Used when creating a new timeline.
3322 231 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3323 231 : let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
3324 231 : "in the context where we call this function, no other task has access to the object",
3325 : );
3326 231 : layers
3327 231 : .open_mut()
3328 231 : .expect("in this context the LayerManager must still be open")
3329 231 : .initialize_empty(Lsn(start_lsn.0));
3330 231 : }
3331 :
3332 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3333 : /// files.
3334 3 : pub(super) async fn load_layer_map(
3335 3 : &self,
3336 3 : disk_consistent_lsn: Lsn,
3337 3 : index_part: IndexPart,
3338 3 : ) -> anyhow::Result<()> {
3339 : use LayerName::*;
3340 : use init::Decision::*;
3341 : use init::{Discovered, DismissedLayer};
3342 :
3343 3 : let mut guard = self
3344 3 : .layers
3345 3 : .write(LayerManagerLockHolder::LoadLayerMap)
3346 3 : .await;
3347 :
3348 3 : let timer = self.metrics.load_layer_map_histo.start_timer();
3349 :
3350 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3351 : // structs representing all files on disk
3352 3 : let timeline_path = self
3353 3 : .conf
3354 3 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3355 3 : let conf = self.conf;
3356 3 : let span = tracing::Span::current();
3357 :
3358 : // Copy to move into the task we're about to spawn
3359 3 : let this = self.myself.upgrade().expect("&self method holds the arc");
3360 :
3361 3 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3362 3 : move || {
3363 3 : let _g = span.entered();
3364 3 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3365 3 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3366 3 : let mut unrecognized_files = Vec::new();
3367 :
3368 3 : let mut path = timeline_path;
3369 :
3370 11 : for discovered in discovered {
3371 8 : let (name, kind) = match discovered {
3372 8 : Discovered::Layer(layer_file_name, local_metadata) => {
3373 8 : discovered_layers.push((layer_file_name, local_metadata));
3374 8 : continue;
3375 : }
3376 0 : Discovered::IgnoredBackup(path) => {
3377 0 : std::fs::remove_file(path)
3378 0 : .or_else(fs_ext::ignore_not_found)
3379 0 : .fatal_err("Removing .old file");
3380 0 : continue;
3381 : }
3382 0 : Discovered::Unknown(file_name) => {
3383 : // we will later error if there are any
3384 0 : unrecognized_files.push(file_name);
3385 0 : continue;
3386 : }
3387 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3388 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3389 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3390 : };
3391 0 : path.push(Utf8Path::new(&name));
3392 0 : init::cleanup(&path, kind)?;
3393 0 : path.pop();
3394 : }
3395 :
3396 3 : if !unrecognized_files.is_empty() {
3397 : // assume that if there are any there are many many.
3398 0 : let n = unrecognized_files.len();
3399 0 : let first = &unrecognized_files[..n.min(10)];
3400 0 : anyhow::bail!(
3401 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3402 : );
3403 3 : }
3404 :
3405 3 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3406 :
3407 3 : let mut loaded_layers = Vec::new();
3408 3 : let mut needs_cleanup = Vec::new();
3409 3 : let mut total_physical_size = 0;
3410 :
3411 11 : for (name, decision) in decided {
3412 8 : let decision = match decision {
3413 8 : Ok(decision) => decision,
3414 0 : Err(DismissedLayer::Future { local }) => {
3415 0 : if let Some(local) = local {
3416 0 : init::cleanup_future_layer(
3417 0 : &local.local_path,
3418 0 : &name,
3419 0 : disk_consistent_lsn,
3420 0 : )?;
3421 0 : }
3422 0 : needs_cleanup.push(name);
3423 0 : continue;
3424 : }
3425 0 : Err(DismissedLayer::LocalOnly(local)) => {
3426 0 : init::cleanup_local_only_file(&name, &local)?;
3427 : // this file never existed remotely, we will have to do rework
3428 0 : continue;
3429 : }
3430 0 : Err(DismissedLayer::BadMetadata(local)) => {
3431 0 : init::cleanup_local_file_for_remote(&local)?;
3432 : // this file never existed remotely, we will have to do rework
3433 0 : continue;
3434 : }
3435 : };
3436 :
3437 8 : match &name {
3438 6 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3439 2 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3440 : }
3441 :
3442 8 : tracing::debug!(layer=%name, ?decision, "applied");
3443 :
3444 8 : let layer = match decision {
3445 8 : Resident { local, remote } => {
3446 8 : total_physical_size += local.file_size;
3447 8 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3448 8 : .drop_eviction_guard()
3449 : }
3450 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3451 : };
3452 :
3453 8 : loaded_layers.push(layer);
3454 : }
3455 3 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3456 3 : }
3457 : })
3458 3 : .await
3459 3 : .map_err(anyhow::Error::new)
3460 3 : .and_then(|x| x)?;
3461 :
3462 3 : let num_layers = loaded_layers.len();
3463 :
3464 3 : guard
3465 3 : .open_mut()
3466 3 : .expect("layermanager must be open during init")
3467 3 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3468 :
3469 3 : self.remote_client
3470 3 : .schedule_layer_file_deletion(&needs_cleanup)?;
3471 3 : self.remote_client
3472 3 : .schedule_index_upload_for_file_changes()?;
3473 : // This barrier orders above DELETEs before any later operations.
3474 : // This is critical because code executing after the barrier might
3475 : // create again objects with the same key that we just scheduled for deletion.
3476 : // For example, if we just scheduled deletion of an image layer "from the future",
3477 : // later compaction might run again and re-create the same image layer.
3478 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3479 : // "same" here means same key range and LSN.
3480 : //
3481 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3482 : // the upload queue may execute the PUT first, then the DELETE.
3483 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3484 : //
3485 : // 1. a future image layer is created and uploaded
3486 : // 2. ps restart
3487 : // 3. the future layer from (1) is deleted during load layer map
3488 : // 4. image layer is re-created and uploaded
3489 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3490 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3491 : //
3492 : // See https://github.com/neondatabase/neon/issues/5878
3493 : //
3494 : // NB: generation numbers naturally protect against this because they disambiguate
3495 : // (1) and (4)
3496 : // TODO: this is basically a no-op now, should we remove it?
3497 3 : self.remote_client.schedule_barrier()?;
3498 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3499 : // on retry.
3500 :
3501 3 : info!(
3502 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3503 : num_layers, disk_consistent_lsn, total_physical_size
3504 : );
3505 :
3506 3 : timer.stop_and_record();
3507 3 : Ok(())
3508 3 : }
3509 :
3510 : /// Retrieve current logical size of the timeline.
3511 : ///
3512 : /// The size could be lagging behind the actual number, in case
3513 : /// the initial size calculation has not been run (gets triggered on the first size access).
3514 : ///
3515 : /// return size and boolean flag that shows if the size is exact
3516 0 : pub(crate) fn get_current_logical_size(
3517 0 : self: &Arc<Self>,
3518 0 : priority: GetLogicalSizePriority,
3519 0 : ctx: &RequestContext,
3520 0 : ) -> logical_size::CurrentLogicalSize {
3521 0 : if !self.tenant_shard_id.is_shard_zero() {
3522 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3523 : // when HTTP API is serving a GET for timeline zero, return zero
3524 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3525 0 : }
3526 :
3527 0 : let current_size = self.current_logical_size.current_size();
3528 0 : debug!("Current size: {current_size:?}");
3529 :
3530 0 : match (current_size.accuracy(), priority) {
3531 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3532 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3533 0 : // background task will eventually deliver an exact value, we're in no rush
3534 0 : }
3535 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3536 : // background task is not ready, but user is asking for it now;
3537 : // => make the background task skip the line
3538 : // (The alternative would be to calculate the size here, but,
3539 : // it can actually take a long time if the user has a lot of rels.
3540 : // And we'll inevitable need it again; So, let the background task do the work.)
3541 0 : match self
3542 0 : .current_logical_size
3543 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3544 0 : .get()
3545 : {
3546 0 : Some(cancel) => cancel.cancel(),
3547 : None => {
3548 0 : match self.current_state() {
3549 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3550 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3551 0 : // Don't make noise.
3552 0 : }
3553 : TimelineState::Loading => {
3554 : // Import does not return an activated timeline.
3555 0 : info!(
3556 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3557 : );
3558 : }
3559 : TimelineState::Active => {
3560 : // activation should be setting the once cell
3561 0 : warn!(
3562 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3563 : );
3564 0 : debug_assert!(false);
3565 : }
3566 : }
3567 : }
3568 : }
3569 : }
3570 : }
3571 :
3572 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3573 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3574 0 : let first = self
3575 0 : .current_logical_size
3576 0 : .did_return_approximate_to_walreceiver
3577 0 : .compare_exchange(
3578 0 : false,
3579 0 : true,
3580 0 : AtomicOrdering::Relaxed,
3581 0 : AtomicOrdering::Relaxed,
3582 0 : )
3583 0 : .is_ok();
3584 0 : if first {
3585 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3586 0 : }
3587 0 : }
3588 0 : }
3589 :
3590 0 : current_size
3591 0 : }
3592 :
3593 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3594 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3595 : // nothing to do for freshly created timelines;
3596 0 : assert_eq!(
3597 0 : self.current_logical_size.current_size().accuracy(),
3598 : logical_size::Accuracy::Exact,
3599 : );
3600 0 : self.current_logical_size.initialized.add_permits(1);
3601 0 : return;
3602 : };
3603 :
3604 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3605 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3606 0 : self.current_logical_size
3607 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3608 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3609 :
3610 0 : let self_clone = Arc::clone(self);
3611 0 : let background_ctx = ctx.detached_child(
3612 0 : TaskKind::InitialLogicalSizeCalculation,
3613 0 : DownloadBehavior::Download,
3614 : );
3615 0 : task_mgr::spawn(
3616 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3617 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3618 0 : self.tenant_shard_id,
3619 0 : Some(self.timeline_id),
3620 0 : "initial size calculation",
3621 : // NB: don't log errors here, task_mgr will do that.
3622 0 : async move {
3623 0 : self_clone
3624 0 : .initial_logical_size_calculation_task(
3625 0 : initial_part_end,
3626 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3627 0 : background_ctx,
3628 0 : )
3629 0 : .await;
3630 0 : Ok(())
3631 0 : }
3632 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)),
3633 : );
3634 0 : }
3635 :
3636 : /// # Cancellation
3637 : ///
3638 : /// This method is sensitive to `Timeline::cancel`.
3639 : ///
3640 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3641 : ///
3642 : /// # Cancel-Safety
3643 : ///
3644 : /// It does Timeline IO, hence this should be polled to completion because
3645 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3646 : /// to reason about.
3647 0 : async fn initial_logical_size_calculation_task(
3648 0 : self: Arc<Self>,
3649 0 : initial_part_end: Lsn,
3650 0 : skip_concurrency_limiter: CancellationToken,
3651 0 : background_ctx: RequestContext,
3652 0 : ) {
3653 0 : scopeguard::defer! {
3654 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3655 : self.current_logical_size.initialized.add_permits(1);
3656 : }
3657 :
3658 0 : let try_once = |attempt: usize| {
3659 0 : let background_ctx = &background_ctx;
3660 0 : let self_ref = &self;
3661 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3662 0 : async move {
3663 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3664 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3665 0 : background_ctx,
3666 : );
3667 :
3668 : use crate::metrics::initial_logical_size::StartCircumstances;
3669 0 : let (_maybe_permit, circumstances) = tokio::select! {
3670 0 : permit = wait_for_permit => {
3671 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3672 : }
3673 0 : _ = self_ref.cancel.cancelled() => {
3674 0 : return Err(CalculateLogicalSizeError::Cancelled);
3675 : }
3676 0 : () = skip_concurrency_limiter.cancelled() => {
3677 : // Some action that is part of a end user interaction requested logical size
3678 : // => break out of the rate limit
3679 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3680 : // but then again what happens if they cancel; also, we should just be using
3681 : // one runtime across the entire process, so, let's leave this for now.
3682 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3683 : }
3684 : };
3685 :
3686 0 : let metrics_guard = if attempt == 1 {
3687 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3688 : } else {
3689 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3690 : };
3691 :
3692 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3693 0 : self_ref.conf.get_vectored_concurrent_io,
3694 0 : self_ref
3695 0 : .gate
3696 0 : .enter()
3697 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3698 : );
3699 :
3700 0 : let calculated_size = self_ref
3701 0 : .logical_size_calculation_task(
3702 0 : initial_part_end,
3703 0 : LogicalSizeCalculationCause::Initial,
3704 0 : background_ctx,
3705 0 : )
3706 0 : .await?;
3707 :
3708 0 : self_ref
3709 0 : .trigger_aux_file_size_computation(
3710 0 : initial_part_end,
3711 0 : background_ctx,
3712 0 : io_concurrency,
3713 0 : )
3714 0 : .await?;
3715 :
3716 : // TODO: add aux file size to logical size
3717 :
3718 0 : Ok((calculated_size, metrics_guard))
3719 0 : }
3720 0 : };
3721 :
3722 0 : let retrying = async {
3723 0 : let mut attempt = 0;
3724 : loop {
3725 0 : attempt += 1;
3726 :
3727 0 : match try_once(attempt).await {
3728 0 : Ok(res) => return ControlFlow::Continue(res),
3729 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3730 : Err(
3731 0 : e @ (CalculateLogicalSizeError::Decode(_)
3732 : | CalculateLogicalSizeError::PageRead(_)),
3733 : ) => {
3734 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3735 : // exponential back-off doesn't make sense at these long intervals;
3736 : // use fixed retry interval with generous jitter instead
3737 0 : let sleep_duration = Duration::from_secs(
3738 0 : u64::try_from(
3739 : // 1hour base
3740 0 : (60_i64 * 60_i64)
3741 0 : // 10min jitter
3742 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3743 : )
3744 0 : .expect("10min < 1hour"),
3745 : );
3746 0 : tokio::select! {
3747 0 : _ = tokio::time::sleep(sleep_duration) => {}
3748 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3749 : }
3750 : }
3751 : }
3752 : }
3753 0 : };
3754 :
3755 0 : let (calculated_size, metrics_guard) = match retrying.await {
3756 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3757 0 : ControlFlow::Break(()) => return,
3758 : };
3759 :
3760 : // we cannot query current_logical_size.current_size() to know the current
3761 : // *negative* value, only truncated to u64.
3762 0 : let added = self
3763 0 : .current_logical_size
3764 0 : .size_added_after_initial
3765 0 : .load(AtomicOrdering::Relaxed);
3766 :
3767 0 : let sum = calculated_size.saturating_add_signed(added);
3768 :
3769 : // set the gauge value before it can be set in `update_current_logical_size`.
3770 0 : self.metrics.current_logical_size_gauge.set(sum);
3771 :
3772 0 : self.current_logical_size
3773 0 : .initial_logical_size
3774 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3775 0 : .ok()
3776 0 : .expect("only this task sets it");
3777 0 : }
3778 :
3779 7 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3780 7 : self: &Arc<Self>,
3781 7 : lsn: Lsn,
3782 7 : cause: LogicalSizeCalculationCause,
3783 7 : ctx: RequestContext,
3784 7 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3785 7 : let (sender, receiver) = oneshot::channel();
3786 7 : let self_clone = Arc::clone(self);
3787 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3788 : // we should stop the size calculation work and return an error.
3789 : // That would require restructuring this function's API to
3790 : // return the result directly, instead of a Receiver for the result.
3791 7 : let ctx = ctx.detached_child(
3792 7 : TaskKind::OndemandLogicalSizeCalculation,
3793 7 : DownloadBehavior::Download,
3794 : );
3795 7 : task_mgr::spawn(
3796 7 : task_mgr::BACKGROUND_RUNTIME.handle(),
3797 7 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3798 7 : self.tenant_shard_id,
3799 7 : Some(self.timeline_id),
3800 7 : "ondemand logical size calculation",
3801 7 : async move {
3802 7 : let res = self_clone
3803 7 : .logical_size_calculation_task(lsn, cause, &ctx)
3804 7 : .await;
3805 7 : let _ = sender.send(res).ok();
3806 7 : Ok(()) // Receiver is responsible for handling errors
3807 7 : }
3808 7 : .in_current_span(),
3809 : );
3810 7 : receiver
3811 7 : }
3812 :
3813 : #[instrument(skip_all)]
3814 : async fn logical_size_calculation_task(
3815 : self: &Arc<Self>,
3816 : lsn: Lsn,
3817 : cause: LogicalSizeCalculationCause,
3818 : ctx: &RequestContext,
3819 : ) -> Result<u64, CalculateLogicalSizeError> {
3820 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3821 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3822 : // accurate relation sizes, and they do not emit consumption metrics.
3823 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3824 :
3825 : let guard = self
3826 : .gate
3827 : .enter()
3828 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3829 :
3830 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3831 : }
3832 :
3833 : /// Calculate the logical size of the database at the latest LSN.
3834 : ///
3835 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3836 : /// especially if we need to download remote layers.
3837 7 : async fn calculate_logical_size(
3838 7 : &self,
3839 7 : up_to_lsn: Lsn,
3840 7 : cause: LogicalSizeCalculationCause,
3841 7 : _guard: &GateGuard,
3842 7 : ctx: &RequestContext,
3843 7 : ) -> Result<u64, CalculateLogicalSizeError> {
3844 7 : info!(
3845 0 : "Calculating logical size for timeline {} at {}",
3846 : self.timeline_id, up_to_lsn
3847 : );
3848 :
3849 7 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3850 : {
3851 0 : return Err(CalculateLogicalSizeError::Cancelled);
3852 7 : }
3853 :
3854 : // See if we've already done the work for initial size calculation.
3855 : // This is a short-cut for timelines that are mostly unused.
3856 7 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3857 0 : return Ok(size);
3858 7 : }
3859 7 : let storage_time_metrics = match cause {
3860 : LogicalSizeCalculationCause::Initial
3861 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3862 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3863 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3864 7 : &self.metrics.imitate_logical_size_histo
3865 : }
3866 : };
3867 7 : let timer = storage_time_metrics.start_timer();
3868 7 : let logical_size = self
3869 7 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3870 7 : .await?;
3871 7 : debug!("calculated logical size: {logical_size}");
3872 7 : timer.stop_and_record();
3873 7 : Ok(logical_size)
3874 7 : }
3875 :
3876 : /// Update current logical size, adding `delta' to the old value.
3877 135285 : fn update_current_logical_size(&self, delta: i64) {
3878 135285 : let logical_size = &self.current_logical_size;
3879 135285 : logical_size.increment_size(delta);
3880 :
3881 : // Also set the value in the prometheus gauge. Note that
3882 : // there is a race condition here: if this is is called by two
3883 : // threads concurrently, the prometheus gauge might be set to
3884 : // one value while current_logical_size is set to the
3885 : // other.
3886 135285 : match logical_size.current_size() {
3887 135285 : CurrentLogicalSize::Exact(ref new_current_size) => self
3888 135285 : .metrics
3889 135285 : .current_logical_size_gauge
3890 135285 : .set(new_current_size.into()),
3891 0 : CurrentLogicalSize::Approximate(_) => {
3892 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3893 0 : // forth between the initial size calculation task.
3894 0 : }
3895 : }
3896 135285 : }
3897 :
3898 1522 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3899 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3900 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3901 : // would overwrite the previous value if they modify different databases.
3902 :
3903 1522 : match count {
3904 561 : MetricsUpdate::Set(count) => {
3905 561 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3906 561 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3907 561 : }
3908 960 : MetricsUpdate::Add(count) => {
3909 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3910 : // it's fine.
3911 960 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3912 960 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3913 960 : // the value reliably.
3914 960 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3915 960 : }
3916 : // Otherwise, ignore this update
3917 : }
3918 1 : MetricsUpdate::Sub(count) => {
3919 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3920 : // it's fine.
3921 1 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3922 1 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3923 1 : // The operation could overflow so we need to normalize the value.
