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 anyhow::{Context, Result, anyhow, bail, ensure};
27 : use arc_swap::{ArcSwap, ArcSwapOption};
28 : use bytes::Bytes;
29 : use camino::Utf8Path;
30 : use chrono::{DateTime, Utc};
31 : use compaction::{CompactionOutcome, GcCompactionCombinedSettings};
32 : use enumset::EnumSet;
33 : use fail::fail_point;
34 : use futures::stream::FuturesUnordered;
35 : use futures::{FutureExt, StreamExt};
36 : use handle::ShardTimelineId;
37 : use layer_manager::Shutdown;
38 : use offload::OffloadError;
39 : use once_cell::sync::Lazy;
40 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
41 : use pageserver_api::key::{
42 : KEY_SIZE, Key, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
43 : SPARSE_RANGE,
44 : };
45 : use pageserver_api::keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning};
46 : use pageserver_api::models::{
47 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
48 : DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest, EvictionPolicy,
49 : InMemoryLayerInfo, LayerMapInfo, LsnLease, PageTraceEvent, RelSizeMigration, TimelineState,
50 : };
51 : use pageserver_api::reltag::{BlockNumber, RelTag};
52 : use pageserver_api::shard::{ShardIdentity, ShardIndex, ShardNumber, TenantShardId};
53 : #[cfg(test)]
54 : use pageserver_api::value::Value;
55 : use postgres_connection::PgConnectionConfig;
56 : use postgres_ffi::v14::xlog_utils;
57 : use postgres_ffi::{WAL_SEGMENT_SIZE, to_pg_timestamp};
58 : use rand::Rng;
59 : use remote_storage::DownloadError;
60 : use serde_with::serde_as;
61 : use storage_broker::BrokerClientChannel;
62 : use tokio::runtime::Handle;
63 : use tokio::sync::mpsc::Sender;
64 : use tokio::sync::{Notify, oneshot, watch};
65 : use tokio_util::sync::CancellationToken;
66 : use tracing::*;
67 : use utils::generation::Generation;
68 : use utils::guard_arc_swap::GuardArcSwap;
69 : use utils::id::TimelineId;
70 : use utils::lsn::{AtomicLsn, Lsn, RecordLsn};
71 : use utils::postgres_client::PostgresClientProtocol;
72 : use utils::rate_limit::RateLimit;
73 : use utils::seqwait::SeqWait;
74 : use utils::simple_rcu::{Rcu, RcuReadGuard};
75 : use utils::sync::gate::{Gate, GateGuard};
76 : use utils::{completion, critical, fs_ext, pausable_failpoint};
77 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
78 :
79 : use self::delete::DeleteTimelineFlow;
80 : pub(super) use self::eviction_task::EvictionTaskTenantState;
81 : use self::eviction_task::EvictionTaskTimelineState;
82 : use self::layer_manager::LayerManager;
83 : use self::logical_size::LogicalSize;
84 : use self::walreceiver::{WalReceiver, WalReceiverConf};
85 : use super::config::TenantConf;
86 : use super::remote_timeline_client::index::{GcCompactionState, IndexPart};
87 : use super::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
88 : use super::secondary::heatmap::HeatMapLayer;
89 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
90 : use super::upload_queue::NotInitialized;
91 : use super::{
92 : AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
93 : debug_assert_current_span_has_tenant_and_timeline_id,
94 : };
95 : use crate::aux_file::AuxFileSizeEstimator;
96 : use crate::config::PageServerConf;
97 : use crate::context::{DownloadBehavior, RequestContext};
98 : use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
99 : use crate::keyspace::{KeyPartitioning, KeySpace};
100 : use crate::l0_flush::{self, L0FlushGlobalState};
101 : use crate::metrics::{
102 : DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_AMORTIZED_GLOBAL, LAYERS_PER_READ_BATCH_GLOBAL,
103 : LAYERS_PER_READ_GLOBAL, ScanLatencyOngoingRecording, TimelineMetrics,
104 : };
105 : use crate::page_service::TenantManagerTypes;
106 : use crate::pgdatadir_mapping::{
107 : CalculateLogicalSizeError, CollectKeySpaceError, DirectoryKind, LsnForTimestamp,
108 : MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
109 : };
110 : use crate::task_mgr::TaskKind;
111 : use crate::tenant::config::{AttachmentMode, TenantConfOpt};
112 : use crate::tenant::gc_result::GcResult;
113 : use crate::tenant::layer_map::{LayerMap, SearchResult};
114 : use crate::tenant::metadata::TimelineMetadata;
115 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
116 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
117 : use crate::tenant::storage_layer::{
118 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
119 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
120 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
121 : ValueReconstructState, ValuesReconstructState,
122 : };
123 : use crate::tenant::tasks::BackgroundLoopKind;
124 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
125 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
126 : use crate::walingest::WalLagCooldown;
127 : use crate::{ZERO_PAGE, task_mgr, walredo};
128 :
129 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
130 : pub(crate) enum FlushLoopState {
131 : NotStarted,
132 : Running {
133 : #[cfg(test)]
134 : expect_initdb_optimization: bool,
135 : #[cfg(test)]
136 : initdb_optimization_count: usize,
137 : },
138 : Exited,
139 : }
140 :
141 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
142 : pub enum ImageLayerCreationMode {
143 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
144 : Try,
145 : /// Force creating the image layers if possible. For now, no image layers will be created
146 : /// for metadata keys. Used in compaction code path with force flag enabled.
147 : Force,
148 : /// Initial ingestion of the data, and no data should be dropped in this function. This
149 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
150 : /// code path.
151 : Initial,
152 : }
153 :
154 : #[derive(Clone, Debug, Default)]
155 : pub enum LastImageLayerCreationStatus {
156 : Incomplete {
157 : /// The last key of the partition (exclusive) that was processed in the last
158 : /// image layer creation attempt. We will continue from this key in the next
159 : /// attempt.
160 : last_key: Key,
161 : },
162 : Complete,
163 : #[default]
164 : Initial,
165 : }
166 :
167 : impl std::fmt::Display for ImageLayerCreationMode {
168 1148 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
169 1148 : write!(f, "{:?}", self)
170 1148 : }
171 : }
172 :
173 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
174 : /// Can be removed after all refactors are done.
175 56 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
176 56 : drop(rlock)
177 56 : }
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 1204 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
182 1204 : drop(rlock)
183 1204 : }
184 :
185 : /// The outward-facing resources required to build a Timeline
186 : pub struct TimelineResources {
187 : pub remote_client: RemoteTimelineClient,
188 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
189 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
190 : pub l0_compaction_trigger: Arc<Notify>,
191 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
192 : }
193 :
194 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
195 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
196 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
197 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
198 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
199 : pub(crate) struct RelSizeCache {
200 : pub(crate) complete_as_of: Lsn,
201 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
202 : }
203 :
204 : pub struct Timeline {
205 : pub(crate) conf: &'static PageServerConf,
206 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
207 :
208 : myself: Weak<Self>,
209 :
210 : pub(crate) tenant_shard_id: TenantShardId,
211 : pub timeline_id: TimelineId,
212 :
213 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
214 : /// Never changes for the lifetime of this [`Timeline`] object.
215 : ///
216 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
217 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
218 : pub(crate) generation: Generation,
219 :
220 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
221 : /// to shards, and is constant through the lifetime of this Timeline.
222 : shard_identity: ShardIdentity,
223 :
224 : pub pg_version: u32,
225 :
226 : /// The tuple has two elements.
227 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
228 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
229 : ///
230 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
231 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
232 : ///
233 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
234 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
235 : /// `PersistentLayerDesc`'s.
236 : ///
237 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
238 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
239 : /// runtime, e.g., during page reconstruction.
240 : ///
241 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
242 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
243 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
244 :
245 : last_freeze_at: AtomicLsn,
246 : // Atomic would be more appropriate here.
247 : last_freeze_ts: RwLock<Instant>,
248 :
249 : pub(crate) standby_horizon: AtomicLsn,
250 :
251 : // WAL redo manager. `None` only for broken tenants.
252 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
253 :
254 : /// Remote storage client.
255 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
256 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
257 :
258 : // What page versions do we hold in the repository? If we get a
259 : // request > last_record_lsn, we need to wait until we receive all
260 : // the WAL up to the request. The SeqWait provides functions for
261 : // that. TODO: If we get a request for an old LSN, such that the
262 : // versions have already been garbage collected away, we should
263 : // throw an error, but we don't track that currently.
264 : //
265 : // last_record_lsn.load().last points to the end of last processed WAL record.
266 : //
267 : // We also remember the starting point of the previous record in
268 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
269 : // first WAL record when the node is started up. But here, we just
270 : // keep track of it.
271 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
272 :
273 : // All WAL records have been processed and stored durably on files on
274 : // local disk, up to this LSN. On crash and restart, we need to re-process
275 : // the WAL starting from this point.
276 : //
277 : // Some later WAL records might have been processed and also flushed to disk
278 : // already, so don't be surprised to see some, but there's no guarantee on
279 : // them yet.
280 : disk_consistent_lsn: AtomicLsn,
281 :
282 : // Parent timeline that this timeline was branched from, and the LSN
283 : // of the branch point.
284 : ancestor_timeline: Option<Arc<Timeline>>,
285 : ancestor_lsn: Lsn,
286 :
287 : // The LSN of gc-compaction that was last applied to this timeline.
288 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
289 :
290 : pub(crate) metrics: Arc<TimelineMetrics>,
291 :
292 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
293 : // in `crate::page_service` writes these metrics.
294 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
295 :
296 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
297 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
298 :
299 : /// Ensures layers aren't frozen by checkpointer between
300 : /// [`Timeline::get_layer_for_write`] and layer reads.
301 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
302 : /// Must always be acquired before the layer map/individual layer lock
303 : /// to avoid deadlock.
304 : ///
305 : /// The state is cleared upon freezing.
306 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
307 :
308 : /// Used to avoid multiple `flush_loop` tasks running
309 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
310 :
311 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
312 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
313 : /// The flush cycle counter is sent back on the layer_flush_done channel when
314 : /// the flush finishes. You can use that to wait for the flush to finish.
315 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
316 : /// read by whoever sends an update
317 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
318 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
319 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
320 :
321 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
322 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
323 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
324 : // the planned GC cutoff.
325 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
326 :
327 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
328 :
329 : // List of child timelines and their branch points. This is needed to avoid
330 : // garbage collecting data that is still needed by the child timelines.
331 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
332 :
333 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
334 :
335 : // It may change across major versions so for simplicity
336 : // keep it after running initdb for a timeline.
337 : // It is needed in checks when we want to error on some operations
338 : // when they are requested for pre-initdb lsn.
339 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
340 : // though let's keep them both for better error visibility.
341 : pub initdb_lsn: Lsn,
342 :
343 : /// The repartitioning result. Allows a single writer and multiple readers.
344 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
345 :
346 : /// Configuration: how often should the partitioning be recalculated.
347 : repartition_threshold: u64,
348 :
349 : last_image_layer_creation_check_at: AtomicLsn,
350 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
351 :
352 : /// Current logical size of the "datadir", at the last LSN.
353 : current_logical_size: LogicalSize,
354 :
355 : /// Information about the last processed message by the WAL receiver,
356 : /// or None if WAL receiver has not received anything for this timeline
357 : /// yet.
358 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
359 : pub walreceiver: Mutex<Option<WalReceiver>>,
360 :
361 : /// Relation size cache
362 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
363 :
364 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
365 :
366 : state: watch::Sender<TimelineState>,
367 :
368 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
369 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
370 : pub delete_progress: TimelineDeleteProgress,
371 :
372 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
373 :
374 : /// Load or creation time information about the disk_consistent_lsn and when the loading
375 : /// happened. Used for consumption metrics.
376 : pub(crate) loaded_at: (Lsn, SystemTime),
377 :
378 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
379 : pub(crate) gate: Gate,
380 :
381 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
382 : /// to the timeline should drop out when this token fires.
383 : pub(crate) cancel: CancellationToken,
384 :
385 : /// Make sure we only have one running compaction at a time in tests.
386 : ///
387 : /// Must only be taken in two places:
388 : /// - [`Timeline::compact`] (this file)
389 : /// - [`delete::delete_local_timeline_directory`]
390 : ///
391 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
392 : compaction_lock: tokio::sync::Mutex<()>,
393 :
394 : /// If true, the last compaction failed.
395 : compaction_failed: AtomicBool,
396 :
397 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
398 : l0_compaction_trigger: Arc<Notify>,
399 :
400 : /// Make sure we only have one running gc at a time.
401 : ///
402 : /// Must only be taken in two places:
403 : /// - [`Timeline::gc`] (this file)
404 : /// - [`delete::delete_local_timeline_directory`]
405 : ///
406 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
407 : gc_lock: tokio::sync::Mutex<()>,
408 :
409 : /// Cloned from [`super::Tenant::pagestream_throttle`] on construction.
410 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
411 :
412 : /// Size estimator for aux file v2
413 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
414 :
415 : /// Some test cases directly place keys into the timeline without actually modifying the directory
416 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
417 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
418 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
419 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
420 : #[cfg(test)]
421 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
422 :
423 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
424 :
425 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
426 :
427 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
428 :
429 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
430 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
431 :
432 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
433 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
434 :
435 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
436 :
437 : /// May host a background Tokio task which downloads all the layers from the current
438 : /// heatmap on demand.
439 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
440 :
441 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
442 : }
443 :
444 : pub(crate) enum PreviousHeatmap {
445 : Active {
446 : heatmap: HeatMapTimeline,
447 : read_at: std::time::Instant,
448 : // End LSN covered by the heatmap if known
449 : end_lsn: Option<Lsn>,
450 : },
451 : Obsolete,
452 : }
453 :
454 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
455 :
456 : pub struct WalReceiverInfo {
457 : pub wal_source_connconf: PgConnectionConfig,
458 : pub last_received_msg_lsn: Lsn,
459 : pub last_received_msg_ts: u128,
460 : }
461 :
462 : /// Information about how much history needs to be retained, needed by
463 : /// Garbage Collection.
464 : #[derive(Default)]
465 : pub(crate) struct GcInfo {
466 : /// Specific LSNs that are needed.
467 : ///
468 : /// Currently, this includes all points where child branches have
469 : /// been forked off from. In the future, could also include
470 : /// explicit user-defined snapshot points.
471 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
472 :
473 : /// The cutoff coordinates, which are combined by selecting the minimum.
474 : pub(crate) cutoffs: GcCutoffs,
475 :
476 : /// Leases granted to particular LSNs.
477 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
478 :
479 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
480 : pub(crate) within_ancestor_pitr: bool,
481 : }
482 :
483 : impl GcInfo {
484 600 : pub(crate) fn min_cutoff(&self) -> Lsn {
485 600 : self.cutoffs.select_min()
486 600 : }
487 :
488 464 : pub(super) fn insert_child(
489 464 : &mut self,
490 464 : child_id: TimelineId,
491 464 : child_lsn: Lsn,
492 464 : is_offloaded: MaybeOffloaded,
493 464 : ) {
494 464 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
495 464 : self.retain_lsns.sort_by_key(|i| i.0);
496 464 : }
497 :
498 8 : pub(super) fn remove_child_maybe_offloaded(
499 8 : &mut self,
500 8 : child_id: TimelineId,
501 8 : maybe_offloaded: MaybeOffloaded,
502 8 : ) -> bool {
503 8 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
504 8 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
505 8 : let mut removed = false;
506 12 : self.retain_lsns.retain(|i| {
507 12 : if removed {
508 4 : return true;
509 8 : }
510 8 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
511 8 : removed |= remove;
512 8 : !remove
513 12 : });
514 8 : removed
515 8 : }
516 :
517 8 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
518 8 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
519 8 : }
520 :
521 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
522 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
523 0 : }
524 464 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
525 464 : self.leases.contains_key(&lsn)
526 464 : }
527 : }
528 :
529 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
530 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
531 : /// between time-based and space-based retention for observability and consumption metrics purposes.
532 : #[derive(Debug, Clone)]
533 : pub(crate) struct GcCutoffs {
534 : /// Calculated from the [`TenantConf::gc_horizon`], this LSN indicates how much
535 : /// history we must keep to retain a specified number of bytes of WAL.
536 : pub(crate) space: Lsn,
537 :
538 : /// Calculated from [`TenantConf::pitr_interval`], this LSN indicates how much
539 : /// history we must keep to enable reading back at least the PITR interval duration.
540 : pub(crate) time: Lsn,
541 : }
542 :
543 : impl Default for GcCutoffs {
544 904 : fn default() -> Self {
545 904 : Self {
546 904 : space: Lsn::INVALID,
547 904 : time: Lsn::INVALID,
548 904 : }
549 904 : }
550 : }
551 :
552 : impl GcCutoffs {
553 600 : fn select_min(&self) -> Lsn {
554 600 : std::cmp::min(self.space, self.time)
555 600 : }
556 : }
557 :
558 : pub(crate) struct TimelineVisitOutcome {
559 : completed_keyspace: KeySpace,
560 : image_covered_keyspace: KeySpace,
561 : }
562 :
563 : /// An error happened in a get() operation.
564 : #[derive(thiserror::Error, Debug)]
565 : pub(crate) enum PageReconstructError {
566 : #[error(transparent)]
567 : Other(anyhow::Error),
568 :
569 : #[error("Ancestor LSN wait error: {0}")]
570 : AncestorLsnTimeout(WaitLsnError),
571 :
572 : #[error("timeline shutting down")]
573 : Cancelled,
574 :
575 : /// An error happened replaying WAL records
576 : #[error(transparent)]
577 : WalRedo(anyhow::Error),
578 :
579 : #[error("{0}")]
580 : MissingKey(MissingKeyError),
581 : }
582 :
583 : impl From<anyhow::Error> for PageReconstructError {
584 0 : fn from(value: anyhow::Error) -> Self {
585 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
586 0 : match value.downcast::<PageReconstructError>() {
587 0 : Ok(pre) => pre,
588 0 : Err(other) => PageReconstructError::Other(other),
589 : }
590 0 : }
591 : }
592 :
593 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
594 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
595 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
596 0 : }
597 : }
598 :
599 : impl From<layer_manager::Shutdown> for PageReconstructError {
600 0 : fn from(_: layer_manager::Shutdown) -> Self {
601 0 : PageReconstructError::Cancelled
602 0 : }
603 : }
604 :
605 : impl GetVectoredError {
606 : #[cfg(test)]
607 12 : pub(crate) fn is_missing_key_error(&self) -> bool {
608 12 : matches!(self, Self::MissingKey(_))
609 12 : }
610 : }
611 :
612 : impl From<layer_manager::Shutdown> for GetVectoredError {
613 0 : fn from(_: layer_manager::Shutdown) -> Self {
614 0 : GetVectoredError::Cancelled
615 0 : }
616 : }
617 :
618 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
619 : /// only and not used by the "real read path".
620 : pub enum ReadPathLayerId {
621 : PersistentLayer(PersistentLayerKey),
622 : InMemoryLayer(Range<Lsn>),
623 : }
624 :
625 : impl std::fmt::Display for ReadPathLayerId {
626 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
627 0 : match self {
628 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
629 0 : ReadPathLayerId::InMemoryLayer(range) => {
630 0 : write!(f, "in-mem {}..{}", range.start, range.end)
631 : }
632 : }
633 0 : }
634 : }
635 : pub struct ReadPath {
636 : keyspace: KeySpace,
637 : lsn: Lsn,
638 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
639 : }
640 :
641 : impl ReadPath {
642 1255306 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
643 1255306 : Self {
644 1255306 : keyspace,
645 1255306 : lsn,
646 1255306 : path: Vec::new(),
647 1255306 : }
648 1255306 : }
649 :
650 1693718 : pub fn record_layer_visit(
651 1693718 : &mut self,
652 1693718 : layer_to_read: &ReadableLayer,
653 1693718 : keyspace_to_read: &KeySpace,
654 1693718 : lsn_range: &Range<Lsn>,
655 1693718 : ) {
656 1693718 : let id = match layer_to_read {
657 480346 : ReadableLayer::PersistentLayer(layer) => {
658 480346 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
659 : }
660 1213372 : ReadableLayer::InMemoryLayer(layer) => {
661 1213372 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
662 : }
663 : };
664 1693718 : self.path
665 1693718 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
666 1693718 : }
667 : }
668 :
669 : impl std::fmt::Display for ReadPath {
670 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
671 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
672 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
673 0 : writeln!(
674 0 : f,
675 0 : "{}: {} {}..{} {}",
676 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
677 0 : )?;
678 : }
679 0 : Ok(())
680 0 : }
681 : }
682 :
683 : #[derive(thiserror::Error)]
684 : pub struct MissingKeyError {
685 : key: Key,
686 : shard: ShardNumber,
687 : cont_lsn: Lsn,
688 : request_lsn: Lsn,
689 : ancestor_lsn: Option<Lsn>,
690 : /// Debug information about the read path if there's an error
691 : read_path: Option<ReadPath>,
692 : backtrace: Option<std::backtrace::Backtrace>,
693 : }
694 :
695 : impl std::fmt::Debug for MissingKeyError {
696 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
697 0 : write!(f, "{}", self)
698 0 : }
699 : }
700 :
701 : impl std::fmt::Display for MissingKeyError {
702 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
703 0 : write!(
704 0 : f,
705 0 : "could not find data for key {} (shard {:?}) at LSN {}, request LSN {}",
706 0 : self.key, self.shard, self.cont_lsn, self.request_lsn
707 0 : )?;
708 :
709 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
710 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
711 0 : }
712 :
713 0 : if let Some(ref read_path) = self.read_path {
714 0 : write!(f, "\n{}", read_path)?;
715 0 : }
716 :
717 0 : if let Some(ref backtrace) = self.backtrace {
718 0 : write!(f, "\n{}", backtrace)?;
719 0 : }
720 :
721 0 : Ok(())
722 0 : }
723 : }
724 :
725 : impl PageReconstructError {
726 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
727 0 : pub(crate) fn is_stopping(&self) -> bool {
728 : use PageReconstructError::*;
729 0 : match self {
730 0 : Cancelled => true,
731 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
732 : }
733 0 : }
734 : }
735 :
736 : #[derive(thiserror::Error, Debug)]
737 : pub(crate) enum CreateImageLayersError {
738 : #[error("timeline shutting down")]
739 : Cancelled,
740 :
741 : #[error("read failed")]
742 : GetVectoredError(#[source] GetVectoredError),
743 :
744 : #[error("reconstruction failed")]
745 : PageReconstructError(#[source] PageReconstructError),
746 :
747 : #[error(transparent)]
748 : Other(#[from] anyhow::Error),
749 : }
750 :
751 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
752 0 : fn from(_: layer_manager::Shutdown) -> Self {
753 0 : CreateImageLayersError::Cancelled
754 0 : }
755 : }
756 :
757 : #[derive(thiserror::Error, Debug, Clone)]
758 : pub(crate) enum FlushLayerError {
759 : /// Timeline cancellation token was cancelled
760 : #[error("timeline shutting down")]
761 : Cancelled,
762 :
763 : /// We tried to flush a layer while the Timeline is in an unexpected state
764 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
765 : NotRunning(FlushLoopState),
766 :
767 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
768 : // loop via a watch channel, where we can only borrow it.
769 : #[error("create image layers (shared)")]
770 : CreateImageLayersError(Arc<CreateImageLayersError>),
771 :
772 : #[error("other (shared)")]
773 : Other(#[from] Arc<anyhow::Error>),
774 : }
775 :
776 : impl FlushLayerError {
777 : // When crossing from generic anyhow errors to this error type, we explicitly check
778 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
779 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
780 0 : let cancelled = timeline.cancel.is_cancelled()
781 : // The upload queue might have been shut down before the official cancellation of the timeline.
782 0 : || err
783 0 : .downcast_ref::<NotInitialized>()
784 0 : .map(NotInitialized::is_stopping)
785 0 : .unwrap_or_default();
786 0 : if cancelled {
787 0 : Self::Cancelled
788 : } else {
789 0 : Self::Other(Arc::new(err))
790 : }
791 0 : }
792 : }
793 :
794 : impl From<layer_manager::Shutdown> for FlushLayerError {
795 0 : fn from(_: layer_manager::Shutdown) -> Self {
796 0 : FlushLayerError::Cancelled
797 0 : }
798 : }
799 :
800 : #[derive(thiserror::Error, Debug)]
801 : pub(crate) enum GetVectoredError {
802 : #[error("timeline shutting down")]
803 : Cancelled,
804 :
805 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
806 : Oversized(u64),
807 :
808 : #[error("requested at invalid LSN: {0}")]
809 : InvalidLsn(Lsn),
810 :
811 : #[error("requested key not found: {0}")]
812 : MissingKey(MissingKeyError),
813 :
814 : #[error("ancestry walk")]
815 : GetReadyAncestorError(#[source] GetReadyAncestorError),
816 :
817 : #[error(transparent)]
818 : Other(#[from] anyhow::Error),
819 : }
820 :
821 : impl From<GetReadyAncestorError> for GetVectoredError {
822 4 : fn from(value: GetReadyAncestorError) -> Self {
823 : use GetReadyAncestorError::*;
824 4 : match value {
825 0 : Cancelled => GetVectoredError::Cancelled,
826 : AncestorLsnTimeout(_) | BadState { .. } => {
827 4 : GetVectoredError::GetReadyAncestorError(value)
828 : }
829 : }
830 4 : }
831 : }
832 :
833 : #[derive(thiserror::Error, Debug)]
834 : pub(crate) enum GetReadyAncestorError {
835 : #[error("ancestor LSN wait error")]
836 : AncestorLsnTimeout(#[from] WaitLsnError),
837 :
838 : #[error("bad state on timeline {timeline_id}: {state:?}")]
839 : BadState {
840 : timeline_id: TimelineId,
841 : state: TimelineState,
842 : },
843 :
844 : #[error("cancelled")]
845 : Cancelled,
846 : }
847 :
848 : #[derive(Clone, Copy)]
849 : pub enum LogicalSizeCalculationCause {
850 : Initial,
851 : ConsumptionMetricsSyntheticSize,
852 : EvictionTaskImitation,
853 : TenantSizeHandler,
854 : }
855 :
856 : pub enum GetLogicalSizePriority {
857 : User,
858 : Background,
859 : }
860 :
861 0 : #[derive(Debug, enumset::EnumSetType)]
862 : pub(crate) enum CompactFlags {
863 : ForceRepartition,
864 : ForceImageLayerCreation,
865 : ForceL0Compaction,
866 : OnlyL0Compaction,
867 : EnhancedGcBottomMostCompaction,
868 : DryRun,
869 : /// Disables compaction yielding e.g. due to high L0 count. This is set e.g. when requesting
870 : /// compaction via HTTP API.
871 : NoYield,
872 : }
873 :
874 : #[serde_with::serde_as]
875 0 : #[derive(Debug, Clone, serde::Deserialize)]
876 : pub(crate) struct CompactRequest {
877 : pub compact_key_range: Option<CompactKeyRange>,
878 : pub compact_lsn_range: Option<CompactLsnRange>,
879 : /// Whether the compaction job should be scheduled.
880 : #[serde(default)]
881 : pub scheduled: bool,
882 : /// Whether the compaction job should be split across key ranges.
883 : #[serde(default)]
884 : pub sub_compaction: bool,
885 : /// Max job size for each subcompaction job.
886 : pub sub_compaction_max_job_size_mb: Option<u64>,
887 : }
888 :
889 : #[derive(Debug, Clone, Default)]
890 : pub(crate) struct CompactOptions {
891 : pub flags: EnumSet<CompactFlags>,
892 : /// If set, the compaction will only compact the key range specified by this option.
893 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
894 : pub compact_key_range: Option<CompactKeyRange>,
895 : /// If set, the compaction will only compact the LSN within this value.
896 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
897 : pub compact_lsn_range: Option<CompactLsnRange>,
898 : /// Enable sub-compaction (split compaction job across key ranges).
899 : /// This option is only used by GC compaction.
900 : pub sub_compaction: bool,
901 : /// Set job size for the GC compaction.
902 : /// This option is only used by GC compaction.
903 : pub sub_compaction_max_job_size_mb: Option<u64>,
904 : }
905 :
906 : impl std::fmt::Debug for Timeline {
907 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
908 0 : write!(f, "Timeline<{}>", self.timeline_id)
909 0 : }
910 : }
911 :
912 : #[derive(thiserror::Error, Debug)]
913 : pub(crate) enum WaitLsnError {
914 : // Called on a timeline which is shutting down
915 : #[error("Shutdown")]
916 : Shutdown,
917 :
918 : // Called on an timeline not in active state or shutting down
919 : #[error("Bad timeline state: {0:?}")]
920 : BadState(TimelineState),
921 :
922 : // Timeout expired while waiting for LSN to catch up with goal.
923 : #[error("{0}")]
924 : Timeout(String),
925 : }
926 :
927 : // The impls below achieve cancellation mapping for errors.
928 : // Perhaps there's a way of achieving this with less cruft.
929 :
930 : impl From<CreateImageLayersError> for CompactionError {
931 0 : fn from(e: CreateImageLayersError) -> Self {
932 0 : match e {
933 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
934 0 : CreateImageLayersError::Other(e) => {
935 0 : CompactionError::Other(e.context("create image layers"))
936 : }
937 0 : _ => CompactionError::Other(e.into()),
938 : }
939 0 : }
940 : }
941 :
942 : impl From<CreateImageLayersError> for FlushLayerError {
943 0 : fn from(e: CreateImageLayersError) -> Self {
944 0 : match e {
945 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
946 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
947 : }
948 0 : }
949 : }
950 :
951 : impl From<PageReconstructError> for CreateImageLayersError {
952 0 : fn from(e: PageReconstructError) -> Self {
953 0 : match e {
954 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
955 0 : _ => CreateImageLayersError::PageReconstructError(e),
956 : }
957 0 : }
958 : }
959 :
960 : impl From<GetVectoredError> for CreateImageLayersError {
961 0 : fn from(e: GetVectoredError) -> Self {
962 0 : match e {
963 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
964 0 : _ => CreateImageLayersError::GetVectoredError(e),
965 : }
966 0 : }
967 : }
968 :
969 : impl From<GetVectoredError> for PageReconstructError {
970 12 : fn from(e: GetVectoredError) -> Self {
971 12 : match e {
972 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
973 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
974 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
975 8 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
976 4 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
977 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
978 : }
979 12 : }
980 : }
981 :
982 : impl From<GetReadyAncestorError> for PageReconstructError {
983 4 : fn from(e: GetReadyAncestorError) -> Self {
984 : use GetReadyAncestorError::*;
985 4 : match e {
986 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
987 4 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
988 0 : Cancelled => PageReconstructError::Cancelled,
989 : }
990 4 : }
991 : }
992 :
993 : pub(crate) enum WaitLsnTimeout {
994 : Custom(Duration),
995 : // Use the [`PageServerConf::wait_lsn_timeout`] default
996 : Default,
997 : }
998 :
999 : pub(crate) enum WaitLsnWaiter<'a> {
1000 : Timeline(&'a Timeline),
1001 : Tenant,
1002 : PageService,
1003 : HttpEndpoint,
1004 : }
1005 :
1006 : /// Argument to [`Timeline::shutdown`].
