Line data Source code
1 : pub(crate) mod analysis;
2 : pub(crate) mod compaction;
3 : pub mod delete;
4 : pub(crate) mod detach_ancestor;
5 : mod eviction_task;
6 : pub(crate) mod handle;
7 : mod heatmap_layers_downloader;
8 : pub(crate) mod import_pgdata;
9 : mod init;
10 : pub mod layer_manager;
11 : pub(crate) mod logical_size;
12 : pub mod offload;
13 : pub mod span;
14 : pub mod uninit;
15 : mod walreceiver;
16 :
17 : use std::array;
18 : use std::cmp::{max, min};
19 : use std::collections::btree_map::Entry;
20 : use std::collections::{BTreeMap, HashMap, HashSet};
21 : use std::ops::{ControlFlow, Deref, Range};
22 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
23 : use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
24 : use std::time::{Duration, Instant, SystemTime};
25 :
26 : use crate::PERF_TRACE_TARGET;
27 : use crate::walredo::RedoAttemptType;
28 : use anyhow::{Context, Result, anyhow, bail, ensure};
29 : use arc_swap::{ArcSwap, ArcSwapOption};
30 : use bytes::Bytes;
31 : use camino::Utf8Path;
32 : use chrono::{DateTime, Utc};
33 : use compaction::{CompactionOutcome, GcCompactionCombinedSettings};
34 : use enumset::EnumSet;
35 : use fail::fail_point;
36 : use futures::stream::FuturesUnordered;
37 : use futures::{FutureExt, StreamExt};
38 : use handle::ShardTimelineId;
39 : use layer_manager::Shutdown;
40 : use offload::OffloadError;
41 : use once_cell::sync::Lazy;
42 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
43 : use pageserver_api::key::{
44 : KEY_SIZE, Key, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
45 : SPARSE_RANGE,
46 : };
47 : use pageserver_api::keyspace::{KeySpaceAccum, KeySpaceRandomAccum, SparseKeyPartitioning};
48 : use pageserver_api::models::{
49 : CompactKeyRange, CompactLsnRange, CompactionAlgorithm, CompactionAlgorithmSettings,
50 : DetachBehavior, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
51 : EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, LsnLease, PageTraceEvent, RelSizeMigration,
52 : TimelineState,
53 : };
54 : use pageserver_api::reltag::{BlockNumber, RelTag};
55 : use pageserver_api::shard::{ShardIdentity, ShardIndex, ShardNumber, TenantShardId};
56 : #[cfg(test)]
57 : use pageserver_api::value::Value;
58 : use postgres_connection::PgConnectionConfig;
59 : use postgres_ffi::v14::xlog_utils;
60 : use postgres_ffi::{WAL_SEGMENT_SIZE, to_pg_timestamp};
61 : use rand::Rng;
62 : use remote_storage::DownloadError;
63 : use serde_with::serde_as;
64 : use storage_broker::BrokerClientChannel;
65 : use tokio::runtime::Handle;
66 : use tokio::sync::mpsc::Sender;
67 : use tokio::sync::{Notify, oneshot, watch};
68 : use tokio_util::sync::CancellationToken;
69 : use tracing::*;
70 : use utils::generation::Generation;
71 : use utils::guard_arc_swap::GuardArcSwap;
72 : use utils::id::TimelineId;
73 : use utils::logging::{MonitorSlowFutureCallback, monitor_slow_future};
74 : use utils::lsn::{AtomicLsn, Lsn, RecordLsn};
75 : use utils::postgres_client::PostgresClientProtocol;
76 : use utils::rate_limit::RateLimit;
77 : use utils::seqwait::SeqWait;
78 : use utils::simple_rcu::{Rcu, RcuReadGuard};
79 : use utils::sync::gate::{Gate, GateGuard};
80 : use utils::{completion, critical, fs_ext, pausable_failpoint};
81 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
82 :
83 : use self::delete::DeleteTimelineFlow;
84 : pub(super) use self::eviction_task::EvictionTaskTenantState;
85 : use self::eviction_task::EvictionTaskTimelineState;
86 : use self::layer_manager::LayerManager;
87 : use self::logical_size::LogicalSize;
88 : use self::walreceiver::{WalReceiver, WalReceiverConf};
89 : use super::remote_timeline_client::RemoteTimelineClient;
90 : use super::remote_timeline_client::index::{GcCompactionState, IndexPart};
91 : use super::secondary::heatmap::HeatMapLayer;
92 : use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
93 : use super::tasks::log_compaction_error;
94 : use super::upload_queue::NotInitialized;
95 : use super::{
96 : AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
97 : debug_assert_current_span_has_tenant_and_timeline_id,
98 : };
99 : use crate::aux_file::AuxFileSizeEstimator;
100 : use crate::config::PageServerConf;
101 : use crate::context::{
102 : DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
103 : };
104 : use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
105 : use crate::keyspace::{KeyPartitioning, KeySpace};
106 : use crate::l0_flush::{self, L0FlushGlobalState};
107 : use crate::metrics::{
108 : DELTAS_PER_READ_GLOBAL, LAYERS_PER_READ_AMORTIZED_GLOBAL, LAYERS_PER_READ_BATCH_GLOBAL,
109 : LAYERS_PER_READ_GLOBAL, ScanLatencyOngoingRecording, TimelineMetrics,
110 : };
111 : use crate::page_service::TenantManagerTypes;
112 : use crate::pgdatadir_mapping::{
113 : CalculateLogicalSizeError, CollectKeySpaceError, DirectoryKind, LsnForTimestamp,
114 : MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
115 : };
116 : use crate::task_mgr::TaskKind;
117 : use crate::tenant::config::AttachmentMode;
118 : use crate::tenant::gc_result::GcResult;
119 : use crate::tenant::layer_map::LayerMap;
120 : use crate::tenant::metadata::TimelineMetadata;
121 : use crate::tenant::storage_layer::delta_layer::DeltaEntry;
122 : use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
123 : use crate::tenant::storage_layer::{
124 : AsLayerDesc, BatchLayerWriter, DeltaLayerWriter, EvictionError, ImageLayerName,
125 : ImageLayerWriter, InMemoryLayer, IoConcurrency, Layer, LayerAccessStatsReset, LayerName,
126 : PersistentLayerDesc, PersistentLayerKey, ResidentLayer, ValueReconstructSituation,
127 : ValueReconstructState, ValuesReconstructState,
128 : };
129 : use crate::tenant::tasks::BackgroundLoopKind;
130 : use crate::tenant::timeline::logical_size::CurrentLogicalSize;
131 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
132 : use crate::walingest::WalLagCooldown;
133 : use crate::{ZERO_PAGE, task_mgr, walredo};
134 :
135 : #[derive(Debug, PartialEq, Eq, Clone, Copy)]
136 : pub(crate) enum FlushLoopState {
137 : NotStarted,
138 : Running {
139 : #[cfg(test)]
140 : expect_initdb_optimization: bool,
141 : #[cfg(test)]
142 : initdb_optimization_count: usize,
143 : },
144 : Exited,
145 : }
146 :
147 : #[derive(Debug, Copy, Clone, PartialEq, Eq)]
148 : pub enum ImageLayerCreationMode {
149 : /// Try to create image layers based on `time_for_new_image_layer`. Used in compaction code path.
150 : Try,
151 : /// Force creating the image layers if possible. For now, no image layers will be created
152 : /// for metadata keys. Used in compaction code path with force flag enabled.
153 : Force,
154 : /// Initial ingestion of the data, and no data should be dropped in this function. This
155 : /// means that no metadata keys should be included in the partitions. Used in flush frozen layer
156 : /// code path.
157 : Initial,
158 : }
159 :
160 : #[derive(Clone, Debug, Default)]
161 : pub enum LastImageLayerCreationStatus {
162 : Incomplete {
163 : /// The last key of the partition (exclusive) that was processed in the last
164 : /// image layer creation attempt. We will continue from this key in the next
165 : /// attempt.
166 : last_key: Key,
167 : },
168 : Complete,
169 : #[default]
170 : Initial,
171 : }
172 :
173 : impl std::fmt::Display for ImageLayerCreationMode {
174 3480 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
175 3480 : write!(f, "{:?}", self)
176 3480 : }
177 : }
178 :
179 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
180 : /// Can be removed after all refactors are done.
181 168 : fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
182 168 : drop(rlock)
183 168 : }
184 :
185 : /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
186 : /// Can be removed after all refactors are done.
187 3648 : fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
188 3648 : drop(rlock)
189 3648 : }
190 :
191 : /// The outward-facing resources required to build a Timeline
192 : pub struct TimelineResources {
193 : pub remote_client: RemoteTimelineClient,
194 : pub pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
195 : pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
196 : pub l0_compaction_trigger: Arc<Notify>,
197 : pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
198 : }
199 :
200 : /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
201 : /// ingestion considerably, because WAL ingestion needs to check on most records if the record
202 : /// implicitly extends the relation. At startup, `complete_as_of` is initialized to the current end
203 : /// of the timeline (disk_consistent_lsn). It's used on reads of relation sizes to check if the
204 : /// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
205 : pub(crate) struct RelSizeCache {
206 : pub(crate) complete_as_of: Lsn,
207 : pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
208 : }
209 :
210 : pub struct Timeline {
211 : pub(crate) conf: &'static PageServerConf,
212 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
213 :
214 : myself: Weak<Self>,
215 :
216 : pub(crate) tenant_shard_id: TenantShardId,
217 : pub timeline_id: TimelineId,
218 :
219 : /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
220 : /// Never changes for the lifetime of this [`Timeline`] object.
221 : ///
222 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
223 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
224 : pub(crate) generation: Generation,
225 :
226 : /// The detailed sharding information from our parent Tenant. This enables us to map keys
227 : /// to shards, and is constant through the lifetime of this Timeline.
228 : shard_identity: ShardIdentity,
229 :
230 : pub pg_version: u32,
231 :
232 : /// The tuple has two elements.
233 : /// 1. `LayerFileManager` keeps track of the various physical representations of the layer files (inmem, local, remote).
234 : /// 2. `LayerMap`, the acceleration data structure for `get_reconstruct_data`.
235 : ///
236 : /// `LayerMap` maps out the `(PAGE,LSN) / (KEY,LSN)` space, which is composed of `(KeyRange, LsnRange)` rectangles.
237 : /// We describe these rectangles through the `PersistentLayerDesc` struct.
238 : ///
239 : /// When we want to reconstruct a page, we first find the `PersistentLayerDesc`'s that we need for page reconstruction,
240 : /// using `LayerMap`. Then, we use `LayerFileManager` to get the `PersistentLayer`'s that correspond to the
241 : /// `PersistentLayerDesc`'s.
242 : ///
243 : /// Hence, it's important to keep things coherent. The `LayerFileManager` must always have an entry for all
244 : /// `PersistentLayerDesc`'s in the `LayerMap`. If it doesn't, `LayerFileManager::get_from_desc` will panic at
245 : /// runtime, e.g., during page reconstruction.
246 : ///
247 : /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
248 : /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
249 : pub(crate) layers: tokio::sync::RwLock<LayerManager>,
250 :
251 : last_freeze_at: AtomicLsn,
252 : // Atomic would be more appropriate here.
253 : last_freeze_ts: RwLock<Instant>,
254 :
255 : pub(crate) standby_horizon: AtomicLsn,
256 :
257 : // WAL redo manager. `None` only for broken tenants.
258 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
259 :
260 : /// Remote storage client.
261 : /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
262 : pub(crate) remote_client: Arc<RemoteTimelineClient>,
263 :
264 : // What page versions do we hold in the repository? If we get a
265 : // request > last_record_lsn, we need to wait until we receive all
266 : // the WAL up to the request. The SeqWait provides functions for
267 : // that. TODO: If we get a request for an old LSN, such that the
268 : // versions have already been garbage collected away, we should
269 : // throw an error, but we don't track that currently.
270 : //
271 : // last_record_lsn.load().last points to the end of last processed WAL record.
272 : //
273 : // We also remember the starting point of the previous record in
274 : // 'last_record_lsn.load().prev'. It's used to set the xl_prev pointer of the
275 : // first WAL record when the node is started up. But here, we just
276 : // keep track of it.
277 : last_record_lsn: SeqWait<RecordLsn, Lsn>,
278 :
279 : // All WAL records have been processed and stored durably on files on
280 : // local disk, up to this LSN. On crash and restart, we need to re-process
281 : // the WAL starting from this point.
282 : //
283 : // Some later WAL records might have been processed and also flushed to disk
284 : // already, so don't be surprised to see some, but there's no guarantee on
285 : // them yet.
286 : disk_consistent_lsn: AtomicLsn,
287 :
288 : // Parent timeline that this timeline was branched from, and the LSN
289 : // of the branch point.