3924 1 : let prev_val =
3925 1 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3926 1 : let res = prev_val.saturating_sub(count);
3927 1 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3928 1 : }
3929 : // Otherwise, ignore this update
3930 : }
3931 : };
3932 :
3933 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3934 1522 : let aux_metric =
3935 1522 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3936 :
3937 1522 : let sum_of_entries = self
3938 1522 : .directory_metrics
3939 1522 : .iter()
3940 12176 : .map(|v| v.load(AtomicOrdering::Relaxed))
3941 1522 : .sum();
3942 : // Set a high general threshold and a lower threshold for the auxiliary files,
3943 : // as we can have large numbers of relations in the db directory.
3944 : const SUM_THRESHOLD: u64 = 5000;
3945 : const AUX_THRESHOLD: u64 = 1000;
3946 1522 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3947 0 : self.metrics
3948 0 : .directory_entries_count_gauge
3949 0 : .set(sum_of_entries);
3950 1522 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3951 0 : metric.set(sum_of_entries);
3952 1522 : }
3953 1522 : }
3954 :
3955 0 : async fn find_layer(
3956 0 : &self,
3957 0 : layer_name: &LayerName,
3958 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3959 0 : let guard = self
3960 0 : .layers
3961 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
3962 0 : .await;
3963 0 : let layer = guard
3964 0 : .layer_map()?
3965 0 : .iter_historic_layers()
3966 0 : .find(|l| &l.layer_name() == layer_name)
3967 0 : .map(|found| guard.get_from_desc(&found));
3968 0 : Ok(layer)
3969 0 : }
3970 :
3971 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
3972 0 : let crnt = self.previous_heatmap.load();
3973 0 : match crnt.as_deref() {
3974 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
3975 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
3976 0 : None => true,
3977 : },
3978 0 : Some(PreviousHeatmap::Obsolete) => false,
3979 0 : None => false,
3980 : }
3981 0 : }
3982 :
3983 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3984 : /// indicating which layers are currently on-disk on the primary.
3985 : ///
3986 : /// None is returned if the Timeline is in a state where uploading a heatmap
3987 : /// doesn't make sense, such as shutting down or initializing. The caller
3988 : /// should treat this as a cue to simply skip doing any heatmap uploading
3989 : /// for this timeline.
3990 8 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3991 8 : if !self.is_active() {
3992 0 : return None;
3993 8 : }
3994 :
3995 8 : let guard = self
3996 8 : .layers
3997 8 : .read(LayerManagerLockHolder::GenerateHeatmap)
3998 8 : .await;
3999 :
4000 : // Firstly, if there's any heatmap left over from when this location
4001 : // was a secondary, take that into account. Keep layers that are:
4002 : // * present in the layer map
4003 : // * visible
4004 : // * non-resident
4005 : // * not evicted since we read the heatmap
4006 : //
4007 : // Without this, a new cold, attached location would clobber the previous
4008 : // heatamp.
4009 8 : let previous_heatmap = self.previous_heatmap.load();
4010 8 : let visible_non_resident = match previous_heatmap.as_deref() {
4011 : Some(PreviousHeatmap::Active {
4012 6 : heatmap, read_at, ..
4013 23 : }) => Some(heatmap.all_layers().filter_map(|hl| {
4014 23 : let desc: PersistentLayerDesc = hl.name.clone().into();
4015 23 : let layer = guard.try_get_from_key(&desc.key())?;
4016 :
4017 23 : if layer.visibility() == LayerVisibilityHint::Covered {
4018 0 : return None;
4019 23 : }
4020 :
4021 23 : if layer.is_likely_resident() {
4022 10 : return None;
4023 13 : }
4024 :
4025 13 : if layer.last_evicted_at().happened_after(*read_at) {
4026 3 : return None;
4027 10 : }
4028 :
4029 10 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
4030 23 : })),
4031 0 : Some(PreviousHeatmap::Obsolete) => None,
4032 2 : None => None,
4033 : };
4034 :
4035 : // Secondly, all currently visible, resident layers are included.
4036 18 : let resident = guard.likely_resident_layers().filter_map(|layer| {
4037 18 : match layer.visibility() {
4038 : LayerVisibilityHint::Visible => {
4039 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
4040 17 : let last_activity_ts = layer.latest_activity();
4041 17 : Some((
4042 17 : layer.layer_desc().clone(),
4043 17 : layer.metadata(),
4044 17 : last_activity_ts,
4045 17 : false, // these layers are not cold
4046 17 : ))
4047 : }
4048 : LayerVisibilityHint::Covered => {
4049 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
4050 1 : None
4051 : }
4052 : }
4053 18 : });
4054 :
4055 8 : let mut layers = match visible_non_resident {
4056 6 : Some(non_resident) => {
4057 6 : let mut non_resident = non_resident.peekable();
4058 6 : if non_resident.peek().is_none() {
4059 2 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
4060 2 : self.previous_heatmap
4061 2 : .store(Some(PreviousHeatmap::Obsolete.into()));
4062 4 : }
4063 :
4064 6 : non_resident.chain(resident).collect::<Vec<_>>()
4065 : }
4066 2 : None => resident.collect::<Vec<_>>(),
4067 : };
4068 :
4069 : // Sort layers in order of which to download first. For a large set of layers to download, we
4070 : // want to prioritize those layers which are most likely to still be in the resident many minutes
4071 : // or hours later:
4072 : // - Cold layers go last for convenience when a human inspects the heatmap.
4073 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
4074 : // only exist for a few minutes before being compacted into L1s.
4075 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
4076 : // the layer is likely to be covered by an image layer during compaction.
4077 60 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
4078 60 : std::cmp::Reverse((
4079 60 : *cold,
4080 60 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
4081 60 : desc.lsn_range.end,
4082 60 : ))
4083 60 : });
4084 :
4085 8 : let layers = layers
4086 8 : .into_iter()
4087 27 : .map(|(desc, meta, atime, cold)| {
4088 27 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
4089 27 : })
4090 8 : .collect();
4091 :
4092 8 : Some(HeatMapTimeline::new(self.timeline_id, layers))
4093 8 : }
4094 :
4095 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
4096 0 : let guard = self
4097 0 : .layers
4098 0 : .read(LayerManagerLockHolder::GenerateHeatmap)
4099 0 : .await;
4100 :
4101 0 : let now = SystemTime::now();
4102 0 : let mut heatmap_layers = Vec::default();
4103 0 : for vl in guard.visible_layers() {
4104 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
4105 0 : continue;
4106 0 : }
4107 :
4108 0 : let hl = HeatMapLayer {
4109 0 : name: vl.layer_desc().layer_name(),
4110 0 : metadata: vl.metadata(),
4111 0 : access_time: now,
4112 0 : cold: true,
4113 0 : };
4114 0 : heatmap_layers.push(hl);
4115 : }
4116 :
4117 0 : tracing::info!(
4118 0 : "Generating unarchival heatmap with {} layers",
4119 0 : heatmap_layers.len()
4120 : );
4121 :
4122 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
4123 0 : PreviousHeatmap::Active {
4124 0 : heatmap,
4125 0 : read_at: Instant::now(),
4126 0 : end_lsn: Some(end_lsn),
4127 0 : }
4128 0 : }
4129 :
4130 : /// Returns true if the given lsn is or was an ancestor branchpoint.
4131 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
4132 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
4133 : // branchpoint in the value in IndexPart::lineage
4134 0 : self.ancestor_lsn == lsn
4135 0 : || (self.ancestor_lsn == Lsn::INVALID
4136 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
4137 0 : }
4138 : }
4139 :
4140 : #[derive(Clone)]
4141 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
4142 : /// In other words, a set of segments in a 2D space.
4143 : ///
4144 : /// This representation has the advatange of avoiding hash map
4145 : /// allocations for uniform queries.
4146 : pub(crate) enum VersionedKeySpaceQuery {
4147 : /// Variant for queries at a single [`Lsn`]
4148 : Uniform { keyspace: KeySpace, lsn: Lsn },
4149 : /// Variant for queries at multiple [`Lsn`]s
4150 : Scattered {
4151 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
4152 : },
4153 : }
4154 :
4155 : impl VersionedKeySpaceQuery {
4156 302203 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
4157 302203 : Self::Uniform { keyspace, lsn }
4158 302203 : }
4159 :
4160 10192 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
4161 10192 : Self::Scattered { keyspaces_at_lsn }
4162 10192 : }
4163 :
4164 : /// Returns the most recent (largest) LSN included in the query.
4165 : /// If any of the LSNs included in the query are invalid, returns
4166 : /// an error instead.
4167 624790 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
4168 624790 : match self {
4169 604406 : Self::Uniform { lsn, .. } => {
4170 604406 : if !lsn.is_valid() {
4171 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4172 604406 : }
4173 :
4174 604406 : Ok(*lsn)
4175 : }
4176 20384 : Self::Scattered { keyspaces_at_lsn } => {
4177 20384 : let mut max_lsn = None;
4178 42218 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4179 42218 : if !lsn.is_valid() {
4180 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4181 42218 : }
4182 42218 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4183 : }
4184 :
4185 20384 : if let Some(computed) = max_lsn {
4186 20384 : Ok(*computed)
4187 : } else {
4188 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4189 : }
4190 : }
4191 : }
4192 624790 : }
4193 :
4194 : /// Returns the total keyspace being queried: the result of projecting
4195 : /// everything in the key dimensions onto the key axis.
4196 323415 : fn total_keyspace(&self) -> KeySpace {
4197 323415 : match self {
4198 303031 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4199 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4200 20384 : .iter()
4201 20384 : .map(|(_lsn, keyspace)| keyspace)
4202 42218 : .fold(KeySpace::default(), |mut acc, v| {
4203 42218 : acc.merge(v);
4204 42218 : acc
4205 42218 : }),
4206 : }
4207 323415 : }
4208 :
4209 : /// Returns LSN for a specific key.
4210 : ///
4211 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4212 395566 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4213 395566 : match self {
4214 322374 : Self::Uniform { lsn, .. } => *lsn,
4215 73192 : Self::Scattered { keyspaces_at_lsn } => {
4216 73192 : keyspaces_at_lsn
4217 73192 : .iter()
4218 418434 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4219 73192 : .expect("Returned key was requested")
4220 : .0
4221 : }
4222 : }
4223 395566 : }
4224 :
4225 : /// Remove any parts of the query (segments) which overlap with the provided
4226 : /// key space (also segments).
4227 966206 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4228 966206 : match self {
4229 945822 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4230 20384 : Self::Scattered { keyspaces_at_lsn } => {
4231 20384 : let mut removed_accum = KeySpaceRandomAccum::new();
4232 42218 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4233 42218 : let removed = keyspace.remove_overlapping_with(to_remove);
4234 42218 : removed_accum.add_keyspace(removed);
4235 42218 : });
4236 :
4237 20384 : removed_accum.to_keyspace()
4238 : }
4239 : }
4240 966206 : }
4241 :
4242 738595 : fn is_empty(&self) -> bool {
4243 738595 : match self {
4244 718211 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4245 20384 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4246 20384 : .iter()
4247 31301 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4248 : }
4249 738595 : }
4250 :
4251 : /// "Lower" the query on the LSN dimension
4252 113806 : fn lower(&mut self, to: Lsn) {
4253 113806 : match self {
4254 113806 : Self::Uniform { lsn, .. } => {
4255 113806 : // If the originally requested LSN is smaller than the starting
4256 113806 : // LSN of the ancestor we are descending into, we need to respect that.
4257 113806 : // Hence the min.
4258 113806 : *lsn = std::cmp::min(*lsn, to);
4259 113806 : }
4260 0 : Self::Scattered { keyspaces_at_lsn } => {
4261 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4262 0 : *lsn = std::cmp::min(*lsn, to);
4263 0 : });
4264 : }
4265 : }
4266 113806 : }
4267 : }
4268 :
4269 : impl std::fmt::Display for VersionedKeySpaceQuery {
4270 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4271 0 : write!(f, "[")?;
4272 :
4273 0 : match self {
4274 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4275 0 : write!(f, "{keyspace} @ {lsn}")?;
4276 : }
4277 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4278 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4279 0 : write!(f, "{keyspace} @ {lsn},")?;
4280 : }
4281 : }
4282 : }
4283 :
4284 0 : write!(f, "]")
4285 0 : }
4286 : }
4287 :
4288 : impl Timeline {
4289 : #[allow(clippy::doc_lazy_continuation)]
4290 : /// Get the data needed to reconstruct all keys in the provided keyspace
4291 : ///
4292 : /// The algorithm is as follows:
4293 : /// 1. While some keys are still not done and there's a timeline to visit:
4294 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4295 : /// 2.1: Build the fringe for the current keyspace
4296 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4297 : /// intersects
4298 : /// 2.3. Pop the timeline from the fringe
4299 : /// 2.4. If the fringe is empty, go back to 1
4300 312395 : async fn get_vectored_reconstruct_data(
4301 312395 : &self,
4302 312395 : mut query: VersionedKeySpaceQuery,
4303 312395 : reconstruct_state: &mut ValuesReconstructState,
4304 312395 : ctx: &RequestContext,
4305 312395 : ) -> Result<(), GetVectoredError> {
4306 312395 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4307 :
4308 : let mut timeline_owned: Arc<Timeline>;
4309 312395 : let mut timeline = self;
4310 :
4311 312394 : let missing_keyspace = loop {
4312 426200 : if self.cancel.is_cancelled() {
4313 0 : return Err(GetVectoredError::Cancelled);
4314 426200 : }
4315 :
4316 : let TimelineVisitOutcome {
4317 426200 : completed_keyspace: completed,
4318 426200 : image_covered_keyspace,
4319 : } = {
4320 426200 : let ctx = RequestContextBuilder::from(ctx)
4321 426200 : .perf_span(|crnt_perf_span| {
4322 0 : info_span!(
4323 : target: PERF_TRACE_TARGET,
4324 0 : parent: crnt_perf_span,
4325 : "PLAN_IO_TIMELINE",
4326 : timeline = %timeline.timeline_id,
4327 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4328 : )
4329 0 : })
4330 426200 : .attached_child();
4331 :
4332 426200 : Self::get_vectored_reconstruct_data_timeline(
4333 426200 : timeline,
4334 426200 : &query,
4335 426200 : reconstruct_state,
4336 426200 : &self.cancel,
4337 426200 : &ctx,
4338 : )
4339 426200 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4340 426200 : .await?
4341 : };
4342 :
4343 426200 : query.remove_overlapping_with(&completed);
4344 :
4345 : // Do not descend into the ancestor timeline for aux files.
4346 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4347 : // stalling compaction.
4348 426200 : query.remove_overlapping_with(&KeySpace {
4349 426200 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4350 426200 : });
4351 :
4352 : // Keyspace is fully retrieved
4353 426200 : if query.is_empty() {
4354 312262 : break None;
4355 113938 : }
4356 :
4357 113938 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4358 : // Not fully retrieved but no ancestor timeline.
4359 132 : break Some(query.total_keyspace());
4360 : };
4361 :
4362 : // Now we see if there are keys covered by the image layer but does not exist in the
4363 : // image layer, which means that the key does not exist.
4364 :
4365 : // The block below will stop the vectored search if any of the keys encountered an image layer
4366 : // which did not contain a snapshot for said key. Since we have already removed all completed
4367 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4368 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4369 : // and stop the search as a result of that.
4370 113806 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4371 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4372 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4373 : // figuring out what is the inherited key range and do a fine-grained pruning.
4374 113806 : removed.remove_overlapping_with(&KeySpace {
4375 113806 : ranges: vec![SPARSE_RANGE],
4376 113806 : });
4377 113806 : if !removed.is_empty() {
4378 0 : break Some(removed);
4379 113806 : }
4380 :
4381 : // Each key range in the original query is at some point in the LSN space.
4382 : // When descending into the ancestor, lower all ranges in the LSN space
4383 : // such that new changes on the parent timeline are not visible.
4384 113806 : query.lower(timeline.ancestor_lsn);
4385 :
4386 113806 : let ctx = RequestContextBuilder::from(ctx)
4387 113806 : .perf_span(|crnt_perf_span| {
4388 0 : info_span!(
4389 : target: PERF_TRACE_TARGET,
4390 0 : parent: crnt_perf_span,
4391 : "GET_ANCESTOR",
4392 : timeline = %timeline.timeline_id,
4393 0 : ancestor = %ancestor_timeline.timeline_id,
4394 : ancestor_lsn = %timeline.ancestor_lsn
4395 : )
4396 0 : })
4397 113806 : .attached_child();
4398 :
4399 113806 : timeline_owned = timeline
4400 113806 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4401 113806 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4402 113806 : .await?;
4403 113805 : timeline = &*timeline_owned;
4404 : };
4405 :
4406 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4407 312394 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4408 132 : let mut missing_keyspace = missing_keyspace;
4409 132 : missing_keyspace.remove_overlapping_with(&KeySpace {
4410 132 : ranges: vec![SPARSE_RANGE],
4411 132 : });
4412 132 : if missing_keyspace.is_empty() {
4413 125 : None
4414 : } else {
4415 7 : Some(missing_keyspace)
4416 : }
4417 : } else {
4418 312262 : None
4419 : };
4420 :
4421 312394 : if let Some(missing_keyspace) = missing_keyspace {
4422 7 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4423 7 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4424 7 : shard: self.shard_identity.number,
4425 7 : original_hwm_lsn,
4426 7 : ancestor_lsn: Some(timeline.ancestor_lsn),
4427 7 : backtrace: None,
4428 7 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4429 7 : query: None,
4430 7 : })));
4431 312387 : }
4432 :
4433 312387 : Ok(())
4434 312395 : }
4435 :
4436 426200 : async fn get_vectored_init_fringe(
4437 426200 : &self,
4438 426200 : query: &VersionedKeySpaceQuery,
4439 426200 : ) -> Result<LayerFringe, GetVectoredError> {
4440 426200 : let mut fringe = LayerFringe::new();
4441 426200 : let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
4442 :
4443 426200 : match query {
4444 416008 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4445 : // LSNs requested by the compute or determined by the pageserver
4446 : // are inclusive. Queries to the layer map use exclusive LSNs.
4447 : // Hence, bump the value before the query - same in the other
4448 : // match arm.
4449 416008 : let cont_lsn = Lsn(lsn.0 + 1);
4450 416008 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4451 : }
4452 10192 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4453 21109 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4454 21109 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4455 21109 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4456 : }
4457 : }
4458 : }
4459 :
4460 426200 : Ok(fringe)
4461 426200 : }
4462 :
4463 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4464 : ///
4465 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4466 : /// the current keyspace. The current keyspace of the search at any given timeline
4467 : /// is the original keyspace minus all the keys that have been completed minus
4468 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4469 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4470 : ///
4471 : /// This is basically a depth-first search visitor implementation where a vertex
4472 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4473 : ///
4474 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4475 : /// and get all the required reconstruct data from the layer in one go.
4476 : ///
4477 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4478 : /// decides how to deal with these two keyspaces.
4479 426200 : async fn get_vectored_reconstruct_data_timeline(
4480 426200 : timeline: &Timeline,
4481 426200 : query: &VersionedKeySpaceQuery,
4482 426200 : reconstruct_state: &mut ValuesReconstructState,
4483 426200 : cancel: &CancellationToken,
4484 426200 : ctx: &RequestContext,
4485 426200 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4486 : // Prevent GC from progressing while visiting the current timeline.
4487 : // If we are GC-ing because a new image layer was added while traversing
4488 : // the timeline, then it will remove layers that are required for fulfilling
4489 : // the current get request (read-path cannot "look back" and notice the new
4490 : // image layer).
4491 426200 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4492 :
4493 : // See `compaction::compact_with_gc` for why we need this.