1007 : #[derive(Debug, Clone, Copy)]
1008 : pub(crate) enum ShutdownMode {
1009 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1010 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1011 : ///
1012 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1013 : /// the call to [`Timeline::shutdown`].
1014 : FreezeAndFlush,
1015 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1016 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1017 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1018 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1019 : Reload,
1020 : /// Shut down immediately, without waiting for any open layers to flush.
1021 : Hard,
1022 : }
1023 :
1024 : enum ImageLayerCreationOutcome {
1025 : /// We generated an image layer
1026 : Generated {
1027 : unfinished_image_layer: ImageLayerWriter,
1028 : },
1029 : /// The key range is empty
1030 : Empty,
1031 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1032 : /// the image layer creation.
1033 : Skip,
1034 : }
1035 :
1036 : /// Public interface functions
1037 : impl Timeline {
1038 : /// Get the LSN where this branch was created
1039 8 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1040 8 : self.ancestor_lsn
1041 8 : }
1042 :
1043 : /// Get the ancestor's timeline id
1044 24 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1045 24 : self.ancestor_timeline
1046 24 : .as_ref()
1047 24 : .map(|ancestor| ancestor.timeline_id)
1048 24 : }
1049 :
1050 : /// Get the ancestor timeline
1051 4 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1052 4 : self.ancestor_timeline.as_ref()
1053 4 : }
1054 :
1055 : /// Get the bytes written since the PITR cutoff on this branch, and
1056 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1057 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1058 0 : let gc_info = self.gc_info.read().unwrap();
1059 0 : let history = self
1060 0 : .get_last_record_lsn()
1061 0 : .checked_sub(gc_info.cutoffs.time)
1062 0 : .unwrap_or(Lsn(0))
1063 0 : .0;
1064 0 : (history, gc_info.within_ancestor_pitr)
1065 0 : }
1066 :
1067 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1068 1704239 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1069 1704239 : self.applied_gc_cutoff_lsn.read()
1070 1704239 : }
1071 :
1072 : /// Read timeline's planned GC cutoff: this is the logical end of history that users
1073 : /// are allowed to read (based on configured PITR), even if physically we have more history.
1074 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Lsn {
1075 0 : self.gc_info.read().unwrap().cutoffs.time
1076 0 : }
1077 :
1078 : /// Look up given page version.
1079 : ///
1080 : /// If a remote layer file is needed, it is downloaded as part of this
1081 : /// call.
1082 : ///
1083 : /// This method enforces [`Self::pagestream_throttle`] internally.
1084 : ///
1085 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1086 : /// abstraction above this needs to store suitable metadata to track what
1087 : /// data exists with what keys, in separate metadata entries. If a
1088 : /// non-existent key is requested, we may incorrectly return a value from
1089 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1090 : /// non-existing key.
1091 : ///
1092 : /// # Cancel-Safety
1093 : ///
1094 : /// This method is cancellation-safe.
1095 : #[inline(always)]
1096 1215130 : pub(crate) async fn get(
1097 1215130 : &self,
1098 1215130 : key: Key,
1099 1215130 : lsn: Lsn,
1100 1215130 : ctx: &RequestContext,
1101 1215130 : ) -> Result<Bytes, PageReconstructError> {
1102 1215130 : if !lsn.is_valid() {
1103 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1104 1215130 : }
1105 1215130 :
1106 1215130 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1107 1215130 : // already checked the key against the shard_identity when looking up the Timeline from
1108 1215130 : // page_service.
1109 1215130 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1110 :
1111 1215130 : let keyspace = KeySpace {
1112 1215130 : ranges: vec![key..key.next()],
1113 1215130 : };
1114 1215130 :
1115 1215130 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1116 :
1117 1215130 : let vectored_res = self
1118 1215130 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
1119 1215130 : .await;
1120 :
1121 1215130 : let key_value = vectored_res?.pop_first();
1122 1215118 : match key_value {
1123 1215094 : Some((got_key, value)) => {
1124 1215094 : if got_key != key {
1125 0 : error!(
1126 0 : "Expected {}, but singular vectored get returned {}",
1127 : key, got_key
1128 : );
1129 0 : Err(PageReconstructError::Other(anyhow!(
1130 0 : "Singular vectored get returned wrong key"
1131 0 : )))
1132 : } else {
1133 1215094 : value
1134 : }
1135 : }
1136 24 : None => Err(PageReconstructError::MissingKey(MissingKeyError {
1137 24 : key,
1138 24 : shard: self.shard_identity.get_shard_number(&key),
1139 24 : cont_lsn: Lsn(0),
1140 24 : request_lsn: lsn,
1141 24 : ancestor_lsn: None,
1142 24 : backtrace: None,
1143 24 : read_path: None,
1144 24 : })),
1145 : }
1146 1215130 : }
1147 :
1148 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1149 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1150 :
1151 : /// Look up multiple page versions at a given LSN
1152 : ///
1153 : /// This naive implementation will be replaced with a more efficient one
1154 : /// which actually vectorizes the read path.
1155 39432 : pub(crate) async fn get_vectored(
1156 39432 : &self,
1157 39432 : keyspace: KeySpace,
1158 39432 : lsn: Lsn,
1159 39432 : io_concurrency: super::storage_layer::IoConcurrency,
1160 39432 : ctx: &RequestContext,
1161 39432 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1162 39432 : if !lsn.is_valid() {
1163 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1164 39432 : }
1165 39432 :
1166 39432 : let key_count = keyspace.total_raw_size().try_into().unwrap();
1167 39432 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1168 0 : return Err(GetVectoredError::Oversized(key_count));
1169 39432 : }
1170 :
1171 78864 : for range in &keyspace.ranges {
1172 39432 : let mut key = range.start;
1173 79380 : while key != range.end {
1174 39948 : assert!(!self.shard_identity.is_key_disposable(&key));
1175 39948 : key = key.next();
1176 : }
1177 : }
1178 :
1179 39432 : trace!(
1180 0 : "get vectored request for {:?}@{} from task kind {:?}",
1181 0 : keyspace,
1182 0 : lsn,
1183 0 : ctx.task_kind(),
1184 : );
1185 :
1186 39432 : let start = crate::metrics::GET_VECTORED_LATENCY
1187 39432 : .for_task_kind(ctx.task_kind())
1188 39432 : .map(|metric| (metric, Instant::now()));
1189 :
1190 39432 : let res = self
1191 39432 : .get_vectored_impl(
1192 39432 : keyspace.clone(),
1193 39432 : lsn,
1194 39432 : &mut ValuesReconstructState::new(io_concurrency),
1195 39432 : ctx,
1196 39432 : )
1197 39432 : .await;
1198 :
1199 39432 : if let Some((metric, start)) = start {
1200 0 : let elapsed = start.elapsed();
1201 0 : metric.observe(elapsed.as_secs_f64());
1202 39432 : }
1203 :
1204 39432 : res
1205 39432 : }
1206 :
1207 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1208 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1209 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1210 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1211 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1212 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1213 : /// the scan operation will not cause OOM in the future.
1214 24 : pub(crate) async fn scan(
1215 24 : &self,
1216 24 : keyspace: KeySpace,
1217 24 : lsn: Lsn,
1218 24 : ctx: &RequestContext,
1219 24 : io_concurrency: super::storage_layer::IoConcurrency,
1220 24 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1221 24 : if !lsn.is_valid() {
1222 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1223 24 : }
1224 24 :
1225 24 : trace!(
1226 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1227 0 : keyspace,
1228 0 : lsn,
1229 0 : ctx.task_kind()
1230 : );
1231 :
1232 : // We should generalize this into Keyspace::contains in the future.
1233 48 : for range in &keyspace.ranges {
1234 24 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1235 24 : || range.end.field1 > METADATA_KEY_END_PREFIX
1236 : {
1237 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1238 0 : "only metadata keyspace can be scanned"
1239 0 : )));
1240 24 : }
1241 : }
1242 :
1243 24 : let start = crate::metrics::SCAN_LATENCY
1244 24 : .for_task_kind(ctx.task_kind())
1245 24 : .map(ScanLatencyOngoingRecording::start_recording);
1246 :
1247 24 : let vectored_res = self
1248 24 : .get_vectored_impl(
1249 24 : keyspace.clone(),
1250 24 : lsn,
1251 24 : &mut ValuesReconstructState::new(io_concurrency),
1252 24 : ctx,
1253 24 : )
1254 24 : .await;
1255 :
1256 24 : if let Some(recording) = start {
1257 0 : recording.observe();
1258 24 : }
1259 :
1260 24 : vectored_res
1261 24 : }
1262 :
1263 1255306 : pub(super) async fn get_vectored_impl(
1264 1255306 : &self,
1265 1255306 : keyspace: KeySpace,
1266 1255306 : lsn: Lsn,
1267 1255306 : reconstruct_state: &mut ValuesReconstructState,
1268 1255306 : ctx: &RequestContext,
1269 1255306 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1270 1255306 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1271 1255306 : Some(ReadPath::new(keyspace.clone(), lsn))
1272 : } else {
1273 0 : None
1274 : };
1275 1255306 : reconstruct_state.read_path = read_path;
1276 :
1277 1255306 : let traversal_res: Result<(), _> = self
1278 1255306 : .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
1279 1255306 : .await;
1280 1255306 : if let Err(err) = traversal_res {
1281 : // Wait for all the spawned IOs to complete.
1282 : // See comments on `spawn_io` inside `storage_layer` for more details.
1283 32 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1284 32 : .into_values()
1285 32 : .map(|state| state.collect_pending_ios())
1286 32 : .collect::<FuturesUnordered<_>>();
1287 32 : while collect_futs.next().await.is_some() {}
1288 32 : return Err(err);
1289 1255274 : };
1290 1255274 :
1291 1255274 : let layers_visited = reconstruct_state.get_layers_visited();
1292 1255274 :
1293 1255274 : let futs = FuturesUnordered::new();
1294 1335978 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1295 1335978 : futs.push({
1296 1335978 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1297 1335978 : async move {
1298 1335978 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1299 :
1300 1335978 : let converted = match state.collect_pending_ios().await {
1301 1335978 : Ok(ok) => ok,
1302 0 : Err(err) => {
1303 0 : return (key, Err(err));
1304 : }
1305 : };
1306 1335978 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1307 1335978 :
1308 1335978 : // The walredo module expects the records to be descending in terms of Lsn.
1309 1335978 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1310 1335978 : debug_assert!(
1311 1335978 : converted
1312 1335978 : .records
1313 1335978 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1314 0 : "{converted:?}"
1315 : );
1316 :
1317 : (
1318 1335978 : key,
1319 1335978 : walredo_self.reconstruct_value(key, lsn, converted).await,
1320 : )
1321 1335978 : }
1322 1335978 : });
1323 1335978 : }
1324 :
1325 1255274 : let results = futs
1326 1255274 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1327 1255274 : .await;
1328 :
1329 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1330 : // when they're missing. Instead they are omitted from the resulting btree
1331 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1332 : // to avoid infinite results.
1333 1255274 : if !results.is_empty() {
1334 1254782 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1335 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1336 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1337 0 : LOG_PACER.lock().unwrap().call(|| {
1338 0 : let num_keys = keyspace.total_raw_size();
1339 0 : let num_pages = results.len();
1340 0 : tracing::info!(
1341 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1342 0 : lsn = %lsn,
1343 0 : "Vectored read for {keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1344 : );
1345 0 : });
1346 1254782 : }
1347 :
1348 : // Records the number of layers visited in a few different ways:
1349 : //
1350 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1351 : // layer directly affects its observed latency.
1352 : //
1353 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1354 : // layer visits and access cost.
1355 : //
1356 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1357 : // read amplification after batching.
1358 1254782 : let layers_visited = layers_visited as f64;
1359 1254782 : let avg_layers_visited = layers_visited / results.len() as f64;
1360 1254782 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1361 2590760 : for _ in &results {
1362 1335978 : self.metrics.layers_per_read.observe(layers_visited);
1363 1335978 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1364 1335978 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1365 1335978 : }
1366 492 : }
1367 :
1368 1255274 : Ok(results)
1369 1255306 : }
1370 :
1371 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1372 548924 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1373 548924 : self.last_record_lsn.load().last
1374 548924 : }
1375 :
1376 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1377 0 : self.last_record_lsn.load().prev
1378 0 : }
1379 :
1380 : /// Atomically get both last and prev.
1381 456 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1382 456 : self.last_record_lsn.load()
1383 456 : }
1384 :
1385 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1386 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1387 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1388 0 : self.last_record_lsn.status_receiver()
1389 0 : }
1390 :
1391 896 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1392 896 : self.disk_consistent_lsn.load()
1393 896 : }
1394 :
1395 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1396 : /// not validated with control plane yet.
1397 : /// See [`Self::get_remote_consistent_lsn_visible`].
1398 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1399 0 : self.remote_client.remote_consistent_lsn_projected()
1400 0 : }
1401 :
1402 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1403 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1404 : /// generation validation in the deletion queue.
1405 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1406 0 : self.remote_client.remote_consistent_lsn_visible()
1407 0 : }
1408 :
1409 : /// The sum of the file size of all historic layers in the layer map.
1410 : /// This method makes no distinction between local and remote layers.
1411 : /// Hence, the result **does not represent local filesystem usage**.
1412 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1413 0 : let guard = self.layers.read().await;
1414 0 : guard.layer_size_sum()
1415 0 : }
1416 :
1417 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1418 0 : self.metrics.resident_physical_size_get()
1419 0 : }
1420 :
1421 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1422 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1423 0 : }
1424 :
1425 : ///
1426 : /// Wait until WAL has been received and processed up to this LSN.
1427 : ///
1428 : /// You should call this before any of the other get_* or list_* functions. Calling
1429 : /// those functions with an LSN that has been processed yet is an error.
1430 : ///
1431 448329 : pub(crate) async fn wait_lsn(
1432 448329 : &self,
1433 448329 : lsn: Lsn,
1434 448329 : who_is_waiting: WaitLsnWaiter<'_>,
1435 448329 : timeout: WaitLsnTimeout,
1436 448329 : ctx: &RequestContext, /* Prepare for use by cancellation */
1437 448329 : ) -> Result<(), WaitLsnError> {
1438 448329 : let state = self.current_state();
1439 448329 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1440 0 : return Err(WaitLsnError::Shutdown);
1441 448329 : } else if !matches!(state, TimelineState::Active) {
1442 0 : return Err(WaitLsnError::BadState(state));
1443 448329 : }
1444 448329 :
1445 448329 : if cfg!(debug_assertions) {
1446 448329 : match ctx.task_kind() {
1447 : TaskKind::WalReceiverManager
1448 : | TaskKind::WalReceiverConnectionHandler
1449 : | TaskKind::WalReceiverConnectionPoller => {
1450 0 : let is_myself = match who_is_waiting {
1451 0 : WaitLsnWaiter::Timeline(waiter) => {
1452 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1453 : }
1454 : WaitLsnWaiter::Tenant
1455 : | WaitLsnWaiter::PageService
1456 0 : | WaitLsnWaiter::HttpEndpoint => unreachable!(
1457 0 : "tenant or page_service context are not expected to have task kind {:?}",
1458 0 : ctx.task_kind()
1459 0 : ),
1460 : };
1461 0 : if is_myself {
1462 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1463 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1464 0 : panic!(
1465 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1466 0 : );
1467 0 : }
1468 0 : } else {
1469 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1470 0 : // our walreceiver task can make progress independent of theirs
1471 0 : }
1472 : }
1473 448329 : _ => {}
1474 : }
1475 0 : }
1476 :
1477 448329 : let timeout = match timeout {
1478 0 : WaitLsnTimeout::Custom(t) => t,
1479 448329 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1480 : };
1481 :
1482 448329 : let _timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1483 448329 :
1484 448329 : match self.last_record_lsn.wait_for_timeout(lsn, timeout).await {
1485 448329 : Ok(()) => Ok(()),
1486 0 : Err(e) => {
1487 : use utils::seqwait::SeqWaitError::*;
1488 0 : match e {
1489 0 : Shutdown => Err(WaitLsnError::Shutdown),
1490 : Timeout => {
1491 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1492 0 : drop(_timer);
1493 0 : let walreceiver_status = self.walreceiver_status();
1494 0 : Err(WaitLsnError::Timeout(format!(
1495 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1496 0 : lsn,
1497 0 : self.get_last_record_lsn(),
1498 0 : self.get_disk_consistent_lsn(),
1499 0 : walreceiver_status,
1500 0 : )))
1501 : }
1502 : }
1503 : }
1504 : }
1505 448329 : }
1506 :
1507 0 : pub(crate) fn walreceiver_status(&self) -> String {
1508 0 : match &*self.walreceiver.lock().unwrap() {
1509 0 : None => "stopping or stopped".to_string(),
1510 0 : Some(walreceiver) => match walreceiver.status() {
1511 0 : Some(status) => status.to_human_readable_string(),
1512 0 : None => "Not active".to_string(),
1513 : },
1514 : }
1515 0 : }
1516 :
1517 : /// Check that it is valid to request operations with that lsn.
1518 464 : pub(crate) fn check_lsn_is_in_scope(
1519 464 : &self,
1520 464 : lsn: Lsn,
1521 464 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1522 464 : ) -> anyhow::Result<()> {
1523 464 : ensure!(
1524 464 : lsn >= **latest_gc_cutoff_lsn,
1525 8 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1526 8 : lsn,
1527 8 : **latest_gc_cutoff_lsn,
1528 : );
1529 456 : Ok(())
1530 464 : }
1531 :
1532 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1533 20 : pub(crate) fn init_lsn_lease(
1534 20 : &self,
1535 20 : lsn: Lsn,
1536 20 : length: Duration,
1537 20 : ctx: &RequestContext,
1538 20 : ) -> anyhow::Result<LsnLease> {
1539 20 : self.make_lsn_lease(lsn, length, true, ctx)
1540 20 : }
1541 :
1542 : /// 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.
1543 8 : pub(crate) fn renew_lsn_lease(
1544 8 : &self,
1545 8 : lsn: Lsn,
1546 8 : length: Duration,
1547 8 : ctx: &RequestContext,
1548 8 : ) -> anyhow::Result<LsnLease> {
1549 8 : self.make_lsn_lease(lsn, length, false, ctx)
1550 8 : }
1551 :
1552 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1553 : ///
1554 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1555 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1556 : ///
1557 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1558 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1559 28 : fn make_lsn_lease(
1560 28 : &self,
1561 28 : lsn: Lsn,
1562 28 : length: Duration,
1563 28 : init: bool,
1564 28 : _ctx: &RequestContext,
1565 28 : ) -> anyhow::Result<LsnLease> {
1566 24 : let lease = {
1567 : // Normalize the requested LSN to be aligned, and move to the first record
1568 : // if it points to the beginning of the page (header).
1569 28 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1570 28 :
1571 28 : let mut gc_info = self.gc_info.write().unwrap();
1572 28 : let planned_cutoff = gc_info.min_cutoff();
1573 28 :
1574 28 : let valid_until = SystemTime::now() + length;
1575 28 :
1576 28 : let entry = gc_info.leases.entry(lsn);
1577 28 :
1578 28 : match entry {
1579 12 : Entry::Occupied(mut occupied) => {
1580 12 : let existing_lease = occupied.get_mut();
1581 12 : if valid_until > existing_lease.valid_until {
1582 4 : existing_lease.valid_until = valid_until;
1583 4 : let dt: DateTime<Utc> = valid_until.into();
1584 4 : info!("lease extended to {}", dt);
1585 : } else {
1586 8 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1587 8 : info!("existing lease covers greater length, valid until {}", dt);
1588 : }
1589 :
1590 12 : existing_lease.clone()
1591 : }
1592 16 : Entry::Vacant(vacant) => {
1593 : // Reject already GC-ed LSN if we are in AttachedSingle and
1594 : // not blocked by the lsn lease deadline.
1595 16 : let validate = {
1596 16 : let conf = self.tenant_conf.load();
1597 16 : conf.location.attach_mode == AttachmentMode::Single
1598 16 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1599 : };
1600 :
1601 16 : if init || validate {
1602 16 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1603 16 : if lsn < *latest_gc_cutoff_lsn {
1604 4 : bail!(
1605 4 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1606 4 : lsn,
1607 4 : *latest_gc_cutoff_lsn
1608 4 : );
1609 12 : }
1610 12 : if lsn < planned_cutoff {
1611 0 : bail!(
1612 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1613 0 : lsn,
1614 0 : planned_cutoff
1615 0 : );
1616 12 : }
1617 0 : }
1618 :
1619 12 : let dt: DateTime<Utc> = valid_until.into();
1620 12 : info!("lease created, valid until {}", dt);
1621 12 : vacant.insert(LsnLease { valid_until }).clone()
1622 : }
1623 : }
1624 : };
1625 :
1626 24 : Ok(lease)
1627 28 : }
1628 :
1629 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1630 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1631 : #[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))]
1632 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1633 : self.freeze0().await
1634 : }
1635 :
1636 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1637 : #[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))]
1638 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1639 : self.freeze_and_flush0().await
1640 : }
1641 :
1642 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1643 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1644 2252 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1645 2252 : let mut g = self.write_lock.lock().await;
1646 2252 : let to_lsn = self.get_last_record_lsn();
1647 2252 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1648 2252 : }
1649 :
1650 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1651 : // polluting the span hierarchy.
1652 2252 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1653 2252 : let token = self.freeze0().await?;
1654 2252 : self.wait_flush_completion(token).await
1655 2252 : }
1656 :
1657 : // Check if an open ephemeral layer should be closed: this provides
1658 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1659 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1660 : // ephemeral layer bytes has been breached.
1661 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1662 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1663 : // If the write lock is held, there is an active wal receiver: rolling open layers
1664 : // is their responsibility while they hold this lock.
1665 0 : return;
1666 : };
1667 :
1668 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1669 : // time, and this was missed.
1670 : // if write_guard.is_none() { return; }
1671 :
1672 0 : let Ok(layers_guard) = self.layers.try_read() else {
1673 : // Don't block if the layer lock is busy
1674 0 : return;
1675 : };
1676 :
1677 0 : let Ok(lm) = layers_guard.layer_map() else {
1678 0 : return;
1679 : };
1680 :
1681 0 : let Some(open_layer) = &lm.open_layer else {
1682 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1683 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1684 : // that didn't result in writes to this shard.
1685 :
1686 : // Must not hold the layers lock while waiting for a flush.
1687 0 : drop(layers_guard);
1688 0 :
1689 0 : let last_record_lsn = self.get_last_record_lsn();
1690 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1691 0 : if last_record_lsn > disk_consistent_lsn {
1692 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1693 : // we are a sharded tenant and have skipped some WAL
1694 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1695 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1696 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1697 : // without any data ingested (yet) doesn't write a remote index as soon as it
1698 : // sees its LSN advance: we only do this if we've been layer-less
1699 : // for some time.
1700 0 : tracing::debug!(
1701 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1702 : disk_consistent_lsn,
1703 : last_record_lsn
1704 : );
1705 :
1706 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1707 : // We know there is no open layer, so we can request freezing without actually
1708 : // freezing anything. This is true even if we have dropped the layers_guard, we
1709 : // still hold the write_guard.
1710 0 : let _ = async {
1711 0 : let token = self
1712 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1713 0 : .await?;
1714 0 : self.wait_flush_completion(token).await
1715 0 : }
1716 0 : .await;
1717 0 : }
1718 0 : }
1719 :
1720 0 : return;
1721 : };
1722 :
1723 0 : let Some(current_size) = open_layer.try_len() else {
1724 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1725 : // read lock to get size should always succeed.
1726 0 : tracing::warn!("Lock conflict while reading size of open layer");
1727 0 : return;
1728 : };
1729 :
1730 0 : let current_lsn = self.get_last_record_lsn();
1731 :
1732 0 : let checkpoint_distance_override = open_layer.tick().await;
1733 :
1734 0 : if let Some(size_override) = checkpoint_distance_override {
1735 0 : if current_size > size_override {
1736 : // This is not harmful, but it only happens in relatively rare cases where
1737 : // time-based checkpoints are not happening fast enough to keep the amount of
1738 : // ephemeral data within configured limits. It's a sign of stress on the system.
1739 0 : tracing::info!(
1740 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
1741 : );
1742 0 : }
1743 0 : }
1744 :
1745 0 : let checkpoint_distance =
1746 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1747 0 :
1748 0 : if self.should_roll(
1749 0 : current_size,
1750 0 : current_size,
1751 0 : checkpoint_distance,
1752 0 : self.get_last_record_lsn(),
1753 0 : self.last_freeze_at.load(),
1754 0 : open_layer.get_opened_at(),
1755 0 : ) {
1756 0 : match open_layer.info() {
1757 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1758 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1759 0 : // happens asynchronously in the background.
1760 0 : tracing::debug!(
1761 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1762 : );
1763 : }
1764 : InMemoryLayerInfo::Open { .. } => {
1765 : // Upgrade to a write lock and freeze the layer
1766 0 : drop(layers_guard);
1767 0 : let res = self
1768 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1769 0 : .await;
1770 :
1771 0 : if let Err(e) = res {
1772 0 : tracing::info!(
1773 0 : "failed to flush frozen layer after background freeze: {e:#}"
1774 : );
1775 0 : }
1776 : }
1777 : }
1778 0 : }
1779 0 : }
1780 :
1781 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1782 : ///
1783 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1784 : /// child timelines that are not offloaded yet.
1785 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1786 0 : if self.remote_client.is_archived() != Some(true) {
1787 0 : return (false, "the timeline is not archived");
1788 0 : }
1789 0 : if !self.remote_client.no_pending_work() {
1790 : // if the remote client is still processing some work, we can't offload
1791 0 : return (false, "the upload queue is not drained yet");
1792 0 : }
1793 0 :
1794 0 : (true, "ok")
1795 0 : }
1796 :
1797 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1798 : /// compaction tasks.
1799 728 : pub(crate) async fn compact(
1800 728 : self: &Arc<Self>,
1801 728 : cancel: &CancellationToken,
1802 728 : flags: EnumSet<CompactFlags>,
1803 728 : ctx: &RequestContext,
1804 728 : ) -> Result<CompactionOutcome, CompactionError> {
1805 728 : self.compact_with_options(
1806 728 : cancel,
1807 728 : CompactOptions {
1808 728 : flags,
1809 728 : compact_key_range: None,
1810 728 : compact_lsn_range: None,
1811 728 : sub_compaction: false,
1812 728 : sub_compaction_max_job_size_mb: None,
1813 728 : },
1814 728 : ctx,
1815 728 : )
1816 728 : .await
1817 728 : }
1818 :
1819 : /// Outermost timeline compaction operation; downloads needed layers.
1820 : ///
1821 : /// NB: the cancellation token is usually from a background task, but can also come from a
1822 : /// request task.
1823 728 : pub(crate) async fn compact_with_options(
1824 728 : self: &Arc<Self>,
1825 728 : cancel: &CancellationToken,
1826 728 : options: CompactOptions,
1827 728 : ctx: &RequestContext,
1828 728 : ) -> Result<CompactionOutcome, CompactionError> {
1829 728 : // Acquire the compaction lock and task semaphore.
1830 728 : //
1831 728 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1832 728 : // out by other background tasks (including image compaction). We request this via
1833 728 : // `BackgroundLoopKind::L0Compaction`.
1834 728 : //
1835 728 : // If this is a regular compaction pass, and L0-only compaction is enabled in the config,
1836 728 : // then we should yield for immediate L0 compaction if necessary while we're waiting for the
1837 728 : // background task semaphore. There's no point yielding otherwise, since we'd just end up
1838 728 : // right back here.
1839 728 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1840 728 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1841 0 : true => BackgroundLoopKind::L0Compaction,
1842 728 : false => BackgroundLoopKind::Compaction,
1843 : };
1844 728 : let yield_for_l0 = !is_l0_only
1845 728 : && self.get_compaction_l0_first()
1846 0 : && !options.flags.contains(CompactFlags::NoYield);
1847 :
1848 728 : let acquire = async move {
1849 728 : let guard = self.compaction_lock.lock().await;
1850 728 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
1851 728 : (guard, permit)
1852 728 : };
1853 :
1854 728 : let (_guard, _permit) = tokio::select! {
1855 728 : (guard, permit) = acquire => (guard, permit),
1856 728 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
1857 0 : return Ok(CompactionOutcome::YieldForL0);
1858 : }
1859 728 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1860 728 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
1861 : };
1862 :
1863 728 : let last_record_lsn = self.get_last_record_lsn();
1864 728 :
1865 728 : // Last record Lsn could be zero in case the timeline was just created
1866 728 : if !last_record_lsn.is_valid() {
1867 0 : warn!(
1868 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
1869 : );
1870 0 : return Ok(CompactionOutcome::Skipped);
1871 728 : }
1872 :
1873 728 : let result = match self.get_compaction_algorithm_settings().kind {
1874 : CompactionAlgorithm::Tiered => {
1875 0 : self.compact_tiered(cancel, ctx).await?;
1876 0 : Ok(CompactionOutcome::Done)
1877 : }
1878 728 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
1879 : };
1880 :
1881 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
1882 0 : match &result {
1883 728 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
1884 0 : Err(e) if e.is_cancel() => {}
1885 0 : Err(CompactionError::ShuttingDown) => {
1886 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
1887 0 : }
1888 0 : Err(CompactionError::AlreadyRunning(_)) => {
1889 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
1890 0 : }
1891 : Err(CompactionError::Other(_)) => {
1892 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1893 : }
1894 : Err(CompactionError::CollectKeySpaceError(_)) => {
1895 : // Cancelled errors are covered by the `Err(e) if e.is_cancel()` branch.
1896 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
1897 : }
1898 : // Don't change the current value on offload failure or shutdown. We don't want to
1899 : // abruptly stall nor resume L0 flushes in these cases.
1900 0 : Err(CompactionError::Offload(_)) => {}
1901 : };
1902 :
1903 728 : result
1904 728 : }
1905 :
1906 : /// Mutate the timeline with a [`TimelineWriter`].
1907 10266392 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
1908 10266392 : TimelineWriter {
1909 10266392 : tl: self,
1910 10266392 : write_guard: self.write_lock.lock().await,
1911 : }
1912 10266392 : }
1913 :
1914 0 : pub(crate) fn activate(
1915 0 : self: &Arc<Self>,
1916 0 : parent: Arc<crate::tenant::Tenant>,
1917 0 : broker_client: BrokerClientChannel,
1918 0 : background_jobs_can_start: Option<&completion::Barrier>,
1919 0 : ctx: &RequestContext,
1920 0 : ) {
1921 0 : if self.tenant_shard_id.is_shard_zero() {
1922 0 : // Logical size is only maintained accurately on shard zero.
1923 0 : self.spawn_initial_logical_size_computation_task(ctx);
1924 0 : }
1925 0 : self.launch_wal_receiver(ctx, broker_client);
1926 0 : self.set_state(TimelineState::Active);
1927 0 : self.launch_eviction_task(parent, background_jobs_can_start);
1928 0 : }
1929 :
1930 : /// After this function returns, there are no timeline-scoped tasks are left running.
1931 : ///
1932 : /// The preferred pattern for is:
1933 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
1934 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
1935 : /// go the extra mile and keep track of JoinHandles
1936 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
1937 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
1938 : ///
1939 : /// For legacy reasons, we still have multiple tasks spawned using
1940 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
1941 : /// We refer to these as "timeline-scoped task_mgr tasks".