290 : ancestor_timeline: Option<Arc<Timeline>>,
291 : ancestor_lsn: Lsn,
292 :
293 : // The LSN of gc-compaction that was last applied to this timeline.
294 : gc_compaction_state: ArcSwap<Option<GcCompactionState>>,
295 :
296 : pub(crate) metrics: Arc<TimelineMetrics>,
297 :
298 : // `Timeline` doesn't write these metrics itself, but it manages the lifetime. Code
299 : // in `crate::page_service` writes these metrics.
300 : pub(crate) query_metrics: crate::metrics::SmgrQueryTimePerTimeline,
301 :
302 : directory_metrics_inited: [AtomicBool; DirectoryKind::KINDS_NUM],
303 : directory_metrics: [AtomicU64; DirectoryKind::KINDS_NUM],
304 :
305 : /// Ensures layers aren't frozen by checkpointer between
306 : /// [`Timeline::get_layer_for_write`] and layer reads.
307 : /// Locked automatically by [`TimelineWriter`] and checkpointer.
308 : /// Must always be acquired before the layer map/individual layer lock
309 : /// to avoid deadlock.
310 : ///
311 : /// The state is cleared upon freezing.
312 : write_lock: tokio::sync::Mutex<Option<TimelineWriterState>>,
313 :
314 : /// Used to avoid multiple `flush_loop` tasks running
315 : pub(super) flush_loop_state: Mutex<FlushLoopState>,
316 :
317 : /// layer_flush_start_tx can be used to wake up the layer-flushing task.
318 : /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
319 : /// The flush cycle counter is sent back on the layer_flush_done channel when
320 : /// the flush finishes. You can use that to wait for the flush to finish.
321 : /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
322 : /// read by whoever sends an update
323 : layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
324 : /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
325 : layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
326 :
327 : // The LSN at which we have executed GC: whereas [`Self::gc_info`] records the LSN at which
328 : // we _intend_ to GC (i.e. the PITR cutoff), this LSN records where we actually last did it.
329 : // Because PITR interval is mutable, it's possible for this LSN to be earlier or later than
330 : // the planned GC cutoff.
331 : pub applied_gc_cutoff_lsn: Rcu<Lsn>,
332 :
333 : pub(crate) gc_compaction_layer_update_lock: tokio::sync::RwLock<()>,
334 :
335 : // List of child timelines and their branch points. This is needed to avoid
336 : // garbage collecting data that is still needed by the child timelines.
337 : pub(crate) gc_info: std::sync::RwLock<GcInfo>,
338 :
339 : pub(crate) last_image_layer_creation_status: ArcSwap<LastImageLayerCreationStatus>,
340 :
341 : // It may change across major versions so for simplicity
342 : // keep it after running initdb for a timeline.
343 : // It is needed in checks when we want to error on some operations
344 : // when they are requested for pre-initdb lsn.
345 : // It can be unified with latest_gc_cutoff_lsn under some "first_valid_lsn",
346 : // though let's keep them both for better error visibility.
347 : pub initdb_lsn: Lsn,
348 :
349 : /// The repartitioning result. Allows a single writer and multiple readers.
350 : pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>,
351 :
352 : /// Configuration: how often should the partitioning be recalculated.
353 : repartition_threshold: u64,
354 :
355 : last_image_layer_creation_check_at: AtomicLsn,
356 : last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
357 :
358 : /// Current logical size of the "datadir", at the last LSN.
359 : current_logical_size: LogicalSize,
360 :
361 : /// Information about the last processed message by the WAL receiver,
362 : /// or None if WAL receiver has not received anything for this timeline
363 : /// yet.
364 : pub last_received_wal: Mutex<Option<WalReceiverInfo>>,
365 : pub walreceiver: Mutex<Option<WalReceiver>>,
366 :
367 : /// Relation size cache
368 : pub(crate) rel_size_cache: RwLock<RelSizeCache>,
369 :
370 : download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
371 :
372 : state: watch::Sender<TimelineState>,
373 :
374 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
375 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
376 : pub delete_progress: TimelineDeleteProgress,
377 :
378 : eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
379 :
380 : /// Load or creation time information about the disk_consistent_lsn and when the loading
381 : /// happened. Used for consumption metrics.
382 : pub(crate) loaded_at: (Lsn, SystemTime),
383 :
384 : /// Gate to prevent shutdown completing while I/O is still happening to this timeline's data
385 : pub(crate) gate: Gate,
386 :
387 : /// Cancellation token scoped to this timeline: anything doing long-running work relating
388 : /// to the timeline should drop out when this token fires.
389 : pub(crate) cancel: CancellationToken,
390 :
391 : /// Make sure we only have one running compaction at a time in tests.
392 : ///
393 : /// Must only be taken in two places:
394 : /// - [`Timeline::compact`] (this file)
395 : /// - [`delete::delete_local_timeline_directory`]
396 : ///
397 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
398 : compaction_lock: tokio::sync::Mutex<()>,
399 :
400 : /// If true, the last compaction failed.
401 : compaction_failed: AtomicBool,
402 :
403 : /// Notifies the tenant compaction loop that there is pending L0 compaction work.
404 : l0_compaction_trigger: Arc<Notify>,
405 :
406 : /// Make sure we only have one running gc at a time.
407 : ///
408 : /// Must only be taken in two places:
409 : /// - [`Timeline::gc`] (this file)
410 : /// - [`delete::delete_local_timeline_directory`]
411 : ///
412 : /// Timeline deletion will acquire both compaction and gc locks in whatever order.
413 : gc_lock: tokio::sync::Mutex<()>,
414 :
415 : /// Cloned from [`super::TenantShard::pagestream_throttle`] on construction.
416 : pub(crate) pagestream_throttle: Arc<crate::tenant::throttle::Throttle>,
417 :
418 : /// Size estimator for aux file v2
419 : pub(crate) aux_file_size_estimator: AuxFileSizeEstimator,
420 :
421 : /// Some test cases directly place keys into the timeline without actually modifying the directory
422 : /// keys (i.e., DB_DIR). The test cases creating such keys will put the keyspaces here, so that
423 : /// these keys won't get garbage-collected during compaction/GC. This field only modifies the dense
424 : /// keyspace return value of `collect_keyspace`. For sparse keyspaces, use AUX keys for testing, and
425 : /// in the future, add `extra_test_sparse_keyspace` if necessary.
426 : #[cfg(test)]
427 : pub(crate) extra_test_dense_keyspace: ArcSwap<KeySpace>,
428 :
429 : pub(crate) l0_flush_global_state: L0FlushGlobalState,
430 :
431 : pub(crate) handles: handle::PerTimelineState<TenantManagerTypes>,
432 :
433 : pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
434 :
435 : /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
436 : pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
437 :
438 : /// If Some, collects GetPage metadata for an ongoing PageTrace.
439 : pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
440 :
441 : pub(super) previous_heatmap: ArcSwapOption<PreviousHeatmap>,
442 :
443 : /// May host a background Tokio task which downloads all the layers from the current
444 : /// heatmap on demand.
445 : heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
446 :
447 : pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
448 :
449 : wait_lsn_log_slow: tokio::sync::Semaphore,
450 : }
451 :
452 : pub(crate) enum PreviousHeatmap {
453 : Active {
454 : heatmap: HeatMapTimeline,
455 : read_at: std::time::Instant,
456 : // End LSN covered by the heatmap if known
457 : end_lsn: Option<Lsn>,
458 : },
459 : Obsolete,
460 : }
461 :
462 : pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
463 :
464 : pub struct WalReceiverInfo {
465 : pub wal_source_connconf: PgConnectionConfig,
466 : pub last_received_msg_lsn: Lsn,
467 : pub last_received_msg_ts: u128,
468 : }
469 :
470 : /// Information about how much history needs to be retained, needed by
471 : /// Garbage Collection.
472 : #[derive(Default)]
473 : pub(crate) struct GcInfo {
474 : /// Specific LSNs that are needed.
475 : ///
476 : /// Currently, this includes all points where child branches have
477 : /// been forked off from. In the future, could also include
478 : /// explicit user-defined snapshot points.
479 : pub(crate) retain_lsns: Vec<(Lsn, TimelineId, MaybeOffloaded)>,
480 :
481 : /// The cutoff coordinates, which are combined by selecting the minimum.
482 : pub(crate) cutoffs: GcCutoffs,
483 :
484 : /// Leases granted to particular LSNs.
485 : pub(crate) leases: BTreeMap<Lsn, LsnLease>,
486 :
487 : /// Whether our branch point is within our ancestor's PITR interval (for cost estimation)
488 : pub(crate) within_ancestor_pitr: bool,
489 : }
490 :
491 : impl GcInfo {
492 1848 : pub(crate) fn min_cutoff(&self) -> Lsn {
493 1848 : self.cutoffs.select_min()
494 1848 : }
495 :
496 1428 : pub(super) fn insert_child(
497 1428 : &mut self,
498 1428 : child_id: TimelineId,
499 1428 : child_lsn: Lsn,
500 1428 : is_offloaded: MaybeOffloaded,
501 1428 : ) {
502 1428 : self.retain_lsns.push((child_lsn, child_id, is_offloaded));
503 1428 : self.retain_lsns.sort_by_key(|i| i.0);
504 1428 : }
505 :
506 24 : pub(super) fn remove_child_maybe_offloaded(
507 24 : &mut self,
508 24 : child_id: TimelineId,
509 24 : maybe_offloaded: MaybeOffloaded,
510 24 : ) -> bool {
511 24 : // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
512 24 : // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
513 24 : let mut removed = false;
514 36 : self.retain_lsns.retain(|i| {
515 36 : if removed {
516 12 : return true;
517 24 : }
518 24 : let remove = i.1 == child_id && i.2 == maybe_offloaded;
519 24 : removed |= remove;
520 24 : !remove
521 36 : });
522 24 : removed
523 24 : }
524 :
525 24 : pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
526 24 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
527 24 : }
528 :
529 0 : pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
530 0 : self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
531 0 : }
532 1428 : pub(crate) fn lsn_covered_by_lease(&self, lsn: Lsn) -> bool {
533 1428 : self.leases.contains_key(&lsn)
534 1428 : }
535 : }
536 :
537 : /// The `GcInfo` component describing which Lsns need to be retained. Functionally, this
538 : /// is a single number (the oldest LSN which we must retain), but it internally distinguishes
539 : /// between time-based and space-based retention for observability and consumption metrics purposes.
540 : #[derive(Debug, Clone)]
541 : pub(crate) struct GcCutoffs {
542 : /// Calculated from the [`pageserver_api::models::TenantConfig::gc_horizon`], this LSN indicates how much
543 : /// history we must keep to retain a specified number of bytes of WAL.
544 : pub(crate) space: Lsn,
545 :
546 : /// Calculated from [`pageserver_api::models::TenantConfig::pitr_interval`], this LSN indicates how much
547 : /// history we must keep to enable reading back at least the PITR interval duration.
548 : pub(crate) time: Lsn,
549 : }
550 :
551 : impl Default for GcCutoffs {
552 2784 : fn default() -> Self {
553 2784 : Self {
554 2784 : space: Lsn::INVALID,
555 2784 : time: Lsn::INVALID,
556 2784 : }
557 2784 : }
558 : }
559 :
560 : impl GcCutoffs {
561 1848 : fn select_min(&self) -> Lsn {
562 1848 : std::cmp::min(self.space, self.time)
563 1848 : }
564 : }
565 :
566 : pub(crate) struct TimelineVisitOutcome {
567 : completed_keyspace: KeySpace,
568 : image_covered_keyspace: KeySpace,
569 : }
570 :
571 : /// An error happened in a get() operation.
572 : #[derive(thiserror::Error, Debug)]
573 : pub(crate) enum PageReconstructError {
574 : #[error(transparent)]
575 : Other(anyhow::Error),
576 :
577 : #[error("Ancestor LSN wait error: {0}")]
578 : AncestorLsnTimeout(WaitLsnError),
579 :
580 : #[error("timeline shutting down")]
581 : Cancelled,
582 :
583 : /// An error happened replaying WAL records
584 : #[error(transparent)]
585 : WalRedo(anyhow::Error),
586 :
587 : #[error("{0}")]
588 : MissingKey(Box<MissingKeyError>),
589 : }
590 :
591 : impl From<anyhow::Error> for PageReconstructError {
592 0 : fn from(value: anyhow::Error) -> Self {
593 0 : // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
594 0 : match value.downcast::<PageReconstructError>() {
595 0 : Ok(pre) => pre,
596 0 : Err(other) => PageReconstructError::Other(other),
597 : }
598 0 : }
599 : }
600 :
601 : impl From<utils::bin_ser::DeserializeError> for PageReconstructError {
602 0 : fn from(value: utils::bin_ser::DeserializeError) -> Self {
603 0 : PageReconstructError::Other(anyhow::Error::new(value).context("deserialization failure"))
604 0 : }
605 : }
606 :
607 : impl From<layer_manager::Shutdown> for PageReconstructError {
608 0 : fn from(_: layer_manager::Shutdown) -> Self {
609 0 : PageReconstructError::Cancelled
610 0 : }
611 : }
612 :
613 : impl GetVectoredError {
614 : #[cfg(test)]
615 36 : pub(crate) fn is_missing_key_error(&self) -> bool {
616 36 : matches!(self, Self::MissingKey(_))
617 36 : }
618 : }
619 :
620 : impl From<layer_manager::Shutdown> for GetVectoredError {
621 0 : fn from(_: layer_manager::Shutdown) -> Self {
622 0 : GetVectoredError::Cancelled
623 0 : }
624 : }
625 :
626 : /// A layer identifier when used in the [`ReadPath`] structure. This enum is for observability purposes
627 : /// only and not used by the "real read path".
