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 std::array;
18 : use std::cmp::{max, min};
19 : use std::collections::btree_map::Entry;
20 : use std::collections::{BTreeMap, HashMap, HashSet};
21 : use std::ops::{ControlFlow, Deref, Range};
22 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
23 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
24 : use std::time::{Duration, Instant, SystemTime};
25 :
26 : use crate::PERF_TRACE_TARGET;
27 : use crate::walredo::RedoAttemptType;
28 : use anyhow::{Context, Result, anyhow, bail, ensure};
29 : use arc_swap::{ArcSwap, ArcSwapOption};
30 : use bytes::Bytes;
31 : use camino::Utf8Path;
32 : use chrono::{DateTime, Utc};
33 : use compaction::{CompactionOutcome, GcCompactionCombinedSettings};
34 : use enumset::EnumSet;
35 : use fail::fail_point;
36 : use futures::stream::FuturesUnordered;
37 : use futures::{FutureExt, StreamExt};
38 : use handle::ShardTimelineId;
39 : use layer_manager::Shutdown;
40 : use offload::OffloadError;
41 : use once_cell::sync::Lazy;
42 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
43 : use pageserver_api::key::{
44 : KEY_SIZE, Key, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
45 : SPARSE_RANGE,
46 : };
47 : use pageserver_api::keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning};
48 : use pageserver_api::models::{
49 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
50 : DetachBehavior, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
51 : EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, LsnLease, PageTraceEvent, RelSizeMigration,
52 : TimelineState,
53 : };
54 : use pageserver_api::reltag::{BlockNumber, RelTag};
55 : use pageserver_api::shard::{ShardIdentity, ShardIndex, ShardNumber, TenantShardId};
56 : #[cfg(test)]
57 : use pageserver_api::value::Value;
58 : use postgres_connection::PgConnectionConfig;
59 : use postgres_ffi::v14::xlog_utils;
60 : use postgres_ffi::{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, fs_ext, pausable_failpoint};
81 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
82 :
83 : use self::delete::DeleteTimelineFlow;
84 : pub(super) use self::eviction_task::EvictionTaskTenantState;
85 : use self::eviction_task::EvictionTaskTimelineState;
86 : use self::layer_manager::LayerManager;
87 : use self::logical_size::LogicalSize;
88 : use self::walreceiver::{WalReceiver, WalReceiverConf};
89 : use super::remote_timeline_client::RemoteTimelineClient;
90 : use super::remote_timeline_client::index::{GcCompactionState, IndexPart};
91 : use super::secondary::heatmap::HeatMapLayer;
92 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
93 : use super::tasks::log_compaction_error;
94 : use super::upload_queue::NotInitialized;
95 : use super::{
96 : AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
97 : debug_assert_current_span_has_tenant_and_timeline_id,
98 : };
99 : use crate::aux_file::AuxFileSizeEstimator;
100 : use crate::config::PageServerConf;
101 : use crate::context::{
102 : DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
103 : };
104 : use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
105 : use crate::keyspace::{KeyPartitioning, KeySpace};
106 : use crate::l0_flush::{self, L0FlushGlobalState};
107 : use crate::metrics::{
108 : DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_AMORTIZED_GLOBAL, LAYERS_PER_READ_BATCH_GLOBAL,
109 : LAYERS_PER_READ_GLOBAL, ScanLatencyOngoingRecording, TimelineMetrics,
110 : };
111 : use crate::page_service::TenantManagerTypes;
112 : use crate::pgdatadir_mapping::{
113 : CalculateLogicalSizeError, CollectKeySpaceError, DirectoryKind, LsnForTimestamp,
114 : MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
115 : };
116 : use crate::task_mgr::TaskKind;
117 : use crate::tenant::config::AttachmentMode;
118 : use crate::tenant::gc_result::GcResult;
119 : use crate::tenant::layer_map::LayerMap;
120 : use crate::tenant::metadata::TimelineMetadata;
121 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
122 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
123 : use crate::tenant::storage_layer::{
124 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
125 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
126 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
127 : ValueReconstructState, ValuesReconstructState,
128 : };
129 : use crate::tenant::tasks::BackgroundLoopKind;
130 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
131 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
132 : use crate::walingest::WalLagCooldown;
133 : use crate::{ZERO_PAGE, task_mgr, walredo};
134 :
135 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
136 : pub(crate) enum FlushLoopState {
137 : NotStarted,
138 : Running {
139 : #[cfg(test)]
140 : expect_initdb_optimization: bool,
141 : #[cfg(test)]
142 : initdb_optimization_count: usize,
143 : },
144 : Exited,
145 : }
146 :
147 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
148 : pub enum ImageLayerCreationMode {
149 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
150 : Try,
151 : /// Force creating the image layers if possible. For now, no image layers will be created
152 : /// for metadata keys. Used in compaction code path with force flag enabled.
153 : Force,
154 : /// Initial ingestion of the data, and no data should be dropped in this function. This
155 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
156 : /// code path.
157 : Initial,
158 : }
159 :
160 : #[derive(Clone, Debug, Default)]
161 : pub enum LastImageLayerCreationStatus {
162 : Incomplete {
163 : /// The last key of the partition (exclusive) that was processed in the last
164 : /// image layer creation attempt. We will continue from this key in the next
165 : /// attempt.
166 : last_key: Key,
167 : },
168 : Complete,
169 : #[default]
170 : Initial,
171 : }
172 :
173 : impl std::fmt::Display for ImageLayerCreationMode {
174 1156 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
175 1156 : write!(f, "{:?}", self)
176 1156 : }
177 : }
178 :
179 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
180 : /// Can be removed after all refactors are done.
181 56 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
182 56 : drop(rlock)
183 56 : }
184 :
185 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
186 : /// Can be removed after all refactors are done.
187 1212 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
188 1212 : drop(rlock)
189 1212 : }
190 :
191 : /// The outward-facing resources required to build a Timeline
192 : pub struct TimelineResources {
193 : pub remote_client: RemoteTimelineClient,
194 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
195 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
196 : pub l0_compaction_trigger: Arc<Notify>,
197 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
198 : }
199 :
200 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
201 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
202 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
203 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
204 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
205 : pub(crate) struct RelSizeCache {
206 : pub(crate) complete_as_of: Lsn,
207 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
208 : }
209 :
210 : pub struct Timeline {
211 : pub(crate) conf: &'static PageServerConf,
212 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
213 :
214 : myself: Weak<Self>,
215 :
216 : pub(crate) tenant_shard_id: TenantShardId,
217 : pub timeline_id: TimelineId,
218 :
219 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
220 : /// Never changes for the lifetime of this [`Timeline`] object.
221 : ///
222 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
223 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
224 : pub(crate) generation: Generation,
225 :
226 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
227 : /// to shards, and is constant through the lifetime of this Timeline.
228 : shard_identity: ShardIdentity,
229 :
230 : pub pg_version: u32,
231 :
232 : /// The tuple has two elements.
233 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
234 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
235 : ///
236 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
237 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
238 : ///
239 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
240 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
241 : /// `PersistentLayerDesc`'s.
242 : ///
243 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
244 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
245 : /// runtime, e.g., during page reconstruction.
246 : ///
247 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
248 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
249 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
250 :
251 : last_freeze_at: AtomicLsn,
252 : // Atomic would be more appropriate here.
253 : last_freeze_ts: RwLock<Instant>,
254 :
255 : pub(crate) standby_horizon: AtomicLsn,
256 :
257 : // WAL redo manager. `None` only for broken tenants.
258 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
259 :
260 : /// Remote storage client.
261 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
262 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
263 :
264 : // What page versions do we hold in the repository? If we get a
265 : // request > last_record_lsn, we need to wait until we receive all
266 : // the WAL up to the request. The SeqWait provides functions for
267 : // that. TODO: If we get a request for an old LSN, such that the
268 : // versions have already been garbage collected away, we should
269 : // throw an error, but we don't track that currently.
270 : //
271 : // last_record_lsn.load().last points to the end of last processed WAL record.
272 : //
273 : // We also remember the starting point of the previous record in
274 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
275 : // first WAL record when the node is started up. But here, we just
276 : // keep track of it.
277 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
278 :
279 : // All WAL records have been processed and stored durably on files on
280 : // local disk, up to this LSN. On crash and restart, we need to re-process
281 : // the WAL starting from this point.
282 : //
283 : // Some later WAL records might have been processed and also flushed to disk
284 : // already, so don't be surprised to see some, but there's no guarantee on
285 : // them yet.
286 : disk_consistent_lsn: AtomicLsn,
287 :
288 : // Parent timeline that this timeline was branched from, and the LSN
289 : // of the branch point.
290 : ancestor_timeline: Option<Arc<Timeline>>,
291 : ancestor_lsn: Lsn,
292 :
293 : // The LSN of gc-compaction that was last applied to this timeline.
294 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
295 :
296 : pub(crate) metrics: Arc<TimelineMetrics>,
297 :
298 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
299 : // in `crate::page_service` writes these metrics.
300 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
301 :
302 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
303 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
304 :
305 : /// Ensures layers aren't frozen by checkpointer between
306 : /// [`Timeline::get_layer_for_write`] and layer reads.
307 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
308 : /// Must always be acquired before the layer map/individual layer lock
309 : /// to avoid deadlock.
310 : ///
311 : /// The state is cleared upon freezing.
312 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
313 :
314 : /// Used to avoid multiple `flush_loop` tasks running
315 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
316 :
317 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
318 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
319 : /// The flush cycle counter is sent back on the layer_flush_done channel when
320 : /// the flush finishes. You can use that to wait for the flush to finish.
321 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
322 : /// read by whoever sends an update
323 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
324 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
325 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
326 :
327 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
328 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
329 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
330 : // the planned GC cutoff.
331 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
332 :
333 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
334 :
335 : // List of child timelines and their branch points. This is needed to avoid
336 : // garbage collecting data that is still needed by the child timelines.
337 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
338 :
339 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
340 :
341 : // It may change across major versions so for simplicity
342 : // keep it after running initdb for a timeline.
343 : // It is needed in checks when we want to error on some operations
344 : // when they are requested for pre-initdb lsn.
345 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
346 : // though let's keep them both for better error visibility.
347 : pub initdb_lsn: Lsn,
348 :
349 : /// The repartitioning result. Allows a single writer and multiple readers.
350 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
351 :
352 : /// Configuration: how often should the partitioning be recalculated.
353 : repartition_threshold: u64,
354 :
355 : last_image_layer_creation_check_at: AtomicLsn,
356 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
357 :
358 : /// Current logical size of the "datadir", at the last LSN.
359 : current_logical_size: LogicalSize,
360 :
361 : /// Information about the last processed message by the WAL receiver,
362 : /// or None if WAL receiver has not received anything for this timeline
363 : /// yet.
364 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
365 : pub walreceiver: Mutex<Option<WalReceiver>>,
366 :
367 : /// Relation size cache
368 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
369 :
370 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
371 :
372 : state: watch::Sender<TimelineState>,
373 :
374 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
375 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
376 : pub delete_progress: TimelineDeleteProgress,
377 :
378 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
379 :
380 : /// Load or creation time information about the disk_consistent_lsn and when the loading
381 : /// happened. Used for consumption metrics.
382 : pub(crate) loaded_at: (Lsn, SystemTime),
383 :
384 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
385 : pub(crate) gate: Gate,
386 :
387 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
388 : /// to the timeline should drop out when this token fires.
389 : pub(crate) cancel: CancellationToken,
390 :
391 : /// Make sure we only have one running compaction at a time in tests.
392 : ///
393 : /// Must only be taken in two places:
394 : /// - [`Timeline::compact`] (this file)
395 : /// - [`delete::delete_local_timeline_directory`]
396 : ///
397 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
398 : compaction_lock: tokio::sync::Mutex<()>,
399 :
400 : /// If true, the last compaction failed.
401 : compaction_failed: AtomicBool,
402 :
403 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
404 : l0_compaction_trigger: Arc<Notify>,
405 :
406 : /// Make sure we only have one running gc at a time.
407 : ///
408 : /// Must only be taken in two places:
409 : /// - [`Timeline::gc`] (this file)
410 : /// - [`delete::delete_local_timeline_directory`]
411 : ///
412 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
413 : gc_lock: tokio::sync::Mutex<()>,
414 :
415 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
416 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
417 :
418 : /// Size estimator for aux file v2
419 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
420 :
421 : /// Some test cases directly place keys into the timeline without actually modifying the directory
422 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
423 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
424 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
425 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
426 : #[cfg(test)]
427 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
428 :
429 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
430 :
431 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
432 :
433 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
434 :
435 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
436 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
437 :
438 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
439 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
440 :
441 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
442 :
443 : /// May host a background Tokio task which downloads all the layers from the current
444 : /// heatmap on demand.
445 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
446 :
447 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
448 :
449 : wait_lsn_log_slow: tokio::sync::Semaphore,
450 : }
451 :
452 : pub(crate) enum PreviousHeatmap {
453 : Active {
454 : heatmap: HeatMapTimeline,
455 : read_at: std::time::Instant,
456 : // End LSN covered by the heatmap if known
457 : end_lsn: Option<Lsn>,
458 : },
459 : Obsolete,
460 : }
461 :
462 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
463 :
464 : pub struct WalReceiverInfo {
465 : pub wal_source_connconf: PgConnectionConfig,
466 : pub last_received_msg_lsn: Lsn,
467 : pub last_received_msg_ts: u128,
468 : }
469 :
470 : /// Information about how much history needs to be retained, needed by
471 : /// Garbage Collection.
472 : #[derive(Default)]
473 : pub(crate) struct GcInfo {
474 : /// Specific LSNs that are needed.
475 : ///
476 : /// Currently, this includes all points where child branches have
477 : /// been forked off from. In the future, could also include
478 : /// explicit user-defined snapshot points.
479 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
480 :
481 : /// The cutoff coordinates, which are combined by selecting the minimum.
482 : pub(crate) cutoffs: GcCutoffs,
483 :
484 : /// Leases granted to particular LSNs.
485 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
486 :
487 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
488 : pub(crate) within_ancestor_pitr: bool,
489 : }
490 :
491 : impl GcInfo {
492 616 : pub(crate) fn min_cutoff(&self) -> Lsn {
493 616 : self.cutoffs.select_min()
494 616 : }
495 :
496 476 : pub(super) fn insert_child(
497 476 : &mut self,
498 476 : child_id: TimelineId,
499 476 : child_lsn: Lsn,
500 476 : is_offloaded: MaybeOffloaded,
501 476 : ) {
502 476 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
503 476 : self.retain_lsns.sort_by_key(|i| i.0);
504 476 : }
505 :
506 8 : pub(super) fn remove_child_maybe_offloaded(
507 8 : &mut self,
508 8 : child_id: TimelineId,
509 8 : maybe_offloaded: MaybeOffloaded,
510 8 : ) -> bool {
511 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
512 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
513 8 : let mut removed = false;
514 12 : self.retain_lsns.retain(|i| {
515 12 : if removed {
516 4 : return true;
517 8 : }
518 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
519 8 : removed |= remove;
520 8 : !remove
521 12 : });
522 8 : removed
523 8 : }
524 :
525 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
526 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
527 8 : }
528 :
529 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
530 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
531 0 : }
532 476 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
533 476 : self.leases.contains_key(&lsn)
534 476 : }
535 : }
536 :
537 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
538 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
539 : /// between time-based and space-based retention for observability and consumption metrics purposes.
540 : #[derive(Debug, Clone)]
541 : pub(crate) struct GcCutoffs {
542 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
543 : /// history we must keep to retain a specified number of bytes of WAL.
544 : pub(crate) space: Lsn,
545 :
546 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates how much
547 : /// history we must keep to enable reading back at least the PITR interval duration.
548 : pub(crate) time: Lsn,
549 : }
550 :
551 : impl Default for GcCutoffs {
552 924 : fn default() -> Self {
553 924 : Self {
554 924 : space: Lsn::INVALID,
555 924 : time: Lsn::INVALID,
556 924 : }
557 924 : }
558 : }
559 :
560 : impl GcCutoffs {
561 616 : fn select_min(&self) -> Lsn {
562 616 : std::cmp::min(self.space, self.time)
563 616 : }
564 : }
565 :
566 : pub(crate) struct TimelineVisitOutcome {
567 : completed_keyspace: KeySpace,
568 : image_covered_keyspace: KeySpace,
569 : }
570 :
571 : /// An error happened in a get() operation.
572 : #[derive(thiserror::Error, Debug)]
573 : pub(crate) enum PageReconstructError {
574 : #[error(transparent)]
575 : Other(anyhow::Error),
576 :
577 : #[error("Ancestor LSN wait error: {0}")]
578 : AncestorLsnTimeout(WaitLsnError),
579 :
580 : #[error("timeline shutting down")]
581 : Cancelled,
582 :
583 : /// An error happened replaying WAL records
584 : #[error(transparent)]
585 : WalRedo(anyhow::Error),
586 :
587 : #[error("{0}")]
588 : MissingKey(MissingKeyError),
589 : }
590 :
591 : impl From<anyhow::Error> for PageReconstructError {
592 0 : fn from(value: anyhow::Error) -> Self {
593 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
594 0 : match value.downcast::<PageReconstructError>() {
595 0 : Ok(pre) => pre,
596 0 : Err(other) => PageReconstructError::Other(other),
597 : }
598 0 : }
599 : }
600 :
601 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
602 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
603 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
604 0 : }
605 : }
606 :
607 : impl From<layer_manager::Shutdown> for PageReconstructError {
608 0 : fn from(_: layer_manager::Shutdown) -> Self {
609 0 : PageReconstructError::Cancelled
610 0 : }
611 : }
612 :
613 : impl GetVectoredError {
614 : #[cfg(test)]
615 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
616 12 : matches!(self, Self::MissingKey(_))
617 12 : }
618 : }
619 :
620 : impl From<layer_manager::Shutdown> for GetVectoredError {
621 0 : fn from(_: layer_manager::Shutdown) -> Self {
622 0 : GetVectoredError::Cancelled
623 0 : }
624 : }
625 :
626 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
627 : /// only and not used by the "real read path".
628 : pub enum ReadPathLayerId {
629 : PersistentLayer(PersistentLayerKey),
630 : InMemoryLayer(Range<Lsn>),
631 : }
632 :
633 : impl std::fmt::Display for ReadPathLayerId {
634 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
635 0 : match self {
636 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
637 0 : ReadPathLayerId::InMemoryLayer(range) => {
638 0 : write!(f, "in-mem {}..{}", range.start, range.end)
639 : }
640 : }
641 0 : }
642 : }
643 : pub struct ReadPath {
644 : keyspace: KeySpace,
645 : lsn: Lsn,
646 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
647 : }
648 :
649 : impl ReadPath {
650 1255534 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
651 1255534 : Self {
652 1255534 : keyspace,
653 1255534 : lsn,
654 1255534 : path: Vec::new(),
655 1255534 : }
656 1255534 : }
657 :
658 1693889 : pub fn record_layer_visit(
659 1693889 : &mut self,
660 1693889 : layer_to_read: &ReadableLayer,
661 1693889 : keyspace_to_read: &KeySpace,
662 1693889 : lsn_range: &Range<Lsn>,
663 1693889 : ) {
664 1693889 : let id = match layer_to_read {
665 480361 : ReadableLayer::PersistentLayer(layer) => {
666 480361 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
667 : }
668 1213528 : ReadableLayer::InMemoryLayer(layer) => {
669 1213528 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
670 : }
671 : };
672 1693889 : self.path
673 1693889 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
674 1693889 : }
675 : }
676 :
677 : impl std::fmt::Display for ReadPath {
678 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
679 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
680 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
681 0 : writeln!(
682 0 : f,
683 0 : "{}: {} {}..{} {}",
684 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
685 0 : )?;
686 : }
687 0 : Ok(())
688 0 : }
689 : }
690 :
691 : #[derive(thiserror::Error)]
692 : pub struct MissingKeyError {
693 : key: Key,
694 : shard: ShardNumber,
695 : cont_lsn: Lsn,
696 : request_lsn: Lsn,
697 : ancestor_lsn: Option<Lsn>,
698 : /// Debug information about the read path if there's an error
699 : read_path: Option<ReadPath>,
700 : backtrace: Option<std::backtrace::Backtrace>,
701 : }
702 :
703 : impl std::fmt::Debug for MissingKeyError {
704 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
705 0 : write!(f, "{}", self)
706 0 : }
707 : }
708 :
709 : impl std::fmt::Display for MissingKeyError {
710 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
711 0 : write!(
712 0 : f,
713 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
714 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
715 0 : )?;
716 :
717 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
718 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
719 0 : }
720 :
721 0 : if let Some(ref read_path) = self.read_path {
722 0 : write!(f, "\n{}", read_path)?;
723 0 : }
724 :
725 0 : if let Some(ref backtrace) = self.backtrace {
726 0 : write!(f, "\n{}", backtrace)?;
727 0 : }
728 :
729 0 : Ok(())
730 0 : }
731 : }
732 :
733 : impl PageReconstructError {
734 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
735 0 : pub(crate) fn is_stopping(&self) -> bool {
736 : use PageReconstructError::*;
737 0 : match self {
738 0 : Cancelled => true,
739 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
740 : }
741 0 : }
742 : }
743 :
744 : #[derive(thiserror::Error, Debug)]
745 : pub(crate) enum CreateImageLayersError {
746 : #[error("timeline shutting down")]
747 : Cancelled,
748 :
749 : #[error("read failed")]
750 : GetVectoredError(#[source] GetVectoredError),
751 :
752 : #[error("reconstruction failed")]
753 : PageReconstructError(#[source] PageReconstructError),
754 :
755 : #[error(transparent)]
756 : Other(#[from] anyhow::Error),
757 : }
758 :
759 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
760 0 : fn from(_: layer_manager::Shutdown) -> Self {
761 0 : CreateImageLayersError::Cancelled
762 0 : }
763 : }
764 :
765 : #[derive(thiserror::Error, Debug, Clone)]
766 : pub(crate) enum FlushLayerError {
767 : /// Timeline cancellation token was cancelled
768 : #[error("timeline shutting down")]
769 : Cancelled,
770 :
771 : /// We tried to flush a layer while the Timeline is in an unexpected state
772 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
773 : NotRunning(FlushLoopState),
774 :
775 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
776 : // loop via a watch channel, where we can only borrow it.
777 : #[error("create image layers (shared)")]
778 : CreateImageLayersError(Arc<CreateImageLayersError>),
779 :
780 : #[error("other (shared)")]
781 : Other(#[from] Arc<anyhow::Error>),
782 : }
783 :
784 : impl FlushLayerError {
785 : // When crossing from generic anyhow errors to this error type, we explicitly check
786 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
787 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
788 0 : let cancelled = timeline.cancel.is_cancelled()
789 : // The upload queue might have been shut down before the official cancellation of the timeline.
790 0 : || err
791 0 : .downcast_ref::<NotInitialized>()
792 0 : .map(NotInitialized::is_stopping)
793 0 : .unwrap_or_default();
794 0 : if cancelled {
795 0 : Self::Cancelled
796 : } else {
797 0 : Self::Other(Arc::new(err))
798 : }
799 0 : }
800 : }
801 :
802 : impl From<layer_manager::Shutdown> for FlushLayerError {
803 0 : fn from(_: layer_manager::Shutdown) -> Self {
804 0 : FlushLayerError::Cancelled
805 0 : }
806 : }
807 :
808 : #[derive(thiserror::Error, Debug)]
809 : pub(crate) enum GetVectoredError {
810 : #[error("timeline shutting down")]
811 : Cancelled,
812 :
813 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
814 : Oversized(u64),
815 :
816 : #[error("requested at invalid LSN: {0}")]
817 : InvalidLsn(Lsn),
818 :
819 : #[error("requested key not found: {0}")]
820 : MissingKey(MissingKeyError),
821 :
822 : #[error("ancestry walk")]
823 : GetReadyAncestorError(#[source] GetReadyAncestorError),
824 :
825 : #[error(transparent)]
826 : Other(#[from] anyhow::Error),
827 : }
828 :
829 : impl From<GetReadyAncestorError> for GetVectoredError {
830 4 : fn from(value: GetReadyAncestorError) -> Self {
831 : use GetReadyAncestorError::*;
832 4 : match value {
833 0 : Cancelled => GetVectoredError::Cancelled,
834 : AncestorLsnTimeout(_) | BadState { .. } => {
835 4 : GetVectoredError::GetReadyAncestorError(value)
836 : }
837 : }
838 4 : }
839 : }
840 :
841 : #[derive(thiserror::Error, Debug)]
842 : pub(crate) enum GetReadyAncestorError {
843 : #[error("ancestor LSN wait error")]
844 : AncestorLsnTimeout(#[from] WaitLsnError),
845 :
846 : #[error("bad state on timeline {timeline_id}: {state:?}")]
847 : BadState {
848 : timeline_id: TimelineId,
849 : state: TimelineState,
850 : },
851 :
852 : #[error("cancelled")]
853 : Cancelled,
854 : }
855 :
856 : #[derive(Clone, Copy)]
857 : pub enum LogicalSizeCalculationCause {
858 : Initial,
859 : ConsumptionMetricsSyntheticSize,
860 : EvictionTaskImitation,
861 : TenantSizeHandler,
862 : }
863 :
864 : pub enum GetLogicalSizePriority {
865 : User,
866 : Background,
867 : }
868 :
869 0 : #[derive(Debug, enumset::EnumSetType)]
870 : pub(crate) enum CompactFlags {
871 : ForceRepartition,
872 : ForceImageLayerCreation,
873 : ForceL0Compaction,
874 : OnlyL0Compaction,
875 : EnhancedGcBottomMostCompaction,
876 : DryRun,
877 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
878 : /// the background compaction task, since we want to aggressively compact down L0 to bound
879 : /// read amplification.
880 : ///
881 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
882 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
883 : /// compaction).
884 : YieldForL0,
885 : }
886 :
887 : #[serde_with::serde_as]
888 0 : #[derive(Debug, Clone, serde::Deserialize)]
889 : pub(crate) struct CompactRequest {
890 : pub compact_key_range: Option<CompactKeyRange>,
891 : pub compact_lsn_range: Option<CompactLsnRange>,
892 : /// Whether the compaction job should be scheduled.
893 : #[serde(default)]
894 : pub scheduled: bool,
895 : /// Whether the compaction job should be split across key ranges.
896 : #[serde(default)]
897 : pub sub_compaction: bool,
898 : /// Max job size for each subcompaction job.
899 : pub sub_compaction_max_job_size_mb: Option<u64>,
900 : }
901 :
902 0 : #[derive(Debug, Clone, serde::Deserialize)]
903 : pub(crate) struct MarkInvisibleRequest {
904 : #[serde(default)]
905 : pub is_visible: Option<bool>,
906 : }
907 :
908 : #[derive(Debug, Clone, Default)]
909 : pub(crate) struct CompactOptions {
910 : pub flags: EnumSet<CompactFlags>,
911 : /// If set, the compaction will only compact the key range specified by this option.
912 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
913 : pub compact_key_range: Option<CompactKeyRange>,
914 : /// If set, the compaction will only compact the LSN within this value.
915 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
916 : pub compact_lsn_range: Option<CompactLsnRange>,
917 : /// Enable sub-compaction (split compaction job across key ranges).
918 : /// This option is only used by GC compaction.
919 : pub sub_compaction: bool,
920 : /// Set job size for the GC compaction.
921 : /// This option is only used by GC compaction.
922 : pub sub_compaction_max_job_size_mb: Option<u64>,
923 : }
924 :
925 : impl std::fmt::Debug for Timeline {
926 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
927 0 : write!(f, "Timeline<{}>", self.timeline_id)
928 0 : }
929 : }
930 :
931 : #[derive(thiserror::Error, Debug)]
932 : pub(crate) enum WaitLsnError {
933 : // Called on a timeline which is shutting down
934 : #[error("Shutdown")]
935 : Shutdown,
936 :
937 : // Called on an timeline not in active state or shutting down
938 : #[error("Bad timeline state: {0:?}")]
939 : BadState(TimelineState),
940 :
941 : // Timeout expired while waiting for LSN to catch up with goal.
942 : #[error("{0}")]
943 : Timeout(String),
944 : }
945 :
946 : // The impls below achieve cancellation mapping for errors.
947 : // Perhaps there's a way of achieving this with less cruft.
948 :
949 : impl From<CreateImageLayersError> for CompactionError {
950 0 : fn from(e: CreateImageLayersError) -> Self {
951 0 : match e {
952 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
953 0 : CreateImageLayersError::Other(e) => {
954 0 : CompactionError::Other(e.context("create image layers"))
955 : }
956 0 : _ => CompactionError::Other(e.into()),
957 : }
958 0 : }
959 : }
960 :
961 : impl From<CreateImageLayersError> for FlushLayerError {
962 0 : fn from(e: CreateImageLayersError) -> Self {
963 0 : match e {
964 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
965 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
966 : }
967 0 : }
968 : }
969 :
970 : impl From<PageReconstructError> for CreateImageLayersError {
971 0 : fn from(e: PageReconstructError) -> Self {
972 0 : match e {
973 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
974 0 : _ => CreateImageLayersError::PageReconstructError(e),
975 : }
976 0 : }
977 : }
978 :
979 : impl From<GetVectoredError> for CreateImageLayersError {
980 0 : fn from(e: GetVectoredError) -> Self {
981 0 : match e {
982 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
983 0 : _ => CreateImageLayersError::GetVectoredError(e),
984 : }
985 0 : }
986 : }
987 :
988 : impl From<GetVectoredError> for PageReconstructError {
989 12 : fn from(e: GetVectoredError) -> Self {
990 12 : match e {
991 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
992 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
993 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
994 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
995 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
996 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
997 : }
998 12 : }
999 : }
1000 :
1001 : impl From<GetReadyAncestorError> for PageReconstructError {
1002 4 : fn from(e: GetReadyAncestorError) -> Self {
1003 : use GetReadyAncestorError::*;
1004 4 : match e {
1005 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1006 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1007 0 : Cancelled => PageReconstructError::Cancelled,
1008 : }
1009 4 : }
1010 : }
1011 :
1012 : pub(crate) enum WaitLsnTimeout {
1013 : Custom(Duration),
1014 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1015 : Default,
1016 : }
1017 :
1018 : pub(crate) enum WaitLsnWaiter<'a> {
1019 : Timeline(&'a Timeline),
1020 : Tenant,
1021 : PageService,
1022 : HttpEndpoint,
1023 : }
1024 :
1025 : /// Argument to [`Timeline::shutdown`].
1026 : #[derive(Debug, Clone, Copy)]
1027 : pub(crate) enum ShutdownMode {
1028 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1029 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1030 : ///
1031 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1032 : /// the call to [`Timeline::shutdown`].
1033 : FreezeAndFlush,
1034 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1035 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1036 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1037 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1038 : Reload,
1039 : /// Shut down immediately, without waiting for any open layers to flush.
1040 : Hard,
1041 : }
1042 :
1043 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1044 : enum ImageLayerCreationOutcome {
1045 : /// We generated an image layer
1046 : Generated {
1047 : unfinished_image_layer: ImageLayerWriter,
1048 : },
1049 : /// The key range is empty
1050 : Empty,
1051 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1052 : /// the image layer creation.