1942 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
1943 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
1944 : /// or [`task_mgr::shutdown_watcher`].
1945 : /// We want to gradually convert the code base away from these.
1946 : ///
1947 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
1948 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
1949 : /// ones that aren't mentioned here):
1950 : /// - [`TaskKind::TimelineDeletionWorker`]
1951 : /// - NB: also used for tenant deletion
1952 : /// - [`TaskKind::RemoteUploadTask`]`
1953 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
1954 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
1955 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
1956 : /// - [`TaskKind::Eviction`]
1957 : /// - [`TaskKind::LayerFlushTask`]
1958 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
1959 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
1960 20 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
1961 20 : debug_assert_current_span_has_tenant_and_timeline_id();
1962 20 :
1963 20 : // Regardless of whether we're going to try_freeze_and_flush
1964 20 : // or not, stop ingesting any more data. Walreceiver only provides
1965 20 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
1966 20 : // So, only after the self.gate.close() below will we know for sure that
1967 20 : // no walreceiver tasks are left.
1968 20 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
1969 20 : // data during the call to `self.freeze_and_flush()` below.
1970 20 : // That's not ideal, but, we don't have the concept of a ChildGuard,
1971 20 : // which is what we'd need to properly model early shutdown of the walreceiver
1972 20 : // task sub-tree before the other Timeline task sub-trees.
1973 20 : let walreceiver = self.walreceiver.lock().unwrap().take();
1974 20 : tracing::debug!(
1975 0 : is_some = walreceiver.is_some(),
1976 0 : "Waiting for WalReceiverManager..."
1977 : );
1978 20 : if let Some(walreceiver) = walreceiver {
1979 0 : walreceiver.cancel();
1980 20 : }
1981 : // ... and inform any waiters for newer LSNs that there won't be any.
1982 20 : self.last_record_lsn.shutdown();
1983 20 :
1984 20 : if let ShutdownMode::FreezeAndFlush = mode {
1985 12 : let do_flush = if let Some((open, frozen)) = self
1986 12 : .layers
1987 12 : .read()
1988 12 : .await
1989 12 : .layer_map()
1990 12 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
1991 12 : .ok()
1992 12 : .filter(|(open, frozen)| *open || *frozen > 0)
1993 : {
1994 0 : if self.remote_client.is_archived() == Some(true) {
1995 : // No point flushing on shutdown for an archived timeline: it is not important
1996 : // to have it nice and fresh after our restart, and trying to flush here might
1997 : // race with trying to offload it (which also stops the flush loop)
1998 0 : false
1999 : } else {
2000 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2001 0 : true
2002 : }
2003 : } else {
2004 : // this is double-shutdown, it'll be a no-op
2005 12 : true
2006 : };
2007 :
2008 : // we shut down walreceiver above, so, we won't add anything more
2009 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2010 : // to reach the disk & upload queue, then shut the upload queue and
2011 : // wait for it to drain.
2012 12 : if do_flush {
2013 12 : match self.freeze_and_flush().await {
2014 : Ok(_) => {
2015 : // drain the upload queue
2016 : // if we did not wait for completion here, it might be our shutdown process
2017 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2018 : // be spawned.
2019 : //
2020 : // what is problematic is the shutting down of RemoteTimelineClient, because
2021 : // obviously it does not make sense to stop while we wait for it, but what
2022 : // about corner cases like s3 suddenly hanging up?
2023 12 : self.remote_client.shutdown().await;
2024 : }
2025 : Err(FlushLayerError::Cancelled) => {
2026 : // this is likely the second shutdown, ignore silently.
2027 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2028 0 : debug_assert!(self.cancel.is_cancelled());
2029 : }
2030 0 : Err(e) => {
2031 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2032 0 : // we have some extra WAL replay to do next time the timeline starts.
2033 0 : warn!("failed to freeze and flush: {e:#}");
2034 : }
2035 : }
2036 :
2037 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2038 : // we also do a final check here to ensure that the queue is empty.
2039 12 : if !self.remote_client.no_pending_work() {
2040 0 : warn!(
2041 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2042 : );
2043 12 : }
2044 0 : }
2045 8 : }
2046 :
2047 20 : if let ShutdownMode::Reload = mode {
2048 : // drain the upload queue
2049 4 : self.remote_client.shutdown().await;
2050 4 : if !self.remote_client.no_pending_work() {
2051 0 : warn!(
2052 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2053 : );
2054 4 : }
2055 16 : }
2056 :
2057 : // Signal any subscribers to our cancellation token to drop out
2058 20 : tracing::debug!("Cancelling CancellationToken");
2059 20 : self.cancel.cancel();
2060 20 :
2061 20 : // If we have a background task downloading heatmap layers stop it.
2062 20 : // The background downloads are sensitive to timeline cancellation (done above),
2063 20 : // so the drain will be immediate.
2064 20 : self.stop_and_drain_heatmap_layers_download().await;
2065 :
2066 : // Ensure Prevent new page service requests from starting.
2067 20 : self.handles.shutdown();
2068 20 :
2069 20 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2070 20 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2071 20 : self.remote_client.stop();
2072 20 :
2073 20 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2074 20 : // to shut down the upload queue tasks.
2075 20 : // TODO: fix that, task management should be encapsulated inside remote_client.
2076 20 : task_mgr::shutdown_tasks(
2077 20 : Some(TaskKind::RemoteUploadTask),
2078 20 : Some(self.tenant_shard_id),
2079 20 : Some(self.timeline_id),
2080 20 : )
2081 20 : .await;
2082 :
2083 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2084 20 : tracing::debug!("Waiting for tasks...");
2085 20 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2086 :
2087 : {
2088 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2089 : // open.
2090 20 : let mut write_guard = self.write_lock.lock().await;
2091 20 : self.layers.write().await.shutdown(&mut write_guard);
2092 20 : }
2093 20 :
2094 20 : // Finally wait until any gate-holders are complete.
2095 20 : //
2096 20 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2097 20 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2098 20 : self.gate.close().await;
2099 :
2100 20 : self.metrics.shutdown();
2101 20 : }
2102 :
2103 908 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2104 908 : match (self.current_state(), new_state) {
2105 908 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2106 4 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2107 : }
2108 0 : (st, TimelineState::Loading) => {
2109 0 : error!("ignoring transition from {st:?} into Loading state");
2110 : }
2111 0 : (TimelineState::Broken { .. }, new_state) => {
2112 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2113 : }
2114 : (TimelineState::Stopping, TimelineState::Active) => {
2115 0 : error!("Not activating a Stopping timeline");
2116 : }
2117 904 : (_, new_state) => {
2118 904 : self.state.send_replace(new_state);
2119 904 : }
2120 : }
2121 908 : }
2122 :
2123 4 : pub(crate) fn set_broken(&self, reason: String) {
2124 4 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2125 4 : let broken_state = TimelineState::Broken {
2126 4 : reason,
2127 4 : backtrace: backtrace_str,
2128 4 : };
2129 4 : self.set_state(broken_state);
2130 4 :
2131 4 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2132 4 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2133 4 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2134 4 : self.cancel.cancel();
2135 4 : }
2136 :
2137 450445 : pub(crate) fn current_state(&self) -> TimelineState {
2138 450445 : self.state.borrow().clone()
2139 450445 : }
2140 :
2141 12 : pub(crate) fn is_broken(&self) -> bool {
2142 12 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2143 12 : }
2144 :
2145 472 : pub(crate) fn is_active(&self) -> bool {
2146 472 : self.current_state() == TimelineState::Active
2147 472 : }
2148 :
2149 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
2150 0 : self.remote_client.is_archived()
2151 0 : }
2152 :
2153 736 : pub(crate) fn is_stopping(&self) -> bool {
2154 736 : self.current_state() == TimelineState::Stopping
2155 736 : }
2156 :
2157 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2158 0 : self.state.subscribe()
2159 0 : }
2160 :
2161 448333 : pub(crate) async fn wait_to_become_active(
2162 448333 : &self,
2163 448333 : _ctx: &RequestContext, // Prepare for use by cancellation
2164 448333 : ) -> Result<(), TimelineState> {
2165 448333 : let mut receiver = self.state.subscribe();
2166 : loop {
2167 448333 : let current_state = receiver.borrow().clone();
2168 448333 : match current_state {
2169 : TimelineState::Loading => {
2170 0 : receiver
2171 0 : .changed()
2172 0 : .await
2173 0 : .expect("holding a reference to self");
2174 : }
2175 : TimelineState::Active { .. } => {
2176 448329 : return Ok(());
2177 : }
2178 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2179 : // There's no chance the timeline can transition back into ::Active
2180 4 : return Err(current_state);
2181 : }
2182 : }
2183 : }
2184 448333 : }
2185 :
2186 0 : pub(crate) async fn layer_map_info(
2187 0 : &self,
2188 0 : reset: LayerAccessStatsReset,
2189 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2190 0 : let guard = self.layers.read().await;
2191 0 : let layer_map = guard.layer_map()?;
2192 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2193 0 : if let Some(open_layer) = &layer_map.open_layer {
2194 0 : in_memory_layers.push(open_layer.info());
2195 0 : }
2196 0 : for frozen_layer in &layer_map.frozen_layers {
2197 0 : in_memory_layers.push(frozen_layer.info());
2198 0 : }
2199 :
2200 0 : let historic_layers = layer_map
2201 0 : .iter_historic_layers()
2202 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2203 0 : .collect();
2204 0 :
2205 0 : Ok(LayerMapInfo {
2206 0 : in_memory_layers,
2207 0 : historic_layers,
2208 0 : })
2209 0 : }
2210 :
2211 : #[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))]
2212 : pub(crate) async fn download_layer(
2213 : &self,
2214 : layer_file_name: &LayerName,
2215 : ctx: &RequestContext,
2216 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2217 : let Some(layer) = self
2218 : .find_layer(layer_file_name)
2219 : .await
2220 0 : .map_err(|e| match e {
2221 0 : layer_manager::Shutdown => {
2222 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2223 0 : }
2224 0 : })?
2225 : else {
2226 : return Ok(None);
2227 : };
2228 :
2229 : layer.download(ctx).await?;
2230 :
2231 : Ok(Some(true))
2232 : }
2233 :
2234 : /// Evict just one layer.
2235 : ///
2236 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2237 0 : pub(crate) async fn evict_layer(
2238 0 : &self,
2239 0 : layer_file_name: &LayerName,
2240 0 : ) -> anyhow::Result<Option<bool>> {
2241 0 : let _gate = self
2242 0 : .gate
2243 0 : .enter()
2244 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2245 :
2246 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2247 0 : return Ok(None);
2248 : };
2249 :
2250 : // curl has this by default
2251 0 : let timeout = std::time::Duration::from_secs(120);
2252 0 :
2253 0 : match local_layer.evict_and_wait(timeout).await {
2254 0 : Ok(()) => Ok(Some(true)),
2255 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2256 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2257 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2258 : }
2259 0 : }
2260 :
2261 9606020 : fn should_roll(
2262 9606020 : &self,
2263 9606020 : layer_size: u64,
2264 9606020 : projected_layer_size: u64,
2265 9606020 : checkpoint_distance: u64,
2266 9606020 : projected_lsn: Lsn,
2267 9606020 : last_freeze_at: Lsn,
2268 9606020 : opened_at: Instant,
2269 9606020 : ) -> bool {
2270 9606020 : let distance = projected_lsn.widening_sub(last_freeze_at);
2271 9606020 :
2272 9606020 : // Rolling the open layer can be triggered by:
2273 9606020 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2274 9606020 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2275 9606020 : // account for how writes are distributed across shards: we expect each node to consume
2276 9606020 : // 1/count of the LSN on average.
2277 9606020 : // 2. The size of the currently open layer.
2278 9606020 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2279 9606020 : // up and suspend activity.
2280 9606020 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2281 0 : info!(
2282 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2283 : projected_lsn, layer_size, distance
2284 : );
2285 :
2286 0 : true
2287 9606020 : } else if projected_layer_size >= checkpoint_distance {
2288 : // NB: this check is relied upon by:
2289 160 : let _ = IndexEntry::validate_checkpoint_distance;
2290 160 : info!(
2291 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2292 : projected_lsn, layer_size, projected_layer_size
2293 : );
2294 :
2295 160 : true
2296 9605860 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2297 0 : info!(
2298 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2299 0 : projected_lsn,
2300 0 : layer_size,
2301 0 : opened_at.elapsed()
2302 : );
2303 :
2304 0 : true
2305 : } else {
2306 9605860 : false
2307 : }
2308 9606020 : }
2309 : }
2310 :
2311 : /// Number of times we will compute partition within a checkpoint distance.
2312 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2313 :
2314 : // Private functions
2315 : impl Timeline {
2316 24 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2317 24 : let tenant_conf = self.tenant_conf.load();
2318 24 : tenant_conf
2319 24 : .tenant_conf
2320 24 : .lsn_lease_length
2321 24 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2322 24 : }
2323 :
2324 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2325 0 : let tenant_conf = self.tenant_conf.load();
2326 0 : tenant_conf
2327 0 : .tenant_conf
2328 0 : .lsn_lease_length_for_ts
2329 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2330 0 : }
2331 :
2332 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2333 0 : let tenant_conf = self.tenant_conf.load();
2334 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2335 0 : }
2336 :
2337 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2338 0 : let tenant_conf = self.tenant_conf.load();
2339 0 : tenant_conf
2340 0 : .tenant_conf
2341 0 : .lazy_slru_download
2342 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2343 0 : }
2344 :
2345 9609016 : fn get_checkpoint_distance(&self) -> u64 {
2346 9609016 : let tenant_conf = self.tenant_conf.load();
2347 9609016 : tenant_conf
2348 9609016 : .tenant_conf
2349 9609016 : .checkpoint_distance
2350 9609016 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2351 9609016 : }
2352 :
2353 9605860 : fn get_checkpoint_timeout(&self) -> Duration {
2354 9605860 : let tenant_conf = self.tenant_conf.load();
2355 9605860 : tenant_conf
2356 9605860 : .tenant_conf
2357 9605860 : .checkpoint_timeout
2358 9605860 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2359 9605860 : }
2360 :
2361 5032 : fn get_compaction_period(&self) -> Duration {
2362 5032 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2363 5032 : tenant_conf
2364 5032 : .compaction_period
2365 5032 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2366 5032 : }
2367 :
2368 1336 : fn get_compaction_target_size(&self) -> u64 {
2369 1336 : let tenant_conf = self.tenant_conf.load();
2370 1336 : tenant_conf
2371 1336 : .tenant_conf
2372 1336 : .compaction_target_size
2373 1336 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2374 1336 : }
2375 :
2376 3140 : fn get_compaction_threshold(&self) -> usize {
2377 3140 : let tenant_conf = self.tenant_conf.load();
2378 3140 : tenant_conf
2379 3140 : .tenant_conf
2380 3140 : .compaction_threshold
2381 3140 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2382 3140 : }
2383 :
2384 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2385 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2386 : /// possible that the index part persists the state while the config doesn't get persisted.
2387 3892 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2388 3892 : let tenant_conf = self.tenant_conf.load();
2389 3892 : tenant_conf
2390 3892 : .tenant_conf
2391 3892 : .rel_size_v2_enabled
2392 3892 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2393 3892 : }
2394 :
2395 4384 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2396 4384 : self.rel_size_v2_status
2397 4384 : .load()
2398 4384 : .as_ref()
2399 4384 : .map(|s| s.as_ref().clone())
2400 4384 : .unwrap_or(RelSizeMigration::Legacy)
2401 4384 : }
2402 :
2403 56 : fn get_compaction_upper_limit(&self) -> usize {
2404 56 : let tenant_conf = self.tenant_conf.load();
2405 56 : tenant_conf
2406 56 : .tenant_conf
2407 56 : .compaction_upper_limit
2408 56 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2409 56 : }
2410 :
2411 728 : pub fn get_compaction_l0_first(&self) -> bool {
2412 728 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2413 728 : tenant_conf
2414 728 : .compaction_l0_first
2415 728 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2416 728 : }
2417 :
2418 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2419 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2420 0 : tenant_conf
2421 0 : .compaction_l0_semaphore
2422 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2423 0 : }
2424 :
2425 2516 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2426 : // Disable L0 flushes by default. This and compaction needs further tuning.
2427 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3
2428 :
2429 : // If compaction is disabled, don't delay.
2430 2516 : if self.get_compaction_period() == Duration::ZERO {
2431 2516 : return None;
2432 0 : }
2433 0 :
2434 0 : let compaction_threshold = self.get_compaction_threshold();
2435 0 : let tenant_conf = self.tenant_conf.load();
2436 0 : let l0_flush_delay_threshold = tenant_conf
2437 0 : .tenant_conf
2438 0 : .l0_flush_delay_threshold
2439 0 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2440 0 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2441 0 :
2442 0 : // 0 disables backpressure.
2443 0 : if l0_flush_delay_threshold == 0 {
2444 0 : return None;
2445 0 : }
2446 0 :
2447 0 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2448 0 : // backpressure flushes below this.
2449 0 : // TODO: the tenant config should have validation to prevent this instead.
2450 0 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2451 0 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2452 2516 : }
2453 :
2454 2516 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2455 : // Disable L0 stalls by default. In ingest benchmarks, we see image compaction take >10
2456 : // minutes, blocking L0 compaction, and we can't stall L0 flushes for that long.
2457 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2458 :
2459 : // If compaction is disabled, don't stall.
2460 2516 : if self.get_compaction_period() == Duration::ZERO {
2461 2516 : return None;
2462 0 : }
2463 0 :
2464 0 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2465 0 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2466 0 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2467 0 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2468 0 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2469 0 : return None;
2470 0 : }
2471 0 :
2472 0 : let compaction_threshold = self.get_compaction_threshold();
2473 0 : let tenant_conf = self.tenant_conf.load();
2474 0 : let l0_flush_stall_threshold = tenant_conf
2475 0 : .tenant_conf
2476 0 : .l0_flush_stall_threshold
2477 0 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2478 0 :
2479 0 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2480 0 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2481 0 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2482 0 : if cfg!(feature = "testing")
2483 0 : && compaction_threshold == 1
2484 0 : && l0_flush_stall_threshold.is_none()
2485 : {
2486 0 : return None;
2487 0 : }
2488 0 :
2489 0 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2490 0 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2491 0 :
2492 0 : // 0 disables backpressure.
2493 0 : if l0_flush_stall_threshold == 0 {
2494 0 : return None;
2495 0 : }
2496 0 :
2497 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2498 0 : // backpressure flushes below this.
2499 0 : // TODO: the tenant config should have validation to prevent this instead.
2500 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2501 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2502 2516 : }
2503 :
2504 2356 : fn get_l0_flush_wait_upload(&self) -> bool {
2505 2356 : let tenant_conf = self.tenant_conf.load();
2506 2356 : tenant_conf
2507 2356 : .tenant_conf
2508 2356 : .l0_flush_wait_upload
2509 2356 : .unwrap_or(self.conf.default_tenant_conf.l0_flush_wait_upload)
2510 2356 : }
2511 :
2512 28 : fn get_image_creation_threshold(&self) -> usize {
2513 28 : let tenant_conf = self.tenant_conf.load();
2514 28 : tenant_conf
2515 28 : .tenant_conf
2516 28 : .image_creation_threshold
2517 28 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2518 28 : }
2519 :
2520 728 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2521 728 : let tenant_conf = &self.tenant_conf.load();
2522 728 : tenant_conf
2523 728 : .tenant_conf
2524 728 : .compaction_algorithm
2525 728 : .as_ref()
2526 728 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2527 728 : .clone()
2528 728 : }
2529 :
2530 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2531 0 : let tenant_conf = self.tenant_conf.load();
2532 0 : tenant_conf
2533 0 : .tenant_conf
2534 0 : .eviction_policy
2535 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2536 0 : }
2537 :
2538 904 : fn get_evictions_low_residence_duration_metric_threshold(
2539 904 : tenant_conf: &TenantConfOpt,
2540 904 : default_tenant_conf: &TenantConf,
2541 904 : ) -> Duration {
2542 904 : tenant_conf
2543 904 : .evictions_low_residence_duration_metric_threshold
2544 904 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2545 904 : }
2546 :
2547 1148 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2548 1148 : let tenant_conf = self.tenant_conf.load();
2549 1148 : tenant_conf
2550 1148 : .tenant_conf
2551 1148 : .image_layer_creation_check_threshold
2552 1148 : .unwrap_or(
2553 1148 : self.conf
2554 1148 : .default_tenant_conf
2555 1148 : .image_layer_creation_check_threshold,
2556 1148 : )
2557 1148 : }
2558 :
2559 0 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
2560 0 : let tenant_conf = &self.tenant_conf.load();
2561 0 : let gc_compaction_enabled = tenant_conf
2562 0 : .tenant_conf
2563 0 : .gc_compaction_enabled
2564 0 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
2565 0 : let gc_compaction_initial_threshold_kb = tenant_conf
2566 0 : .tenant_conf
2567 0 : .gc_compaction_initial_threshold_kb
2568 0 : .unwrap_or(
2569 0 : self.conf
2570 0 : .default_tenant_conf
2571 0 : .gc_compaction_initial_threshold_kb,
2572 0 : );
2573 0 : let gc_compaction_ratio_percent = tenant_conf
2574 0 : .tenant_conf
2575 0 : .gc_compaction_ratio_percent
2576 0 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
2577 0 : GcCompactionCombinedSettings {
2578 0 : gc_compaction_enabled,
2579 0 : gc_compaction_initial_threshold_kb,
2580 0 : gc_compaction_ratio_percent,
2581 0 : }
2582 0 : }
2583 :
2584 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2585 0 : let tenant_conf = self.tenant_conf.load();
2586 0 : tenant_conf
2587 0 : .tenant_conf
2588 0 : .image_creation_preempt_threshold
2589 0 : .unwrap_or(
2590 0 : self.conf
2591 0 : .default_tenant_conf
2592 0 : .image_creation_preempt_threshold,
2593 0 : )
2594 0 : }
2595 :
2596 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2597 : ///
2598 : /// Priority order is:
2599 : /// 1. Tenant config override
2600 : /// 2. Default value for tenant config override
2601 : /// 3. Pageserver config override
2602 : /// 4. Pageserver config default
2603 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2604 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2605 0 : tenant_conf
2606 0 : .wal_receiver_protocol_override
2607 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2608 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2609 0 : }
2610 :
2611 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2612 0 : // NB: Most tenant conf options are read by background loops, so,
2613 0 : // changes will automatically be picked up.
2614 0 :
2615 0 : // The threshold is embedded in the metric. So, we need to update it.
2616 0 : {
2617 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2618 0 : &new_conf.tenant_conf,
2619 0 : &self.conf.default_tenant_conf,
2620 0 : );
2621 0 :
2622 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2623 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2624 0 :
2625 0 : let timeline_id_str = self.timeline_id.to_string();
2626 0 :
2627 0 : self.remote_client.update_config(&new_conf.location);
2628 0 :
2629 0 : self.metrics
2630 0 : .evictions_with_low_residence_duration
2631 0 : .write()
2632 0 : .unwrap()
2633 0 : .change_threshold(
2634 0 : &tenant_id_str,
2635 0 : &shard_id_str,
2636 0 : &timeline_id_str,
2637 0 : new_threshold,
2638 0 : );
2639 0 : }
2640 0 : }
2641 :
2642 : /// Open a Timeline handle.
2643 : ///
2644 : /// Loads the metadata for the timeline into memory, but not the layer map.
2645 : #[allow(clippy::too_many_arguments)]
2646 904 : pub(super) fn new(
2647 904 : conf: &'static PageServerConf,
2648 904 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2649 904 : metadata: &TimelineMetadata,
2650 904 : previous_heatmap: Option<PreviousHeatmap>,
2651 904 : ancestor: Option<Arc<Timeline>>,
2652 904 : timeline_id: TimelineId,
2653 904 : tenant_shard_id: TenantShardId,
2654 904 : generation: Generation,
2655 904 : shard_identity: ShardIdentity,
2656 904 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2657 904 : resources: TimelineResources,
2658 904 : pg_version: u32,
2659 904 : state: TimelineState,
2660 904 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2661 904 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2662 904 : gc_compaction_state: Option<GcCompactionState>,
2663 904 : rel_size_v2_status: Option<RelSizeMigration>,
2664 904 : cancel: CancellationToken,
2665 904 : ) -> Arc<Self> {
2666 904 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2667 904 : let (state, _) = watch::channel(state);
2668 904 :
2669 904 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2670 904 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2671 904 :
2672 904 : let evictions_low_residence_duration_metric_threshold = {
2673 904 : let loaded_tenant_conf = tenant_conf.load();
2674 904 : Self::get_evictions_low_residence_duration_metric_threshold(
2675 904 : &loaded_tenant_conf.tenant_conf,
2676 904 : &conf.default_tenant_conf,
2677 904 : )
2678 : };
2679 :
2680 904 : if let Some(ancestor) = &ancestor {
2681 460 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2682 460 : // If we construct an explicit timeline object, it's obviously not offloaded
2683 460 : let is_offloaded = MaybeOffloaded::No;
2684 460 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2685 460 : }
2686 :
2687 904 : Arc::new_cyclic(|myself| {
2688 904 : let metrics = Arc::new(TimelineMetrics::new(
2689 904 : &tenant_shard_id,
2690 904 : &timeline_id,
2691 904 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2692 904 : "mtime",
2693 904 : evictions_low_residence_duration_metric_threshold,
2694 904 : ),
2695 904 : ));
2696 904 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2697 :
2698 904 : let mut result = Timeline {
2699 904 : conf,
2700 904 : tenant_conf,
2701 904 : myself: myself.clone(),
2702 904 : timeline_id,
2703 904 : tenant_shard_id,
2704 904 : generation,
2705 904 : shard_identity,
2706 904 : pg_version,
2707 904 : layers: Default::default(),
2708 904 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2709 904 :
2710 904 : walredo_mgr,
2711 904 : walreceiver: Mutex::new(None),
2712 904 :
2713 904 : remote_client: Arc::new(resources.remote_client),
2714 904 :
2715 904 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2716 904 : last_record_lsn: SeqWait::new(RecordLsn {
2717 904 : last: disk_consistent_lsn,
2718 904 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2719 904 : }),
2720 904 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2721 904 :
2722 904 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
2723 904 :
2724 904 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2725 904 : last_freeze_ts: RwLock::new(Instant::now()),
2726 904 :
2727 904 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2728 904 :
2729 904 : ancestor_timeline: ancestor,
2730 904 : ancestor_lsn: metadata.ancestor_lsn(),
2731 904 :
2732 904 : metrics,
2733 904 :
2734 904 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2735 904 : &tenant_shard_id,
2736 904 : &timeline_id,
2737 904 : resources.pagestream_throttle_metrics,
2738 904 : ),
2739 904 :
2740 7232 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2741 7232 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
2742 904 :
2743 904 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2744 904 :
2745 904 : layer_flush_start_tx,
2746 904 : layer_flush_done_tx,
2747 904 :
2748 904 : write_lock: tokio::sync::Mutex::new(None),
2749 904 :
2750 904 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2751 904 :
2752 904 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2753 904 : LastImageLayerCreationStatus::default(),
2754 904 : )),
2755 904 :
2756 904 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2757 904 : initdb_lsn: metadata.initdb_lsn(),
2758 904 :
2759 904 : current_logical_size: if disk_consistent_lsn.is_valid() {
2760 : // we're creating timeline data with some layer files existing locally,
2761 : // need to recalculate timeline's logical size based on data in the layers.
2762 468 : LogicalSize::deferred_initial(disk_consistent_lsn)
2763 : } else {
2764 : // we're creating timeline data without any layers existing locally,
2765 : // initial logical size is 0.
2766 436 : LogicalSize::empty_initial()
2767 : },
2768 :
2769 904 : partitioning: GuardArcSwap::new((
2770 904 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2771 904 : Lsn(0),
2772 904 : )),
2773 904 : repartition_threshold: 0,
2774 904 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2775 904 : last_image_layer_creation_check_instant: Mutex::new(None),
2776 904 :
2777 904 : last_received_wal: Mutex::new(None),
2778 904 : rel_size_cache: RwLock::new(RelSizeCache {
2779 904 : complete_as_of: disk_consistent_lsn,
2780 904 : map: HashMap::new(),
2781 904 : }),
2782 904 :
2783 904 : download_all_remote_layers_task_info: RwLock::new(None),
2784 904 :
2785 904 : state,
2786 904 :
2787 904 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2788 904 : EvictionTaskTimelineState::default(),
2789 904 : ),
2790 904 : delete_progress: TimelineDeleteProgress::default(),
2791 904 :
2792 904 : cancel,
2793 904 : gate: Gate::default(),
2794 904 :
2795 904 : compaction_lock: tokio::sync::Mutex::default(),
2796 904 : compaction_failed: AtomicBool::default(),
2797 904 : l0_compaction_trigger: resources.l0_compaction_trigger,
2798 904 : gc_lock: tokio::sync::Mutex::default(),
2799 904 :
2800 904 : standby_horizon: AtomicLsn::new(0),
2801 904 :
2802 904 : pagestream_throttle: resources.pagestream_throttle,
2803 904 :
2804 904 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2805 904 :
2806 904 : #[cfg(test)]
2807 904 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
2808 904 :
2809 904 : l0_flush_global_state: resources.l0_flush_global_state,
2810 904 :
2811 904 : handles: Default::default(),
2812 904 :
2813 904 : attach_wal_lag_cooldown,
2814 904 :
2815 904 : create_idempotency,
2816 904 :
2817 904 : page_trace: Default::default(),
2818 904 :
2819 904 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
2820 904 :
2821 904 : heatmap_layers_downloader: Mutex::new(None),
2822 904 :
2823 904 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
2824 904 : };
2825 904 :
2826 904 : result.repartition_threshold =
2827 904 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
2828 904 :
2829 904 : result
2830 904 : .metrics
2831 904 : .last_record_lsn_gauge
2832 904 : .set(disk_consistent_lsn.0 as i64);
2833 904 : result
2834 904 : })
2835 904 : }
2836 :
2837 1312 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
2838 1312 : let Ok(guard) = self.gate.enter() else {
2839 0 : info!("cannot start flush loop when the timeline gate has already been closed");
2840 0 : return;
2841 : };
2842 1312 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
2843 1312 : match *flush_loop_state {
2844 892 : FlushLoopState::NotStarted => (),
2845 : FlushLoopState::Running { .. } => {
2846 420 : info!(
2847 0 : "skipping attempt to start flush_loop twice {}/{}",
2848 0 : self.tenant_shard_id, self.timeline_id
2849 : );
2850 420 : return;
2851 : }
2852 : FlushLoopState::Exited => {
2853 0 : info!(
2854 0 : "ignoring attempt to restart exited flush_loop {}/{}",
2855 0 : self.tenant_shard_id, self.timeline_id
2856 : );
2857 0 : return;
2858 : }
2859 : }
2860 :
2861 892 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
2862 892 : let self_clone = Arc::clone(self);
2863 892 :
2864 892 : debug!("spawning flush loop");
2865 892 : *flush_loop_state = FlushLoopState::Running {
2866 892 : #[cfg(test)]
2867 892 : expect_initdb_optimization: false,
2868 892 : #[cfg(test)]
2869 892 : initdb_optimization_count: 0,
2870 892 : };
2871 892 : task_mgr::spawn(
2872 892 : task_mgr::BACKGROUND_RUNTIME.handle(),
2873 892 : task_mgr::TaskKind::LayerFlushTask,
2874 892 : self.tenant_shard_id,
2875 892 : Some(self.timeline_id),
2876 892 : "layer flush task",
2877 892 : async move {
2878 892 : let _guard = guard;
2879 892 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
2880 892 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
2881 20 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
2882 20 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
2883 20 : *flush_loop_state = FlushLoopState::Exited;
2884 20 : Ok(())
2885 20 : }
2886 892 : .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))
2887 : );
2888 1312 : }
2889 :
2890 0 : pub(crate) fn update_gc_compaction_state(
2891 0 : &self,
2892 0 : gc_compaction_state: GcCompactionState,
2893 0 : ) -> anyhow::Result<()> {
2894 0 : self.gc_compaction_state
2895 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
2896 0 : self.remote_client
2897 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
2898 0 : }
2899 :
2900 0 : pub(crate) fn update_rel_size_v2_status(
2901 0 : &self,
2902 0 : rel_size_v2_status: RelSizeMigration,
2903 0 : ) -> anyhow::Result<()> {
2904 0 : self.rel_size_v2_status
2905 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
2906 0 : self.remote_client
2907 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
2908 0 : }
2909 :
2910 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
2911 0 : self.gc_compaction_state.load_full().as_ref().clone()
2912 0 : }
2913 :
2914 : /// Creates and starts the wal receiver.