4494 426200 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4495 :
4496 : // Initialize the fringe
4497 426200 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4498 :
4499 426200 : let mut completed_keyspace = KeySpace::default();
4500 426200 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4501 :
4502 871401 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4503 445201 : if cancel.is_cancelled() {
4504 0 : return Err(GetVectoredError::Cancelled);
4505 445201 : }
4506 :
4507 445201 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4508 445201 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4509 445201 : }
4510 :
4511 : // Visit the layer and plan IOs for it
4512 445201 : let next_cont_lsn = lsn_range.start;
4513 445201 : layer_to_read
4514 445201 : .get_values_reconstruct_data(
4515 445201 : keyspace_to_read.clone(),
4516 445201 : lsn_range,
4517 445201 : reconstruct_state,
4518 445201 : ctx,
4519 445201 : )
4520 445201 : .await?;
4521 :
4522 445201 : let mut unmapped_keyspace = keyspace_to_read;
4523 445201 : let cont_lsn = next_cont_lsn;
4524 :
4525 445201 : reconstruct_state.on_layer_visited(&layer_to_read);
4526 :
4527 445201 : let (keys_done_last_step, keys_with_image_coverage) =
4528 445201 : reconstruct_state.consume_done_keys();
4529 445201 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4530 445201 : completed_keyspace.merge(&keys_done_last_step);
4531 445201 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4532 14120 : unmapped_keyspace
4533 14120 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4534 14120 : image_covered_keyspace.add_range(keys_with_image_coverage);
4535 431081 : }
4536 :
4537 : // Query the layer map for the next layers to read.
4538 : //
4539 : // Do not descent any further if the last layer we visited
4540 : // completed all keys in the keyspace it inspected. This is not
4541 : // required for correctness, but avoids visiting extra layers
4542 : // which turns out to be a perf bottleneck in some cases.
4543 445201 : if !unmapped_keyspace.is_empty() {
4544 128487 : let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
4545 128487 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4546 :
4547 : // It's safe to drop the layer map lock after planning the next round of reads.
4548 : // The fringe keeps readable handles for the layers which are safe to read even
4549 : // if layers were compacted or flushed.
4550 : //
4551 : // The more interesting consideration is: "Why is the read algorithm still correct
4552 : // if the layer map changes while it is operating?". Doing a vectored read on a
4553 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4554 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4555 : // range at *a particular point in time*. It is fine for the answer to be different
4556 : // at two different time points.
4557 128487 : drop(guard);
4558 316714 : }
4559 : }
4560 :
4561 426200 : Ok(TimelineVisitOutcome {
4562 426200 : completed_keyspace,
4563 426200 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4564 426200 : })
4565 426200 : }
4566 :
4567 113806 : async fn get_ready_ancestor_timeline(
4568 113806 : &self,
4569 113806 : ancestor: &Arc<Timeline>,
4570 113806 : ctx: &RequestContext,
4571 113806 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4572 : // It's possible that the ancestor timeline isn't active yet, or
4573 : // is active but hasn't yet caught up to the branch point. Wait
4574 : // for it.
4575 : //
4576 : // This cannot happen while the pageserver is running normally,
4577 : // because you cannot create a branch from a point that isn't
4578 : // present in the pageserver yet. However, we don't wait for the
4579 : // branch point to be uploaded to cloud storage before creating
4580 : // a branch. I.e., the branch LSN need not be remote consistent
4581 : // for the branching operation to succeed.
4582 : //
4583 : // Hence, if we try to load a tenant in such a state where
4584 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4585 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4586 : // then we will need to wait for the ancestor timeline to
4587 : // re-stream WAL up to branch_lsn before we access it.
4588 : //
4589 : // How can a tenant get in such a state?
4590 : // - ungraceful pageserver process exit
4591 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4592 : //
4593 : // NB: this could be avoided by requiring
4594 : // branch_lsn >= remote_consistent_lsn
4595 : // during branch creation.
4596 113806 : match ancestor.wait_to_become_active(ctx).await {
4597 113805 : Ok(()) => {}
4598 : Err(TimelineState::Stopping) => {
4599 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4600 0 : return Err(GetReadyAncestorError::Cancelled);
4601 : }
4602 1 : Err(state) => {
4603 1 : return Err(GetReadyAncestorError::BadState {
4604 1 : timeline_id: ancestor.timeline_id,
4605 1 : state,
4606 1 : });
4607 : }
4608 : }
4609 113805 : ancestor
4610 113805 : .wait_lsn(
4611 113805 : self.ancestor_lsn,
4612 113805 : WaitLsnWaiter::Timeline(self),
4613 113805 : WaitLsnTimeout::Default,
4614 113805 : ctx,
4615 113805 : )
4616 113805 : .await
4617 113805 : .map_err(|e| match e {
4618 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4619 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4620 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4621 0 : timeline_id: ancestor.timeline_id,
4622 0 : state,
4623 0 : },
4624 0 : })?;
4625 :
4626 113805 : Ok(ancestor.clone())
4627 113806 : }
4628 :
4629 148592 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4630 148592 : &self.shard_identity
4631 148592 : }
4632 :
4633 : #[inline(always)]
4634 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4635 0 : ShardTimelineId {
4636 0 : shard_index: ShardIndex {
4637 0 : shard_number: self.shard_identity.number,
4638 0 : shard_count: self.shard_identity.count,
4639 0 : },
4640 0 : timeline_id: self.timeline_id,
4641 0 : }
4642 0 : }
4643 :
4644 : /// Returns a non-frozen open in-memory layer for ingestion.
4645 : ///
4646 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4647 : /// this function without holding the mutex.
4648 659 : async fn get_layer_for_write(
4649 659 : &self,
4650 659 : lsn: Lsn,
4651 659 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4652 659 : ctx: &RequestContext,
4653 659 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4654 659 : let mut guard = self
4655 659 : .layers
4656 659 : .write(LayerManagerLockHolder::GetLayerForWrite)
4657 659 : .await;
4658 :
4659 659 : let last_record_lsn = self.get_last_record_lsn();
4660 659 : ensure!(
4661 659 : lsn > last_record_lsn,
4662 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4663 : lsn,
4664 : last_record_lsn,
4665 : );
4666 :
4667 659 : let layer = guard
4668 659 : .open_mut()?
4669 659 : .get_layer_for_write(
4670 659 : lsn,
4671 659 : self.conf,
4672 659 : self.timeline_id,
4673 659 : self.tenant_shard_id,
4674 659 : &self.gate,
4675 659 : &self.cancel,
4676 659 : ctx,
4677 : )
4678 659 : .await?;
4679 659 : Ok(layer)
4680 659 : }
4681 :
4682 2639558 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4683 2639558 : assert!(new_lsn.is_aligned());
4684 :
4685 2639558 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4686 2639558 : self.last_record_lsn.advance(new_lsn);
4687 2639558 : }
4688 :
4689 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4690 : ///
4691 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4692 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4693 610 : async fn freeze_inmem_layer_at(
4694 610 : &self,
4695 610 : at: Lsn,
4696 610 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4697 610 : ) -> Result<u64, FlushLayerError> {
4698 610 : let frozen = {
4699 610 : let mut guard = self
4700 610 : .layers
4701 610 : .write(LayerManagerLockHolder::TryFreezeLayer)
4702 610 : .await;
4703 610 : guard
4704 610 : .open_mut()?
4705 610 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4706 610 : .await
4707 : };
4708 :
4709 610 : if frozen {
4710 596 : let now = Instant::now();
4711 596 : *(self.last_freeze_ts.write().unwrap()) = now;
4712 596 : }
4713 :
4714 : // Increment the flush cycle counter and wake up the flush task.
4715 : // Remember the new value, so that when we listen for the flush
4716 : // to finish, we know when the flush that we initiated has
4717 : // finished, instead of some other flush that was started earlier.
4718 610 : let mut my_flush_request = 0;
4719 :
4720 610 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4721 610 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4722 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4723 610 : }
4724 :
4725 610 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4726 610 : my_flush_request = *counter + 1;
4727 610 : *counter = my_flush_request;
4728 610 : *lsn = std::cmp::max(at, *lsn);
4729 610 : });
4730 :
4731 610 : assert_ne!(my_flush_request, 0);
4732 :
4733 610 : Ok(my_flush_request)
4734 610 : }
4735 :
4736 : /// Layer flusher task's main loop.
4737 231 : async fn flush_loop(
4738 231 : self: &Arc<Self>,
4739 231 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4740 231 : ctx: &RequestContext,
4741 231 : ) {
4742 : // Always notify waiters about the flush loop exiting since the loop might stop
4743 : // when the timeline hasn't been cancelled.
4744 231 : let scopeguard_rx = layer_flush_start_rx.clone();
4745 231 : scopeguard::defer! {
4746 : let (flush_counter, _) = *scopeguard_rx.borrow();
4747 : let _ = self
4748 : .layer_flush_done_tx
4749 : .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
4750 : }
4751 :
4752 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4753 231 : let mut watch_l0 = match self
4754 231 : .layers
4755 231 : .read(LayerManagerLockHolder::FlushLoop)
4756 231 : .await
4757 231 : .layer_map()
4758 : {
4759 231 : Ok(lm) => lm.watch_level0_deltas(),
4760 0 : Err(Shutdown) => return,
4761 : };
4762 :
4763 231 : info!("started flush loop");
4764 : loop {
4765 838 : tokio::select! {
4766 838 : _ = self.cancel.cancelled() => {
4767 5 : info!("shutting down layer flush task due to Timeline::cancel");
4768 5 : break;
4769 : },
4770 838 : _ = layer_flush_start_rx.changed() => {}
4771 : }
4772 607 : trace!("waking up");
4773 607 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4774 :
4775 : // The highest LSN to which we flushed in the loop over frozen layers
4776 607 : let mut flushed_to_lsn = Lsn(0);
4777 :
4778 607 : let result = loop {
4779 1202 : if self.cancel.is_cancelled() {
4780 0 : info!("dropping out of flush loop for timeline shutdown");
4781 0 : return;
4782 1202 : }
4783 :
4784 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4785 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4786 1202 : if flushed_to_lsn >= frozen_to_lsn {
4787 592 : break Ok(());
4788 610 : }
4789 :
4790 : // Fetch the next layer to flush, if any.
4791 610 : let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
4792 610 : let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
4793 610 : let Ok(lm) = layers.layer_map() else {
4794 0 : info!("dropping out of flush loop for timeline shutdown");
4795 0 : return;
4796 : };
4797 610 : let l0_count = lm.level0_deltas().len();
4798 610 : let frozen_count = lm.frozen_layers.len();
4799 610 : let frozen_size: u64 = lm
4800 610 : .frozen_layers
4801 610 : .iter()
4802 610 : .map(|l| l.estimated_in_mem_size())
4803 610 : .sum();
4804 610 : let open_layer_size: u64 = lm
4805 610 : .open_layer
4806 610 : .as_ref()
4807 610 : .map(|l| l.estimated_in_mem_size())
4808 610 : .unwrap_or(0);
4809 610 : let layer = lm.frozen_layers.front().cloned();
4810 610 : (layer, l0_count, frozen_count, frozen_size, open_layer_size)
4811 : // drop 'layers' lock
4812 : };
4813 610 : let Some(layer) = layer else {
4814 14 : break Ok(());
4815 : };
4816 :
4817 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4818 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4819 596 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4820 0 : if l0_count >= stall_threshold {
4821 0 : warn!(
4822 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4823 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4824 : );
4825 0 : let stall_timer = self
4826 0 : .metrics
4827 0 : .flush_delay_histo
4828 0 : .start_timer()
4829 0 : .record_on_drop();
4830 0 : tokio::select! {
4831 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4832 0 : if let Ok(l0) = result.as_deref() {
4833 0 : let delay = stall_timer.elapsed().as_secs_f64();
4834 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4835 0 : }
4836 : },
4837 0 : _ = self.cancel.cancelled() => {},
4838 : }
4839 0 : continue; // check again
4840 0 : }
4841 596 : }
4842 :
4843 : // Flush the layer.
4844 596 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4845 596 : match self.flush_frozen_layer(layer, ctx).await {
4846 595 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4847 : Err(FlushLayerError::Cancelled) => {
4848 0 : info!("dropping out of flush loop for timeline shutdown");
4849 0 : return;
4850 : }
4851 1 : err @ Err(
4852 : FlushLayerError::NotRunning(_)
4853 : | FlushLayerError::Other(_)
4854 : | FlushLayerError::CreateImageLayersError(_),
4855 : ) => {
4856 1 : error!("could not flush frozen layer: {err:?}");
4857 1 : break err.map(|_| ());
4858 : }
4859 : }
4860 595 : let flush_duration = flush_timer.stop_and_record();
4861 :
4862 : // Notify the tenant compaction loop if L0 compaction is needed.
4863 595 : let l0_count = *watch_l0.borrow();
4864 595 : if l0_count >= self.get_compaction_threshold() {
4865 238 : self.l0_compaction_trigger.notify_one();
4866 357 : }
4867 :
4868 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4869 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4870 : // ingestion by having ephemeral layer rolls wait for flushes.
4871 595 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4872 3 : if l0_count >= delay_threshold {
4873 0 : let delay = flush_duration.as_secs_f64();
4874 0 : info!(
4875 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4876 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
4877 : );
4878 0 : let _delay_timer = self
4879 0 : .metrics
4880 0 : .flush_delay_histo
4881 0 : .start_timer()
4882 0 : .record_on_drop();
4883 0 : tokio::select! {
4884 0 : _ = tokio::time::sleep(flush_duration) => {},
4885 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4886 0 : _ = self.cancel.cancelled() => {},
4887 : }
4888 3 : }
4889 592 : }
4890 : };
4891 :
4892 : // Unsharded tenants should never advance their LSN beyond the end of the
4893 : // highest layer they write: such gaps between layer data and the frozen LSN
4894 : // are only legal on sharded tenants.
4895 607 : debug_assert!(
4896 607 : self.shard_identity.count.count() > 1
4897 600 : || flushed_to_lsn >= frozen_to_lsn
4898 14 : || !flushed_to_lsn.is_valid()
4899 : );
4900 :
4901 607 : if flushed_to_lsn < frozen_to_lsn
4902 15 : && self.shard_identity.count.count() > 1
4903 1 : && result.is_ok()
4904 : {
4905 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4906 : // to us via layer_flush_start_rx, then advance it here.
4907 : //
4908 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4909 : // never encounter a gap in the wal.
4910 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4911 0 : tracing::debug!(
4912 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
4913 : );
4914 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4915 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4916 0 : tracing::warn!(
4917 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
4918 : );
4919 0 : }
4920 0 : }
4921 607 : }
4922 :
4923 : // Notify any listeners that we're done
4924 607 : let _ = self
4925 607 : .layer_flush_done_tx
4926 607 : .send_replace((flush_counter, result));
4927 : }
4928 5 : }
4929 :
4930 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4931 570 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4932 570 : let mut rx = self.layer_flush_done_tx.subscribe();
4933 : loop {
4934 : {
4935 1153 : let (last_result_counter, last_result) = &*rx.borrow();
4936 1153 : if *last_result_counter >= request {
4937 570 : if let Err(err) = last_result {
4938 : // We already logged the original error in
4939 : // flush_loop. We cannot propagate it to the caller
4940 : // here, because it might not be Cloneable
4941 1 : return Err(err.clone());
4942 : } else {
4943 569 : return Ok(());
4944 : }
4945 583 : }
4946 : }
4947 583 : trace!("waiting for flush to complete");
4948 583 : tokio::select! {
4949 583 : rx_e = rx.changed() => {
4950 583 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4951 : },
4952 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4953 : // the notification from [`flush_loop`] that it completed.
4954 583 : _ = self.cancel.cancelled() => {
4955 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4956 0 : return Ok(())
4957 : }
4958 : };
4959 583 : trace!("done")
4960 : }
4961 570 : }
4962 :
4963 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4964 : ///
4965 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4966 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4967 : async fn flush_frozen_layer(
4968 : self: &Arc<Self>,
4969 : frozen_layer: Arc<InMemoryLayer>,
4970 : ctx: &RequestContext,
4971 : ) -> Result<Lsn, FlushLayerError> {
4972 : debug_assert_current_span_has_tenant_and_timeline_id();
4973 :
4974 : // As a special case, when we have just imported an image into the repository,
4975 : // instead of writing out a L0 delta layer, we directly write out image layer
4976 : // files instead. This is possible as long as *all* the data imported into the
4977 : // repository have the same LSN.
4978 : let lsn_range = frozen_layer.get_lsn_range();
4979 :
4980 : // Whether to directly create image layers for this flush, or flush them as delta layers
4981 : let create_image_layer =
4982 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4983 :
4984 : #[cfg(test)]
4985 : {
4986 : match &mut *self.flush_loop_state.lock().unwrap() {
4987 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4988 : panic!("flush loop not running")
4989 : }
4990 : FlushLoopState::Running {
4991 : expect_initdb_optimization,
4992 : initdb_optimization_count,
4993 : ..
4994 : } => {
4995 : if create_image_layer {
4996 : *initdb_optimization_count += 1;
4997 : } else {
4998 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4999 : }
5000 : }
5001 : }
5002 : }
5003 :
5004 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
5005 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
5006 : // require downloading anything during initial import.
5007 : let ((rel_partition, metadata_partition), _lsn) = self
5008 : .repartition(
5009 : self.initdb_lsn,
5010 : self.get_compaction_target_size(),
5011 : EnumSet::empty(),
5012 : ctx,
5013 : )
5014 : .await
5015 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
5016 :
5017 : if self.cancel.is_cancelled() {
5018 : return Err(FlushLayerError::Cancelled);
5019 : }
5020 :
5021 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
5022 : // So that the key ranges don't overlap.
5023 : let mut partitions = KeyPartitioning::default();
5024 : partitions.parts.extend(rel_partition.parts);
5025 : if !metadata_partition.parts.is_empty() {
5026 : assert_eq!(
5027 : metadata_partition.parts.len(),
5028 : 1,
5029 : "currently sparse keyspace should only contain a single metadata keyspace"
5030 : );
5031 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
5032 : // every single key within the keyspace, and therefore, it's safe to force converting it
5033 : // into a dense keyspace before calling this function.
5034 : partitions
5035 : .parts
5036 : .extend(metadata_partition.into_dense().parts);
5037 : }
5038 :
5039 : let mut layers_to_upload = Vec::new();
5040 : let (generated_image_layers, is_complete) = self
5041 : .create_image_layers(
5042 : &partitions,
5043 : self.initdb_lsn,
5044 : ImageLayerCreationMode::Initial,
5045 : ctx,
5046 : LastImageLayerCreationStatus::Initial,
5047 : false, // don't yield for L0, we're flushing L0
5048 : )
5049 : .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
5050 : .await?;
5051 : debug_assert!(
5052 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
5053 : "init image generation mode must fully cover the keyspace"
5054 : );
5055 : layers_to_upload.extend(generated_image_layers);
5056 :
5057 : (layers_to_upload, None)
5058 : } else {
5059 : // Normal case, write out a L0 delta layer file.
5060 : // `create_delta_layer` will not modify the layer map.
5061 : // We will remove frozen layer and add delta layer in one atomic operation later.
5062 : let Some(layer) = self
5063 : .create_delta_layer(&frozen_layer, None, ctx)
5064 : .await
5065 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
5066 : else {
5067 : panic!("delta layer cannot be empty if no filter is applied");
5068 : };
5069 : (
5070 : // FIXME: even though we have a single image and single delta layer assumption
5071 : // we push them to vec
5072 : vec![layer.clone()],
5073 : Some(layer),
5074 : )
5075 : };
5076 :
5077 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
5078 :
5079 : if self.cancel.is_cancelled() {
5080 : return Err(FlushLayerError::Cancelled);
5081 : }
5082 :
5083 1 : fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
5084 1 : Err(FlushLayerError::Other(anyhow!("failpoint").into()))
5085 1 : });
5086 :
5087 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
5088 :
5089 : // The new on-disk layers are now in the layer map. We can remove the
5090 : // in-memory layer from the map now. The flushed layer is stored in
5091 : // the mapping in `create_delta_layer`.