1053 : Skip,
1054 : }
1055 :
1056 : /// Public interface functions
1057 : impl Timeline {
1058 : /// Get the LSN where this branch was created
1059 88 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1060 88 : self.ancestor_lsn
1061 88 : }
1062 :
1063 : /// Get the ancestor's timeline id
1064 152 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1065 152 : self.ancestor_timeline
1066 152 : .as_ref()
1067 152 : .map(|ancestor| ancestor.timeline_id)
1068 152 : }
1069 :
1070 : /// Get the ancestor timeline
1071 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1072 4 : self.ancestor_timeline.as_ref()
1073 4 : }
1074 :
1075 : /// Get the bytes written since the PITR cutoff on this branch, and
1076 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1077 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1078 0 : let gc_info = self.gc_info.read().unwrap();
1079 0 : let history = self
1080 0 : .get_last_record_lsn()
1081 0 : .checked_sub(gc_info.cutoffs.time)
1082 0 : .unwrap_or(Lsn(0))
1083 0 : .0;
1084 0 : (history, gc_info.within_ancestor_pitr)
1085 0 : }
1086 :
1087 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1088 1707985 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1089 1707985 : self.applied_gc_cutoff_lsn.read()
1090 1707985 : }
1091 :
1092 : /// Read timeline's planned GC cutoff: this is the logical end of history that users
1093 : /// are allowed to read (based on configured PITR), even if physically we have more history.
1094 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Lsn {
1095 0 : self.gc_info.read().unwrap().cutoffs.time
1096 0 : }
1097 :
1098 : /// Look up given page version.
1099 : ///
1100 : /// If a remote layer file is needed, it is downloaded as part of this
1101 : /// call.
1102 : ///
1103 : /// This method enforces [`Self::pagestream_throttle`] internally.
1104 : ///
1105 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1106 : /// abstraction above this needs to store suitable metadata to track what
1107 : /// data exists with what keys, in separate metadata entries. If a
1108 : /// non-existent key is requested, we may incorrectly return a value from
1109 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1110 : /// non-existing key.
1111 : ///
1112 : /// # Cancel-Safety
1113 : ///
1114 : /// This method is cancellation-safe.
1115 : #[inline(always)]
1116 1215302 : pub(crate) async fn get(
1117 1215302 : &self,
1118 1215302 : key: Key,
1119 1215302 : lsn: Lsn,
1120 1215302 : ctx: &RequestContext,
1121 1215302 : ) -> Result<Bytes, PageReconstructError> {
1122 1215302 : if !lsn.is_valid() {
1123 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1124 1215302 : }
1125 1215302 :
1126 1215302 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1127 1215302 : // already checked the key against the shard_identity when looking up the Timeline from
1128 1215302 : // page_service.
1129 1215302 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1130 :
1131 1215302 : let keyspace = KeySpace {
1132 1215302 : ranges: vec![key..key.next()],
1133 1215302 : };
1134 1215302 :
1135 1215302 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1136 :
1137 1215302 : let vectored_res = self
1138 1215302 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1139 1215302 : .await;
1140 :
1141 1215302 : let key_value = vectored_res?.pop_first();
1142 1215290 : match key_value {
1143 1215266 : Some((got_key, value)) => {
1144 1215266 : if got_key != key {
1145 0 : error!(
1146 0 : "Expected {}, but singular vectored get returned {}",
1147 : key, got_key
1148 : );
1149 0 : Err(PageReconstructError::Other(anyhow!(
1150 0 : "Singular vectored get returned wrong key"
1151 0 : )))
1152 : } else {
1153 1215266 : value
1154 : }
1155 : }
1156 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1157 24 : key,
1158 24 : shard: self.shard_identity.get_shard_number(&key),
1159 24 : cont_lsn: Lsn(0),
1160 24 : request_lsn: lsn,
1161 24 : ancestor_lsn: None,
1162 24 : backtrace: None,
1163 24 : read_path: None,
1164 24 : })),
1165 : }
1166 1215302 : }
1167 :
1168 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1169 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1170 :
1171 : /// Look up multiple page versions at a given LSN
1172 : ///
1173 : /// This naive implementation will be replaced with a more efficient one
1174 : /// which actually vectorizes the read path.
1175 39480 : pub(crate) async fn get_vectored(
1176 39480 : &self,
1177 39480 : keyspace: KeySpace,
1178 39480 : lsn: Lsn,
1179 39480 : io_concurrency: super::storage_layer::IoConcurrency,
1180 39480 : ctx: &RequestContext,
1181 39480 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1182 39480 : if !lsn.is_valid() {
1183 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1184 39480 : }
1185 39480 :
1186 39480 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1187 39480 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1188 0 : return Err(GetVectoredError::Oversized(key_count));
1189 39480 : }
1190 :
1191 78960 : for range in &keyspace.ranges {
1192 39480 : let mut key = range.start;
1193 79484 : while key != range.end {
1194 40004 : assert!(!self.shard_identity.is_key_disposable(&key));
1195 40004 : key = key.next();
1196 : }
1197 : }
1198 :
1199 39480 : trace!(
1200 0 : "get vectored request for {:?}@{} from task kind {:?}",
1201 0 : keyspace,
1202 0 : lsn,
1203 0 : ctx.task_kind(),
1204 : );
1205 :
1206 39480 : let start = crate::metrics::GET_VECTORED_LATENCY
1207 39480 : .for_task_kind(ctx.task_kind())
1208 39480 : .map(|metric| (metric, Instant::now()));
1209 :
1210 39480 : let res = self
1211 39480 : .get_vectored_impl(
1212 39480 : keyspace.clone(),
1213 39480 : lsn,
1214 39480 : &mut ValuesReconstructState::new(io_concurrency),
1215 39480 : ctx,
1216 39480 : )
1217 39480 : .await;
1218 :
1219 39480 : if let Some((metric, start)) = start {
1220 0 : let elapsed = start.elapsed();
1221 0 : metric.observe(elapsed.as_secs_f64());
1222 39480 : }
1223 :
1224 39480 : res
1225 39480 : }
1226 :
1227 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1228 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1229 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1230 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1231 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1232 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1233 : /// the scan operation will not cause OOM in the future.
1234 24 : pub(crate) async fn scan(
1235 24 : &self,
1236 24 : keyspace: KeySpace,
1237 24 : lsn: Lsn,
1238 24 : ctx: &RequestContext,
1239 24 : io_concurrency: super::storage_layer::IoConcurrency,
1240 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1241 24 : if !lsn.is_valid() {
1242 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1243 24 : }
1244 24 :
1245 24 : trace!(
1246 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1247 0 : keyspace,
1248 0 : lsn,
1249 0 : ctx.task_kind()
1250 : );
1251 :
1252 : // We should generalize this into Keyspace::contains in the future.
1253 48 : for range in &keyspace.ranges {
1254 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1255 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1256 : {
1257 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1258 0 : "only metadata keyspace can be scanned"
1259 0 : )));
1260 24 : }
1261 : }
1262 :
1263 24 : let start = crate::metrics::SCAN_LATENCY
1264 24 : .for_task_kind(ctx.task_kind())
1265 24 : .map(ScanLatencyOngoingRecording::start_recording);
1266 :
1267 24 : let vectored_res = self
1268 24 : .get_vectored_impl(
1269 24 : keyspace.clone(),
1270 24 : lsn,
1271 24 : &mut ValuesReconstructState::new(io_concurrency),
1272 24 : ctx,
1273 24 : )
1274 24 : .await;
1275 :
1276 24 : if let Some(recording) = start {
1277 0 : recording.observe();
1278 24 : }
1279 :
1280 24 : vectored_res
1281 24 : }
1282 :
1283 1255534 : pub(super) async fn get_vectored_impl(
1284 1255534 : &self,
1285 1255534 : keyspace: KeySpace,
1286 1255534 : lsn: Lsn,
1287 1255534 : reconstruct_state: &mut ValuesReconstructState,
1288 1255534 : ctx: &RequestContext,
1289 1255534 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1290 1255534 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1291 1255534 : Some(ReadPath::new(keyspace.clone(), lsn))
1292 : } else {
1293 0 : None
1294 : };
1295 1255534 : reconstruct_state.read_path = read_path;
1296 :
1297 1255534 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1298 0 : RedoAttemptType::LegacyCompaction
1299 : } else {
1300 1255534 : RedoAttemptType::ReadPage
1301 : };
1302 :
1303 1255534 : let traversal_res: Result<(), _> = {
1304 1255534 : let ctx = RequestContextBuilder::from(ctx)
1305 1255534 : .perf_span(|crnt_perf_span| {
1306 0 : info_span!(
1307 : target: PERF_TRACE_TARGET,
1308 0 : parent: crnt_perf_span,
1309 : "PLAN_IO",
1310 : )
1311 1255534 : })
1312 1255534 : .attached_child();
1313 1255534 :
1314 1255534 : self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, &ctx)
1315 1255534 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1316 1255534 : .await
1317 : };
1318 :
1319 1255534 : if let Err(err) = traversal_res {
1320 : // Wait for all the spawned IOs to complete.
1321 : // See comments on `spawn_io` inside `storage_layer` for more details.
1322 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1323 32 : .into_values()
1324 32 : .map(|state| state.collect_pending_ios())
1325 32 : .collect::<FuturesUnordered<_>>();
1326 32 : while collect_futs.next().await.is_some() {}
1327 32 : return Err(err);
1328 1255502 : };
1329 1255502 :
1330 1255502 : let layers_visited = reconstruct_state.get_layers_visited();
1331 1255502 :
1332 1255502 : let ctx = RequestContextBuilder::from(ctx)
1333 1255502 : .perf_span(|crnt_perf_span| {
1334 0 : info_span!(
1335 : target: PERF_TRACE_TARGET,
1336 0 : parent: crnt_perf_span,
1337 : "RECONSTRUCT",
1338 : )
1339 1255502 : })
1340 1255502 : .attached_child();
1341 1255502 :
1342 1255502 : let futs = FuturesUnordered::new();
1343 1336206 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1344 1336206 : futs.push({
1345 1336206 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1346 1336206 : let ctx = RequestContextBuilder::from(&ctx)
1347 1336206 : .perf_span(|crnt_perf_span| {
1348 0 : info_span!(
1349 : target: PERF_TRACE_TARGET,
1350 0 : parent: crnt_perf_span,
1351 : "RECONSTRUCT_KEY",
1352 : key = %key,
1353 : )
1354 1336206 : })
1355 1336206 : .attached_child();
1356 1336206 :
1357 1336206 : async move {
1358 1336206 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1359 :
1360 1336206 : let res = state
1361 1336206 : .collect_pending_ios()
1362 1336206 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1363 0 : info_span!(
1364 : target: PERF_TRACE_TARGET,
1365 0 : parent: crnt_perf_span,
1366 : "WAIT_FOR_IO_COMPLETIONS",
1367 : )
1368 1336206 : })
1369 1336206 : .await;
1370 :
1371 1336206 : let converted = match res {
1372 1336206 : Ok(ok) => ok,
1373 0 : Err(err) => {
1374 0 : return (key, Err(err));
1375 : }
1376 : };
1377 1336206 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1378 1336206 :
1379 1336206 : // The walredo module expects the records to be descending in terms of Lsn.
1380 1336206 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1381 1336206 : debug_assert!(
1382 1336206 : converted
1383 1336206 : .records
1384 1336206 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1385 0 : "{converted:?}"
1386 : );
1387 :
1388 1336206 : let walredo_deltas = converted.num_deltas();
1389 1336206 : let walredo_res = walredo_self
1390 1336206 : .reconstruct_value(key, lsn, converted, redo_attempt_type)
1391 1336206 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1392 0 : info_span!(
1393 : target: PERF_TRACE_TARGET,
1394 0 : parent: crnt_perf_span,
1395 : "WALREDO",
1396 : deltas = %walredo_deltas,
1397 : )
1398 1336206 : })
1399 1336206 : .await;
1400 :
1401 1336206 : (key, walredo_res)
1402 1336206 : }
1403 1336206 : });
1404 1336206 : }
1405 :
1406 1255502 : let results = futs
1407 1255502 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1408 1255502 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1409 1255502 : .await;
1410 :
1411 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1412 : // when they're missing. Instead they are omitted from the resulting btree
1413 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1414 : // to avoid infinite results.
1415 1255502 : if !results.is_empty() {
1416 1255002 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1417 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1418 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1419 0 : LOG_PACER.lock().unwrap().call(|| {
1420 0 : let num_keys = keyspace.total_raw_size();
1421 0 : let num_pages = results.len();
1422 0 : tracing::info!(
1423 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1424 0 : lsn = %lsn,
1425 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1426 : );
1427 0 : });
1428 1255002 : }
1429 :
1430 : // Records the number of layers visited in a few different ways:
1431 : //
1432 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1433 : // layer directly affects its observed latency.
1434 : //
1435 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1436 : // layer visits and access cost.
1437 : //
1438 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1439 : // read amplification after batching.
1440 1255002 : let layers_visited = layers_visited as f64;
1441 1255002 : let avg_layers_visited = layers_visited / results.len() as f64;
1442 1255002 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1443 2591208 : for _ in &results {
1444 1336206 : self.metrics.layers_per_read.observe(layers_visited);
1445 1336206 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1446 1336206 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1447 1336206 : }
1448 500 : }
1449 :
1450 1255502 : Ok(results)
1451 1255534 : }
1452 :
1453 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1454 549092 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1455 549092 : self.last_record_lsn.load().last
1456 549092 : }
1457 :
1458 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1459 0 : self.last_record_lsn.load().prev
1460 0 : }
1461 :
1462 : /// Atomically get both last and prev.
1463 468 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1464 468 : self.last_record_lsn.load()
1465 468 : }
1466 :
1467 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1468 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1469 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1470 0 : self.last_record_lsn.status_receiver()
1471 0 : }
1472 :
1473 916 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1474 916 : self.disk_consistent_lsn.load()
1475 916 : }
1476 :
1477 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1478 : /// not validated with control plane yet.
1479 : /// See [`Self::get_remote_consistent_lsn_visible`].
1480 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1481 0 : self.remote_client.remote_consistent_lsn_projected()
1482 0 : }
1483 :
1484 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1485 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1486 : /// generation validation in the deletion queue.
1487 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1488 0 : self.remote_client.remote_consistent_lsn_visible()
1489 0 : }
1490 :
1491 : /// The sum of the file size of all historic layers in the layer map.
1492 : /// This method makes no distinction between local and remote layers.
1493 : /// Hence, the result **does not represent local filesystem usage**.
1494 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1495 0 : let guard = self.layers.read().await;
1496 0 : guard.layer_size_sum()
1497 0 : }
1498 :
1499 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1500 0 : self.metrics.resident_physical_size_get()
1501 0 : }
1502 :
1503 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1504 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1505 0 : }
1506 :
1507 : ///
1508 : /// Wait until WAL has been received and processed up to this LSN.
1509 : ///
1510 : /// You should call this before any of the other get_* or list_* functions. Calling
1511 : /// those functions with an LSN that has been processed yet is an error.
1512 : ///
1513 451799 : pub(crate) async fn wait_lsn(
1514 451799 : &self,
1515 451799 : lsn: Lsn,
1516 451799 : who_is_waiting: WaitLsnWaiter<'_>,
1517 451799 : timeout: WaitLsnTimeout,
1518 451799 : ctx: &RequestContext, /* Prepare for use by cancellation */
1519 451799 : ) -> Result<(), WaitLsnError> {
1520 451799 : let state = self.current_state();
1521 451799 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1522 0 : return Err(WaitLsnError::Shutdown);
1523 451799 : } else if !matches!(state, TimelineState::Active) {
1524 0 : return Err(WaitLsnError::BadState(state));
1525 451799 : }
1526 451799 :
1527 451799 : if cfg!(debug_assertions) {
1528 451799 : match ctx.task_kind() {
1529 : TaskKind::WalReceiverManager
1530 : | TaskKind::WalReceiverConnectionHandler
1531 : | TaskKind::WalReceiverConnectionPoller => {
1532 0 : let is_myself = match who_is_waiting {
1533 0 : WaitLsnWaiter::Timeline(waiter) => {
1534 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1535 : }
1536 : WaitLsnWaiter::Tenant
1537 : | WaitLsnWaiter::PageService
1538 0 : | WaitLsnWaiter::HttpEndpoint => unreachable!(
1539 0 : "tenant or page_service context are not expected to have task kind {:?}",
1540 0 : ctx.task_kind()
1541 0 : ),
1542 : };
1543 0 : if is_myself {
1544 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1545 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1546 0 : panic!(
1547 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1548 0 : );
1549 0 : }
1550 0 : } else {
1551 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1552 0 : // our walreceiver task can make progress independent of theirs
1553 0 : }
1554 : }
1555 451799 : _ => {}
1556 : }
1557 0 : }
1558 :
1559 451799 : let timeout = match timeout {
1560 0 : WaitLsnTimeout::Custom(t) => t,
1561 451799 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1562 : };
1563 :
1564 451799 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1565 451799 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1566 451799 :
1567 451799 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1568 451799 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1569 451799 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1570 451799 : let log_slow_threshold = Duration::from_secs(1);
1571 451799 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1572 451799 : let log_slow_period = Duration::from_secs(10);
1573 451799 : let mut logging_permit = None;
1574 451799 : let wait_for_timeout = monitor_slow_future(
1575 451799 : log_slow_threshold,
1576 451799 : log_slow_period,
1577 451799 : wait_for_timeout,
1578 451799 : |MonitorSlowFutureCallback {
1579 : ready,
1580 : is_slow,
1581 : elapsed_total,
1582 : elapsed_since_last_callback,
1583 451799 : }| {
1584 451799 : self.metrics
1585 451799 : .wait_lsn_in_progress_micros
1586 451799 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1587 451799 : if !is_slow {
1588 451799 : return;
1589 0 : }
1590 0 : // It's slow, see if we should log it.
1591 0 : // (We limit the logging to one per invocation per timeline to avoid excessive
1592 0 : // logging during an extended broker / networking outage that affects all timelines.)
1593 0 : if logging_permit.is_none() {
1594 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1595 0 : }
1596 0 : if logging_permit.is_none() {
1597 0 : return;
1598 0 : }
1599 0 : // We log it.
1600 0 : if ready {
1601 0 : info!(
1602 0 : "slow wait_lsn completed after {:.3}s",
1603 0 : elapsed_total.as_secs_f64()
1604 : );
1605 : } else {
1606 0 : info!(
1607 0 : "slow wait_lsn still running for {:.3}s",
1608 0 : elapsed_total.as_secs_f64()
1609 : );
1610 : }
1611 451799 : },
1612 451799 : );
1613 451799 : let res = wait_for_timeout.await;
1614 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1615 451799 : drop(logging_permit);
1616 451799 : drop(start_finish_counterpair_guard);
1617 451799 : drop(timer);
1618 451799 : match res {
1619 451799 : Ok(()) => Ok(()),
1620 0 : Err(e) => {
1621 : use utils::seqwait::SeqWaitError::*;
1622 0 : match e {
1623 0 : Shutdown => Err(WaitLsnError::Shutdown),
1624 : Timeout => {
1625 0 : let walreceiver_status = self.walreceiver_status();
1626 0 : Err(WaitLsnError::Timeout(format!(
1627 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1628 0 : lsn,
1629 0 : self.get_last_record_lsn(),
1630 0 : self.get_disk_consistent_lsn(),
1631 0 : walreceiver_status,
1632 0 : )))
1633 : }
1634 : }
1635 : }
1636 : }
1637 451799 : }
1638 :
1639 0 : pub(crate) fn walreceiver_status(&self) -> String {
1640 0 : match &*self.walreceiver.lock().unwrap() {
1641 0 : None => "stopping or stopped".to_string(),
1642 0 : Some(walreceiver) => match walreceiver.status() {
1643 0 : Some(status) => status.to_human_readable_string(),
1644 0 : None => "Not active".to_string(),
1645 : },
1646 : }
1647 0 : }
1648 :
1649 : /// Check that it is valid to request operations with that lsn.
1650 476 : pub(crate) fn check_lsn_is_in_scope(
1651 476 : &self,
1652 476 : lsn: Lsn,
1653 476 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1654 476 : ) -> anyhow::Result<()> {
1655 476 : ensure!(
1656 476 : lsn >= **latest_gc_cutoff_lsn,
1657 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1658 8 : lsn,
1659 8 : **latest_gc_cutoff_lsn,
1660 : );
1661 468 : Ok(())
1662 476 : }
1663 :
1664 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1665 20 : pub(crate) fn init_lsn_lease(
1666 20 : &self,
1667 20 : lsn: Lsn,
1668 20 : length: Duration,
1669 20 : ctx: &RequestContext,
1670 20 : ) -> anyhow::Result<LsnLease> {
1671 20 : self.make_lsn_lease(lsn, length, true, ctx)
1672 20 : }
1673 :
1674 : /// 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.
1675 8 : pub(crate) fn renew_lsn_lease(
1676 8 : &self,
1677 8 : lsn: Lsn,
1678 8 : length: Duration,
1679 8 : ctx: &RequestContext,
1680 8 : ) -> anyhow::Result<LsnLease> {
1681 8 : self.make_lsn_lease(lsn, length, false, ctx)
1682 8 : }
1683 :
1684 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1685 : ///
1686 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1687 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1688 : ///
1689 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1690 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1691 28 : fn make_lsn_lease(
1692 28 : &self,
1693 28 : lsn: Lsn,
1694 28 : length: Duration,
1695 28 : init: bool,
1696 28 : _ctx: &RequestContext,
1697 28 : ) -> anyhow::Result<LsnLease> {
1698 24 : let lease = {
1699 : // Normalize the requested LSN to be aligned, and move to the first record
1700 : // if it points to the beginning of the page (header).
1701 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1702 28 :
1703 28 : let mut gc_info = self.gc_info.write().unwrap();
1704 28 : let planned_cutoff = gc_info.min_cutoff();
1705 28 :
1706 28 : let valid_until = SystemTime::now() + length;
1707 28 :
1708 28 : let entry = gc_info.leases.entry(lsn);
1709 28 :
1710 28 : match entry {
1711 12 : Entry::Occupied(mut occupied) => {
1712 12 : let existing_lease = occupied.get_mut();
1713 12 : if valid_until > existing_lease.valid_until {
1714 4 : existing_lease.valid_until = valid_until;
1715 4 : let dt: DateTime<Utc> = valid_until.into();
1716 4 : info!("lease extended to {}", dt);
1717 : } else {
1718 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1719 8 : info!("existing lease covers greater length, valid until {}", dt);
1720 : }
1721 :
1722 12 : existing_lease.clone()
1723 : }
1724 16 : Entry::Vacant(vacant) => {
1725 : // Reject already GC-ed LSN if we are in AttachedSingle and
1726 : // not blocked by the lsn lease deadline.
1727 16 : let validate = {
1728 16 : let conf = self.tenant_conf.load();
1729 16 : conf.location.attach_mode == AttachmentMode::Single
1730 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1731 : };
1732 :
1733 16 : if init || validate {
1734 16 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1735 16 : if lsn < *latest_gc_cutoff_lsn {
1736 4 : bail!(
1737 4 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1738 4 : lsn,
1739 4 : *latest_gc_cutoff_lsn
1740 4 : );
1741 12 : }
1742 12 : if lsn < planned_cutoff {
1743 0 : bail!(
1744 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1745 0 : lsn,
1746 0 : planned_cutoff
1747 0 : );
1748 12 : }
1749 0 : }
1750 :
1751 12 : let dt: DateTime<Utc> = valid_until.into();
1752 12 : info!("lease created, valid until {}", dt);
1753 12 : vacant.insert(LsnLease { valid_until }).clone()
1754 : }
1755 : }
1756 : };
1757 :
1758 24 : Ok(lease)
1759 28 : }
1760 :
1761 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1762 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1763 : #[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))]
1764 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1765 : self.freeze0().await
1766 : }
1767 :
1768 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1769 : #[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))]
1770 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1771 : self.freeze_and_flush0().await
1772 : }
1773 :
1774 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1775 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1776 2260 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1777 2260 : let mut g = self.write_lock.lock().await;
1778 2260 : let to_lsn = self.get_last_record_lsn();
1779 2260 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1780 2260 : }
1781 :
1782 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1783 : // polluting the span hierarchy.
1784 2260 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1785 2260 : let token = self.freeze0().await?;
1786 2260 : self.wait_flush_completion(token).await
1787 2260 : }
1788 :
1789 : // Check if an open ephemeral layer should be closed: this provides
1790 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1791 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1792 : // ephemeral layer bytes has been breached.
1793 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1794 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1795 : // If the write lock is held, there is an active wal receiver: rolling open layers
1796 : // is their responsibility while they hold this lock.
1797 0 : return;
1798 : };
1799 :
1800 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1801 : // time, and this was missed.
1802 : // if write_guard.is_none() { return; }
1803 :
1804 0 : let Ok(layers_guard) = self.layers.try_read() else {
1805 : // Don't block if the layer lock is busy
1806 0 : return;
1807 : };
1808 :
1809 0 : let Ok(lm) = layers_guard.layer_map() else {
1810 0 : return;
1811 : };
1812 :
1813 0 : let Some(open_layer) = &lm.open_layer else {
1814 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1815 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1816 : // that didn't result in writes to this shard.
1817 :
1818 : // Must not hold the layers lock while waiting for a flush.
1819 0 : drop(layers_guard);
1820 0 :
1821 0 : let last_record_lsn = self.get_last_record_lsn();
1822 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1823 0 : if last_record_lsn > disk_consistent_lsn {
1824 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1825 : // we are a sharded tenant and have skipped some WAL
1826 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1827 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1828 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1829 : // without any data ingested (yet) doesn't write a remote index as soon as it
1830 : // sees its LSN advance: we only do this if we've been layer-less
1831 : // for some time.
1832 0 : tracing::debug!(
1833 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1834 : disk_consistent_lsn,
1835 : last_record_lsn
1836 : );
1837 :
1838 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1839 : // We know there is no open layer, so we can request freezing without actually
1840 : // freezing anything. This is true even if we have dropped the layers_guard, we
1841 : // still hold the write_guard.
1842 0 : let _ = async {
1843 0 : let token = self
1844 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1845 0 : .await?;
1846 0 : self.wait_flush_completion(token).await
1847 0 : }
1848 0 : .await;
1849 0 : }
1850 0 : }
1851 :
1852 0 : return;
1853 : };
1854 :
1855 0 : let Some(current_size) = open_layer.try_len() else {
1856 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1857 : // read lock to get size should always succeed.
1858 0 : tracing::warn!("Lock conflict while reading size of open layer");
1859 0 : return;
1860 : };
1861 :
1862 0 : let current_lsn = self.get_last_record_lsn();
1863 :
1864 0 : let checkpoint_distance_override = open_layer.tick().await;
1865 :
1866 0 : if let Some(size_override) = checkpoint_distance_override {
1867 0 : if current_size > size_override {
1868 : // This is not harmful, but it only happens in relatively rare cases where
1869 : // time-based checkpoints are not happening fast enough to keep the amount of
1870 : // ephemeral data within configured limits. It's a sign of stress on the system.
1871 0 : tracing::info!(
1872 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
1873 : );
1874 0 : }
1875 0 : }
1876 :
1877 0 : let checkpoint_distance =
1878 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1879 0 :
1880 0 : if self.should_roll(
1881 0 : current_size,
1882 0 : current_size,
1883 0 : checkpoint_distance,
1884 0 : self.get_last_record_lsn(),
1885 0 : self.last_freeze_at.load(),
1886 0 : open_layer.get_opened_at(),
1887 0 : ) {
1888 0 : match open_layer.info() {
1889 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1890 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1891 0 : // happens asynchronously in the background.
1892 0 : tracing::debug!(
1893 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1894 : );
1895 : }
1896 : InMemoryLayerInfo::Open { .. } => {
1897 : // Upgrade to a write lock and freeze the layer
1898 0 : drop(layers_guard);
1899 0 : let res = self
1900 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1901 0 : .await;
1902 :
1903 0 : if let Err(e) = res {
1904 0 : tracing::info!(
1905 0 : "failed to flush frozen layer after background freeze: {e:#}"
1906 : );
1907 0 : }
1908 : }
1909 : }
1910 0 : }
1911 0 : }
1912 :
1913 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1914 : ///
1915 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1916 : /// child timelines that are not offloaded yet.
1917 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1918 0 : if self.remote_client.is_archived() != Some(true) {
1919 0 : return (false, "the timeline is not archived");
1920 0 : }
1921 0 : if !self.remote_client.no_pending_work() {
1922 : // if the remote client is still processing some work, we can't offload
1923 0 : return (false, "the upload queue is not drained yet");
1924 0 : }
1925 0 :
1926 0 : (true, "ok")
1927 0 : }
1928 :
1929 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1930 : /// compaction tasks.
1931 728 : pub(crate) async fn compact(
1932 728 : self: &Arc<Self>,
1933 728 : cancel: &CancellationToken,
1934 728 : flags: EnumSet<CompactFlags>,
1935 728 : ctx: &RequestContext,
1936 728 : ) -> Result<CompactionOutcome, CompactionError> {
1937 728 : let res = self
1938 728 : .compact_with_options(
1939 728 : cancel,
1940 728 : CompactOptions {
1941 728 : flags,
1942 728 : compact_key_range: None,
1943 728 : compact_lsn_range: None,
1944 728 : sub_compaction: false,
1945 728 : sub_compaction_max_job_size_mb: None,
1946 728 : },
1947 728 : ctx,
1948 728 : )
1949 728 : .await;
1950 728 : if let Err(err) = &res {
1951 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
1952 728 : }
1953 728 : res
1954 728 : }
1955 :
1956 : /// Outermost timeline compaction operation; downloads needed layers.
1957 : ///
1958 : /// NB: the cancellation token is usually from a background task, but can also come from a
1959 : /// request task.
1960 728 : pub(crate) async fn compact_with_options(
1961 728 : self: &Arc<Self>,
1962 728 : cancel: &CancellationToken,
1963 728 : options: CompactOptions,
1964 728 : ctx: &RequestContext,
1965 728 : ) -> Result<CompactionOutcome, CompactionError> {
1966 728 : // Acquire the compaction lock and task semaphore.
1967 728 : //
1968 728 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1969 728 : // out by other background tasks (including image compaction). We request this via
1970 728 : // `BackgroundLoopKind::L0Compaction`.
1971 728 : //
1972 728 : // Yield for pending L0 compaction while waiting for the semaphore.
1973 728 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1974 728 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1975 0 : true => BackgroundLoopKind::L0Compaction,
1976 728 : false => BackgroundLoopKind::Compaction,
1977 : };
1978 728 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
1979 728 : if yield_for_l0 {
1980 : // If this is an L0 pass, it doesn't make sense to yield for L0.
1981 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
1982 : // If `compaction_l0_first` is disabled, there's no point yielding.