2915 : ///
2916 : /// This function is expected to be called at most once per Timeline's lifecycle
2917 : /// when the timeline is activated.
2918 0 : fn launch_wal_receiver(
2919 0 : self: &Arc<Self>,
2920 0 : ctx: &RequestContext,
2921 0 : broker_client: BrokerClientChannel,
2922 0 : ) {
2923 0 : info!(
2924 0 : "launching WAL receiver for timeline {} of tenant {}",
2925 0 : self.timeline_id, self.tenant_shard_id
2926 : );
2927 :
2928 0 : let tenant_conf = self.tenant_conf.load();
2929 0 : let wal_connect_timeout = tenant_conf
2930 0 : .tenant_conf
2931 0 : .walreceiver_connect_timeout
2932 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
2933 0 : let lagging_wal_timeout = tenant_conf
2934 0 : .tenant_conf
2935 0 : .lagging_wal_timeout
2936 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
2937 0 : let max_lsn_wal_lag = tenant_conf
2938 0 : .tenant_conf
2939 0 : .max_lsn_wal_lag
2940 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
2941 0 :
2942 0 : let mut guard = self.walreceiver.lock().unwrap();
2943 0 : assert!(
2944 0 : guard.is_none(),
2945 0 : "multiple launches / re-launches of WAL receiver are not supported"
2946 : );
2947 0 : *guard = Some(WalReceiver::start(
2948 0 : Arc::clone(self),
2949 0 : WalReceiverConf {
2950 0 : protocol: self.resolve_wal_receiver_protocol(),
2951 0 : wal_connect_timeout,
2952 0 : lagging_wal_timeout,
2953 0 : max_lsn_wal_lag,
2954 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
2955 0 : availability_zone: self.conf.availability_zone.clone(),
2956 0 : ingest_batch_size: self.conf.ingest_batch_size,
2957 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
2958 0 : },
2959 0 : broker_client,
2960 0 : ctx,
2961 0 : ));
2962 0 : }
2963 :
2964 : /// Initialize with an empty layer map. Used when creating a new timeline.
2965 892 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
2966 892 : let mut layers = self.layers.try_write().expect(
2967 892 : "in the context where we call this function, no other task has access to the object",
2968 892 : );
2969 892 : layers
2970 892 : .open_mut()
2971 892 : .expect("in this context the LayerManager must still be open")
2972 892 : .initialize_empty(Lsn(start_lsn.0));
2973 892 : }
2974 :
2975 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
2976 : /// files.
2977 12 : pub(super) async fn load_layer_map(
2978 12 : &self,
2979 12 : disk_consistent_lsn: Lsn,
2980 12 : index_part: IndexPart,
2981 12 : ) -> anyhow::Result<()> {
2982 : use LayerName::*;
2983 : use init::Decision::*;
2984 : use init::{Discovered, DismissedLayer};
2985 :
2986 12 : let mut guard = self.layers.write().await;
2987 :
2988 12 : let timer = self.metrics.load_layer_map_histo.start_timer();
2989 12 :
2990 12 : // Scan timeline directory and create ImageLayerName and DeltaFilename
2991 12 : // structs representing all files on disk
2992 12 : let timeline_path = self
2993 12 : .conf
2994 12 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
2995 12 : let conf = self.conf;
2996 12 : let span = tracing::Span::current();
2997 12 :
2998 12 : // Copy to move into the task we're about to spawn
2999 12 : let this = self.myself.upgrade().expect("&self method holds the arc");
3000 :
3001 12 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3002 12 : move || {
3003 12 : let _g = span.entered();
3004 12 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3005 12 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3006 12 : let mut unrecognized_files = Vec::new();
3007 12 :
3008 12 : let mut path = timeline_path;
3009 :
3010 44 : for discovered in discovered {
3011 32 : let (name, kind) = match discovered {
3012 32 : Discovered::Layer(layer_file_name, local_metadata) => {
3013 32 : discovered_layers.push((layer_file_name, local_metadata));
3014 32 : continue;
3015 : }
3016 0 : Discovered::IgnoredBackup(path) => {
3017 0 : std::fs::remove_file(path)
3018 0 : .or_else(fs_ext::ignore_not_found)
3019 0 : .fatal_err("Removing .old file");
3020 0 : continue;
3021 : }
3022 0 : Discovered::Unknown(file_name) => {
3023 0 : // we will later error if there are any
3024 0 : unrecognized_files.push(file_name);
3025 0 : continue;
3026 : }
3027 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3028 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3029 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3030 : };
3031 0 : path.push(Utf8Path::new(&name));
3032 0 : init::cleanup(&path, kind)?;
3033 0 : path.pop();
3034 : }
3035 :
3036 12 : if !unrecognized_files.is_empty() {
3037 : // assume that if there are any there are many many.
3038 0 : let n = unrecognized_files.len();
3039 0 : let first = &unrecognized_files[..n.min(10)];
3040 0 : anyhow::bail!(
3041 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3042 0 : );
3043 12 : }
3044 12 :
3045 12 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3046 12 :
3047 12 : let mut loaded_layers = Vec::new();
3048 12 : let mut needs_cleanup = Vec::new();
3049 12 : let mut total_physical_size = 0;
3050 :
3051 44 : for (name, decision) in decided {
3052 32 : let decision = match decision {
3053 32 : Ok(decision) => decision,
3054 0 : Err(DismissedLayer::Future { local }) => {
3055 0 : if let Some(local) = local {
3056 0 : init::cleanup_future_layer(
3057 0 : &local.local_path,
3058 0 : &name,
3059 0 : disk_consistent_lsn,
3060 0 : )?;
3061 0 : }
3062 0 : needs_cleanup.push(name);
3063 0 : continue;
3064 : }
3065 0 : Err(DismissedLayer::LocalOnly(local)) => {
3066 0 : init::cleanup_local_only_file(&name, &local)?;
3067 : // this file never existed remotely, we will have to do rework
3068 0 : continue;
3069 : }
3070 0 : Err(DismissedLayer::BadMetadata(local)) => {
3071 0 : init::cleanup_local_file_for_remote(&local)?;
3072 : // this file never existed remotely, we will have to do rework
3073 0 : continue;
3074 : }
3075 : };
3076 :
3077 32 : match &name {
3078 24 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3079 8 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3080 : }
3081 :
3082 32 : tracing::debug!(layer=%name, ?decision, "applied");
3083 :
3084 32 : let layer = match decision {
3085 32 : Resident { local, remote } => {
3086 32 : total_physical_size += local.file_size;
3087 32 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3088 32 : .drop_eviction_guard()
3089 : }
3090 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3091 : };
3092 :
3093 32 : loaded_layers.push(layer);
3094 : }
3095 12 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3096 12 : }
3097 12 : })
3098 12 : .await
3099 12 : .map_err(anyhow::Error::new)
3100 12 : .and_then(|x| x)?;
3101 :
3102 12 : let num_layers = loaded_layers.len();
3103 12 :
3104 12 : guard
3105 12 : .open_mut()
3106 12 : .expect("layermanager must be open during init")
3107 12 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3108 12 :
3109 12 : self.remote_client
3110 12 : .schedule_layer_file_deletion(&needs_cleanup)?;
3111 12 : self.remote_client
3112 12 : .schedule_index_upload_for_file_changes()?;
3113 : // This barrier orders above DELETEs before any later operations.
3114 : // This is critical because code executing after the barrier might
3115 : // create again objects with the same key that we just scheduled for deletion.
3116 : // For example, if we just scheduled deletion of an image layer "from the future",
3117 : // later compaction might run again and re-create the same image layer.
3118 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3119 : // "same" here means same key range and LSN.
3120 : //
3121 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3122 : // the upload queue may execute the PUT first, then the DELETE.
3123 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3124 : //
3125 : // 1. a future image layer is created and uploaded
3126 : // 2. ps restart
3127 : // 3. the future layer from (1) is deleted during load layer map
3128 : // 4. image layer is re-created and uploaded
3129 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3130 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3131 : //
3132 : // See https://github.com/neondatabase/neon/issues/5878
3133 : //
3134 : // NB: generation numbers naturally protect against this because they disambiguate
3135 : // (1) and (4)
3136 : // TODO: this is basically a no-op now, should we remove it?
3137 12 : self.remote_client.schedule_barrier()?;
3138 : // Tenant::create_timeline will wait for these uploads to happen before returning, or
3139 : // on retry.
3140 :
3141 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3142 12 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3143 12 : self.update_layer_visibility().await?;
3144 :
3145 12 : info!(
3146 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3147 : num_layers, disk_consistent_lsn, total_physical_size
3148 : );
3149 :
3150 12 : timer.stop_and_record();
3151 12 : Ok(())
3152 12 : }
3153 :
3154 : /// Retrieve current logical size of the timeline.
3155 : ///
3156 : /// The size could be lagging behind the actual number, in case
3157 : /// the initial size calculation has not been run (gets triggered on the first size access).
3158 : ///
3159 : /// return size and boolean flag that shows if the size is exact
3160 0 : pub(crate) fn get_current_logical_size(
3161 0 : self: &Arc<Self>,
3162 0 : priority: GetLogicalSizePriority,
3163 0 : ctx: &RequestContext,
3164 0 : ) -> logical_size::CurrentLogicalSize {
3165 0 : if !self.tenant_shard_id.is_shard_zero() {
3166 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3167 : // when HTTP API is serving a GET for timeline zero, return zero
3168 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3169 0 : }
3170 0 :
3171 0 : let current_size = self.current_logical_size.current_size();
3172 0 : debug!("Current size: {current_size:?}");
3173 :
3174 0 : match (current_size.accuracy(), priority) {
3175 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3176 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3177 0 : // background task will eventually deliver an exact value, we're in no rush
3178 0 : }
3179 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3180 : // background task is not ready, but user is asking for it now;
3181 : // => make the background task skip the line
3182 : // (The alternative would be to calculate the size here, but,
3183 : // it can actually take a long time if the user has a lot of rels.
3184 : // And we'll inevitable need it again; So, let the background task do the work.)
3185 0 : match self
3186 0 : .current_logical_size
3187 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3188 0 : .get()
3189 : {
3190 0 : Some(cancel) => cancel.cancel(),
3191 : None => {
3192 0 : match self.current_state() {
3193 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3194 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3195 0 : // Don't make noise.
3196 0 : }
3197 : TimelineState::Loading => {
3198 : // Import does not return an activated timeline.
3199 0 : info!(
3200 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3201 : );
3202 : }
3203 : TimelineState::Active => {
3204 : // activation should be setting the once cell
3205 0 : warn!(
3206 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3207 : );
3208 0 : debug_assert!(false);
3209 : }
3210 : }
3211 : }
3212 : }
3213 : }
3214 : }
3215 :
3216 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3217 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3218 0 : let first = self
3219 0 : .current_logical_size
3220 0 : .did_return_approximate_to_walreceiver
3221 0 : .compare_exchange(
3222 0 : false,
3223 0 : true,
3224 0 : AtomicOrdering::Relaxed,
3225 0 : AtomicOrdering::Relaxed,
3226 0 : )
3227 0 : .is_ok();
3228 0 : if first {
3229 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3230 0 : }
3231 0 : }
3232 0 : }
3233 :
3234 0 : current_size
3235 0 : }
3236 :
3237 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3238 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3239 : // nothing to do for freshly created timelines;
3240 0 : assert_eq!(
3241 0 : self.current_logical_size.current_size().accuracy(),
3242 0 : logical_size::Accuracy::Exact,
3243 0 : );
3244 0 : self.current_logical_size.initialized.add_permits(1);
3245 0 : return;
3246 : };
3247 :
3248 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3249 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3250 0 : self.current_logical_size
3251 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3252 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3253 0 :
3254 0 : let self_clone = Arc::clone(self);
3255 0 : let background_ctx = ctx.detached_child(
3256 0 : TaskKind::InitialLogicalSizeCalculation,
3257 0 : DownloadBehavior::Download,
3258 0 : );
3259 0 : task_mgr::spawn(
3260 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3261 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3262 0 : self.tenant_shard_id,
3263 0 : Some(self.timeline_id),
3264 0 : "initial size calculation",
3265 : // NB: don't log errors here, task_mgr will do that.
3266 0 : async move {
3267 0 : self_clone
3268 0 : .initial_logical_size_calculation_task(
3269 0 : initial_part_end,
3270 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3271 0 : background_ctx,
3272 0 : )
3273 0 : .await;
3274 0 : Ok(())
3275 0 : }
3276 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)),
3277 : );
3278 0 : }
3279 :
3280 : /// # Cancellation
3281 : ///
3282 : /// This method is sensitive to `Timeline::cancel`.
3283 : ///
3284 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3285 : ///
3286 : /// # Cancel-Safety
3287 : ///
3288 : /// It does Timeline IO, hence this should be polled to completion because
3289 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3290 : /// to reason about.
3291 0 : async fn initial_logical_size_calculation_task(
3292 0 : self: Arc<Self>,
3293 0 : initial_part_end: Lsn,
3294 0 : skip_concurrency_limiter: CancellationToken,
3295 0 : background_ctx: RequestContext,
3296 0 : ) {
3297 0 : scopeguard::defer! {
3298 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3299 0 : self.current_logical_size.initialized.add_permits(1);
3300 0 : }
3301 0 :
3302 0 : let try_once = |attempt: usize| {
3303 0 : let background_ctx = &background_ctx;
3304 0 : let self_ref = &self;
3305 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3306 0 : async move {
3307 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3308 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3309 0 : background_ctx,
3310 0 : );
3311 :
3312 : use crate::metrics::initial_logical_size::StartCircumstances;
3313 0 : let (_maybe_permit, circumstances) = tokio::select! {
3314 0 : permit = wait_for_permit => {
3315 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3316 : }
3317 0 : _ = self_ref.cancel.cancelled() => {
3318 0 : return Err(CalculateLogicalSizeError::Cancelled);
3319 : }
3320 0 : () = skip_concurrency_limiter.cancelled() => {
3321 : // Some action that is part of a end user interaction requested logical size
3322 : // => break out of the rate limit
3323 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3324 : // but then again what happens if they cancel; also, we should just be using
3325 : // one runtime across the entire process, so, let's leave this for now.
3326 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3327 : }
3328 : };
3329 :
3330 0 : let metrics_guard = if attempt == 1 {
3331 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3332 : } else {
3333 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3334 : };
3335 :
3336 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3337 0 : self_ref.conf,
3338 0 : self_ref
3339 0 : .gate
3340 0 : .enter()
3341 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3342 : );
3343 :
3344 0 : let calculated_size = self_ref
3345 0 : .logical_size_calculation_task(
3346 0 : initial_part_end,
3347 0 : LogicalSizeCalculationCause::Initial,
3348 0 : background_ctx,
3349 0 : )
3350 0 : .await?;
3351 :
3352 0 : self_ref
3353 0 : .trigger_aux_file_size_computation(
3354 0 : initial_part_end,
3355 0 : background_ctx,
3356 0 : io_concurrency,
3357 0 : )
3358 0 : .await?;
3359 :
3360 : // TODO: add aux file size to logical size
3361 :
3362 0 : Ok((calculated_size, metrics_guard))
3363 0 : }
3364 0 : };
3365 :
3366 0 : let retrying = async {
3367 0 : let mut attempt = 0;
3368 : loop {
3369 0 : attempt += 1;
3370 0 :
3371 0 : match try_once(attempt).await {
3372 0 : Ok(res) => return ControlFlow::Continue(res),
3373 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3374 : Err(
3375 0 : e @ (CalculateLogicalSizeError::Decode(_)
3376 0 : | CalculateLogicalSizeError::PageRead(_)),
3377 0 : ) => {
3378 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3379 : // exponential back-off doesn't make sense at these long intervals;
3380 : // use fixed retry interval with generous jitter instead
3381 0 : let sleep_duration = Duration::from_secs(
3382 0 : u64::try_from(
3383 0 : // 1hour base
3384 0 : (60_i64 * 60_i64)
3385 0 : // 10min jitter
3386 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3387 0 : )
3388 0 : .expect("10min < 1hour"),
3389 0 : );
3390 0 : tokio::select! {
3391 0 : _ = tokio::time::sleep(sleep_duration) => {}
3392 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3393 : }
3394 : }
3395 : }
3396 : }
3397 0 : };
3398 :
3399 0 : let (calculated_size, metrics_guard) = match retrying.await {
3400 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3401 0 : ControlFlow::Break(()) => return,
3402 : };
3403 :
3404 : // we cannot query current_logical_size.current_size() to know the current
3405 : // *negative* value, only truncated to u64.
3406 0 : let added = self
3407 0 : .current_logical_size
3408 0 : .size_added_after_initial
3409 0 : .load(AtomicOrdering::Relaxed);
3410 0 :
3411 0 : let sum = calculated_size.saturating_add_signed(added);
3412 0 :
3413 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3414 0 : self.metrics.current_logical_size_gauge.set(sum);
3415 0 :
3416 0 : self.current_logical_size
3417 0 : .initial_logical_size
3418 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3419 0 : .ok()
3420 0 : .expect("only this task sets it");
3421 0 : }
3422 :
3423 0 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3424 0 : self: &Arc<Self>,
3425 0 : lsn: Lsn,
3426 0 : cause: LogicalSizeCalculationCause,
3427 0 : ctx: RequestContext,
3428 0 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3429 0 : let (sender, receiver) = oneshot::channel();
3430 0 : let self_clone = Arc::clone(self);
3431 0 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3432 0 : // we should stop the size calculation work and return an error.
3433 0 : // That would require restructuring this function's API to
3434 0 : // return the result directly, instead of a Receiver for the result.
3435 0 : let ctx = ctx.detached_child(
3436 0 : TaskKind::OndemandLogicalSizeCalculation,
3437 0 : DownloadBehavior::Download,
3438 0 : );
3439 0 : task_mgr::spawn(
3440 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3441 0 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3442 0 : self.tenant_shard_id,
3443 0 : Some(self.timeline_id),
3444 0 : "ondemand logical size calculation",
3445 0 : async move {
3446 0 : let res = self_clone
3447 0 : .logical_size_calculation_task(lsn, cause, &ctx)
3448 0 : .await;
3449 0 : let _ = sender.send(res).ok();
3450 0 : Ok(()) // Receiver is responsible for handling errors
3451 0 : }
3452 0 : .in_current_span(),
3453 0 : );
3454 0 : receiver
3455 0 : }
3456 :
3457 : #[instrument(skip_all)]
3458 : async fn logical_size_calculation_task(
3459 : self: &Arc<Self>,
3460 : lsn: Lsn,
3461 : cause: LogicalSizeCalculationCause,
3462 : ctx: &RequestContext,
3463 : ) -> Result<u64, CalculateLogicalSizeError> {
3464 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3465 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3466 : // accurate relation sizes, and they do not emit consumption metrics.
3467 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3468 :
3469 : let guard = self
3470 : .gate
3471 : .enter()
3472 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3473 :
3474 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3475 : }
3476 :
3477 : /// Calculate the logical size of the database at the latest LSN.
3478 : ///
3479 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3480 : /// especially if we need to download remote layers.
3481 0 : async fn calculate_logical_size(
3482 0 : &self,
3483 0 : up_to_lsn: Lsn,
3484 0 : cause: LogicalSizeCalculationCause,
3485 0 : _guard: &GateGuard,
3486 0 : ctx: &RequestContext,
3487 0 : ) -> Result<u64, CalculateLogicalSizeError> {
3488 0 : info!(
3489 0 : "Calculating logical size for timeline {} at {}",
3490 : self.timeline_id, up_to_lsn
3491 : );
3492 :
3493 0 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3494 : {
3495 0 : return Err(CalculateLogicalSizeError::Cancelled);
3496 0 : }
3497 :
3498 : // See if we've already done the work for initial size calculation.
3499 : // This is a short-cut for timelines that are mostly unused.
3500 0 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3501 0 : return Ok(size);
3502 0 : }
3503 0 : let storage_time_metrics = match cause {
3504 : LogicalSizeCalculationCause::Initial
3505 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3506 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3507 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3508 0 : &self.metrics.imitate_logical_size_histo
3509 : }
3510 : };
3511 0 : let timer = storage_time_metrics.start_timer();
3512 0 : let logical_size = self
3513 0 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3514 0 : .await?;
3515 0 : debug!("calculated logical size: {logical_size}");
3516 0 : timer.stop_and_record();
3517 0 : Ok(logical_size)
3518 0 : }
3519 :
3520 : /// Update current logical size, adding `delta' to the old value.
3521 541140 : fn update_current_logical_size(&self, delta: i64) {
3522 541140 : let logical_size = &self.current_logical_size;
3523 541140 : logical_size.increment_size(delta);
3524 541140 :
3525 541140 : // Also set the value in the prometheus gauge. Note that
3526 541140 : // there is a race condition here: if this is is called by two
3527 541140 : // threads concurrently, the prometheus gauge might be set to
3528 541140 : // one value while current_logical_size is set to the
3529 541140 : // other.
3530 541140 : match logical_size.current_size() {
3531 541140 : CurrentLogicalSize::Exact(ref new_current_size) => self
3532 541140 : .metrics
3533 541140 : .current_logical_size_gauge
3534 541140 : .set(new_current_size.into()),
3535 0 : CurrentLogicalSize::Approximate(_) => {
3536 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3537 0 : // forth between the initial size calculation task.
3538 0 : }
3539 : }
3540 541140 : }
3541 :
3542 5988 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3543 5988 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3544 5988 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3545 5988 : // would overwrite the previous value if they modify different databases.
3546 5988 :
3547 5988 : match count {
3548 2144 : MetricsUpdate::Set(count) => {
3549 2144 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3550 2144 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3551 2144 : }
3552 3840 : MetricsUpdate::Add(count) => {
3553 3840 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3554 3840 : // it's fine.
3555 3840 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3556 3840 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3557 3840 : // the value reliably.
3558 3840 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3559 3840 : }
3560 : // Otherwise, ignore this update
3561 : }
3562 4 : MetricsUpdate::Sub(count) => {
3563 4 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3564 4 : // it's fine.
3565 4 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3566 4 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3567 4 : // The operation could overflow so we need to normalize the value.
3568 4 : let prev_val =
3569 4 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3570 4 : let res = prev_val.saturating_sub(count);
3571 4 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3572 4 : }
3573 : // Otherwise, ignore this update
3574 : }
3575 : };
3576 :
3577 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3578 5988 : let aux_metric =
3579 5988 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3580 5988 :
3581 5988 : let sum_of_entries = self
3582 5988 : .directory_metrics
3583 5988 : .iter()
3584 47904 : .map(|v| v.load(AtomicOrdering::Relaxed))
3585 5988 : .sum();
3586 : // Set a high general threshold and a lower threshold for the auxiliary files,
3587 : // as we can have large numbers of relations in the db directory.
3588 : const SUM_THRESHOLD: u64 = 5000;
3589 : const AUX_THRESHOLD: u64 = 1000;
3590 5988 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3591 0 : self.metrics
3592 0 : .directory_entries_count_gauge
3593 0 : .set(sum_of_entries);
3594 5988 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3595 0 : metric.set(sum_of_entries);
3596 5988 : }
3597 5988 : }
3598 :
3599 0 : async fn find_layer(
3600 0 : &self,
3601 0 : layer_name: &LayerName,
3602 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3603 0 : let guard = self.layers.read().await;
3604 0 : let layer = guard
3605 0 : .layer_map()?
3606 0 : .iter_historic_layers()
3607 0 : .find(|l| &l.layer_name() == layer_name)
3608 0 : .map(|found| guard.get_from_desc(&found));
3609 0 : Ok(layer)
3610 0 : }
3611 :
3612 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
3613 0 : let crnt = self.previous_heatmap.load();
3614 0 : match crnt.as_deref() {
3615 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
3616 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
3617 0 : None => true,
3618 : },
3619 0 : Some(PreviousHeatmap::Obsolete) => false,
3620 0 : None => false,
3621 : }
3622 0 : }
3623 :
3624 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3625 : /// indicating which layers are currently on-disk on the primary.
3626 : ///
3627 : /// None is returned if the Timeline is in a state where uploading a heatmap
3628 : /// doesn't make sense, such as shutting down or initializing. The caller
3629 : /// should treat this as a cue to simply skip doing any heatmap uploading
3630 : /// for this timeline.
3631 32 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3632 32 : if !self.is_active() {
3633 0 : return None;
3634 32 : }
3635 :
3636 32 : let guard = self.layers.read().await;
3637 :
3638 : // Firstly, if there's any heatmap left over from when this location
3639 : // was a secondary, take that into account. Keep layers that are:
3640 : // * present in the layer map
3641 : // * visible
3642 : // * non-resident
3643 : // * not evicted since we read the heatmap
3644 : //
3645 : // Without this, a new cold, attached location would clobber the previous
3646 : // heatamp.
3647 32 : let previous_heatmap = self.previous_heatmap.load();
3648 32 : let visible_non_resident = match previous_heatmap.as_deref() {
3649 : Some(PreviousHeatmap::Active {
3650 24 : heatmap, read_at, ..
3651 92 : }) => Some(heatmap.all_layers().filter_map(|hl| {
3652 92 : let desc: PersistentLayerDesc = hl.name.clone().into();
3653 92 : let layer = guard.try_get_from_key(&desc.key())?;
3654 :
3655 92 : if layer.visibility() == LayerVisibilityHint::Covered {
3656 0 : return None;
3657 92 : }
3658 92 :
3659 92 : if layer.is_likely_resident() {
3660 40 : return None;
3661 52 : }
3662 52 :
3663 52 : if layer.last_evicted_at().happened_after(*read_at) {
3664 12 : return None;
3665 40 : }
3666 40 :
3667 40 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
3668 92 : })),
3669 0 : Some(PreviousHeatmap::Obsolete) => None,
3670 8 : None => None,
3671 : };
3672 :
3673 : // Secondly, all currently visible, resident layers are included.
3674 72 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3675 72 : match layer.visibility() {
3676 : LayerVisibilityHint::Visible => {
3677 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3678 68 : let last_activity_ts = layer.latest_activity();
3679 68 : Some((
3680 68 : layer.layer_desc().clone(),
3681 68 : layer.metadata(),
3682 68 : last_activity_ts,
3683 68 : false, // these layers are not cold
3684 68 : ))
3685 : }
3686 : LayerVisibilityHint::Covered => {
3687 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3688 4 : None
3689 : }
3690 : }
3691 72 : });
3692 :
3693 32 : let mut layers = match visible_non_resident {
3694 24 : Some(non_resident) => {
3695 24 : let mut non_resident = non_resident.peekable();
3696 24 : if non_resident.peek().is_none() {
3697 8 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
3698 8 : self.previous_heatmap
3699 8 : .store(Some(PreviousHeatmap::Obsolete.into()));
3700 16 : }
3701 :
3702 24 : non_resident.chain(resident).collect::<Vec<_>>()
3703 : }
3704 8 : None => resident.collect::<Vec<_>>(),
3705 : };
3706 :
3707 : // Sort layers in order of which to download first. For a large set of layers to download, we
3708 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3709 : // or hours later:
3710 : // - Cold layers go last for convenience when a human inspects the heatmap.
3711 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3712 : // only exist for a few minutes before being compacted into L1s.
3713 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3714 : // the layer is likely to be covered by an image layer during compaction.
3715 240 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
3716 240 : std::cmp::Reverse((
3717 240 : *cold,
3718 240 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3719 240 : desc.lsn_range.end,
3720 240 : ))
3721 240 : });
3722 32 :
3723 32 : let layers = layers
3724 32 : .into_iter()
3725 108 : .map(|(desc, meta, atime, cold)| {
3726 108 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
3727 108 : })
3728 32 : .collect();
3729 32 :
3730 32 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3731 32 : }
3732 :
3733 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
3734 0 : let guard = self.layers.read().await;
3735 :
3736 0 : let now = SystemTime::now();
3737 0 : let mut heatmap_layers = Vec::default();
3738 0 : for vl in guard.visible_layers() {
3739 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
3740 0 : continue;
3741 0 : }
3742 0 :
3743 0 : let hl = HeatMapLayer {
3744 0 : name: vl.layer_desc().layer_name(),
3745 0 : metadata: vl.metadata(),
3746 0 : access_time: now,
3747 0 : cold: true,
3748 0 : };
3749 0 : heatmap_layers.push(hl);
3750 : }
3751 :
3752 0 : tracing::info!(
3753 0 : "Generating unarchival heatmap with {} layers",
3754 0 : heatmap_layers.len()
3755 : );
3756 :
3757 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
3758 0 : PreviousHeatmap::Active {
3759 0 : heatmap,
3760 0 : read_at: Instant::now(),
3761 0 : end_lsn: Some(end_lsn),
3762 0 : }
3763 0 : }
3764 :
3765 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3766 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3767 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3768 0 : // branchpoint in the value in IndexPart::lineage
3769 0 : self.ancestor_lsn == lsn
3770 0 : || (self.ancestor_lsn == Lsn::INVALID
3771 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3772 0 : }
3773 : }
3774 :
3775 : impl Timeline {
3776 : #[allow(clippy::doc_lazy_continuation)]
3777 : /// Get the data needed to reconstruct all keys in the provided keyspace
3778 : ///
3779 : /// The algorithm is as follows:
3780 : /// 1. While some keys are still not done and there's a timeline to visit:
3781 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
3782 : /// 2.1: Build the fringe for the current keyspace
3783 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
3784 : /// intersects
3785 : /// 2.3. Pop the timeline from the fringe
3786 : /// 2.4. If the fringe is empty, go back to 1
3787 1255306 : async fn get_vectored_reconstruct_data(
3788 1255306 : &self,
3789 1255306 : mut keyspace: KeySpace,
3790 1255306 : request_lsn: Lsn,
3791 1255306 : reconstruct_state: &mut ValuesReconstructState,
3792 1255306 : ctx: &RequestContext,
3793 1255306 : ) -> Result<(), GetVectoredError> {
3794 1255306 : let mut timeline_owned: Arc<Timeline>;
3795 1255306 : let mut timeline = self;
3796 1255306 :
3797 1255306 : let mut cont_lsn = Lsn(request_lsn.0 + 1);
3798 :
3799 1255302 : let missing_keyspace = loop {
3800 1703635 : if self.cancel.is_cancelled() {
3801 0 : return Err(GetVectoredError::Cancelled);
3802 1703635 : }
3803 :
3804 : let TimelineVisitOutcome {
3805 1703635 : completed_keyspace: completed,
3806 1703635 : image_covered_keyspace,
3807 1703635 : } = Self::get_vectored_reconstruct_data_timeline(
3808 1703635 : timeline,
3809 1703635 : keyspace.clone(),
3810 1703635 : cont_lsn,
3811 1703635 : reconstruct_state,
3812 1703635 : &self.cancel,
3813 1703635 : ctx,
3814 1703635 : )
3815 1703635 : .await?;
3816 :
3817 1703635 : keyspace.remove_overlapping_with(&completed);
3818 1703635 :
3819 1703635 : // Do not descend into the ancestor timeline for aux files.