5092 : {
5093 : let mut guard = self
5094 : .layers
5095 : .write(LayerManagerLockHolder::FlushFrozenLayer)
5096 : .await;
5097 :
5098 : guard.open_mut()?.finish_flush_l0_layer(
5099 : delta_layer_to_add.as_ref(),
5100 : &frozen_layer,
5101 : &self.metrics,
5102 : );
5103 :
5104 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
5105 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
5106 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
5107 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
5108 : }
5109 : // release lock on 'layers'
5110 : };
5111 :
5112 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
5113 : // a compaction can delete the file and then it won't be available for uploads any more.
5114 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
5115 : // race situation.
5116 : // See https://github.com/neondatabase/neon/issues/4526
5117 : pausable_failpoint!("flush-frozen-pausable");
5118 :
5119 : // This failpoint is used by another test case `test_pageserver_recovery`.
5120 : fail_point!("flush-frozen-exit");
5121 :
5122 : Ok(Lsn(lsn_range.end.0 - 1))
5123 : }
5124 :
5125 : /// Return true if the value changed
5126 : ///
5127 : /// This function must only be used from the layer flush task.
5128 595 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
5129 595 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
5130 595 : assert!(
5131 595 : new_value >= old_value,
5132 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
5133 : );
5134 :
5135 595 : self.metrics
5136 595 : .disk_consistent_lsn_gauge
5137 595 : .set(new_value.0 as i64);
5138 595 : new_value != old_value
5139 595 : }
5140 :
5141 : /// Update metadata file
5142 622 : fn schedule_uploads(
5143 622 : &self,
5144 622 : disk_consistent_lsn: Lsn,
5145 622 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
5146 622 : ) -> anyhow::Result<()> {
5147 : // We can only save a valid 'prev_record_lsn' value on disk if we
5148 : // flushed *all* in-memory changes to disk. We only track
5149 : // 'prev_record_lsn' in memory for the latest processed record, so we
5150 : // don't remember what the correct value that corresponds to some old
5151 : // LSN is. But if we flush everything, then the value corresponding
5152 : // current 'last_record_lsn' is correct and we can store it on disk.
5153 : let RecordLsn {
5154 622 : last: last_record_lsn,
5155 622 : prev: prev_record_lsn,
5156 622 : } = self.last_record_lsn.load();
5157 622 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
5158 558 : Some(prev_record_lsn)
5159 : } else {
5160 64 : None
5161 : };
5162 :
5163 622 : let update = crate::tenant::metadata::MetadataUpdate::new(
5164 622 : disk_consistent_lsn,
5165 622 : ondisk_prev_record_lsn,
5166 622 : *self.applied_gc_cutoff_lsn.read(),
5167 : );
5168 :
5169 622 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
5170 0 : "{}",
5171 0 : x.unwrap()
5172 : ));
5173 :
5174 1223 : for layer in layers_to_upload {
5175 601 : self.remote_client.schedule_layer_file_upload(layer)?;
5176 : }
5177 622 : self.remote_client
5178 622 : .schedule_index_upload_for_metadata_update(&update)?;
5179 :
5180 622 : Ok(())
5181 622 : }
5182 :
5183 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
5184 0 : self.remote_client
5185 0 : .preserve_initdb_archive(
5186 0 : &self.tenant_shard_id.tenant_id,
5187 0 : &self.timeline_id,
5188 0 : &self.cancel,
5189 0 : )
5190 0 : .await
5191 0 : }
5192 :
5193 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
5194 : // in layer map immediately. The caller is responsible to put it into the layer map.
5195 486 : async fn create_delta_layer(
5196 486 : self: &Arc<Self>,
5197 486 : frozen_layer: &Arc<InMemoryLayer>,
5198 486 : key_range: Option<Range<Key>>,
5199 486 : ctx: &RequestContext,
5200 486 : ) -> anyhow::Result<Option<ResidentLayer>> {
5201 486 : let self_clone = Arc::clone(self);
5202 486 : let frozen_layer = Arc::clone(frozen_layer);
5203 486 : let ctx = ctx.attached_child();
5204 486 : let work = async move {
5205 486 : let Some((desc, path)) = frozen_layer
5206 486 : .write_to_disk(
5207 486 : &ctx,
5208 486 : key_range,
5209 486 : self_clone.l0_flush_global_state.inner(),
5210 486 : &self_clone.gate,
5211 486 : self_clone.cancel.clone(),
5212 486 : )
5213 486 : .await?
5214 : else {
5215 0 : return Ok(None);
5216 : };
5217 486 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5218 :
5219 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5220 : // We just need to fsync the directory in which these inodes are linked,
5221 : // which we know to be the timeline directory.
5222 : //
5223 : // We use fatal_err() below because the after write_to_disk returns with success,
5224 : // the in-memory state of the filesystem already has the layer file in its final place,
5225 : // and subsequent pageserver code could think it's durable while it really isn't.
5226 486 : let timeline_dir = VirtualFile::open(
5227 486 : &self_clone
5228 486 : .conf
5229 486 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5230 486 : &ctx,
5231 486 : )
5232 486 : .await
5233 486 : .fatal_err("VirtualFile::open for timeline dir fsync");
5234 486 : timeline_dir
5235 486 : .sync_all()
5236 486 : .await
5237 486 : .fatal_err("VirtualFile::sync_all timeline dir");
5238 486 : anyhow::Ok(Some(new_delta))
5239 486 : };
5240 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5241 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5242 : use crate::virtual_file::io_engine::IoEngine;
5243 486 : match crate::virtual_file::io_engine::get() {
5244 0 : IoEngine::NotSet => panic!("io engine not set"),
5245 : IoEngine::StdFs => {
5246 0 : let span = tracing::info_span!("blocking");
5247 0 : tokio::task::spawn_blocking({
5248 0 : move || Handle::current().block_on(work.instrument(span))
5249 : })
5250 0 : .await
5251 0 : .context("spawn_blocking")
5252 0 : .and_then(|x| x)
5253 : }
5254 : #[cfg(target_os = "linux")]
5255 486 : IoEngine::TokioEpollUring => work.await,
5256 : }
5257 486 : }
5258 :
5259 302 : async fn repartition(
5260 302 : &self,
5261 302 : lsn: Lsn,
5262 302 : partition_size: u64,
5263 302 : flags: EnumSet<CompactFlags>,
5264 302 : ctx: &RequestContext,
5265 302 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
5266 302 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5267 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5268 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5269 : // and hence before the compaction task starts.
5270 0 : return Err(RepartitionError::Other(anyhow!(
5271 0 : "repartition() called concurrently"
5272 0 : )));
5273 : };
5274 302 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5275 302 : if lsn < *partition_lsn {
5276 0 : return Err(RepartitionError::Other(anyhow!(
5277 0 : "repartition() called with LSN going backwards, this should not happen"
5278 0 : )));
5279 302 : }
5280 :
5281 302 : let distance = lsn.0 - partition_lsn.0;
5282 302 : if *partition_lsn != Lsn(0)
5283 141 : && distance <= self.repartition_threshold
5284 141 : && !flags.contains(CompactFlags::ForceRepartition)
5285 : {
5286 134 : debug!(
5287 : distance,
5288 : threshold = self.repartition_threshold,
5289 0 : "no repartitioning needed"
5290 : );
5291 134 : return Ok((
5292 134 : (dense_partition.clone(), sparse_partition.clone()),
5293 134 : *partition_lsn,
5294 134 : ));
5295 168 : }
5296 :
5297 168 : let (dense_ks, sparse_ks) = self
5298 168 : .collect_keyspace(lsn, ctx)
5299 168 : .await
5300 168 : .map_err(RepartitionError::CollectKeyspace)?;
5301 168 : let dense_partitioning = dense_ks.partition(
5302 168 : &self.shard_identity,
5303 168 : partition_size,
5304 168 : postgres_ffi::BLCKSZ as u64,
5305 : );
5306 168 : let sparse_partitioning = SparseKeyPartitioning {
5307 168 : parts: vec![sparse_ks],
5308 168 : }; // no partitioning for metadata keys for now
5309 168 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5310 168 : guard.write(result.clone());
5311 168 : Ok(result)
5312 302 : }
5313 :
5314 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5315 7 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
5316 7 : let threshold = self.get_image_creation_threshold();
5317 :
5318 7 : let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
5319 7 : let Ok(layers) = guard.layer_map() else {
5320 0 : return false;
5321 : };
5322 :
5323 7 : let mut max_deltas = 0;
5324 14 : for part_range in &partition.ranges {
5325 7 : let image_coverage = layers.image_coverage(part_range, lsn);
5326 14 : for (img_range, last_img) in image_coverage {
5327 7 : let img_lsn = if let Some(last_img) = last_img {
5328 0 : last_img.get_lsn_range().end
5329 : } else {
5330 7 : Lsn(0)
5331 : };
5332 : // Let's consider an example:
5333 : //
5334 : // delta layer with LSN range 71-81
5335 : // delta layer with LSN range 81-91
5336 : // delta layer with LSN range 91-101
5337 : // image layer at LSN 100
5338 : //
5339 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5340 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5341 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5342 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5343 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5344 7 : if img_lsn < lsn {
5345 7 : let num_deltas =
5346 7 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5347 :
5348 7 : max_deltas = max_deltas.max(num_deltas);
5349 7 : if num_deltas >= threshold {
5350 0 : debug!(
5351 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5352 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5353 : );
5354 0 : return true;
5355 7 : }
5356 0 : }
5357 : }
5358 : }
5359 :
5360 7 : debug!(
5361 : max_deltas,
5362 0 : "none of the partitioned ranges had >= {threshold} deltas"
5363 : );
5364 7 : false
5365 7 : }
5366 :
5367 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5368 : /// so that at most one image layer will be produced from this function.
5369 : #[allow(clippy::too_many_arguments)]
5370 123 : async fn create_image_layer_for_rel_blocks(
5371 123 : self: &Arc<Self>,
5372 123 : partition: &KeySpace,
5373 123 : mut image_layer_writer: ImageLayerWriter,
5374 123 : lsn: Lsn,
5375 123 : ctx: &RequestContext,
5376 123 : img_range: Range<Key>,
5377 123 : io_concurrency: IoConcurrency,
5378 123 : progress: Option<(usize, usize)>,
5379 123 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5380 123 : let mut wrote_keys = false;
5381 :
5382 123 : let mut key_request_accum = KeySpaceAccum::new();
5383 817 : for range in &partition.ranges {
5384 694 : let mut key = range.start;
5385 1504 : while key < range.end {
5386 : // Decide whether to retain this key: usually we do, but sharded tenants may
5387 : // need to drop keys that don't belong to them. If we retain the key, add it
5388 : // to `key_request_accum` for later issuing a vectored get
5389 810 : if self.shard_identity.is_key_disposable(&key) {
5390 0 : debug!(
5391 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5392 : key,
5393 0 : self.shard_identity.get_shard_number(&key)
5394 : );
5395 810 : } else {
5396 810 : key_request_accum.add_key(key);
5397 810 : }
5398 :
5399 810 : let last_key_in_range = key.next() == range.end;
5400 810 : key = key.next();
5401 :
5402 : // Maybe flush `key_rest_accum`
5403 810 : if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
5404 810 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5405 : {
5406 694 : let query =
5407 694 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5408 :
5409 694 : let results = self
5410 694 : .get_vectored(query, io_concurrency.clone(), ctx)
5411 694 : .await?;
5412 :
5413 694 : if self.cancel.is_cancelled() {
5414 0 : return Err(CreateImageLayersError::Cancelled);
5415 694 : }
5416 :
5417 1504 : for (img_key, img) in results {
5418 810 : let img = match img {
5419 810 : Ok(img) => img,
5420 0 : Err(err) => {
5421 : // If we fail to reconstruct a VM or FSM page, we can zero the
5422 : // page without losing any actual user data. That seems better
5423 : // than failing repeatedly and getting stuck.
5424 : //
5425 : // We had a bug at one point, where we truncated the FSM and VM
5426 : // in the pageserver, but the Postgres didn't know about that
5427 : // and continued to generate incremental WAL records for pages
5428 : // that didn't exist in the pageserver. Trying to replay those
5429 : // WAL records failed to find the previous image of the page.
5430 : // This special case allows us to recover from that situation.
5431 : // See https://github.com/neondatabase/neon/issues/2601.
5432 : //
5433 : // Unfortunately we cannot do this for the main fork, or for
5434 : // any metadata keys, keys, as that would lead to actual data
5435 : // loss.
5436 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5437 0 : warn!(
5438 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5439 : );
5440 0 : ZERO_PAGE.clone()
5441 : } else {
5442 0 : return Err(CreateImageLayersError::from(err));
5443 : }
5444 : }
5445 : };
5446 :
5447 : // Write all the keys we just read into our new image layer.
5448 810 : image_layer_writer.put_image(img_key, img, ctx).await?;
5449 810 : wrote_keys = true;
5450 : }
5451 116 : }
5452 : }
5453 : }
5454 :
5455 123 : let progress_report = progress
5456 123 : .map(|(idx, total)| format!("({idx}/{total}) "))
5457 123 : .unwrap_or_default();
5458 123 : if wrote_keys {
5459 : // Normal path: we have written some data into the new image layer for this
5460 : // partition, so flush it to disk.
5461 123 : info!(
5462 0 : "{} produced image layer for rel {}",
5463 : progress_report,
5464 0 : ImageLayerName {
5465 0 : key_range: img_range.clone(),
5466 0 : lsn
5467 0 : },
5468 : );
5469 123 : Ok(ImageLayerCreationOutcome::Generated {
5470 123 : unfinished_image_layer: image_layer_writer,
5471 123 : })
5472 : } else {
5473 0 : tracing::debug!(
5474 0 : "{} no data in range {}-{}",
5475 : progress_report,
5476 : img_range.start,
5477 : img_range.end
5478 : );
5479 0 : Ok(ImageLayerCreationOutcome::Empty)
5480 : }
5481 123 : }
5482 :
5483 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5484 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5485 : /// would not be too large to fit in a single image layer.
5486 : ///
5487 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5488 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5489 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5490 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5491 : #[allow(clippy::too_many_arguments)]
5492 118 : async fn create_image_layer_for_metadata_keys(
5493 118 : self: &Arc<Self>,
5494 118 : partition: &KeySpace,
5495 118 : mut image_layer_writer: ImageLayerWriter,
5496 118 : lsn: Lsn,
5497 118 : ctx: &RequestContext,
5498 118 : img_range: Range<Key>,
5499 118 : mode: ImageLayerCreationMode,
5500 118 : io_concurrency: IoConcurrency,
5501 118 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5502 : // Metadata keys image layer creation.
5503 118 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5504 118 : let begin = Instant::now();
5505 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5506 : // not contain too many keys, otherwise this takes a lot of memory.
5507 118 : let data = self
5508 118 : .get_vectored_impl(
5509 118 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5510 118 : &mut reconstruct_state,
5511 118 : ctx,
5512 118 : )
5513 118 : .await?;
5514 118 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5515 118 : let mut new_data = BTreeMap::new();
5516 118 : let mut total_kb_retrieved = 0;
5517 118 : let mut total_keys_retrieved = 0;
5518 5124 : for (k, v) in data {
5519 5006 : let v = v?;
5520 5006 : total_kb_retrieved += KEY_SIZE + v.len();
5521 5006 : total_keys_retrieved += 1;
5522 5006 : new_data.insert(k, v);
5523 : }
5524 118 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5525 : };
5526 118 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5527 118 : let elapsed = begin.elapsed();
5528 :
5529 118 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5530 118 : info!(
5531 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",
5532 0 : elapsed.as_secs_f64()
5533 : );
5534 :
5535 118 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5536 1 : return Ok(ImageLayerCreationOutcome::Skip);
5537 117 : }
5538 117 : if self.cancel.is_cancelled() {
5539 0 : return Err(CreateImageLayersError::Cancelled);
5540 117 : }
5541 117 : let mut wrote_any_image = false;
5542 5123 : for (k, v) in data {
5543 5006 : if v.is_empty() {
5544 : // the key has been deleted, it does not need an image
5545 : // in metadata keyspace, an empty image == tombstone
5546 4 : continue;
5547 5002 : }
5548 5002 : wrote_any_image = true;
5549 :
5550 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5551 :
5552 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5553 : // on the normal data path either.
5554 5002 : image_layer_writer.put_image(k, v, ctx).await?;
5555 : }
5556 :
5557 117 : if wrote_any_image {
5558 : // Normal path: we have written some data into the new image layer for this
5559 : // partition, so flush it to disk.
5560 6 : info!(
5561 0 : "created image layer for metadata {}",
5562 0 : ImageLayerName {
5563 0 : key_range: img_range.clone(),
5564 0 : lsn
5565 0 : }
5566 : );
5567 6 : Ok(ImageLayerCreationOutcome::Generated {
5568 6 : unfinished_image_layer: image_layer_writer,
5569 6 : })
5570 : } else {
5571 111 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5572 111 : Ok(ImageLayerCreationOutcome::Empty)
5573 : }
5574 118 : }
5575 :
5576 : /// Predicate function which indicates whether we should check if new image layers
5577 : /// are required. Since checking if new image layers are required is expensive in
5578 : /// terms of CPU, we only do it in the following cases:
5579 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
5580 : /// 2. If enough time has passed since the last check:
5581 : /// 1. For large tenants, we wish to perform the check more often since they
5582 : /// suffer from the lack of image layers
5583 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5584 190 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5585 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
5586 :
5587 190 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5588 190 : let distance = lsn
5589 190 : .checked_sub(last_checks_at)
5590 190 : .expect("Attempt to compact with LSN going backwards");
5591 190 : let min_distance =
5592 190 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5593 :
5594 190 : let distance_based_decision = distance.0 >= min_distance;
5595 :
5596 190 : let mut time_based_decision = false;
5597 190 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5598 190 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
5599 139 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
5600 : {
5601 0 : self.get_checkpoint_timeout()
5602 : } else {
5603 139 : Duration::from_secs(3600 * 48)
5604 : };
5605 :
5606 139 : time_based_decision = match *last_check_instant {
5607 29 : Some(last_check) => {
5608 29 : let elapsed = last_check.elapsed();
5609 29 : elapsed >= check_required_after
5610 : }
5611 110 : None => true,
5612 : };
5613 51 : }
5614 :
5615 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5616 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5617 : // check.
5618 190 : let decision = distance_based_decision || time_based_decision;
5619 :
5620 190 : if decision {
5621 111 : self.last_image_layer_creation_check_at.store(lsn);
5622 111 : *last_check_instant = Some(Instant::now());
5623 111 : }
5624 :
5625 190 : decision
5626 190 : }
5627 :
5628 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5629 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5630 : ///
5631 : /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
5632 190 : async fn create_image_layers(
5633 190 : self: &Arc<Timeline>,
5634 190 : partitioning: &KeyPartitioning,
5635 190 : lsn: Lsn,
5636 190 : mode: ImageLayerCreationMode,
5637 190 : ctx: &RequestContext,
5638 190 : last_status: LastImageLayerCreationStatus,
5639 190 : yield_for_l0: bool,
5640 190 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5641 190 : let timer = self.metrics.create_images_time_histo.start_timer();
5642 :
5643 190 : if partitioning.parts.is_empty() {
5644 0 : warn!("no partitions to create image layers for");
5645 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5646 190 : }
5647 :
5648 : // We need to avoid holes between generated image layers.
5649 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5650 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5651 : //
5652 : // How such hole between partitions can appear?
5653 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5654 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5655 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5656 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5657 190 : let mut start = Key::MIN;
5658 :
5659 190 : let check_for_image_layers =
5660 190 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5661 0 : info!(
5662 0 : "resuming image layer creation: last_status=incomplete, continue from {}",
5663 : last_key
5664 : );
5665 0 : true
5666 : } else {
5667 190 : self.should_check_if_image_layers_required(lsn)
5668 : };
5669 :
5670 190 : let mut batch_image_writer = BatchLayerWriter::new(self.conf);
5671 :
5672 190 : let mut all_generated = true;
5673 :
5674 190 : let mut partition_processed = 0;
5675 190 : let mut total_partitions = partitioning.parts.len();
5676 190 : let mut last_partition_processed = None;
5677 190 : let mut partition_parts = partitioning.parts.clone();
5678 :
5679 190 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5680 : // We need to skip the partitions that have already been processed.