628 : pub enum ReadPathLayerId {
629 : PersistentLayer(PersistentLayerKey),
630 : InMemoryLayer(Range<Lsn>),
631 : }
632 :
633 : impl std::fmt::Display for ReadPathLayerId {
634 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
635 0 : match self {
636 0 : ReadPathLayerId::PersistentLayer(key) => write!(f, "{}", key),
637 0 : ReadPathLayerId::InMemoryLayer(range) => {
638 0 : write!(f, "in-mem {}..{}", range.start, range.end)
639 : }
640 : }
641 0 : }
642 : }
643 : pub struct ReadPath {
644 : keyspace: KeySpace,
645 : lsn: Lsn,
646 : path: Vec<(ReadPathLayerId, KeySpace, Range<Lsn>)>,
647 : }
648 :
649 : impl ReadPath {
650 3778772 : pub fn new(keyspace: KeySpace, lsn: Lsn) -> Self {
651 3778772 : Self {
652 3778772 : keyspace,
653 3778772 : lsn,
654 3778772 : path: Vec::new(),
655 3778772 : }
656 3778772 : }
657 :
658 5316626 : pub fn record_layer_visit(
659 5316626 : &mut self,
660 5316626 : layer_to_read: &ReadableLayer,
661 5316626 : keyspace_to_read: &KeySpace,
662 5316626 : lsn_range: &Range<Lsn>,
663 5316626 : ) {
664 5316626 : let id = match layer_to_read {
665 1599415 : ReadableLayer::PersistentLayer(layer) => {
666 1599415 : ReadPathLayerId::PersistentLayer(layer.layer_desc().key())
667 : }
668 3717211 : ReadableLayer::InMemoryLayer(layer) => {
669 3717211 : ReadPathLayerId::InMemoryLayer(layer.get_lsn_range())
670 : }
671 : };
672 5316626 : self.path
673 5316626 : .push((id, keyspace_to_read.clone(), lsn_range.clone()));
674 5316626 : }
675 : }
676 :
677 : impl std::fmt::Display for ReadPath {
678 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
679 0 : writeln!(f, "Read path for {} at lsn {}:", self.keyspace, self.lsn)?;
680 0 : for (idx, (layer_id, keyspace, lsn_range)) in self.path.iter().enumerate() {
681 0 : writeln!(
682 0 : f,
683 0 : "{}: {} {}..{} {}",
684 0 : idx, layer_id, lsn_range.start, lsn_range.end, keyspace
685 0 : )?;
686 : }
687 0 : Ok(())
688 0 : }
689 : }
690 :
691 : #[derive(thiserror::Error)]
692 : pub struct MissingKeyError {
693 : keyspace: KeySpace,
694 : shard: ShardNumber,
695 : query: Option<VersionedKeySpaceQuery>,
696 : // This is largest request LSN from the get page request batch
697 : original_hwm_lsn: Lsn,
698 : ancestor_lsn: Option<Lsn>,
699 : /// Debug information about the read path if there's an error
700 : read_path: Option<ReadPath>,
701 : backtrace: Option<std::backtrace::Backtrace>,
702 : }
703 :
704 : impl MissingKeyError {
705 84 : fn enrich(&mut self, query: VersionedKeySpaceQuery) {
706 84 : self.query = Some(query);
707 84 : }
708 : }
709 :
710 : impl std::fmt::Debug for MissingKeyError {
711 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
712 0 : write!(f, "{}", self)
713 0 : }
714 : }
715 :
716 : impl std::fmt::Display for MissingKeyError {
717 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
718 0 : write!(
719 0 : f,
720 0 : "could not find data for key {} (shard {:?}), original HWM LSN {}",
721 0 : self.keyspace, self.shard, self.original_hwm_lsn
722 0 : )?;
723 :
724 0 : if let Some(ref ancestor_lsn) = self.ancestor_lsn {
725 0 : write!(f, ", ancestor {}", ancestor_lsn)?;
726 0 : }
727 :
728 0 : if let Some(ref query) = self.query {
729 0 : write!(f, ", query {}", query)?;
730 0 : }
731 :
732 0 : if let Some(ref read_path) = self.read_path {
733 0 : write!(f, "\n{}", read_path)?;
734 0 : }
735 :
736 0 : if let Some(ref backtrace) = self.backtrace {
737 0 : write!(f, "\n{}", backtrace)?;
738 0 : }
739 :
740 0 : Ok(())
741 0 : }
742 : }
743 :
744 : impl PageReconstructError {
745 : /// Returns true if this error indicates a tenant/timeline shutdown alike situation
746 0 : pub(crate) fn is_stopping(&self) -> bool {
747 : use PageReconstructError::*;
748 0 : match self {
749 0 : Cancelled => true,
750 0 : Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
751 : }
752 0 : }
753 : }
754 :
755 : #[derive(thiserror::Error, Debug)]
756 : pub(crate) enum CreateImageLayersError {
757 : #[error("timeline shutting down")]
758 : Cancelled,
759 :
760 : #[error("read failed")]
761 : GetVectoredError(#[source] GetVectoredError),
762 :
763 : #[error("reconstruction failed")]
764 : PageReconstructError(#[source] PageReconstructError),
765 :
766 : #[error(transparent)]
767 : Other(#[from] anyhow::Error),
768 : }
769 :
770 : impl From<layer_manager::Shutdown> for CreateImageLayersError {
771 0 : fn from(_: layer_manager::Shutdown) -> Self {
772 0 : CreateImageLayersError::Cancelled
773 0 : }
774 : }
775 :
776 : #[derive(thiserror::Error, Debug, Clone)]
777 : pub(crate) enum FlushLayerError {
778 : /// Timeline cancellation token was cancelled
779 : #[error("timeline shutting down")]
780 : Cancelled,
781 :
782 : /// We tried to flush a layer while the Timeline is in an unexpected state
783 : #[error("cannot flush frozen layers when flush_loop is not running, state is {0:?}")]
784 : NotRunning(FlushLoopState),
785 :
786 : // Arc<> the following non-clonable error types: we must be Clone-able because the flush error is propagated from the flush
787 : // loop via a watch channel, where we can only borrow it.
788 : #[error("create image layers (shared)")]
789 : CreateImageLayersError(Arc<CreateImageLayersError>),
790 :
791 : #[error("other (shared)")]
792 : Other(#[from] Arc<anyhow::Error>),
793 : }
794 :
795 : impl FlushLayerError {
796 : // When crossing from generic anyhow errors to this error type, we explicitly check
797 : // for timeline cancellation to avoid logging inoffensive shutdown errors as warn/err.
798 0 : fn from_anyhow(timeline: &Timeline, err: anyhow::Error) -> Self {
799 0 : let cancelled = timeline.cancel.is_cancelled()
800 : // The upload queue might have been shut down before the official cancellation of the timeline.
801 0 : || err
802 0 : .downcast_ref::<NotInitialized>()
803 0 : .map(NotInitialized::is_stopping)
804 0 : .unwrap_or_default();
805 0 : if cancelled {
806 0 : Self::Cancelled
807 : } else {
808 0 : Self::Other(Arc::new(err))
809 : }
810 0 : }
811 : }
812 :
813 : impl From<layer_manager::Shutdown> for FlushLayerError {
814 0 : fn from(_: layer_manager::Shutdown) -> Self {
815 0 : FlushLayerError::Cancelled
816 0 : }
817 : }
818 :
819 : #[derive(thiserror::Error, Debug)]
820 : pub(crate) enum GetVectoredError {
821 : #[error("timeline shutting down")]
822 : Cancelled,
823 :
824 : #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
825 : Oversized(u64),
826 :
827 : #[error("requested at invalid LSN: {0}")]
828 : InvalidLsn(Lsn),
829 :
830 : #[error("requested key not found: {0}")]
831 : MissingKey(Box<MissingKeyError>),
832 :
833 : #[error("ancestry walk")]
834 : GetReadyAncestorError(#[source] GetReadyAncestorError),
835 :
836 : #[error(transparent)]
837 : Other(#[from] anyhow::Error),
838 : }
839 :
840 : impl From<GetReadyAncestorError> for GetVectoredError {
841 12 : fn from(value: GetReadyAncestorError) -> Self {
842 : use GetReadyAncestorError::*;
843 12 : match value {
844 0 : Cancelled => GetVectoredError::Cancelled,
845 : AncestorLsnTimeout(_) | BadState { .. } => {
846 12 : GetVectoredError::GetReadyAncestorError(value)
847 : }
848 : }
849 12 : }
850 : }
851 :
852 : #[derive(thiserror::Error, Debug)]
853 : pub(crate) enum GetReadyAncestorError {
854 : #[error("ancestor LSN wait error")]
855 : AncestorLsnTimeout(#[from] WaitLsnError),
856 :
857 : #[error("bad state on timeline {timeline_id}: {state:?}")]
858 : BadState {
859 : timeline_id: TimelineId,
860 : state: TimelineState,
861 : },
862 :
863 : #[error("cancelled")]
864 : Cancelled,
865 : }
866 :
867 : #[derive(Clone, Copy)]
868 : pub enum LogicalSizeCalculationCause {
869 : Initial,
870 : ConsumptionMetricsSyntheticSize,
871 : EvictionTaskImitation,
872 : TenantSizeHandler,
873 : }
874 :
875 : pub enum GetLogicalSizePriority {
876 : User,
877 : Background,
878 : }
879 :
880 0 : #[derive(Debug, enumset::EnumSetType)]
881 : pub(crate) enum CompactFlags {
882 : ForceRepartition,
883 : ForceImageLayerCreation,
884 : ForceL0Compaction,
885 : OnlyL0Compaction,
886 : EnhancedGcBottomMostCompaction,
887 : DryRun,
888 : /// Makes image compaction yield if there's pending L0 compaction. This should always be used in
889 : /// the background compaction task, since we want to aggressively compact down L0 to bound
890 : /// read amplification.
891 : ///
892 : /// It only makes sense to use this when `compaction_l0_first` is enabled (such that we yield to
893 : /// an L0 compaction pass), and without `OnlyL0Compaction` (L0 compaction shouldn't yield for L0
894 : /// compaction).
895 : YieldForL0,
896 : }
897 :
898 : #[serde_with::serde_as]
899 0 : #[derive(Debug, Clone, serde::Deserialize)]
900 : pub(crate) struct CompactRequest {
901 : pub compact_key_range: Option<CompactKeyRange>,
902 : pub compact_lsn_range: Option<CompactLsnRange>,
903 : /// Whether the compaction job should be scheduled.
904 : #[serde(default)]
905 : pub scheduled: bool,
906 : /// Whether the compaction job should be split across key ranges.
907 : #[serde(default)]
908 : pub sub_compaction: bool,
909 : /// Max job size for each subcompaction job.
910 : pub sub_compaction_max_job_size_mb: Option<u64>,
911 : }
912 :
913 0 : #[derive(Debug, Clone, serde::Deserialize)]
914 : pub(crate) struct MarkInvisibleRequest {
915 : #[serde(default)]
916 : pub is_visible: Option<bool>,
917 : }
918 :
919 : #[derive(Debug, Clone, Default)]
920 : pub(crate) struct CompactOptions {
921 : pub flags: EnumSet<CompactFlags>,
922 : /// If set, the compaction will only compact the key range specified by this option.
923 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
924 : pub compact_key_range: Option<CompactKeyRange>,
925 : /// If set, the compaction will only compact the LSN within this value.
926 : /// This option is only used by GC compaction. For the full explanation, see [`compaction::GcCompactJob`].
927 : pub compact_lsn_range: Option<CompactLsnRange>,
928 : /// Enable sub-compaction (split compaction job across key ranges).
929 : /// This option is only used by GC compaction.
930 : pub sub_compaction: bool,
931 : /// Set job size for the GC compaction.