3820 1703635 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
3821 1703635 : // stalling compaction.
3822 1703635 : keyspace.remove_overlapping_with(&KeySpace {
3823 1703635 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
3824 1703635 : });
3825 1703635 :
3826 1703635 : // Keyspace is fully retrieved
3827 1703635 : if keyspace.is_empty() {
3828 1254794 : break None;
3829 448841 : }
3830 :
3831 448841 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
3832 : // Not fully retrieved but no ancestor timeline.
3833 508 : break Some(keyspace);
3834 : };
3835 :
3836 : // Now we see if there are keys covered by the image layer but does not exist in the
3837 : // image layer, which means that the key does not exist.
3838 :
3839 : // The block below will stop the vectored search if any of the keys encountered an image layer
3840 : // which did not contain a snapshot for said key. Since we have already removed all completed
3841 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
3842 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
3843 : // and stop the search as a result of that.
3844 448333 : let mut removed = keyspace.remove_overlapping_with(&image_covered_keyspace);
3845 448333 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
3846 448333 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
3847 448333 : // figuring out what is the inherited key range and do a fine-grained pruning.
3848 448333 : removed.remove_overlapping_with(&KeySpace {
3849 448333 : ranges: vec![SPARSE_RANGE],
3850 448333 : });
3851 448333 : if !removed.is_empty() {
3852 0 : break Some(removed);
3853 448333 : }
3854 448333 : // If we reached this point, `remove_overlapping_with` should not have made any change to the
3855 448333 : // keyspace.
3856 448333 :
3857 448333 : // Take the min to avoid reconstructing a page with data newer than request Lsn.
3858 448333 : cont_lsn = std::cmp::min(Lsn(request_lsn.0 + 1), Lsn(timeline.ancestor_lsn.0 + 1));
3859 448333 : timeline_owned = timeline
3860 448333 : .get_ready_ancestor_timeline(ancestor_timeline, ctx)
3861 448333 : .await?;
3862 448329 : timeline = &*timeline_owned;
3863 : };
3864 :
3865 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
3866 1255302 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
3867 508 : let mut missing_keyspace = missing_keyspace;
3868 508 : missing_keyspace.remove_overlapping_with(&KeySpace {
3869 508 : ranges: vec![SPARSE_RANGE],
3870 508 : });
3871 508 : if missing_keyspace.is_empty() {
3872 480 : None
3873 : } else {
3874 28 : Some(missing_keyspace)
3875 : }
3876 : } else {
3877 1254794 : None
3878 : };
3879 :
3880 1255302 : if let Some(missing_keyspace) = missing_keyspace {
3881 28 : return Err(GetVectoredError::MissingKey(MissingKeyError {
3882 28 : key: missing_keyspace.start().unwrap(), /* better if we can store the full keyspace */
3883 28 : shard: self
3884 28 : .shard_identity
3885 28 : .get_shard_number(&missing_keyspace.start().unwrap()),
3886 28 : cont_lsn,
3887 28 : request_lsn,
3888 28 : ancestor_lsn: Some(timeline.ancestor_lsn),
3889 28 : backtrace: None,
3890 28 : read_path: std::mem::take(&mut reconstruct_state.read_path),
3891 28 : }));
3892 1255274 : }
3893 1255274 :
3894 1255274 : Ok(())
3895 1255306 : }
3896 :
3897 : /// Collect the reconstruct data for a keyspace from the specified timeline.
3898 : ///
3899 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
3900 : /// the current keyspace. The current keyspace of the search at any given timeline
3901 : /// is the original keyspace minus all the keys that have been completed minus
3902 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
3903 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
3904 : ///
3905 : /// This is basically a depth-first search visitor implementation where a vertex
3906 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
3907 : ///
3908 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
3909 : /// and get all the required reconstruct data from the layer in one go.
3910 : ///
3911 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
3912 : /// decides how to deal with these two keyspaces.
3913 1703635 : async fn get_vectored_reconstruct_data_timeline(
3914 1703635 : timeline: &Timeline,
3915 1703635 : keyspace: KeySpace,
3916 1703635 : mut cont_lsn: Lsn,
3917 1703635 : reconstruct_state: &mut ValuesReconstructState,
3918 1703635 : cancel: &CancellationToken,
3919 1703635 : ctx: &RequestContext,
3920 1703635 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
3921 1703635 : let mut unmapped_keyspace = keyspace.clone();
3922 1703635 : let mut fringe = LayerFringe::new();
3923 1703635 :
3924 1703635 : let mut completed_keyspace = KeySpace::default();
3925 1703635 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
3926 1703635 :
3927 1703635 : // Prevent GC from progressing while visiting the current timeline.
3928 1703635 : // If we are GC-ing because a new image layer was added while traversing
3929 1703635 : // the timeline, then it will remove layers that are required for fulfilling
3930 1703635 : // the current get request (read-path cannot "look back" and notice the new
3931 1703635 : // image layer).
3932 1703635 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
3933 :
3934 : // See `compaction::compact_with_gc` for why we need this.
3935 1703635 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
3936 :
3937 : loop {
3938 3397353 : if cancel.is_cancelled() {
3939 0 : return Err(GetVectoredError::Cancelled);
3940 3397353 : }
3941 3397353 :
3942 3397353 : let (keys_done_last_step, keys_with_image_coverage) =
3943 3397353 : reconstruct_state.consume_done_keys();
3944 3397353 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
3945 3397353 : completed_keyspace.merge(&keys_done_last_step);
3946 3397353 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
3947 45100 : unmapped_keyspace
3948 45100 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
3949 45100 : image_covered_keyspace.add_range(keys_with_image_coverage);
3950 3352253 : }
3951 :
3952 : // Do not descent any further if the last layer we visited
3953 : // completed all keys in the keyspace it inspected. This is not
3954 : // required for correctness, but avoids visiting extra layers
3955 : // which turns out to be a perf bottleneck in some cases.
3956 3397353 : if !unmapped_keyspace.is_empty() {
3957 2146599 : let guard = timeline.layers.read().await;
3958 2146599 : let layers = guard.layer_map()?;
3959 :
3960 2147995 : for range in unmapped_keyspace.ranges.iter() {
3961 2147995 : let results = layers.range_search(range.clone(), cont_lsn);
3962 2147995 :
3963 2147995 : results
3964 2147995 : .found
3965 2147995 : .into_iter()
3966 2147995 : .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
3967 1694654 : (
3968 1694654 : guard.upgrade(layer),
3969 1694654 : keyspace_accum.to_keyspace(),
3970 1694654 : lsn_floor..cont_lsn,
3971 1694654 : )
3972 2147995 : })
3973 2147995 : .for_each(|(layer, keyspace, lsn_range)| {
3974 1694654 : fringe.update(layer, keyspace, lsn_range)
3975 2147995 : });
3976 2147995 : }
3977 :
3978 : // It's safe to drop the layer map lock after planning the next round of reads.
3979 : // The fringe keeps readable handles for the layers which are safe to read even
3980 : // if layers were compacted or flushed.
3981 : //
3982 : // The more interesting consideration is: "Why is the read algorithm still correct
3983 : // if the layer map changes while it is operating?". Doing a vectored read on a
3984 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
3985 : // covered by the read. The layer map tells us how to move the lsn downwards for a
3986 : // range at *a particular point in time*. It is fine for the answer to be different
3987 : // at two different time points.
3988 2146599 : drop(guard);
3989 1250754 : }
3990 :
3991 3397353 : if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
3992 1693718 : if let Some(ref mut read_path) = reconstruct_state.read_path {
3993 1693718 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
3994 1693718 : }
3995 1693718 : let next_cont_lsn = lsn_range.start;
3996 1693718 : layer_to_read
3997 1693718 : .get_values_reconstruct_data(
3998 1693718 : keyspace_to_read.clone(),
3999 1693718 : lsn_range,
4000 1693718 : reconstruct_state,
4001 1693718 : ctx,
4002 1693718 : )
4003 1693718 : .await?;
4004 :
4005 1693718 : unmapped_keyspace = keyspace_to_read;
4006 1693718 : cont_lsn = next_cont_lsn;
4007 1693718 :
4008 1693718 : reconstruct_state.on_layer_visited(&layer_to_read);
4009 : } else {
4010 1703635 : break;
4011 1703635 : }
4012 1703635 : }
4013 1703635 :
4014 1703635 : Ok(TimelineVisitOutcome {
4015 1703635 : completed_keyspace,
4016 1703635 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4017 1703635 : })
4018 1703635 : }
4019 :
4020 448333 : async fn get_ready_ancestor_timeline(
4021 448333 : &self,
4022 448333 : ancestor: &Arc<Timeline>,
4023 448333 : ctx: &RequestContext,
4024 448333 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4025 448333 : // It's possible that the ancestor timeline isn't active yet, or
4026 448333 : // is active but hasn't yet caught up to the branch point. Wait
4027 448333 : // for it.
4028 448333 : //
4029 448333 : // This cannot happen while the pageserver is running normally,
4030 448333 : // because you cannot create a branch from a point that isn't
4031 448333 : // present in the pageserver yet. However, we don't wait for the
4032 448333 : // branch point to be uploaded to cloud storage before creating
4033 448333 : // a branch. I.e., the branch LSN need not be remote consistent
4034 448333 : // for the branching operation to succeed.
4035 448333 : //
4036 448333 : // Hence, if we try to load a tenant in such a state where
4037 448333 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4038 448333 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4039 448333 : // then we will need to wait for the ancestor timeline to
4040 448333 : // re-stream WAL up to branch_lsn before we access it.
4041 448333 : //
4042 448333 : // How can a tenant get in such a state?
4043 448333 : // - ungraceful pageserver process exit
4044 448333 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4045 448333 : //
4046 448333 : // NB: this could be avoided by requiring
4047 448333 : // branch_lsn >= remote_consistent_lsn
4048 448333 : // during branch creation.
4049 448333 : match ancestor.wait_to_become_active(ctx).await {
4050 448329 : Ok(()) => {}
4051 : Err(TimelineState::Stopping) => {
4052 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4053 0 : return Err(GetReadyAncestorError::Cancelled);
4054 : }
4055 4 : Err(state) => {
4056 4 : return Err(GetReadyAncestorError::BadState {
4057 4 : timeline_id: ancestor.timeline_id,
4058 4 : state,
4059 4 : });
4060 : }
4061 : }
4062 448329 : ancestor
4063 448329 : .wait_lsn(
4064 448329 : self.ancestor_lsn,
4065 448329 : WaitLsnWaiter::Timeline(self),
4066 448329 : WaitLsnTimeout::Default,
4067 448329 : ctx,
4068 448329 : )
4069 448329 : .await
4070 448329 : .map_err(|e| match e {
4071 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4072 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4073 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4074 0 : timeline_id: ancestor.timeline_id,
4075 0 : state,
4076 0 : },
4077 448329 : })?;
4078 :
4079 448329 : Ok(ancestor.clone())
4080 448333 : }
4081 :
4082 594312 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4083 594312 : &self.shard_identity
4084 594312 : }
4085 :
4086 : #[inline(always)]
4087 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4088 0 : ShardTimelineId {
4089 0 : shard_index: ShardIndex {
4090 0 : shard_number: self.shard_identity.number,
4091 0 : shard_count: self.shard_identity.count,
4092 0 : },
4093 0 : timeline_id: self.timeline_id,
4094 0 : }
4095 0 : }
4096 :
4097 : /// Returns a non-frozen open in-memory layer for ingestion.
4098 : ///
4099 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4100 : /// this function without holding the mutex.
4101 2604 : async fn get_layer_for_write(
4102 2604 : &self,
4103 2604 : lsn: Lsn,
4104 2604 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4105 2604 : ctx: &RequestContext,
4106 2604 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4107 2604 : let mut guard = self.layers.write().await;
4108 :
4109 2604 : let last_record_lsn = self.get_last_record_lsn();
4110 2604 : ensure!(
4111 2604 : lsn > last_record_lsn,
4112 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4113 : lsn,
4114 : last_record_lsn,
4115 : );
4116 :
4117 2604 : let layer = guard
4118 2604 : .open_mut()?
4119 2604 : .get_layer_for_write(
4120 2604 : lsn,
4121 2604 : self.conf,
4122 2604 : self.timeline_id,
4123 2604 : self.tenant_shard_id,
4124 2604 : &self.gate,
4125 2604 : ctx,
4126 2604 : )
4127 2604 : .await?;
4128 2604 : Ok(layer)
4129 2604 : }
4130 :
4131 10558196 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4132 10558196 : assert!(new_lsn.is_aligned());
4133 :
4134 10558196 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4135 10558196 : self.last_record_lsn.advance(new_lsn);
4136 10558196 : }
4137 :
4138 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4139 : ///
4140 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4141 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4142 2412 : async fn freeze_inmem_layer_at(
4143 2412 : &self,
4144 2412 : at: Lsn,
4145 2412 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4146 2412 : ) -> Result<u64, FlushLayerError> {
4147 2412 : let frozen = {
4148 2412 : let mut guard = self.layers.write().await;
4149 2412 : guard
4150 2412 : .open_mut()?
4151 2412 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4152 2412 : .await
4153 : };
4154 :
4155 2412 : if frozen {
4156 2356 : let now = Instant::now();
4157 2356 : *(self.last_freeze_ts.write().unwrap()) = now;
4158 2356 : }
4159 :
4160 : // Increment the flush cycle counter and wake up the flush task.
4161 : // Remember the new value, so that when we listen for the flush
4162 : // to finish, we know when the flush that we initiated has
4163 : // finished, instead of some other flush that was started earlier.
4164 2412 : let mut my_flush_request = 0;
4165 2412 :
4166 2412 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4167 2412 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4168 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4169 2412 : }
4170 2412 :
4171 2412 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4172 2412 : my_flush_request = *counter + 1;
4173 2412 : *counter = my_flush_request;
4174 2412 : *lsn = std::cmp::max(at, *lsn);
4175 2412 : });
4176 2412 :
4177 2412 : assert_ne!(my_flush_request, 0);
4178 :
4179 2412 : Ok(my_flush_request)
4180 2412 : }
4181 :
4182 : /// Layer flusher task's main loop.
4183 892 : async fn flush_loop(
4184 892 : self: &Arc<Self>,
4185 892 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4186 892 : ctx: &RequestContext,
4187 892 : ) {
4188 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4189 892 : let mut watch_l0 = match self.layers.read().await.layer_map() {
4190 892 : Ok(lm) => lm.watch_level0_deltas(),
4191 0 : Err(Shutdown) => return,
4192 : };
4193 :
4194 892 : info!("started flush loop");
4195 : loop {
4196 3242 : tokio::select! {
4197 3242 : _ = self.cancel.cancelled() => {
4198 20 : info!("shutting down layer flush task due to Timeline::cancel");
4199 20 : break;
4200 : },
4201 3242 : _ = layer_flush_start_rx.changed() => {}
4202 2350 : }
4203 2350 : trace!("waking up");
4204 2350 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4205 2350 :
4206 2350 : // The highest LSN to which we flushed in the loop over frozen layers
4207 2350 : let mut flushed_to_lsn = Lsn(0);
4208 :
4209 2350 : let result = loop {
4210 4706 : if self.cancel.is_cancelled() {
4211 0 : info!("dropping out of flush loop for timeline shutdown");
4212 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
4213 : // anyone waiting on that will respect self.cancel as well: they will stop
4214 : // waiting at the same time we as drop out of this loop.
4215 0 : return;
4216 4706 : }
4217 4706 :
4218 4706 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4219 4706 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4220 4706 : if flushed_to_lsn >= frozen_to_lsn {
4221 2294 : break Ok(());
4222 2412 : }
4223 :
4224 : // Fetch the next layer to flush, if any.
4225 2412 : let (layer, l0_count, frozen_count, frozen_size) = {
4226 2412 : let layers = self.layers.read().await;
4227 2412 : let Ok(lm) = layers.layer_map() else {
4228 0 : info!("dropping out of flush loop for timeline shutdown");
4229 0 : return;
4230 : };
4231 2412 : let l0_count = lm.level0_deltas().len();
4232 2412 : let frozen_count = lm.frozen_layers.len();
4233 2412 : let frozen_size: u64 = lm
4234 2412 : .frozen_layers
4235 2412 : .iter()
4236 2420 : .map(|l| l.estimated_in_mem_size())
4237 2412 : .sum();
4238 2412 : let layer = lm.frozen_layers.front().cloned();
4239 2412 : (layer, l0_count, frozen_count, frozen_size)
4240 2412 : // drop 'layers' lock
4241 2412 : };
4242 2412 : let Some(layer) = layer else {
4243 56 : break Ok(());
4244 : };
4245 :
4246 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4247 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4248 2356 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4249 0 : if l0_count >= stall_threshold {
4250 0 : warn!(
4251 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4252 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4253 : );
4254 0 : let stall_timer = self
4255 0 : .metrics
4256 0 : .flush_delay_histo
4257 0 : .start_timer()
4258 0 : .record_on_drop();
4259 0 : tokio::select! {
4260 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4261 0 : if let Ok(l0) = result.as_deref() {
4262 0 : let delay = stall_timer.elapsed().as_secs_f64();
4263 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4264 0 : }
4265 : },
4266 0 : _ = self.cancel.cancelled() => {},
4267 : }
4268 0 : continue; // check again
4269 0 : }
4270 2356 : }
4271 :
4272 : // Flush the layer.
4273 2356 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4274 2356 : match self.flush_frozen_layer(layer, ctx).await {
4275 2356 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4276 : Err(FlushLayerError::Cancelled) => {
4277 0 : info!("dropping out of flush loop for timeline shutdown");
4278 0 : return;
4279 : }
4280 0 : err @ Err(
4281 0 : FlushLayerError::NotRunning(_)
4282 0 : | FlushLayerError::Other(_)
4283 0 : | FlushLayerError::CreateImageLayersError(_),
4284 0 : ) => {
4285 0 : error!("could not flush frozen layer: {err:?}");
4286 0 : break err.map(|_| ());
4287 : }
4288 : }
4289 2356 : let flush_duration = flush_timer.stop_and_record();
4290 2356 :
4291 2356 : // Notify the tenant compaction loop if L0 compaction is needed.
4292 2356 : let l0_count = *watch_l0.borrow();
4293 2356 : if l0_count >= self.get_compaction_threshold() {
4294 956 : self.l0_compaction_trigger.notify_one();
4295 1400 : }
4296 :
4297 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4298 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4299 : // ingestion by having ephemeral layer rolls wait for flushes.
4300 2356 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4301 0 : if l0_count >= delay_threshold {
4302 0 : let delay = flush_duration.as_secs_f64();
4303 0 : info!(
4304 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4305 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4306 : );
4307 0 : let _delay_timer = self
4308 0 : .metrics
4309 0 : .flush_delay_histo
4310 0 : .start_timer()
4311 0 : .record_on_drop();
4312 0 : tokio::select! {
4313 0 : _ = tokio::time::sleep(flush_duration) => {},
4314 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4315 0 : _ = self.cancel.cancelled() => {},
4316 : }
4317 0 : }
4318 2356 : }
4319 : };
4320 :
4321 : // Unsharded tenants should never advance their LSN beyond the end of the
4322 : // highest layer they write: such gaps between layer data and the frozen LSN
4323 : // are only legal on sharded tenants.
4324 2350 : debug_assert!(
4325 2350 : self.shard_identity.count.count() > 1
4326 2350 : || flushed_to_lsn >= frozen_to_lsn
4327 56 : || !flushed_to_lsn.is_valid()
4328 : );
4329 :
4330 2350 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4331 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4332 : // to us via layer_flush_start_rx, then advance it here.
4333 : //
4334 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4335 : // never encounter a gap in the wal.
4336 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4337 0 : tracing::debug!(
4338 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
4339 : );
4340 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4341 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4342 0 : tracing::warn!(
4343 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
4344 : );
4345 0 : }
4346 0 : }
4347 2350 : }
4348 :
4349 : // Notify any listeners that we're done
4350 2350 : let _ = self
4351 2350 : .layer_flush_done_tx
4352 2350 : .send_replace((flush_counter, result));
4353 : }
4354 20 : }
4355 :
4356 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4357 2252 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4358 2252 : let mut rx = self.layer_flush_done_tx.subscribe();
4359 : loop {
4360 : {
4361 4576 : let (last_result_counter, last_result) = &*rx.borrow();
4362 4576 : if *last_result_counter >= request {
4363 2252 : if let Err(err) = last_result {
4364 : // We already logged the original error in
4365 : // flush_loop. We cannot propagate it to the caller
4366 : // here, because it might not be Cloneable
4367 0 : return Err(err.clone());
4368 : } else {
4369 2252 : return Ok(());
4370 : }
4371 2324 : }
4372 2324 : }
4373 2324 : trace!("waiting for flush to complete");
4374 2324 : tokio::select! {
4375 2324 : rx_e = rx.changed() => {
4376 2324 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4377 : },
4378 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4379 : // the notification from [`flush_loop`] that it completed.
4380 2324 : _ = self.cancel.cancelled() => {
4381 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4382 0 : return Ok(())
4383 : }
4384 : };
4385 2324 : trace!("done")
4386 : }
4387 2252 : }
4388 :
4389 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4390 : ///
4391 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4392 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4393 : async fn flush_frozen_layer(
4394 : self: &Arc<Self>,
4395 : frozen_layer: Arc<InMemoryLayer>,
4396 : ctx: &RequestContext,
4397 : ) -> Result<Lsn, FlushLayerError> {
4398 : debug_assert_current_span_has_tenant_and_timeline_id();
4399 :
4400 : // As a special case, when we have just imported an image into the repository,
4401 : // instead of writing out a L0 delta layer, we directly write out image layer
4402 : // files instead. This is possible as long as *all* the data imported into the
4403 : // repository have the same LSN.
4404 : let lsn_range = frozen_layer.get_lsn_range();
4405 :
4406 : // Whether to directly create image layers for this flush, or flush them as delta layers
4407 : let create_image_layer =
4408 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4409 :
4410 : #[cfg(test)]
4411 : {
4412 : match &mut *self.flush_loop_state.lock().unwrap() {
4413 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4414 : panic!("flush loop not running")
4415 : }
4416 : FlushLoopState::Running {
4417 : expect_initdb_optimization,
4418 : initdb_optimization_count,
4419 : ..
4420 : } => {
4421 : if create_image_layer {
4422 : *initdb_optimization_count += 1;
4423 : } else {
4424 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4425 : }
4426 : }
4427 : }
4428 : }
4429 :
4430 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4431 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4432 : // require downloading anything during initial import.
4433 : let ((rel_partition, metadata_partition), _lsn) = self
4434 : .repartition(
4435 : self.initdb_lsn,
4436 : self.get_compaction_target_size(),
4437 : EnumSet::empty(),
4438 : ctx,
4439 : )
4440 : .await
4441 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4442 :
4443 : if self.cancel.is_cancelled() {
4444 : return Err(FlushLayerError::Cancelled);
4445 : }
4446 :
4447 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4448 : // So that the key ranges don't overlap.
4449 : let mut partitions = KeyPartitioning::default();
4450 : partitions.parts.extend(rel_partition.parts);
4451 : if !metadata_partition.parts.is_empty() {
4452 : assert_eq!(
4453 : metadata_partition.parts.len(),
4454 : 1,
4455 : "currently sparse keyspace should only contain a single metadata keyspace"
4456 : );
4457 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4458 : // every single key within the keyspace, and therefore, it's safe to force converting it
4459 : // into a dense keyspace before calling this function.
4460 : partitions
4461 : .parts
4462 : .extend(metadata_partition.into_dense().parts);
4463 : }
4464 :
4465 : let mut layers_to_upload = Vec::new();
4466 : let (generated_image_layers, is_complete) = self
4467 : .create_image_layers(
4468 : &partitions,
4469 : self.initdb_lsn,
4470 : ImageLayerCreationMode::Initial,
4471 : ctx,
4472 : LastImageLayerCreationStatus::Initial,
4473 : false, // don't yield for L0, we're flushing L0
4474 : )
4475 : .await?;
4476 : debug_assert!(
4477 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4478 : "init image generation mode must fully cover the keyspace"
4479 : );
4480 : layers_to_upload.extend(generated_image_layers);
4481 :
4482 : (layers_to_upload, None)
4483 : } else {
4484 : // Normal case, write out a L0 delta layer file.
4485 : // `create_delta_layer` will not modify the layer map.
4486 : // We will remove frozen layer and add delta layer in one atomic operation later.
4487 : let Some(layer) = self
4488 : .create_delta_layer(&frozen_layer, None, ctx)
4489 : .await
4490 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4491 : else {
4492 : panic!("delta layer cannot be empty if no filter is applied");
4493 : };
4494 : (
4495 : // FIXME: even though we have a single image and single delta layer assumption
4496 : // we push them to vec
4497 : vec![layer.clone()],
4498 : Some(layer),
4499 : )
4500 : };
4501 :
4502 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4503 :
4504 : if self.cancel.is_cancelled() {
4505 : return Err(FlushLayerError::Cancelled);
4506 : }
4507 :
4508 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4509 :
4510 : // The new on-disk layers are now in the layer map. We can remove the
4511 : // in-memory layer from the map now. The flushed layer is stored in
4512 : // the mapping in `create_delta_layer`.
4513 : {
4514 : let mut guard = self.layers.write().await;
4515 :
4516 : guard.open_mut()?.finish_flush_l0_layer(
4517 : delta_layer_to_add.as_ref(),
4518 : &frozen_layer,
4519 : &self.metrics,
4520 : );
4521 :
4522 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4523 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4524 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4525 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4526 : }
4527 : // release lock on 'layers'
4528 : };
4529 :
4530 : // Backpressure mechanism: wait with continuation of the flush loop until we have uploaded all layer files.
4531 : // This makes us refuse ingest until the new layers have been persisted to the remote
4532 : // TODO: remove this, and rely on l0_flush_{delay,stall}_threshold instead.
4533 : if self.get_l0_flush_wait_upload() {
4534 : let start = Instant::now();
4535 : self.remote_client
4536 : .wait_completion()
4537 : .await
4538 0 : .map_err(|e| match e {
4539 : WaitCompletionError::UploadQueueShutDownOrStopped
4540 : | WaitCompletionError::NotInitialized(
4541 : NotInitialized::ShuttingDown | NotInitialized::Stopped,
4542 0 : ) => FlushLayerError::Cancelled,
4543 : WaitCompletionError::NotInitialized(NotInitialized::Uninitialized) => {
4544 0 : FlushLayerError::Other(anyhow!(e).into())
4545 : }
4546 0 : })?;
4547 : let duration = start.elapsed().as_secs_f64();
4548 : self.metrics.flush_wait_upload_time_gauge_add(duration);
4549 : }
4550 :
4551 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4552 : // a compaction can delete the file and then it won't be available for uploads any more.
4553 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4554 : // race situation.
4555 : // See https://github.com/neondatabase/neon/issues/4526
4556 : pausable_failpoint!("flush-frozen-pausable");
4557 :
4558 : // This failpoint is used by another test case `test_pageserver_recovery`.
4559 : fail_point!("flush-frozen-exit");
4560 :
4561 : Ok(Lsn(lsn_range.end.0 - 1))
4562 : }
4563 :
4564 : /// Return true if the value changed
4565 : ///
4566 : /// This function must only be used from the layer flush task.
4567 2356 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4568 2356 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4569 2356 : assert!(
4570 2356 : new_value >= old_value,
4571 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
4572 : );
4573 :
4574 2356 : self.metrics
4575 2356 : .disk_consistent_lsn_gauge
4576 2356 : .set(new_value.0 as i64);
4577 2356 : new_value != old_value
4578 2356 : }
4579 :
4580 : /// Update metadata file
4581 2456 : fn schedule_uploads(
4582 2456 : &self,
4583 2456 : disk_consistent_lsn: Lsn,
4584 2456 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4585 2456 : ) -> anyhow::Result<()> {
4586 2456 : // We can only save a valid 'prev_record_lsn' value on disk if we
4587 2456 : // flushed *all* in-memory changes to disk. We only track
4588 2456 : // 'prev_record_lsn' in memory for the latest processed record, so we
4589 2456 : // don't remember what the correct value that corresponds to some old
4590 2456 : // LSN is. But if we flush everything, then the value corresponding
4591 2456 : // current 'last_record_lsn' is correct and we can store it on disk.
4592 2456 : let RecordLsn {
4593 2456 : last: last_record_lsn,
4594 2456 : prev: prev_record_lsn,
4595 2456 : } = self.last_record_lsn.load();
4596 2456 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4597 2201 : Some(prev_record_lsn)
4598 : } else {
4599 255 : None
4600 : };
4601 :
4602 2456 : let update = crate::tenant::metadata::MetadataUpdate::new(
4603 2456 : disk_consistent_lsn,
4604 2456 : ondisk_prev_record_lsn,
4605 2456 : *self.applied_gc_cutoff_lsn.read(),
4606 2456 : );
4607 2456 :
4608 2456 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4609 0 : "{}",
4610 0 : x.unwrap()
4611 2456 : ));
4612 :
4613 4836 : for layer in layers_to_upload {
4614 2380 : self.remote_client.schedule_layer_file_upload(layer)?;
4615 : }
4616 2456 : self.remote_client
4617 2456 : .schedule_index_upload_for_metadata_update(&update)?;
4618 :
4619 2456 : Ok(())
4620 2456 : }
4621 :
4622 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4623 0 : self.remote_client
4624 0 : .preserve_initdb_archive(
4625 0 : &self.tenant_shard_id.tenant_id,
4626 0 : &self.timeline_id,
4627 0 : &self.cancel,
4628 0 : )
4629 0 : .await
4630 0 : }
4631 :
4632 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4633 : // in layer map immediately. The caller is responsible to put it into the layer map.