5681 0 : let mut found = false;
5682 0 : for (i, partition) in partition_parts.iter().enumerate() {
5683 0 : if last_key <= partition.end().unwrap() {
5684 : // ```plain
5685 : // |------|--------|----------|------|
5686 : // ^last_key
5687 : // ^start from this partition
5688 : // ```
5689 : // Why `i+1` instead of `i`?
5690 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5691 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5692 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5693 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5694 0 : if i + 1 >= total_partitions {
5695 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5696 : // iteration of image layer creation should return a complete status.
5697 0 : break; // with found=false
5698 0 : }
5699 0 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5700 0 : total_partitions = partition_parts.len();
5701 : // Update the start key to the partition start.
5702 0 : start = partition_parts[0].start().unwrap();
5703 0 : found = true;
5704 0 : break;
5705 0 : }
5706 : }
5707 0 : if !found {
5708 : // Last key is within the last partition, or larger than all partitions.
5709 0 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5710 0 : }
5711 190 : }
5712 :
5713 190 : let total = partition_parts.len();
5714 392 : for (idx, partition) in partition_parts.iter().enumerate() {
5715 392 : if self.cancel.is_cancelled() {
5716 0 : return Err(CreateImageLayersError::Cancelled);
5717 392 : }
5718 392 : partition_processed += 1;
5719 392 : let img_range = start..partition.ranges.last().unwrap().end;
5720 392 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5721 392 : if compact_metadata {
5722 760 : for range in &partition.ranges {
5723 570 : assert!(
5724 570 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5725 570 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5726 0 : "metadata keys must be partitioned separately"
5727 : );
5728 : }
5729 190 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5730 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5731 : // might mess up with evictions.
5732 72 : start = img_range.end;
5733 72 : continue;
5734 118 : }
5735 : // For initial and force modes, we always generate image layers for metadata keys.
5736 202 : } else if let ImageLayerCreationMode::Try = mode {
5737 : // check_for_image_layers = false -> skip
5738 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5739 79 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
5740 79 : start = img_range.end;
5741 79 : continue;
5742 0 : }
5743 123 : }
5744 241 : if let ImageLayerCreationMode::Force = mode {
5745 : // When forced to create image layers, we might try and create them where they already
5746 : // exist. This mode is only used in tests/debug.
5747 14 : let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
5748 14 : if layers.contains_key(&PersistentLayerKey {
5749 14 : key_range: img_range.clone(),
5750 14 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5751 14 : is_delta: false,
5752 14 : }) {
5753 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5754 : // in the future.
5755 0 : tracing::info!(
5756 0 : "Skipping image layer at {lsn} {}..{}, already exists",
5757 : img_range.start,
5758 : img_range.end
5759 : );
5760 0 : start = img_range.end;
5761 0 : continue;
5762 14 : }
5763 227 : }
5764 :
5765 241 : let image_layer_writer = ImageLayerWriter::new(
5766 241 : self.conf,
5767 241 : self.timeline_id,
5768 241 : self.tenant_shard_id,
5769 241 : &img_range,
5770 241 : lsn,
5771 241 : &self.gate,
5772 241 : self.cancel.clone(),
5773 241 : ctx,
5774 241 : )
5775 241 : .await
5776 241 : .map_err(CreateImageLayersError::Other)?;
5777 :
5778 241 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5779 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5780 0 : "failpoint image-layer-writer-fail-before-finish"
5781 0 : )))
5782 0 : });
5783 :
5784 241 : let io_concurrency = IoConcurrency::spawn_from_conf(
5785 241 : self.conf.get_vectored_concurrent_io,
5786 241 : self.gate
5787 241 : .enter()
5788 241 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5789 : );
5790 :
5791 241 : let outcome = if !compact_metadata {
5792 123 : self.create_image_layer_for_rel_blocks(
5793 123 : partition,
5794 123 : image_layer_writer,
5795 123 : lsn,
5796 123 : ctx,
5797 123 : img_range.clone(),
5798 123 : io_concurrency,
5799 123 : Some((idx, total)),
5800 123 : )
5801 123 : .await?
5802 : } else {
5803 118 : self.create_image_layer_for_metadata_keys(
5804 118 : partition,
5805 118 : image_layer_writer,
5806 118 : lsn,
5807 118 : ctx,
5808 118 : img_range.clone(),
5809 118 : mode,
5810 118 : io_concurrency,
5811 118 : )
5812 118 : .await?
5813 : };
5814 241 : match outcome {
5815 111 : ImageLayerCreationOutcome::Empty => {
5816 111 : // No data in this partition, so we don't need to create an image layer (for now).
5817 111 : // The next image layer should cover this key range, so we don't advance the `start`
5818 111 : // key.
5819 111 : }
5820 : ImageLayerCreationOutcome::Generated {
5821 129 : unfinished_image_layer,
5822 129 : } => {
5823 129 : batch_image_writer.add_unfinished_image_writer(
5824 129 : unfinished_image_layer,
5825 129 : img_range.clone(),
5826 129 : lsn,
5827 129 : );
5828 129 : // The next image layer should be generated right after this one.
5829 129 : start = img_range.end;
5830 129 : }
5831 1 : ImageLayerCreationOutcome::Skip => {
5832 1 : // We don't need to create an image layer for this partition.
5833 1 : // The next image layer should NOT cover this range, otherwise
5834 1 : // the keyspace becomes empty (reads don't go past image layers).
5835 1 : start = img_range.end;
5836 1 : }
5837 : }
5838 :
5839 241 : if let ImageLayerCreationMode::Try = mode {
5840 : // We have at least made some progress
5841 1 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5842 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5843 : // image layer generation taking too long time and blocking L0 compaction. So in this
5844 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5845 : // if there are too many of them.
5846 0 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5847 0 : * self.get_compaction_threshold();
5848 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
5849 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
5850 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
5851 0 : if image_preempt_threshold != 0 {
5852 0 : let should_yield = self
5853 0 : .l0_compaction_trigger
5854 0 : .notified()
5855 0 : .now_or_never()
5856 0 : .is_some();
5857 0 : if should_yield {
5858 0 : tracing::info!(
5859 0 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
5860 0 : partition.start().unwrap(),
5861 0 : partition.end().unwrap()
5862 : );
5863 0 : last_partition_processed = Some(partition.clone());
5864 0 : all_generated = false;
5865 0 : break;
5866 0 : }
5867 0 : }
5868 1 : }
5869 240 : }
5870 : }
5871 :
5872 190 : let image_layers = batch_image_writer
5873 190 : .finish(self, ctx)
5874 190 : .await
5875 190 : .map_err(CreateImageLayersError::Other)?;
5876 :
5877 190 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
5878 :
5879 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5880 : // now they are being scheduled outside of write lock; current way is inconsistent with
5881 : // compaction lock order.
5882 190 : guard
5883 190 : .open_mut()?
5884 190 : .track_new_image_layers(&image_layers, &self.metrics);
5885 190 : drop_layer_manager_wlock(guard);
5886 190 : let duration = timer.stop_and_record();
5887 :
5888 : // Creating image layers may have caused some previously visible layers to be covered
5889 190 : if !image_layers.is_empty() {
5890 117 : self.update_layer_visibility().await?;
5891 73 : }
5892 :
5893 190 : let total_layer_size = image_layers
5894 190 : .iter()
5895 190 : .map(|l| l.metadata().file_size)
5896 190 : .sum::<u64>();
5897 :
5898 190 : if !image_layers.is_empty() {
5899 117 : info!(
5900 0 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5901 0 : image_layers.len(),
5902 : total_layer_size,
5903 0 : duration.as_secs_f64(),
5904 : partition_processed,
5905 : total_partitions
5906 : );
5907 73 : }
5908 :
5909 : Ok((
5910 190 : image_layers,
5911 190 : if all_generated {
5912 190 : LastImageLayerCreationStatus::Complete
5913 : } else {
5914 : LastImageLayerCreationStatus::Incomplete {
5915 0 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5916 0 : last_partition_processed.end().unwrap_or(Key::MIN)
5917 : } else {
5918 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5919 0 : Key::MIN
5920 : },
5921 : }
5922 : },
5923 : ))
5924 190 : }
5925 :
5926 : /// Wait until the background initial logical size calculation is complete, or
5927 : /// this Timeline is shut down. Calling this function will cause the initial
5928 : /// logical size calculation to skip waiting for the background jobs barrier.
5929 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5930 0 : if !self.shard_identity.is_shard_zero() {
5931 : // We don't populate logical size on shard >0: skip waiting for it.
5932 0 : return;
5933 0 : }
5934 :
5935 0 : if self.remote_client.is_deleting() {
5936 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5937 0 : return;
5938 0 : }
5939 :
5940 0 : if self.current_logical_size.current_size().is_exact() {
5941 : // root timelines are initialized with exact count, but never start the background
5942 : // calculation
5943 0 : return;
5944 0 : }
5945 :
5946 0 : if self.cancel.is_cancelled() {
5947 : // We already requested stopping the tenant, so we cannot wait for the logical size
5948 : // calculation to complete given the task might have been already cancelled.
5949 0 : return;
5950 0 : }
5951 :
5952 0 : if let Some(await_bg_cancel) = self
5953 0 : .current_logical_size
5954 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5955 0 : .get()
5956 0 : {
5957 0 : await_bg_cancel.cancel();
5958 0 : } else {
5959 : // We should not wait if we were not able to explicitly instruct
5960 : // the logical size cancellation to skip the concurrency limit semaphore.
5961 : // TODO: this is an unexpected case. We should restructure so that it
5962 : // can't happen.
5963 0 : tracing::warn!(
5964 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5965 : );
5966 0 : debug_assert!(false);
5967 : }
5968 :
5969 0 : tokio::select!(
5970 0 : _ = self.current_logical_size.initialized.acquire() => {},
5971 0 : _ = self.cancel.cancelled() => {}
5972 : )
5973 0 : }
5974 :
5975 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5976 : /// Timelines layers up to the ancestor_lsn.
5977 : ///
5978 : /// Requires a timeline that:
5979 : /// - has an ancestor to detach from
5980 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5981 : /// a technical requirement
5982 : ///
5983 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5984 : /// polled again until completion.
5985 : ///
5986 : /// During the operation all timelines sharing the data with this timeline will be reparented
5987 : /// from our ancestor to be branches of this timeline.
5988 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5989 0 : self: &Arc<Timeline>,
5990 0 : tenant: &crate::tenant::TenantShard,
5991 0 : options: detach_ancestor::Options,
5992 0 : behavior: DetachBehavior,
5993 0 : ctx: &RequestContext,
5994 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5995 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
5996 0 : }
5997 :
5998 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5999 : /// reparents any reparentable children of previous ancestor.
6000 : ///
6001 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
6002 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
6003 : /// successfully, tenant must be reloaded.
6004 : ///
6005 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
6006 : /// resetting the tenant.
6007 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
6008 0 : self: &Arc<Timeline>,
6009 0 : tenant: &crate::tenant::TenantShard,
6010 0 : prepared: detach_ancestor::PreparedTimelineDetach,
6011 0 : ancestor_timeline_id: TimelineId,
6012 0 : ancestor_lsn: Lsn,
6013 0 : behavior: DetachBehavior,
6014 0 : ctx: &RequestContext,
6015 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
6016 0 : detach_ancestor::detach_and_reparent(
6017 0 : self,
6018 0 : tenant,
6019 0 : prepared,
6020 0 : ancestor_timeline_id,
6021 0 : ancestor_lsn,
6022 0 : behavior,
6023 0 : ctx,
6024 0 : )
6025 0 : .await
6026 0 : }
6027 :
6028 : /// Final step which unblocks the GC.
6029 : ///
6030 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
6031 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
6032 0 : self: &Arc<Timeline>,
6033 0 : tenant: &crate::tenant::TenantShard,
6034 0 : attempt: detach_ancestor::Attempt,
6035 0 : ctx: &RequestContext,
6036 0 : ) -> Result<(), detach_ancestor::Error> {
6037 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
6038 0 : }
6039 : }
6040 :
6041 : impl Drop for Timeline {
6042 5 : fn drop(&mut self) {
6043 5 : if let Some(ancestor) = &self.ancestor_timeline {
6044 : // This lock should never be poisoned, but in case it is we do a .map() instead of
6045 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
6046 2 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
6047 2 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
6048 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
6049 0 : "Couldn't remove retain_lsn entry from timeline's parent on drop: already removed");
6050 2 : }
6051 0 : }
6052 3 : }
6053 5 : info!(
6054 0 : "Timeline {} for tenant {} is being dropped",
6055 : self.timeline_id, self.tenant_shard_id.tenant_id
6056 : );
6057 5 : }
6058 : }
6059 :
6060 : /// Top-level failure to compact.
6061 : #[derive(Debug, thiserror::Error)]
6062 : pub(crate) enum CompactionError {
6063 : #[error("The timeline or pageserver is shutting down")]
6064 : ShuttingDown,
6065 : #[error(transparent)]
6066 : Other(anyhow::Error),
6067 : }
6068 :
6069 : impl CompactionError {
6070 : /// Errors that can be ignored, i.e., cancel and shutdown.
6071 0 : pub fn is_cancel(&self) -> bool {
6072 0 : matches!(self, Self::ShuttingDown)
6073 0 : }
6074 :
6075 0 : pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
6076 0 : if err.is_cancel() {
6077 0 : Self::ShuttingDown
6078 : } else {
6079 0 : Self::Other(err.into_anyhow())
6080 : }
6081 0 : }
6082 : }
6083 :
6084 : impl From<super::upload_queue::NotInitialized> for CompactionError {
6085 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
6086 0 : match value {
6087 : super::upload_queue::NotInitialized::Uninitialized => {
6088 0 : CompactionError::Other(anyhow::anyhow!(value))
6089 : }
6090 : super::upload_queue::NotInitialized::ShuttingDown
6091 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
6092 : }
6093 0 : }
6094 : }
6095 :
6096 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
6097 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
6098 0 : match e {
6099 : super::storage_layer::layer::DownloadError::TimelineShutdown
6100 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
6101 0 : CompactionError::ShuttingDown
6102 : }
6103 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
6104 : | super::storage_layer::layer::DownloadError::DownloadRequired
6105 : | super::storage_layer::layer::DownloadError::NotFile(_)
6106 : | super::storage_layer::layer::DownloadError::DownloadFailed
6107 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
6108 0 : CompactionError::Other(anyhow::anyhow!(e))
6109 : }
6110 : #[cfg(test)]
6111 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
6112 0 : CompactionError::Other(anyhow::anyhow!(e))
6113 : }
6114 : }
6115 0 : }
6116 : }
6117 :
6118 : impl From<layer_manager::Shutdown> for CompactionError {
6119 0 : fn from(_: layer_manager::Shutdown) -> Self {
6120 0 : CompactionError::ShuttingDown
6121 0 : }
6122 : }
6123 :
6124 : impl From<super::storage_layer::errors::PutError> for CompactionError {
6125 0 : fn from(e: super::storage_layer::errors::PutError) -> Self {
6126 0 : if e.is_cancel() {
6127 0 : CompactionError::ShuttingDown
6128 : } else {
6129 0 : CompactionError::Other(e.into_anyhow())
6130 : }
6131 0 : }
6132 : }
6133 :
6134 : #[serde_as]
6135 : #[derive(serde::Serialize)]
6136 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
6137 :
6138 : #[derive(Default)]
6139 : enum DurationRecorder {
6140 : #[default]
6141 : NotStarted,
6142 : Recorded(RecordedDuration, tokio::time::Instant),
6143 : }
6144 :
6145 : impl DurationRecorder {
6146 115 : fn till_now(&self) -> DurationRecorder {
6147 115 : match self {
6148 : DurationRecorder::NotStarted => {
6149 0 : panic!("must only call on recorded measurements")
6150 : }
6151 115 : DurationRecorder::Recorded(_, ended) => {
6152 115 : let now = tokio::time::Instant::now();
6153 115 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
6154 : }
6155 : }
6156 115 : }
6157 138 : fn into_recorded(self) -> Option<RecordedDuration> {
6158 138 : match self {
6159 0 : DurationRecorder::NotStarted => None,
6160 138 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
6161 : }
6162 138 : }
6163 : }
6164 :
6165 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
6166 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
6167 : /// the layer descriptor requires the user to provide the ranges, which should cover all
6168 : /// keys specified in the `data` field.
6169 : #[cfg(test)]
6170 : #[derive(Clone)]
6171 : pub struct DeltaLayerTestDesc {
6172 : pub lsn_range: Range<Lsn>,
6173 : pub key_range: Range<Key>,
6174 : pub data: Vec<(Key, Lsn, Value)>,
6175 : }
6176 :
6177 : #[cfg(test)]
6178 : #[derive(Clone)]
6179 : pub struct InMemoryLayerTestDesc {
6180 : pub lsn_range: Range<Lsn>,
6181 : pub data: Vec<(Key, Lsn, Value)>,
6182 : pub is_open: bool,
6183 : }
6184 :
6185 : #[cfg(test)]
6186 : impl DeltaLayerTestDesc {
6187 2 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
6188 2 : Self {
6189 2 : lsn_range,
6190 2 : key_range,
6191 2 : data,
6192 2 : }
6193 2 : }
6194 :
6195 45 : pub fn new_with_inferred_key_range(
6196 45 : lsn_range: Range<Lsn>,
6197 45 : data: Vec<(Key, Lsn, Value)>,
6198 45 : ) -> Self {
6199 45 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
6200 45 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
6201 45 : Self {
6202 45 : key_range: (*key_min)..(key_max.next()),
6203 45 : lsn_range,
6204 45 : data,
6205 45 : }
6206 45 : }
6207 :
6208 5 : pub(crate) fn layer_name(&self) -> LayerName {
6209 5 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6210 5 : key_range: self.key_range.clone(),
6211 5 : lsn_range: self.lsn_range.clone(),
6212 5 : })
6213 5 : }
6214 : }
6215 :
6216 : impl Timeline {
6217 23 : async fn finish_compact_batch(
6218 23 : self: &Arc<Self>,
6219 23 : new_deltas: &[ResidentLayer],
6220 23 : new_images: &[ResidentLayer],
6221 23 : layers_to_remove: &[Layer],
6222 23 : ) -> Result<(), CompactionError> {
6223 23 : let mut guard = tokio::select! {
6224 23 : guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
6225 23 : _ = self.cancel.cancelled() => {
6226 0 : return Err(CompactionError::ShuttingDown);
6227 : }
6228 : };
6229 :
6230 23 : let mut duplicated_layers = HashSet::new();
6231 :
6232 23 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6233 :
6234 186 : for l in new_deltas {
6235 163 : if guard.contains(l.as_ref()) {
6236 : // expected in tests
6237 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6238 :
6239 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6240 : // `guard` on self.layers. as of writing this, there are no error returns except
6241 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6242 : // because we have not implemented L0 => L0 compaction.
6243 0 : duplicated_layers.insert(l.layer_desc().key());
6244 163 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6245 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6246 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6247 0 : )));
6248 163 : } else {
6249 163 : insert_layers.push(l.clone());
6250 163 : }
6251 : }
6252 :
6253 : // only remove those inputs which were not outputs
6254 23 : let remove_layers: Vec<Layer> = layers_to_remove
6255 23 : .iter()
6256 201 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6257 23 : .cloned()
6258 23 : .collect();
6259 :
6260 23 : if !new_images.is_empty() {
6261 0 : guard
6262 0 : .open_mut()?