932 : /// This option is only used by GC compaction.
933 : pub sub_compaction_max_job_size_mb: Option<u64>,
934 : }
935 :
936 : impl std::fmt::Debug for Timeline {
937 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
938 0 : write!(f, "Timeline<{}>", self.timeline_id)
939 0 : }
940 : }
941 :
942 : #[derive(thiserror::Error, Debug, Clone)]
943 : pub(crate) enum WaitLsnError {
944 : // Called on a timeline which is shutting down
945 : #[error("Shutdown")]
946 : Shutdown,
947 :
948 : // Called on an timeline not in active state or shutting down
949 : #[error("Bad timeline state: {0:?}")]
950 : BadState(TimelineState),
951 :
952 : // Timeout expired while waiting for LSN to catch up with goal.
953 : #[error("{0}")]
954 : Timeout(String),
955 : }
956 :
957 : // The impls below achieve cancellation mapping for errors.
958 : // Perhaps there's a way of achieving this with less cruft.
959 :
960 : impl From<CreateImageLayersError> for CompactionError {
961 0 : fn from(e: CreateImageLayersError) -> Self {
962 0 : match e {
963 0 : CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
964 0 : CreateImageLayersError::Other(e) => {
965 0 : CompactionError::Other(e.context("create image layers"))
966 : }
967 0 : _ => CompactionError::Other(e.into()),
968 : }
969 0 : }
970 : }
971 :
972 : impl From<CreateImageLayersError> for FlushLayerError {
973 0 : fn from(e: CreateImageLayersError) -> Self {
974 0 : match e {
975 0 : CreateImageLayersError::Cancelled => FlushLayerError::Cancelled,
976 0 : any => FlushLayerError::CreateImageLayersError(Arc::new(any)),
977 : }
978 0 : }
979 : }
980 :
981 : impl From<PageReconstructError> for CreateImageLayersError {
982 0 : fn from(e: PageReconstructError) -> Self {
983 0 : match e {
984 0 : PageReconstructError::Cancelled => CreateImageLayersError::Cancelled,
985 0 : _ => CreateImageLayersError::PageReconstructError(e),
986 : }
987 0 : }
988 : }
989 :
990 : impl From<GetVectoredError> for CreateImageLayersError {
991 0 : fn from(e: GetVectoredError) -> Self {
992 0 : match e {
993 0 : GetVectoredError::Cancelled => CreateImageLayersError::Cancelled,
994 0 : _ => CreateImageLayersError::GetVectoredError(e),
995 : }
996 0 : }
997 : }
998 :
999 : impl From<GetVectoredError> for PageReconstructError {
1000 36 : fn from(e: GetVectoredError) -> Self {
1001 36 : match e {
1002 0 : GetVectoredError::Cancelled => PageReconstructError::Cancelled,
1003 0 : GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
1004 0 : err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
1005 24 : GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
1006 12 : GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
1007 0 : GetVectoredError::Other(err) => PageReconstructError::Other(err),
1008 : }
1009 36 : }
1010 : }
1011 :
1012 : impl From<GetReadyAncestorError> for PageReconstructError {
1013 12 : fn from(e: GetReadyAncestorError) -> Self {
1014 : use GetReadyAncestorError::*;
1015 12 : match e {
1016 0 : AncestorLsnTimeout(wait_err) => PageReconstructError::AncestorLsnTimeout(wait_err),
1017 12 : bad_state @ BadState { .. } => PageReconstructError::Other(anyhow::anyhow!(bad_state)),
1018 0 : Cancelled => PageReconstructError::Cancelled,
1019 : }
1020 12 : }
1021 : }
1022 :
1023 : pub(crate) enum WaitLsnTimeout {
1024 : Custom(Duration),
1025 : // Use the [`PageServerConf::wait_lsn_timeout`] default
1026 : Default,
1027 : }
1028 :
1029 : pub(crate) enum WaitLsnWaiter<'a> {
1030 : Timeline(&'a Timeline),
1031 : Tenant,
1032 : PageService,
1033 : HttpEndpoint,
1034 : }
1035 :
1036 : /// Argument to [`Timeline::shutdown`].
1037 : #[derive(Debug, Clone, Copy)]
1038 : pub(crate) enum ShutdownMode {
1039 : /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
1040 : /// also to remote storage. This method can easily take multiple seconds for a busy timeline.
1041 : ///
1042 : /// While we are flushing, we continue to accept read I/O for LSNs ingested before
1043 : /// the call to [`Timeline::shutdown`].
1044 : FreezeAndFlush,
1045 : /// Only flush the layers to the remote storage without freezing any open layers. Flush the deletion
1046 : /// queue. This is the mode used by ancestor detach and any other operations that reloads a tenant
1047 : /// but not increasing the generation number. Note that this mode cannot be used at tenant shutdown,
1048 : /// as flushing the deletion queue at that time will cause shutdown-in-progress errors.
1049 : Reload,
1050 : /// Shut down immediately, without waiting for any open layers to flush.
1051 : Hard,
1052 : }
1053 :
1054 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1055 : enum ImageLayerCreationOutcome {
1056 : /// We generated an image layer
1057 : Generated {
1058 : unfinished_image_layer: ImageLayerWriter,
1059 : },
1060 : /// The key range is empty
1061 : Empty,
1062 : /// (Only used in metadata image layer creation), after reading the metadata keys, we decide to skip
1063 : /// the image layer creation.
1064 : Skip,
1065 : }
1066 :
1067 : /// Public interface functions
1068 : impl Timeline {
1069 : /// Get the LSN where this branch was created
1070 264 : pub(crate) fn get_ancestor_lsn(&self) -> Lsn {
1071 264 : self.ancestor_lsn
1072 264 : }
1073 :
1074 : /// Get the ancestor's timeline id
1075 456 : pub(crate) fn get_ancestor_timeline_id(&self) -> Option<TimelineId> {
1076 456 : self.ancestor_timeline
1077 456 : .as_ref()
1078 456 : .map(|ancestor| ancestor.timeline_id)
1079 456 : }
1080 :
1081 : /// Get the ancestor timeline
1082 12 : pub(crate) fn ancestor_timeline(&self) -> Option<&Arc<Timeline>> {
1083 12 : self.ancestor_timeline.as_ref()
1084 12 : }
1085 :
1086 : /// Get the bytes written since the PITR cutoff on this branch, and
1087 : /// whether this branch's ancestor_lsn is within its parent's PITR.
1088 0 : pub(crate) fn get_pitr_history_stats(&self) -> (u64, bool) {
1089 0 : let gc_info = self.gc_info.read().unwrap();
1090 0 : let history = self
1091 0 : .get_last_record_lsn()
1092 0 : .checked_sub(gc_info.cutoffs.time)
1093 0 : .unwrap_or(Lsn(0))
1094 0 : .0;
1095 0 : (history, gc_info.within_ancestor_pitr)
1096 0 : }
1097 :
1098 : /// Read timeline's GC cutoff: this is the LSN at which GC has started to happen
1099 5133624 : pub(crate) fn get_applied_gc_cutoff_lsn(&self) -> RcuReadGuard<Lsn> {
1100 5133624 : self.applied_gc_cutoff_lsn.read()
1101 5133624 : }
1102 :
1103 : /// Read timeline's planned GC cutoff: this is the logical end of history that users
1104 : /// are allowed to read (based on configured PITR), even if physically we have more history.
1105 0 : pub(crate) fn get_gc_cutoff_lsn(&self) -> Lsn {
1106 0 : self.gc_info.read().unwrap().cutoffs.time
1107 0 : }
1108 :
1109 : /// Look up given page version.
1110 : ///
1111 : /// If a remote layer file is needed, it is downloaded as part of this
1112 : /// call.
1113 : ///
1114 : /// This method enforces [`Self::pagestream_throttle`] internally.
1115 : ///
1116 : /// NOTE: It is considered an error to 'get' a key that doesn't exist. The
1117 : /// abstraction above this needs to store suitable metadata to track what
1118 : /// data exists with what keys, in separate metadata entries. If a
1119 : /// non-existent key is requested, we may incorrectly return a value from
1120 : /// an ancestor branch, for example, or waste a lot of cycles chasing the
1121 : /// non-existing key.
1122 : ///
1123 : /// # Cancel-Safety
1124 : ///
1125 : /// This method is cancellation-safe.
1126 : #[inline(always)]
1127 3645992 : pub(crate) async fn get(
1128 3645992 : &self,
1129 3645992 : key: Key,
1130 3645992 : lsn: Lsn,
1131 3645992 : ctx: &RequestContext,
1132 3645992 : ) -> Result<Bytes, PageReconstructError> {
1133 3645992 : if !lsn.is_valid() {
1134 0 : return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
1135 3645992 : }
1136 3645992 :
1137 3645992 : // This check is debug-only because of the cost of hashing, and because it's a double-check: we
1138 3645992 : // already checked the key against the shard_identity when looking up the Timeline from
1139 3645992 : // page_service.
1140 3645992 : debug_assert!(!self.shard_identity.is_key_disposable(&key));
1141 :
1142 3645992 : let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
1143 3645992 :
1144 3645992 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
1145 :
1146 3645992 : let vectored_res = self
1147 3645992 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
1148 3645992 : .await;
1149 :
1150 3645992 : let key_value = vectored_res?.pop_first();
1151 3645956 : match key_value {
1152 3645884 : Some((got_key, value)) => {
1153 3645884 : if got_key != key {
1154 0 : error!(
1155 0 : "Expected {}, but singular vectored get returned {}",
1156 : key, got_key
1157 : );
1158 0 : Err(PageReconstructError::Other(anyhow!(
1159 0 : "Singular vectored get returned wrong key"
1160 0 : )))
1161 : } else {
1162 3645884 : value
1163 : }
1164 : }
1165 72 : None => Err(PageReconstructError::MissingKey(Box::new(
1166 72 : MissingKeyError {
1167 72 : keyspace: KeySpace::single(key..key.next()),
1168 72 : shard: self.shard_identity.get_shard_number(&key),
1169 72 : original_hwm_lsn: lsn,
1170 72 : ancestor_lsn: None,
1171 72 : backtrace: None,
1172 72 : read_path: None,
1173 72 : query: None,
1174 72 : },
1175 72 : ))),
1176 : }
1177 3645992 : }
1178 :
1179 : pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
1180 : pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
1181 :
1182 : /// Look up multiple page versions at a given LSN
1183 : ///
1184 : /// This naive implementation will be replaced with a more efficient one
1185 : /// which actually vectorizes the read path.
1186 130512 : pub(crate) async fn get_vectored(
1187 130512 : &self,
1188 130512 : query: VersionedKeySpaceQuery,
1189 130512 : io_concurrency: super::storage_layer::IoConcurrency,
1190 130512 : ctx: &RequestContext,
1191 130512 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1192 130512 : let total_keyspace = query.total_keyspace();
1193 130512 :
1194 130512 : let key_count = total_keyspace.total_raw_size().try_into().unwrap();
1195 130512 : if key_count > Timeline::MAX_GET_VECTORED_KEYS {
1196 0 : return Err(GetVectoredError::Oversized(key_count));
1197 130512 : }
1198 :
1199 410052 : for range in &total_keyspace.ranges {
1200 279540 : let mut key = range.start;
1201 783636 : while key != range.end {
1202 504096 : assert!(!self.shard_identity.is_key_disposable(&key));
1203 504096 : key = key.next();
1204 : }
1205 : }
1206 :
1207 130512 : trace!(
1208 0 : "get vectored query {} from task kind {:?}",
1209 0 : query,
1210 0 : ctx.task_kind(),
1211 : );
1212 :
1213 130512 : let start = crate::metrics::GET_VECTORED_LATENCY
1214 130512 : .for_task_kind(ctx.task_kind())
1215 130512 : .map(|metric| (metric, Instant::now()));
1216 :
1217 130512 : let res = self
1218 130512 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1219 130512 : .await;
1220 :
1221 130512 : if let Some((metric, start)) = start {
1222 0 : let elapsed = start.elapsed();
1223 0 : metric.observe(elapsed.as_secs_f64());
1224 130512 : }
1225 :
1226 130512 : res
1227 130512 : }
1228 :
1229 : /// Scan the keyspace and return all existing key-values in the keyspace. This currently uses vectored
1230 : /// get underlying. Normal vectored get would throw an error when a key in the keyspace is not found
1231 : /// during the search, but for the scan interface, it returns all existing key-value pairs, and does
1232 : /// not expect each single key in the key space will be found. The semantics is closer to the RocksDB
1233 : /// scan iterator interface. We could optimize this interface later to avoid some checks in the vectored
1234 : /// get path to maintain and split the probing and to-be-probe keyspace. We also need to ensure that
1235 : /// the scan operation will not cause OOM in the future.