4634 1936 : async fn create_delta_layer(
4635 1936 : self: &Arc<Self>,
4636 1936 : frozen_layer: &Arc<InMemoryLayer>,
4637 1936 : key_range: Option<Range<Key>>,
4638 1936 : ctx: &RequestContext,
4639 1936 : ) -> anyhow::Result<Option<ResidentLayer>> {
4640 1936 : let self_clone = Arc::clone(self);
4641 1936 : let frozen_layer = Arc::clone(frozen_layer);
4642 1936 : let ctx = ctx.attached_child();
4643 1936 : let work = async move {
4644 1936 : let Some((desc, path)) = frozen_layer
4645 1936 : .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
4646 1936 : .await?
4647 : else {
4648 0 : return Ok(None);
4649 : };
4650 1936 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
4651 :
4652 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
4653 : // We just need to fsync the directory in which these inodes are linked,
4654 : // which we know to be the timeline directory.
4655 : //
4656 : // We use fatal_err() below because the after write_to_disk returns with success,
4657 : // the in-memory state of the filesystem already has the layer file in its final place,
4658 : // and subsequent pageserver code could think it's durable while it really isn't.
4659 1936 : let timeline_dir = VirtualFile::open(
4660 1936 : &self_clone
4661 1936 : .conf
4662 1936 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
4663 1936 : &ctx,
4664 1936 : )
4665 1936 : .await
4666 1936 : .fatal_err("VirtualFile::open for timeline dir fsync");
4667 1936 : timeline_dir
4668 1936 : .sync_all()
4669 1936 : .await
4670 1936 : .fatal_err("VirtualFile::sync_all timeline dir");
4671 1936 : anyhow::Ok(Some(new_delta))
4672 1936 : };
4673 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
4674 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
4675 : use crate::virtual_file::io_engine::IoEngine;
4676 1936 : match crate::virtual_file::io_engine::get() {
4677 0 : IoEngine::NotSet => panic!("io engine not set"),
4678 : IoEngine::StdFs => {
4679 968 : let span = tracing::info_span!("blocking");
4680 968 : tokio::task::spawn_blocking({
4681 968 : move || Handle::current().block_on(work.instrument(span))
4682 968 : })
4683 968 : .await
4684 968 : .context("spawn_blocking")
4685 968 : .and_then(|x| x)
4686 : }
4687 : #[cfg(target_os = "linux")]
4688 968 : IoEngine::TokioEpollUring => work.await,
4689 : }
4690 1936 : }
4691 :
4692 1148 : async fn repartition(
4693 1148 : &self,
4694 1148 : lsn: Lsn,
4695 1148 : partition_size: u64,
4696 1148 : flags: EnumSet<CompactFlags>,
4697 1148 : ctx: &RequestContext,
4698 1148 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
4699 1148 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
4700 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
4701 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
4702 : // and hence before the compaction task starts.
4703 0 : return Err(CompactionError::Other(anyhow!(
4704 0 : "repartition() called concurrently"
4705 0 : )));
4706 : };
4707 1148 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
4708 1148 : if lsn < *partition_lsn {
4709 0 : return Err(CompactionError::Other(anyhow!(
4710 0 : "repartition() called with LSN going backwards, this should not happen"
4711 0 : )));
4712 1148 : }
4713 1148 :
4714 1148 : let distance = lsn.0 - partition_lsn.0;
4715 1148 : if *partition_lsn != Lsn(0)
4716 524 : && distance <= self.repartition_threshold
4717 524 : && !flags.contains(CompactFlags::ForceRepartition)
4718 : {
4719 496 : debug!(
4720 : distance,
4721 : threshold = self.repartition_threshold,
4722 0 : "no repartitioning needed"
4723 : );
4724 496 : return Ok((
4725 496 : (dense_partition.clone(), sparse_partition.clone()),
4726 496 : *partition_lsn,
4727 496 : ));
4728 652 : }
4729 :
4730 652 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
4731 652 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
4732 652 : let sparse_partitioning = SparseKeyPartitioning {
4733 652 : parts: vec![sparse_ks],
4734 652 : }; // no partitioning for metadata keys for now
4735 652 : let result = ((dense_partitioning, sparse_partitioning), lsn);
4736 652 : guard.write(result.clone());
4737 652 : Ok(result)
4738 1148 : }
4739 :
4740 : // Is it time to create a new image layer for the given partition? True if we want to generate.
4741 28 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
4742 28 : let threshold = self.get_image_creation_threshold();
4743 :
4744 28 : let guard = self.layers.read().await;
4745 28 : let Ok(layers) = guard.layer_map() else {
4746 0 : return false;
4747 : };
4748 :
4749 28 : let mut max_deltas = 0;
4750 56 : for part_range in &partition.ranges {
4751 28 : let image_coverage = layers.image_coverage(part_range, lsn);
4752 56 : for (img_range, last_img) in image_coverage {
4753 28 : let img_lsn = if let Some(last_img) = last_img {
4754 0 : last_img.get_lsn_range().end
4755 : } else {
4756 28 : Lsn(0)
4757 : };
4758 : // Let's consider an example:
4759 : //
4760 : // delta layer with LSN range 71-81
4761 : // delta layer with LSN range 81-91
4762 : // delta layer with LSN range 91-101
4763 : // image layer at LSN 100
4764 : //
4765 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
4766 : // there's no need to create a new one. We check this case explicitly, to avoid passing
4767 : // a bogus range to count_deltas below, with start > end. It's even possible that there
4768 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
4769 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
4770 28 : if img_lsn < lsn {
4771 28 : let num_deltas =
4772 28 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
4773 28 :
4774 28 : max_deltas = max_deltas.max(num_deltas);
4775 28 : if num_deltas >= threshold {
4776 0 : debug!(
4777 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
4778 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
4779 : );
4780 0 : return true;
4781 28 : }
4782 0 : }
4783 : }
4784 : }
4785 :
4786 28 : debug!(
4787 : max_deltas,
4788 0 : "none of the partitioned ranges had >= {threshold} deltas"
4789 : );
4790 28 : false
4791 28 : }
4792 :
4793 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
4794 : /// so that at most one image layer will be produced from this function.
4795 : #[allow(clippy::too_many_arguments)]
4796 472 : async fn create_image_layer_for_rel_blocks(
4797 472 : self: &Arc<Self>,
4798 472 : partition: &KeySpace,
4799 472 : mut image_layer_writer: ImageLayerWriter,
4800 472 : lsn: Lsn,
4801 472 : ctx: &RequestContext,
4802 472 : img_range: Range<Key>,
4803 472 : io_concurrency: IoConcurrency,
4804 472 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4805 472 : let mut wrote_keys = false;
4806 472 :
4807 472 : let mut key_request_accum = KeySpaceAccum::new();
4808 3128 : for range in &partition.ranges {
4809 2656 : let mut key = range.start;
4810 5756 : while key < range.end {
4811 : // Decide whether to retain this key: usually we do, but sharded tenants may
4812 : // need to drop keys that don't belong to them. If we retain the key, add it
4813 : // to `key_request_accum` for later issuing a vectored get
4814 3100 : if self.shard_identity.is_key_disposable(&key) {
4815 0 : debug!(
4816 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
4817 0 : key,
4818 0 : self.shard_identity.get_shard_number(&key)
4819 : );
4820 3100 : } else {
4821 3100 : key_request_accum.add_key(key);
4822 3100 : }
4823 :
4824 3100 : let last_key_in_range = key.next() == range.end;
4825 3100 : key = key.next();
4826 3100 :
4827 3100 : // Maybe flush `key_rest_accum`
4828 3100 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
4829 3100 : || (last_key_in_range && key_request_accum.raw_size() > 0)
4830 : {
4831 2656 : let results = self
4832 2656 : .get_vectored(
4833 2656 : key_request_accum.consume_keyspace(),
4834 2656 : lsn,
4835 2656 : io_concurrency.clone(),
4836 2656 : ctx,
4837 2656 : )
4838 2656 : .await?;
4839 :
4840 2656 : if self.cancel.is_cancelled() {
4841 0 : return Err(CreateImageLayersError::Cancelled);
4842 2656 : }
4843 :
4844 5756 : for (img_key, img) in results {
4845 3100 : let img = match img {
4846 3100 : Ok(img) => img,
4847 0 : Err(err) => {
4848 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
4849 0 : // page without losing any actual user data. That seems better
4850 0 : // than failing repeatedly and getting stuck.
4851 0 : //
4852 0 : // We had a bug at one point, where we truncated the FSM and VM
4853 0 : // in the pageserver, but the Postgres didn't know about that
4854 0 : // and continued to generate incremental WAL records for pages
4855 0 : // that didn't exist in the pageserver. Trying to replay those
4856 0 : // WAL records failed to find the previous image of the page.
4857 0 : // This special case allows us to recover from that situation.
4858 0 : // See https://github.com/neondatabase/neon/issues/2601.
4859 0 : //
4860 0 : // Unfortunately we cannot do this for the main fork, or for
4861 0 : // any metadata keys, keys, as that would lead to actual data
4862 0 : // loss.
4863 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
4864 0 : warn!(
4865 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
4866 : );
4867 0 : ZERO_PAGE.clone()
4868 : } else {
4869 0 : return Err(CreateImageLayersError::from(err));
4870 : }
4871 : }
4872 : };
4873 :
4874 : // Write all the keys we just read into our new image layer.
4875 3100 : image_layer_writer.put_image(img_key, img, ctx).await?;
4876 3100 : wrote_keys = true;
4877 : }
4878 444 : }
4879 : }
4880 : }
4881 :
4882 472 : if wrote_keys {
4883 : // Normal path: we have written some data into the new image layer for this
4884 : // partition, so flush it to disk.
4885 472 : info!(
4886 0 : "produced image layer for rel {}",
4887 0 : ImageLayerName {
4888 0 : key_range: img_range.clone(),
4889 0 : lsn
4890 0 : },
4891 : );
4892 472 : Ok(ImageLayerCreationOutcome::Generated {
4893 472 : unfinished_image_layer: image_layer_writer,
4894 472 : })
4895 : } else {
4896 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4897 0 : Ok(ImageLayerCreationOutcome::Empty)
4898 : }
4899 472 : }
4900 :
4901 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
4902 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
4903 : /// would not be too large to fit in a single image layer.
4904 : ///
4905 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
4906 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
4907 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
4908 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
4909 : #[allow(clippy::too_many_arguments)]
4910 452 : async fn create_image_layer_for_metadata_keys(
4911 452 : self: &Arc<Self>,
4912 452 : partition: &KeySpace,
4913 452 : mut image_layer_writer: ImageLayerWriter,
4914 452 : lsn: Lsn,
4915 452 : ctx: &RequestContext,
4916 452 : img_range: Range<Key>,
4917 452 : mode: ImageLayerCreationMode,
4918 452 : io_concurrency: IoConcurrency,
4919 452 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
4920 452 : // Metadata keys image layer creation.
4921 452 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
4922 452 : let begin = Instant::now();
4923 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
4924 : // not contain too many keys, otherwise this takes a lot of memory.
4925 452 : let data = self
4926 452 : .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
4927 452 : .await?;
4928 452 : let (data, total_kb_retrieved, total_keys_retrieved) = {
4929 452 : let mut new_data = BTreeMap::new();
4930 452 : let mut total_kb_retrieved = 0;
4931 452 : let mut total_keys_retrieved = 0;
4932 20476 : for (k, v) in data {
4933 20024 : let v = v?;
4934 20024 : total_kb_retrieved += KEY_SIZE + v.len();
4935 20024 : total_keys_retrieved += 1;
4936 20024 : new_data.insert(k, v);
4937 : }
4938 452 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
4939 452 : };
4940 452 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
4941 452 : let elapsed = begin.elapsed();
4942 452 :
4943 452 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
4944 452 : info!(
4945 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",
4946 0 : elapsed.as_secs_f64()
4947 : );
4948 :
4949 452 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
4950 4 : return Ok(ImageLayerCreationOutcome::Skip);
4951 448 : }
4952 448 : if self.cancel.is_cancelled() {
4953 0 : return Err(CreateImageLayersError::Cancelled);
4954 448 : }
4955 448 : let mut wrote_any_image = false;
4956 20472 : for (k, v) in data {
4957 20024 : if v.is_empty() {
4958 : // the key has been deleted, it does not need an image
4959 : // in metadata keyspace, an empty image == tombstone
4960 16 : continue;
4961 20008 : }
4962 20008 : wrote_any_image = true;
4963 20008 :
4964 20008 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
4965 20008 :
4966 20008 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
4967 20008 : // on the normal data path either.
4968 20008 : image_layer_writer.put_image(k, v, ctx).await?;
4969 : }
4970 :
4971 448 : if wrote_any_image {
4972 : // Normal path: we have written some data into the new image layer for this
4973 : // partition, so flush it to disk.
4974 24 : info!(
4975 0 : "created image layer for metadata {}",
4976 0 : ImageLayerName {
4977 0 : key_range: img_range.clone(),
4978 0 : lsn
4979 0 : }
4980 : );
4981 24 : Ok(ImageLayerCreationOutcome::Generated {
4982 24 : unfinished_image_layer: image_layer_writer,
4983 24 : })
4984 : } else {
4985 424 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
4986 424 : Ok(ImageLayerCreationOutcome::Empty)
4987 : }
4988 452 : }
4989 :
4990 : /// Predicate function which indicates whether we should check if new image layers
4991 : /// are required. Since checking if new image layers are required is expensive in
4992 : /// terms of CPU, we only do it in the following cases:
4993 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
4994 : /// 2. If enough time has passed since the last check:
4995 : /// 1. For large tenants, we wish to perform the check more often since they
4996 : /// suffer from the lack of image layers
4997 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
4998 1148 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
4999 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
5000 :
5001 1148 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5002 1148 : let distance = lsn
5003 1148 : .checked_sub(last_checks_at)
5004 1148 : .expect("Attempt to compact with LSN going backwards");
5005 1148 : let min_distance =
5006 1148 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5007 1148 :
5008 1148 : let distance_based_decision = distance.0 >= min_distance;
5009 1148 :
5010 1148 : let mut time_based_decision = false;
5011 1148 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5012 1148 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
5013 944 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
5014 : {
5015 0 : self.get_checkpoint_timeout()
5016 : } else {
5017 944 : Duration::from_secs(3600 * 48)
5018 : };
5019 :
5020 944 : time_based_decision = match *last_check_instant {
5021 524 : Some(last_check) => {
5022 524 : let elapsed = last_check.elapsed();
5023 524 : elapsed >= check_required_after
5024 : }
5025 420 : None => true,
5026 : };
5027 204 : }
5028 :
5029 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5030 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5031 : // check.
5032 1148 : let decision = distance_based_decision || time_based_decision;
5033 :
5034 1148 : if decision {
5035 424 : self.last_image_layer_creation_check_at.store(lsn);
5036 424 : *last_check_instant = Some(Instant::now());
5037 724 : }
5038 :
5039 1148 : decision
5040 1148 : }
5041 :
5042 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5043 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5044 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
5045 : async fn create_image_layers(
5046 : self: &Arc<Timeline>,
5047 : partitioning: &KeyPartitioning,
5048 : lsn: Lsn,
5049 : mode: ImageLayerCreationMode,
5050 : ctx: &RequestContext,
5051 : last_status: LastImageLayerCreationStatus,
5052 : yield_for_l0: bool,
5053 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5054 : let timer = self.metrics.create_images_time_histo.start_timer();
5055 :
5056 : if partitioning.parts.is_empty() {
5057 : warn!("no partitions to create image layers for");
5058 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5059 : }
5060 :
5061 : // We need to avoid holes between generated image layers.
5062 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5063 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5064 : //
5065 : // How such hole between partitions can appear?
5066 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5067 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5068 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5069 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5070 : let mut start = Key::MIN;
5071 :
5072 : let check_for_image_layers =
5073 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5074 : info!(
5075 : "resuming image layer creation: last_status=incomplete, continue from {}",
5076 : last_key
5077 : );
5078 : true
5079 : } else {
5080 : self.should_check_if_image_layers_required(lsn)
5081 : };
5082 :
5083 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
5084 :
5085 : let mut all_generated = true;
5086 :
5087 : let mut partition_processed = 0;
5088 : let mut total_partitions = partitioning.parts.len();
5089 : let mut last_partition_processed = None;
5090 : let mut partition_parts = partitioning.parts.clone();
5091 :
5092 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5093 : // We need to skip the partitions that have already been processed.
5094 : let mut found = false;
5095 : for (i, partition) in partition_parts.iter().enumerate() {
5096 : if last_key <= partition.end().unwrap() {
5097 : // ```plain
5098 : // |------|--------|----------|------|
5099 : // ^last_key
5100 : // ^start from this partition
5101 : // ```
5102 : // Why `i+1` instead of `i`?
5103 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5104 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5105 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5106 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5107 : if i + 1 >= total_partitions {
5108 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5109 : // iteration of image layer creation should return a complete status.
5110 : break; // with found=false
5111 : }
5112 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5113 : total_partitions = partition_parts.len();
5114 : // Update the start key to the partition start.
5115 : start = partition_parts[0].start().unwrap();
5116 : found = true;
5117 : break;
5118 : }
5119 : }
5120 : if !found {
5121 : // Last key is within the last partition, or larger than all partitions.
5122 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5123 : }
5124 : }
5125 :
5126 : for partition in partition_parts.iter() {
5127 : if self.cancel.is_cancelled() {
5128 : return Err(CreateImageLayersError::Cancelled);
5129 : }
5130 : partition_processed += 1;
5131 : let img_range = start..partition.ranges.last().unwrap().end;
5132 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5133 : if compact_metadata {
5134 : for range in &partition.ranges {
5135 : assert!(
5136 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5137 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5138 : "metadata keys must be partitioned separately"
5139 : );
5140 : }
5141 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5142 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5143 : // might mess up with evictions.
5144 : start = img_range.end;
5145 : continue;
5146 : }
5147 : // For initial and force modes, we always generate image layers for metadata keys.
5148 : } else if let ImageLayerCreationMode::Try = mode {
5149 : // check_for_image_layers = false -> skip
5150 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5151 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
5152 : start = img_range.end;
5153 : continue;
5154 : }
5155 : }
5156 : if let ImageLayerCreationMode::Force = mode {
5157 : // When forced to create image layers, we might try and create them where they already
5158 : // exist. This mode is only used in tests/debug.
5159 : let layers = self.layers.read().await;
5160 : if layers.contains_key(&PersistentLayerKey {
5161 : key_range: img_range.clone(),
5162 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5163 : is_delta: false,
5164 : }) {
5165 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5166 : // in the future.
5167 : tracing::info!(
5168 : "Skipping image layer at {lsn} {}..{}, already exists",
5169 : img_range.start,
5170 : img_range.end
5171 : );
5172 : start = img_range.end;
5173 : continue;
5174 : }
5175 : }
5176 :
5177 : let image_layer_writer = ImageLayerWriter::new(
5178 : self.conf,
5179 : self.timeline_id,
5180 : self.tenant_shard_id,
5181 : &img_range,
5182 : lsn,
5183 : ctx,
5184 : )
5185 : .await?;
5186 :
5187 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5188 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5189 0 : "failpoint image-layer-writer-fail-before-finish"
5190 0 : )))
5191 0 : });
5192 :
5193 : let io_concurrency = IoConcurrency::spawn_from_conf(
5194 : self.conf,
5195 : self.gate
5196 : .enter()
5197 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5198 : );
5199 :
5200 : let outcome = if !compact_metadata {
5201 : self.create_image_layer_for_rel_blocks(
5202 : partition,
5203 : image_layer_writer,
5204 : lsn,
5205 : ctx,
5206 : img_range.clone(),
5207 : io_concurrency,
5208 : )
5209 : .await?
5210 : } else {
5211 : self.create_image_layer_for_metadata_keys(
5212 : partition,
5213 : image_layer_writer,
5214 : lsn,
5215 : ctx,
5216 : img_range.clone(),
5217 : mode,
5218 : io_concurrency,
5219 : )
5220 : .await?
5221 : };
5222 : match outcome {
5223 : ImageLayerCreationOutcome::Empty => {
5224 : // No data in this partition, so we don't need to create an image layer (for now).
5225 : // The next image layer should cover this key range, so we don't advance the `start`
5226 : // key.
5227 : }
5228 : ImageLayerCreationOutcome::Generated {
5229 : unfinished_image_layer,
5230 : } => {
5231 : batch_image_writer.add_unfinished_image_writer(
5232 : unfinished_image_layer,
5233 : img_range.clone(),
5234 : lsn,
5235 : );
5236 : // The next image layer should be generated right after this one.
5237 : start = img_range.end;
5238 : }
5239 : ImageLayerCreationOutcome::Skip => {
5240 : // We don't need to create an image layer for this partition.
5241 : // The next image layer should NOT cover this range, otherwise
5242 : // the keyspace becomes empty (reads don't go past image layers).
5243 : start = img_range.end;
5244 : }
5245 : }
5246 :
5247 : if let ImageLayerCreationMode::Try = mode {
5248 : // We have at least made some progress
5249 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5250 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5251 : // image layer generation taking too long time and blocking L0 compaction. So in this
5252 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5253 : // if there are too many of them.
5254 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5255 : * self.get_compaction_threshold();
5256 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
5257 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
5258 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
5259 : if image_preempt_threshold != 0 {
5260 : let should_yield = self
5261 : .l0_compaction_trigger
5262 : .notified()
5263 : .now_or_never()
5264 : .is_some();
5265 : if should_yield {
5266 : tracing::info!(
5267 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
5268 : partition.start().unwrap(),
5269 : partition.end().unwrap()
5270 : );
5271 : last_partition_processed = Some(partition.clone());
5272 : all_generated = false;
5273 : break;
5274 : }
5275 : }
5276 : }
5277 : }
5278 : }
5279 :
5280 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5281 :
5282 : let mut guard = self.layers.write().await;
5283 :
5284 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5285 : // now they are being scheduled outside of write lock; current way is inconsistent with
5286 : // compaction lock order.
5287 : guard
5288 : .open_mut()?
5289 : .track_new_image_layers(&image_layers, &self.metrics);
5290 : drop_wlock(guard);
5291 : let duration = timer.stop_and_record();
5292 :
5293 : // Creating image layers may have caused some previously visible layers to be covered
5294 : if !image_layers.is_empty() {
5295 : self.update_layer_visibility().await?;
5296 : }
5297 :
5298 : let total_layer_size = image_layers
5299 : .iter()
5300 496 : .map(|l| l.metadata().file_size)
5301 : .sum::<u64>();
5302 :
5303 : if !image_layers.is_empty() {
5304 : info!(
5305 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5306 : image_layers.len(),
5307 : total_layer_size,
5308 : duration.as_secs_f64(),
5309 : partition_processed,
5310 : total_partitions
5311 : );
5312 : }
5313 :
5314 : Ok((
5315 : image_layers,
5316 : if all_generated {
5317 : LastImageLayerCreationStatus::Complete
5318 : } else {
5319 : LastImageLayerCreationStatus::Incomplete {
5320 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5321 : last_partition_processed.end().unwrap_or(Key::MIN)
5322 : } else {
5323 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5324 : Key::MIN
5325 : },
5326 : }
5327 : },
5328 : ))
5329 : }
5330 :
5331 : /// Wait until the background initial logical size calculation is complete, or
5332 : /// this Timeline is shut down. Calling this function will cause the initial
5333 : /// logical size calculation to skip waiting for the background jobs barrier.
5334 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5335 0 : if !self.shard_identity.is_shard_zero() {
5336 : // We don't populate logical size on shard >0: skip waiting for it.
5337 0 : return;
5338 0 : }
5339 0 :
5340 0 : if self.remote_client.is_deleting() {
5341 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5342 0 : return;
5343 0 : }
5344 0 :
5345 0 : if self.current_logical_size.current_size().is_exact() {
5346 : // root timelines are initialized with exact count, but never start the background
5347 : // calculation
5348 0 : return;
5349 0 : }
5350 :
5351 0 : if let Some(await_bg_cancel) = self
5352 0 : .current_logical_size
5353 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5354 0 : .get()
5355 0 : {
5356 0 : await_bg_cancel.cancel();
5357 0 : } else {
5358 : // We should not wait if we were not able to explicitly instruct
5359 : // the logical size cancellation to skip the concurrency limit semaphore.
5360 : // TODO: this is an unexpected case. We should restructure so that it
5361 : // can't happen.
5362 0 : tracing::warn!(
5363 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5364 : );
5365 0 : debug_assert!(false);
5366 : }
5367 :
5368 0 : tokio::select!(
5369 0 : _ = self.current_logical_size.initialized.acquire() => {},
5370 0 : _ = self.cancel.cancelled() => {}
5371 : )
5372 0 : }
5373 :
5374 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5375 : /// Timelines layers up to the ancestor_lsn.
5376 : ///
5377 : /// Requires a timeline that:
5378 : /// - has an ancestor to detach from
5379 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5380 : /// a technical requirement
5381 : ///
5382 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5383 : /// polled again until completion.
5384 : ///
5385 : /// During the operation all timelines sharing the data with this timeline will be reparented
5386 : /// from our ancestor to be branches of this timeline.
5387 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5388 0 : self: &Arc<Timeline>,
5389 0 : tenant: &crate::tenant::Tenant,
5390 0 : options: detach_ancestor::Options,
5391 0 : ctx: &RequestContext,
5392 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5393 0 : detach_ancestor::prepare(self, tenant, options, ctx).await
5394 0 : }
5395 :
5396 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5397 : /// reparents any reparentable children of previous ancestor.
5398 : ///
5399 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5400 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5401 : /// successfully, tenant must be reloaded.
5402 : ///
5403 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5404 : /// resetting the tenant.
5405 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5406 0 : self: &Arc<Timeline>,
5407 0 : tenant: &crate::tenant::Tenant,
5408 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5409 0 : ctx: &RequestContext,
5410 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5411 0 : detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
5412 0 : }
5413 :
5414 : /// Final step which unblocks the GC.
5415 : ///
5416 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5417 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5418 0 : self: &Arc<Timeline>,
5419 0 : tenant: &crate::tenant::Tenant,
5420 0 : attempt: detach_ancestor::Attempt,
5421 0 : ctx: &RequestContext,
5422 0 : ) -> Result<(), detach_ancestor::Error> {
5423 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5424 0 : }
5425 : }
5426 :
5427 : impl Drop for Timeline {
5428 20 : fn drop(&mut self) {
5429 20 : if let Some(ancestor) = &self.ancestor_timeline {
5430 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5431 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5432 8 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5433 8 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5434 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5435 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5436 8 : }
5437 0 : }
5438 12 : }
5439 20 : info!(
5440 0 : "Timeline {} for tenant {} is being dropped",
5441 : self.timeline_id, self.tenant_shard_id.tenant_id
5442 : );
5443 20 : }
5444 : }
5445 :
5446 : /// Top-level failure to compact.
5447 : #[derive(Debug, thiserror::Error)]
5448 : pub(crate) enum CompactionError {
5449 : #[error("The timeline or pageserver is shutting down")]
5450 : ShuttingDown,
5451 : /// Compaction tried to offload a timeline and failed
5452 : #[error("Failed to offload timeline: {0}")]
5453 : Offload(OffloadError),
5454 : /// Compaction cannot be done right now; page reconstruction and so on.
5455 : #[error("Failed to collect keyspace: {0}")]
5456 : CollectKeySpaceError(#[from] CollectKeySpaceError),
5457 : #[error(transparent)]
5458 : Other(anyhow::Error),
5459 : #[error("Compaction already running: {0}")]
5460 : AlreadyRunning(&'static str),
5461 : }
5462 :
5463 : impl CompactionError {
5464 : /// Errors that can be ignored, i.e., cancel and shutdown.
5465 0 : pub fn is_cancel(&self) -> bool {
5466 0 : matches!(
5467 0 : self,
5468 : Self::ShuttingDown
5469 : | Self::AlreadyRunning(_)
5470 : | Self::CollectKeySpaceError(CollectKeySpaceError::Cancelled)
5471 : | Self::CollectKeySpaceError(CollectKeySpaceError::PageRead(
5472 : PageReconstructError::Cancelled
5473 : ))
5474 : | Self::Offload(OffloadError::Cancelled)
5475 : )
5476 0 : }
5477 :
5478 : /// Critical errors that indicate data corruption.
5479 0 : pub fn is_critical(&self) -> bool {
5480 0 : matches!(
5481 0 : self,
5482 : Self::CollectKeySpaceError(
5483 : CollectKeySpaceError::Decode(_)
5484 : | CollectKeySpaceError::PageRead(
5485 : PageReconstructError::MissingKey(_) | PageReconstructError::WalRedo(_),
5486 : )
5487 : )
5488 : )
5489 0 : }
5490 : }
5491 :
5492 : impl From<OffloadError> for CompactionError {
5493 0 : fn from(e: OffloadError) -> Self {
5494 0 : match e {
5495 0 : OffloadError::Cancelled => Self::ShuttingDown,
5496 0 : _ => Self::Offload(e),
5497 : }
5498 0 : }
5499 : }
5500 :
5501 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5502 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5503 0 : match value {
5504 : super::upload_queue::NotInitialized::Uninitialized => {
5505 0 : CompactionError::Other(anyhow::anyhow!(value))
5506 : }
5507 : super::upload_queue::NotInitialized::ShuttingDown
5508 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5509 : }
5510 0 : }
5511 : }
5512 :
5513 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5514 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5515 0 : match e {
5516 : super::storage_layer::layer::DownloadError::TimelineShutdown
5517 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5518 0 : CompactionError::ShuttingDown
5519 : }
5520 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5521 : | super::storage_layer::layer::DownloadError::DownloadRequired
5522 : | super::storage_layer::layer::DownloadError::NotFile(_)
5523 : | super::storage_layer::layer::DownloadError::DownloadFailed
5524 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5525 0 : CompactionError::Other(anyhow::anyhow!(e))
5526 : }
5527 : #[cfg(test)]
5528 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5529 0 : CompactionError::Other(anyhow::anyhow!(e))
5530 : }
5531 : }
5532 0 : }
5533 : }
5534 :
5535 : impl From<layer_manager::Shutdown> for CompactionError {
5536 0 : fn from(_: layer_manager::Shutdown) -> Self {
5537 0 : CompactionError::ShuttingDown
5538 0 : }
5539 : }
5540 :
5541 : #[serde_as]
5542 392 : #[derive(serde::Serialize)]
5543 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5544 :
5545 : #[derive(Default)]
5546 : enum DurationRecorder {
5547 : #[default]
5548 : NotStarted,
5549 : Recorded(RecordedDuration, tokio::time::Instant),
5550 : }
5551 :
5552 : impl DurationRecorder {
5553 1008 : fn till_now(&self) -> DurationRecorder {
5554 1008 : match self {
5555 : DurationRecorder::NotStarted => {
5556 0 : panic!("must only call on recorded measurements")
5557 : }
5558 1008 : DurationRecorder::Recorded(_, ended) => {
5559 1008 : let now = tokio::time::Instant::now();
5560 1008 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5561 1008 : }
5562 1008 : }
5563 1008 : }
5564 392 : fn into_recorded(self) -> Option<RecordedDuration> {
5565 392 : match self {
5566 0 : DurationRecorder::NotStarted => None,
5567 392 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5568 : }
5569 392 : }
5570 : }
5571 :
5572 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5573 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5574 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5575 : /// keys specified in the `data` field.