1983 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
1984 728 : }
1985 :
1986 728 : let acquire = async move {
1987 728 : let guard = self.compaction_lock.lock().await;
1988 728 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
1989 728 : (guard, permit)
1990 728 : };
1991 :
1992 728 : let (_guard, _permit) = tokio::select! {
1993 728 : (guard, permit) = acquire => (guard, permit),
1994 728 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
1995 0 : return Ok(CompactionOutcome::YieldForL0);
1996 : }
1997 728 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1998 728 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1999 : };
2000 :
2001 728 : let last_record_lsn = self.get_last_record_lsn();
2002 728 :
2003 728 : // Last record Lsn could be zero in case the timeline was just created
2004 728 : if !last_record_lsn.is_valid() {
2005 0 : warn!(
2006 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2007 : );
2008 0 : return Ok(CompactionOutcome::Skipped);
2009 728 : }
2010 :
2011 728 : let result = match self.get_compaction_algorithm_settings().kind {
2012 : CompactionAlgorithm::Tiered => {
2013 0 : self.compact_tiered(cancel, ctx).await?;
2014 0 : Ok(CompactionOutcome::Done)
2015 : }
2016 728 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2017 : };
2018 :
2019 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2020 0 : match &result {
2021 728 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2022 0 : Err(e) if e.is_cancel() => {}
2023 0 : Err(CompactionError::ShuttingDown) => {
2024 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
2025 0 : }
2026 0 : Err(CompactionError::AlreadyRunning(_)) => {
2027 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
2028 0 : }
2029 : Err(CompactionError::Other(_)) => {
2030 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
2031 : }
2032 : Err(CompactionError::CollectKeySpaceError(_)) => {
2033 : // Cancelled errors are covered by the `Err(e) if e.is_cancel()` branch.
2034 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
2035 : }
2036 : // Don't change the current value on offload failure or shutdown. We don't want to
2037 : // abruptly stall nor resume L0 flushes in these cases.
2038 0 : Err(CompactionError::Offload(_)) => {}
2039 : };
2040 :
2041 728 : result
2042 728 : }
2043 :
2044 : /// Mutate the timeline with a [`TimelineWriter`].
2045 10266400 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2046 10266400 : TimelineWriter {
2047 10266400 : tl: self,
2048 10266400 : write_guard: self.write_lock.lock().await,
2049 : }
2050 10266400 : }
2051 :
2052 0 : pub(crate) fn activate(
2053 0 : self: &Arc<Self>,
2054 0 : parent: Arc<crate::tenant::Tenant>,
2055 0 : broker_client: BrokerClientChannel,
2056 0 : background_jobs_can_start: Option<&completion::Barrier>,
2057 0 : ctx: &RequestContext,
2058 0 : ) {
2059 0 : if self.tenant_shard_id.is_shard_zero() {
2060 0 : // Logical size is only maintained accurately on shard zero.
2061 0 : self.spawn_initial_logical_size_computation_task(ctx);
2062 0 : }
2063 0 : self.launch_wal_receiver(ctx, broker_client);
2064 0 : self.set_state(TimelineState::Active);
2065 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2066 0 : }
2067 :
2068 : /// After this function returns, there are no timeline-scoped tasks are left running.
2069 : ///
2070 : /// The preferred pattern for is:
2071 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2072 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2073 : /// go the extra mile and keep track of JoinHandles
2074 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2075 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2076 : ///
2077 : /// For legacy reasons, we still have multiple tasks spawned using
2078 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2079 : /// We refer to these as "timeline-scoped task_mgr tasks".
2080 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2081 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2082 : /// or [`task_mgr::shutdown_watcher`].
2083 : /// We want to gradually convert the code base away from these.
2084 : ///
2085 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2086 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2087 : /// ones that aren't mentioned here):
2088 : /// - [`TaskKind::TimelineDeletionWorker`]
2089 : /// - NB: also used for tenant deletion
2090 : /// - [`TaskKind::RemoteUploadTask`]`
2091 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2092 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2093 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2094 : /// - [`TaskKind::Eviction`]
2095 : /// - [`TaskKind::LayerFlushTask`]
2096 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2097 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2098 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2099 20 : debug_assert_current_span_has_tenant_and_timeline_id();
2100 20 :
2101 20 : // Regardless of whether we're going to try_freeze_and_flush
2102 20 : // or not, stop ingesting any more data. Walreceiver only provides
2103 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
2104 20 : // So, only after the self.gate.close() below will we know for sure that
2105 20 : // no walreceiver tasks are left.
2106 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
2107 20 : // data during the call to `self.freeze_and_flush()` below.
2108 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
2109 20 : // which is what we'd need to properly model early shutdown of the walreceiver
2110 20 : // task sub-tree before the other Timeline task sub-trees.
2111 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
2112 20 : tracing::debug!(
2113 0 : is_some = walreceiver.is_some(),
2114 0 : "Waiting for WalReceiverManager..."
2115 : );
2116 20 : if let Some(walreceiver) = walreceiver {
2117 0 : walreceiver.cancel();
2118 20 : }
2119 : // ... and inform any waiters for newer LSNs that there won't be any.
2120 20 : self.last_record_lsn.shutdown();
2121 20 :
2122 20 : if let ShutdownMode::FreezeAndFlush = mode {
2123 12 : let do_flush = if let Some((open, frozen)) = self
2124 12 : .layers
2125 12 : .read()
2126 12 : .await
2127 12 : .layer_map()
2128 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2129 12 : .ok()
2130 12 : .filter(|(open, frozen)| *open || *frozen > 0)
2131 : {
2132 0 : if self.remote_client.is_archived() == Some(true) {
2133 : // No point flushing on shutdown for an archived timeline: it is not important
2134 : // to have it nice and fresh after our restart, and trying to flush here might
2135 : // race with trying to offload it (which also stops the flush loop)
2136 0 : false
2137 : } else {
2138 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2139 0 : true
2140 : }
2141 : } else {
2142 : // this is double-shutdown, it'll be a no-op
2143 12 : true
2144 : };
2145 :
2146 : // we shut down walreceiver above, so, we won't add anything more
2147 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2148 : // to reach the disk & upload queue, then shut the upload queue and
2149 : // wait for it to drain.
2150 12 : if do_flush {
2151 12 : match self.freeze_and_flush().await {
2152 : Ok(_) => {
2153 : // drain the upload queue
2154 : // if we did not wait for completion here, it might be our shutdown process
2155 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2156 : // be spawned.
2157 : //
2158 : // what is problematic is the shutting down of RemoteTimelineClient, because
2159 : // obviously it does not make sense to stop while we wait for it, but what
2160 : // about corner cases like s3 suddenly hanging up?
2161 12 : self.remote_client.shutdown().await;
2162 : }
2163 : Err(FlushLayerError::Cancelled) => {
2164 : // this is likely the second shutdown, ignore silently.
2165 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2166 0 : debug_assert!(self.cancel.is_cancelled());
2167 : }
2168 0 : Err(e) => {
2169 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2170 0 : // we have some extra WAL replay to do next time the timeline starts.
2171 0 : warn!("failed to freeze and flush: {e:#}");
2172 : }
2173 : }
2174 :
2175 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2176 : // we also do a final check here to ensure that the queue is empty.
2177 12 : if !self.remote_client.no_pending_work() {
2178 0 : warn!(
2179 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2180 : );
2181 12 : }
2182 0 : }
2183 8 : }
2184 :
2185 20 : if let ShutdownMode::Reload = mode {
2186 : // drain the upload queue
2187 4 : self.remote_client.shutdown().await;
2188 4 : if !self.remote_client.no_pending_work() {
2189 0 : warn!(
2190 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2191 : );
2192 4 : }
2193 16 : }
2194 :
2195 : // Signal any subscribers to our cancellation token to drop out
2196 20 : tracing::debug!("Cancelling CancellationToken");
2197 20 : self.cancel.cancel();
2198 20 :
2199 20 : // If we have a background task downloading heatmap layers stop it.
2200 20 : // The background downloads are sensitive to timeline cancellation (done above),
2201 20 : // so the drain will be immediate.
2202 20 : self.stop_and_drain_heatmap_layers_download().await;
2203 :
2204 : // Ensure Prevent new page service requests from starting.
2205 20 : self.handles.shutdown();
2206 20 :
2207 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2208 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2209 20 : self.remote_client.stop();
2210 20 :
2211 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2212 20 : // to shut down the upload queue tasks.
2213 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
2214 20 : task_mgr::shutdown_tasks(
2215 20 : Some(TaskKind::RemoteUploadTask),
2216 20 : Some(self.tenant_shard_id),
2217 20 : Some(self.timeline_id),
2218 20 : )
2219 20 : .await;
2220 :
2221 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2222 20 : tracing::debug!("Waiting for tasks...");
2223 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2224 :
2225 : {
2226 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2227 : // open.
2228 20 : let mut write_guard = self.write_lock.lock().await;
2229 20 : self.layers.write().await.shutdown(&mut write_guard);
2230 20 : }
2231 20 :
2232 20 : // Finally wait until any gate-holders are complete.
2233 20 : //
2234 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2235 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2236 20 : self.gate.close().await;
2237 :
2238 20 : self.metrics.shutdown();
2239 20 : }
2240 :
2241 928 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2242 928 : match (self.current_state(), new_state) {
2243 928 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2244 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2245 : }
2246 0 : (st, TimelineState::Loading) => {
2247 0 : error!("ignoring transition from {st:?} into Loading state");
2248 : }
2249 0 : (TimelineState::Broken { .. }, new_state) => {
2250 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2251 : }
2252 : (TimelineState::Stopping, TimelineState::Active) => {
2253 0 : error!("Not activating a Stopping timeline");
2254 : }
2255 924 : (_, new_state) => {
2256 924 : self.state.send_replace(new_state);
2257 924 : }
2258 : }
2259 928 : }
2260 :
2261 4 : pub(crate) fn set_broken(&self, reason: String) {
2262 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2263 4 : let broken_state = TimelineState::Broken {
2264 4 : reason,
2265 4 : backtrace: backtrace_str,
2266 4 : };
2267 4 : self.set_state(broken_state);
2268 4 :
2269 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2270 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2271 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2272 4 : self.cancel.cancel();
2273 4 : }
2274 :
2275 453967 : pub(crate) fn current_state(&self) -> TimelineState {
2276 453967 : self.state.borrow().clone()
2277 453967 : }
2278 :
2279 12 : pub(crate) fn is_broken(&self) -> bool {
2280 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2281 12 : }
2282 :
2283 504 : pub(crate) fn is_active(&self) -> bool {
2284 504 : self.current_state() == TimelineState::Active
2285 504 : }
2286 :
2287 32 : pub(crate) fn is_archived(&self) -> Option<bool> {
2288 32 : self.remote_client.is_archived()
2289 32 : }
2290 :
2291 32 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2292 32 : self.remote_client.is_invisible()
2293 32 : }
2294 :
2295 736 : pub(crate) fn is_stopping(&self) -> bool {
2296 736 : self.current_state() == TimelineState::Stopping
2297 736 : }
2298 :
2299 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2300 0 : self.state.subscribe()
2301 0 : }
2302 :
2303 451803 : pub(crate) async fn wait_to_become_active(
2304 451803 : &self,
2305 451803 : _ctx: &RequestContext, // Prepare for use by cancellation
2306 451803 : ) -> Result<(), TimelineState> {
2307 451803 : let mut receiver = self.state.subscribe();
2308 : loop {
2309 451803 : let current_state = receiver.borrow().clone();
2310 451803 : match current_state {
2311 : TimelineState::Loading => {
2312 0 : receiver
2313 0 : .changed()
2314 0 : .await
2315 0 : .expect("holding a reference to self");
2316 : }
2317 : TimelineState::Active => {
2318 451799 : return Ok(());
2319 : }
2320 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2321 : // There's no chance the timeline can transition back into ::Active
2322 4 : return Err(current_state);
2323 : }
2324 : }
2325 : }
2326 451803 : }
2327 :
2328 0 : pub(crate) async fn layer_map_info(
2329 0 : &self,
2330 0 : reset: LayerAccessStatsReset,
2331 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2332 0 : let guard = self.layers.read().await;
2333 0 : let layer_map = guard.layer_map()?;
2334 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2335 0 : if let Some(open_layer) = &layer_map.open_layer {
2336 0 : in_memory_layers.push(open_layer.info());
2337 0 : }
2338 0 : for frozen_layer in &layer_map.frozen_layers {
2339 0 : in_memory_layers.push(frozen_layer.info());
2340 0 : }
2341 :
2342 0 : let historic_layers = layer_map
2343 0 : .iter_historic_layers()
2344 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2345 0 : .collect();
2346 0 :
2347 0 : Ok(LayerMapInfo {
2348 0 : in_memory_layers,
2349 0 : historic_layers,
2350 0 : })
2351 0 : }
2352 :
2353 : #[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))]
2354 : pub(crate) async fn download_layer(
2355 : &self,
2356 : layer_file_name: &LayerName,
2357 : ctx: &RequestContext,
2358 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2359 : let Some(layer) = self
2360 : .find_layer(layer_file_name)
2361 : .await
2362 0 : .map_err(|e| match e {
2363 0 : layer_manager::Shutdown => {
2364 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2365 0 : }
2366 0 : })?
2367 : else {
2368 : return Ok(None);
2369 : };
2370 :
2371 : layer.download(ctx).await?;
2372 :
2373 : Ok(Some(true))
2374 : }
2375 :
2376 : /// Evict just one layer.
2377 : ///
2378 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2379 0 : pub(crate) async fn evict_layer(
2380 0 : &self,
2381 0 : layer_file_name: &LayerName,
2382 0 : ) -> anyhow::Result<Option<bool>> {
2383 0 : let _gate = self
2384 0 : .gate
2385 0 : .enter()
2386 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2387 :
2388 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2389 0 : return Ok(None);
2390 : };
2391 :
2392 : // curl has this by default
2393 0 : let timeout = std::time::Duration::from_secs(120);
2394 0 :
2395 0 : match local_layer.evict_and_wait(timeout).await {
2396 0 : Ok(()) => Ok(Some(true)),
2397 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2398 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2399 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2400 : }
2401 0 : }
2402 :
2403 9606020 : fn should_roll(
2404 9606020 : &self,
2405 9606020 : layer_size: u64,
2406 9606020 : projected_layer_size: u64,
2407 9606020 : checkpoint_distance: u64,
2408 9606020 : projected_lsn: Lsn,
2409 9606020 : last_freeze_at: Lsn,
2410 9606020 : opened_at: Instant,
2411 9606020 : ) -> bool {
2412 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2413 9606020 :
2414 9606020 : // Rolling the open layer can be triggered by:
2415 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2416 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2417 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2418 9606020 : // 1/count of the LSN on average.
2419 9606020 : // 2. The size of the currently open layer.
2420 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2421 9606020 : // up and suspend activity.
2422 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2423 0 : info!(
2424 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2425 : projected_lsn, layer_size, distance
2426 : );
2427 :
2428 0 : true
2429 9606020 : } else if projected_layer_size >= checkpoint_distance {
2430 : // NB: this check is relied upon by:
2431 160 : let _ = IndexEntry::validate_checkpoint_distance;
2432 160 : info!(
2433 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2434 : projected_lsn, layer_size, projected_layer_size
2435 : );
2436 :
2437 160 : true
2438 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2439 0 : info!(
2440 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2441 0 : projected_lsn,
2442 0 : layer_size,
2443 0 : opened_at.elapsed()
2444 : );
2445 :
2446 0 : true
2447 : } else {
2448 9605860 : false
2449 : }
2450 9606020 : }
2451 : }
2452 :
2453 : /// Number of times we will compute partition within a checkpoint distance.
2454 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2455 :
2456 : // Private functions
2457 : impl Timeline {
2458 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2459 24 : let tenant_conf = self.tenant_conf.load();
2460 24 : tenant_conf
2461 24 : .tenant_conf
2462 24 : .lsn_lease_length
2463 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2464 24 : }
2465 :
2466 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2467 0 : let tenant_conf = self.tenant_conf.load();
2468 0 : tenant_conf
2469 0 : .tenant_conf
2470 0 : .lsn_lease_length_for_ts
2471 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2472 0 : }
2473 :
2474 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2475 0 : let tenant_conf = self.tenant_conf.load();
2476 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2477 0 : }
2478 :
2479 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2480 0 : let tenant_conf = self.tenant_conf.load();
2481 0 : tenant_conf
2482 0 : .tenant_conf
2483 0 : .lazy_slru_download
2484 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2485 0 : }
2486 :
2487 : /// Checks if a get page request should get perf tracing
2488 : ///
2489 : /// The configuration priority is: tenant config override, default tenant config,
2490 : /// pageserver config.
2491 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2492 0 : let tenant_conf = self.tenant_conf.load();
2493 0 : let ratio = tenant_conf
2494 0 : .tenant_conf
2495 0 : .sampling_ratio
2496 0 : .flatten()
2497 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2498 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2499 0 :
2500 0 : match ratio {
2501 0 : Some(r) => {
2502 0 : if r.numerator == 0 {
2503 0 : false
2504 : } else {
2505 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2506 : }
2507 : }
2508 0 : None => false,
2509 : }
2510 0 : }
2511 :
2512 9609044 : fn get_checkpoint_distance(&self) -> u64 {
2513 9609044 : let tenant_conf = self.tenant_conf.load();
2514 9609044 : tenant_conf
2515 9609044 : .tenant_conf
2516 9609044 : .checkpoint_distance
2517 9609044 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2518 9609044 : }
2519 :
2520 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2521 9605860 : let tenant_conf = self.tenant_conf.load();
2522 9605860 : tenant_conf
2523 9605860 : .tenant_conf
2524 9605860 : .checkpoint_timeout
2525 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2526 9605860 : }
2527 :
2528 5048 : fn get_compaction_period(&self) -> Duration {
2529 5048 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2530 5048 : tenant_conf
2531 5048 : .compaction_period
2532 5048 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2533 5048 : }
2534 :
2535 1352 : fn get_compaction_target_size(&self) -> u64 {
2536 1352 : let tenant_conf = self.tenant_conf.load();
2537 1352 : tenant_conf
2538 1352 : .tenant_conf
2539 1352 : .compaction_target_size
2540 1352 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2541 1352 : }
2542 :
2543 3156 : fn get_compaction_threshold(&self) -> usize {
2544 3156 : let tenant_conf = self.tenant_conf.load();
2545 3156 : tenant_conf
2546 3156 : .tenant_conf
2547 3156 : .compaction_threshold
2548 3156 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2549 3156 : }
2550 :
2551 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2552 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2553 : /// possible that the index part persists the state while the config doesn't get persisted.
2554 3892 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2555 3892 : let tenant_conf = self.tenant_conf.load();
2556 3892 : tenant_conf
2557 3892 : .tenant_conf
2558 3892 : .rel_size_v2_enabled
2559 3892 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2560 3892 : }
2561 :
2562 4396 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2563 4396 : self.rel_size_v2_status
2564 4396 : .load()
2565 4396 : .as_ref()
2566 4396 : .map(|s| s.as_ref().clone())
2567 4396 : .unwrap_or(RelSizeMigration::Legacy)
2568 4396 : }
2569 :
2570 56 : fn get_compaction_upper_limit(&self) -> usize {
2571 56 : let tenant_conf = self.tenant_conf.load();
2572 56 : tenant_conf
2573 56 : .tenant_conf
2574 56 : .compaction_upper_limit
2575 56 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2576 56 : }
2577 :
2578 0 : pub fn get_compaction_l0_first(&self) -> bool {
2579 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2580 0 : tenant_conf
2581 0 : .compaction_l0_first
2582 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2583 0 : }
2584 :
2585 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2586 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2587 0 : tenant_conf
2588 0 : .compaction_l0_semaphore
2589 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2590 0 : }
2591 :
2592 2524 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2593 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2594 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2595 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2596 :
2597 : // If compaction is disabled, don't delay.
2598 2524 : if self.get_compaction_period() == Duration::ZERO {
2599 2520 : return None;
2600 4 : }
2601 4 :
2602 4 : let compaction_threshold = self.get_compaction_threshold();
2603 4 : let tenant_conf = self.tenant_conf.load();
2604 4 : let l0_flush_delay_threshold = tenant_conf
2605 4 : .tenant_conf
2606 4 : .l0_flush_delay_threshold
2607 4 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2608 4 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2609 4 :
2610 4 : // 0 disables backpressure.
2611 4 : if l0_flush_delay_threshold == 0 {
2612 0 : return None;
2613 4 : }
2614 4 :
2615 4 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2616 4 : // backpressure flushes below this.
2617 4 : // TODO: the tenant config should have validation to prevent this instead.
2618 4 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2619 4 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2620 2524 : }
2621 :
2622 2524 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2623 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2624 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2625 : // sibling timelines. We need more confidence before enabling this.
2626 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2627 :
2628 : // If compaction is disabled, don't stall.
2629 2524 : if self.get_compaction_period() == Duration::ZERO {
2630 2520 : return None;
2631 4 : }
2632 4 :
2633 4 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2634 4 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2635 4 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2636 4 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2637 4 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2638 0 : return None;
2639 4 : }
2640 4 :
2641 4 : let compaction_threshold = self.get_compaction_threshold();
2642 4 : let tenant_conf = self.tenant_conf.load();
2643 4 : let l0_flush_stall_threshold = tenant_conf
2644 4 : .tenant_conf
2645 4 : .l0_flush_stall_threshold
2646 4 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2647 4 :
2648 4 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2649 4 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2650 4 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2651 4 : if cfg!(feature = "testing")
2652 4 : && compaction_threshold == 1
2653 0 : && l0_flush_stall_threshold.is_none()
2654 : {
2655 0 : return None;
2656 4 : }
2657 4 :
2658 4 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2659 4 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2660 4 :
2661 4 : // 0 disables backpressure.
2662 4 : if l0_flush_stall_threshold == 0 {
2663 4 : return None;
2664 0 : }
2665 0 :
2666 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2667 0 : // backpressure flushes below this.
2668 0 : // TODO: the tenant config should have validation to prevent this instead.
2669 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2670 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2671 2524 : }
2672 :
2673 28 : fn get_image_creation_threshold(&self) -> usize {
2674 28 : let tenant_conf = self.tenant_conf.load();
2675 28 : tenant_conf
2676 28 : .tenant_conf
2677 28 : .image_creation_threshold
2678 28 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2679 28 : }
2680 :
2681 728 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2682 728 : let tenant_conf = &self.tenant_conf.load();
2683 728 : tenant_conf
2684 728 : .tenant_conf
2685 728 : .compaction_algorithm
2686 728 : .as_ref()
2687 728 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2688 728 : .clone()
2689 728 : }
2690 :
2691 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2692 0 : let tenant_conf = self.tenant_conf.load();
2693 0 : tenant_conf
2694 0 : .tenant_conf
2695 0 : .eviction_policy
2696 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2697 0 : }
2698 :
2699 924 : fn get_evictions_low_residence_duration_metric_threshold(
2700 924 : tenant_conf: &pageserver_api::models::TenantConfig,
2701 924 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
2702 924 : ) -> Duration {
2703 924 : tenant_conf
2704 924 : .evictions_low_residence_duration_metric_threshold
2705 924 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2706 924 : }
2707 :
2708 1156 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2709 1156 : let tenant_conf = self.tenant_conf.load();
2710 1156 : tenant_conf
2711 1156 : .tenant_conf
2712 1156 : .image_layer_creation_check_threshold
2713 1156 : .unwrap_or(
2714 1156 : self.conf
2715 1156 : .default_tenant_conf
2716 1156 : .image_layer_creation_check_threshold,
2717 1156 : )
2718 1156 : }
2719 :
2720 0 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
2721 0 : let tenant_conf = &self.tenant_conf.load();
2722 0 : let gc_compaction_enabled = tenant_conf
2723 0 : .tenant_conf
2724 0 : .gc_compaction_enabled
2725 0 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
2726 0 : let gc_compaction_initial_threshold_kb = tenant_conf
2727 0 : .tenant_conf
2728 0 : .gc_compaction_initial_threshold_kb
2729 0 : .unwrap_or(
2730 0 : self.conf
2731 0 : .default_tenant_conf
2732 0 : .gc_compaction_initial_threshold_kb,
2733 0 : );
2734 0 : let gc_compaction_ratio_percent = tenant_conf
2735 0 : .tenant_conf
2736 0 : .gc_compaction_ratio_percent
2737 0 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
2738 0 : GcCompactionCombinedSettings {
2739 0 : gc_compaction_enabled,
2740 0 : gc_compaction_initial_threshold_kb,
2741 0 : gc_compaction_ratio_percent,
2742 0 : }
2743 0 : }
2744 :
2745 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2746 0 : let tenant_conf = self.tenant_conf.load();
2747 0 : tenant_conf
2748 0 : .tenant_conf
2749 0 : .image_creation_preempt_threshold
2750 0 : .unwrap_or(
2751 0 : self.conf
2752 0 : .default_tenant_conf
2753 0 : .image_creation_preempt_threshold,
2754 0 : )
2755 0 : }
2756 :
2757 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2758 : ///
2759 : /// Priority order is:
2760 : /// 1. Tenant config override
2761 : /// 2. Default value for tenant config override
2762 : /// 3. Pageserver config override
2763 : /// 4. Pageserver config default
2764 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2765 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2766 0 : tenant_conf
2767 0 : .wal_receiver_protocol_override
2768 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2769 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2770 0 : }
2771 :
2772 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2773 0 : // NB: Most tenant conf options are read by background loops, so,
2774 0 : // changes will automatically be picked up.
2775 0 :
2776 0 : // The threshold is embedded in the metric. So, we need to update it.
2777 0 : {
2778 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2779 0 : &new_conf.tenant_conf,
2780 0 : &self.conf.default_tenant_conf,
2781 0 : );
2782 0 :
2783 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2784 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2785 0 :
2786 0 : let timeline_id_str = self.timeline_id.to_string();
2787 0 :
2788 0 : self.remote_client.update_config(&new_conf.location);
2789 0 :
2790 0 : self.metrics
2791 0 : .evictions_with_low_residence_duration
2792 0 : .write()
2793 0 : .unwrap()
2794 0 : .change_threshold(
2795 0 : &tenant_id_str,
2796 0 : &shard_id_str,
2797 0 : &timeline_id_str,
2798 0 : new_threshold,
2799 0 : );
2800 0 : }
2801 0 : }
2802 :
2803 : /// Open a Timeline handle.
2804 : ///
2805 : /// Loads the metadata for the timeline into memory, but not the layer map.
2806 : #[allow(clippy::too_many_arguments)]
2807 924 : pub(super) fn new(
2808 924 : conf: &'static PageServerConf,
2809 924 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2810 924 : metadata: &TimelineMetadata,
2811 924 : previous_heatmap: Option<PreviousHeatmap>,
2812 924 : ancestor: Option<Arc<Timeline>>,
2813 924 : timeline_id: TimelineId,
2814 924 : tenant_shard_id: TenantShardId,
2815 924 : generation: Generation,
2816 924 : shard_identity: ShardIdentity,
2817 924 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2818 924 : resources: TimelineResources,
2819 924 : pg_version: u32,
2820 924 : state: TimelineState,
2821 924 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2822 924 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2823 924 : gc_compaction_state: Option<GcCompactionState>,
2824 924 : rel_size_v2_status: Option<RelSizeMigration>,
2825 924 : cancel: CancellationToken,
2826 924 : ) -> Arc<Self> {
2827 924 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2828 924 : let (state, _) = watch::channel(state);
2829 924 :
2830 924 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2831 924 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2832 924 :
2833 924 : let evictions_low_residence_duration_metric_threshold = {
2834 924 : let loaded_tenant_conf = tenant_conf.load();
2835 924 : Self::get_evictions_low_residence_duration_metric_threshold(
2836 924 : &loaded_tenant_conf.tenant_conf,
2837 924 : &conf.default_tenant_conf,
2838 924 : )
2839 : };
2840 :
2841 924 : if let Some(ancestor) = &ancestor {
2842 472 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2843 472 : // If we construct an explicit timeline object, it's obviously not offloaded
2844 472 : let is_offloaded = MaybeOffloaded::No;
2845 472 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2846 472 : }
2847 :
2848 924 : Arc::new_cyclic(|myself| {
2849 924 : let metrics = Arc::new(TimelineMetrics::new(
2850 924 : &tenant_shard_id,
2851 924 : &timeline_id,
2852 924 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2853 924 : "mtime",
2854 924 : evictions_low_residence_duration_metric_threshold,
2855 924 : ),
2856 924 : ));
2857 924 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2858 :
2859 924 : let mut result = Timeline {
2860 924 : conf,
2861 924 : tenant_conf,
2862 924 : myself: myself.clone(),
2863 924 : timeline_id,
2864 924 : tenant_shard_id,
2865 924 : generation,
2866 924 : shard_identity,
2867 924 : pg_version,
2868 924 : layers: Default::default(),
2869 924 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2870 924 :
2871 924 : walredo_mgr,
2872 924 : walreceiver: Mutex::new(None),
2873 924 :
2874 924 : remote_client: Arc::new(resources.remote_client),
2875 924 :
2876 924 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2877 924 : last_record_lsn: SeqWait::new(RecordLsn {
2878 924 : last: disk_consistent_lsn,
2879 924 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2880 924 : }),
2881 924 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2882 924 :
2883 924 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
2884 924 :
2885 924 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2886 924 : last_freeze_ts: RwLock::new(Instant::now()),
2887 924 :
2888 924 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2889 924 :
2890 924 : ancestor_timeline: ancestor,
2891 924 : ancestor_lsn: metadata.ancestor_lsn(),
2892 924 :
2893 924 : metrics,
2894 924 :
2895 924 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2896 924 : &tenant_shard_id,
2897 924 : &timeline_id,
2898 924 : resources.pagestream_throttle_metrics,
2899 924 : ),
2900 924 :
2901 7392 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2902 7392 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
2903 924 :
2904 924 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2905 924 :
2906 924 : layer_flush_start_tx,
2907 924 : layer_flush_done_tx,
2908 924 :
2909 924 : write_lock: tokio::sync::Mutex::new(None),
2910 924 :
2911 924 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2912 924 :
2913 924 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2914 924 : LastImageLayerCreationStatus::default(),
2915 924 : )),
2916 924 :
2917 924 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2918 924 : initdb_lsn: metadata.initdb_lsn(),
2919 924 :
2920 924 : current_logical_size: if disk_consistent_lsn.is_valid() {
2921 : // we're creating timeline data with some layer files existing locally,
2922 : // need to recalculate timeline's logical size based on data in the layers.
2923 480 : LogicalSize::deferred_initial(disk_consistent_lsn)
2924 : } else {
2925 : // we're creating timeline data without any layers existing locally,
2926 : // initial logical size is 0.