1236 72 : pub(crate) async fn scan(
1237 72 : &self,
1238 72 : keyspace: KeySpace,
1239 72 : lsn: Lsn,
1240 72 : ctx: &RequestContext,
1241 72 : io_concurrency: super::storage_layer::IoConcurrency,
1242 72 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1243 72 : if !lsn.is_valid() {
1244 0 : return Err(GetVectoredError::InvalidLsn(lsn));
1245 72 : }
1246 72 :
1247 72 : trace!(
1248 0 : "key-value scan request for {:?}@{} from task kind {:?}",
1249 0 : keyspace,
1250 0 : lsn,
1251 0 : ctx.task_kind()
1252 : );
1253 :
1254 : // We should generalize this into Keyspace::contains in the future.
1255 144 : for range in &keyspace.ranges {
1256 72 : if range.start.field1 < METADATA_KEY_BEGIN_PREFIX
1257 72 : || range.end.field1 > METADATA_KEY_END_PREFIX
1258 : {
1259 0 : return Err(GetVectoredError::Other(anyhow::anyhow!(
1260 0 : "only metadata keyspace can be scanned"
1261 0 : )));
1262 72 : }
1263 : }
1264 :
1265 72 : let start = crate::metrics::SCAN_LATENCY
1266 72 : .for_task_kind(ctx.task_kind())
1267 72 : .map(ScanLatencyOngoingRecording::start_recording);
1268 72 :
1269 72 : let query = VersionedKeySpaceQuery::uniform(keyspace, lsn);
1270 :
1271 72 : let vectored_res = self
1272 72 : .get_vectored_impl(query, &mut ValuesReconstructState::new(io_concurrency), ctx)
1273 72 : .await;
1274 :
1275 72 : if let Some(recording) = start {
1276 0 : recording.observe();
1277 72 : }
1278 :
1279 72 : vectored_res
1280 72 : }
1281 :
1282 3778772 : pub(super) async fn get_vectored_impl(
1283 3778772 : &self,
1284 3778772 : query: VersionedKeySpaceQuery,
1285 3778772 : reconstruct_state: &mut ValuesReconstructState,
1286 3778772 : ctx: &RequestContext,
1287 3778772 : ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
1288 3778772 : if query.is_empty() {
1289 0 : return Ok(BTreeMap::default());
1290 3778772 : }
1291 :
1292 3778772 : let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
1293 : Some(ReadPath::new(
1294 3778772 : query.total_keyspace(),
1295 3778772 : query.high_watermark_lsn()?,
1296 : ))
1297 : } else {
1298 0 : None
1299 : };
1300 :
1301 3778772 : reconstruct_state.read_path = read_path;
1302 :
1303 3778772 : let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
1304 0 : RedoAttemptType::LegacyCompaction
1305 : } else {
1306 3778772 : RedoAttemptType::ReadPage
1307 : };
1308 :
1309 3778772 : let traversal_res: Result<(), _> = {
1310 3778772 : let ctx = RequestContextBuilder::from(ctx)
1311 3778772 : .perf_span(|crnt_perf_span| {
1312 0 : info_span!(
1313 : target: PERF_TRACE_TARGET,
1314 0 : parent: crnt_perf_span,
1315 : "PLAN_IO",
1316 : )
1317 3778772 : })
1318 3778772 : .attached_child();
1319 3778772 :
1320 3778772 : self.get_vectored_reconstruct_data(query.clone(), reconstruct_state, &ctx)
1321 3778772 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1322 3778772 : .await
1323 : };
1324 :
1325 3778772 : if let Err(err) = traversal_res {
1326 : // Wait for all the spawned IOs to complete.
1327 : // See comments on `spawn_io` inside `storage_layer` for more details.
1328 96 : let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
1329 96 : .into_values()
1330 96 : .map(|state| state.collect_pending_ios())
1331 96 : .collect::<FuturesUnordered<_>>();
1332 96 : while collect_futs.next().await.is_some() {}
1333 :
1334 : // Enrich the missing key error with the original query.
1335 96 : if let GetVectoredError::MissingKey(mut missing_err) = err {
1336 84 : missing_err.enrich(query.clone());
1337 84 : return Err(GetVectoredError::MissingKey(missing_err));
1338 12 : }
1339 12 :
1340 12 : return Err(err);
1341 3778676 : };
1342 3778676 :
1343 3778676 : let layers_visited = reconstruct_state.get_layers_visited();
1344 3778676 :
1345 3778676 : let ctx = RequestContextBuilder::from(ctx)
1346 3778676 : .perf_span(|crnt_perf_span| {
1347 0 : info_span!(
1348 : target: PERF_TRACE_TARGET,
1349 0 : parent: crnt_perf_span,
1350 : "RECONSTRUCT",
1351 : )
1352 3778676 : })
1353 3778676 : .attached_child();
1354 3778676 :
1355 3778676 : let futs = FuturesUnordered::new();
1356 4392788 : for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
1357 4392788 : let req_lsn_for_key = query.map_key_to_lsn(&key);
1358 4392788 :
1359 4392788 : futs.push({
1360 4392788 : let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
1361 4392788 : let ctx = RequestContextBuilder::from(&ctx)
1362 4392788 : .perf_span(|crnt_perf_span| {
1363 0 : info_span!(
1364 : target: PERF_TRACE_TARGET,
1365 0 : parent: crnt_perf_span,
1366 : "RECONSTRUCT_KEY",
1367 : key = %key,
1368 : )
1369 4392788 : })
1370 4392788 : .attached_child();
1371 4392788 :
1372 4392788 : async move {
1373 4392788 : assert_eq!(state.situation, ValueReconstructSituation::Complete);
1374 :
1375 4392788 : let res = state
1376 4392788 : .collect_pending_ios()
1377 4392788 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1378 0 : info_span!(
1379 : target: PERF_TRACE_TARGET,
1380 0 : parent: crnt_perf_span,
1381 : "WAIT_FOR_IO_COMPLETIONS",
1382 : )
1383 4392788 : })
1384 4392788 : .await;
1385 :
1386 4392788 : let converted = match res {
1387 4392788 : Ok(ok) => ok,
1388 0 : Err(err) => {
1389 0 : return (key, Err(err));
1390 : }
1391 : };
1392 4392788 : DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
1393 4392788 :
1394 4392788 : // The walredo module expects the records to be descending in terms of Lsn.
1395 4392788 : // And we submit the IOs in that order, so, there shuold be no need to sort here.
1396 4392788 : debug_assert!(
1397 4392788 : converted
1398 4392788 : .records
1399 16839048 : .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
1400 0 : "{converted:?}"
1401 : );
1402 :
1403 4392788 : let walredo_deltas = converted.num_deltas();
1404 4392788 : let walredo_res = walredo_self
1405 4392788 : .reconstruct_value(key, req_lsn_for_key, converted, redo_attempt_type)
1406 4392788 : .maybe_perf_instrument(&ctx, |crnt_perf_span| {
1407 0 : info_span!(
1408 : target: PERF_TRACE_TARGET,
1409 0 : parent: crnt_perf_span,
1410 : "WALREDO",
1411 : deltas = %walredo_deltas,
1412 : )
1413 4392788 : })
1414 4392788 : .await;
1415 :
1416 4392788 : (key, walredo_res)
1417 4392788 : }
1418 4392788 : });
1419 4392788 : }
1420 :
1421 3778676 : let results = futs
1422 3778676 : .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
1423 3778676 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
1424 3778676 : .await;
1425 :
1426 : // For aux file keys (v1 or v2) the vectored read path does not return an error
1427 : // when they're missing. Instead they are omitted from the resulting btree
1428 : // (this is a requirement, not a bug). Skip updating the metric in these cases
1429 : // to avoid infinite results.
1430 3778676 : if !results.is_empty() {
1431 3777164 : if layers_visited >= Self::LAYERS_VISITED_WARN_THRESHOLD {
1432 0 : let total_keyspace = query.total_keyspace();
1433 0 : let max_request_lsn = query.high_watermark_lsn().expect("Validated previously");
1434 0 :
1435 0 : static LOG_PACER: Lazy<Mutex<RateLimit>> =
1436 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(60))));
1437 0 : LOG_PACER.lock().unwrap().call(|| {
1438 0 : let num_keys = total_keyspace.total_raw_size();
1439 0 : let num_pages = results.len();
1440 0 : tracing::info!(
1441 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1442 0 : lsn = %max_request_lsn,
1443 0 : "Vectored read for {total_keyspace} visited {layers_visited} layers. Returned {num_pages}/{num_keys} pages.",
1444 : );
1445 0 : });
1446 3777164 : }
1447 :
1448 : // Records the number of layers visited in a few different ways:
1449 : //
1450 : // * LAYERS_PER_READ: all layers count towards every read in the batch, because each
1451 : // layer directly affects its observed latency.
1452 : //
1453 : // * LAYERS_PER_READ_BATCH: all layers count towards each batch, to get the per-batch
1454 : // layer visits and access cost.
1455 : //
1456 : // * LAYERS_PER_READ_AMORTIZED: the average layer count per read, to get the amortized
1457 : // read amplification after batching.
1458 3777164 : let layers_visited = layers_visited as f64;
1459 3777164 : let avg_layers_visited = layers_visited / results.len() as f64;
1460 3777164 : LAYERS_PER_READ_BATCH_GLOBAL.observe(layers_visited);
1461 8169952 : for _ in &results {
1462 4392788 : self.metrics.layers_per_read.observe(layers_visited);
1463 4392788 : LAYERS_PER_READ_GLOBAL.observe(layers_visited);
1464 4392788 : LAYERS_PER_READ_AMORTIZED_GLOBAL.observe(avg_layers_visited);
1465 4392788 : }
1466 1512 : }
1467 :
1468 3778676 : Ok(results)
1469 3778772 : }
1470 :
1471 : /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
1472 1647420 : pub(crate) fn get_last_record_lsn(&self) -> Lsn {
1473 1647420 : self.last_record_lsn.load().last
1474 1647420 : }
1475 :
1476 0 : pub(crate) fn get_prev_record_lsn(&self) -> Lsn {
1477 0 : self.last_record_lsn.load().prev
1478 0 : }
1479 :
1480 : /// Atomically get both last and prev.
1481 1404 : pub(crate) fn get_last_record_rlsn(&self) -> RecordLsn {
1482 1404 : self.last_record_lsn.load()
1483 1404 : }
1484 :
1485 : /// Subscribe to callers of wait_lsn(). The value of the channel is None if there are no
1486 : /// wait_lsn() calls in progress, and Some(Lsn) if there is an active waiter for wait_lsn().
1487 0 : pub(crate) fn subscribe_for_wait_lsn_updates(&self) -> watch::Receiver<Option<Lsn>> {
1488 0 : self.last_record_lsn.status_receiver()
1489 0 : }
1490 :
1491 2760 : pub(crate) fn get_disk_consistent_lsn(&self) -> Lsn {
1492 2760 : self.disk_consistent_lsn.load()
1493 2760 : }
1494 :
1495 : /// remote_consistent_lsn from the perspective of the tenant's current generation,
1496 : /// not validated with control plane yet.
1497 : /// See [`Self::get_remote_consistent_lsn_visible`].
1498 0 : pub(crate) fn get_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
1499 0 : self.remote_client.remote_consistent_lsn_projected()
1500 0 : }
1501 :
1502 : /// remote_consistent_lsn which the tenant is guaranteed not to go backward from,
1503 : /// i.e. a value of remote_consistent_lsn_projected which has undergone
1504 : /// generation validation in the deletion queue.
1505 0 : pub(crate) fn get_remote_consistent_lsn_visible(&self) -> Option<Lsn> {
1506 0 : self.remote_client.remote_consistent_lsn_visible()
1507 0 : }
1508 :
1509 : /// The sum of the file size of all historic layers in the layer map.
1510 : /// This method makes no distinction between local and remote layers.
1511 : /// Hence, the result **does not represent local filesystem usage**.
1512 0 : pub(crate) async fn layer_size_sum(&self) -> u64 {
1513 0 : let guard = self.layers.read().await;
1514 0 : guard.layer_size_sum()
1515 0 : }
1516 :
1517 0 : pub(crate) fn resident_physical_size(&self) -> u64 {
1518 0 : self.metrics.resident_physical_size_get()
1519 0 : }
1520 :
1521 0 : pub(crate) fn get_directory_metrics(&self) -> [u64; DirectoryKind::KINDS_NUM] {
1522 0 : array::from_fn(|idx| self.directory_metrics[idx].load(AtomicOrdering::Relaxed))
1523 0 : }
1524 :
1525 : ///
1526 : /// Wait until WAL has been received and processed up to this LSN.