5576 : #[cfg(test)]
5577 : #[derive(Clone)]
5578 : pub struct DeltaLayerTestDesc {
5579 : pub lsn_range: Range<Lsn>,
5580 : pub key_range: Range<Key>,
5581 : pub data: Vec<(Key, Lsn, Value)>,
5582 : }
5583 :
5584 : #[cfg(test)]
5585 : #[derive(Clone)]
5586 : pub struct InMemoryLayerTestDesc {
5587 : pub lsn_range: Range<Lsn>,
5588 : pub data: Vec<(Key, Lsn, Value)>,
5589 : pub is_open: bool,
5590 : }
5591 :
5592 : #[cfg(test)]
5593 : impl DeltaLayerTestDesc {
5594 8 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5595 8 : Self {
5596 8 : lsn_range,
5597 8 : key_range,
5598 8 : data,
5599 8 : }
5600 8 : }
5601 :
5602 176 : pub fn new_with_inferred_key_range(
5603 176 : lsn_range: Range<Lsn>,
5604 176 : data: Vec<(Key, Lsn, Value)>,
5605 176 : ) -> Self {
5606 440 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5607 440 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5608 176 : Self {
5609 176 : key_range: (*key_min)..(key_max.next()),
5610 176 : lsn_range,
5611 176 : data,
5612 176 : }
5613 176 : }
5614 :
5615 20 : pub(crate) fn layer_name(&self) -> LayerName {
5616 20 : LayerName::Delta(super::storage_layer::DeltaLayerName {
5617 20 : key_range: self.key_range.clone(),
5618 20 : lsn_range: self.lsn_range.clone(),
5619 20 : })
5620 20 : }
5621 : }
5622 :
5623 : impl Timeline {
5624 56 : async fn finish_compact_batch(
5625 56 : self: &Arc<Self>,
5626 56 : new_deltas: &[ResidentLayer],
5627 56 : new_images: &[ResidentLayer],
5628 56 : layers_to_remove: &[Layer],
5629 56 : ) -> Result<(), CompactionError> {
5630 56 : let mut guard = tokio::select! {
5631 56 : guard = self.layers.write() => guard,
5632 56 : _ = self.cancel.cancelled() => {
5633 0 : return Err(CompactionError::ShuttingDown);
5634 : }
5635 : };
5636 :
5637 56 : let mut duplicated_layers = HashSet::new();
5638 56 :
5639 56 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
5640 :
5641 672 : for l in new_deltas {
5642 616 : if guard.contains(l.as_ref()) {
5643 : // expected in tests
5644 0 : tracing::error!(layer=%l, "duplicated L1 layer");
5645 :
5646 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
5647 : // `guard` on self.layers. as of writing this, there are no error returns except
5648 : // for compact_level0_phase1 creating an L0, which does not happen in practice
5649 : // because we have not implemented L0 => L0 compaction.
5650 0 : duplicated_layers.insert(l.layer_desc().key());
5651 616 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
5652 0 : return Err(CompactionError::Other(anyhow::anyhow!(
5653 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
5654 0 : )));
5655 616 : } else {
5656 616 : insert_layers.push(l.clone());
5657 616 : }
5658 : }
5659 :
5660 : // only remove those inputs which were not outputs
5661 56 : let remove_layers: Vec<Layer> = layers_to_remove
5662 56 : .iter()
5663 804 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
5664 56 : .cloned()
5665 56 : .collect();
5666 56 :
5667 56 : if !new_images.is_empty() {
5668 0 : guard
5669 0 : .open_mut()?
5670 0 : .track_new_image_layers(new_images, &self.metrics);
5671 56 : }
5672 :
5673 56 : guard
5674 56 : .open_mut()?
5675 56 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
5676 56 :
5677 56 : self.remote_client
5678 56 : .schedule_compaction_update(&remove_layers, new_deltas)?;
5679 :
5680 56 : drop_wlock(guard);
5681 56 :
5682 56 : Ok(())
5683 56 : }
5684 :
5685 0 : async fn rewrite_layers(
5686 0 : self: &Arc<Self>,
5687 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
5688 0 : mut drop_layers: Vec<Layer>,
5689 0 : ) -> Result<(), CompactionError> {
5690 0 : let mut guard = self.layers.write().await;
5691 :
5692 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
5693 : // to avoid double-removing, and avoid rewriting something that was removed.
5694 0 : replace_layers.retain(|(l, _)| guard.contains(l));
5695 0 : drop_layers.retain(|l| guard.contains(l));
5696 0 :
5697 0 : guard
5698 0 : .open_mut()?
5699 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
5700 0 :
5701 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
5702 0 :
5703 0 : self.remote_client
5704 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
5705 :
5706 0 : Ok(())
5707 0 : }
5708 :
5709 : /// Schedules the uploads of the given image layers
5710 728 : fn upload_new_image_layers(
5711 728 : self: &Arc<Self>,
5712 728 : new_images: impl IntoIterator<Item = ResidentLayer>,
5713 728 : ) -> Result<(), super::upload_queue::NotInitialized> {
5714 780 : for layer in new_images {
5715 52 : self.remote_client.schedule_layer_file_upload(layer)?;
5716 : }
5717 : // should any new image layer been created, not uploading index_part will
5718 : // result in a mismatch between remote_physical_size and layermap calculated
5719 : // size, which will fail some tests, but should not be an issue otherwise.
5720 728 : self.remote_client
5721 728 : .schedule_index_upload_for_file_changes()?;
5722 728 : Ok(())
5723 728 : }
5724 :
5725 0 : async fn find_gc_time_cutoff(
5726 0 : &self,
5727 0 : now: SystemTime,
5728 0 : pitr: Duration,
5729 0 : cancel: &CancellationToken,
5730 0 : ctx: &RequestContext,
5731 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
5732 0 : debug_assert_current_span_has_tenant_and_timeline_id();
5733 0 : if self.shard_identity.is_shard_zero() {
5734 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
5735 0 : let time_range = if pitr == Duration::ZERO {
5736 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
5737 : } else {
5738 0 : pitr
5739 : };
5740 :
5741 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
5742 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
5743 0 : let timestamp = to_pg_timestamp(time_cutoff);
5744 :
5745 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
5746 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
5747 0 : LsnForTimestamp::Future(lsn) => {
5748 0 : // The timestamp is in the future. That sounds impossible,
5749 0 : // but what it really means is that there hasn't been
5750 0 : // any commits since the cutoff timestamp.
5751 0 : //
5752 0 : // In this case we should use the LSN of the most recent commit,
5753 0 : // which is implicitly the last LSN in the log.
5754 0 : debug!("future({})", lsn);
5755 0 : Some(self.get_last_record_lsn())
5756 : }
5757 0 : LsnForTimestamp::Past(lsn) => {
5758 0 : debug!("past({})", lsn);
5759 0 : None
5760 : }
5761 0 : LsnForTimestamp::NoData(lsn) => {
5762 0 : debug!("nodata({})", lsn);
5763 0 : None
5764 : }
5765 : };
5766 0 : Ok(time_cutoff)
5767 : } else {
5768 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
5769 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
5770 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
5771 : // space cutoff that we would also have respected ourselves.
5772 0 : match self
5773 0 : .remote_client
5774 0 : .download_foreign_index(ShardNumber(0), cancel)
5775 0 : .await
5776 : {
5777 0 : Ok((index_part, index_generation, _index_mtime)) => {
5778 0 : tracing::info!(
5779 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
5780 0 : index_part.metadata.latest_gc_cutoff_lsn()
5781 : );
5782 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
5783 : }
5784 : Err(DownloadError::NotFound) => {
5785 : // This is unexpected, because during timeline creations shard zero persists to remote
5786 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
5787 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
5788 : // state, then shard zero should _eventually_ write an index when it GCs.
5789 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
5790 0 : Ok(None)
5791 : }
5792 0 : Err(e) => {
5793 0 : // TODO: this function should return a different error type than page reconstruct error
5794 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
5795 : }
5796 : }
5797 :
5798 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
5799 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
5800 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
5801 : // new location. This is legal in principle, but problematic in practice because it might result
5802 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
5803 : // because they have GC'd past the branch point.
5804 : }
5805 0 : }
5806 :
5807 : /// Find the Lsns above which layer files need to be retained on
5808 : /// garbage collection.
5809 : ///
5810 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
5811 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
5812 : /// the space-based retention.
5813 : ///
5814 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
5815 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
5816 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
5817 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
5818 : /// in the response as the GC cutoff point for the timeline.
5819 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
5820 : pub(super) async fn find_gc_cutoffs(
5821 : &self,
5822 : now: SystemTime,
5823 : space_cutoff: Lsn,
5824 : pitr: Duration,
5825 : cancel: &CancellationToken,
5826 : ctx: &RequestContext,
5827 : ) -> Result<GcCutoffs, PageReconstructError> {
5828 : let _timer = self
5829 : .metrics
5830 : .find_gc_cutoffs_histo
5831 : .start_timer()
5832 : .record_on_drop();
5833 :
5834 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
5835 :
5836 : if cfg!(test) {
5837 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
5838 : if pitr == Duration::ZERO {
5839 : return Ok(GcCutoffs {
5840 : time: self.get_last_record_lsn(),
5841 : space: space_cutoff,
5842 : });
5843 : }
5844 : }
5845 :
5846 : // Calculate a time-based limit on how much to retain:
5847 : // - if PITR interval is set, then this is our cutoff.
5848 : // - if PITR interval is not set, then we do a lookup
5849 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
5850 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
5851 :
5852 : Ok(match (pitr, time_cutoff) {
5853 : (Duration::ZERO, Some(time_cutoff)) => {
5854 : // PITR is not set. Retain the size-based limit, or the default time retention,
5855 : // whichever requires less data.
5856 : GcCutoffs {
5857 : time: self.get_last_record_lsn(),
5858 : space: std::cmp::max(time_cutoff, space_cutoff),
5859 : }
5860 : }
5861 : (Duration::ZERO, None) => {
5862 : // PITR is not set, and time lookup failed
5863 : GcCutoffs {
5864 : time: self.get_last_record_lsn(),
5865 : space: space_cutoff,
5866 : }
5867 : }
5868 : (_, None) => {
5869 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
5870 : // cannot advance beyond what was already GC'd, and respect space-based retention
5871 : GcCutoffs {
5872 : time: *self.get_applied_gc_cutoff_lsn(),
5873 : space: space_cutoff,
5874 : }
5875 : }
5876 : (_, Some(time_cutoff)) => {
5877 : // PITR interval is set and we looked up timestamp successfully. Ignore
5878 : // size based retention and make time cutoff authoritative
5879 : GcCutoffs {
5880 : time: time_cutoff,
5881 : space: time_cutoff,
5882 : }
5883 : }
5884 : })
5885 : }
5886 :
5887 : /// Garbage collect layer files on a timeline that are no longer needed.
5888 : ///
5889 : /// Currently, we don't make any attempt at removing unneeded page versions
5890 : /// within a layer file. We can only remove the whole file if it's fully
5891 : /// obsolete.
5892 8 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
5893 : // this is most likely the background tasks, but it might be the spawned task from
5894 : // immediate_gc
5895 8 : let _g = tokio::select! {
5896 8 : guard = self.gc_lock.lock() => guard,
5897 8 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
5898 : };
5899 8 : let timer = self.metrics.garbage_collect_histo.start_timer();
5900 8 :
5901 8 : fail_point!("before-timeline-gc");
5902 8 :
5903 8 : // Is the timeline being deleted?
5904 8 : if self.is_stopping() {
5905 0 : return Err(GcError::TimelineCancelled);
5906 8 : }
5907 8 :
5908 8 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
5909 8 : let gc_info = self.gc_info.read().unwrap();
5910 8 :
5911 8 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
5912 8 : let time_cutoff = gc_info.cutoffs.time;
5913 8 : let retain_lsns = gc_info
5914 8 : .retain_lsns
5915 8 : .iter()
5916 8 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
5917 8 : .collect();
5918 8 :
5919 8 : // Gets the maximum LSN that holds the valid lease.
5920 8 : //
5921 8 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
5922 8 : // Here, we do not check for stale leases again.
5923 8 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
5924 8 :
5925 8 : (
5926 8 : space_cutoff,
5927 8 : time_cutoff,
5928 8 : retain_lsns,
5929 8 : max_lsn_with_valid_lease,
5930 8 : )
5931 8 : };
5932 8 :
5933 8 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
5934 8 : let standby_horizon = self.standby_horizon.load();
5935 8 : // Hold GC for the standby, but as a safety guard do it only within some
5936 8 : // reasonable lag.
5937 8 : if standby_horizon != Lsn::INVALID {
5938 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
5939 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
5940 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
5941 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
5942 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
5943 : } else {
5944 0 : warn!(
5945 0 : "standby is lagging for more than {}MB, not holding gc for it",
5946 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
5947 : )
5948 : }
5949 0 : }
5950 8 : }
5951 :
5952 : // Reset standby horizon to ignore it if it is not updated till next GC.
5953 : // It is an easy way to unset it when standby disappears without adding
5954 : // more conf options.
5955 8 : self.standby_horizon.store(Lsn::INVALID);
5956 8 : self.metrics
5957 8 : .standby_horizon_gauge
5958 8 : .set(Lsn::INVALID.0 as i64);
5959 :
5960 8 : let res = self
5961 8 : .gc_timeline(
5962 8 : space_cutoff,
5963 8 : time_cutoff,
5964 8 : retain_lsns,
5965 8 : max_lsn_with_valid_lease,
5966 8 : new_gc_cutoff,
5967 8 : )
5968 8 : .instrument(
5969 8 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
5970 : )
5971 8 : .await?;
5972 :
5973 : // only record successes
5974 8 : timer.stop_and_record();
5975 8 :
5976 8 : Ok(res)
5977 8 : }
5978 :
5979 8 : async fn gc_timeline(
5980 8 : &self,
5981 8 : space_cutoff: Lsn,
5982 8 : time_cutoff: Lsn,
5983 8 : retain_lsns: Vec<Lsn>,
5984 8 : max_lsn_with_valid_lease: Option<Lsn>,
5985 8 : new_gc_cutoff: Lsn,
5986 8 : ) -> Result<GcResult, GcError> {
5987 8 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
5988 8 :
5989 8 : let now = SystemTime::now();
5990 8 : let mut result: GcResult = GcResult::default();
5991 8 :
5992 8 : // Nothing to GC. Return early.
5993 8 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
5994 8 : if latest_gc_cutoff >= new_gc_cutoff {
5995 0 : info!(
5996 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
5997 : );
5998 0 : return Ok(result);
5999 8 : }
6000 :
6001 : // We need to ensure that no one tries to read page versions or create
6002 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6003 : // for details. This will block until the old value is no longer in use.
6004 : //
6005 : // The GC cutoff should only ever move forwards.
6006 8 : let waitlist = {
6007 8 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6008 8 : if *write_guard > new_gc_cutoff {
6009 0 : return Err(GcError::BadLsn {
6010 0 : why: format!(
6011 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6012 0 : *write_guard, new_gc_cutoff
6013 0 : ),
6014 0 : });
6015 8 : }
6016 8 :
6017 8 : write_guard.store_and_unlock(new_gc_cutoff)
6018 8 : };
6019 8 : waitlist.wait().await;
6020 :
6021 8 : info!("GC starting");
6022 :
6023 8 : debug!("retain_lsns: {:?}", retain_lsns);
6024 :
6025 8 : let mut layers_to_remove = Vec::new();
6026 :
6027 : // Scan all layers in the timeline (remote or on-disk).
6028 : //
6029 : // Garbage collect the layer if all conditions are satisfied:
6030 : // 1. it is older than cutoff LSN;
6031 : // 2. it is older than PITR interval;
6032 : // 3. it doesn't need to be retained for 'retain_lsns';
6033 : // 4. it does not need to be kept for LSNs holding valid leases.
6034 : // 5. newer on-disk image layers cover the layer's whole key range
6035 : //
6036 : // TODO holding a write lock is too agressive and avoidable
6037 8 : let mut guard = self.layers.write().await;
6038 8 : let layers = guard.layer_map()?;
6039 48 : 'outer: for l in layers.iter_historic_layers() {
6040 48 : result.layers_total += 1;
6041 48 :
6042 48 : // 1. Is it newer than GC horizon cutoff point?
6043 48 : if l.get_lsn_range().end > space_cutoff {
6044 4 : info!(
6045 0 : "keeping {} because it's newer than space_cutoff {}",
6046 0 : l.layer_name(),
6047 : space_cutoff,
6048 : );
6049 4 : result.layers_needed_by_cutoff += 1;
6050 4 : continue 'outer;
6051 44 : }
6052 44 :
6053 44 : // 2. It is newer than PiTR cutoff point?
6054 44 : if l.get_lsn_range().end > time_cutoff {
6055 0 : info!(
6056 0 : "keeping {} because it's newer than time_cutoff {}",
6057 0 : l.layer_name(),
6058 : time_cutoff,
6059 : );
6060 0 : result.layers_needed_by_pitr += 1;
6061 0 : continue 'outer;
6062 44 : }
6063 :
6064 : // 3. Is it needed by a child branch?
6065 : // NOTE With that we would keep data that
6066 : // might be referenced by child branches forever.
6067 : // We can track this in child timeline GC and delete parent layers when
6068 : // they are no longer needed. This might be complicated with long inheritance chains.
6069 : //
6070 : // TODO Vec is not a great choice for `retain_lsns`
6071 44 : for retain_lsn in &retain_lsns {
6072 : // start_lsn is inclusive
6073 0 : if &l.get_lsn_range().start <= retain_lsn {
6074 0 : info!(
6075 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6076 0 : l.layer_name(),
6077 0 : retain_lsn,
6078 0 : l.is_incremental(),
6079 : );
6080 0 : result.layers_needed_by_branches += 1;
6081 0 : continue 'outer;
6082 0 : }
6083 : }
6084 :
6085 : // 4. Is there a valid lease that requires us to keep this layer?
6086 44 : if let Some(lsn) = &max_lsn_with_valid_lease {
6087 : // keep if layer start <= any of the lease
6088 36 : if &l.get_lsn_range().start <= lsn {
6089 28 : info!(
6090 0 : "keeping {} because there is a valid lease preventing GC at {}",
6091 0 : l.layer_name(),
6092 : lsn,
6093 : );
6094 28 : result.layers_needed_by_leases += 1;
6095 28 : continue 'outer;
6096 8 : }
6097 8 : }
6098 :
6099 : // 5. Is there a later on-disk layer for this relation?
6100 : //
6101 : // The end-LSN is exclusive, while disk_consistent_lsn is
6102 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6103 : // OK for a delta layer to have end LSN 101, but if the end LSN
6104 : // is 102, then it might not have been fully flushed to disk
6105 : // before crash.
6106 : //
6107 : // For example, imagine that the following layers exist:
6108 : //
6109 : // 1000 - image (A)
6110 : // 1000-2000 - delta (B)
6111 : // 2000 - image (C)
6112 : // 2000-3000 - delta (D)
6113 : // 3000 - image (E)
6114 : //
6115 : // If GC horizon is at 2500, we can remove layers A and B, but
6116 : // we cannot remove C, even though it's older than 2500, because
6117 : // the delta layer 2000-3000 depends on it.
6118 16 : if !layers
6119 16 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6120 : {
6121 12 : info!("keeping {} because it is the latest layer", l.layer_name());
6122 12 : result.layers_not_updated += 1;
6123 12 : continue 'outer;
6124 4 : }
6125 4 :
6126 4 : // We didn't find any reason to keep this file, so remove it.
6127 4 : info!(
6128 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6129 0 : l.layer_name(),
6130 0 : l.is_incremental(),
6131 : );
6132 4 : layers_to_remove.push(l);
6133 : }
6134 :
6135 8 : if !layers_to_remove.is_empty() {
6136 : // Persist the new GC cutoff value before we actually remove anything.
6137 : // This unconditionally schedules also an index_part.json update, even though, we will
6138 : // be doing one a bit later with the unlinked gc'd layers.
6139 4 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6140 4 : self.schedule_uploads(disk_consistent_lsn, None)
6141 4 : .map_err(|e| {
6142 0 : if self.cancel.is_cancelled() {
6143 0 : GcError::TimelineCancelled
6144 : } else {
6145 0 : GcError::Remote(e)
6146 : }
6147 4 : })?;
6148 :
6149 4 : let gc_layers = layers_to_remove
6150 4 : .iter()
6151 4 : .map(|x| guard.get_from_desc(x))
6152 4 : .collect::<Vec<Layer>>();
6153 4 :
6154 4 : result.layers_removed = gc_layers.len() as u64;
6155 4 :
6156 4 : self.remote_client.schedule_gc_update(&gc_layers)?;
6157 :
6158 4 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
6159 4 :
6160 4 : #[cfg(feature = "testing")]
6161 4 : {
6162 4 : result.doomed_layers = gc_layers;
6163 4 : }
6164 4 : }
6165 :
6166 8 : info!(
6167 0 : "GC completed removing {} layers, cutoff {}",
6168 : result.layers_removed, new_gc_cutoff
6169 : );
6170 :
6171 8 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
6172 8 : Ok(result)
6173 8 : }
6174 :
6175 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6176 1340058 : async fn reconstruct_value(
6177 1340058 : &self,
6178 1340058 : key: Key,
6179 1340058 : request_lsn: Lsn,
6180 1340058 : mut data: ValueReconstructState,
6181 1340058 : ) -> Result<Bytes, PageReconstructError> {
6182 1340058 : // Perform WAL redo if needed
6183 1340058 : data.records.reverse();
6184 1340058 :
6185 1340058 : // If we have a page image, and no WAL, we're all set
6186 1340058 : if data.records.is_empty() {
6187 1338382 : if let Some((img_lsn, img)) = &data.img {
6188 1338382 : trace!(
6189 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6190 : key, img_lsn, request_lsn,
6191 : );
6192 1338382 : Ok(img.clone())
6193 : } else {
6194 0 : Err(PageReconstructError::from(anyhow!(
6195 0 : "base image for {key} at {request_lsn} not found"
6196 0 : )))
6197 : }
6198 : } else {
6199 : // We need to do WAL redo.
6200 : //
6201 : // If we don't have a base image, then the oldest WAL record better initialize
6202 : // the page
6203 1676 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6204 0 : Err(PageReconstructError::from(anyhow!(
6205 0 : "Base image for {} at {} not found, but got {} WAL records",
6206 0 : key,
6207 0 : request_lsn,
6208 0 : data.records.len()
6209 0 : )))
6210 : } else {
6211 1676 : if data.img.is_some() {
6212 1536 : trace!(
6213 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6214 0 : data.records.len(),
6215 : key,
6216 : request_lsn
6217 : );
6218 : } else {
6219 140 : trace!(
6220 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
6221 0 : data.records.len(),
6222 : key,
6223 : request_lsn
6224 : );
6225 : };
6226 1676 : let res = self
6227 1676 : .walredo_mgr
6228 1676 : .as_ref()
6229 1676 : .context("timeline has no walredo manager")
6230 1676 : .map_err(PageReconstructError::WalRedo)?
6231 1676 : .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
6232 1676 : .await;
6233 1676 : let img = match res {
6234 1676 : Ok(img) => img,
6235 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6236 0 : Err(walredo::Error::Other(err)) => {
6237 0 : critical!("walredo failure during page reconstruction: {err:?}");
6238 0 : return Err(PageReconstructError::WalRedo(
6239 0 : err.context("reconstruct a page image"),
6240 0 : ));
6241 : }
6242 : };
6243 1676 : Ok(img)
6244 : }
6245 : }
6246 1340058 : }
6247 :
6248 0 : pub(crate) async fn spawn_download_all_remote_layers(
6249 0 : self: Arc<Self>,
6250 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6251 0 : ctx: &RequestContext,
6252 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6253 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6254 :
6255 : // this is not really needed anymore; it has tests which really check the return value from
6256 : // http api. it would be better not to maintain this anymore.
6257 :
6258 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6259 0 : if let Some(st) = &*status_guard {
6260 0 : match &st.state {
6261 : DownloadRemoteLayersTaskState::Running => {
6262 0 : return Err(st.clone());
6263 : }
6264 : DownloadRemoteLayersTaskState::ShutDown
6265 0 : | DownloadRemoteLayersTaskState::Completed => {
6266 0 : *status_guard = None;
6267 0 : }
6268 : }
6269 0 : }
6270 :
6271 0 : let self_clone = Arc::clone(&self);
6272 0 : let task_ctx = ctx.detached_child(
6273 0 : TaskKind::DownloadAllRemoteLayers,
6274 0 : DownloadBehavior::Download,
6275 0 : );
6276 0 : let task_id = task_mgr::spawn(
6277 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6278 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6279 0 : self.tenant_shard_id,
6280 0 : Some(self.timeline_id),
6281 0 : "download all remote layers task",
6282 0 : async move {
6283 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
6284 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6285 0 : match &mut *status_guard {
6286 : None => {
6287 0 : warn!("tasks status is supposed to be Some(), since we are running");
6288 : }
6289 0 : Some(st) => {
6290 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6291 0 : if st.task_id != exp_task_id {
6292 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6293 0 : } else {
6294 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6295 0 : }
6296 : }
6297 : };
6298 0 : Ok(())
6299 0 : }
6300 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))
6301 : );
6302 :
6303 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6304 0 : task_id: format!("{task_id}"),
6305 0 : state: DownloadRemoteLayersTaskState::Running,
6306 0 : total_layer_count: 0,
6307 0 : successful_download_count: 0,
6308 0 : failed_download_count: 0,
6309 0 : };
6310 0 : *status_guard = Some(initial_info.clone());
6311 0 :
6312 0 : Ok(initial_info)
6313 0 : }
6314 :
6315 0 : async fn download_all_remote_layers(
6316 0 : self: &Arc<Self>,
6317 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6318 0 : ctx: &RequestContext,
6319 0 : ) {
6320 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6321 :
6322 0 : let remaining = {
6323 0 : let guard = self.layers.read().await;
6324 0 : let Ok(lm) = guard.layer_map() else {
6325 : // technically here we could look into iterating accessible layers, but downloading
6326 : // all layers of a shutdown timeline makes no sense regardless.
6327 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6328 0 : return;
6329 : };
6330 0 : lm.iter_historic_layers()
6331 0 : .map(|desc| guard.get_from_desc(&desc))
6332 0 : .collect::<Vec<_>>()
6333 0 : };
6334 0 : let total_layer_count = remaining.len();
6335 :
6336 : macro_rules! lock_status {
6337 : ($st:ident) => {
6338 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6339 : let st = st
6340 : .as_mut()
6341 : .expect("this function is only called after the task has been spawned");
6342 : assert_eq!(
6343 : st.task_id,
6344 : format!(
6345 : "{}",
6346 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6347 : )
6348 : );
6349 : let $st = st;
6350 : };
6351 : }
6352 :
6353 : {
6354 0 : lock_status!(st);
6355 0 : st.total_layer_count = total_layer_count as u64;
6356 0 : }
6357 0 :
6358 0 : let mut remaining = remaining.into_iter();
6359 0 : let mut have_remaining = true;
6360 0 : let mut js = tokio::task::JoinSet::new();
6361 0 :
6362 0 : let cancel = task_mgr::shutdown_token();
6363 0 :
6364 0 : let limit = request.max_concurrent_downloads;
6365 :
6366 : loop {
6367 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6368 0 : let Some(next) = remaining.next() else {
6369 0 : have_remaining = false;
6370 0 : break;
6371 : };
6372 :
6373 0 : let span = tracing::info_span!("download", layer = %next);
6374 :
6375 0 : let ctx = ctx.attached_child();
6376 0 : js.spawn(
6377 0 : async move {
6378 0 : let res = next.download(&ctx).await;
6379 0 : (next, res)
6380 0 : }
6381 0 : .instrument(span),
6382 0 : );
6383 0 : }
6384 :
6385 0 : while let Some(res) = js.join_next().await {
6386 0 : match res {
6387 : Ok((_, Ok(_))) => {
6388 0 : lock_status!(st);
6389 0 : st.successful_download_count += 1;
6390 : }
6391 0 : Ok((layer, Err(e))) => {
6392 0 : tracing::error!(%layer, "download failed: {e:#}");
6393 0 : lock_status!(st);
6394 0 : st.failed_download_count += 1;
6395 : }
6396 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6397 0 : Err(je) if je.is_panic() => {
6398 0 : lock_status!(st);
6399 0 : st.failed_download_count += 1;
6400 : }
6401 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6402 : }
6403 : }
6404 :
6405 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6406 0 : break;
6407 0 : }
6408 : }
6409 :
6410 : {
6411 0 : lock_status!(st);
6412 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6413 : }
6414 0 : }
6415 :
6416 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6417 0 : &self,
6418 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6419 0 : self.download_all_remote_layers_task_info
6420 0 : .read()
6421 0 : .unwrap()
6422 0 : .clone()
6423 0 : }
6424 : }
6425 :
6426 : impl Timeline {
6427 : /// Returns non-remote layers for eviction.
6428 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6429 0 : let guard = self.layers.read().await;
6430 0 : let mut max_layer_size: Option<u64> = None;
6431 0 :
6432 0 : let resident_layers = guard
6433 0 : .likely_resident_layers()
6434 0 : .map(|layer| {
6435 0 : let file_size = layer.layer_desc().file_size;
6436 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6437 0 :
6438 0 : let last_activity_ts = layer.latest_activity();
6439 0 :
6440 0 : EvictionCandidate {
6441 0 : layer: layer.to_owned().into(),
6442 0 : last_activity_ts,
6443 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6444 0 : visibility: layer.visibility(),
6445 0 : }
6446 0 : })
6447 0 : .collect();
6448 0 :
6449 0 : DiskUsageEvictionInfo {
6450 0 : max_layer_size,
6451 0 : resident_layers,
6452 0 : }
6453 0 : }
6454 :
6455 3748 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6456 3748 : ShardIndex {
6457 3748 : shard_number: self.tenant_shard_id.shard_number,
6458 3748 : shard_count: self.tenant_shard_id.shard_count,
6459 3748 : }
6460 3748 : }
6461 :
6462 : /// Persistently blocks gc for `Manual` reason.
6463 : ///
6464 : /// Returns true if no such block existed before, false otherwise.
6465 0 : pub(crate) async fn block_gc(&self, tenant: &super::Tenant) -> anyhow::Result<bool> {
6466 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6467 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6468 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6469 0 : }
6470 :
6471 : /// Persistently unblocks gc for `Manual` reason.
6472 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::Tenant) -> anyhow::Result<()> {
6473 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6474 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6475 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6476 0 : }
6477 :
6478 : #[cfg(test)]
6479 100 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6480 100 : self.last_record_lsn.advance(new_lsn);
6481 100 : }
6482 :
6483 : #[cfg(test)]
6484 4 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6485 4 : self.disk_consistent_lsn.store(new_value);
6486 4 : }
6487 :
6488 : /// Force create an image layer and place it into the layer map.
6489 : ///
6490 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6491 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6492 : /// placed into the layer map in one run AND be validated.