2927 444 : LogicalSize::empty_initial()
2928 : },
2929 :
2930 924 : partitioning: GuardArcSwap::new((
2931 924 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2932 924 : Lsn(0),
2933 924 : )),
2934 924 : repartition_threshold: 0,
2935 924 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2936 924 : last_image_layer_creation_check_instant: Mutex::new(None),
2937 924 :
2938 924 : last_received_wal: Mutex::new(None),
2939 924 : rel_size_cache: RwLock::new(RelSizeCache {
2940 924 : complete_as_of: disk_consistent_lsn,
2941 924 : map: HashMap::new(),
2942 924 : }),
2943 924 :
2944 924 : download_all_remote_layers_task_info: RwLock::new(None),
2945 924 :
2946 924 : state,
2947 924 :
2948 924 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2949 924 : EvictionTaskTimelineState::default(),
2950 924 : ),
2951 924 : delete_progress: TimelineDeleteProgress::default(),
2952 924 :
2953 924 : cancel,
2954 924 : gate: Gate::default(),
2955 924 :
2956 924 : compaction_lock: tokio::sync::Mutex::default(),
2957 924 : compaction_failed: AtomicBool::default(),
2958 924 : l0_compaction_trigger: resources.l0_compaction_trigger,
2959 924 : gc_lock: tokio::sync::Mutex::default(),
2960 924 :
2961 924 : standby_horizon: AtomicLsn::new(0),
2962 924 :
2963 924 : pagestream_throttle: resources.pagestream_throttle,
2964 924 :
2965 924 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2966 924 :
2967 924 : #[cfg(test)]
2968 924 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2969 924 :
2970 924 : l0_flush_global_state: resources.l0_flush_global_state,
2971 924 :
2972 924 : handles: Default::default(),
2973 924 :
2974 924 : attach_wal_lag_cooldown,
2975 924 :
2976 924 : create_idempotency,
2977 924 :
2978 924 : page_trace: Default::default(),
2979 924 :
2980 924 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
2981 924 :
2982 924 : heatmap_layers_downloader: Mutex::new(None),
2983 924 :
2984 924 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
2985 924 :
2986 924 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
2987 924 : };
2988 924 :
2989 924 : result.repartition_threshold =
2990 924 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2991 924 :
2992 924 : result
2993 924 : .metrics
2994 924 : .last_record_lsn_gauge
2995 924 : .set(disk_consistent_lsn.0 as i64);
2996 924 : result
2997 924 : })
2998 924 : }
2999 :
3000 1340 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3001 1340 : let Ok(guard) = self.gate.enter() else {
3002 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3003 0 : return;
3004 : };
3005 1340 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3006 1340 : match *flush_loop_state {
3007 912 : FlushLoopState::NotStarted => (),
3008 : FlushLoopState::Running { .. } => {
3009 428 : info!(
3010 0 : "skipping attempt to start flush_loop twice {}/{}",
3011 0 : self.tenant_shard_id, self.timeline_id
3012 : );
3013 428 : return;
3014 : }
3015 : FlushLoopState::Exited => {
3016 0 : info!(
3017 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3018 0 : self.tenant_shard_id, self.timeline_id
3019 : );
3020 0 : return;
3021 : }
3022 : }
3023 :
3024 912 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3025 912 : let self_clone = Arc::clone(self);
3026 912 :
3027 912 : debug!("spawning flush loop");
3028 912 : *flush_loop_state = FlushLoopState::Running {
3029 912 : #[cfg(test)]
3030 912 : expect_initdb_optimization: false,
3031 912 : #[cfg(test)]
3032 912 : initdb_optimization_count: 0,
3033 912 : };
3034 912 : task_mgr::spawn(
3035 912 : task_mgr::BACKGROUND_RUNTIME.handle(),
3036 912 : task_mgr::TaskKind::LayerFlushTask,
3037 912 : self.tenant_shard_id,
3038 912 : Some(self.timeline_id),
3039 912 : "layer flush task",
3040 912 : async move {
3041 912 : let _guard = guard;
3042 912 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3043 912 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3044 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3045 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3046 20 : *flush_loop_state = FlushLoopState::Exited;
3047 20 : Ok(())
3048 20 : }
3049 912 : .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))
3050 : );
3051 1340 : }
3052 :
3053 0 : pub(crate) fn update_gc_compaction_state(
3054 0 : &self,
3055 0 : gc_compaction_state: GcCompactionState,
3056 0 : ) -> anyhow::Result<()> {
3057 0 : self.gc_compaction_state
3058 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3059 0 : self.remote_client
3060 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3061 0 : }
3062 :
3063 0 : pub(crate) fn update_rel_size_v2_status(
3064 0 : &self,
3065 0 : rel_size_v2_status: RelSizeMigration,
3066 0 : ) -> anyhow::Result<()> {
3067 0 : self.rel_size_v2_status
3068 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3069 0 : self.remote_client
3070 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3071 0 : }
3072 :
3073 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3074 0 : self.gc_compaction_state.load_full().as_ref().clone()
3075 0 : }
3076 :
3077 : /// Creates and starts the wal receiver.
3078 : ///
3079 : /// This function is expected to be called at most once per Timeline's lifecycle
3080 : /// when the timeline is activated.
3081 0 : fn launch_wal_receiver(
3082 0 : self: &Arc<Self>,
3083 0 : ctx: &RequestContext,
3084 0 : broker_client: BrokerClientChannel,
3085 0 : ) {
3086 0 : info!(
3087 0 : "launching WAL receiver for timeline {} of tenant {}",
3088 0 : self.timeline_id, self.tenant_shard_id
3089 : );
3090 :
3091 0 : let tenant_conf = self.tenant_conf.load();
3092 0 : let wal_connect_timeout = tenant_conf
3093 0 : .tenant_conf
3094 0 : .walreceiver_connect_timeout
3095 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3096 0 : let lagging_wal_timeout = tenant_conf
3097 0 : .tenant_conf
3098 0 : .lagging_wal_timeout
3099 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3100 0 : let max_lsn_wal_lag = tenant_conf
3101 0 : .tenant_conf
3102 0 : .max_lsn_wal_lag
3103 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3104 0 :
3105 0 : let mut guard = self.walreceiver.lock().unwrap();
3106 0 : assert!(
3107 0 : guard.is_none(),
3108 0 : "multiple launches / re-launches of WAL receiver are not supported"
3109 : );
3110 0 : *guard = Some(WalReceiver::start(
3111 0 : Arc::clone(self),
3112 0 : WalReceiverConf {
3113 0 : protocol: self.resolve_wal_receiver_protocol(),
3114 0 : wal_connect_timeout,
3115 0 : lagging_wal_timeout,
3116 0 : max_lsn_wal_lag,
3117 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3118 0 : availability_zone: self.conf.availability_zone.clone(),
3119 0 : ingest_batch_size: self.conf.ingest_batch_size,
3120 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3121 0 : },
3122 0 : broker_client,
3123 0 : ctx,
3124 0 : ));
3125 0 : }
3126 :
3127 : /// Initialize with an empty layer map. Used when creating a new timeline.
3128 912 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3129 912 : let mut layers = self.layers.try_write().expect(
3130 912 : "in the context where we call this function, no other task has access to the object",
3131 912 : );
3132 912 : layers
3133 912 : .open_mut()
3134 912 : .expect("in this context the LayerManager must still be open")
3135 912 : .initialize_empty(Lsn(start_lsn.0));
3136 912 : }
3137 :
3138 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3139 : /// files.
3140 12 : pub(super) async fn load_layer_map(
3141 12 : &self,
3142 12 : disk_consistent_lsn: Lsn,
3143 12 : index_part: IndexPart,
3144 12 : ) -> anyhow::Result<()> {
3145 : use LayerName::*;
3146 : use init::Decision::*;
3147 : use init::{Discovered, DismissedLayer};
3148 :
3149 12 : let mut guard = self.layers.write().await;
3150 :
3151 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
3152 12 :
3153 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3154 12 : // structs representing all files on disk
3155 12 : let timeline_path = self
3156 12 : .conf
3157 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3158 12 : let conf = self.conf;
3159 12 : let span = tracing::Span::current();
3160 12 :
3161 12 : // Copy to move into the task we're about to spawn
3162 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
3163 :
3164 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3165 12 : move || {
3166 12 : let _g = span.entered();
3167 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3168 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3169 12 : let mut unrecognized_files = Vec::new();
3170 12 :
3171 12 : let mut path = timeline_path;
3172 :
3173 44 : for discovered in discovered {
3174 32 : let (name, kind) = match discovered {
3175 32 : Discovered::Layer(layer_file_name, local_metadata) => {
3176 32 : discovered_layers.push((layer_file_name, local_metadata));
3177 32 : continue;
3178 : }
3179 0 : Discovered::IgnoredBackup(path) => {
3180 0 : std::fs::remove_file(path)
3181 0 : .or_else(fs_ext::ignore_not_found)
3182 0 : .fatal_err("Removing .old file");
3183 0 : continue;
3184 : }
3185 0 : Discovered::Unknown(file_name) => {
3186 0 : // we will later error if there are any
3187 0 : unrecognized_files.push(file_name);
3188 0 : continue;
3189 : }
3190 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3191 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3192 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3193 : };
3194 0 : path.push(Utf8Path::new(&name));
3195 0 : init::cleanup(&path, kind)?;
3196 0 : path.pop();
3197 : }
3198 :
3199 12 : if !unrecognized_files.is_empty() {
3200 : // assume that if there are any there are many many.
3201 0 : let n = unrecognized_files.len();
3202 0 : let first = &unrecognized_files[..n.min(10)];
3203 0 : anyhow::bail!(
3204 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3205 0 : );
3206 12 : }
3207 12 :
3208 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3209 12 :
3210 12 : let mut loaded_layers = Vec::new();
3211 12 : let mut needs_cleanup = Vec::new();
3212 12 : let mut total_physical_size = 0;
3213 :
3214 44 : for (name, decision) in decided {
3215 32 : let decision = match decision {
3216 32 : Ok(decision) => decision,
3217 0 : Err(DismissedLayer::Future { local }) => {
3218 0 : if let Some(local) = local {
3219 0 : init::cleanup_future_layer(
3220 0 : &local.local_path,
3221 0 : &name,
3222 0 : disk_consistent_lsn,
3223 0 : )?;
3224 0 : }
3225 0 : needs_cleanup.push(name);
3226 0 : continue;
3227 : }
3228 0 : Err(DismissedLayer::LocalOnly(local)) => {
3229 0 : init::cleanup_local_only_file(&name, &local)?;
3230 : // this file never existed remotely, we will have to do rework
3231 0 : continue;
3232 : }
3233 0 : Err(DismissedLayer::BadMetadata(local)) => {
3234 0 : init::cleanup_local_file_for_remote(&local)?;
3235 : // this file never existed remotely, we will have to do rework
3236 0 : continue;
3237 : }
3238 : };
3239 :
3240 32 : match &name {
3241 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3242 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3243 : }
3244 :
3245 32 : tracing::debug!(layer=%name, ?decision, "applied");
3246 :
3247 32 : let layer = match decision {
3248 32 : Resident { local, remote } => {
3249 32 : total_physical_size += local.file_size;
3250 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3251 32 : .drop_eviction_guard()
3252 : }
3253 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3254 : };
3255 :
3256 32 : loaded_layers.push(layer);
3257 : }
3258 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3259 12 : }
3260 12 : })
3261 12 : .await
3262 12 : .map_err(anyhow::Error::new)
3263 12 : .and_then(|x| x)?;
3264 :
3265 12 : let num_layers = loaded_layers.len();
3266 12 :
3267 12 : guard
3268 12 : .open_mut()
3269 12 : .expect("layermanager must be open during init")
3270 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3271 12 :
3272 12 : self.remote_client
3273 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
3274 12 : self.remote_client
3275 12 : .schedule_index_upload_for_file_changes()?;
3276 : // This barrier orders above DELETEs before any later operations.
3277 : // This is critical because code executing after the barrier might
3278 : // create again objects with the same key that we just scheduled for deletion.
3279 : // For example, if we just scheduled deletion of an image layer "from the future",
3280 : // later compaction might run again and re-create the same image layer.
3281 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3282 : // "same" here means same key range and LSN.
3283 : //
3284 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3285 : // the upload queue may execute the PUT first, then the DELETE.
3286 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3287 : //
3288 : // 1. a future image layer is created and uploaded
3289 : // 2. ps restart
3290 : // 3. the future layer from (1) is deleted during load layer map
3291 : // 4. image layer is re-created and uploaded
3292 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3293 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3294 : //
3295 : // See https://github.com/neondatabase/neon/issues/5878
3296 : //
3297 : // NB: generation numbers naturally protect against this because they disambiguate
3298 : // (1) and (4)
3299 : // TODO: this is basically a no-op now, should we remove it?
3300 12 : self.remote_client.schedule_barrier()?;
3301 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
3302 : // on retry.
3303 :
3304 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3305 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3306 12 : self.update_layer_visibility().await?;
3307 :
3308 12 : info!(
3309 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3310 : num_layers, disk_consistent_lsn, total_physical_size
3311 : );
3312 :
3313 12 : timer.stop_and_record();
3314 12 : Ok(())
3315 12 : }
3316 :
3317 : /// Retrieve current logical size of the timeline.
3318 : ///
3319 : /// The size could be lagging behind the actual number, in case
3320 : /// the initial size calculation has not been run (gets triggered on the first size access).
3321 : ///
3322 : /// return size and boolean flag that shows if the size is exact
3323 0 : pub(crate) fn get_current_logical_size(
3324 0 : self: &Arc<Self>,
3325 0 : priority: GetLogicalSizePriority,
3326 0 : ctx: &RequestContext,
3327 0 : ) -> logical_size::CurrentLogicalSize {
3328 0 : if !self.tenant_shard_id.is_shard_zero() {
3329 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3330 : // when HTTP API is serving a GET for timeline zero, return zero
3331 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3332 0 : }
3333 0 :
3334 0 : let current_size = self.current_logical_size.current_size();
3335 0 : debug!("Current size: {current_size:?}");
3336 :
3337 0 : match (current_size.accuracy(), priority) {
3338 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3339 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3340 0 : // background task will eventually deliver an exact value, we're in no rush
3341 0 : }
3342 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3343 : // background task is not ready, but user is asking for it now;
3344 : // => make the background task skip the line
3345 : // (The alternative would be to calculate the size here, but,
3346 : // it can actually take a long time if the user has a lot of rels.
3347 : // And we'll inevitable need it again; So, let the background task do the work.)
3348 0 : match self
3349 0 : .current_logical_size
3350 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3351 0 : .get()
3352 : {
3353 0 : Some(cancel) => cancel.cancel(),
3354 : None => {
3355 0 : match self.current_state() {
3356 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3357 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3358 0 : // Don't make noise.
3359 0 : }
3360 : TimelineState::Loading => {
3361 : // Import does not return an activated timeline.
3362 0 : info!(
3363 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3364 : );
3365 : }
3366 : TimelineState::Active => {
3367 : // activation should be setting the once cell
3368 0 : warn!(
3369 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3370 : );
3371 0 : debug_assert!(false);
3372 : }
3373 : }
3374 : }
3375 : }
3376 : }
3377 : }
3378 :
3379 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3380 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3381 0 : let first = self
3382 0 : .current_logical_size
3383 0 : .did_return_approximate_to_walreceiver
3384 0 : .compare_exchange(
3385 0 : false,
3386 0 : true,
3387 0 : AtomicOrdering::Relaxed,
3388 0 : AtomicOrdering::Relaxed,
3389 0 : )
3390 0 : .is_ok();
3391 0 : if first {
3392 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3393 0 : }
3394 0 : }
3395 0 : }
3396 :
3397 0 : current_size
3398 0 : }
3399 :
3400 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3401 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3402 : // nothing to do for freshly created timelines;
3403 0 : assert_eq!(
3404 0 : self.current_logical_size.current_size().accuracy(),
3405 0 : logical_size::Accuracy::Exact,
3406 0 : );
3407 0 : self.current_logical_size.initialized.add_permits(1);
3408 0 : return;
3409 : };
3410 :
3411 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3412 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3413 0 : self.current_logical_size
3414 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3415 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3416 0 :
3417 0 : let self_clone = Arc::clone(self);
3418 0 : let background_ctx = ctx.detached_child(
3419 0 : TaskKind::InitialLogicalSizeCalculation,
3420 0 : DownloadBehavior::Download,
3421 0 : );
3422 0 : task_mgr::spawn(
3423 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3424 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3425 0 : self.tenant_shard_id,
3426 0 : Some(self.timeline_id),
3427 0 : "initial size calculation",
3428 : // NB: don't log errors here, task_mgr will do that.
3429 0 : async move {
3430 0 : self_clone
3431 0 : .initial_logical_size_calculation_task(
3432 0 : initial_part_end,
3433 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3434 0 : background_ctx,
3435 0 : )
3436 0 : .await;
3437 0 : Ok(())
3438 0 : }
3439 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)),
3440 : );
3441 0 : }
3442 :
3443 : /// # Cancellation
3444 : ///
3445 : /// This method is sensitive to `Timeline::cancel`.
3446 : ///
3447 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3448 : ///
3449 : /// # Cancel-Safety
3450 : ///
3451 : /// It does Timeline IO, hence this should be polled to completion because
3452 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3453 : /// to reason about.
3454 0 : async fn initial_logical_size_calculation_task(
3455 0 : self: Arc<Self>,
3456 0 : initial_part_end: Lsn,
3457 0 : skip_concurrency_limiter: CancellationToken,
3458 0 : background_ctx: RequestContext,
3459 0 : ) {
3460 0 : scopeguard::defer! {
3461 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3462 0 : self.current_logical_size.initialized.add_permits(1);
3463 0 : }
3464 0 :
3465 0 : let try_once = |attempt: usize| {
3466 0 : let background_ctx = &background_ctx;
3467 0 : let self_ref = &self;
3468 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3469 0 : async move {
3470 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3471 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3472 0 : background_ctx,
3473 0 : );
3474 :
3475 : use crate::metrics::initial_logical_size::StartCircumstances;
3476 0 : let (_maybe_permit, circumstances) = tokio::select! {
3477 0 : permit = wait_for_permit => {
3478 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3479 : }
3480 0 : _ = self_ref.cancel.cancelled() => {
3481 0 : return Err(CalculateLogicalSizeError::Cancelled);
3482 : }
3483 0 : () = skip_concurrency_limiter.cancelled() => {
3484 : // Some action that is part of a end user interaction requested logical size
3485 : // => break out of the rate limit
3486 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3487 : // but then again what happens if they cancel; also, we should just be using
3488 : // one runtime across the entire process, so, let's leave this for now.
3489 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3490 : }
3491 : };
3492 :
3493 0 : let metrics_guard = if attempt == 1 {
3494 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3495 : } else {
3496 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3497 : };
3498 :
3499 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3500 0 : self_ref.conf,
3501 0 : self_ref
3502 0 : .gate
3503 0 : .enter()
3504 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3505 : );
3506 :
3507 0 : let calculated_size = self_ref
3508 0 : .logical_size_calculation_task(
3509 0 : initial_part_end,
3510 0 : LogicalSizeCalculationCause::Initial,
3511 0 : background_ctx,
3512 0 : )
3513 0 : .await?;
3514 :
3515 0 : self_ref
3516 0 : .trigger_aux_file_size_computation(
3517 0 : initial_part_end,
3518 0 : background_ctx,
3519 0 : io_concurrency,
3520 0 : )
3521 0 : .await?;
3522 :
3523 : // TODO: add aux file size to logical size
3524 :
3525 0 : Ok((calculated_size, metrics_guard))
3526 0 : }
3527 0 : };
3528 :
3529 0 : let retrying = async {
3530 0 : let mut attempt = 0;
3531 : loop {
3532 0 : attempt += 1;
3533 0 :
3534 0 : match try_once(attempt).await {
3535 0 : Ok(res) => return ControlFlow::Continue(res),
3536 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3537 : Err(
3538 0 : e @ (CalculateLogicalSizeError::Decode(_)
3539 0 : | CalculateLogicalSizeError::PageRead(_)),
3540 0 : ) => {
3541 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3542 : // exponential back-off doesn't make sense at these long intervals;
3543 : // use fixed retry interval with generous jitter instead
3544 0 : let sleep_duration = Duration::from_secs(
3545 0 : u64::try_from(
3546 0 : // 1hour base
3547 0 : (60_i64 * 60_i64)
3548 0 : // 10min jitter
3549 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3550 0 : )
3551 0 : .expect("10min < 1hour"),
3552 0 : );
3553 0 : tokio::select! {
3554 0 : _ = tokio::time::sleep(sleep_duration) => {}
3555 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3556 : }
3557 : }
3558 : }
3559 : }
3560 0 : };
3561 :
3562 0 : let (calculated_size, metrics_guard) = match retrying.await {
3563 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3564 0 : ControlFlow::Break(()) => return,
3565 : };
3566 :
3567 : // we cannot query current_logical_size.current_size() to know the current
3568 : // *negative* value, only truncated to u64.
3569 0 : let added = self
3570 0 : .current_logical_size
3571 0 : .size_added_after_initial
3572 0 : .load(AtomicOrdering::Relaxed);
3573 0 :
3574 0 : let sum = calculated_size.saturating_add_signed(added);
3575 0 :
3576 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3577 0 : self.metrics.current_logical_size_gauge.set(sum);
3578 0 :
3579 0 : self.current_logical_size
3580 0 : .initial_logical_size
3581 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3582 0 : .ok()
3583 0 : .expect("only this task sets it");
3584 0 : }
3585 :
3586 28 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3587 28 : self: &Arc<Self>,
3588 28 : lsn: Lsn,
3589 28 : cause: LogicalSizeCalculationCause,
3590 28 : ctx: RequestContext,
3591 28 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3592 28 : let (sender, receiver) = oneshot::channel();
3593 28 : let self_clone = Arc::clone(self);
3594 28 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3595 28 : // we should stop the size calculation work and return an error.
3596 28 : // That would require restructuring this function's API to
3597 28 : // return the result directly, instead of a Receiver for the result.
3598 28 : let ctx = ctx.detached_child(
3599 28 : TaskKind::OndemandLogicalSizeCalculation,
3600 28 : DownloadBehavior::Download,
3601 28 : );
3602 28 : task_mgr::spawn(
3603 28 : task_mgr::BACKGROUND_RUNTIME.handle(),
3604 28 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3605 28 : self.tenant_shard_id,
3606 28 : Some(self.timeline_id),
3607 28 : "ondemand logical size calculation",
3608 28 : async move {
3609 28 : let res = self_clone
3610 28 : .logical_size_calculation_task(lsn, cause, &ctx)
3611 28 : .await;
3612 28 : let _ = sender.send(res).ok();
3613 28 : Ok(()) // Receiver is responsible for handling errors
3614 28 : }
3615 28 : .in_current_span(),
3616 28 : );
3617 28 : receiver
3618 28 : }
3619 :
3620 : #[instrument(skip_all)]
3621 : async fn logical_size_calculation_task(
3622 : self: &Arc<Self>,
3623 : lsn: Lsn,
3624 : cause: LogicalSizeCalculationCause,
3625 : ctx: &RequestContext,
3626 : ) -> Result<u64, CalculateLogicalSizeError> {
3627 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3628 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3629 : // accurate relation sizes, and they do not emit consumption metrics.
3630 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3631 :
3632 : let guard = self
3633 : .gate
3634 : .enter()
3635 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3636 :
3637 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3638 : }
3639 :
3640 : /// Calculate the logical size of the database at the latest LSN.
3641 : ///
3642 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3643 : /// especially if we need to download remote layers.
3644 28 : async fn calculate_logical_size(
3645 28 : &self,
3646 28 : up_to_lsn: Lsn,
3647 28 : cause: LogicalSizeCalculationCause,
3648 28 : _guard: &GateGuard,
3649 28 : ctx: &RequestContext,
3650 28 : ) -> Result<u64, CalculateLogicalSizeError> {
3651 28 : info!(
3652 0 : "Calculating logical size for timeline {} at {}",
3653 : self.timeline_id, up_to_lsn
3654 : );
3655 :
3656 28 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3657 : {
3658 0 : return Err(CalculateLogicalSizeError::Cancelled);
3659 28 : }
3660 :
3661 : // See if we've already done the work for initial size calculation.
3662 : // This is a short-cut for timelines that are mostly unused.
3663 28 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3664 0 : return Ok(size);
3665 28 : }
3666 28 : let storage_time_metrics = match cause {
3667 : LogicalSizeCalculationCause::Initial
3668 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3669 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3670 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3671 28 : &self.metrics.imitate_logical_size_histo
3672 : }
3673 : };
3674 28 : let timer = storage_time_metrics.start_timer();
3675 28 : let logical_size = self
3676 28 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3677 28 : .await?;
3678 28 : debug!("calculated logical size: {logical_size}");
3679 28 : timer.stop_and_record();
3680 28 : Ok(logical_size)
3681 28 : }
3682 :
3683 : /// Update current logical size, adding `delta' to the old value.
3684 541140 : fn update_current_logical_size(&self, delta: i64) {
3685 541140 : let logical_size = &self.current_logical_size;
3686 541140 : logical_size.increment_size(delta);
3687 541140 :
3688 541140 : // Also set the value in the prometheus gauge. Note that
3689 541140 : // there is a race condition here: if this is is called by two
3690 541140 : // threads concurrently, the prometheus gauge might be set to
3691 541140 : // one value while current_logical_size is set to the
3692 541140 : // other.
3693 541140 : match logical_size.current_size() {
3694 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3695 541140 : .metrics
3696 541140 : .current_logical_size_gauge
3697 541140 : .set(new_current_size.into()),
3698 0 : CurrentLogicalSize::Approximate(_) => {
3699 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3700 0 : // forth between the initial size calculation task.
3701 0 : }
3702 : }
3703 541140 : }
3704 :
3705 6028 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3706 6028 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3707 6028 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3708 6028 : // would overwrite the previous value if they modify different databases.
3709 6028 :
3710 6028 : match count {
3711 2184 : MetricsUpdate::Set(count) => {
3712 2184 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3713 2184 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3714 2184 : }
3715 3840 : MetricsUpdate::Add(count) => {
3716 3840 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3717 3840 : // it's fine.
3718 3840 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3719 3840 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3720 3840 : // the value reliably.
3721 3840 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3722 3840 : }
3723 : // Otherwise, ignore this update
3724 : }
3725 4 : MetricsUpdate::Sub(count) => {
3726 4 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3727 4 : // it's fine.
3728 4 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3729 4 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3730 4 : // The operation could overflow so we need to normalize the value.
3731 4 : let prev_val =
3732 4 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3733 4 : let res = prev_val.saturating_sub(count);
3734 4 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3735 4 : }
3736 : // Otherwise, ignore this update
3737 : }
3738 : };
3739 :
3740 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3741 6028 : let aux_metric =
3742 6028 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3743 6028 :
3744 6028 : let sum_of_entries = self
3745 6028 : .directory_metrics
3746 6028 : .iter()
3747 48224 : .map(|v| v.load(AtomicOrdering::Relaxed))
3748 6028 : .sum();
3749 : // Set a high general threshold and a lower threshold for the auxiliary files,
3750 : // as we can have large numbers of relations in the db directory.
3751 : const SUM_THRESHOLD: u64 = 5000;
3752 : const AUX_THRESHOLD: u64 = 1000;
3753 6028 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3754 0 : self.metrics
3755 0 : .directory_entries_count_gauge
3756 0 : .set(sum_of_entries);
3757 6028 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3758 0 : metric.set(sum_of_entries);
3759 6028 : }
3760 6028 : }
3761 :
3762 0 : async fn find_layer(
3763 0 : &self,
3764 0 : layer_name: &LayerName,
3765 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3766 0 : let guard = self.layers.read().await;
3767 0 : let layer = guard
3768 0 : .layer_map()?
3769 0 : .iter_historic_layers()
3770 0 : .find(|l| &l.layer_name() == layer_name)
3771 0 : .map(|found| guard.get_from_desc(&found));
3772 0 : Ok(layer)
3773 0 : }
3774 :
3775 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
3776 0 : let crnt = self.previous_heatmap.load();
3777 0 : match crnt.as_deref() {
3778 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
3779 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
3780 0 : None => true,
3781 : },
3782 0 : Some(PreviousHeatmap::Obsolete) => false,
3783 0 : None => false,
3784 : }
3785 0 : }
3786 :
3787 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3788 : /// indicating which layers are currently on-disk on the primary.
3789 : ///
3790 : /// None is returned if the Timeline is in a state where uploading a heatmap
3791 : /// doesn't make sense, such as shutting down or initializing. The caller
3792 : /// should treat this as a cue to simply skip doing any heatmap uploading
3793 : /// for this timeline.
3794 32 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3795 32 : if !self.is_active() {
3796 0 : return None;
3797 32 : }
3798 :
3799 32 : let guard = self.layers.read().await;
3800 :
3801 : // Firstly, if there's any heatmap left over from when this location
3802 : // was a secondary, take that into account. Keep layers that are:
3803 : // * present in the layer map
3804 : // * visible
3805 : // * non-resident
3806 : // * not evicted since we read the heatmap
3807 : //
3808 : // Without this, a new cold, attached location would clobber the previous
3809 : // heatamp.
3810 32 : let previous_heatmap = self.previous_heatmap.load();
3811 32 : let visible_non_resident = match previous_heatmap.as_deref() {
3812 : Some(PreviousHeatmap::Active {
3813 24 : heatmap, read_at, ..
3814 92 : }) => Some(heatmap.all_layers().filter_map(|hl| {
3815 92 : let desc: PersistentLayerDesc = hl.name.clone().into();
3816 92 : let layer = guard.try_get_from_key(&desc.key())?;
3817 :
3818 92 : if layer.visibility() == LayerVisibilityHint::Covered {
3819 0 : return None;
3820 92 : }
3821 92 :
3822 92 : if layer.is_likely_resident() {
3823 40 : return None;
3824 52 : }
3825 52 :
3826 52 : if layer.last_evicted_at().happened_after(*read_at) {
3827 12 : return None;
3828 40 : }
3829 40 :
3830 40 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
3831 92 : })),
3832 0 : Some(PreviousHeatmap::Obsolete) => None,
3833 8 : None => None,
3834 : };
3835 :
3836 : // Secondly, all currently visible, resident layers are included.
3837 72 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3838 72 : match layer.visibility() {
3839 : LayerVisibilityHint::Visible => {
3840 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3841 68 : let last_activity_ts = layer.latest_activity();
3842 68 : Some((
3843 68 : layer.layer_desc().clone(),
3844 68 : layer.metadata(),
3845 68 : last_activity_ts,
3846 68 : false, // these layers are not cold
3847 68 : ))
3848 : }
3849 : LayerVisibilityHint::Covered => {
3850 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3851 4 : None
3852 : }
3853 : }
3854 72 : });
3855 :
3856 32 : let mut layers = match visible_non_resident {
3857 24 : Some(non_resident) => {
3858 24 : let mut non_resident = non_resident.peekable();
3859 24 : if non_resident.peek().is_none() {
3860 8 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
3861 8 : self.previous_heatmap
3862 8 : .store(Some(PreviousHeatmap::Obsolete.into()));
3863 16 : }
3864 :
3865 24 : non_resident.chain(resident).collect::<Vec<_>>()
3866 : }
3867 8 : None => resident.collect::<Vec<_>>(),
3868 : };
3869 :
3870 : // Sort layers in order of which to download first. For a large set of layers to download, we
3871 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3872 : // or hours later:
3873 : // - Cold layers go last for convenience when a human inspects the heatmap.
3874 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3875 : // only exist for a few minutes before being compacted into L1s.