6263 0 : .track_new_image_layers(new_images, &self.metrics);
6264 23 : }
6265 :
6266 23 : guard
6267 23 : .open_mut()?
6268 23 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6269 :
6270 23 : self.remote_client
6271 23 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6272 :
6273 23 : drop_layer_manager_wlock(guard);
6274 :
6275 23 : Ok(())
6276 23 : }
6277 :
6278 0 : async fn rewrite_layers(
6279 0 : self: &Arc<Self>,
6280 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6281 0 : mut drop_layers: Vec<Layer>,
6282 0 : ) -> Result<(), CompactionError> {
6283 0 : let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
6284 :
6285 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6286 : // to avoid double-removing, and avoid rewriting something that was removed.
6287 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6288 0 : drop_layers.retain(|l| guard.contains(l));
6289 :
6290 0 : guard
6291 0 : .open_mut()?
6292 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6293 :
6294 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6295 :
6296 0 : self.remote_client
6297 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6298 :
6299 0 : Ok(())
6300 0 : }
6301 :
6302 : /// Schedules the uploads of the given image layers
6303 80 : fn upload_new_image_layers(
6304 80 : self: &Arc<Self>,
6305 80 : new_images: impl IntoIterator<Item = ResidentLayer>,
6306 80 : ) -> Result<(), super::upload_queue::NotInitialized> {
6307 93 : for layer in new_images {
6308 13 : self.remote_client.schedule_layer_file_upload(layer)?;
6309 : }
6310 : // should any new image layer been created, not uploading index_part will
6311 : // result in a mismatch between remote_physical_size and layermap calculated
6312 : // size, which will fail some tests, but should not be an issue otherwise.
6313 80 : self.remote_client
6314 80 : .schedule_index_upload_for_file_changes()?;
6315 80 : Ok(())
6316 80 : }
6317 :
6318 0 : async fn find_gc_time_cutoff(
6319 0 : &self,
6320 0 : now: SystemTime,
6321 0 : pitr: Duration,
6322 0 : cancel: &CancellationToken,
6323 0 : ctx: &RequestContext,
6324 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6325 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6326 0 : if self.shard_identity.is_shard_zero() {
6327 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6328 0 : let time_range = if pitr == Duration::ZERO {
6329 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6330 : } else {
6331 0 : pitr
6332 : };
6333 :
6334 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6335 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6336 0 : let timestamp = to_pg_timestamp(time_cutoff);
6337 :
6338 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6339 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6340 0 : LsnForTimestamp::Future(lsn) => {
6341 : // The timestamp is in the future. That sounds impossible,
6342 : // but what it really means is that there hasn't been
6343 : // any commits since the cutoff timestamp.
6344 : //
6345 : // In this case we should use the LSN of the most recent commit,
6346 : // which is implicitly the last LSN in the log.
6347 0 : debug!("future({})", lsn);
6348 0 : Some(self.get_last_record_lsn())
6349 : }
6350 0 : LsnForTimestamp::Past(lsn) => {
6351 0 : debug!("past({})", lsn);
6352 0 : None
6353 : }
6354 0 : LsnForTimestamp::NoData(lsn) => {
6355 0 : debug!("nodata({})", lsn);
6356 0 : None
6357 : }
6358 : };
6359 0 : Ok(time_cutoff)
6360 : } else {
6361 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6362 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6363 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6364 : // space cutoff that we would also have respected ourselves.
6365 0 : match self
6366 0 : .remote_client
6367 0 : .download_foreign_index(ShardNumber(0), cancel)
6368 0 : .await
6369 : {
6370 0 : Ok((index_part, index_generation, _index_mtime)) => {
6371 0 : tracing::info!(
6372 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6373 0 : index_part.metadata.latest_gc_cutoff_lsn()
6374 : );
6375 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6376 : }
6377 : Err(DownloadError::NotFound) => {
6378 : // This is unexpected, because during timeline creations shard zero persists to remote
6379 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6380 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6381 : // state, then shard zero should _eventually_ write an index when it GCs.
6382 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6383 0 : Ok(None)
6384 : }
6385 0 : Err(e) => {
6386 : // TODO: this function should return a different error type than page reconstruct error
6387 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6388 : }
6389 : }
6390 :
6391 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6392 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6393 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6394 : // new location. This is legal in principle, but problematic in practice because it might result
6395 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6396 : // because they have GC'd past the branch point.
6397 : }
6398 0 : }
6399 :
6400 : /// Find the Lsns above which layer files need to be retained on
6401 : /// garbage collection.
6402 : ///
6403 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6404 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6405 : /// the space-based retention.
6406 : ///
6407 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6408 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6409 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6410 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6411 : /// in the response as the GC cutoff point for the timeline.
6412 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6413 : pub(super) async fn find_gc_cutoffs(
6414 : &self,
6415 : now: SystemTime,
6416 : space_cutoff: Lsn,
6417 : pitr: Duration,
6418 : cancel: &CancellationToken,
6419 : ctx: &RequestContext,
6420 : ) -> Result<GcCutoffs, PageReconstructError> {
6421 : let _timer = self
6422 : .metrics
6423 : .find_gc_cutoffs_histo
6424 : .start_timer()
6425 : .record_on_drop();
6426 :
6427 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6428 :
6429 : if cfg!(test) && pitr == Duration::ZERO {
6430 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6431 : return Ok(GcCutoffs {
6432 : time: Some(self.get_last_record_lsn()),
6433 : space: space_cutoff,
6434 : });
6435 : }
6436 :
6437 : // Calculate a time-based limit on how much to retain:
6438 : // - if PITR interval is set, then this is our cutoff.
6439 : // - if PITR interval is not set, then we do a lookup
6440 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6441 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6442 :
6443 : Ok(match (pitr, time_cutoff) {
6444 : (Duration::ZERO, Some(time_cutoff)) => {
6445 : // PITR is not set. Retain the size-based limit, or the default time retention,
6446 : // whichever requires less data.
6447 : GcCutoffs {
6448 : time: Some(self.get_last_record_lsn()),
6449 : space: std::cmp::max(time_cutoff, space_cutoff),
6450 : }
6451 : }
6452 : (Duration::ZERO, None) => {
6453 : // PITR is not set, and time lookup failed
6454 : GcCutoffs {
6455 : time: Some(self.get_last_record_lsn()),
6456 : space: space_cutoff,
6457 : }
6458 : }
6459 : (_, None) => {
6460 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6461 : // cannot advance beyond what was already GC'd, and respect space-based retention
6462 : GcCutoffs {
6463 : time: Some(*self.get_applied_gc_cutoff_lsn()),
6464 : space: space_cutoff,
6465 : }
6466 : }
6467 : (_, Some(time_cutoff)) => {
6468 : // PITR interval is set and we looked up timestamp successfully. Ignore
6469 : // size based retention and make time cutoff authoritative
6470 : GcCutoffs {
6471 : time: Some(time_cutoff),
6472 : space: time_cutoff,
6473 : }
6474 : }
6475 : })
6476 : }
6477 :
6478 : /// Garbage collect layer files on a timeline that are no longer needed.
6479 : ///
6480 : /// Currently, we don't make any attempt at removing unneeded page versions
6481 : /// within a layer file. We can only remove the whole file if it's fully
6482 : /// obsolete.
6483 377 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6484 : // this is most likely the background tasks, but it might be the spawned task from
6485 : // immediate_gc
6486 377 : let _g = tokio::select! {
6487 377 : guard = self.gc_lock.lock() => guard,
6488 377 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6489 : };
6490 377 : let timer = self.metrics.garbage_collect_histo.start_timer();
6491 :
6492 377 : fail_point!("before-timeline-gc");
6493 :
6494 : // Is the timeline being deleted?
6495 377 : if self.is_stopping() {
6496 0 : return Err(GcError::TimelineCancelled);
6497 377 : }
6498 :
6499 377 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6500 377 : let gc_info = self.gc_info.read().unwrap();
6501 :
6502 377 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6503 377 : let time_cutoff = gc_info.cutoffs.time;
6504 377 : let retain_lsns = gc_info
6505 377 : .retain_lsns
6506 377 : .iter()
6507 377 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6508 377 : .collect();
6509 :
6510 : // Gets the maximum LSN that holds the valid lease.
6511 : //
6512 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6513 : // Here, we do not check for stale leases again.
6514 377 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6515 :
6516 377 : (
6517 377 : space_cutoff,
6518 377 : time_cutoff,
6519 377 : retain_lsns,
6520 377 : max_lsn_with_valid_lease,
6521 377 : )
6522 : };
6523 :
6524 377 : let mut new_gc_cutoff = space_cutoff.min(time_cutoff.unwrap_or_default());
6525 377 : let standby_horizon = self.standby_horizon.load();
6526 : // Hold GC for the standby, but as a safety guard do it only within some
6527 : // reasonable lag.
6528 377 : if standby_horizon != Lsn::INVALID {
6529 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6530 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6531 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6532 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6533 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6534 : } else {
6535 0 : warn!(
6536 0 : "standby is lagging for more than {}MB, not holding gc for it",
6537 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6538 : )
6539 : }
6540 0 : }
6541 377 : }
6542 :
6543 : // Reset standby horizon to ignore it if it is not updated till next GC.
6544 : // It is an easy way to unset it when standby disappears without adding
6545 : // more conf options.
6546 377 : self.standby_horizon.store(Lsn::INVALID);
6547 377 : self.metrics
6548 377 : .standby_horizon_gauge
6549 377 : .set(Lsn::INVALID.0 as i64);
6550 :
6551 377 : let res = self
6552 377 : .gc_timeline(
6553 377 : space_cutoff,
6554 377 : time_cutoff,
6555 377 : retain_lsns,
6556 377 : max_lsn_with_valid_lease,
6557 377 : new_gc_cutoff,
6558 : )
6559 377 : .instrument(
6560 377 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6561 : )
6562 377 : .await?;
6563 :
6564 : // only record successes
6565 377 : timer.stop_and_record();
6566 :
6567 377 : Ok(res)
6568 377 : }
6569 :
6570 377 : async fn gc_timeline(
6571 377 : &self,
6572 377 : space_cutoff: Lsn,
6573 377 : time_cutoff: Option<Lsn>, // None if uninitialized
6574 377 : retain_lsns: Vec<Lsn>,
6575 377 : max_lsn_with_valid_lease: Option<Lsn>,
6576 377 : new_gc_cutoff: Lsn,
6577 377 : ) -> Result<GcResult, GcError> {
6578 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6579 :
6580 377 : let now = SystemTime::now();
6581 377 : let mut result: GcResult = GcResult::default();
6582 :
6583 : // Nothing to GC. Return early.
6584 377 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6585 377 : if latest_gc_cutoff >= new_gc_cutoff {
6586 11 : info!(
6587 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6588 : );
6589 11 : return Ok(result);
6590 366 : }
6591 :
6592 366 : let Some(time_cutoff) = time_cutoff else {
6593 : // The GC cutoff should have been computed by now, but let's be defensive.
6594 0 : info!("Nothing to GC: time_cutoff not yet computed");
6595 0 : return Ok(result);
6596 : };
6597 :
6598 : // We need to ensure that no one tries to read page versions or create
6599 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6600 : // for details. This will block until the old value is no longer in use.
6601 : //
6602 : // The GC cutoff should only ever move forwards.
6603 366 : let waitlist = {
6604 366 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6605 366 : if *write_guard > new_gc_cutoff {
6606 0 : return Err(GcError::BadLsn {
6607 0 : why: format!(
6608 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6609 0 : *write_guard, new_gc_cutoff
6610 0 : ),
6611 0 : });
6612 366 : }
6613 :
6614 366 : write_guard.store_and_unlock(new_gc_cutoff)
6615 : };
6616 366 : waitlist.wait().await;
6617 :
6618 366 : info!("GC starting");
6619 :
6620 366 : debug!("retain_lsns: {:?}", retain_lsns);
6621 :
6622 366 : let max_retain_lsn = retain_lsns.iter().max();
6623 :
6624 : // Scan all layers in the timeline (remote or on-disk).
6625 : //
6626 : // Garbage collect the layer if all conditions are satisfied:
6627 : // 1. it is older than cutoff LSN;
6628 : // 2. it is older than PITR interval;
6629 : // 3. it doesn't need to be retained for 'retain_lsns';
6630 : // 4. it does not need to be kept for LSNs holding valid leases.
6631 : // 5. newer on-disk image layers cover the layer's whole key range
6632 366 : let layers_to_remove = {
6633 366 : let mut layers_to_remove = Vec::new();
6634 :
6635 366 : let guard = self
6636 366 : .layers
6637 366 : .read(LayerManagerLockHolder::GarbageCollection)
6638 366 : .await;
6639 366 : let layers = guard.layer_map()?;
6640 6209 : 'outer: for l in layers.iter_historic_layers() {
6641 6209 : result.layers_total += 1;
6642 :
6643 : // 1. Is it newer than GC horizon cutoff point?
6644 6209 : if l.get_lsn_range().end > space_cutoff {
6645 371 : debug!(
6646 0 : "keeping {} because it's newer than space_cutoff {}",
6647 0 : l.layer_name(),
6648 : space_cutoff,
6649 : );
6650 371 : result.layers_needed_by_cutoff += 1;
6651 371 : continue 'outer;
6652 5838 : }
6653 :
6654 : // 2. It is newer than PiTR cutoff point?
6655 5838 : if l.get_lsn_range().end > time_cutoff {
6656 0 : debug!(
6657 0 : "keeping {} because it's newer than time_cutoff {}",
6658 0 : l.layer_name(),
6659 : time_cutoff,
6660 : );
6661 0 : result.layers_needed_by_pitr += 1;
6662 0 : continue 'outer;
6663 5838 : }
6664 :
6665 : // 3. Is it needed by a child branch?
6666 : // NOTE With that we would keep data that
6667 : // might be referenced by child branches forever.
6668 : // We can track this in child timeline GC and delete parent layers when
6669 : // they are no longer needed. This might be complicated with long inheritance chains.
6670 5838 : if let Some(retain_lsn) = max_retain_lsn {
6671 : // start_lsn is inclusive
6672 6 : if &l.get_lsn_range().start <= retain_lsn {
6673 6 : debug!(
6674 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6675 0 : l.layer_name(),
6676 : retain_lsn,
6677 0 : l.is_incremental(),
6678 : );
6679 6 : result.layers_needed_by_branches += 1;
6680 6 : continue 'outer;
6681 0 : }
6682 5832 : }
6683 :
6684 : // 4. Is there a valid lease that requires us to keep this layer?
6685 5832 : if let Some(lsn) = &max_lsn_with_valid_lease {
6686 : // keep if layer start <= any of the lease
6687 9 : if &l.get_lsn_range().start <= lsn {
6688 7 : debug!(
6689 0 : "keeping {} because there is a valid lease preventing GC at {}",
6690 0 : l.layer_name(),
6691 : lsn,
6692 : );
6693 7 : result.layers_needed_by_leases += 1;
6694 7 : continue 'outer;
6695 2 : }
6696 5823 : }
6697 :
6698 : // 5. Is there a later on-disk layer for this relation?
6699 : //
6700 : // The end-LSN is exclusive, while disk_consistent_lsn is
6701 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6702 : // OK for a delta layer to have end LSN 101, but if the end LSN
6703 : // is 102, then it might not have been fully flushed to disk
6704 : // before crash.
6705 : //
6706 : // For example, imagine that the following layers exist:
6707 : //
6708 : // 1000 - image (A)
6709 : // 1000-2000 - delta (B)
6710 : // 2000 - image (C)
6711 : // 2000-3000 - delta (D)
6712 : // 3000 - image (E)
6713 : //
6714 : // If GC horizon is at 2500, we can remove layers A and B, but
6715 : // we cannot remove C, even though it's older than 2500, because
6716 : // the delta layer 2000-3000 depends on it.
6717 5825 : if !layers
6718 5825 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6719 : {
6720 5821 : debug!("keeping {} because it is the latest layer", l.layer_name());
6721 5821 : result.layers_not_updated += 1;
6722 5821 : continue 'outer;
6723 4 : }
6724 :
6725 : // We didn't find any reason to keep this file, so remove it.
6726 4 : info!(
6727 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6728 0 : l.layer_name(),
6729 0 : l.is_incremental(),
6730 : );
6731 4 : layers_to_remove.push(l);
6732 : }
6733 :
6734 366 : layers_to_remove
6735 : };
6736 :
6737 366 : if !layers_to_remove.is_empty() {
6738 : // Persist the new GC cutoff value before we actually remove anything.
6739 : // This unconditionally schedules also an index_part.json update, even though, we will
6740 : // be doing one a bit later with the unlinked gc'd layers.
6741 3 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6742 3 : self.schedule_uploads(disk_consistent_lsn, None)
6743 3 : .map_err(|e| {
6744 0 : if self.cancel.is_cancelled() {
6745 0 : GcError::TimelineCancelled
6746 : } else {
6747 0 : GcError::Remote(e)
6748 : }
6749 0 : })?;
6750 :
6751 3 : let mut guard = self
6752 3 : .layers
6753 3 : .write(LayerManagerLockHolder::GarbageCollection)
6754 3 : .await;
6755 :
6756 3 : let gc_layers = layers_to_remove
6757 3 : .iter()
6758 4 : .flat_map(|desc| guard.try_get_from_key(&desc.key()).cloned())
6759 3 : .collect::<Vec<Layer>>();
6760 :
6761 3 : result.layers_removed = gc_layers.len() as u64;
6762 :
6763 3 : self.remote_client.schedule_gc_update(&gc_layers)?;
6764 3 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
6765 :
6766 : #[cfg(feature = "testing")]
6767 3 : {
6768 3 : result.doomed_layers = gc_layers;
6769 3 : }
6770 363 : }
6771 :
6772 366 : info!(
6773 0 : "GC completed removing {} layers, cutoff {}",
6774 : result.layers_removed, new_gc_cutoff
6775 : );
6776 :
6777 366 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
6778 366 : Ok(result)
6779 377 : }
6780 :
6781 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6782 364899 : async fn reconstruct_value(
6783 364899 : &self,
6784 364899 : key: Key,
6785 364899 : request_lsn: Lsn,
6786 364899 : mut data: ValueReconstructState,
6787 364899 : redo_attempt_type: RedoAttemptType,
6788 364899 : ) -> Result<Bytes, PageReconstructError> {
6789 : // Perform WAL redo if needed
6790 364899 : data.records.reverse();
6791 :
6792 364899 : let fire_critical_error = match redo_attempt_type {
6793 363566 : RedoAttemptType::ReadPage => true,
6794 0 : RedoAttemptType::LegacyCompaction => true,
6795 1333 : RedoAttemptType::GcCompaction => false,
6796 : };
6797 :
6798 : // If we have a page image, and no WAL, we're all set
6799 364899 : if data.records.is_empty() {
6800 338125 : if let Some((img_lsn, img)) = &data.img {
6801 338125 : trace!(
6802 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6803 : key, img_lsn, request_lsn,
6804 : );
6805 338125 : Ok(img.clone())
6806 : } else {
6807 0 : Err(PageReconstructError::from(anyhow!(
6808 0 : "base image for {key} at {request_lsn} not found"
6809 0 : )))
6810 : }
6811 : } else {
6812 : // We need to do WAL redo.
6813 : //
6814 : // If we don't have a base image, then the oldest WAL record better initialize
6815 : // the page
6816 26774 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6817 0 : Err(PageReconstructError::from(anyhow!(
6818 0 : "Base image for {} at {} not found, but got {} WAL records",
6819 0 : key,
6820 0 : request_lsn,
6821 0 : data.records.len()
6822 0 : )))
6823 : } else {
6824 26774 : if data.img.is_some() {
6825 13029 : trace!(
6826 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6827 0 : data.records.len(),
6828 : key,
6829 : request_lsn
6830 : );
6831 : } else {
6832 13745 : trace!(
6833 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
6834 0 : data.records.len(),
6835 : key,
6836 : request_lsn
6837 : );
6838 : };
6839 26774 : let res = self
6840 26774 : .walredo_mgr
6841 26774 : .as_ref()
6842 26774 : .context("timeline has no walredo manager")
6843 26774 : .map_err(PageReconstructError::WalRedo)?