6493 : #[cfg(test)]
6494 128 : pub(super) async fn force_create_image_layer(
6495 128 : self: &Arc<Timeline>,
6496 128 : lsn: Lsn,
6497 128 : mut images: Vec<(Key, Bytes)>,
6498 128 : check_start_lsn: Option<Lsn>,
6499 128 : ctx: &RequestContext,
6500 128 : ) -> anyhow::Result<()> {
6501 128 : let last_record_lsn = self.get_last_record_lsn();
6502 128 : assert!(
6503 128 : lsn <= last_record_lsn,
6504 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6505 : );
6506 128 : if let Some(check_start_lsn) = check_start_lsn {
6507 128 : assert!(lsn >= check_start_lsn);
6508 0 : }
6509 380 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6510 128 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6511 128 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6512 128 : let mut image_layer_writer = ImageLayerWriter::new(
6513 128 : self.conf,
6514 128 : self.timeline_id,
6515 128 : self.tenant_shard_id,
6516 128 : &(min_key..end_key),
6517 128 : lsn,
6518 128 : ctx,
6519 128 : )
6520 128 : .await?;
6521 636 : for (key, img) in images {
6522 508 : image_layer_writer.put_image(key, img, ctx).await?;
6523 : }
6524 128 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6525 128 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6526 128 : info!("force created image layer {}", image_layer.local_path());
6527 : {
6528 128 : let mut guard = self.layers.write().await;
6529 128 : guard
6530 128 : .open_mut()
6531 128 : .unwrap()
6532 128 : .force_insert_layer(image_layer.clone());
6533 128 : }
6534 128 :
6535 128 : // Update remote_timeline_client state to reflect existence of this layer
6536 128 : self.remote_client
6537 128 : .schedule_layer_file_upload(image_layer)
6538 128 : .unwrap();
6539 128 :
6540 128 : Ok(())
6541 128 : }
6542 :
6543 : /// Force create a delta layer and place it into the layer map.
6544 : ///
6545 : /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]
6546 : /// or [`Tenant::create_test_timeline_with_layers`] to ensure all these layers are
6547 : /// placed into the layer map in one run AND be validated.
6548 : #[cfg(test)]
6549 184 : pub(super) async fn force_create_delta_layer(
6550 184 : self: &Arc<Timeline>,
6551 184 : mut deltas: DeltaLayerTestDesc,
6552 184 : check_start_lsn: Option<Lsn>,
6553 184 : ctx: &RequestContext,
6554 184 : ) -> anyhow::Result<()> {
6555 184 : let last_record_lsn = self.get_last_record_lsn();
6556 184 : deltas
6557 184 : .data
6558 264 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6559 184 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6560 184 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6561 632 : for (_, lsn, _) in &deltas.data {
6562 448 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6563 : }
6564 184 : assert!(
6565 184 : deltas.lsn_range.end <= last_record_lsn,
6566 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6567 : deltas.lsn_range.end,
6568 : last_record_lsn
6569 : );
6570 184 : if let Some(check_start_lsn) = check_start_lsn {
6571 184 : assert!(deltas.lsn_range.start >= check_start_lsn);
6572 0 : }
6573 184 : let mut delta_layer_writer = DeltaLayerWriter::new(
6574 184 : self.conf,
6575 184 : self.timeline_id,
6576 184 : self.tenant_shard_id,
6577 184 : deltas.key_range.start,
6578 184 : deltas.lsn_range,
6579 184 : ctx,
6580 184 : )
6581 184 : .await?;
6582 632 : for (key, lsn, val) in deltas.data {
6583 448 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6584 : }
6585 184 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6586 184 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6587 184 : info!("force created delta layer {}", delta_layer.local_path());
6588 : {
6589 184 : let mut guard = self.layers.write().await;
6590 184 : guard
6591 184 : .open_mut()
6592 184 : .unwrap()
6593 184 : .force_insert_layer(delta_layer.clone());
6594 184 : }
6595 184 :
6596 184 : // Update remote_timeline_client state to reflect existence of this layer
6597 184 : self.remote_client
6598 184 : .schedule_layer_file_upload(delta_layer)
6599 184 : .unwrap();
6600 184 :
6601 184 : Ok(())
6602 184 : }
6603 :
6604 : /// Force create an in-memory layer and place them into the layer map.
6605 : #[cfg(test)]
6606 8 : pub(super) async fn force_create_in_memory_layer(
6607 8 : self: &Arc<Timeline>,
6608 8 : mut in_memory: InMemoryLayerTestDesc,
6609 8 : check_start_lsn: Option<Lsn>,
6610 8 : ctx: &RequestContext,
6611 8 : ) -> anyhow::Result<()> {
6612 : use utils::bin_ser::BeSer;
6613 :
6614 : // Validate LSNs
6615 8 : if let Some(check_start_lsn) = check_start_lsn {
6616 8 : assert!(in_memory.lsn_range.start >= check_start_lsn);
6617 0 : }
6618 :
6619 8 : let last_record_lsn = self.get_last_record_lsn();
6620 8 : let layer_end_lsn = if in_memory.is_open {
6621 4 : in_memory
6622 4 : .data
6623 4 : .iter()
6624 40 : .map(|(_key, lsn, _value)| lsn)
6625 4 : .max()
6626 4 : .cloned()
6627 : } else {
6628 4 : Some(in_memory.lsn_range.end)
6629 : };
6630 :
6631 8 : if let Some(end) = layer_end_lsn {
6632 8 : assert!(
6633 8 : end <= last_record_lsn,
6634 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6635 : end,
6636 : last_record_lsn,
6637 : );
6638 0 : }
6639 :
6640 80 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
6641 80 : assert!(*lsn >= in_memory.lsn_range.start);
6642 80 : assert!(*lsn < in_memory.lsn_range.end);
6643 80 : });
6644 8 :
6645 8 : // Build the batch
6646 8 : in_memory
6647 8 : .data
6648 72 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6649 8 :
6650 8 : let data = in_memory
6651 8 : .data
6652 8 : .into_iter()
6653 80 : .map(|(key, lsn, value)| {
6654 80 : let value_size = value.serialized_size().unwrap() as usize;
6655 80 : (key.to_compact(), lsn, value_size, value)
6656 80 : })
6657 8 : .collect::<Vec<_>>();
6658 8 :
6659 8 : let batch = SerializedValueBatch::from_values(data);
6660 :
6661 : // Create the in-memory layer and write the batch into it
6662 8 : let layer = InMemoryLayer::create(
6663 8 : self.conf,
6664 8 : self.timeline_id,
6665 8 : self.tenant_shard_id,
6666 8 : in_memory.lsn_range.start,
6667 8 : &self.gate,
6668 8 : ctx,
6669 8 : )
6670 8 : .await
6671 8 : .unwrap();
6672 8 :
6673 8 : layer.put_batch(batch, ctx).await.unwrap();
6674 8 : if !in_memory.is_open {
6675 4 : layer.freeze(in_memory.lsn_range.end).await;
6676 4 : }
6677 :
6678 8 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
6679 :
6680 : // Link the layer to the layer map
6681 : {
6682 8 : let mut guard = self.layers.write().await;
6683 8 : let layer_map = guard.open_mut().unwrap();
6684 8 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
6685 8 : }
6686 8 :
6687 8 : Ok(())
6688 8 : }
6689 :
6690 : /// Return all keys at the LSN in the image layers
6691 : #[cfg(test)]
6692 12 : pub(crate) async fn inspect_image_layers(
6693 12 : self: &Arc<Timeline>,
6694 12 : lsn: Lsn,
6695 12 : ctx: &RequestContext,
6696 12 : io_concurrency: IoConcurrency,
6697 12 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
6698 12 : let mut all_data = Vec::new();
6699 12 : let guard = self.layers.read().await;
6700 68 : for layer in guard.layer_map()?.iter_historic_layers() {
6701 68 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
6702 16 : let layer = guard.get_from_desc(&layer);
6703 16 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
6704 16 : layer
6705 16 : .get_values_reconstruct_data(
6706 16 : KeySpace::single(Key::MIN..Key::MAX),
6707 16 : lsn..Lsn(lsn.0 + 1),
6708 16 : &mut reconstruct_data,
6709 16 : ctx,
6710 16 : )
6711 16 : .await?;
6712 132 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
6713 132 : let v = v.collect_pending_ios().await?;
6714 132 : all_data.push((k, v.img.unwrap().1));
6715 : }
6716 52 : }
6717 : }
6718 12 : all_data.sort();
6719 12 : Ok(all_data)
6720 12 : }
6721 :
6722 : /// Get all historic layer descriptors in the layer map
6723 : #[cfg(test)]
6724 48 : pub(crate) async fn inspect_historic_layers(
6725 48 : self: &Arc<Timeline>,
6726 48 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
6727 48 : let mut layers = Vec::new();
6728 48 : let guard = self.layers.read().await;
6729 228 : for layer in guard.layer_map()?.iter_historic_layers() {
6730 228 : layers.push(layer.key());
6731 228 : }
6732 48 : Ok(layers)
6733 48 : }
6734 :
6735 : #[cfg(test)]
6736 20 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
6737 20 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
6738 20 : keyspace.merge(&ks);
6739 20 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
6740 20 : }
6741 : }
6742 :
6743 : /// Tracking writes ingestion does to a particular in-memory layer.
6744 : ///
6745 : /// Cleared upon freezing a layer.
6746 : pub(crate) struct TimelineWriterState {
6747 : open_layer: Arc<InMemoryLayer>,
6748 : current_size: u64,
6749 : // Previous Lsn which passed through
6750 : prev_lsn: Option<Lsn>,
6751 : // Largest Lsn which passed through the current writer
6752 : max_lsn: Option<Lsn>,
6753 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
6754 : cached_last_freeze_at: Lsn,
6755 : }
6756 :
6757 : impl TimelineWriterState {
6758 2604 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
6759 2604 : Self {
6760 2604 : open_layer,
6761 2604 : current_size,
6762 2604 : prev_lsn: None,
6763 2604 : max_lsn: None,
6764 2604 : cached_last_freeze_at: last_freeze_at,
6765 2604 : }
6766 2604 : }
6767 : }
6768 :
6769 : /// Various functions to mutate the timeline.
6770 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
6771 : // This is probably considered a bad practice in Rust and should be fixed eventually,
6772 : // but will cause large code changes.
6773 : pub(crate) struct TimelineWriter<'a> {
6774 : tl: &'a Timeline,
6775 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
6776 : }
6777 :
6778 : impl Deref for TimelineWriter<'_> {
6779 : type Target = Timeline;
6780 :
6781 19796508 : fn deref(&self) -> &Self::Target {
6782 19796508 : self.tl
6783 19796508 : }
6784 : }
6785 :
6786 : #[derive(PartialEq)]
6787 : enum OpenLayerAction {
6788 : Roll,
6789 : Open,
6790 : None,
6791 : }
6792 :
6793 : impl TimelineWriter<'_> {
6794 9608476 : async fn handle_open_layer_action(
6795 9608476 : &mut self,
6796 9608476 : at: Lsn,
6797 9608476 : action: OpenLayerAction,
6798 9608476 : ctx: &RequestContext,
6799 9608476 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
6800 9608476 : match action {
6801 : OpenLayerAction::Roll => {
6802 160 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
6803 160 : self.roll_layer(freeze_at).await?;
6804 160 : self.open_layer(at, ctx).await?;
6805 : }
6806 2444 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
6807 : OpenLayerAction::None => {
6808 9605872 : assert!(self.write_guard.is_some());
6809 : }
6810 : }
6811 :
6812 9608476 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
6813 9608476 : }
6814 :
6815 2604 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
6816 2604 : let layer = self
6817 2604 : .tl
6818 2604 : .get_layer_for_write(at, &self.write_guard, ctx)
6819 2604 : .await?;
6820 2604 : let initial_size = layer.size().await?;
6821 :
6822 2604 : let last_freeze_at = self.last_freeze_at.load();
6823 2604 : self.write_guard.replace(TimelineWriterState::new(
6824 2604 : layer,
6825 2604 : initial_size,
6826 2604 : last_freeze_at,
6827 2604 : ));
6828 2604 :
6829 2604 : Ok(())
6830 2604 : }
6831 :
6832 160 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
6833 160 : let current_size = self.write_guard.as_ref().unwrap().current_size;
6834 :
6835 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
6836 : // to propagate the backpressure up into WAL ingestion.
6837 160 : let l0_count = self
6838 160 : .tl
6839 160 : .layers
6840 160 : .read()
6841 160 : .await
6842 160 : .layer_map()?
6843 160 : .level0_deltas()
6844 160 : .len();
6845 160 : let wait_thresholds = [
6846 160 : self.get_l0_flush_delay_threshold(),
6847 160 : self.get_l0_flush_stall_threshold(),
6848 160 : ];
6849 160 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
6850 :
6851 : // self.write_guard will be taken by the freezing
6852 160 : let flush_id = self
6853 160 : .tl
6854 160 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
6855 160 : .await?;
6856 :
6857 160 : assert!(self.write_guard.is_none());
6858 :
6859 160 : if let Some(wait_threshold) = wait_threshold {
6860 0 : if l0_count >= wait_threshold {
6861 0 : debug!(
6862 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
6863 : );
6864 0 : self.tl.wait_flush_completion(flush_id).await?;
6865 0 : }
6866 160 : }
6867 :
6868 160 : if current_size >= self.get_checkpoint_distance() * 2 {
6869 0 : warn!("Flushed oversized open layer with size {}", current_size)
6870 160 : }
6871 :
6872 160 : Ok(())
6873 160 : }
6874 :
6875 9608476 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
6876 9608476 : let state = &*self.write_guard;
6877 9608476 : let Some(state) = &state else {
6878 2444 : return OpenLayerAction::Open;
6879 : };
6880 :
6881 : #[cfg(feature = "testing")]
6882 9606032 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
6883 : // this check and assertion are not really needed because
6884 : // LayerManager::try_freeze_in_memory_layer will always clear out the
6885 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
6886 : // is no TimelineWriterState.
6887 0 : assert!(
6888 0 : state.open_layer.end_lsn.get().is_some(),
6889 0 : "our open_layer must be outdated"
6890 : );
6891 :
6892 : // this would be a memory leak waiting to happen because the in-memory layer always has
6893 : // an index
6894 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
6895 9606032 : }
6896 9606032 :
6897 9606032 : if state.prev_lsn == Some(lsn) {
6898 : // Rolling mid LSN is not supported by [downstream code].
6899 : // Hence, only roll at LSN boundaries.
6900 : //
6901 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
6902 12 : return OpenLayerAction::None;
6903 9606020 : }
6904 9606020 :
6905 9606020 : if state.current_size == 0 {
6906 : // Don't roll empty layers
6907 0 : return OpenLayerAction::None;
6908 9606020 : }
6909 9606020 :
6910 9606020 : if self.tl.should_roll(
6911 9606020 : state.current_size,
6912 9606020 : state.current_size + new_value_size,
6913 9606020 : self.get_checkpoint_distance(),
6914 9606020 : lsn,
6915 9606020 : state.cached_last_freeze_at,
6916 9606020 : state.open_layer.get_opened_at(),
6917 9606020 : ) {
6918 160 : OpenLayerAction::Roll
6919 : } else {
6920 9605860 : OpenLayerAction::None
6921 : }
6922 9608476 : }
6923 :
6924 : /// Put a batch of keys at the specified Lsns.
6925 9608472 : pub(crate) async fn put_batch(
6926 9608472 : &mut self,
6927 9608472 : batch: SerializedValueBatch,
6928 9608472 : ctx: &RequestContext,
6929 9608472 : ) -> anyhow::Result<()> {
6930 9608472 : if !batch.has_data() {
6931 0 : return Ok(());
6932 9608472 : }
6933 9608472 :
6934 9608472 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
6935 9608472 : // We don't assert this in release builds, since key ownership policies may change over
6936 9608472 : // time. Stray keys will be removed during compaction.
6937 9608472 : if cfg!(debug_assertions) {
6938 19789888 : for metadata in &batch.metadata {
6939 10181416 : if let ValueMeta::Serialized(metadata) = metadata {
6940 10181416 : let key = Key::from_compact(metadata.key);
6941 10181416 : assert!(
6942 10181416 : self.shard_identity.is_key_local(&key)
6943 0 : || self.shard_identity.is_key_global(&key),
6944 0 : "key {key} does not belong on shard {}",
6945 0 : self.shard_identity.shard_index()
6946 : );
6947 0 : }
6948 : }
6949 0 : }
6950 :
6951 9608472 : let batch_max_lsn = batch.max_lsn;
6952 9608472 : let buf_size: u64 = batch.buffer_size() as u64;
6953 9608472 :
6954 9608472 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
6955 9608472 : let layer = self
6956 9608472 : .handle_open_layer_action(batch_max_lsn, action, ctx)
6957 9608472 : .await?;
6958 :
6959 9608472 : let res = layer.put_batch(batch, ctx).await;
6960 :
6961 9608472 : if res.is_ok() {
6962 9608472 : // Update the current size only when the entire write was ok.
6963 9608472 : // In case of failures, we may have had partial writes which
6964 9608472 : // render the size tracking out of sync. That's ok because
6965 9608472 : // the checkpoint distance should be significantly smaller
6966 9608472 : // than the S3 single shot upload limit of 5GiB.
6967 9608472 : let state = self.write_guard.as_mut().unwrap();
6968 9608472 :
6969 9608472 : state.current_size += buf_size;
6970 9608472 : state.prev_lsn = Some(batch_max_lsn);
6971 9608472 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
6972 9608472 : }
6973 :
6974 9608472 : res
6975 9608472 : }
6976 :
6977 : #[cfg(test)]
6978 : /// Test helper, for tests that would like to poke individual values without composing a batch
6979 8780308 : pub(crate) async fn put(
6980 8780308 : &mut self,
6981 8780308 : key: Key,
6982 8780308 : lsn: Lsn,
6983 8780308 : value: &Value,
6984 8780308 : ctx: &RequestContext,
6985 8780308 : ) -> anyhow::Result<()> {
6986 : use utils::bin_ser::BeSer;
6987 8780308 : if !key.is_valid_key_on_write_path() {
6988 0 : bail!(
6989 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
6990 0 : key
6991 0 : );
6992 8780308 : }
6993 8780308 : let val_ser_size = value.serialized_size().unwrap() as usize;
6994 8780308 : let batch = SerializedValueBatch::from_values(vec![(
6995 8780308 : key.to_compact(),
6996 8780308 : lsn,
6997 8780308 : val_ser_size,
6998 8780308 : value.clone(),
6999 8780308 : )]);
7000 8780308 :
7001 8780308 : self.put_batch(batch, ctx).await
7002 8780308 : }
7003 :
7004 4 : pub(crate) async fn delete_batch(
7005 4 : &mut self,
7006 4 : batch: &[(Range<Key>, Lsn)],
7007 4 : ctx: &RequestContext,
7008 4 : ) -> anyhow::Result<()> {
7009 4 : if let Some((_, lsn)) = batch.first() {
7010 4 : let action = self.get_open_layer_action(*lsn, 0);
7011 4 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7012 4 : layer.put_tombstones(batch).await?;
7013 0 : }
7014 :
7015 4 : Ok(())
7016 4 : }
7017 :
7018 : /// Track the end of the latest digested WAL record.
7019 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7020 : ///
7021 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7022 : ///
7023 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7024 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7025 : /// the latest and can be read.
7026 10558196 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7027 10558196 : self.tl.finish_write(new_lsn);
7028 10558196 : }
7029 :
7030 541140 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7031 541140 : self.tl.update_current_logical_size(delta)
7032 541140 : }
7033 : }
7034 :
7035 : // We need TimelineWriter to be send in upcoming conversion of
7036 : // Timeline::layers to tokio::sync::RwLock.
7037 : #[test]
7038 4 : fn is_send() {
7039 4 : fn _assert_send<T: Send>() {}
7040 4 : _assert_send::<TimelineWriter<'_>>();
7041 4 : }
7042 :
7043 : #[cfg(test)]
7044 : mod tests {
7045 : use std::sync::Arc;
7046 :
7047 : use pageserver_api::key::Key;
7048 : use pageserver_api::value::Value;
7049 : use std::iter::Iterator;
7050 : use tracing::Instrument;
7051 : use utils::id::TimelineId;
7052 : use utils::lsn::Lsn;
7053 :
7054 : use super::HeatMapTimeline;
7055 : use crate::context::RequestContextBuilder;
7056 : use crate::tenant::harness::{TenantHarness, test_img};
7057 : use crate::tenant::layer_map::LayerMap;
7058 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7059 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7060 : use crate::tenant::{PreviousHeatmap, Timeline};
7061 :
7062 20 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7063 20 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7064 20 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7065 100 : for (l, r) in lhs_rhs {
7066 80 : assert_eq!(l.name, r.name);
7067 80 : assert_eq!(l.metadata, r.metadata);
7068 : }
7069 20 : }
7070 :
7071 : #[tokio::test]
7072 4 : async fn test_heatmap_generation() {
7073 4 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7074 4 :
7075 4 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7076 4 : Lsn(0x10)..Lsn(0x20),
7077 4 : vec![(
7078 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7079 4 : Lsn(0x11),
7080 4 : Value::Image(test_img("foo")),
7081 4 : )],
7082 4 : );
7083 4 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7084 4 : Lsn(0x10)..Lsn(0x20),
7085 4 : vec![(
7086 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7087 4 : Lsn(0x11),
7088 4 : Value::Image(test_img("foo")),
7089 4 : )],
7090 4 : );
7091 4 : let l0_delta = DeltaLayerTestDesc::new(
7092 4 : Lsn(0x20)..Lsn(0x30),
7093 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7094 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7095 4 : vec![(
7096 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7097 4 : Lsn(0x25),
7098 4 : Value::Image(test_img("foo")),
7099 4 : )],
7100 4 : );
7101 4 : let delta_layers = vec![
7102 4 : covered_delta.clone(),
7103 4 : visible_delta.clone(),
7104 4 : l0_delta.clone(),
7105 4 : ];
7106 4 :
7107 4 : let image_layer = (
7108 4 : Lsn(0x40),
7109 4 : vec![(
7110 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7111 4 : test_img("bar"),
7112 4 : )],
7113 4 : );
7114 4 : let image_layers = vec![image_layer];
7115 4 :
7116 4 : let (tenant, ctx) = harness.load().await;
7117 4 : let timeline = tenant
7118 4 : .create_test_timeline_with_layers(
7119 4 : TimelineId::generate(),
7120 4 : Lsn(0x10),
7121 4 : 14,
7122 4 : &ctx,
7123 4 : Vec::new(), // in-memory layers
7124 4 : delta_layers,
7125 4 : image_layers,
7126 4 : Lsn(0x100),
7127 4 : )
7128 4 : .await
7129 4 : .unwrap();
7130 4 : let ctx = &ctx.with_scope_timeline(&timeline);
7131 4 :
7132 4 : // Layer visibility is an input to heatmap generation, so refresh it first
7133 4 : timeline.update_layer_visibility().await.unwrap();
7134 4 :
7135 4 : let heatmap = timeline
7136 4 : .generate_heatmap()
7137 4 : .await
7138 4 : .expect("Infallible while timeline is not shut down");
7139 4 :
7140 4 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
7141 4 :
7142 4 : // L0 should come last
7143 4 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
7144 4 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
7145 4 :
7146 4 : let mut last_lsn = Lsn::MAX;
7147 20 : for layer in heatmap_layers {
7148 4 : // Covered layer should be omitted
7149 16 : assert!(layer.name != covered_delta.layer_name());
7150 4 :
7151 16 : let layer_lsn = match &layer.name {
7152 8 : LayerName::Delta(d) => d.lsn_range.end,
7153 8 : LayerName::Image(i) => i.lsn,
7154 4 : };
7155 4 :
7156 4 : // Apart from L0s, newest Layers should come first
7157 16 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
7158 12 : assert!(layer_lsn <= last_lsn);
7159 12 : last_lsn = layer_lsn;
7160 4 : }
7161 4 : }
7162 4 :
7163 4 : // Evict all the layers and stash the old heatmap in the timeline.
7164 4 : // This simulates a migration to a cold secondary location.
7165 4 :
7166 4 : let guard = timeline.layers.read().await;
7167 4 : let mut all_layers = Vec::new();
7168 4 : let forever = std::time::Duration::from_secs(120);
7169 20 : for layer in guard.likely_resident_layers() {
7170 20 : all_layers.push(layer.clone());
7171 20 : layer.evict_and_wait(forever).await.unwrap();
7172 4 : }
7173 4 : drop(guard);
7174 4 :
7175 4 : timeline
7176 4 : .previous_heatmap
7177 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
7178 4 : heatmap: heatmap.clone(),
7179 4 : read_at: std::time::Instant::now(),
7180 4 : end_lsn: None,
7181 4 : })));
7182 4 :
7183 4 : // Generate a new heatmap and assert that it contains the same layers as the old one.
7184 4 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
7185 4 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
7186 4 :
7187 4 : // Download each layer one by one. Generate the heatmap at each step and check
7188 4 : // that it's stable.
7189 24 : for layer in all_layers {
7190 20 : if layer.visibility() == LayerVisibilityHint::Covered {
7191 4 : continue;
7192 16 : }
7193 16 :
7194 16 : eprintln!("Downloading {layer} and re-generating heatmap");
7195 16 :
7196 16 : let ctx = &RequestContextBuilder::extend(ctx)
7197 16 : .download_behavior(crate::context::DownloadBehavior::Download)
7198 16 : .build();
7199 4 :
7200 16 : let _resident = layer
7201 16 : .download_and_keep_resident(ctx)
7202 16 : .instrument(tracing::info_span!(
7203 16 : parent: None,
7204 4 : "download_layer",
7205 4 : tenant_id = %timeline.tenant_shard_id.tenant_id,
7206 0 : shard_id = %timeline.tenant_shard_id.shard_slug(),
7207 0 : timeline_id = %timeline.timeline_id
7208 4 : ))
7209 16 : .await
7210 16 : .unwrap();
7211 4 :
7212 16 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
7213 16 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
7214 4 : }
7215 4 :
7216 4 : // Everything from the post-migration heatmap is now resident.
7217 4 : // Check that we drop it from memory.
7218 4 : assert!(matches!(
7219 4 : timeline.previous_heatmap.load().as_deref(),
7220 4 : Some(PreviousHeatmap::Obsolete)
7221 4 : ));
7222 4 : }
7223 :
7224 : #[tokio::test]
7225 4 : async fn test_previous_heatmap_obsoletion() {
7226 4 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
7227 4 : .await
7228 4 : .unwrap();
7229 4 :
7230 4 : let l0_delta = DeltaLayerTestDesc::new(
7231 4 : Lsn(0x20)..Lsn(0x30),
7232 4 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7233 4 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7234 4 : vec![(
7235 4 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7236 4 : Lsn(0x25),
7237 4 : Value::Image(test_img("foo")),
7238 4 : )],
7239 4 : );
7240 4 :
7241 4 : let image_layer = (
7242 4 : Lsn(0x40),
7243 4 : vec![(
7244 4 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7245 4 : test_img("bar"),
7246 4 : )],
7247 4 : );
7248 4 :
7249 4 : let delta_layers = vec![l0_delta];
7250 4 : let image_layers = vec![image_layer];
7251 4 :
7252 4 : let (tenant, ctx) = harness.load().await;
7253 4 : let timeline = tenant
7254 4 : .create_test_timeline_with_layers(
7255 4 : TimelineId::generate(),
7256 4 : Lsn(0x10),
7257 4 : 14,
7258 4 : &ctx,
7259 4 : Vec::new(), // in-memory layers
7260 4 : delta_layers,
7261 4 : image_layers,
7262 4 : Lsn(0x100),
7263 4 : )
7264 4 : .await
7265 4 : .unwrap();
7266 4 :
7267 4 : // Layer visibility is an input to heatmap generation, so refresh it first
7268 4 : timeline.update_layer_visibility().await.unwrap();
7269 4 :
7270 4 : let heatmap = timeline
7271 4 : .generate_heatmap()
7272 4 : .await
7273 4 : .expect("Infallible while timeline is not shut down");
7274 4 :
7275 4 : // Both layers should be in the heatmap
7276 4 : assert!(heatmap.all_layers().count() > 0);
7277 4 :
7278 4 : // Now simulate a migration.
7279 4 : timeline
7280 4 : .previous_heatmap
7281 4 : .store(Some(Arc::new(PreviousHeatmap::Active {
7282 4 : heatmap: heatmap.clone(),
7283 4 : read_at: std::time::Instant::now(),
7284 4 : end_lsn: None,
7285 4 : })));
7286 4 :
7287 4 : // Evict all the layers in the previous heatmap
7288 4 : let guard = timeline.layers.read().await;
7289 4 : let forever = std::time::Duration::from_secs(120);
7290 12 : for layer in guard.likely_resident_layers() {
7291 12 : layer.evict_and_wait(forever).await.unwrap();
7292 4 : }
7293 4 : drop(guard);
7294 4 :
7295 4 : // Generate a new heatmap and check that the previous heatmap
7296 4 : // has been marked obsolete.
7297 4 : let post_eviction_heatmap = timeline
7298 4 : .generate_heatmap()
7299 4 : .await
7300 4 : .expect("Infallible while timeline is not shut down");
7301 4 :
7302 4 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
7303 4 : assert!(matches!(
7304 4 : timeline.previous_heatmap.load().as_deref(),
7305 4 : Some(PreviousHeatmap::Obsolete)
7306 4 : ));
7307 4 : }
7308 :
7309 : #[tokio::test]
7310 4 : async fn two_layer_eviction_attempts_at_the_same_time() {
7311 4 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
7312 4 : .await
7313 4 : .unwrap();
7314 4 :
7315 4 : let (tenant, ctx) = harness.load().await;
7316 4 : let timeline = tenant
7317 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
7318 4 : .await
7319 4 : .unwrap();
7320 4 :
7321 4 : let layer = find_some_layer(&timeline).await;
7322 4 : let layer = layer
7323 4 : .keep_resident()
7324 4 : .await
7325 4 : .expect("no download => no downloading errors")
7326 4 : .drop_eviction_guard();
7327 4 :
7328 4 : let forever = std::time::Duration::from_secs(120);
7329 4 :
7330 4 : let first = layer.evict_and_wait(forever);
7331 4 : let second = layer.evict_and_wait(forever);
7332 4 :
7333 4 : let (first, second) = tokio::join!(first, second);
7334 4 :
7335 4 : let res = layer.keep_resident().await;
7336 4 : assert!(res.is_none(), "{res:?}");
7337 4 :
7338 4 : match (first, second) {
7339 4 : (Ok(()), Ok(())) => {
7340 4 : // because there are no more timeline locks being taken on eviction path, we can
7341 4 : // witness all three outcomes here.
7342 4 : }
7343 4 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7344 0 : // if one completes before the other, this is fine just as well.
7345 0 : }
7346 4 : other => unreachable!("unexpected {:?}", other),
7347 4 : }
7348 4 : }
7349 :
7350 4 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7351 4 : let layers = timeline.layers.read().await;
7352 4 : let desc = layers
7353 4 : .layer_map()
7354 4 : .unwrap()
7355 4 : .iter_historic_layers()
7356 4 : .next()
7357 4 : .expect("must find one layer to evict");
7358 4 :
7359 4 : layers.get_from_desc(&desc)
7360 4 : }
7361 : }
|