3876 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3877 : // the layer is likely to be covered by an image layer during compaction.
3878 238 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
3879 238 : std::cmp::Reverse((
3880 238 : *cold,
3881 238 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3882 238 : desc.lsn_range.end,
3883 238 : ))
3884 238 : });
3885 32 :
3886 32 : let layers = layers
3887 32 : .into_iter()
3888 108 : .map(|(desc, meta, atime, cold)| {
3889 108 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
3890 108 : })
3891 32 : .collect();
3892 32 :
3893 32 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3894 32 : }
3895 :
3896 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
3897 0 : let guard = self.layers.read().await;
3898 :
3899 0 : let now = SystemTime::now();
3900 0 : let mut heatmap_layers = Vec::default();
3901 0 : for vl in guard.visible_layers() {
3902 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
3903 0 : continue;
3904 0 : }
3905 0 :
3906 0 : let hl = HeatMapLayer {
3907 0 : name: vl.layer_desc().layer_name(),
3908 0 : metadata: vl.metadata(),
3909 0 : access_time: now,
3910 0 : cold: true,
3911 0 : };
3912 0 : heatmap_layers.push(hl);
3913 : }
3914 :
3915 0 : tracing::info!(
3916 0 : "Generating unarchival heatmap with {} layers",
3917 0 : heatmap_layers.len()
3918 : );
3919 :
3920 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
3921 0 : PreviousHeatmap::Active {
3922 0 : heatmap,
3923 0 : read_at: Instant::now(),
3924 0 : end_lsn: Some(end_lsn),
3925 0 : }
3926 0 : }
3927 :
3928 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3929 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3930 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3931 0 : // branchpoint in the value in IndexPart::lineage
3932 0 : self.ancestor_lsn == lsn
3933 0 : || (self.ancestor_lsn == Lsn::INVALID
3934 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3935 0 : }
3936 : }
3937 :
3938 : impl Timeline {
3939 : #[allow(clippy::doc_lazy_continuation)]
3940 : /// Get the data needed to reconstruct all keys in the provided keyspace
3941 : ///
3942 : /// The algorithm is as follows:
3943 : /// 1. While some keys are still not done and there's a timeline to visit:
3944 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3945 : /// 2.1: Build the fringe for the current keyspace
3946 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3947 : /// intersects
3948 : /// 2.3. Pop the timeline from the fringe
3949 : /// 2.4. If the fringe is empty, go back to 1
3950 1255534 : async fn get_vectored_reconstruct_data(
3951 1255534 : &self,
3952 1255534 : mut keyspace: KeySpace,
3953 1255534 : request_lsn: Lsn,
3954 1255534 : reconstruct_state: &mut ValuesReconstructState,
3955 1255534 : ctx: &RequestContext,
3956 1255534 : ) -> Result<(), GetVectoredError> {
3957 1255534 : let mut timeline_owned: Arc<Timeline>;
3958 1255534 : let mut timeline = self;
3959 1255534 :
3960 1255534 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3961 :
3962 1255530 : let missing_keyspace = loop {
3963 1707333 : if self.cancel.is_cancelled() {
3964 0 : return Err(GetVectoredError::Cancelled);
3965 1707333 : }
3966 :
3967 : let TimelineVisitOutcome {
3968 1707333 : completed_keyspace: completed,
3969 1707333 : image_covered_keyspace,
3970 : } = {
3971 1707333 : let ctx = RequestContextBuilder::from(ctx)
3972 1707333 : .perf_span(|crnt_perf_span| {
3973 0 : info_span!(
3974 : target: PERF_TRACE_TARGET,
3975 0 : parent: crnt_perf_span,
3976 : "PLAN_IO_TIMELINE",
3977 : timeline = %timeline.timeline_id,
3978 : lsn = %cont_lsn,
3979 : )
3980 1707333 : })
3981 1707333 : .attached_child();
3982 1707333 :
3983 1707333 : Self::get_vectored_reconstruct_data_timeline(
3984 1707333 : timeline,
3985 1707333 : keyspace.clone(),
3986 1707333 : cont_lsn,
3987 1707333 : reconstruct_state,
3988 1707333 : &self.cancel,
3989 1707333 : &ctx,
3990 1707333 : )
3991 1707333 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
3992 1707333 : .await?
3993 : };
3994 :
3995 1707333 : keyspace.remove_overlapping_with(&completed);
3996 1707333 :
3997 1707333 : // Do not descend into the ancestor timeline for aux files.
3998 1707333 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3999 1707333 : // stalling compaction.
4000 1707333 : keyspace.remove_overlapping_with(&KeySpace {
4001 1707333 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4002 1707333 : });
4003 1707333 :
4004 1707333 : // Keyspace is fully retrieved
4005 1707333 : if keyspace.is_empty() {
4006 1255014 : break None;
4007 452319 : }
4008 :
4009 452319 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4010 : // Not fully retrieved but no ancestor timeline.
4011 516 : break Some(keyspace);
4012 : };
4013 :
4014 : // Now we see if there are keys covered by the image layer but does not exist in the
4015 : // image layer, which means that the key does not exist.
4016 :
4017 : // The block below will stop the vectored search if any of the keys encountered an image layer
4018 : // which did not contain a snapshot for said key. Since we have already removed all completed
4019 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4020 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4021 : // and stop the search as a result of that.
4022 451803 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
4023 451803 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4024 451803 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4025 451803 : // figuring out what is the inherited key range and do a fine-grained pruning.
4026 451803 : removed.remove_overlapping_with(&KeySpace {
4027 451803 : ranges: vec![SPARSE_RANGE],
4028 451803 : });
4029 451803 : if !removed.is_empty() {
4030 0 : break Some(removed);
4031 451803 : }
4032 451803 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
4033 451803 : // keyspace.
4034 451803 :
4035 451803 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
4036 451803 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
4037 451803 :
4038 451803 : let ctx = RequestContextBuilder::from(ctx)
4039 451803 : .perf_span(|crnt_perf_span| {
4040 0 : info_span!(
4041 : target: PERF_TRACE_TARGET,
4042 0 : parent: crnt_perf_span,
4043 : "GET_ANCESTOR",
4044 : timeline = %timeline.timeline_id,
4045 : lsn = %cont_lsn,
4046 0 : ancestor = %ancestor_timeline.timeline_id,
4047 : ancestor_lsn = %timeline.ancestor_lsn
4048 : )
4049 451803 : })
4050 451803 : .attached_child();
4051 :
4052 451803 : timeline_owned = timeline
4053 451803 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4054 451803 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4055 451803 : .await?;
4056 451799 : timeline = &*timeline_owned;
4057 : };
4058 :
4059 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4060 1255530 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4061 516 : let mut missing_keyspace = missing_keyspace;
4062 516 : missing_keyspace.remove_overlapping_with(&KeySpace {
4063 516 : ranges: vec![SPARSE_RANGE],
4064 516 : });
4065 516 : if missing_keyspace.is_empty() {
4066 488 : None
4067 : } else {
4068 28 : Some(missing_keyspace)
4069 : }
4070 : } else {
4071 1255014 : None
4072 : };
4073 :
4074 1255530 : if let Some(missing_keyspace) = missing_keyspace {
4075 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
4076 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
4077 28 : shard: self
4078 28 : .shard_identity
4079 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
4080 28 : cont_lsn,
4081 28 : request_lsn,
4082 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
4083 28 : backtrace: None,
4084 28 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4085 28 : }));
4086 1255502 : }
4087 1255502 :
4088 1255502 : Ok(())
4089 1255534 : }
4090 :
4091 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4092 : ///
4093 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4094 : /// the current keyspace. The current keyspace of the search at any given timeline
4095 : /// is the original keyspace minus all the keys that have been completed minus
4096 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4097 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4098 : ///
4099 : /// This is basically a depth-first search visitor implementation where a vertex
4100 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4101 : ///
4102 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4103 : /// and get all the required reconstruct data from the layer in one go.
4104 : ///
4105 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4106 : /// decides how to deal with these two keyspaces.
4107 1707333 : async fn get_vectored_reconstruct_data_timeline(
4108 1707333 : timeline: &Timeline,
4109 1707333 : keyspace: KeySpace,
4110 1707333 : mut cont_lsn: Lsn,
4111 1707333 : reconstruct_state: &mut ValuesReconstructState,
4112 1707333 : cancel: &CancellationToken,
4113 1707333 : ctx: &RequestContext,
4114 1707333 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4115 1707333 : // Prevent GC from progressing while visiting the current timeline.
4116 1707333 : // If we are GC-ing because a new image layer was added while traversing
4117 1707333 : // the timeline, then it will remove layers that are required for fulfilling
4118 1707333 : // the current get request (read-path cannot "look back" and notice the new
4119 1707333 : // image layer).
4120 1707333 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4121 :
4122 : // See `compaction::compact_with_gc` for why we need this.
4123 1707333 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4124 :
4125 : // Initialize the fringe
4126 1707333 : let mut fringe = {
4127 1707333 : let mut fringe = LayerFringe::new();
4128 :
4129 1707333 : let guard = timeline.layers.read().await;
4130 1707333 : guard.update_search_fringe(&keyspace, cont_lsn, &mut fringe)?;
4131 :
4132 1707333 : fringe
4133 1707333 : };
4134 1707333 :
4135 1707333 : let mut completed_keyspace = KeySpace::default();
4136 1707333 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4137 :
4138 3401222 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4139 1693889 : if cancel.is_cancelled() {
4140 0 : return Err(GetVectoredError::Cancelled);
4141 1693889 : }
4142 :
4143 1693889 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4144 1693889 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4145 1693889 : }
4146 :
4147 : // Visit the layer and plan IOs for it
4148 1693889 : let next_cont_lsn = lsn_range.start;
4149 1693889 : layer_to_read
4150 1693889 : .get_values_reconstruct_data(
4151 1693889 : keyspace_to_read.clone(),
4152 1693889 : lsn_range,
4153 1693889 : reconstruct_state,
4154 1693889 : ctx,
4155 1693889 : )
4156 1693889 : .await?;
4157 :
4158 1693889 : let mut unmapped_keyspace = keyspace_to_read;
4159 1693889 : cont_lsn = next_cont_lsn;
4160 1693889 :
4161 1693889 : reconstruct_state.on_layer_visited(&layer_to_read);
4162 1693889 :
4163 1693889 : let (keys_done_last_step, keys_with_image_coverage) =
4164 1693889 : reconstruct_state.consume_done_keys();
4165 1693889 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4166 1693889 : completed_keyspace.merge(&keys_done_last_step);
4167 1693889 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4168 45128 : unmapped_keyspace
4169 45128 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4170 45128 : image_covered_keyspace.add_range(keys_with_image_coverage);
4171 1648761 : }
4172 :
4173 : // Query the layer map for the next layers to read.
4174 : //
4175 : // Do not descent any further if the last layer we visited
4176 : // completed all keys in the keyspace it inspected. This is not
4177 : // required for correctness, but avoids visiting extra layers
4178 : // which turns out to be a perf bottleneck in some cases.
4179 1693889 : if !unmapped_keyspace.is_empty() {
4180 442935 : let guard = timeline.layers.read().await;
4181 442935 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4182 :
4183 : // It's safe to drop the layer map lock after planning the next round of reads.
4184 : // The fringe keeps readable handles for the layers which are safe to read even
4185 : // if layers were compacted or flushed.
4186 : //
4187 : // The more interesting consideration is: "Why is the read algorithm still correct
4188 : // if the layer map changes while it is operating?". Doing a vectored read on a
4189 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4190 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4191 : // range at *a particular point in time*. It is fine for the answer to be different
4192 : // at two different time points.
4193 442935 : drop(guard);
4194 1250954 : }
4195 : }
4196 :
4197 1707333 : Ok(TimelineVisitOutcome {
4198 1707333 : completed_keyspace,
4199 1707333 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4200 1707333 : })
4201 1707333 : }
4202 :
4203 451803 : async fn get_ready_ancestor_timeline(
4204 451803 : &self,
4205 451803 : ancestor: &Arc<Timeline>,
4206 451803 : ctx: &RequestContext,
4207 451803 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4208 451803 : // It's possible that the ancestor timeline isn't active yet, or
4209 451803 : // is active but hasn't yet caught up to the branch point. Wait
4210 451803 : // for it.
4211 451803 : //
4212 451803 : // This cannot happen while the pageserver is running normally,
4213 451803 : // because you cannot create a branch from a point that isn't
4214 451803 : // present in the pageserver yet. However, we don't wait for the
4215 451803 : // branch point to be uploaded to cloud storage before creating
4216 451803 : // a branch. I.e., the branch LSN need not be remote consistent
4217 451803 : // for the branching operation to succeed.
4218 451803 : //
4219 451803 : // Hence, if we try to load a tenant in such a state where
4220 451803 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4221 451803 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4222 451803 : // then we will need to wait for the ancestor timeline to
4223 451803 : // re-stream WAL up to branch_lsn before we access it.
4224 451803 : //
4225 451803 : // How can a tenant get in such a state?
4226 451803 : // - ungraceful pageserver process exit
4227 451803 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4228 451803 : //
4229 451803 : // NB: this could be avoided by requiring
4230 451803 : // branch_lsn >= remote_consistent_lsn
4231 451803 : // during branch creation.
4232 451803 : match ancestor.wait_to_become_active(ctx).await {
4233 451799 : Ok(()) => {}
4234 : Err(TimelineState::Stopping) => {
4235 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4236 0 : return Err(GetReadyAncestorError::Cancelled);
4237 : }
4238 4 : Err(state) => {
4239 4 : return Err(GetReadyAncestorError::BadState {
4240 4 : timeline_id: ancestor.timeline_id,
4241 4 : state,
4242 4 : });
4243 : }
4244 : }
4245 451799 : ancestor
4246 451799 : .wait_lsn(
4247 451799 : self.ancestor_lsn,
4248 451799 : WaitLsnWaiter::Timeline(self),
4249 451799 : WaitLsnTimeout::Default,
4250 451799 : ctx,
4251 451799 : )
4252 451799 : .await
4253 451799 : .map_err(|e| match e {
4254 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4255 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4256 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4257 0 : timeline_id: ancestor.timeline_id,
4258 0 : state,
4259 0 : },
4260 451799 : })?;
4261 :
4262 451799 : Ok(ancestor.clone())
4263 451803 : }
4264 :
4265 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4266 594312 : &self.shard_identity
4267 594312 : }
4268 :
4269 : #[inline(always)]
4270 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4271 0 : ShardTimelineId {
4272 0 : shard_index: ShardIndex {
4273 0 : shard_number: self.shard_identity.number,
4274 0 : shard_count: self.shard_identity.count,
4275 0 : },
4276 0 : timeline_id: self.timeline_id,
4277 0 : }
4278 0 : }
4279 :
4280 : /// Returns a non-frozen open in-memory layer for ingestion.
4281 : ///
4282 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4283 : /// this function without holding the mutex.
4284 2612 : async fn get_layer_for_write(
4285 2612 : &self,
4286 2612 : lsn: Lsn,
4287 2612 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4288 2612 : ctx: &RequestContext,
4289 2612 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4290 2612 : let mut guard = self.layers.write().await;
4291 :
4292 2612 : let last_record_lsn = self.get_last_record_lsn();
4293 2612 : ensure!(
4294 2612 : lsn > last_record_lsn,
4295 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4296 : lsn,
4297 : last_record_lsn,
4298 : );
4299 :
4300 2612 : let layer = guard
4301 2612 : .open_mut()?
4302 2612 : .get_layer_for_write(
4303 2612 : lsn,
4304 2612 : self.conf,
4305 2612 : self.timeline_id,
4306 2612 : self.tenant_shard_id,
4307 2612 : &self.gate,
4308 2612 : &self.cancel,
4309 2612 : ctx,
4310 2612 : )
4311 2612 : .await?;
4312 2612 : Ok(layer)
4313 2612 : }
4314 :
4315 10558204 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4316 10558204 : assert!(new_lsn.is_aligned());
4317 :
4318 10558204 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4319 10558204 : self.last_record_lsn.advance(new_lsn);
4320 10558204 : }
4321 :
4322 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4323 : ///
4324 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4325 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4326 2420 : async fn freeze_inmem_layer_at(
4327 2420 : &self,
4328 2420 : at: Lsn,
4329 2420 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4330 2420 : ) -> Result<u64, FlushLayerError> {
4331 2420 : let frozen = {
4332 2420 : let mut guard = self.layers.write().await;
4333 2420 : guard
4334 2420 : .open_mut()?
4335 2420 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4336 2420 : .await
4337 : };
4338 :
4339 2420 : if frozen {
4340 2364 : let now = Instant::now();
4341 2364 : *(self.last_freeze_ts.write().unwrap()) = now;
4342 2364 : }
4343 :
4344 : // Increment the flush cycle counter and wake up the flush task.
4345 : // Remember the new value, so that when we listen for the flush
4346 : // to finish, we know when the flush that we initiated has
4347 : // finished, instead of some other flush that was started earlier.
4348 2420 : let mut my_flush_request = 0;
4349 2420 :
4350 2420 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4351 2420 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4352 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4353 2420 : }
4354 2420 :
4355 2420 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4356 2420 : my_flush_request = *counter + 1;
4357 2420 : *counter = my_flush_request;
4358 2420 : *lsn = std::cmp::max(at, *lsn);
4359 2420 : });
4360 2420 :
4361 2420 : assert_ne!(my_flush_request, 0);
4362 :
4363 2420 : Ok(my_flush_request)
4364 2420 : }
4365 :
4366 : /// Layer flusher task's main loop.
4367 912 : async fn flush_loop(
4368 912 : self: &Arc<Self>,
4369 912 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4370 912 : ctx: &RequestContext,
4371 912 : ) {
4372 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4373 912 : let mut watch_l0 = match self.layers.read().await.layer_map() {
4374 912 : Ok(lm) => lm.watch_level0_deltas(),
4375 0 : Err(Shutdown) => return,
4376 : };
4377 :
4378 912 : info!("started flush loop");
4379 : loop {
4380 3279 : tokio::select! {
4381 3279 : _ = self.cancel.cancelled() => {
4382 20 : info!("shutting down layer flush task due to Timeline::cancel");
4383 20 : break;
4384 : },
4385 3279 : _ = layer_flush_start_rx.changed() => {}
4386 2367 : }
4387 2367 : trace!("waking up");
4388 2367 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4389 2367 :
4390 2367 : // The highest LSN to which we flushed in the loop over frozen layers
4391 2367 : let mut flushed_to_lsn = Lsn(0);
4392 :
4393 2367 : let result = loop {
4394 4731 : if self.cancel.is_cancelled() {
4395 0 : info!("dropping out of flush loop for timeline shutdown");
4396 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
4397 : // anyone waiting on that will respect self.cancel as well: they will stop
4398 : // waiting at the same time we as drop out of this loop.
4399 0 : return;
4400 4731 : }
4401 4731 :
4402 4731 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4403 4731 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4404 4731 : if flushed_to_lsn >= frozen_to_lsn {
4405 2311 : break Ok(());
4406 2420 : }
4407 :
4408 : // Fetch the next layer to flush, if any.
4409 2420 : let (layer, l0_count, frozen_count, frozen_size) = {
4410 2420 : let layers = self.layers.read().await;
4411 2420 : let Ok(lm) = layers.layer_map() else {
4412 0 : info!("dropping out of flush loop for timeline shutdown");
4413 0 : return;
4414 : };
4415 2420 : let l0_count = lm.level0_deltas().len();
4416 2420 : let frozen_count = lm.frozen_layers.len();
4417 2420 : let frozen_size: u64 = lm
4418 2420 : .frozen_layers
4419 2420 : .iter()
4420 2420 : .map(|l| l.estimated_in_mem_size())
4421 2420 : .sum();
4422 2420 : let layer = lm.frozen_layers.front().cloned();
4423 2420 : (layer, l0_count, frozen_count, frozen_size)
4424 2420 : // drop 'layers' lock
4425 2420 : };
4426 2420 : let Some(layer) = layer else {
4427 56 : break Ok(());
4428 : };
4429 :
4430 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4431 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4432 2364 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4433 0 : if l0_count >= stall_threshold {
4434 0 : warn!(
4435 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4436 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4437 : );
4438 0 : let stall_timer = self
4439 0 : .metrics
4440 0 : .flush_delay_histo
4441 0 : .start_timer()
4442 0 : .record_on_drop();
4443 0 : tokio::select! {
4444 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4445 0 : if let Ok(l0) = result.as_deref() {
4446 0 : let delay = stall_timer.elapsed().as_secs_f64();
4447 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4448 0 : }
4449 : },
4450 0 : _ = self.cancel.cancelled() => {},
4451 : }
4452 0 : continue; // check again
4453 0 : }
4454 2364 : }
4455 :
4456 : // Flush the layer.
4457 2364 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4458 2364 : match self.flush_frozen_layer(layer, ctx).await {
4459 2364 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4460 : Err(FlushLayerError::Cancelled) => {
4461 0 : info!("dropping out of flush loop for timeline shutdown");
4462 0 : return;
4463 : }
4464 0 : err @ Err(
4465 0 : FlushLayerError::NotRunning(_)
4466 0 : | FlushLayerError::Other(_)
4467 0 : | FlushLayerError::CreateImageLayersError(_),
4468 0 : ) => {
4469 0 : error!("could not flush frozen layer: {err:?}");
4470 0 : break err.map(|_| ());
4471 : }
4472 : }
4473 2364 : let flush_duration = flush_timer.stop_and_record();
4474 2364 :
4475 2364 : // Notify the tenant compaction loop if L0 compaction is needed.
4476 2364 : let l0_count = *watch_l0.borrow();
4477 2364 : if l0_count >= self.get_compaction_threshold() {
4478 956 : self.l0_compaction_trigger.notify_one();
4479 1408 : }
4480 :
4481 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4482 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4483 : // ingestion by having ephemeral layer rolls wait for flushes.
4484 2364 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4485 4 : if l0_count >= delay_threshold {
4486 0 : let delay = flush_duration.as_secs_f64();
4487 0 : info!(
4488 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4489 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4490 : );
4491 0 : let _delay_timer = self
4492 0 : .metrics
4493 0 : .flush_delay_histo
4494 0 : .start_timer()
4495 0 : .record_on_drop();
4496 0 : tokio::select! {
4497 0 : _ = tokio::time::sleep(flush_duration) => {},
4498 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4499 0 : _ = self.cancel.cancelled() => {},
4500 : }
4501 4 : }
4502 2360 : }
4503 : };
4504 :
4505 : // Unsharded tenants should never advance their LSN beyond the end of the
4506 : // highest layer they write: such gaps between layer data and the frozen LSN
4507 : // are only legal on sharded tenants.
4508 2367 : debug_assert!(
4509 2367 : self.shard_identity.count.count() > 1
4510 2367 : || flushed_to_lsn >= frozen_to_lsn
4511 56 : || !flushed_to_lsn.is_valid()
4512 : );
4513 :
4514 2367 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4515 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4516 : // to us via layer_flush_start_rx, then advance it here.
4517 : //
4518 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4519 : // never encounter a gap in the wal.
4520 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4521 0 : tracing::debug!(
4522 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
4523 : );
4524 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4525 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4526 0 : tracing::warn!(
4527 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
4528 : );
4529 0 : }
4530 0 : }
4531 2367 : }
4532 :
4533 : // Notify any listeners that we're done
4534 2367 : let _ = self
4535 2367 : .layer_flush_done_tx
4536 2367 : .send_replace((flush_counter, result));
4537 : }
4538 20 : }
4539 :
4540 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4541 2260 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4542 2260 : let mut rx = self.layer_flush_done_tx.subscribe();
4543 : loop {
4544 : {
4545 4592 : let (last_result_counter, last_result) = &*rx.borrow();
4546 4592 : if *last_result_counter >= request {
4547 2260 : if let Err(err) = last_result {
4548 : // We already logged the original error in
4549 : // flush_loop. We cannot propagate it to the caller
4550 : // here, because it might not be Cloneable
4551 0 : return Err(err.clone());
4552 : } else {
4553 2260 : return Ok(());
4554 : }
4555 2332 : }
4556 2332 : }
4557 2332 : trace!("waiting for flush to complete");
4558 2332 : tokio::select! {
4559 2332 : rx_e = rx.changed() => {
4560 2332 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4561 : },
4562 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4563 : // the notification from [`flush_loop`] that it completed.
4564 2332 : _ = self.cancel.cancelled() => {
4565 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4566 0 : return Ok(())
4567 : }
4568 : };
4569 2332 : trace!("done")
4570 : }
4571 2260 : }
4572 :
4573 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4574 : ///
4575 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4576 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4577 : async fn flush_frozen_layer(
4578 : self: &Arc<Self>,
4579 : frozen_layer: Arc<InMemoryLayer>,
4580 : ctx: &RequestContext,
4581 : ) -> Result<Lsn, FlushLayerError> {
4582 : debug_assert_current_span_has_tenant_and_timeline_id();
4583 :
4584 : // As a special case, when we have just imported an image into the repository,
4585 : // instead of writing out a L0 delta layer, we directly write out image layer
4586 : // files instead. This is possible as long as *all* the data imported into the
4587 : // repository have the same LSN.
4588 : let lsn_range = frozen_layer.get_lsn_range();
4589 :
4590 : // Whether to directly create image layers for this flush, or flush them as delta layers
4591 : let create_image_layer =
4592 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4593 :
4594 : #[cfg(test)]
4595 : {
4596 : match &mut *self.flush_loop_state.lock().unwrap() {
4597 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4598 : panic!("flush loop not running")
4599 : }
4600 : FlushLoopState::Running {
4601 : expect_initdb_optimization,
4602 : initdb_optimization_count,
4603 : ..
4604 : } => {
4605 : if create_image_layer {
4606 : *initdb_optimization_count += 1;
4607 : } else {
4608 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4609 : }
4610 : }
4611 : }
4612 : }
4613 :
4614 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4615 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4616 : // require downloading anything during initial import.
4617 : let ((rel_partition, metadata_partition), _lsn) = self
4618 : .repartition(
4619 : self.initdb_lsn,
4620 : self.get_compaction_target_size(),
4621 : EnumSet::empty(),
4622 : ctx,
4623 : )
4624 : .await
4625 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4626 :
4627 : if self.cancel.is_cancelled() {
4628 : return Err(FlushLayerError::Cancelled);
4629 : }
4630 :
4631 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4632 : // So that the key ranges don't overlap.
4633 : let mut partitions = KeyPartitioning::default();
4634 : partitions.parts.extend(rel_partition.parts);
4635 : if !metadata_partition.parts.is_empty() {
4636 : assert_eq!(
4637 : metadata_partition.parts.len(),
4638 : 1,
4639 : "currently sparse keyspace should only contain a single metadata keyspace"
4640 : );
4641 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4642 : // every single key within the keyspace, and therefore, it's safe to force converting it
4643 : // into a dense keyspace before calling this function.
4644 : partitions
4645 : .parts
4646 : .extend(metadata_partition.into_dense().parts);
4647 : }
4648 :
4649 : let mut layers_to_upload = Vec::new();
4650 : let (generated_image_layers, is_complete) = self
4651 : .create_image_layers(
4652 : &partitions,
4653 : self.initdb_lsn,
4654 : ImageLayerCreationMode::Initial,
4655 : ctx,
4656 : LastImageLayerCreationStatus::Initial,
4657 : false, // don't yield for L0, we're flushing L0
4658 : )
4659 : .await?;
4660 : debug_assert!(
4661 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4662 : "init image generation mode must fully cover the keyspace"
4663 : );
4664 : layers_to_upload.extend(generated_image_layers);
4665 :
4666 : (layers_to_upload, None)
4667 : } else {
4668 : // Normal case, write out a L0 delta layer file.
4669 : // `create_delta_layer` will not modify the layer map.
4670 : // We will remove frozen layer and add delta layer in one atomic operation later.
4671 : let Some(layer) = self
4672 : .create_delta_layer(&frozen_layer, None, ctx)
4673 : .await
4674 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4675 : else {
4676 : panic!("delta layer cannot be empty if no filter is applied");
4677 : };
4678 : (
4679 : // FIXME: even though we have a single image and single delta layer assumption
4680 : // we push them to vec
4681 : vec![layer.clone()],
4682 : Some(layer),
4683 : )
4684 : };
4685 :
4686 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4687 :
4688 : if self.cancel.is_cancelled() {
4689 : return Err(FlushLayerError::Cancelled);
4690 : }
4691 :
4692 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4693 :
4694 : // The new on-disk layers are now in the layer map. We can remove the
4695 : // in-memory layer from the map now. The flushed layer is stored in
4696 : // the mapping in `create_delta_layer`.
4697 : {
4698 : let mut guard = self.layers.write().await;
4699 :
4700 : guard.open_mut()?.finish_flush_l0_layer(
4701 : delta_layer_to_add.as_ref(),
4702 : &frozen_layer,
4703 : &self.metrics,
4704 : );
4705 :
4706 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4707 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4708 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4709 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4710 : }
4711 : // release lock on 'layers'
4712 : };
4713 :
4714 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4715 : // a compaction can delete the file and then it won't be available for uploads any more.
4716 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4717 : // race situation.
4718 : // See https://github.com/neondatabase/neon/issues/4526
4719 : pausable_failpoint!("flush-frozen-pausable");
4720 :
4721 : // This failpoint is used by another test case `test_pageserver_recovery`.
4722 : fail_point!("flush-frozen-exit");
4723 :
4724 : Ok(Lsn(lsn_range.end.0 - 1))
4725 : }
4726 :
4727 : /// Return true if the value changed
4728 : ///
4729 : /// This function must only be used from the layer flush task.
4730 2364 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4731 2364 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4732 2364 : assert!(
4733 2364 : new_value >= old_value,
4734 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
4735 : );
4736 :
4737 2364 : self.metrics
4738 2364 : .disk_consistent_lsn_gauge
4739 2364 : .set(new_value.0 as i64);
4740 2364 : new_value != old_value
4741 2364 : }
4742 :
4743 : /// Update metadata file
4744 2464 : fn schedule_uploads(
4745 2464 : &self,
4746 2464 : disk_consistent_lsn: Lsn,
4747 2464 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4748 2464 : ) -> anyhow::Result<()> {
4749 2464 : // We can only save a valid 'prev_record_lsn' value on disk if we
4750 2464 : // flushed *all* in-memory changes to disk. We only track
4751 2464 : // 'prev_record_lsn' in memory for the latest processed record, so we
4752 2464 : // don't remember what the correct value that corresponds to some old
4753 2464 : // LSN is. But if we flush everything, then the value corresponding
4754 2464 : // current 'last_record_lsn' is correct and we can store it on disk.