6844 26774 : .request_redo(
6845 26774 : key,
6846 26774 : request_lsn,
6847 26774 : data.img,
6848 26774 : data.records,
6849 26774 : self.pg_version,
6850 26774 : redo_attempt_type,
6851 : )
6852 26774 : .await;
6853 26773 : let img = match res {
6854 26773 : Ok(img) => img,
6855 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6856 1 : Err(walredo::Error::Other(err)) => {
6857 1 : if fire_critical_error {
6858 0 : critical_timeline!(
6859 0 : self.tenant_shard_id,
6860 0 : self.timeline_id,
6861 0 : "walredo failure during page reconstruction: {err:?}"
6862 : );
6863 1 : }
6864 1 : return Err(PageReconstructError::WalRedo(
6865 1 : err.context("reconstruct a page image"),
6866 1 : ));
6867 : }
6868 : };
6869 26773 : Ok(img)
6870 : }
6871 : }
6872 364899 : }
6873 :
6874 0 : pub(crate) async fn spawn_download_all_remote_layers(
6875 0 : self: Arc<Self>,
6876 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6877 0 : ctx: &RequestContext,
6878 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6879 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6880 :
6881 : // this is not really needed anymore; it has tests which really check the return value from
6882 : // http api. it would be better not to maintain this anymore.
6883 :
6884 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6885 0 : if let Some(st) = &*status_guard {
6886 0 : match &st.state {
6887 : DownloadRemoteLayersTaskState::Running => {
6888 0 : return Err(st.clone());
6889 : }
6890 : DownloadRemoteLayersTaskState::ShutDown
6891 0 : | DownloadRemoteLayersTaskState::Completed => {
6892 0 : *status_guard = None;
6893 0 : }
6894 : }
6895 0 : }
6896 :
6897 0 : let self_clone = Arc::clone(&self);
6898 0 : let task_ctx = ctx.detached_child(
6899 0 : TaskKind::DownloadAllRemoteLayers,
6900 0 : DownloadBehavior::Download,
6901 : );
6902 0 : let task_id = task_mgr::spawn(
6903 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6904 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6905 0 : self.tenant_shard_id,
6906 0 : Some(self.timeline_id),
6907 0 : "download all remote layers task",
6908 0 : async move {
6909 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
6910 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6911 0 : match &mut *status_guard {
6912 : None => {
6913 0 : warn!("tasks status is supposed to be Some(), since we are running");
6914 : }
6915 0 : Some(st) => {
6916 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6917 0 : if st.task_id != exp_task_id {
6918 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6919 0 : } else {
6920 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6921 0 : }
6922 : }
6923 : };
6924 0 : Ok(())
6925 0 : }
6926 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))
6927 : );
6928 :
6929 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6930 0 : task_id: format!("{task_id}"),
6931 0 : state: DownloadRemoteLayersTaskState::Running,
6932 0 : total_layer_count: 0,
6933 0 : successful_download_count: 0,
6934 0 : failed_download_count: 0,
6935 0 : };
6936 0 : *status_guard = Some(initial_info.clone());
6937 :
6938 0 : Ok(initial_info)
6939 0 : }
6940 :
6941 0 : async fn download_all_remote_layers(
6942 0 : self: &Arc<Self>,
6943 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6944 0 : ctx: &RequestContext,
6945 0 : ) {
6946 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6947 :
6948 0 : let remaining = {
6949 0 : let guard = self
6950 0 : .layers
6951 0 : .read(LayerManagerLockHolder::GetLayerMapInfo)
6952 0 : .await;
6953 0 : let Ok(lm) = guard.layer_map() else {
6954 : // technically here we could look into iterating accessible layers, but downloading
6955 : // all layers of a shutdown timeline makes no sense regardless.
6956 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6957 0 : return;
6958 : };
6959 0 : lm.iter_historic_layers()
6960 0 : .map(|desc| guard.get_from_desc(&desc))
6961 0 : .collect::<Vec<_>>()
6962 : };
6963 0 : let total_layer_count = remaining.len();
6964 :
6965 : macro_rules! lock_status {
6966 : ($st:ident) => {
6967 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6968 : let st = st
6969 : .as_mut()
6970 : .expect("this function is only called after the task has been spawned");
6971 : assert_eq!(
6972 : st.task_id,
6973 : format!(
6974 : "{}",
6975 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6976 : )
6977 : );
6978 : let $st = st;
6979 : };
6980 : }
6981 :
6982 : {
6983 0 : lock_status!(st);
6984 0 : st.total_layer_count = total_layer_count as u64;
6985 : }
6986 :
6987 0 : let mut remaining = remaining.into_iter();
6988 0 : let mut have_remaining = true;
6989 0 : let mut js = tokio::task::JoinSet::new();
6990 :
6991 0 : let cancel = task_mgr::shutdown_token();
6992 :
6993 0 : let limit = request.max_concurrent_downloads;
6994 :
6995 : loop {
6996 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6997 0 : let Some(next) = remaining.next() else {
6998 0 : have_remaining = false;
6999 0 : break;
7000 : };
7001 :
7002 0 : let span = tracing::info_span!("download", layer = %next);
7003 :
7004 0 : let ctx = ctx.attached_child();
7005 0 : js.spawn(
7006 0 : async move {
7007 0 : let res = next.download(&ctx).await;
7008 0 : (next, res)
7009 0 : }
7010 0 : .instrument(span),
7011 : );
7012 : }
7013 :
7014 0 : while let Some(res) = js.join_next().await {
7015 0 : match res {
7016 : Ok((_, Ok(_))) => {
7017 0 : lock_status!(st);
7018 0 : st.successful_download_count += 1;
7019 : }
7020 0 : Ok((layer, Err(e))) => {
7021 0 : tracing::error!(%layer, "download failed: {e:#}");
7022 0 : lock_status!(st);
7023 0 : st.failed_download_count += 1;
7024 : }
7025 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
7026 0 : Err(je) if je.is_panic() => {
7027 0 : lock_status!(st);
7028 0 : st.failed_download_count += 1;
7029 : }
7030 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
7031 : }
7032 : }
7033 :
7034 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
7035 0 : break;
7036 0 : }
7037 : }
7038 :
7039 : {
7040 0 : lock_status!(st);
7041 0 : st.state = DownloadRemoteLayersTaskState::Completed;
7042 : }
7043 0 : }
7044 :
7045 0 : pub(crate) fn get_download_all_remote_layers_task_info(
7046 0 : &self,
7047 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
7048 0 : self.download_all_remote_layers_task_info
7049 0 : .read()
7050 0 : .unwrap()
7051 0 : .clone()
7052 0 : }
7053 : }
7054 :
7055 : impl Timeline {
7056 : /// Returns non-remote layers for eviction.
7057 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
7058 0 : let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
7059 0 : let mut max_layer_size: Option<u64> = None;
7060 :
7061 0 : let resident_layers = guard
7062 0 : .likely_resident_layers()
7063 0 : .map(|layer| {
7064 0 : let file_size = layer.layer_desc().file_size;
7065 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
7066 :
7067 0 : let last_activity_ts = layer.latest_activity();
7068 :
7069 0 : EvictionCandidate {
7070 0 : layer: layer.to_owned().into(),
7071 0 : last_activity_ts,
7072 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
7073 0 : visibility: layer.visibility(),
7074 0 : }
7075 0 : })
7076 0 : .collect();
7077 :
7078 0 : DiskUsageEvictionInfo {
7079 0 : max_layer_size,
7080 0 : resident_layers,
7081 0 : }
7082 0 : }
7083 :
7084 961 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
7085 961 : ShardIndex {
7086 961 : shard_number: self.tenant_shard_id.shard_number,
7087 961 : shard_count: self.tenant_shard_id.shard_count,
7088 961 : }
7089 961 : }
7090 :
7091 : /// Persistently blocks gc for `Manual` reason.
7092 : ///
7093 : /// Returns true if no such block existed before, false otherwise.
7094 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
7095 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7096 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7097 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
7098 0 : }
7099 :
7100 : /// Persistently unblocks gc for `Manual` reason.
7101 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
7102 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
7103 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
7104 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
7105 0 : }
7106 :
7107 : #[cfg(test)]
7108 31 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
7109 31 : self.last_record_lsn.advance(new_lsn);
7110 31 : }
7111 :
7112 : #[cfg(test)]
7113 2 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
7114 2 : self.disk_consistent_lsn.store(new_value);
7115 2 : }
7116 :
7117 : /// Force create an image layer and place it into the layer map.
7118 : ///
7119 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7120 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7121 : /// placed into the layer map in one run AND be validated.
7122 : #[cfg(test)]
7123 36 : pub(super) async fn force_create_image_layer(
7124 36 : self: &Arc<Timeline>,
7125 36 : lsn: Lsn,
7126 36 : mut images: Vec<(Key, Bytes)>,
7127 36 : check_start_lsn: Option<Lsn>,
7128 36 : ctx: &RequestContext,
7129 36 : ) -> anyhow::Result<()> {
7130 36 : let last_record_lsn = self.get_last_record_lsn();
7131 36 : assert!(
7132 36 : lsn <= last_record_lsn,
7133 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
7134 : );
7135 36 : if let Some(check_start_lsn) = check_start_lsn {
7136 36 : assert!(lsn >= check_start_lsn);
7137 0 : }
7138 240 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
7139 36 : let min_key = *images.first().map(|(k, _)| k).unwrap();
7140 36 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
7141 36 : let mut image_layer_writer = ImageLayerWriter::new(
7142 36 : self.conf,
7143 36 : self.timeline_id,
7144 36 : self.tenant_shard_id,
7145 36 : &(min_key..end_key),
7146 36 : lsn,
7147 36 : &self.gate,
7148 36 : self.cancel.clone(),
7149 36 : ctx,
7150 36 : )
7151 36 : .await?;
7152 312 : for (key, img) in images {
7153 276 : image_layer_writer.put_image(key, img, ctx).await?;
7154 : }
7155 36 : let (desc, path) = image_layer_writer.finish(ctx).await?;
7156 36 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7157 36 : info!("force created image layer {}", image_layer.local_path());
7158 : {
7159 36 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7160 36 : guard
7161 36 : .open_mut()
7162 36 : .unwrap()
7163 36 : .force_insert_layer(image_layer.clone());
7164 : }
7165 :
7166 : // Update remote_timeline_client state to reflect existence of this layer
7167 36 : self.remote_client
7168 36 : .schedule_layer_file_upload(image_layer)
7169 36 : .unwrap();
7170 :
7171 36 : Ok(())
7172 36 : }
7173 :
7174 : /// Force create a delta layer and place it into the layer map.
7175 : ///
7176 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
7177 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
7178 : /// placed into the layer map in one run AND be validated.
7179 : #[cfg(test)]
7180 50 : pub(super) async fn force_create_delta_layer(
7181 50 : self: &Arc<Timeline>,
7182 50 : mut deltas: DeltaLayerTestDesc,
7183 50 : check_start_lsn: Option<Lsn>,
7184 50 : ctx: &RequestContext,
7185 50 : ) -> anyhow::Result<()> {
7186 50 : let last_record_lsn = self.get_last_record_lsn();
7187 50 : deltas
7188 50 : .data
7189 124364 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7190 50 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
7191 50 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
7192 10464 : for (_, lsn, _) in &deltas.data {
7193 10414 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
7194 : }
7195 50 : assert!(
7196 50 : deltas.lsn_range.end <= last_record_lsn,
7197 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7198 : deltas.lsn_range.end,
7199 : last_record_lsn
7200 : );
7201 50 : if let Some(check_start_lsn) = check_start_lsn {
7202 50 : assert!(deltas.lsn_range.start >= check_start_lsn);
7203 0 : }
7204 50 : let mut delta_layer_writer = DeltaLayerWriter::new(
7205 50 : self.conf,
7206 50 : self.timeline_id,
7207 50 : self.tenant_shard_id,
7208 50 : deltas.key_range.start,
7209 50 : deltas.lsn_range,
7210 50 : &self.gate,
7211 50 : self.cancel.clone(),
7212 50 : ctx,
7213 50 : )
7214 50 : .await?;
7215 10464 : for (key, lsn, val) in deltas.data {
7216 10414 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
7217 : }
7218 50 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
7219 50 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
7220 50 : info!("force created delta layer {}", delta_layer.local_path());
7221 : {
7222 50 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7223 50 : guard
7224 50 : .open_mut()
7225 50 : .unwrap()
7226 50 : .force_insert_layer(delta_layer.clone());
7227 : }
7228 :
7229 : // Update remote_timeline_client state to reflect existence of this layer
7230 50 : self.remote_client
7231 50 : .schedule_layer_file_upload(delta_layer)
7232 50 : .unwrap();
7233 :
7234 50 : Ok(())
7235 50 : }
7236 :
7237 : /// Force create an in-memory layer and place them into the layer map.
7238 : #[cfg(test)]
7239 4 : pub(super) async fn force_create_in_memory_layer(
7240 4 : self: &Arc<Timeline>,
7241 4 : mut in_memory: InMemoryLayerTestDesc,
7242 4 : check_start_lsn: Option<Lsn>,
7243 4 : ctx: &RequestContext,
7244 4 : ) -> anyhow::Result<()> {
7245 : use utils::bin_ser::BeSer;
7246 :
7247 : // Validate LSNs
7248 4 : if let Some(check_start_lsn) = check_start_lsn {
7249 4 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7250 0 : }
7251 :
7252 4 : let last_record_lsn = self.get_last_record_lsn();
7253 4 : let layer_end_lsn = if in_memory.is_open {
7254 1 : in_memory
7255 1 : .data
7256 1 : .iter()
7257 1 : .map(|(_key, lsn, _value)| lsn)
7258 1 : .max()
7259 1 : .cloned()
7260 : } else {
7261 3 : Some(in_memory.lsn_range.end)
7262 : };
7263 :
7264 4 : if let Some(end) = layer_end_lsn {
7265 4 : assert!(
7266 4 : end <= last_record_lsn,
7267 0 : "advance last record lsn before inserting a layer, end_lsn={end}, last_record_lsn={last_record_lsn}",
7268 : );
7269 0 : }
7270 :
7271 19820 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7272 19820 : assert!(*lsn >= in_memory.lsn_range.start);
7273 19820 : assert!(*lsn < in_memory.lsn_range.end);
7274 19820 : });
7275 :
7276 : // Build the batch
7277 4 : in_memory
7278 4 : .data
7279 273384 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7280 :
7281 4 : let data = in_memory
7282 4 : .data
7283 4 : .into_iter()
7284 19820 : .map(|(key, lsn, value)| {
7285 19820 : let value_size = value.serialized_size().unwrap() as usize;
7286 19820 : (key.to_compact(), lsn, value_size, value)
7287 19820 : })
7288 4 : .collect::<Vec<_>>();
7289 :
7290 4 : let batch = SerializedValueBatch::from_values(data);
7291 :
7292 : // Create the in-memory layer and write the batch into it
7293 4 : let layer = InMemoryLayer::create(
7294 4 : self.conf,
7295 4 : self.timeline_id,
7296 4 : self.tenant_shard_id,
7297 4 : in_memory.lsn_range.start,
7298 4 : &self.gate,
7299 4 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7300 4 : &CancellationToken::new(),
7301 4 : ctx,
7302 4 : )
7303 4 : .await
7304 4 : .unwrap();
7305 :
7306 4 : layer.put_batch(batch, ctx).await.unwrap();
7307 4 : if !in_memory.is_open {
7308 3 : layer.freeze(in_memory.lsn_range.end).await;
7309 1 : }
7310 :
7311 4 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7312 :
7313 : // Link the layer to the layer map
7314 : {
7315 4 : let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
7316 4 : let layer_map = guard.open_mut().unwrap();
7317 4 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7318 : }
7319 :
7320 4 : Ok(())
7321 4 : }
7322 :
7323 : /// Return all keys at the LSN in the image layers
7324 : #[cfg(test)]
7325 3 : pub(crate) async fn inspect_image_layers(
7326 3 : self: &Arc<Timeline>,
7327 3 : lsn: Lsn,
7328 3 : ctx: &RequestContext,
7329 3 : io_concurrency: IoConcurrency,
7330 3 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7331 3 : let mut all_data = Vec::new();
7332 3 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7333 17 : for layer in guard.layer_map()?.iter_historic_layers() {
7334 17 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7335 4 : let layer = guard.get_from_desc(&layer);
7336 4 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7337 4 : layer
7338 4 : .get_values_reconstruct_data(
7339 4 : KeySpace::single(Key::MIN..Key::MAX),
7340 4 : lsn..Lsn(lsn.0 + 1),
7341 4 : &mut reconstruct_data,
7342 4 : ctx,
7343 4 : )
7344 4 : .await?;
7345 33 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7346 33 : let v = v.collect_pending_ios().await?;
7347 33 : all_data.push((k, v.img.unwrap().1));
7348 : }
7349 13 : }
7350 : }
7351 3 : all_data.sort();
7352 3 : Ok(all_data)
7353 3 : }
7354 :
7355 : /// Get all historic layer descriptors in the layer map
7356 : #[cfg(test)]
7357 12 : pub(crate) async fn inspect_historic_layers(
7358 12 : self: &Arc<Timeline>,
7359 12 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7360 12 : let mut layers = Vec::new();
7361 12 : let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
7362 57 : for layer in guard.layer_map()?.iter_historic_layers() {
7363 57 : layers.push(layer.key());
7364 57 : }
7365 12 : Ok(layers)
7366 12 : }
7367 :
7368 : #[cfg(test)]
7369 5 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7370 5 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7371 5 : keyspace.merge(&ks);
7372 5 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7373 5 : }
7374 : }
7375 :
7376 : /// Tracking writes ingestion does to a particular in-memory layer.
7377 : ///
7378 : /// Cleared upon freezing a layer.
7379 : pub(crate) struct TimelineWriterState {
7380 : open_layer: Arc<InMemoryLayer>,
7381 : current_size: u64,
7382 : // Previous Lsn which passed through
7383 : prev_lsn: Option<Lsn>,
7384 : // Largest Lsn which passed through the current writer
7385 : max_lsn: Option<Lsn>,
7386 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7387 : cached_last_freeze_at: Lsn,
7388 : }
7389 :
7390 : impl TimelineWriterState {
7391 659 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7392 659 : Self {
7393 659 : open_layer,
7394 659 : current_size,
7395 659 : prev_lsn: None,
7396 659 : max_lsn: None,
7397 659 : cached_last_freeze_at: last_freeze_at,
7398 659 : }
7399 659 : }
7400 : }
7401 :
7402 : /// Various functions to mutate the timeline.
7403 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7404 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7405 : // but will cause large code changes.
7406 : pub(crate) struct TimelineWriter<'a> {
7407 : tl: &'a Timeline,
7408 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7409 : }
7410 :
7411 : impl Deref for TimelineWriter<'_> {
7412 : type Target = Timeline;
7413 :
7414 4949213 : fn deref(&self) -> &Self::Target {
7415 4949213 : self.tl
7416 4949213 : }
7417 : }
7418 :
7419 : #[derive(PartialEq)]
7420 : enum OpenLayerAction {
7421 : Roll,
7422 : Open,
7423 : None,
7424 : }
7425 :
7426 : impl TimelineWriter<'_> {
7427 2402128 : async fn handle_open_layer_action(
7428 2402128 : &mut self,
7429 2402128 : at: Lsn,
7430 2402128 : action: OpenLayerAction,
7431 2402128 : ctx: &RequestContext,
7432 2402128 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7433 2402128 : match action {
7434 : OpenLayerAction::Roll => {
7435 40 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7436 40 : self.roll_layer(freeze_at).await?;
7437 40 : self.open_layer(at, ctx).await?;
7438 : }
7439 619 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7440 : OpenLayerAction::None => {
7441 2401469 : assert!(self.write_guard.is_some());
7442 : }
7443 : }
7444 :
7445 2402128 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7446 2402128 : }
7447 :
7448 659 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7449 659 : let layer = self
7450 659 : .tl
7451 659 : .get_layer_for_write(at, &self.write_guard, ctx)
7452 659 : .await?;
7453 659 : let initial_size = layer.len();
7454 :
7455 659 : let last_freeze_at = self.last_freeze_at.load();
7456 659 : self.write_guard.replace(TimelineWriterState::new(
7457 659 : layer,
7458 659 : initial_size,
7459 659 : last_freeze_at,
7460 659 : ));
7461 :
7462 659 : Ok(())
7463 659 : }
7464 :
7465 40 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7466 40 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7467 :
7468 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7469 : // to propagate the backpressure up into WAL ingestion.