1527 : ///
1528 : /// You should call this before any of the other get_* or list_* functions. Calling
1529 : /// those functions with an LSN that has been processed yet is an error.
1530 : ///
1531 1352896 : pub(crate) async fn wait_lsn(
1532 1352896 : &self,
1533 1352896 : lsn: Lsn,
1534 1352896 : who_is_waiting: WaitLsnWaiter<'_>,
1535 1352896 : timeout: WaitLsnTimeout,
1536 1352896 : ctx: &RequestContext, /* Prepare for use by cancellation */
1537 1352896 : ) -> Result<(), WaitLsnError> {
1538 1352896 : let state = self.current_state();
1539 1352896 : if self.cancel.is_cancelled() || matches!(state, TimelineState::Stopping) {
1540 0 : return Err(WaitLsnError::Shutdown);
1541 1352896 : } else if !matches!(state, TimelineState::Active) {
1542 0 : return Err(WaitLsnError::BadState(state));
1543 1352896 : }
1544 1352896 :
1545 1352896 : if cfg!(debug_assertions) {
1546 1352896 : match ctx.task_kind() {
1547 : TaskKind::WalReceiverManager
1548 : | TaskKind::WalReceiverConnectionHandler
1549 : | TaskKind::WalReceiverConnectionPoller => {
1550 0 : let is_myself = match who_is_waiting {
1551 0 : WaitLsnWaiter::Timeline(waiter) => {
1552 0 : Weak::ptr_eq(&waiter.myself, &self.myself)
1553 : }
1554 : WaitLsnWaiter::Tenant
1555 : | WaitLsnWaiter::PageService
1556 0 : | WaitLsnWaiter::HttpEndpoint => unreachable!(
1557 0 : "tenant or page_service context are not expected to have task kind {:?}",
1558 0 : ctx.task_kind()
1559 0 : ),
1560 : };
1561 0 : if is_myself {
1562 0 : if let Err(current) = self.last_record_lsn.would_wait_for(lsn) {
1563 : // walingest is the only one that can advance last_record_lsn; it should make sure to never reach here
1564 0 : panic!(
1565 0 : "this timeline's walingest task is calling wait_lsn({lsn}) but we only have last_record_lsn={current}; would deadlock"
1566 0 : );
1567 0 : }
1568 0 : } else {
1569 0 : // if another timeline's is waiting for us, there's no deadlock risk because
1570 0 : // our walreceiver task can make progress independent of theirs
1571 0 : }
1572 : }
1573 1352896 : _ => {}
1574 : }
1575 0 : }
1576 :
1577 1352896 : let timeout = match timeout {
1578 0 : WaitLsnTimeout::Custom(t) => t,
1579 1352896 : WaitLsnTimeout::Default => self.conf.wait_lsn_timeout,
1580 : };
1581 :
1582 1352896 : let timer = crate::metrics::WAIT_LSN_TIME.start_timer();
1583 1352896 : let start_finish_counterpair_guard = self.metrics.wait_lsn_start_finish_counterpair.guard();
1584 1352896 :
1585 1352896 : let wait_for_timeout = self.last_record_lsn.wait_for_timeout(lsn, timeout);
1586 1352896 : let wait_for_timeout = std::pin::pin!(wait_for_timeout);
1587 1352896 : // Use threshold of 1 because even 1 second of wait for ingest is very much abnormal.
1588 1352896 : let log_slow_threshold = Duration::from_secs(1);
1589 1352896 : // Use period of 10 to avoid flooding logs during an outage that affects all timelines.
1590 1352896 : let log_slow_period = Duration::from_secs(10);
1591 1352896 : let mut logging_permit = None;
1592 1352896 : let wait_for_timeout = monitor_slow_future(
1593 1352896 : log_slow_threshold,
1594 1352896 : log_slow_period,
1595 1352896 : wait_for_timeout,
1596 1352896 : |MonitorSlowFutureCallback {
1597 : ready,
1598 : is_slow,
1599 : elapsed_total,
1600 : elapsed_since_last_callback,
1601 1352896 : }| {
1602 1352896 : self.metrics
1603 1352896 : .wait_lsn_in_progress_micros
1604 1352896 : .inc_by(u64::try_from(elapsed_since_last_callback.as_micros()).unwrap());
1605 1352896 : if !is_slow {
1606 1352896 : return;
1607 0 : }
1608 0 : // It's slow, see if we should log it.
1609 0 : // (We limit the logging to one per invocation per timeline to avoid excessive
1610 0 : // logging during an extended broker / networking outage that affects all timelines.)
1611 0 : if logging_permit.is_none() {
1612 0 : logging_permit = self.wait_lsn_log_slow.try_acquire().ok();
1613 0 : }
1614 0 : if logging_permit.is_none() {
1615 0 : return;
1616 0 : }
1617 0 : // We log it.
1618 0 : if ready {
1619 0 : info!(
1620 0 : "slow wait_lsn completed after {:.3}s",
1621 0 : elapsed_total.as_secs_f64()
1622 : );
1623 : } else {
1624 0 : info!(
1625 0 : "slow wait_lsn still running for {:.3}s",
1626 0 : elapsed_total.as_secs_f64()
1627 : );
1628 : }
1629 1352896 : },
1630 1352896 : );
1631 1352896 : let res = wait_for_timeout.await;
1632 : // don't count the time spent waiting for lock below, and also in walreceiver.status(), towards the wait_lsn_time_histo
1633 1352896 : drop(logging_permit);
1634 1352896 : drop(start_finish_counterpair_guard);
1635 1352896 : drop(timer);
1636 1352896 : match res {
1637 1352896 : Ok(()) => Ok(()),
1638 0 : Err(e) => {
1639 : use utils::seqwait::SeqWaitError::*;
1640 0 : match e {
1641 0 : Shutdown => Err(WaitLsnError::Shutdown),
1642 : Timeout => {
1643 0 : let walreceiver_status = self.walreceiver_status();
1644 0 : Err(WaitLsnError::Timeout(format!(
1645 0 : "Timed out while waiting for WAL record at LSN {} to arrive, last_record_lsn {} disk consistent LSN={}, WalReceiver status: {}",
1646 0 : lsn,
1647 0 : self.get_last_record_lsn(),
1648 0 : self.get_disk_consistent_lsn(),
1649 0 : walreceiver_status,
1650 0 : )))
1651 : }
1652 : }
1653 : }
1654 : }
1655 1352896 : }
1656 :
1657 0 : pub(crate) fn walreceiver_status(&self) -> String {
1658 0 : match &*self.walreceiver.lock().unwrap() {
1659 0 : None => "stopping or stopped".to_string(),
1660 0 : Some(walreceiver) => match walreceiver.status() {
1661 0 : Some(status) => status.to_human_readable_string(),
1662 0 : None => "Not active".to_string(),
1663 : },
1664 : }
1665 0 : }
1666 :
1667 : /// Check that it is valid to request operations with that lsn.
1668 1428 : pub(crate) fn check_lsn_is_in_scope(
1669 1428 : &self,
1670 1428 : lsn: Lsn,
1671 1428 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1672 1428 : ) -> anyhow::Result<()> {
1673 1428 : ensure!(
1674 1428 : lsn >= **latest_gc_cutoff_lsn,
1675 24 : "LSN {} is earlier than latest GC cutoff {} (we might've already garbage collected needed data)",
1676 24 : lsn,
1677 24 : **latest_gc_cutoff_lsn,
1678 : );
1679 1404 : Ok(())
1680 1428 : }
1681 :
1682 : /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
1683 60 : pub(crate) fn init_lsn_lease(
1684 60 : &self,
1685 60 : lsn: Lsn,
1686 60 : length: Duration,
1687 60 : ctx: &RequestContext,
1688 60 : ) -> anyhow::Result<LsnLease> {
1689 60 : self.make_lsn_lease(lsn, length, true, ctx)
1690 60 : }
1691 :
1692 : /// 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.
1693 24 : pub(crate) fn renew_lsn_lease(
1694 24 : &self,
1695 24 : lsn: Lsn,
1696 24 : length: Duration,
1697 24 : ctx: &RequestContext,
1698 24 : ) -> anyhow::Result<LsnLease> {
1699 24 : self.make_lsn_lease(lsn, length, false, ctx)
1700 24 : }
1701 :
1702 : /// Obtains a temporary lease blocking garbage collection for the given LSN.
1703 : ///
1704 : /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
1705 : /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
1706 : ///
1707 : /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
1708 : /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
1709 84 : fn make_lsn_lease(
1710 84 : &self,
1711 84 : lsn: Lsn,
1712 84 : length: Duration,
1713 84 : init: bool,
1714 84 : _ctx: &RequestContext,
1715 84 : ) -> anyhow::Result<LsnLease> {
1716 72 : let lease = {
1717 : // Normalize the requested LSN to be aligned, and move to the first record
1718 : // if it points to the beginning of the page (header).
1719 84 : let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
1720 84 :
1721 84 : let mut gc_info = self.gc_info.write().unwrap();
1722 84 : let planned_cutoff = gc_info.min_cutoff();
1723 84 :
1724 84 : let valid_until = SystemTime::now() + length;
1725 84 :
1726 84 : let entry = gc_info.leases.entry(lsn);
1727 84 :
1728 84 : match entry {
1729 36 : Entry::Occupied(mut occupied) => {
1730 36 : let existing_lease = occupied.get_mut();
1731 36 : if valid_until > existing_lease.valid_until {
1732 12 : existing_lease.valid_until = valid_until;
1733 12 : let dt: DateTime<Utc> = valid_until.into();
1734 12 : info!("lease extended to {}", dt);
1735 : } else {
1736 24 : let dt: DateTime<Utc> = existing_lease.valid_until.into();
1737 24 : info!("existing lease covers greater length, valid until {}", dt);
1738 : }
1739 :
1740 36 : existing_lease.clone()
1741 : }
1742 48 : Entry::Vacant(vacant) => {
1743 : // Reject already GC-ed LSN if we are in AttachedSingle and
1744 : // not blocked by the lsn lease deadline.
1745 48 : let validate = {
1746 48 : let conf = self.tenant_conf.load();
1747 48 : conf.location.attach_mode == AttachmentMode::Single
1748 48 : && !conf.is_gc_blocked_by_lsn_lease_deadline()
1749 : };
1750 :
1751 48 : if init || validate {
1752 48 : let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
1753 48 : if lsn < *latest_gc_cutoff_lsn {
1754 12 : bail!(
1755 12 : "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
1756 12 : lsn,
1757 12 : *latest_gc_cutoff_lsn
1758 12 : );
1759 36 : }
1760 36 : if lsn < planned_cutoff {
1761 0 : bail!(
1762 0 : "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
1763 0 : lsn,
1764 0 : planned_cutoff
1765 0 : );
1766 36 : }
1767 0 : }
1768 :
1769 36 : let dt: DateTime<Utc> = valid_until.into();
1770 36 : info!("lease created, valid until {}", dt);
1771 36 : vacant.insert(LsnLease { valid_until }).clone()
1772 : }
1773 : }
1774 : };
1775 :
1776 72 : Ok(lease)
1777 84 : }
1778 :
1779 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1780 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1781 : #[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))]
1782 : pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
1783 : self.freeze0().await
1784 : }
1785 :
1786 : /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
1787 : #[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))]
1788 : pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
1789 : self.freeze_and_flush0().await
1790 : }
1791 :
1792 : /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
1793 : /// Returns the flush request ID which can be awaited with wait_flush_completion().
1794 6792 : pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
1795 6792 : let mut g = self.write_lock.lock().await;
1796 6792 : let to_lsn = self.get_last_record_lsn();
1797 6792 : self.freeze_inmem_layer_at(to_lsn, &mut g).await
1798 6792 : }
1799 :
1800 : // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
1801 : // polluting the span hierarchy.
1802 6792 : pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
1803 6792 : let token = self.freeze0().await?;
1804 6792 : self.wait_flush_completion(token).await
1805 6792 : }
1806 :
1807 : // Check if an open ephemeral layer should be closed: this provides
1808 : // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
1809 : // an ephemeral layer open forever when idle. It also freezes layers if the global limit on
1810 : // ephemeral layer bytes has been breached.
1811 0 : pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
1812 0 : let Ok(mut write_guard) = self.write_lock.try_lock() else {
1813 : // If the write lock is held, there is an active wal receiver: rolling open layers
1814 : // is their responsibility while they hold this lock.
1815 0 : return;
1816 : };
1817 :
1818 : // FIXME: why not early exit? because before #7927 the state would had been cleared every
1819 : // time, and this was missed.