4755 2464 : let RecordLsn {
4756 2464 : last: last_record_lsn,
4757 2464 : prev: prev_record_lsn,
4758 2464 : } = self.last_record_lsn.load();
4759 2464 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4760 2209 : Some(prev_record_lsn)
4761 : } else {
4762 255 : None
4763 : };
4764 :
4765 2464 : let update = crate::tenant::metadata::MetadataUpdate::new(
4766 2464 : disk_consistent_lsn,
4767 2464 : ondisk_prev_record_lsn,
4768 2464 : *self.applied_gc_cutoff_lsn.read(),
4769 2464 : );
4770 2464 :
4771 2464 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4772 0 : "{}",
4773 0 : x.unwrap()
4774 2464 : ));
4775 :
4776 4852 : for layer in layers_to_upload {
4777 2388 : self.remote_client.schedule_layer_file_upload(layer)?;
4778 : }
4779 2464 : self.remote_client
4780 2464 : .schedule_index_upload_for_metadata_update(&update)?;
4781 :
4782 2464 : Ok(())
4783 2464 : }
4784 :
4785 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4786 0 : self.remote_client
4787 0 : .preserve_initdb_archive(
4788 0 : &self.tenant_shard_id.tenant_id,
4789 0 : &self.timeline_id,
4790 0 : &self.cancel,
4791 0 : )
4792 0 : .await
4793 0 : }
4794 :
4795 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4796 : // in layer map immediately. The caller is responsible to put it into the layer map.
4797 1936 : async fn create_delta_layer(
4798 1936 : self: &Arc<Self>,
4799 1936 : frozen_layer: &Arc<InMemoryLayer>,
4800 1936 : key_range: Option<Range<Key>>,
4801 1936 : ctx: &RequestContext,
4802 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4803 1936 : let self_clone = Arc::clone(self);
4804 1936 : let frozen_layer = Arc::clone(frozen_layer);
4805 1936 : let ctx = ctx.attached_child();
4806 1936 : let work = async move {
4807 1936 : let Some((desc, path)) = frozen_layer
4808 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4809 1936 : .await?
4810 : else {
4811 0 : return Ok(None);
4812 : };
4813 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4814 :
4815 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4816 : // We just need to fsync the directory in which these inodes are linked,
4817 : // which we know to be the timeline directory.
4818 : //
4819 : // We use fatal_err() below because the after write_to_disk returns with success,
4820 : // the in-memory state of the filesystem already has the layer file in its final place,
4821 : // and subsequent pageserver code could think it's durable while it really isn't.
4822 1936 : let timeline_dir = VirtualFile::open(
4823 1936 : &self_clone
4824 1936 : .conf
4825 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4826 1936 : &ctx,
4827 1936 : )
4828 1936 : .await
4829 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4830 1936 : timeline_dir
4831 1936 : .sync_all()
4832 1936 : .await
4833 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4834 1936 : anyhow::Ok(Some(new_delta))
4835 1936 : };
4836 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4837 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4838 : use crate::virtual_file::io_engine::IoEngine;
4839 1936 : match crate::virtual_file::io_engine::get() {
4840 0 : IoEngine::NotSet => panic!("io engine not set"),
4841 : IoEngine::StdFs => {
4842 968 : let span = tracing::info_span!("blocking");
4843 968 : tokio::task::spawn_blocking({
4844 968 : move || Handle::current().block_on(work.instrument(span))
4845 968 : })
4846 968 : .await
4847 968 : .context("spawn_blocking")
4848 968 : .and_then(|x| x)
4849 : }
4850 : #[cfg(target_os = "linux")]
4851 968 : IoEngine::TokioEpollUring => work.await,
4852 : }
4853 1936 : }
4854 :
4855 1156 : async fn repartition(
4856 1156 : &self,
4857 1156 : lsn: Lsn,
4858 1156 : partition_size: u64,
4859 1156 : flags: EnumSet<CompactFlags>,
4860 1156 : ctx: &RequestContext,
4861 1156 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4862 1156 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4863 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4864 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4865 : // and hence before the compaction task starts.
4866 0 : return Err(CompactionError::Other(anyhow!(
4867 0 : "repartition() called concurrently"
4868 0 : )));
4869 : };
4870 1156 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4871 1156 : if lsn < *partition_lsn {
4872 0 : return Err(CompactionError::Other(anyhow!(
4873 0 : "repartition() called with LSN going backwards, this should not happen"
4874 0 : )));
4875 1156 : }
4876 1156 :
4877 1156 : let distance = lsn.0 - partition_lsn.0;
4878 1156 : if *partition_lsn != Lsn(0)
4879 524 : && distance <= self.repartition_threshold
4880 524 : && !flags.contains(CompactFlags::ForceRepartition)
4881 : {
4882 496 : debug!(
4883 : distance,
4884 : threshold = self.repartition_threshold,
4885 0 : "no repartitioning needed"
4886 : );
4887 496 : return Ok((
4888 496 : (dense_partition.clone(), sparse_partition.clone()),
4889 496 : *partition_lsn,
4890 496 : ));
4891 660 : }
4892 :
4893 660 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4894 660 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4895 660 : let sparse_partitioning = SparseKeyPartitioning {
4896 660 : parts: vec![sparse_ks],
4897 660 : }; // no partitioning for metadata keys for now
4898 660 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4899 660 : guard.write(result.clone());
4900 660 : Ok(result)
4901 1156 : }
4902 :
4903 : // Is it time to create a new image layer for the given partition? True if we want to generate.
4904 28 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4905 28 : let threshold = self.get_image_creation_threshold();
4906 :
4907 28 : let guard = self.layers.read().await;
4908 28 : let Ok(layers) = guard.layer_map() else {
4909 0 : return false;
4910 : };
4911 :
4912 28 : let mut max_deltas = 0;
4913 56 : for part_range in &partition.ranges {
4914 28 : let image_coverage = layers.image_coverage(part_range, lsn);
4915 56 : for (img_range, last_img) in image_coverage {
4916 28 : let img_lsn = if let Some(last_img) = last_img {
4917 0 : last_img.get_lsn_range().end
4918 : } else {
4919 28 : Lsn(0)
4920 : };
4921 : // Let's consider an example:
4922 : //
4923 : // delta layer with LSN range 71-81
4924 : // delta layer with LSN range 81-91
4925 : // delta layer with LSN range 91-101
4926 : // image layer at LSN 100
4927 : //
4928 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4929 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4930 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4931 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4932 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4933 28 : if img_lsn < lsn {
4934 28 : let num_deltas =
4935 28 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4936 28 :
4937 28 : max_deltas = max_deltas.max(num_deltas);
4938 28 : if num_deltas >= threshold {
4939 0 : debug!(
4940 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4941 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4942 : );
4943 0 : return true;
4944 28 : }
4945 0 : }
4946 : }
4947 : }
4948 :
4949 28 : debug!(
4950 : max_deltas,
4951 0 : "none of the partitioned ranges had >= {threshold} deltas"
4952 : );
4953 28 : false
4954 28 : }
4955 :
4956 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4957 : /// so that at most one image layer will be produced from this function.
4958 : #[allow(clippy::too_many_arguments)]
4959 480 : async fn create_image_layer_for_rel_blocks(
4960 480 : self: &Arc<Self>,
4961 480 : partition: &KeySpace,
4962 480 : mut image_layer_writer: ImageLayerWriter,
4963 480 : lsn: Lsn,
4964 480 : ctx: &RequestContext,
4965 480 : img_range: Range<Key>,
4966 480 : io_concurrency: IoConcurrency,
4967 480 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4968 480 : let mut wrote_keys = false;
4969 480 :
4970 480 : let mut key_request_accum = KeySpaceAccum::new();
4971 3184 : for range in &partition.ranges {
4972 2704 : let mut key = range.start;
4973 5860 : while key < range.end {
4974 : // Decide whether to retain this key: usually we do, but sharded tenants may
4975 : // need to drop keys that don't belong to them. If we retain the key, add it
4976 : // to `key_request_accum` for later issuing a vectored get
4977 3156 : if self.shard_identity.is_key_disposable(&key) {
4978 0 : debug!(
4979 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4980 0 : key,
4981 0 : self.shard_identity.get_shard_number(&key)
4982 : );
4983 3156 : } else {
4984 3156 : key_request_accum.add_key(key);
4985 3156 : }
4986 :
4987 3156 : let last_key_in_range = key.next() == range.end;
4988 3156 : key = key.next();
4989 3156 :
4990 3156 : // Maybe flush `key_rest_accum`
4991 3156 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4992 3156 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4993 : {
4994 2704 : let results = self
4995 2704 : .get_vectored(
4996 2704 : key_request_accum.consume_keyspace(),
4997 2704 : lsn,
4998 2704 : io_concurrency.clone(),
4999 2704 : ctx,
5000 2704 : )
5001 2704 : .await?;
5002 :
5003 2704 : if self.cancel.is_cancelled() {
5004 0 : return Err(CreateImageLayersError::Cancelled);
5005 2704 : }
5006 :
5007 5860 : for (img_key, img) in results {
5008 3156 : let img = match img {
5009 3156 : Ok(img) => img,
5010 0 : Err(err) => {
5011 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
5012 0 : // page without losing any actual user data. That seems better
5013 0 : // than failing repeatedly and getting stuck.
5014 0 : //
5015 0 : // We had a bug at one point, where we truncated the FSM and VM
5016 0 : // in the pageserver, but the Postgres didn't know about that
5017 0 : // and continued to generate incremental WAL records for pages
5018 0 : // that didn't exist in the pageserver. Trying to replay those
5019 0 : // WAL records failed to find the previous image of the page.
5020 0 : // This special case allows us to recover from that situation.
5021 0 : // See https://github.com/neondatabase/neon/issues/2601.
5022 0 : //
5023 0 : // Unfortunately we cannot do this for the main fork, or for
5024 0 : // any metadata keys, keys, as that would lead to actual data
5025 0 : // loss.
5026 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5027 0 : warn!(
5028 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5029 : );
5030 0 : ZERO_PAGE.clone()
5031 : } else {
5032 0 : return Err(CreateImageLayersError::from(err));
5033 : }
5034 : }
5035 : };
5036 :
5037 : // Write all the keys we just read into our new image layer.
5038 3156 : image_layer_writer.put_image(img_key, img, ctx).await?;
5039 3156 : wrote_keys = true;
5040 : }
5041 452 : }
5042 : }
5043 : }
5044 :
5045 480 : if wrote_keys {
5046 : // Normal path: we have written some data into the new image layer for this
5047 : // partition, so flush it to disk.
5048 480 : info!(
5049 0 : "produced image layer for rel {}",
5050 0 : ImageLayerName {
5051 0 : key_range: img_range.clone(),
5052 0 : lsn
5053 0 : },
5054 : );
5055 480 : Ok(ImageLayerCreationOutcome::Generated {
5056 480 : unfinished_image_layer: image_layer_writer,
5057 480 : })
5058 : } else {
5059 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5060 0 : Ok(ImageLayerCreationOutcome::Empty)
5061 : }
5062 480 : }
5063 :
5064 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5065 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5066 : /// would not be too large to fit in a single image layer.
5067 : ///
5068 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5069 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5070 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5071 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5072 : #[allow(clippy::too_many_arguments)]
5073 460 : async fn create_image_layer_for_metadata_keys(
5074 460 : self: &Arc<Self>,
5075 460 : partition: &KeySpace,
5076 460 : mut image_layer_writer: ImageLayerWriter,
5077 460 : lsn: Lsn,
5078 460 : ctx: &RequestContext,
5079 460 : img_range: Range<Key>,
5080 460 : mode: ImageLayerCreationMode,
5081 460 : io_concurrency: IoConcurrency,
5082 460 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5083 460 : // Metadata keys image layer creation.
5084 460 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5085 460 : let begin = Instant::now();
5086 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5087 : // not contain too many keys, otherwise this takes a lot of memory.
5088 460 : let data = self
5089 460 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
5090 460 : .await?;
5091 460 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5092 460 : let mut new_data = BTreeMap::new();
5093 460 : let mut total_kb_retrieved = 0;
5094 460 : let mut total_keys_retrieved = 0;
5095 20484 : for (k, v) in data {
5096 20024 : let v = v?;
5097 20024 : total_kb_retrieved += KEY_SIZE + v.len();
5098 20024 : total_keys_retrieved += 1;
5099 20024 : new_data.insert(k, v);
5100 : }
5101 460 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5102 460 : };
5103 460 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5104 460 : let elapsed = begin.elapsed();
5105 460 :
5106 460 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5107 460 : info!(
5108 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",
5109 0 : elapsed.as_secs_f64()
5110 : );
5111 :
5112 460 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5113 4 : return Ok(ImageLayerCreationOutcome::Skip);
5114 456 : }
5115 456 : if self.cancel.is_cancelled() {
5116 0 : return Err(CreateImageLayersError::Cancelled);
5117 456 : }
5118 456 : let mut wrote_any_image = false;
5119 20480 : for (k, v) in data {
5120 20024 : if v.is_empty() {
5121 : // the key has been deleted, it does not need an image
5122 : // in metadata keyspace, an empty image == tombstone
5123 16 : continue;
5124 20008 : }
5125 20008 : wrote_any_image = true;
5126 20008 :
5127 20008 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5128 20008 :
5129 20008 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5130 20008 : // on the normal data path either.
5131 20008 : image_layer_writer.put_image(k, v, ctx).await?;
5132 : }
5133 :
5134 456 : if wrote_any_image {
5135 : // Normal path: we have written some data into the new image layer for this
5136 : // partition, so flush it to disk.
5137 24 : info!(
5138 0 : "created image layer for metadata {}",
5139 0 : ImageLayerName {
5140 0 : key_range: img_range.clone(),
5141 0 : lsn
5142 0 : }
5143 : );
5144 24 : Ok(ImageLayerCreationOutcome::Generated {
5145 24 : unfinished_image_layer: image_layer_writer,
5146 24 : })
5147 : } else {
5148 432 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5149 432 : Ok(ImageLayerCreationOutcome::Empty)
5150 : }
5151 460 : }
5152 :
5153 : /// Predicate function which indicates whether we should check if new image layers
5154 : /// are required. Since checking if new image layers are required is expensive in
5155 : /// terms of CPU, we only do it in the following cases:
5156 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
5157 : /// 2. If enough time has passed since the last check:
5158 : /// 1. For large tenants, we wish to perform the check more often since they
5159 : /// suffer from the lack of image layers
5160 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5161 1156 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5162 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
5163 :
5164 1156 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5165 1156 : let distance = lsn
5166 1156 : .checked_sub(last_checks_at)
5167 1156 : .expect("Attempt to compact with LSN going backwards");
5168 1156 : let min_distance =
5169 1156 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5170 1156 :
5171 1156 : let distance_based_decision = distance.0 >= min_distance;
5172 1156 :
5173 1156 : let mut time_based_decision = false;
5174 1156 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5175 1156 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
5176 952 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
5177 : {
5178 0 : self.get_checkpoint_timeout()
5179 : } else {
5180 952 : Duration::from_secs(3600 * 48)
5181 : };
5182 :
5183 952 : time_based_decision = match *last_check_instant {
5184 524 : Some(last_check) => {
5185 524 : let elapsed = last_check.elapsed();
5186 524 : elapsed >= check_required_after
5187 : }
5188 428 : None => true,
5189 : };
5190 204 : }
5191 :
5192 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5193 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5194 : // check.
5195 1156 : let decision = distance_based_decision || time_based_decision;
5196 :
5197 1156 : if decision {
5198 432 : self.last_image_layer_creation_check_at.store(lsn);
5199 432 : *last_check_instant = Some(Instant::now());
5200 724 : }
5201 :
5202 1156 : decision
5203 1156 : }
5204 :
5205 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5206 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5207 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
5208 : async fn create_image_layers(
5209 : self: &Arc<Timeline>,
5210 : partitioning: &KeyPartitioning,
5211 : lsn: Lsn,
5212 : mode: ImageLayerCreationMode,
5213 : ctx: &RequestContext,
5214 : last_status: LastImageLayerCreationStatus,
5215 : yield_for_l0: bool,
5216 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5217 : let timer = self.metrics.create_images_time_histo.start_timer();
5218 :
5219 : if partitioning.parts.is_empty() {
5220 : warn!("no partitions to create image layers for");
5221 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5222 : }
5223 :
5224 : // We need to avoid holes between generated image layers.
5225 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5226 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5227 : //
5228 : // How such hole between partitions can appear?
5229 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5230 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5231 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5232 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5233 : let mut start = Key::MIN;
5234 :
5235 : let check_for_image_layers =
5236 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5237 : info!(
5238 : "resuming image layer creation: last_status=incomplete, continue from {}",
5239 : last_key
5240 : );
5241 : true
5242 : } else {
5243 : self.should_check_if_image_layers_required(lsn)
5244 : };
5245 :
5246 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
5247 :
5248 : let mut all_generated = true;
5249 :
5250 : let mut partition_processed = 0;
5251 : let mut total_partitions = partitioning.parts.len();
5252 : let mut last_partition_processed = None;
5253 : let mut partition_parts = partitioning.parts.clone();
5254 :
5255 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5256 : // We need to skip the partitions that have already been processed.
5257 : let mut found = false;
5258 : for (i, partition) in partition_parts.iter().enumerate() {
5259 : if last_key <= partition.end().unwrap() {
5260 : // ```plain
5261 : // |------|--------|----------|------|
5262 : // ^last_key
5263 : // ^start from this partition
5264 : // ```
5265 : // Why `i+1` instead of `i`?
5266 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5267 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5268 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5269 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5270 : if i + 1 >= total_partitions {
5271 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5272 : // iteration of image layer creation should return a complete status.
5273 : break; // with found=false
5274 : }
5275 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5276 : total_partitions = partition_parts.len();
5277 : // Update the start key to the partition start.
5278 : start = partition_parts[0].start().unwrap();
5279 : found = true;
5280 : break;
5281 : }
5282 : }
5283 : if !found {
5284 : // Last key is within the last partition, or larger than all partitions.
5285 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5286 : }
5287 : }
5288 :
5289 : for partition in partition_parts.iter() {
5290 : if self.cancel.is_cancelled() {
5291 : return Err(CreateImageLayersError::Cancelled);
5292 : }
5293 : partition_processed += 1;
5294 : let img_range = start..partition.ranges.last().unwrap().end;
5295 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5296 : if compact_metadata {
5297 : for range in &partition.ranges {
5298 : assert!(
5299 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5300 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5301 : "metadata keys must be partitioned separately"
5302 : );
5303 : }
5304 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5305 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5306 : // might mess up with evictions.
5307 : start = img_range.end;
5308 : continue;
5309 : }
5310 : // For initial and force modes, we always generate image layers for metadata keys.
5311 : } else if let ImageLayerCreationMode::Try = mode {
5312 : // check_for_image_layers = false -> skip
5313 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5314 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
5315 : start = img_range.end;
5316 : continue;
5317 : }
5318 : }
5319 : if let ImageLayerCreationMode::Force = mode {
5320 : // When forced to create image layers, we might try and create them where they already
5321 : // exist. This mode is only used in tests/debug.
5322 : let layers = self.layers.read().await;
5323 : if layers.contains_key(&PersistentLayerKey {
5324 : key_range: img_range.clone(),
5325 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5326 : is_delta: false,
5327 : }) {
5328 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5329 : // in the future.
5330 : tracing::info!(
5331 : "Skipping image layer at {lsn} {}..{}, already exists",
5332 : img_range.start,
5333 : img_range.end
5334 : );
5335 : start = img_range.end;
5336 : continue;
5337 : }
5338 : }
5339 :
5340 : let image_layer_writer = ImageLayerWriter::new(
5341 : self.conf,
5342 : self.timeline_id,
5343 : self.tenant_shard_id,
5344 : &img_range,
5345 : lsn,
5346 : ctx,
5347 : )
5348 : .await?;
5349 :
5350 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5351 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5352 0 : "failpoint image-layer-writer-fail-before-finish"
5353 0 : )))
5354 0 : });
5355 :
5356 : let io_concurrency = IoConcurrency::spawn_from_conf(
5357 : self.conf,
5358 : self.gate
5359 : .enter()
5360 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5361 : );
5362 :
5363 : let outcome = if !compact_metadata {
5364 : self.create_image_layer_for_rel_blocks(
5365 : partition,
5366 : image_layer_writer,
5367 : lsn,
5368 : ctx,
5369 : img_range.clone(),
5370 : io_concurrency,
5371 : )
5372 : .await?
5373 : } else {
5374 : self.create_image_layer_for_metadata_keys(
5375 : partition,
5376 : image_layer_writer,
5377 : lsn,
5378 : ctx,
5379 : img_range.clone(),
5380 : mode,
5381 : io_concurrency,
5382 : )
5383 : .await?
5384 : };
5385 : match outcome {
5386 : ImageLayerCreationOutcome::Empty => {
5387 : // No data in this partition, so we don't need to create an image layer (for now).
5388 : // The next image layer should cover this key range, so we don't advance the `start`
5389 : // key.
5390 : }
5391 : ImageLayerCreationOutcome::Generated {
5392 : unfinished_image_layer,
5393 : } => {
5394 : batch_image_writer.add_unfinished_image_writer(
5395 : unfinished_image_layer,
5396 : img_range.clone(),
5397 : lsn,
5398 : );
5399 : // The next image layer should be generated right after this one.
5400 : start = img_range.end;
5401 : }
5402 : ImageLayerCreationOutcome::Skip => {
5403 : // We don't need to create an image layer for this partition.
5404 : // The next image layer should NOT cover this range, otherwise
5405 : // the keyspace becomes empty (reads don't go past image layers).
5406 : start = img_range.end;
5407 : }
5408 : }
5409 :
5410 : if let ImageLayerCreationMode::Try = mode {
5411 : // We have at least made some progress
5412 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5413 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5414 : // image layer generation taking too long time and blocking L0 compaction. So in this
5415 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5416 : // if there are too many of them.
5417 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5418 : * self.get_compaction_threshold();
5419 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
5420 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
5421 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
5422 : if image_preempt_threshold != 0 {
5423 : let should_yield = self
5424 : .l0_compaction_trigger
5425 : .notified()
5426 : .now_or_never()
5427 : .is_some();
5428 : if should_yield {
5429 : tracing::info!(
5430 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
5431 : partition.start().unwrap(),
5432 : partition.end().unwrap()
5433 : );
5434 : last_partition_processed = Some(partition.clone());
5435 : all_generated = false;
5436 : break;
5437 : }
5438 : }
5439 : }
5440 : }
5441 : }
5442 :
5443 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5444 :
5445 : let mut guard = self.layers.write().await;
5446 :
5447 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5448 : // now they are being scheduled outside of write lock; current way is inconsistent with
5449 : // compaction lock order.
5450 : guard
5451 : .open_mut()?
5452 : .track_new_image_layers(&image_layers, &self.metrics);
5453 : drop_wlock(guard);
5454 : let duration = timer.stop_and_record();
5455 :
5456 : // Creating image layers may have caused some previously visible layers to be covered
5457 : if !image_layers.is_empty() {
5458 : self.update_layer_visibility().await?;
5459 : }
5460 :
5461 : let total_layer_size = image_layers
5462 : .iter()
5463 504 : .map(|l| l.metadata().file_size)
5464 : .sum::<u64>();
5465 :
5466 : if !image_layers.is_empty() {
5467 : info!(
5468 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5469 : image_layers.len(),
5470 : total_layer_size,
5471 : duration.as_secs_f64(),
5472 : partition_processed,
5473 : total_partitions
5474 : );
5475 : }
5476 :
5477 : Ok((
5478 : image_layers,
5479 : if all_generated {
5480 : LastImageLayerCreationStatus::Complete
5481 : } else {
5482 : LastImageLayerCreationStatus::Incomplete {
5483 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5484 : last_partition_processed.end().unwrap_or(Key::MIN)
5485 : } else {
5486 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5487 : Key::MIN
5488 : },
5489 : }
5490 : },
5491 : ))
5492 : }
5493 :
5494 : /// Wait until the background initial logical size calculation is complete, or
5495 : /// this Timeline is shut down. Calling this function will cause the initial
5496 : /// logical size calculation to skip waiting for the background jobs barrier.
5497 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5498 0 : if !self.shard_identity.is_shard_zero() {
5499 : // We don't populate logical size on shard >0: skip waiting for it.
5500 0 : return;
5501 0 : }
5502 0 :
5503 0 : if self.remote_client.is_deleting() {
5504 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5505 0 : return;
5506 0 : }
5507 0 :
5508 0 : if self.current_logical_size.current_size().is_exact() {
5509 : // root timelines are initialized with exact count, but never start the background
5510 : // calculation
5511 0 : return;
5512 0 : }
5513 :
5514 0 : if let Some(await_bg_cancel) = self
5515 0 : .current_logical_size
5516 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5517 0 : .get()
5518 0 : {
5519 0 : await_bg_cancel.cancel();
5520 0 : } else {
5521 : // We should not wait if we were not able to explicitly instruct
5522 : // the logical size cancellation to skip the concurrency limit semaphore.
5523 : // TODO: this is an unexpected case. We should restructure so that it
5524 : // can't happen.
5525 0 : tracing::warn!(
5526 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5527 : );
5528 0 : debug_assert!(false);
5529 : }
5530 :
5531 0 : tokio::select!(
5532 0 : _ = self.current_logical_size.initialized.acquire() => {},
5533 0 : _ = self.cancel.cancelled() => {}
5534 : )
5535 0 : }
5536 :
5537 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5538 : /// Timelines layers up to the ancestor_lsn.
5539 : ///
5540 : /// Requires a timeline that:
5541 : /// - has an ancestor to detach from
5542 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5543 : /// a technical requirement
5544 : ///
5545 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5546 : /// polled again until completion.
5547 : ///
5548 : /// During the operation all timelines sharing the data with this timeline will be reparented
5549 : /// from our ancestor to be branches of this timeline.
5550 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5551 0 : self: &Arc<Timeline>,
5552 0 : tenant: &crate::tenant::Tenant,
5553 0 : options: detach_ancestor::Options,
5554 0 : behavior: DetachBehavior,
5555 0 : ctx: &RequestContext,
5556 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5557 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
5558 0 : }
5559 :
5560 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5561 : /// reparents any reparentable children of previous ancestor.
5562 : ///
5563 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5564 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5565 : /// successfully, tenant must be reloaded.
5566 : ///
5567 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5568 : /// resetting the tenant.
5569 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5570 0 : self: &Arc<Timeline>,
5571 0 : tenant: &crate::tenant::Tenant,
5572 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5573 0 : ancestor_timeline_id: TimelineId,
5574 0 : ancestor_lsn: Lsn,
5575 0 : behavior: DetachBehavior,
5576 0 : ctx: &RequestContext,
5577 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5578 0 : detach_ancestor::detach_and_reparent(
5579 0 : self,
5580 0 : tenant,
5581 0 : prepared,
5582 0 : ancestor_timeline_id,
5583 0 : ancestor_lsn,
5584 0 : behavior,
5585 0 : ctx,
5586 0 : )
5587 0 : .await
5588 0 : }
5589 :
5590 : /// Final step which unblocks the GC.
5591 : ///
5592 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5593 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5594 0 : self: &Arc<Timeline>,
5595 0 : tenant: &crate::tenant::Tenant,
5596 0 : attempt: detach_ancestor::Attempt,
5597 0 : ctx: &RequestContext,
5598 0 : ) -> Result<(), detach_ancestor::Error> {
5599 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5600 0 : }
5601 : }
5602 :
5603 : impl Drop for Timeline {
5604 20 : fn drop(&mut self) {
5605 20 : if let Some(ancestor) = &self.ancestor_timeline {
5606 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5607 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5608 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5609 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5610 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5611 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5612 8 : }
5613 0 : }
5614 12 : }
5615 20 : info!(
5616 0 : "Timeline {} for tenant {} is being dropped",
5617 : self.timeline_id, self.tenant_shard_id.tenant_id
5618 : );
5619 20 : }
5620 : }
5621 :
5622 : /// Top-level failure to compact.
5623 : #[derive(Debug, thiserror::Error)]
5624 : pub(crate) enum CompactionError {
5625 : #[error("The timeline or pageserver is shutting down")]
5626 : ShuttingDown,
5627 : /// Compaction tried to offload a timeline and failed
5628 : #[error("Failed to offload timeline: {0}")]
5629 : Offload(OffloadError),
5630 : /// Compaction cannot be done right now; page reconstruction and so on.
5631 : #[error("Failed to collect keyspace: {0}")]
5632 : CollectKeySpaceError(#[from] CollectKeySpaceError),
5633 : #[error(transparent)]
5634 : Other(anyhow::Error),
5635 : #[error("Compaction already running: {0}")]
5636 : AlreadyRunning(&'static str),
5637 : }
5638 :
5639 : impl CompactionError {
5640 : /// Errors that can be ignored, i.e., cancel and shutdown.
5641 0 : pub fn is_cancel(&self) -> bool {
5642 0 : matches!(
5643 0 : self,
5644 : Self::ShuttingDown
5645 : | Self::AlreadyRunning(_)
5646 : | Self::CollectKeySpaceError(CollectKeySpaceError::Cancelled)
5647 : | Self::CollectKeySpaceError(CollectKeySpaceError::PageRead(
5648 : PageReconstructError::Cancelled
5649 : ))
5650 : | Self::Offload(OffloadError::Cancelled)
5651 : )
5652 0 : }
5653 :
5654 : /// Critical errors that indicate data corruption.
5655 0 : pub fn is_critical(&self) -> bool {
5656 0 : matches!(
5657 0 : self,
5658 : Self::CollectKeySpaceError(
5659 : CollectKeySpaceError::Decode(_)
5660 : | CollectKeySpaceError::PageRead(
5661 : PageReconstructError::MissingKey(_) | PageReconstructError::WalRedo(_),
5662 : )
5663 : )
5664 : )
5665 0 : }
5666 : }
5667 :
5668 : impl From<OffloadError> for CompactionError {
5669 0 : fn from(e: OffloadError) -> Self {
5670 0 : match e {
5671 0 : OffloadError::Cancelled => Self::ShuttingDown,
5672 0 : _ => Self::Offload(e),
5673 : }
5674 0 : }
5675 : }
5676 :
5677 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5678 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5679 0 : match value {
5680 : super::upload_queue::NotInitialized::Uninitialized => {
5681 0 : CompactionError::Other(anyhow::anyhow!(value))
5682 : }
5683 : super::upload_queue::NotInitialized::ShuttingDown
5684 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5685 : }
5686 0 : }
5687 : }
5688 :
5689 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5690 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5691 0 : match e {
5692 : super::storage_layer::layer::DownloadError::TimelineShutdown
5693 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5694 0 : CompactionError::ShuttingDown
5695 : }
5696 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5697 : | super::storage_layer::layer::DownloadError::DownloadRequired
5698 : | super::storage_layer::layer::DownloadError::NotFile(_)
5699 : | super::storage_layer::layer::DownloadError::DownloadFailed
5700 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5701 0 : CompactionError::Other(anyhow::anyhow!(e))
5702 : }
5703 : #[cfg(test)]
5704 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5705 0 : CompactionError::Other(anyhow::anyhow!(e))
5706 : }
5707 : }
5708 0 : }
5709 : }
5710 :
5711 : impl From<layer_manager::Shutdown> for CompactionError {
5712 0 : fn from(_: layer_manager::Shutdown) -> Self {
5713 0 : CompactionError::ShuttingDown
5714 0 : }
5715 : }
5716 :
5717 : #[serde_as]
5718 392 : #[derive(serde::Serialize)]
5719 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5720 :
5721 : #[derive(Default)]
5722 : enum DurationRecorder {
5723 : #[default]
5724 : NotStarted,
5725 : Recorded(RecordedDuration, tokio::time::Instant),
5726 : }
5727 :
5728 : impl DurationRecorder {
5729 1008 : fn till_now(&self) -> DurationRecorder {
5730 1008 : match self {
5731 : DurationRecorder::NotStarted => {
5732 0 : panic!("must only call on recorded measurements")
5733 : }
5734 1008 : DurationRecorder::Recorded(_, ended) => {
5735 1008 : let now = tokio::time::Instant::now();
5736 1008 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5737 1008 : }
5738 1008 : }
5739 1008 : }
5740 392 : fn into_recorded(self) -> Option<RecordedDuration> {
5741 392 : match self {
5742 0 : DurationRecorder::NotStarted => None,
5743 392 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5744 : }
5745 392 : }
5746 : }
5747 :
5748 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5749 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5750 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5751 : /// keys specified in the `data` field.