7470 40 : let l0_count = self
7471 40 : .tl
7472 40 : .layers
7473 40 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7474 40 : .await
7475 40 : .layer_map()?
7476 40 : .level0_deltas()
7477 40 : .len();
7478 40 : let wait_thresholds = [
7479 40 : self.get_l0_flush_delay_threshold(),
7480 40 : self.get_l0_flush_stall_threshold(),
7481 40 : ];
7482 40 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7483 :
7484 : // self.write_guard will be taken by the freezing
7485 40 : let flush_id = self
7486 40 : .tl
7487 40 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7488 40 : .await?;
7489 :
7490 40 : assert!(self.write_guard.is_none());
7491 :
7492 40 : if let Some(wait_threshold) = wait_threshold {
7493 0 : if l0_count >= wait_threshold {
7494 0 : debug!(
7495 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7496 : );
7497 0 : self.tl.wait_flush_completion(flush_id).await?;
7498 0 : }
7499 40 : }
7500 :
7501 40 : if current_size >= self.get_checkpoint_distance() * 2 {
7502 0 : warn!("Flushed oversized open layer with size {}", current_size)
7503 40 : }
7504 :
7505 40 : Ok(())
7506 40 : }
7507 :
7508 2402128 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7509 2402128 : let state = &*self.write_guard;
7510 2402128 : let Some(state) = &state else {
7511 619 : return OpenLayerAction::Open;
7512 : };
7513 :
7514 : #[cfg(feature = "testing")]
7515 2401509 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7516 : // this check and assertion are not really needed because
7517 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7518 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7519 : // is no TimelineWriterState.
7520 0 : assert!(
7521 0 : state.open_layer.end_lsn.get().is_some(),
7522 0 : "our open_layer must be outdated"
7523 : );
7524 :
7525 : // this would be a memory leak waiting to happen because the in-memory layer always has
7526 : // an index
7527 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7528 2401509 : }
7529 :
7530 2401509 : if state.prev_lsn == Some(lsn) {
7531 : // Rolling mid LSN is not supported by [downstream code].
7532 : // Hence, only roll at LSN boundaries.
7533 : //
7534 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7535 3 : return OpenLayerAction::None;
7536 2401506 : }
7537 :
7538 2401506 : if state.current_size == 0 {
7539 : // Don't roll empty layers
7540 0 : return OpenLayerAction::None;
7541 2401506 : }
7542 :
7543 2401506 : if self.tl.should_roll(
7544 2401506 : state.current_size,
7545 2401506 : state.current_size + new_value_size,
7546 2401506 : self.get_checkpoint_distance(),
7547 2401506 : lsn,
7548 2401506 : state.cached_last_freeze_at,
7549 2401506 : state.open_layer.get_opened_at(),
7550 : ) {
7551 40 : OpenLayerAction::Roll
7552 : } else {
7553 2401466 : OpenLayerAction::None
7554 : }
7555 2402128 : }
7556 :
7557 : /// Put a batch of keys at the specified Lsns.
7558 2402127 : pub(crate) async fn put_batch(
7559 2402127 : &mut self,
7560 2402127 : batch: SerializedValueBatch,
7561 2402127 : ctx: &RequestContext,
7562 2402127 : ) -> anyhow::Result<()> {
7563 2402127 : if !batch.has_data() {
7564 0 : return Ok(());
7565 2402127 : }
7566 :
7567 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7568 : // We don't assert this in release builds, since key ownership policies may change over
7569 : // time. Stray keys will be removed during compaction.
7570 2402127 : if cfg!(debug_assertions) {
7571 4947521 : for metadata in &batch.metadata {
7572 2545394 : if let ValueMeta::Serialized(metadata) = metadata {
7573 2545394 : let key = Key::from_compact(metadata.key);
7574 2545394 : assert!(
7575 2545394 : self.shard_identity.is_key_local(&key)
7576 12 : || self.shard_identity.is_key_global(&key),
7577 0 : "key {key} does not belong on shard {}",
7578 0 : self.shard_identity.shard_index()
7579 : );
7580 0 : }
7581 : }
7582 0 : }
7583 :
7584 2402127 : let batch_max_lsn = batch.max_lsn;
7585 2402127 : let buf_size: u64 = batch.buffer_size() as u64;
7586 :
7587 2402127 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7588 2402127 : let layer = self
7589 2402127 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7590 2402127 : .await?;
7591 :
7592 2402127 : let res = layer.put_batch(batch, ctx).await;
7593 :
7594 2402127 : if res.is_ok() {
7595 2402127 : // Update the current size only when the entire write was ok.
7596 2402127 : // In case of failures, we may have had partial writes which
7597 2402127 : // render the size tracking out of sync. That's ok because
7598 2402127 : // the checkpoint distance should be significantly smaller
7599 2402127 : // than the S3 single shot upload limit of 5GiB.
7600 2402127 : let state = self.write_guard.as_mut().unwrap();
7601 2402127 :
7602 2402127 : state.current_size += buf_size;
7603 2402127 : state.prev_lsn = Some(batch_max_lsn);
7604 2402127 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7605 2402127 : }
7606 :
7607 2402127 : res
7608 2402127 : }
7609 :
7610 : #[cfg(test)]
7611 : /// Test helper, for tests that would like to poke individual values without composing a batch
7612 2195079 : pub(crate) async fn put(
7613 2195079 : &mut self,
7614 2195079 : key: Key,
7615 2195079 : lsn: Lsn,
7616 2195079 : value: &Value,
7617 2195079 : ctx: &RequestContext,
7618 2195079 : ) -> anyhow::Result<()> {
7619 : use utils::bin_ser::BeSer;
7620 2195079 : if !key.is_valid_key_on_write_path() {
7621 0 : bail!(
7622 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7623 : key
7624 : );
7625 2195079 : }
7626 2195079 : let val_ser_size = value.serialized_size().unwrap() as usize;
7627 2195079 : let batch = SerializedValueBatch::from_values(vec![(
7628 2195079 : key.to_compact(),
7629 2195079 : lsn,
7630 2195079 : val_ser_size,
7631 2195079 : value.clone(),
7632 2195079 : )]);
7633 :
7634 2195079 : self.put_batch(batch, ctx).await
7635 2195079 : }
7636 :
7637 1 : pub(crate) async fn delete_batch(
7638 1 : &mut self,
7639 1 : batch: &[(Range<Key>, Lsn)],
7640 1 : ctx: &RequestContext,
7641 1 : ) -> anyhow::Result<()> {
7642 1 : if let Some((_, lsn)) = batch.first() {
7643 1 : let action = self.get_open_layer_action(*lsn, 0);
7644 1 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7645 1 : layer.put_tombstones(batch).await?;
7646 0 : }
7647 :
7648 1 : Ok(())
7649 1 : }
7650 :
7651 : /// Track the end of the latest digested WAL record.
7652 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7653 : ///
7654 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7655 : ///
7656 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7657 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7658 : /// the latest and can be read.
7659 2639558 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7660 2639558 : self.tl.finish_write(new_lsn);
7661 2639558 : }
7662 :
7663 135285 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7664 135285 : self.tl.update_current_logical_size(delta)
7665 135285 : }
7666 : }
7667 :
7668 : // We need TimelineWriter to be send in upcoming conversion of
7669 : // Timeline::layers to tokio::sync::RwLock.
7670 : #[test]
7671 1 : fn is_send() {
7672 1 : fn _assert_send<T: Send>() {}
7673 1 : _assert_send::<TimelineWriter<'_>>();
7674 1 : }
7675 :
7676 : #[cfg(test)]
7677 : mod tests {
7678 : use std::sync::Arc;
7679 :
7680 : use pageserver_api::key::Key;
7681 : use postgres_ffi::PgMajorVersion;
7682 : use std::iter::Iterator;
7683 : use tracing::Instrument;
7684 : use utils::id::TimelineId;
7685 : use utils::lsn::Lsn;
7686 : use wal_decoder::models::value::Value;
7687 :
7688 : use super::HeatMapTimeline;
7689 : use crate::context::RequestContextBuilder;
7690 : use crate::tenant::harness::{TenantHarness, test_img};
7691 : use crate::tenant::layer_map::LayerMap;
7692 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7693 : use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
7694 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7695 : use crate::tenant::{PreviousHeatmap, Timeline};
7696 :
7697 5 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7698 5 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7699 5 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7700 25 : for (l, r) in lhs_rhs {
7701 20 : assert_eq!(l.name, r.name);
7702 20 : assert_eq!(l.metadata, r.metadata);
7703 : }
7704 5 : }
7705 :
7706 : #[tokio::test]
7707 1 : async fn test_heatmap_generation() {
7708 1 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7709 :
7710 1 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7711 1 : Lsn(0x10)..Lsn(0x20),
7712 1 : vec![(
7713 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7714 1 : Lsn(0x11),
7715 1 : Value::Image(test_img("foo")),
7716 1 : )],
7717 : );
7718 1 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7719 1 : Lsn(0x10)..Lsn(0x20),
7720 1 : vec![(
7721 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7722 1 : Lsn(0x11),
7723 1 : Value::Image(test_img("foo")),
7724 1 : )],
7725 : );
7726 1 : let l0_delta = DeltaLayerTestDesc::new(
7727 1 : Lsn(0x20)..Lsn(0x30),
7728 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7729 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7730 1 : vec![(
7731 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7732 1 : Lsn(0x25),
7733 1 : Value::Image(test_img("foo")),
7734 1 : )],
7735 : );
7736 1 : let delta_layers = vec![
7737 1 : covered_delta.clone(),
7738 1 : visible_delta.clone(),
7739 1 : l0_delta.clone(),
7740 : ];
7741 :
7742 1 : let image_layer = (
7743 1 : Lsn(0x40),
7744 1 : vec![(
7745 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7746 1 : test_img("bar"),
7747 1 : )],
7748 1 : );
7749 1 : let image_layers = vec![image_layer];
7750 :
7751 1 : let (tenant, ctx) = harness.load().await;
7752 1 : let timeline = tenant
7753 1 : .create_test_timeline_with_layers(
7754 1 : TimelineId::generate(),
7755 1 : Lsn(0x10),
7756 1 : PgMajorVersion::PG14,
7757 1 : &ctx,
7758 1 : Vec::new(), // in-memory layers
7759 1 : delta_layers,
7760 1 : image_layers,
7761 1 : Lsn(0x100),
7762 1 : )
7763 1 : .await
7764 1 : .unwrap();
7765 1 : let ctx = &ctx.with_scope_timeline(&timeline);
7766 :
7767 : // Layer visibility is an input to heatmap generation, so refresh it first
7768 1 : timeline.update_layer_visibility().await.unwrap();
7769 :
7770 1 : let heatmap = timeline
7771 1 : .generate_heatmap()
7772 1 : .await
7773 1 : .expect("Infallible while timeline is not shut down");
7774 :
7775 1 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
7776 :
7777 : // L0 should come last
7778 1 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
7779 1 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
7780 :
7781 1 : let mut last_lsn = Lsn::MAX;
7782 5 : for layer in heatmap_layers {
7783 : // Covered layer should be omitted
7784 4 : assert!(layer.name != covered_delta.layer_name());
7785 :
7786 4 : let layer_lsn = match &layer.name {
7787 2 : LayerName::Delta(d) => d.lsn_range.end,
7788 2 : LayerName::Image(i) => i.lsn,
7789 : };
7790 :
7791 : // Apart from L0s, newest Layers should come first
7792 4 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
7793 3 : assert!(layer_lsn <= last_lsn);
7794 3 : last_lsn = layer_lsn;
7795 1 : }
7796 : }
7797 :
7798 : // Evict all the layers and stash the old heatmap in the timeline.
7799 : // This simulates a migration to a cold secondary location.
7800 :
7801 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
7802 1 : let mut all_layers = Vec::new();
7803 1 : let forever = std::time::Duration::from_secs(120);
7804 5 : for layer in guard.likely_resident_layers() {
7805 5 : all_layers.push(layer.clone());
7806 5 : layer.evict_and_wait(forever).await.unwrap();
7807 : }
7808 1 : drop(guard);
7809 :
7810 1 : timeline
7811 1 : .previous_heatmap
7812 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
7813 1 : heatmap: heatmap.clone(),
7814 1 : read_at: std::time::Instant::now(),
7815 1 : end_lsn: None,
7816 1 : })));
7817 :
7818 : // Generate a new heatmap and assert that it contains the same layers as the old one.
7819 1 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
7820 1 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
7821 :
7822 : // Download each layer one by one. Generate the heatmap at each step and check
7823 : // that it's stable.
7824 6 : for layer in all_layers {
7825 5 : if layer.visibility() == LayerVisibilityHint::Covered {
7826 1 : continue;
7827 4 : }
7828 1 :
7829 4 : eprintln!("Downloading {layer} and re-generating heatmap");
7830 1 :
7831 4 : let ctx = &RequestContextBuilder::from(ctx)
7832 4 : .download_behavior(crate::context::DownloadBehavior::Download)
7833 4 : .attached_child();
7834 1 :
7835 4 : let _resident = layer
7836 4 : .download_and_keep_resident(ctx)
7837 4 : .instrument(tracing::info_span!(
7838 4 : parent: None,
7839 1 : "download_layer",
7840 1 : tenant_id = %timeline.tenant_shard_id.tenant_id,
7841 1 : shard_id = %timeline.tenant_shard_id.shard_slug(),
7842 1 : timeline_id = %timeline.timeline_id
7843 1 : ))
7844 4 : .await
7845 4 : .unwrap();
7846 1 :
7847 4 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
7848 4 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
7849 1 : }
7850 1 :
7851 1 : // Everything from the post-migration heatmap is now resident.
7852 1 : // Check that we drop it from memory.
7853 1 : assert!(matches!(
7854 1 : timeline.previous_heatmap.load().as_deref(),
7855 1 : Some(PreviousHeatmap::Obsolete)
7856 1 : ));
7857 1 : }
7858 :
7859 : #[tokio::test]
7860 1 : async fn test_previous_heatmap_obsoletion() {
7861 1 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
7862 1 : .await
7863 1 : .unwrap();
7864 :
7865 1 : let l0_delta = DeltaLayerTestDesc::new(
7866 1 : Lsn(0x20)..Lsn(0x30),
7867 1 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7868 1 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7869 1 : vec![(
7870 1 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7871 1 : Lsn(0x25),
7872 1 : Value::Image(test_img("foo")),
7873 1 : )],
7874 : );
7875 :
7876 1 : let image_layer = (
7877 1 : Lsn(0x40),
7878 1 : vec![(
7879 1 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7880 1 : test_img("bar"),
7881 1 : )],
7882 1 : );
7883 :
7884 1 : let delta_layers = vec![l0_delta];
7885 1 : let image_layers = vec![image_layer];
7886 :
7887 1 : let (tenant, ctx) = harness.load().await;
7888 1 : let timeline = tenant
7889 1 : .create_test_timeline_with_layers(
7890 1 : TimelineId::generate(),
7891 1 : Lsn(0x10),
7892 1 : PgMajorVersion::PG14,
7893 1 : &ctx,
7894 1 : Vec::new(), // in-memory layers
7895 1 : delta_layers,
7896 1 : image_layers,
7897 1 : Lsn(0x100),
7898 1 : )
7899 1 : .await
7900 1 : .unwrap();
7901 :
7902 : // Layer visibility is an input to heatmap generation, so refresh it first
7903 1 : timeline.update_layer_visibility().await.unwrap();
7904 :
7905 1 : let heatmap = timeline
7906 1 : .generate_heatmap()
7907 1 : .await
7908 1 : .expect("Infallible while timeline is not shut down");
7909 :
7910 : // Both layers should be in the heatmap
7911 1 : assert!(heatmap.all_layers().count() > 0);
7912 :
7913 : // Now simulate a migration.
7914 1 : timeline
7915 1 : .previous_heatmap
7916 1 : .store(Some(Arc::new(PreviousHeatmap::Active {
7917 1 : heatmap: heatmap.clone(),
7918 1 : read_at: std::time::Instant::now(),
7919 1 : end_lsn: None,
7920 1 : })));
7921 :
7922 : // Evict all the layers in the previous heatmap
7923 1 : let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
7924 1 : let forever = std::time::Duration::from_secs(120);
7925 3 : for layer in guard.likely_resident_layers() {
7926 3 : layer.evict_and_wait(forever).await.unwrap();
7927 : }
7928 1 : drop(guard);
7929 :
7930 : // Generate a new heatmap and check that the previous heatmap
7931 : // has been marked obsolete.
7932 1 : let post_eviction_heatmap = timeline
7933 1 : .generate_heatmap()
7934 1 : .await
7935 1 : .expect("Infallible while timeline is not shut down");
7936 :
7937 1 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
7938 1 : assert!(matches!(
7939 1 : timeline.previous_heatmap.load().as_deref(),
7940 1 : Some(PreviousHeatmap::Obsolete)
7941 1 : ));
7942 1 : }
7943 :
7944 : #[tokio::test]
7945 1 : async fn two_layer_eviction_attempts_at_the_same_time() {
7946 1 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
7947 1 : .await
7948 1 : .unwrap();
7949 :
7950 1 : let (tenant, ctx) = harness.load().await;
7951 1 : let timeline = tenant
7952 1 : .create_test_timeline(
7953 1 : TimelineId::generate(),
7954 1 : Lsn(0x10),
7955 1 : PgMajorVersion::PG14,
7956 1 : &ctx,
7957 1 : )
7958 1 : .await
7959 1 : .unwrap();
7960 :
7961 1 : let layer = find_some_layer(&timeline).await;
7962 1 : let layer = layer
7963 1 : .keep_resident()
7964 1 : .await
7965 1 : .expect("no download => no downloading errors")
7966 1 : .drop_eviction_guard();
7967 :
7968 1 : let forever = std::time::Duration::from_secs(120);
7969 :
7970 1 : let first = layer.evict_and_wait(forever);
7971 1 : let second = layer.evict_and_wait(forever);
7972 :
7973 1 : let (first, second) = tokio::join!(first, second);
7974 :
7975 1 : let res = layer.keep_resident().await;
7976 1 : assert!(res.is_none(), "{res:?}");
7977 :
7978 1 : match (first, second) {
7979 1 : (Ok(()), Ok(())) => {
7980 1 : // because there are no more timeline locks being taken on eviction path, we can
7981 1 : // witness all three outcomes here.
7982 1 : }
7983 1 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7984 0 : // if one completes before the other, this is fine just as well.
7985 0 : }
7986 1 : other => unreachable!("unexpected {:?}", other),
7987 1 : }
7988 1 : }
7989 :
7990 1 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7991 1 : let layers = timeline
7992 1 : .layers
7993 1 : .read(LayerManagerLockHolder::GetLayerMapInfo)
7994 1 : .await;
7995 1 : let desc = layers
7996 1 : .layer_map()
7997 1 : .unwrap()
7998 1 : .iter_historic_layers()
7999 1 : .next()
8000 1 : .expect("must find one layer to evict");
8001 :
8002 1 : layers.get_from_desc(&desc)
8003 1 : }
8004 : }
|