1820 : // if write_guard.is_none() { return; }
1821 :
1822 0 : let Ok(layers_guard) = self.layers.try_read() else {
1823 : // Don't block if the layer lock is busy
1824 0 : return;
1825 : };
1826 :
1827 0 : let Ok(lm) = layers_guard.layer_map() else {
1828 0 : return;
1829 : };
1830 :
1831 0 : let Some(open_layer) = &lm.open_layer else {
1832 : // If there is no open layer, we have no layer freezing to do. However, we might need to generate
1833 : // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
1834 : // that didn't result in writes to this shard.
1835 :
1836 : // Must not hold the layers lock while waiting for a flush.
1837 0 : drop(layers_guard);
1838 0 :
1839 0 : let last_record_lsn = self.get_last_record_lsn();
1840 0 : let disk_consistent_lsn = self.get_disk_consistent_lsn();
1841 0 : if last_record_lsn > disk_consistent_lsn {
1842 : // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
1843 : // we are a sharded tenant and have skipped some WAL
1844 0 : let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
1845 0 : if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
1846 : // Only do this if have been layer-less longer than get_checkpoint_timeout, so that a shard
1847 : // without any data ingested (yet) doesn't write a remote index as soon as it
1848 : // sees its LSN advance: we only do this if we've been layer-less
1849 : // for some time.
1850 0 : tracing::debug!(
1851 0 : "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
1852 : disk_consistent_lsn,
1853 : last_record_lsn
1854 : );
1855 :
1856 : // The flush loop will update remote consistent LSN as well as disk consistent LSN.
1857 : // We know there is no open layer, so we can request freezing without actually
1858 : // freezing anything. This is true even if we have dropped the layers_guard, we
1859 : // still hold the write_guard.
1860 0 : let _ = async {
1861 0 : let token = self
1862 0 : .freeze_inmem_layer_at(last_record_lsn, &mut write_guard)
1863 0 : .await?;
1864 0 : self.wait_flush_completion(token).await
1865 0 : }
1866 0 : .await;
1867 0 : }
1868 0 : }
1869 :
1870 0 : return;
1871 : };
1872 :
1873 0 : let Some(current_size) = open_layer.try_len() else {
1874 : // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
1875 : // read lock to get size should always succeed.
1876 0 : tracing::warn!("Lock conflict while reading size of open layer");
1877 0 : return;
1878 : };
1879 :
1880 0 : let current_lsn = self.get_last_record_lsn();
1881 :
1882 0 : let checkpoint_distance_override = open_layer.tick().await;
1883 :
1884 0 : if let Some(size_override) = checkpoint_distance_override {
1885 0 : if current_size > size_override {
1886 : // This is not harmful, but it only happens in relatively rare cases where
1887 : // time-based checkpoints are not happening fast enough to keep the amount of
1888 : // ephemeral data within configured limits. It's a sign of stress on the system.
1889 0 : tracing::info!(
1890 0 : "Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure"
1891 : );
1892 0 : }
1893 0 : }
1894 :
1895 0 : let checkpoint_distance =
1896 0 : checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
1897 0 :
1898 0 : if self.should_roll(
1899 0 : current_size,
1900 0 : current_size,
1901 0 : checkpoint_distance,
1902 0 : self.get_last_record_lsn(),
1903 0 : self.last_freeze_at.load(),
1904 0 : open_layer.get_opened_at(),
1905 0 : ) {
1906 0 : match open_layer.info() {
1907 0 : InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
1908 0 : // We may reach this point if the layer was already frozen by not yet flushed: flushing
1909 0 : // happens asynchronously in the background.
1910 0 : tracing::debug!(
1911 0 : "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
1912 : );
1913 : }
1914 : InMemoryLayerInfo::Open { .. } => {
1915 : // Upgrade to a write lock and freeze the layer
1916 0 : drop(layers_guard);
1917 0 : let res = self
1918 0 : .freeze_inmem_layer_at(current_lsn, &mut write_guard)
1919 0 : .await;
1920 :
1921 0 : if let Err(e) = res {
1922 0 : tracing::info!(
1923 0 : "failed to flush frozen layer after background freeze: {e:#}"
1924 : );
1925 0 : }
1926 : }
1927 : }
1928 0 : }
1929 0 : }
1930 :
1931 : /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
1932 : ///
1933 : /// This is neccessary but not sufficient for offloading of the timeline as it might have
1934 : /// child timelines that are not offloaded yet.
1935 0 : pub(crate) fn can_offload(&self) -> (bool, &'static str) {
1936 0 : if self.remote_client.is_archived() != Some(true) {
1937 0 : return (false, "the timeline is not archived");
1938 0 : }
1939 0 : if !self.remote_client.no_pending_work() {
1940 : // if the remote client is still processing some work, we can't offload
1941 0 : return (false, "the upload queue is not drained yet");
1942 0 : }
1943 0 :
1944 0 : (true, "ok")
1945 0 : }
1946 :
1947 : /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
1948 : /// compaction tasks.
1949 2184 : pub(crate) async fn compact(
1950 2184 : self: &Arc<Self>,
1951 2184 : cancel: &CancellationToken,
1952 2184 : flags: EnumSet<CompactFlags>,
1953 2184 : ctx: &RequestContext,
1954 2184 : ) -> Result<CompactionOutcome, CompactionError> {
1955 2184 : let res = self
1956 2184 : .compact_with_options(
1957 2184 : cancel,
1958 2184 : CompactOptions {
1959 2184 : flags,
1960 2184 : compact_key_range: None,
1961 2184 : compact_lsn_range: None,
1962 2184 : sub_compaction: false,
1963 2184 : sub_compaction_max_job_size_mb: None,
1964 2184 : },
1965 2184 : ctx,
1966 2184 : )
1967 2184 : .await;
1968 2184 : if let Err(err) = &res {
1969 0 : log_compaction_error(err, None, cancel.is_cancelled(), false);
1970 2184 : }
1971 2184 : res
1972 2184 : }
1973 :
1974 : /// Outermost timeline compaction operation; downloads needed layers.
1975 : ///
1976 : /// NB: the cancellation token is usually from a background task, but can also come from a
1977 : /// request task.
1978 2184 : pub(crate) async fn compact_with_options(
1979 2184 : self: &Arc<Self>,
1980 2184 : cancel: &CancellationToken,
1981 2184 : options: CompactOptions,
1982 2184 : ctx: &RequestContext,
1983 2184 : ) -> Result<CompactionOutcome, CompactionError> {
1984 2184 : // Acquire the compaction lock and task semaphore.
1985 2184 : //
1986 2184 : // L0-only compaction uses a separate semaphore (if enabled) to make sure it isn't starved
1987 2184 : // out by other background tasks (including image compaction). We request this via
1988 2184 : // `BackgroundLoopKind::L0Compaction`.
1989 2184 : //
1990 2184 : // Yield for pending L0 compaction while waiting for the semaphore.
1991 2184 : let is_l0_only = options.flags.contains(CompactFlags::OnlyL0Compaction);
1992 2184 : let semaphore_kind = match is_l0_only && self.get_compaction_l0_semaphore() {
1993 0 : true => BackgroundLoopKind::L0Compaction,
1994 2184 : false => BackgroundLoopKind::Compaction,
1995 : };
1996 2184 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
1997 2184 : if yield_for_l0 {
1998 : // If this is an L0 pass, it doesn't make sense to yield for L0.
1999 0 : debug_assert!(!is_l0_only, "YieldForL0 during L0 pass");
2000 : // If `compaction_l0_first` is disabled, there's no point yielding.
2001 0 : debug_assert!(self.get_compaction_l0_first(), "YieldForL0 without L0 pass");
2002 2184 : }
2003 :
2004 2184 : let acquire = async move {
2005 2184 : let guard = self.compaction_lock.lock().await;
2006 2184 : let permit = super::tasks::acquire_concurrency_permit(semaphore_kind, ctx).await;
2007 2184 : (guard, permit)
2008 2184 : };
2009 :
2010 2184 : let (_guard, _permit) = tokio::select! {
2011 2184 : (guard, permit) = acquire => (guard, permit),
2012 2184 : _ = self.l0_compaction_trigger.notified(), if yield_for_l0 => {
2013 0 : return Ok(CompactionOutcome::YieldForL0);
2014 : }
2015 2184 : _ = self.cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2016 2184 : _ = cancel.cancelled() => return Ok(CompactionOutcome::Skipped),
2017 : };
2018 :
2019 2184 : let last_record_lsn = self.get_last_record_lsn();
2020 2184 :
2021 2184 : // Last record Lsn could be zero in case the timeline was just created
2022 2184 : if !last_record_lsn.is_valid() {
2023 0 : warn!(
2024 0 : "Skipping compaction for potentially just initialized timeline, it has invalid last record lsn: {last_record_lsn}"
2025 : );
2026 0 : return Ok(CompactionOutcome::Skipped);
2027 2184 : }
2028 :
2029 2184 : let result = match self.get_compaction_algorithm_settings().kind {
2030 : CompactionAlgorithm::Tiered => {
2031 0 : self.compact_tiered(cancel, ctx).await?;
2032 0 : Ok(CompactionOutcome::Done)
2033 : }
2034 2184 : CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
2035 : };
2036 :
2037 : // Signal compaction failure to avoid L0 flush stalls when it's broken.
2038 0 : match &result {
2039 2184 : Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
2040 0 : Err(e) if e.is_cancel() => {}
2041 0 : Err(CompactionError::ShuttingDown) => {
2042 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
2043 0 : }
2044 0 : Err(CompactionError::AlreadyRunning(_)) => {
2045 0 : // Covered by the `Err(e) if e.is_cancel()` branch.
2046 0 : }
2047 : Err(CompactionError::Other(_)) => {
2048 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
2049 : }
2050 : Err(CompactionError::CollectKeySpaceError(_)) => {
2051 : // Cancelled errors are covered by the `Err(e) if e.is_cancel()` branch.
2052 0 : self.compaction_failed.store(true, AtomicOrdering::Relaxed)
2053 : }
2054 : // Don't change the current value on offload failure or shutdown. We don't want to
2055 : // abruptly stall nor resume L0 flushes in these cases.
2056 0 : Err(CompactionError::Offload(_)) => {}
2057 : };
2058 :
2059 2184 : result
2060 2184 : }
2061 :
2062 : /// Mutate the timeline with a [`TimelineWriter`].
2063 30799212 : pub(crate) async fn writer(&self) -> TimelineWriter<'_> {
2064 30799212 : TimelineWriter {
2065 30799212 : tl: self,
2066 30799212 : write_guard: self.write_lock.lock().await,
2067 : }
2068 30799212 : }
2069 :
2070 0 : pub(crate) fn activate(
2071 0 : self: &Arc<Self>,
2072 0 : parent: Arc<crate::tenant::TenantShard>,
2073 0 : broker_client: BrokerClientChannel,
2074 0 : background_jobs_can_start: Option<&completion::Barrier>,
2075 0 : ctx: &RequestContext,
2076 0 : ) {
2077 0 : if self.tenant_shard_id.is_shard_zero() {
2078 0 : // Logical size is only maintained accurately on shard zero.
2079 0 : self.spawn_initial_logical_size_computation_task(ctx);
2080 0 : }
2081 0 : self.launch_wal_receiver(ctx, broker_client);
2082 0 : self.set_state(TimelineState::Active);
2083 0 : self.launch_eviction_task(parent, background_jobs_can_start);
2084 0 : }
2085 :
2086 : /// After this function returns, there are no timeline-scoped tasks are left running.
2087 : ///
2088 : /// The preferred pattern for is:
2089 : /// - in any spawned tasks, keep Timeline::guard open + Timeline::cancel / child token
2090 : /// - if early shutdown (not just cancellation) of a sub-tree of tasks is required,
2091 : /// go the extra mile and keep track of JoinHandles
2092 : /// - Keep track of JoinHandles using a passed-down `Arc<Mutex<Option<JoinSet>>>` or similar,
2093 : /// instead of spawning directly on a runtime. It is a more composable / testable pattern.
2094 : ///
2095 : /// For legacy reasons, we still have multiple tasks spawned using
2096 : /// `task_mgr::spawn(X, Some(tenant_id), Some(timeline_id))`.
2097 : /// We refer to these as "timeline-scoped task_mgr tasks".
2098 : /// Some of these tasks are already sensitive to Timeline::cancel while others are
2099 : /// not sensitive to Timeline::cancel and instead respect [`task_mgr::shutdown_token`]
2100 : /// or [`task_mgr::shutdown_watcher`].
2101 : /// We want to gradually convert the code base away from these.
2102 : ///
2103 : /// Here is an inventory of timeline-scoped task_mgr tasks that are still sensitive to
2104 : /// `task_mgr::shutdown_{token,watcher}` (there are also tenant-scoped and global-scoped
2105 : /// ones that aren't mentioned here):
2106 : /// - [`TaskKind::TimelineDeletionWorker`]
2107 : /// - NB: also used for tenant deletion
2108 : /// - [`TaskKind::RemoteUploadTask`]`
2109 : /// - [`TaskKind::InitialLogicalSizeCalculation`]
2110 : /// - [`TaskKind::DownloadAllRemoteLayers`] (can we get rid of it?)