5752 : #[cfg(test)]
5753 : #[derive(Clone)]
5754 : pub struct DeltaLayerTestDesc {
5755 : pub lsn_range: Range<Lsn>,
5756 : pub key_range: Range<Key>,
5757 : pub data: Vec<(Key, Lsn, Value)>,
5758 : }
5759 :
5760 : #[cfg(test)]
5761 : #[derive(Clone)]
5762 : pub struct InMemoryLayerTestDesc {
5763 : pub lsn_range: Range<Lsn>,
5764 : pub data: Vec<(Key, Lsn, Value)>,
5765 : pub is_open: bool,
5766 : }
5767 :
5768 : #[cfg(test)]
5769 : impl DeltaLayerTestDesc {
5770 8 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5771 8 : Self {
5772 8 : lsn_range,
5773 8 : key_range,
5774 8 : data,
5775 8 : }
5776 8 : }
5777 :
5778 180 : pub fn new_with_inferred_key_range(
5779 180 : lsn_range: Range<Lsn>,
5780 180 : data: Vec<(Key, Lsn, Value)>,
5781 180 : ) -> Self {
5782 464 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5783 464 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5784 180 : Self {
5785 180 : key_range: (*key_min)..(key_max.next()),
5786 180 : lsn_range,
5787 180 : data,
5788 180 : }
5789 180 : }
5790 :
5791 20 : pub(crate) fn layer_name(&self) -> LayerName {
5792 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5793 20 : key_range: self.key_range.clone(),
5794 20 : lsn_range: self.lsn_range.clone(),
5795 20 : })
5796 20 : }
5797 : }
5798 :
5799 : impl Timeline {
5800 56 : async fn finish_compact_batch(
5801 56 : self: &Arc<Self>,
5802 56 : new_deltas: &[ResidentLayer],
5803 56 : new_images: &[ResidentLayer],
5804 56 : layers_to_remove: &[Layer],
5805 56 : ) -> Result<(), CompactionError> {
5806 56 : let mut guard = tokio::select! {
5807 56 : guard = self.layers.write() => guard,
5808 56 : _ = self.cancel.cancelled() => {
5809 0 : return Err(CompactionError::ShuttingDown);
5810 : }
5811 : };
5812 :
5813 56 : let mut duplicated_layers = HashSet::new();
5814 56 :
5815 56 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5816 :
5817 672 : for l in new_deltas {
5818 616 : if guard.contains(l.as_ref()) {
5819 : // expected in tests
5820 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5821 :
5822 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5823 : // `guard` on self.layers. as of writing this, there are no error returns except
5824 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5825 : // because we have not implemented L0 => L0 compaction.
5826 0 : duplicated_layers.insert(l.layer_desc().key());
5827 616 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5828 0 : return Err(CompactionError::Other(anyhow::anyhow!(
5829 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
5830 0 : )));
5831 616 : } else {
5832 616 : insert_layers.push(l.clone());
5833 616 : }
5834 : }
5835 :
5836 : // only remove those inputs which were not outputs
5837 56 : let remove_layers: Vec<Layer> = layers_to_remove
5838 56 : .iter()
5839 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5840 56 : .cloned()
5841 56 : .collect();
5842 56 :
5843 56 : if !new_images.is_empty() {
5844 0 : guard
5845 0 : .open_mut()?
5846 0 : .track_new_image_layers(new_images, &self.metrics);
5847 56 : }
5848 :
5849 56 : guard
5850 56 : .open_mut()?
5851 56 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5852 56 :
5853 56 : self.remote_client
5854 56 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5855 :
5856 56 : drop_wlock(guard);
5857 56 :
5858 56 : Ok(())
5859 56 : }
5860 :
5861 0 : async fn rewrite_layers(
5862 0 : self: &Arc<Self>,
5863 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5864 0 : mut drop_layers: Vec<Layer>,
5865 0 : ) -> Result<(), CompactionError> {
5866 0 : let mut guard = self.layers.write().await;
5867 :
5868 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5869 : // to avoid double-removing, and avoid rewriting something that was removed.
5870 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5871 0 : drop_layers.retain(|l| guard.contains(l));
5872 0 :
5873 0 : guard
5874 0 : .open_mut()?
5875 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5876 0 :
5877 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5878 0 :
5879 0 : self.remote_client
5880 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5881 :
5882 0 : Ok(())
5883 0 : }
5884 :
5885 : /// Schedules the uploads of the given image layers
5886 728 : fn upload_new_image_layers(
5887 728 : self: &Arc<Self>,
5888 728 : new_images: impl IntoIterator<Item = ResidentLayer>,
5889 728 : ) -> Result<(), super::upload_queue::NotInitialized> {
5890 780 : for layer in new_images {
5891 52 : self.remote_client.schedule_layer_file_upload(layer)?;
5892 : }
5893 : // should any new image layer been created, not uploading index_part will
5894 : // result in a mismatch between remote_physical_size and layermap calculated
5895 : // size, which will fail some tests, but should not be an issue otherwise.
5896 728 : self.remote_client
5897 728 : .schedule_index_upload_for_file_changes()?;
5898 728 : Ok(())
5899 728 : }
5900 :
5901 0 : async fn find_gc_time_cutoff(
5902 0 : &self,
5903 0 : now: SystemTime,
5904 0 : pitr: Duration,
5905 0 : cancel: &CancellationToken,
5906 0 : ctx: &RequestContext,
5907 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5908 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5909 0 : if self.shard_identity.is_shard_zero() {
5910 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5911 0 : let time_range = if pitr == Duration::ZERO {
5912 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5913 : } else {
5914 0 : pitr
5915 : };
5916 :
5917 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5918 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5919 0 : let timestamp = to_pg_timestamp(time_cutoff);
5920 :
5921 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5922 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5923 0 : LsnForTimestamp::Future(lsn) => {
5924 0 : // The timestamp is in the future. That sounds impossible,
5925 0 : // but what it really means is that there hasn't been
5926 0 : // any commits since the cutoff timestamp.
5927 0 : //
5928 0 : // In this case we should use the LSN of the most recent commit,
5929 0 : // which is implicitly the last LSN in the log.
5930 0 : debug!("future({})", lsn);
5931 0 : Some(self.get_last_record_lsn())
5932 : }
5933 0 : LsnForTimestamp::Past(lsn) => {
5934 0 : debug!("past({})", lsn);
5935 0 : None
5936 : }
5937 0 : LsnForTimestamp::NoData(lsn) => {
5938 0 : debug!("nodata({})", lsn);
5939 0 : None
5940 : }
5941 : };
5942 0 : Ok(time_cutoff)
5943 : } else {
5944 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5945 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5946 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5947 : // space cutoff that we would also have respected ourselves.
5948 0 : match self
5949 0 : .remote_client
5950 0 : .download_foreign_index(ShardNumber(0), cancel)
5951 0 : .await
5952 : {
5953 0 : Ok((index_part, index_generation, _index_mtime)) => {
5954 0 : tracing::info!(
5955 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5956 0 : index_part.metadata.latest_gc_cutoff_lsn()
5957 : );
5958 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5959 : }
5960 : Err(DownloadError::NotFound) => {
5961 : // This is unexpected, because during timeline creations shard zero persists to remote
5962 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5963 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5964 : // state, then shard zero should _eventually_ write an index when it GCs.
5965 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5966 0 : Ok(None)
5967 : }
5968 0 : Err(e) => {
5969 0 : // TODO: this function should return a different error type than page reconstruct error
5970 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5971 : }
5972 : }
5973 :
5974 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5975 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5976 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5977 : // new location. This is legal in principle, but problematic in practice because it might result
5978 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5979 : // because they have GC'd past the branch point.
5980 : }
5981 0 : }
5982 :
5983 : /// Find the Lsns above which layer files need to be retained on
5984 : /// garbage collection.
5985 : ///
5986 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5987 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5988 : /// the space-based retention.
5989 : ///
5990 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5991 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5992 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5993 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5994 : /// in the response as the GC cutoff point for the timeline.
5995 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5996 : pub(super) async fn find_gc_cutoffs(
5997 : &self,
5998 : now: SystemTime,
5999 : space_cutoff: Lsn,
6000 : pitr: Duration,
6001 : cancel: &CancellationToken,
6002 : ctx: &RequestContext,
6003 : ) -> Result<GcCutoffs, PageReconstructError> {
6004 : let _timer = self
6005 : .metrics
6006 : .find_gc_cutoffs_histo
6007 : .start_timer()
6008 : .record_on_drop();
6009 :
6010 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6011 :
6012 : if cfg!(test) {
6013 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6014 : if pitr == Duration::ZERO {
6015 : return Ok(GcCutoffs {
6016 : time: self.get_last_record_lsn(),
6017 : space: space_cutoff,
6018 : });
6019 : }
6020 : }
6021 :
6022 : // Calculate a time-based limit on how much to retain:
6023 : // - if PITR interval is set, then this is our cutoff.
6024 : // - if PITR interval is not set, then we do a lookup
6025 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6026 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6027 :
6028 : Ok(match (pitr, time_cutoff) {
6029 : (Duration::ZERO, Some(time_cutoff)) => {
6030 : // PITR is not set. Retain the size-based limit, or the default time retention,
6031 : // whichever requires less data.
6032 : GcCutoffs {
6033 : time: self.get_last_record_lsn(),
6034 : space: std::cmp::max(time_cutoff, space_cutoff),
6035 : }
6036 : }
6037 : (Duration::ZERO, None) => {
6038 : // PITR is not set, and time lookup failed
6039 : GcCutoffs {
6040 : time: self.get_last_record_lsn(),
6041 : space: space_cutoff,
6042 : }
6043 : }
6044 : (_, None) => {
6045 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6046 : // cannot advance beyond what was already GC'd, and respect space-based retention
6047 : GcCutoffs {
6048 : time: *self.get_applied_gc_cutoff_lsn(),
6049 : space: space_cutoff,
6050 : }
6051 : }
6052 : (_, Some(time_cutoff)) => {
6053 : // PITR interval is set and we looked up timestamp successfully. Ignore
6054 : // size based retention and make time cutoff authoritative
6055 : GcCutoffs {
6056 : time: time_cutoff,
6057 : space: time_cutoff,
6058 : }
6059 : }
6060 : })
6061 : }
6062 :
6063 : /// Garbage collect layer files on a timeline that are no longer needed.
6064 : ///
6065 : /// Currently, we don't make any attempt at removing unneeded page versions
6066 : /// within a layer file. We can only remove the whole file if it's fully
6067 : /// obsolete.
6068 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6069 : // this is most likely the background tasks, but it might be the spawned task from
6070 : // immediate_gc
6071 8 : let _g = tokio::select! {
6072 8 : guard = self.gc_lock.lock() => guard,
6073 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6074 : };
6075 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
6076 8 :
6077 8 : fail_point!("before-timeline-gc");
6078 8 :
6079 8 : // Is the timeline being deleted?
6080 8 : if self.is_stopping() {
6081 0 : return Err(GcError::TimelineCancelled);
6082 8 : }
6083 8 :
6084 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6085 8 : let gc_info = self.gc_info.read().unwrap();
6086 8 :
6087 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6088 8 : let time_cutoff = gc_info.cutoffs.time;
6089 8 : let retain_lsns = gc_info
6090 8 : .retain_lsns
6091 8 : .iter()
6092 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6093 8 : .collect();
6094 8 :
6095 8 : // Gets the maximum LSN that holds the valid lease.
6096 8 : //
6097 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6098 8 : // Here, we do not check for stale leases again.
6099 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6100 8 :
6101 8 : (
6102 8 : space_cutoff,
6103 8 : time_cutoff,
6104 8 : retain_lsns,
6105 8 : max_lsn_with_valid_lease,
6106 8 : )
6107 8 : };
6108 8 :
6109 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
6110 8 : let standby_horizon = self.standby_horizon.load();
6111 8 : // Hold GC for the standby, but as a safety guard do it only within some
6112 8 : // reasonable lag.
6113 8 : if standby_horizon != Lsn::INVALID {
6114 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6115 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6116 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6117 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6118 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6119 : } else {
6120 0 : warn!(
6121 0 : "standby is lagging for more than {}MB, not holding gc for it",
6122 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6123 : )
6124 : }
6125 0 : }
6126 8 : }
6127 :
6128 : // Reset standby horizon to ignore it if it is not updated till next GC.
6129 : // It is an easy way to unset it when standby disappears without adding
6130 : // more conf options.
6131 8 : self.standby_horizon.store(Lsn::INVALID);
6132 8 : self.metrics
6133 8 : .standby_horizon_gauge
6134 8 : .set(Lsn::INVALID.0 as i64);
6135 :
6136 8 : let res = self
6137 8 : .gc_timeline(
6138 8 : space_cutoff,
6139 8 : time_cutoff,
6140 8 : retain_lsns,
6141 8 : max_lsn_with_valid_lease,
6142 8 : new_gc_cutoff,
6143 8 : )
6144 8 : .instrument(
6145 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6146 : )
6147 8 : .await?;
6148 :
6149 : // only record successes
6150 8 : timer.stop_and_record();
6151 8 :
6152 8 : Ok(res)
6153 8 : }
6154 :
6155 8 : async fn gc_timeline(
6156 8 : &self,
6157 8 : space_cutoff: Lsn,
6158 8 : time_cutoff: Lsn,
6159 8 : retain_lsns: Vec<Lsn>,
6160 8 : max_lsn_with_valid_lease: Option<Lsn>,
6161 8 : new_gc_cutoff: Lsn,
6162 8 : ) -> Result<GcResult, GcError> {
6163 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6164 8 :
6165 8 : let now = SystemTime::now();
6166 8 : let mut result: GcResult = GcResult::default();
6167 8 :
6168 8 : // Nothing to GC. Return early.
6169 8 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6170 8 : if latest_gc_cutoff >= new_gc_cutoff {
6171 0 : info!(
6172 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6173 : );
6174 0 : return Ok(result);
6175 8 : }
6176 :
6177 : // We need to ensure that no one tries to read page versions or create
6178 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6179 : // for details. This will block until the old value is no longer in use.
6180 : //
6181 : // The GC cutoff should only ever move forwards.
6182 8 : let waitlist = {
6183 8 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6184 8 : if *write_guard > new_gc_cutoff {
6185 0 : return Err(GcError::BadLsn {
6186 0 : why: format!(
6187 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6188 0 : *write_guard, new_gc_cutoff
6189 0 : ),
6190 0 : });
6191 8 : }
6192 8 :
6193 8 : write_guard.store_and_unlock(new_gc_cutoff)
6194 8 : };
6195 8 : waitlist.wait().await;
6196 :
6197 8 : info!("GC starting");
6198 :
6199 8 : debug!("retain_lsns: {:?}", retain_lsns);
6200 :
6201 8 : let mut layers_to_remove = Vec::new();
6202 :
6203 : // Scan all layers in the timeline (remote or on-disk).
6204 : //
6205 : // Garbage collect the layer if all conditions are satisfied:
6206 : // 1. it is older than cutoff LSN;
6207 : // 2. it is older than PITR interval;
6208 : // 3. it doesn't need to be retained for 'retain_lsns';
6209 : // 4. it does not need to be kept for LSNs holding valid leases.
6210 : // 5. newer on-disk image layers cover the layer's whole key range
6211 : //
6212 : // TODO holding a write lock is too agressive and avoidable
6213 8 : let mut guard = self.layers.write().await;
6214 8 : let layers = guard.layer_map()?;
6215 48 : 'outer: for l in layers.iter_historic_layers() {
6216 48 : result.layers_total += 1;
6217 48 :
6218 48 : // 1. Is it newer than GC horizon cutoff point?
6219 48 : if l.get_lsn_range().end > space_cutoff {
6220 4 : info!(
6221 0 : "keeping {} because it's newer than space_cutoff {}",
6222 0 : l.layer_name(),
6223 : space_cutoff,
6224 : );
6225 4 : result.layers_needed_by_cutoff += 1;
6226 4 : continue 'outer;
6227 44 : }
6228 44 :
6229 44 : // 2. It is newer than PiTR cutoff point?
6230 44 : if l.get_lsn_range().end > time_cutoff {
6231 0 : info!(
6232 0 : "keeping {} because it's newer than time_cutoff {}",
6233 0 : l.layer_name(),
6234 : time_cutoff,
6235 : );
6236 0 : result.layers_needed_by_pitr += 1;
6237 0 : continue 'outer;
6238 44 : }
6239 :
6240 : // 3. Is it needed by a child branch?
6241 : // NOTE With that we would keep data that
6242 : // might be referenced by child branches forever.
6243 : // We can track this in child timeline GC and delete parent layers when
6244 : // they are no longer needed. This might be complicated with long inheritance chains.
6245 : //
6246 : // TODO Vec is not a great choice for `retain_lsns`
6247 44 : for retain_lsn in &retain_lsns {
6248 : // start_lsn is inclusive
6249 0 : if &l.get_lsn_range().start <= retain_lsn {
6250 0 : info!(
6251 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6252 0 : l.layer_name(),
6253 0 : retain_lsn,
6254 0 : l.is_incremental(),
6255 : );
6256 0 : result.layers_needed_by_branches += 1;
6257 0 : continue 'outer;
6258 0 : }
6259 : }
6260 :
6261 : // 4. Is there a valid lease that requires us to keep this layer?
6262 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
6263 : // keep if layer start <= any of the lease
6264 36 : if &l.get_lsn_range().start <= lsn {
6265 28 : info!(
6266 0 : "keeping {} because there is a valid lease preventing GC at {}",
6267 0 : l.layer_name(),
6268 : lsn,
6269 : );
6270 28 : result.layers_needed_by_leases += 1;
6271 28 : continue 'outer;
6272 8 : }
6273 8 : }
6274 :
6275 : // 5. Is there a later on-disk layer for this relation?
6276 : //
6277 : // The end-LSN is exclusive, while disk_consistent_lsn is
6278 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6279 : // OK for a delta layer to have end LSN 101, but if the end LSN
6280 : // is 102, then it might not have been fully flushed to disk
6281 : // before crash.
6282 : //
6283 : // For example, imagine that the following layers exist:
6284 : //
6285 : // 1000 - image (A)
6286 : // 1000-2000 - delta (B)
6287 : // 2000 - image (C)
6288 : // 2000-3000 - delta (D)
6289 : // 3000 - image (E)
6290 : //
6291 : // If GC horizon is at 2500, we can remove layers A and B, but
6292 : // we cannot remove C, even though it's older than 2500, because
6293 : // the delta layer 2000-3000 depends on it.
6294 16 : if !layers
6295 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6296 : {
6297 12 : info!("keeping {} because it is the latest layer", l.layer_name());
6298 12 : result.layers_not_updated += 1;
6299 12 : continue 'outer;
6300 4 : }
6301 4 :
6302 4 : // We didn't find any reason to keep this file, so remove it.
6303 4 : info!(
6304 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6305 0 : l.layer_name(),
6306 0 : l.is_incremental(),
6307 : );
6308 4 : layers_to_remove.push(l);
6309 : }
6310 :
6311 8 : if !layers_to_remove.is_empty() {
6312 : // Persist the new GC cutoff value before we actually remove anything.
6313 : // This unconditionally schedules also an index_part.json update, even though, we will
6314 : // be doing one a bit later with the unlinked gc'd layers.
6315 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6316 4 : self.schedule_uploads(disk_consistent_lsn, None)
6317 4 : .map_err(|e| {
6318 0 : if self.cancel.is_cancelled() {
6319 0 : GcError::TimelineCancelled
6320 : } else {
6321 0 : GcError::Remote(e)
6322 : }
6323 4 : })?;
6324 :
6325 4 : let gc_layers = layers_to_remove
6326 4 : .iter()
6327 4 : .map(|x| guard.get_from_desc(x))
6328 4 : .collect::<Vec<Layer>>();
6329 4 :
6330 4 : result.layers_removed = gc_layers.len() as u64;
6331 4 :
6332 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
6333 :
6334 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
6335 4 :
6336 4 : #[cfg(feature = "testing")]
6337 4 : {
6338 4 : result.doomed_layers = gc_layers;
6339 4 : }
6340 4 : }
6341 :
6342 8 : info!(
6343 0 : "GC completed removing {} layers, cutoff {}",
6344 : result.layers_removed, new_gc_cutoff
6345 : );
6346 :
6347 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
6348 8 : Ok(result)
6349 8 : }
6350 :
6351 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6352 1340354 : async fn reconstruct_value(
6353 1340354 : &self,
6354 1340354 : key: Key,
6355 1340354 : request_lsn: Lsn,
6356 1340354 : mut data: ValueReconstructState,
6357 1340354 : redo_attempt_type: RedoAttemptType,
6358 1340354 : ) -> Result<Bytes, PageReconstructError> {
6359 1340354 : // Perform WAL redo if needed
6360 1340354 : data.records.reverse();
6361 :
6362 1340354 : let fire_critical_error = match redo_attempt_type {
6363 1336206 : RedoAttemptType::ReadPage => true,
6364 0 : RedoAttemptType::LegacyCompaction => true,
6365 4148 : RedoAttemptType::GcCompaction => false,
6366 : };
6367 :
6368 : // If we have a page image, and no WAL, we're all set
6369 1340354 : if data.records.is_empty() {
6370 1338666 : if let Some((img_lsn, img)) = &data.img {
6371 1338666 : trace!(
6372 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6373 : key, img_lsn, request_lsn,
6374 : );
6375 1338666 : Ok(img.clone())
6376 : } else {
6377 0 : Err(PageReconstructError::from(anyhow!(
6378 0 : "base image for {key} at {request_lsn} not found"
6379 0 : )))
6380 : }
6381 : } else {
6382 : // We need to do WAL redo.
6383 : //
6384 : // If we don't have a base image, then the oldest WAL record better initialize
6385 : // the page
6386 1688 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6387 0 : Err(PageReconstructError::from(anyhow!(
6388 0 : "Base image for {} at {} not found, but got {} WAL records",
6389 0 : key,
6390 0 : request_lsn,
6391 0 : data.records.len()
6392 0 : )))
6393 : } else {
6394 1688 : if data.img.is_some() {
6395 1544 : trace!(
6396 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6397 0 : data.records.len(),
6398 : key,
6399 : request_lsn
6400 : );
6401 : } else {
6402 144 : trace!(
6403 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
6404 0 : data.records.len(),
6405 : key,
6406 : request_lsn
6407 : );
6408 : };
6409 1688 : let res = self
6410 1688 : .walredo_mgr
6411 1688 : .as_ref()
6412 1688 : .context("timeline has no walredo manager")
6413 1688 : .map_err(PageReconstructError::WalRedo)?
6414 1688 : .request_redo(
6415 1688 : key,
6416 1688 : request_lsn,
6417 1688 : data.img,
6418 1688 : data.records,
6419 1688 : self.pg_version,
6420 1688 : redo_attempt_type,
6421 1688 : )
6422 1688 : .await;
6423 1684 : let img = match res {
6424 1684 : Ok(img) => img,
6425 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6426 4 : Err(walredo::Error::Other(err)) => {
6427 4 : if fire_critical_error {
6428 0 : critical!("walredo failure during page reconstruction: {err:?}");
6429 4 : }
6430 4 : return Err(PageReconstructError::WalRedo(
6431 4 : err.context("reconstruct a page image"),
6432 4 : ));
6433 : }
6434 : };
6435 1684 : Ok(img)
6436 : }
6437 : }
6438 1340354 : }
6439 :
6440 0 : pub(crate) async fn spawn_download_all_remote_layers(
6441 0 : self: Arc<Self>,
6442 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6443 0 : ctx: &RequestContext,
6444 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6445 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6446 :
6447 : // this is not really needed anymore; it has tests which really check the return value from
6448 : // http api. it would be better not to maintain this anymore.
6449 :
6450 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6451 0 : if let Some(st) = &*status_guard {
6452 0 : match &st.state {
6453 : DownloadRemoteLayersTaskState::Running => {
6454 0 : return Err(st.clone());
6455 : }
6456 : DownloadRemoteLayersTaskState::ShutDown
6457 0 : | DownloadRemoteLayersTaskState::Completed => {
6458 0 : *status_guard = None;
6459 0 : }
6460 : }
6461 0 : }
6462 :
6463 0 : let self_clone = Arc::clone(&self);
6464 0 : let task_ctx = ctx.detached_child(
6465 0 : TaskKind::DownloadAllRemoteLayers,
6466 0 : DownloadBehavior::Download,
6467 0 : );
6468 0 : let task_id = task_mgr::spawn(
6469 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6470 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6471 0 : self.tenant_shard_id,
6472 0 : Some(self.timeline_id),
6473 0 : "download all remote layers task",
6474 0 : async move {
6475 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
6476 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6477 0 : match &mut *status_guard {
6478 : None => {
6479 0 : warn!("tasks status is supposed to be Some(), since we are running");
6480 : }
6481 0 : Some(st) => {
6482 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6483 0 : if st.task_id != exp_task_id {
6484 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6485 0 : } else {
6486 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6487 0 : }
6488 : }
6489 : };
6490 0 : Ok(())
6491 0 : }
6492 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))
6493 : );
6494 :
6495 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6496 0 : task_id: format!("{task_id}"),
6497 0 : state: DownloadRemoteLayersTaskState::Running,
6498 0 : total_layer_count: 0,
6499 0 : successful_download_count: 0,
6500 0 : failed_download_count: 0,
6501 0 : };
6502 0 : *status_guard = Some(initial_info.clone());
6503 0 :
6504 0 : Ok(initial_info)
6505 0 : }
6506 :
6507 0 : async fn download_all_remote_layers(
6508 0 : self: &Arc<Self>,
6509 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6510 0 : ctx: &RequestContext,
6511 0 : ) {
6512 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6513 :
6514 0 : let remaining = {
6515 0 : let guard = self.layers.read().await;
6516 0 : let Ok(lm) = guard.layer_map() else {
6517 : // technically here we could look into iterating accessible layers, but downloading
6518 : // all layers of a shutdown timeline makes no sense regardless.
6519 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6520 0 : return;
6521 : };
6522 0 : lm.iter_historic_layers()
6523 0 : .map(|desc| guard.get_from_desc(&desc))
6524 0 : .collect::<Vec<_>>()
6525 0 : };
6526 0 : let total_layer_count = remaining.len();
6527 :
6528 : macro_rules! lock_status {
6529 : ($st:ident) => {
6530 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6531 : let st = st
6532 : .as_mut()
6533 : .expect("this function is only called after the task has been spawned");
6534 : assert_eq!(
6535 : st.task_id,
6536 : format!(
6537 : "{}",
6538 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6539 : )
6540 : );
6541 : let $st = st;
6542 : };
6543 : }
6544 :
6545 : {
6546 0 : lock_status!(st);
6547 0 : st.total_layer_count = total_layer_count as u64;
6548 0 : }
6549 0 :
6550 0 : let mut remaining = remaining.into_iter();
6551 0 : let mut have_remaining = true;
6552 0 : let mut js = tokio::task::JoinSet::new();
6553 0 :
6554 0 : let cancel = task_mgr::shutdown_token();
6555 0 :
6556 0 : let limit = request.max_concurrent_downloads;
6557 :
6558 : loop {
6559 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6560 0 : let Some(next) = remaining.next() else {
6561 0 : have_remaining = false;
6562 0 : break;
6563 : };
6564 :
6565 0 : let span = tracing::info_span!("download", layer = %next);
6566 :
6567 0 : let ctx = ctx.attached_child();
6568 0 : js.spawn(
6569 0 : async move {
6570 0 : let res = next.download(&ctx).await;
6571 0 : (next, res)
6572 0 : }
6573 0 : .instrument(span),
6574 0 : );
6575 0 : }
6576 :
6577 0 : while let Some(res) = js.join_next().await {
6578 0 : match res {
6579 : Ok((_, Ok(_))) => {
6580 0 : lock_status!(st);
6581 0 : st.successful_download_count += 1;
6582 : }
6583 0 : Ok((layer, Err(e))) => {
6584 0 : tracing::error!(%layer, "download failed: {e:#}");
6585 0 : lock_status!(st);
6586 0 : st.failed_download_count += 1;
6587 : }
6588 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6589 0 : Err(je) if je.is_panic() => {
6590 0 : lock_status!(st);
6591 0 : st.failed_download_count += 1;
6592 : }
6593 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6594 : }
6595 : }
6596 :
6597 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6598 0 : break;
6599 0 : }
6600 : }
6601 :
6602 : {
6603 0 : lock_status!(st);
6604 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6605 : }
6606 0 : }
6607 :
6608 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6609 0 : &self,
6610 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6611 0 : self.download_all_remote_layers_task_info
6612 0 : .read()
6613 0 : .unwrap()
6614 0 : .clone()
6615 0 : }
6616 : }
6617 :
6618 : impl Timeline {
6619 : /// Returns non-remote layers for eviction.
6620 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6621 0 : let guard = self.layers.read().await;
6622 0 : let mut max_layer_size: Option<u64> = None;
6623 0 :
6624 0 : let resident_layers = guard
6625 0 : .likely_resident_layers()
6626 0 : .map(|layer| {
6627 0 : let file_size = layer.layer_desc().file_size;
6628 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6629 0 :
6630 0 : let last_activity_ts = layer.latest_activity();
6631 0 :
6632 0 : EvictionCandidate {
6633 0 : layer: layer.to_owned().into(),
6634 0 : last_activity_ts,
6635 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6636 0 : visibility: layer.visibility(),
6637 0 : }
6638 0 : })
6639 0 : .collect();
6640 0 :
6641 0 : DiskUsageEvictionInfo {
6642 0 : max_layer_size,
6643 0 : resident_layers,
6644 0 : }
6645 0 : }
6646 :
6647 3764 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6648 3764 : ShardIndex {
6649 3764 : shard_number: self.tenant_shard_id.shard_number,
6650 3764 : shard_count: self.tenant_shard_id.shard_count,
6651 3764 : }
6652 3764 : }
6653 :
6654 : /// Persistently blocks gc for `Manual` reason.
6655 : ///
6656 : /// Returns true if no such block existed before, false otherwise.
6657 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
6658 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6659 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6660 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6661 0 : }
6662 :
6663 : /// Persistently unblocks gc for `Manual` reason.
6664 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
6665 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6666 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6667 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6668 0 : }
6669 :
6670 : #[cfg(test)]
6671 120 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6672 120 : self.last_record_lsn.advance(new_lsn);
6673 120 : }
6674 :
6675 : #[cfg(test)]
6676 4 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6677 4 : self.disk_consistent_lsn.store(new_value);
6678 4 : }
6679 :
6680 : /// Force create an image layer and place it into the layer map.
6681 : ///
6682 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6683 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6684 : /// placed into the layer map in one run AND be validated.
6685 : #[cfg(test)]
6686 132 : pub(super) async fn force_create_image_layer(
6687 132 : self: &Arc<Timeline>,
6688 132 : lsn: Lsn,
6689 132 : mut images: Vec<(Key, Bytes)>,
6690 132 : check_start_lsn: Option<Lsn>,
6691 132 : ctx: &RequestContext,
6692 132 : ) -> anyhow::Result<()> {
6693 132 : let last_record_lsn = self.get_last_record_lsn();
6694 132 : assert!(
6695 132 : lsn <= last_record_lsn,
6696 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6697 : );
6698 132 : if let Some(check_start_lsn) = check_start_lsn {
6699 132 : assert!(lsn >= check_start_lsn);
6700 0 : }
6701 416 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6702 132 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6703 132 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6704 132 : let mut image_layer_writer = ImageLayerWriter::new(
6705 132 : self.conf,
6706 132 : self.timeline_id,
6707 132 : self.tenant_shard_id,
6708 132 : &(min_key..end_key),
6709 132 : lsn,
6710 132 : ctx,
6711 132 : )
6712 132 : .await?;
6713 680 : for (key, img) in images {
6714 548 : image_layer_writer.put_image(key, img, ctx).await?;
6715 : }
6716 132 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6717 132 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6718 132 : info!("force created image layer {}", image_layer.local_path());
6719 : {
6720 132 : let mut guard = self.layers.write().await;
6721 132 : guard
6722 132 : .open_mut()
6723 132 : .unwrap()
6724 132 : .force_insert_layer(image_layer.clone());
6725 132 : }
6726 132 :
6727 132 : // Update remote_timeline_client state to reflect existence of this layer
6728 132 : self.remote_client
6729 132 : .schedule_layer_file_upload(image_layer)
6730 132 : .unwrap();
6731 132 :
6732 132 : Ok(())
6733 132 : }
6734 :
6735 : /// Force create a delta layer and place it into the layer map.
6736 : ///
6737 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6738 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6739 : /// placed into the layer map in one run AND be validated.
6740 : #[cfg(test)]
6741 188 : pub(super) async fn force_create_delta_layer(
6742 188 : self: &Arc<Timeline>,
6743 188 : mut deltas: DeltaLayerTestDesc,
6744 188 : check_start_lsn: Option<Lsn>,
6745 188 : ctx: &RequestContext,
6746 188 : ) -> anyhow::Result<()> {
6747 188 : let last_record_lsn = self.get_last_record_lsn();
6748 188 : deltas
6749 188 : .data
6750 284 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6751 188 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6752 188 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6753 660 : for (_, lsn, _) in &deltas.data {
6754 472 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6755 : }
6756 188 : assert!(
6757 188 : deltas.lsn_range.end <= last_record_lsn,
6758 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6759 : deltas.lsn_range.end,
6760 : last_record_lsn
6761 : );
6762 188 : if let Some(check_start_lsn) = check_start_lsn {
6763 188 : assert!(deltas.lsn_range.start >= check_start_lsn);
6764 0 : }
6765 188 : let mut delta_layer_writer = DeltaLayerWriter::new(
6766 188 : self.conf,
6767 188 : self.timeline_id,
6768 188 : self.tenant_shard_id,
6769 188 : deltas.key_range.start,
6770 188 : deltas.lsn_range,
6771 188 : ctx,
6772 188 : )
6773 188 : .await?;
6774 660 : for (key, lsn, val) in deltas.data {
6775 472 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6776 : }
6777 188 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6778 188 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6779 188 : info!("force created delta layer {}", delta_layer.local_path());
6780 : {
6781 188 : let mut guard = self.layers.write().await;
6782 188 : guard
6783 188 : .open_mut()
6784 188 : .unwrap()
6785 188 : .force_insert_layer(delta_layer.clone());
6786 188 : }
6787 188 :
6788 188 : // Update remote_timeline_client state to reflect existence of this layer
6789 188 : self.remote_client
6790 188 : .schedule_layer_file_upload(delta_layer)
6791 188 : .unwrap();
6792 188 :
6793 188 : Ok(())
6794 188 : }
6795 :
6796 : /// Force create an in-memory layer and place them into the layer map.
6797 : #[cfg(test)]
6798 8 : pub(super) async fn force_create_in_memory_layer(
6799 8 : self: &Arc<Timeline>,
6800 8 : mut in_memory: InMemoryLayerTestDesc,
6801 8 : check_start_lsn: Option<Lsn>,
6802 8 : ctx: &RequestContext,
6803 8 : ) -> anyhow::Result<()> {
6804 : use utils::bin_ser::BeSer;
6805 :
6806 : // Validate LSNs
6807 8 : if let Some(check_start_lsn) = check_start_lsn {
6808 8 : assert!(in_memory.lsn_range.start >= check_start_lsn);
6809 0 : }
6810 :
6811 8 : let last_record_lsn = self.get_last_record_lsn();
6812 8 : let layer_end_lsn = if in_memory.is_open {
6813 4 : in_memory
6814 4 : .data
6815 4 : .iter()
6816 40 : .map(|(_key, lsn, _value)| lsn)
6817 4 : .max()
6818 4 : .cloned()
6819 : } else {
6820 4 : Some(in_memory.lsn_range.end)
6821 : };
6822 :
6823 8 : if let Some(end) = layer_end_lsn {
6824 8 : assert!(
6825 8 : end <= last_record_lsn,
6826 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6827 : end,
6828 : last_record_lsn,
6829 : );
6830 0 : }
6831 :
6832 80 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
6833 80 : assert!(*lsn >= in_memory.lsn_range.start);
6834 80 : assert!(*lsn < in_memory.lsn_range.end);
6835 80 : });
6836 8 :
6837 8 : // Build the batch
6838 8 : in_memory
6839 8 : .data
6840 72 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6841 8 :
6842 8 : let data = in_memory
6843 8 : .data
6844 8 : .into_iter()
6845 80 : .map(|(key, lsn, value)| {
6846 80 : let value_size = value.serialized_size().unwrap() as usize;
6847 80 : (key.to_compact(), lsn, value_size, value)
6848 80 : })
6849 8 : .collect::<Vec<_>>();
6850 8 :
6851 8 : let batch = SerializedValueBatch::from_values(data);
6852 :
6853 : // Create the in-memory layer and write the batch into it
6854 8 : let layer = InMemoryLayer::create(
6855 8 : self.conf,
6856 8 : self.timeline_id,
6857 8 : self.tenant_shard_id,
6858 8 : in_memory.lsn_range.start,
6859 8 : &self.gate,
6860 8 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
6861 8 : &CancellationToken::new(),
6862 8 : ctx,
6863 8 : )
6864 8 : .await
6865 8 : .unwrap();
6866 8 :
6867 8 : layer.put_batch(batch, ctx).await.unwrap();
6868 8 : if !in_memory.is_open {
6869 4 : layer.freeze(in_memory.lsn_range.end).await;
6870 4 : }
6871 :
6872 8 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
6873 :
6874 : // Link the layer to the layer map
6875 : {
6876 8 : let mut guard = self.layers.write().await;
6877 8 : let layer_map = guard.open_mut().unwrap();
6878 8 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
6879 8 : }
6880 8 :
6881 8 : Ok(())
6882 8 : }
6883 :
6884 : /// Return all keys at the LSN in the image layers
6885 : #[cfg(test)]
6886 12 : pub(crate) async fn inspect_image_layers(
6887 12 : self: &Arc<Timeline>,
6888 12 : lsn: Lsn,
6889 12 : ctx: &RequestContext,
6890 12 : io_concurrency: IoConcurrency,
6891 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
6892 12 : let mut all_data = Vec::new();
6893 12 : let guard = self.layers.read().await;
6894 68 : for layer in guard.layer_map()?.iter_historic_layers() {
6895 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6896 16 : let layer = guard.get_from_desc(&layer);
6897 16 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6898 16 : layer
6899 16 : .get_values_reconstruct_data(
6900 16 : KeySpace::single(Key::MIN..Key::MAX),
6901 16 : lsn..Lsn(lsn.0 + 1),
6902 16 : &mut reconstruct_data,
6903 16 : ctx,
6904 16 : )
6905 16 : .await?;
6906 132 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6907 132 : let v = v.collect_pending_ios().await?;
6908 132 : all_data.push((k, v.img.unwrap().1));
6909 : }
6910 52 : }
6911 : }
6912 12 : all_data.sort();
6913 12 : Ok(all_data)
6914 12 : }
6915 :
6916 : /// Get all historic layer descriptors in the layer map
6917 : #[cfg(test)]
6918 48 : pub(crate) async fn inspect_historic_layers(
6919 48 : self: &Arc<Timeline>,
6920 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6921 48 : let mut layers = Vec::new();
6922 48 : let guard = self.layers.read().await;
6923 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6924 228 : layers.push(layer.key());
6925 228 : }
6926 48 : Ok(layers)
6927 48 : }
6928 :
6929 : #[cfg(test)]
6930 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6931 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6932 20 : keyspace.merge(&ks);
6933 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6934 20 : }
6935 : }
6936 :
6937 : /// Tracking writes ingestion does to a particular in-memory layer.
6938 : ///
6939 : /// Cleared upon freezing a layer.
6940 : pub(crate) struct TimelineWriterState {
6941 : open_layer: Arc<InMemoryLayer>,
6942 : current_size: u64,
6943 : // Previous Lsn which passed through
6944 : prev_lsn: Option<Lsn>,
6945 : // Largest Lsn which passed through the current writer
6946 : max_lsn: Option<Lsn>,
6947 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6948 : cached_last_freeze_at: Lsn,
6949 : }
6950 :
6951 : impl TimelineWriterState {
6952 2612 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6953 2612 : Self {
6954 2612 : open_layer,
6955 2612 : current_size,
6956 2612 : prev_lsn: None,
6957 2612 : max_lsn: None,
6958 2612 : cached_last_freeze_at: last_freeze_at,
6959 2612 : }
6960 2612 : }
6961 : }
6962 :
6963 : /// Various functions to mutate the timeline.
6964 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6965 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6966 : // but will cause large code changes.
6967 : pub(crate) struct TimelineWriter<'a> {
6968 : tl: &'a Timeline,
6969 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6970 : }
6971 :
6972 : impl Deref for TimelineWriter<'_> {
6973 : type Target = Timeline;
6974 :
6975 19796612 : fn deref(&self) -> &Self::Target {
6976 19796612 : self.tl
6977 19796612 : }
6978 : }
6979 :
6980 : #[derive(PartialEq)]
6981 : enum OpenLayerAction {
6982 : Roll,
6983 : Open,
6984 : None,
6985 : }
6986 :
6987 : impl TimelineWriter<'_> {
6988 9608484 : async fn handle_open_layer_action(
6989 9608484 : &mut self,
6990 9608484 : at: Lsn,
6991 9608484 : action: OpenLayerAction,
6992 9608484 : ctx: &RequestContext,
6993 9608484 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6994 9608484 : match action {
6995 : OpenLayerAction::Roll => {
6996 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6997 160 : self.roll_layer(freeze_at).await?;
6998 160 : self.open_layer(at, ctx).await?;
6999 : }
7000 2452 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7001 : OpenLayerAction::None => {
7002 9605872 : assert!(self.write_guard.is_some());
7003 : }
7004 : }
7005 :
7006 9608484 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7007 9608484 : }
7008 :
7009 2612 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7010 2612 : let layer = self
7011 2612 : .tl
7012 2612 : .get_layer_for_write(at, &self.write_guard, ctx)
7013 2612 : .await?;
7014 2612 : let initial_size = layer.size().await?;
7015 :
7016 2612 : let last_freeze_at = self.last_freeze_at.load();
7017 2612 : self.write_guard.replace(TimelineWriterState::new(
7018 2612 : layer,
7019 2612 : initial_size,
7020 2612 : last_freeze_at,
7021 2612 : ));
7022 2612 :
7023 2612 : Ok(())
7024 2612 : }
7025 :
7026 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7027 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7028 :
7029 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7030 : // to propagate the backpressure up into WAL ingestion.
7031 160 : let l0_count = self
7032 160 : .tl
7033 160 : .layers
7034 160 : .read()
7035 160 : .await
7036 160 : .layer_map()?
7037 160 : .level0_deltas()
7038 160 : .len();
7039 160 : let wait_thresholds = [
7040 160 : self.get_l0_flush_delay_threshold(),
7041 160 : self.get_l0_flush_stall_threshold(),
7042 160 : ];
7043 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7044 :
7045 : // self.write_guard will be taken by the freezing
7046 160 : let flush_id = self
7047 160 : .tl
7048 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7049 160 : .await?;
7050 :
7051 160 : assert!(self.write_guard.is_none());
7052 :
7053 160 : if let Some(wait_threshold) = wait_threshold {
7054 0 : if l0_count >= wait_threshold {
7055 0 : debug!(
7056 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7057 : );
7058 0 : self.tl.wait_flush_completion(flush_id).await?;
7059 0 : }
7060 160 : }
7061 :
7062 160 : if current_size >= self.get_checkpoint_distance() * 2 {
7063 0 : warn!("Flushed oversized open layer with size {}", current_size)
7064 160 : }
7065 :
7066 160 : Ok(())
7067 160 : }
7068 :
7069 9608484 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7070 9608484 : let state = &*self.write_guard;
7071 9608484 : let Some(state) = &state else {
7072 2452 : return OpenLayerAction::Open;
7073 : };
7074 :
7075 : #[cfg(feature = "testing")]
7076 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7077 : // this check and assertion are not really needed because
7078 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7079 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7080 : // is no TimelineWriterState.
7081 0 : assert!(
7082 0 : state.open_layer.end_lsn.get().is_some(),
7083 0 : "our open_layer must be outdated"
7084 : );
7085 :
7086 : // this would be a memory leak waiting to happen because the in-memory layer always has
7087 : // an index
7088 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7089 9606032 : }
7090 9606032 :
7091 9606032 : if state.prev_lsn == Some(lsn) {
7092 : // Rolling mid LSN is not supported by [downstream code].
7093 : // Hence, only roll at LSN boundaries.
7094 : //
7095 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7096 12 : return OpenLayerAction::None;
7097 9606020 : }
7098 9606020 :
7099 9606020 : if state.current_size == 0 {
7100 : // Don't roll empty layers
7101 0 : return OpenLayerAction::None;
7102 9606020 : }
7103 9606020 :
7104 9606020 : if self.tl.should_roll(
7105 9606020 : state.current_size,
7106 9606020 : state.current_size + new_value_size,
7107 9606020 : self.get_checkpoint_distance(),
7108 9606020 : lsn,
7109 9606020 : state.cached_last_freeze_at,
7110 9606020 : state.open_layer.get_opened_at(),
7111 9606020 : ) {
7112 160 : OpenLayerAction::Roll
7113 : } else {
7114 9605860 : OpenLayerAction::None
7115 : }
7116 9608484 : }
7117 :
7118 : /// Put a batch of keys at the specified Lsns.
7119 9608480 : pub(crate) async fn put_batch(
7120 9608480 : &mut self,
7121 9608480 : batch: SerializedValueBatch,
7122 9608480 : ctx: &RequestContext,
7123 9608480 : ) -> anyhow::Result<()> {
7124 9608480 : if !batch.has_data() {
7125 0 : return Ok(());
7126 9608480 : }
7127 9608480 :
7128 9608480 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7129 9608480 : // We don't assert this in release builds, since key ownership policies may change over
7130 9608480 : // time. Stray keys will be removed during compaction.
7131 9608480 : if cfg!(debug_assertions) {
7132 19789952 : for metadata in &batch.metadata {
7133 10181472 : if let ValueMeta::Serialized(metadata) = metadata {
7134 10181472 : let key = Key::from_compact(metadata.key);
7135 10181472 : assert!(
7136 10181472 : self.shard_identity.is_key_local(&key)
7137 0 : || self.shard_identity.is_key_global(&key),
7138 0 : "key {key} does not belong on shard {}",
7139 0 : self.shard_identity.shard_index()
7140 : );
7141 0 : }
7142 : }
7143 0 : }
7144 :
7145 9608480 : let batch_max_lsn = batch.max_lsn;
7146 9608480 : let buf_size: u64 = batch.buffer_size() as u64;
7147 9608480 :
7148 9608480 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7149 9608480 : let layer = self
7150 9608480 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7151 9608480 : .await?;
7152 :
7153 9608480 : let res = layer.put_batch(batch, ctx).await;
7154 :
7155 9608480 : if res.is_ok() {
7156 9608480 : // Update the current size only when the entire write was ok.
7157 9608480 : // In case of failures, we may have had partial writes which
7158 9608480 : // render the size tracking out of sync. That's ok because
7159 9608480 : // the checkpoint distance should be significantly smaller
7160 9608480 : // than the S3 single shot upload limit of 5GiB.
7161 9608480 : let state = self.write_guard.as_mut().unwrap();
7162 9608480 :
7163 9608480 : state.current_size += buf_size;
7164 9608480 : state.prev_lsn = Some(batch_max_lsn);
7165 9608480 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7166 9608480 : }
7167 :
7168 9608480 : res
7169 9608480 : }
7170 :
7171 : #[cfg(test)]
7172 : /// Test helper, for tests that would like to poke individual values without composing a batch
7173 8780308 : pub(crate) async fn put(
7174 8780308 : &mut self,
7175 8780308 : key: Key,
7176 8780308 : lsn: Lsn,
7177 8780308 : value: &Value,
7178 8780308 : ctx: &RequestContext,
7179 8780308 : ) -> anyhow::Result<()> {
7180 : use utils::bin_ser::BeSer;
7181 8780308 : if !key.is_valid_key_on_write_path() {
7182 0 : bail!(
7183 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7184 0 : key
7185 0 : );
7186 8780308 : }
7187 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
7188 8780308 : let batch = SerializedValueBatch::from_values(vec![(
7189 8780308 : key.to_compact(),
7190 8780308 : lsn,
7191 8780308 : val_ser_size,
7192 8780308 : value.clone(),
7193 8780308 : )]);
7194 8780308 :
7195 8780308 : self.put_batch(batch, ctx).await
7196 8780308 : }
7197 :
7198 4 : pub(crate) async fn delete_batch(
7199 4 : &mut self,
7200 4 : batch: &[(Range<Key>, Lsn)],
7201 4 : ctx: &RequestContext,
7202 4 : ) -> anyhow::Result<()> {
7203 4 : if let Some((_, lsn)) = batch.first() {
7204 4 : let action = self.get_open_layer_action(*lsn, 0);
7205 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7206 4 : layer.put_tombstones(batch).await?;
7207 0 : }
7208 :
7209 4 : Ok(())
7210 4 : }
7211 :
7212 : /// Track the end of the latest digested WAL record.
7213 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7214 : ///
7215 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7216 : ///
7217 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7218 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7219 : /// the latest and can be read.
7220 10558204 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7221 10558204 : self.tl.finish_write(new_lsn);
7222 10558204 : }
7223 :
7224 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7225 541140 : self.tl.update_current_logical_size(delta)
7226 541140 : }
7227 : }
7228 :
7229 : // We need TimelineWriter to be send in upcoming conversion of
7230 : // Timeline::layers to tokio::sync::RwLock.
7231 : #[test]
7232 4 : fn is_send() {
7233 4 : fn _assert_send<T: Send>() {}
7234 4 : _assert_send::<TimelineWriter<'_>>();
7235 4 : }
7236 :
7237 : #[cfg(test)]
7238 : mod tests {
7239 : use std::sync::Arc;
7240 :
7241 : use pageserver_api::key::Key;
7242 : use pageserver_api::value::Value;
7243 : use std::iter::Iterator;
7244 : use tracing::Instrument;
7245 : use utils::id::TimelineId;
7246 : use utils::lsn::Lsn;
7247 :
7248 : use super::HeatMapTimeline;
7249 : use crate::context::RequestContextBuilder;
7250 : use crate::tenant::harness::{TenantHarness, test_img};
7251 : use crate::tenant::layer_map::LayerMap;
7252 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7253 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7254 : use crate::tenant::{PreviousHeatmap, Timeline};
7255 :
7256 20 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7257 20 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7258 20 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7259 100 : for (l, r) in lhs_rhs {
7260 80 : assert_eq!(l.name, r.name);
7261 80 : assert_eq!(l.metadata, r.metadata);
7262 : }
7263 20 : }
7264 :
7265 : #[tokio::test]
7266 4 : async fn test_heatmap_generation() {
7267 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7268 4 :
7269 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7270 4 : Lsn(0x10)..Lsn(0x20),
7271 4 : vec![(
7272 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7273 4 : Lsn(0x11),
7274 4 : Value::Image(test_img("foo")),
7275 4 : )],
7276 4 : );
7277 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7278 4 : Lsn(0x10)..Lsn(0x20),
7279 4 : vec![(
7280 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7281 4 : Lsn(0x11),
7282 4 : Value::Image(test_img("foo")),
7283 4 : )],
7284 4 : );
7285 4 : let l0_delta = DeltaLayerTestDesc::new(
7286 4 : Lsn(0x20)..Lsn(0x30),
7287 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7288 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7289 4 : vec![(
7290 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7291 4 : Lsn(0x25),
7292 4 : Value::Image(test_img("foo")),
7293 4 : )],
7294 4 : );
7295 4 : let delta_layers = vec![
7296 4 : covered_delta.clone(),
7297 4 : visible_delta.clone(),
7298 4 : l0_delta.clone(),
7299 4 : ];
7300 4 :
7301 4 : let image_layer = (
7302 4 : Lsn(0x40),
7303 4 : vec![(
7304 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7305 4 : test_img("bar"),
7306 4 : )],
7307 4 : );
7308 4 : let image_layers = vec![image_layer];
7309 4 :
7310 4 : let (tenant, ctx) = harness.load().await;
7311 4 : let timeline = tenant
7312 4 : .create_test_timeline_with_layers(
7313 4 : TimelineId::generate(),
7314 4 : Lsn(0x10),
7315 4 : 14,
7316 4 : &ctx,
7317 4 : Vec::new(), // in-memory layers
7318 4 : delta_layers,
7319 4 : image_layers,
7320 4 : Lsn(0x100),
7321 4 : )
7322 4 : .await
7323 4 : .unwrap();
7324 4 : let ctx = &ctx.with_scope_timeline(&timeline);
7325 4 :
7326 4 : // Layer visibility is an input to heatmap generation, so refresh it first
7327 4 : timeline.update_layer_visibility().await.unwrap();
7328 4 :
7329 4 : let heatmap = timeline
7330 4 : .generate_heatmap()
7331 4 : .await
7332 4 : .expect("Infallible while timeline is not shut down");
7333 4 :
7334 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
7335 4 :
7336 4 : // L0 should come last
7337 4 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
7338 4 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
7339 4 :
7340 4 : let mut last_lsn = Lsn::MAX;
7341 20 : for layer in heatmap_layers {
7342 4 : // Covered layer should be omitted
7343 16 : assert!(layer.name != covered_delta.layer_name());
7344 4 :
7345 16 : let layer_lsn = match &layer.name {
7346 8 : LayerName::Delta(d) => d.lsn_range.end,
7347 8 : LayerName::Image(i) => i.lsn,
7348 4 : };
7349 4 :
7350 4 : // Apart from L0s, newest Layers should come first
7351 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
7352 12 : assert!(layer_lsn <= last_lsn);
7353 12 : last_lsn = layer_lsn;
7354 4 : }
7355 4 : }
7356 4 :
7357 4 : // Evict all the layers and stash the old heatmap in the timeline.
7358 4 : // This simulates a migration to a cold secondary location.
7359 4 :
7360 4 : let guard = timeline.layers.read().await;
7361 4 : let mut all_layers = Vec::new();
7362 4 : let forever = std::time::Duration::from_secs(120);
7363 20 : for layer in guard.likely_resident_layers() {
7364 20 : all_layers.push(layer.clone());
7365 20 : layer.evict_and_wait(forever).await.unwrap();
7366 4 : }
7367 4 : drop(guard);
7368 4 :
7369 4 : timeline
7370 4 : .previous_heatmap
7371 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
7372 4 : heatmap: heatmap.clone(),
7373 4 : read_at: std::time::Instant::now(),
7374 4 : end_lsn: None,
7375 4 : })));
7376 4 :
7377 4 : // Generate a new heatmap and assert that it contains the same layers as the old one.
7378 4 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
7379 4 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
7380 4 :
7381 4 : // Download each layer one by one. Generate the heatmap at each step and check
7382 4 : // that it's stable.
7383 24 : for layer in all_layers {
7384 20 : if layer.visibility() == LayerVisibilityHint::Covered {
7385 4 : continue;
7386 16 : }
7387 16 :
7388 16 : eprintln!("Downloading {layer} and re-generating heatmap");
7389 16 :
7390 16 : let ctx = &RequestContextBuilder::from(ctx)
7391 16 : .download_behavior(crate::context::DownloadBehavior::Download)
7392 16 : .attached_child();
7393 4 :
7394 16 : let _resident = layer
7395 16 : .download_and_keep_resident(ctx)
7396 16 : .instrument(tracing::info_span!(
7397 16 : parent: None,
7398 4 : "download_layer",
7399 4 : tenant_id = %timeline.tenant_shard_id.tenant_id,
7400 0 : shard_id = %timeline.tenant_shard_id.shard_slug(),
7401 0 : timeline_id = %timeline.timeline_id
7402 4 : ))
7403 16 : .await
7404 16 : .unwrap();
7405 4 :
7406 16 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
7407 16 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
7408 4 : }
7409 4 :
7410 4 : // Everything from the post-migration heatmap is now resident.
7411 4 : // Check that we drop it from memory.
7412 4 : assert!(matches!(
7413 4 : timeline.previous_heatmap.load().as_deref(),
7414 4 : Some(PreviousHeatmap::Obsolete)
7415 4 : ));
7416 4 : }
7417 :
7418 : #[tokio::test]
7419 4 : async fn test_previous_heatmap_obsoletion() {
7420 4 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
7421 4 : .await
7422 4 : .unwrap();
7423 4 :
7424 4 : let l0_delta = DeltaLayerTestDesc::new(
7425 4 : Lsn(0x20)..Lsn(0x30),
7426 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7427 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7428 4 : vec![(
7429 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7430 4 : Lsn(0x25),
7431 4 : Value::Image(test_img("foo")),
7432 4 : )],
7433 4 : );
7434 4 :
7435 4 : let image_layer = (
7436 4 : Lsn(0x40),
7437 4 : vec![(
7438 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7439 4 : test_img("bar"),
7440 4 : )],
7441 4 : );
7442 4 :
7443 4 : let delta_layers = vec![l0_delta];
7444 4 : let image_layers = vec![image_layer];
7445 4 :
7446 4 : let (tenant, ctx) = harness.load().await;
7447 4 : let timeline = tenant
7448 4 : .create_test_timeline_with_layers(
7449 4 : TimelineId::generate(),
7450 4 : Lsn(0x10),
7451 4 : 14,
7452 4 : &ctx,
7453 4 : Vec::new(), // in-memory layers
7454 4 : delta_layers,
7455 4 : image_layers,
7456 4 : Lsn(0x100),
7457 4 : )
7458 4 : .await
7459 4 : .unwrap();
7460 4 :
7461 4 : // Layer visibility is an input to heatmap generation, so refresh it first
7462 4 : timeline.update_layer_visibility().await.unwrap();
7463 4 :
7464 4 : let heatmap = timeline
7465 4 : .generate_heatmap()
7466 4 : .await
7467 4 : .expect("Infallible while timeline is not shut down");
7468 4 :
7469 4 : // Both layers should be in the heatmap
7470 4 : assert!(heatmap.all_layers().count() > 0);
7471 4 :
7472 4 : // Now simulate a migration.
7473 4 : timeline
7474 4 : .previous_heatmap
7475 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
7476 4 : heatmap: heatmap.clone(),
7477 4 : read_at: std::time::Instant::now(),
7478 4 : end_lsn: None,
7479 4 : })));
7480 4 :
7481 4 : // Evict all the layers in the previous heatmap
7482 4 : let guard = timeline.layers.read().await;
7483 4 : let forever = std::time::Duration::from_secs(120);
7484 12 : for layer in guard.likely_resident_layers() {
7485 12 : layer.evict_and_wait(forever).await.unwrap();
7486 4 : }
7487 4 : drop(guard);
7488 4 :
7489 4 : // Generate a new heatmap and check that the previous heatmap
7490 4 : // has been marked obsolete.
7491 4 : let post_eviction_heatmap = timeline
7492 4 : .generate_heatmap()
7493 4 : .await
7494 4 : .expect("Infallible while timeline is not shut down");
7495 4 :
7496 4 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
7497 4 : assert!(matches!(
7498 4 : timeline.previous_heatmap.load().as_deref(),
7499 4 : Some(PreviousHeatmap::Obsolete)
7500 4 : ));
7501 4 : }
7502 :
7503 : #[tokio::test]
7504 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
7505 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
7506 4 : .await
7507 4 : .unwrap();
7508 4 :
7509 4 : let (tenant, ctx) = harness.load().await;
7510 4 : let timeline = tenant
7511 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
7512 4 : .await
7513 4 : .unwrap();
7514 4 :
7515 4 : let layer = find_some_layer(&timeline).await;
7516 4 : let layer = layer
7517 4 : .keep_resident()
7518 4 : .await
7519 4 : .expect("no download => no downloading errors")
7520 4 : .drop_eviction_guard();
7521 4 :
7522 4 : let forever = std::time::Duration::from_secs(120);
7523 4 :
7524 4 : let first = layer.evict_and_wait(forever);
7525 4 : let second = layer.evict_and_wait(forever);
7526 4 :
7527 4 : let (first, second) = tokio::join!(first, second);
7528 4 :
7529 4 : let res = layer.keep_resident().await;
7530 4 : assert!(res.is_none(), "{res:?}");
7531 4 :
7532 4 : match (first, second) {
7533 4 : (Ok(()), Ok(())) => {
7534 4 : // because there are no more timeline locks being taken on eviction path, we can
7535 4 : // witness all three outcomes here.
7536 4 : }
7537 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7538 0 : // if one completes before the other, this is fine just as well.
7539 0 : }
7540 4 : other => unreachable!("unexpected {:?}", other),
7541 4 : }
7542 4 : }
7543 :
7544 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7545 4 : let layers = timeline.layers.read().await;
7546 4 : let desc = layers
7547 4 : .layer_map()
7548 4 : .unwrap()
7549 4 : .iter_historic_layers()
7550 4 : .next()
7551 4 : .expect("must find one layer to evict");
7552 4 :
7553 4 : layers.get_from_desc(&desc)
7554 4 : }
7555 : }
|