2111 : // Inventory of timeline-scoped task_mgr tasks that use spawn but aren't sensitive:
2112 : /// - [`TaskKind::Eviction`]
2113 : /// - [`TaskKind::LayerFlushTask`]
2114 : /// - [`TaskKind::OndemandLogicalSizeCalculation`]
2115 : /// - [`TaskKind::GarbageCollector`] (immediate_gc is timeline-scoped)
2116 60 : pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
2117 60 : debug_assert_current_span_has_tenant_and_timeline_id();
2118 60 :
2119 60 : // Regardless of whether we're going to try_freeze_and_flush
2120 60 : // or not, stop ingesting any more data. Walreceiver only provides
2121 60 : // cancellation but no "wait until gone", because it uses the Timeline::gate.
2122 60 : // So, only after the self.gate.close() below will we know for sure that
2123 60 : // no walreceiver tasks are left.
2124 60 : // For `try_freeze_and_flush=true`, this means that we might still be ingesting
2125 60 : // data during the call to `self.freeze_and_flush()` below.
2126 60 : // That's not ideal, but, we don't have the concept of a ChildGuard,
2127 60 : // which is what we'd need to properly model early shutdown of the walreceiver
2128 60 : // task sub-tree before the other Timeline task sub-trees.
2129 60 : let walreceiver = self.walreceiver.lock().unwrap().take();
2130 60 : tracing::debug!(
2131 0 : is_some = walreceiver.is_some(),
2132 0 : "Waiting for WalReceiverManager..."
2133 : );
2134 60 : if let Some(walreceiver) = walreceiver {
2135 0 : walreceiver.cancel();
2136 60 : }
2137 : // ... and inform any waiters for newer LSNs that there won't be any.
2138 60 : self.last_record_lsn.shutdown();
2139 60 :
2140 60 : if let ShutdownMode::FreezeAndFlush = mode {
2141 36 : let do_flush = if let Some((open, frozen)) = self
2142 36 : .layers
2143 36 : .read()
2144 36 : .await
2145 36 : .layer_map()
2146 36 : .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
2147 36 : .ok()
2148 36 : .filter(|(open, frozen)| *open || *frozen > 0)
2149 : {
2150 0 : if self.remote_client.is_archived() == Some(true) {
2151 : // No point flushing on shutdown for an archived timeline: it is not important
2152 : // to have it nice and fresh after our restart, and trying to flush here might
2153 : // race with trying to offload it (which also stops the flush loop)
2154 0 : false
2155 : } else {
2156 0 : tracing::info!(?open, frozen, "flushing and freezing on shutdown");
2157 0 : true
2158 : }
2159 : } else {
2160 : // this is double-shutdown, it'll be a no-op
2161 36 : true
2162 : };
2163 :
2164 : // we shut down walreceiver above, so, we won't add anything more
2165 : // to the InMemoryLayer; freeze it and wait for all frozen layers
2166 : // to reach the disk & upload queue, then shut the upload queue and
2167 : // wait for it to drain.
2168 36 : if do_flush {
2169 36 : match self.freeze_and_flush().await {
2170 : Ok(_) => {
2171 : // drain the upload queue
2172 : // if we did not wait for completion here, it might be our shutdown process
2173 : // didn't wait for remote uploads to complete at all, as new tasks can forever
2174 : // be spawned.
2175 : //
2176 : // what is problematic is the shutting down of RemoteTimelineClient, because
2177 : // obviously it does not make sense to stop while we wait for it, but what
2178 : // about corner cases like s3 suddenly hanging up?
2179 36 : self.remote_client.shutdown().await;
2180 : }
2181 : Err(FlushLayerError::Cancelled) => {
2182 : // this is likely the second shutdown, ignore silently.
2183 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2184 0 : debug_assert!(self.cancel.is_cancelled());
2185 : }
2186 0 : Err(e) => {
2187 0 : // Non-fatal. Shutdown is infallible. Failures to flush just mean that
2188 0 : // we have some extra WAL replay to do next time the timeline starts.
2189 0 : warn!("failed to freeze and flush: {e:#}");
2190 : }
2191 : }
2192 :
2193 : // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
2194 : // we also do a final check here to ensure that the queue is empty.
2195 36 : if !self.remote_client.no_pending_work() {
2196 0 : warn!(
2197 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2198 : );
2199 36 : }
2200 0 : }
2201 24 : }
2202 :
2203 60 : if let ShutdownMode::Reload = mode {
2204 : // drain the upload queue
2205 12 : self.remote_client.shutdown().await;
2206 12 : if !self.remote_client.no_pending_work() {
2207 0 : warn!(
2208 0 : "still have pending work in remote upload queue, but continuing shutting down anyways"
2209 : );
2210 12 : }
2211 48 : }
2212 :
2213 : // Signal any subscribers to our cancellation token to drop out
2214 60 : tracing::debug!("Cancelling CancellationToken");
2215 60 : self.cancel.cancel();
2216 60 :
2217 60 : // If we have a background task downloading heatmap layers stop it.
2218 60 : // The background downloads are sensitive to timeline cancellation (done above),
2219 60 : // so the drain will be immediate.
2220 60 : self.stop_and_drain_heatmap_layers_download().await;
2221 :
2222 : // Ensure Prevent new page service requests from starting.
2223 60 : self.handles.shutdown();
2224 60 :
2225 60 : // Transition the remote_client into a state where it's only useful for timeline deletion.
2226 60 : // (The deletion use case is why we can't just hook up remote_client to Self::cancel).)
2227 60 : self.remote_client.stop();
2228 60 :
2229 60 : // As documented in remote_client.stop()'s doc comment, it's our responsibility
2230 60 : // to shut down the upload queue tasks.
2231 60 : // TODO: fix that, task management should be encapsulated inside remote_client.
2232 60 : task_mgr::shutdown_tasks(
2233 60 : Some(TaskKind::RemoteUploadTask),
2234 60 : Some(self.tenant_shard_id),
2235 60 : Some(self.timeline_id),
2236 60 : )
2237 60 : .await;
2238 :
2239 : // TODO: work toward making this a no-op. See this function's doc comment for more context.
2240 60 : tracing::debug!("Waiting for tasks...");
2241 60 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), Some(self.timeline_id)).await;
2242 :
2243 : {
2244 : // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
2245 : // open.
2246 60 : let mut write_guard = self.write_lock.lock().await;
2247 60 : self.layers.write().await.shutdown(&mut write_guard);
2248 60 : }
2249 60 :
2250 60 : // Finally wait until any gate-holders are complete.
2251 60 : //
2252 60 : // TODO: once above shutdown_tasks is a no-op, we can close the gate before calling shutdown_tasks
2253 60 : // and use a TBD variant of shutdown_tasks that asserts that there were no tasks left.
2254 60 : self.gate.close().await;
2255 :
2256 60 : self.metrics.shutdown();
2257 60 : }
2258 :
2259 2796 : pub(crate) fn set_state(&self, new_state: TimelineState) {
2260 2796 : match (self.current_state(), new_state) {
2261 2796 : (equal_state_1, equal_state_2) if equal_state_1 == equal_state_2 => {
2262 12 : info!("Ignoring new state, equal to the existing one: {equal_state_2:?}");
2263 : }
2264 0 : (st, TimelineState::Loading) => {
2265 0 : error!("ignoring transition from {st:?} into Loading state");
2266 : }
2267 0 : (TimelineState::Broken { .. }, new_state) => {
2268 0 : error!("Ignoring state update {new_state:?} for broken timeline");
2269 : }
2270 : (TimelineState::Stopping, TimelineState::Active) => {
2271 0 : error!("Not activating a Stopping timeline");
2272 : }
2273 2784 : (_, new_state) => {
2274 2784 : self.state.send_replace(new_state);
2275 2784 : }
2276 : }
2277 2796 : }
2278 :
2279 12 : pub(crate) fn set_broken(&self, reason: String) {
2280 12 : let backtrace_str: String = format!("{}", std::backtrace::Backtrace::force_capture());
2281 12 : let broken_state = TimelineState::Broken {
2282 12 : reason,
2283 12 : backtrace: backtrace_str,
2284 12 : };
2285 12 : self.set_state(broken_state);
2286 12 :
2287 12 : // Although the Broken state is not equivalent to shutdown() (shutdown will be called
2288 12 : // later when this tenant is detach or the process shuts down), firing the cancellation token
2289 12 : // here avoids the need for other tasks to watch for the Broken state explicitly.
2290 12 : self.cancel.cancel();
2291 12 : }
2292 :
2293 1359412 : pub(crate) fn current_state(&self) -> TimelineState {
2294 1359412 : self.state.borrow().clone()
2295 1359412 : }
2296 :
2297 36 : pub(crate) fn is_broken(&self) -> bool {
2298 36 : matches!(&*self.state.borrow(), TimelineState::Broken { .. })
2299 36 : }
2300 :
2301 1512 : pub(crate) fn is_active(&self) -> bool {
2302 1512 : self.current_state() == TimelineState::Active
2303 1512 : }
2304 :
2305 96 : pub(crate) fn is_archived(&self) -> Option<bool> {
2306 96 : self.remote_client.is_archived()
2307 96 : }
2308 :
2309 96 : pub(crate) fn is_invisible(&self) -> Option<bool> {
2310 96 : self.remote_client.is_invisible()
2311 96 : }
2312 :
2313 2208 : pub(crate) fn is_stopping(&self) -> bool {
2314 2208 : self.current_state() == TimelineState::Stopping
2315 2208 : }
2316 :
2317 0 : pub(crate) fn subscribe_for_state_updates(&self) -> watch::Receiver<TimelineState> {
2318 0 : self.state.subscribe()
2319 0 : }
2320 :
2321 1352908 : pub(crate) async fn wait_to_become_active(
2322 1352908 : &self,
2323 1352908 : _ctx: &RequestContext, // Prepare for use by cancellation
2324 1352908 : ) -> Result<(), TimelineState> {
2325 1352908 : let mut receiver = self.state.subscribe();
2326 : loop {
2327 1352908 : let current_state = receiver.borrow().clone();
2328 1352908 : match current_state {
2329 : TimelineState::Loading => {
2330 0 : receiver
2331 0 : .changed()
2332 0 : .await
2333 0 : .expect("holding a reference to self");
2334 : }
2335 : TimelineState::Active => {
2336 1352896 : return Ok(());
2337 : }
2338 : TimelineState::Broken { .. } | TimelineState::Stopping => {
2339 : // There's no chance the timeline can transition back into ::Active
2340 12 : return Err(current_state);
2341 : }
2342 : }
2343 : }
2344 1352908 : }
2345 :
2346 0 : pub(crate) async fn layer_map_info(
2347 0 : &self,
2348 0 : reset: LayerAccessStatsReset,
2349 0 : ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
2350 0 : let guard = self.layers.read().await;
2351 0 : let layer_map = guard.layer_map()?;
2352 0 : let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
2353 0 : if let Some(open_layer) = &layer_map.open_layer {
2354 0 : in_memory_layers.push(open_layer.info());
2355 0 : }
2356 0 : for frozen_layer in &layer_map.frozen_layers {
2357 0 : in_memory_layers.push(frozen_layer.info());
2358 0 : }
2359 :
2360 0 : let historic_layers = layer_map
2361 0 : .iter_historic_layers()
2362 0 : .map(|desc| guard.get_from_desc(&desc).info(reset))
2363 0 : .collect();
2364 0 :
2365 0 : Ok(LayerMapInfo {
2366 0 : in_memory_layers,
2367 0 : historic_layers,
2368 0 : })
2369 0 : }
2370 :
2371 : #[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))]
2372 : pub(crate) async fn download_layer(
2373 : &self,
2374 : layer_file_name: &LayerName,
2375 : ctx: &RequestContext,
2376 : ) -> Result<Option<bool>, super::storage_layer::layer::DownloadError> {
2377 : let Some(layer) = self
2378 : .find_layer(layer_file_name)
2379 : .await
2380 0 : .map_err(|e| match e {
2381 0 : layer_manager::Shutdown => {
2382 0 : super::storage_layer::layer::DownloadError::TimelineShutdown
2383 0 : }
2384 0 : })?
2385 : else {
2386 : return Ok(None);
2387 : };
2388 :
2389 : layer.download(ctx).await?;
2390 :
2391 : Ok(Some(true))
2392 : }
2393 :
2394 : /// Evict just one layer.
2395 : ///
2396 : /// Returns `Ok(None)` in the case where the layer could not be found by its `layer_file_name`.
2397 0 : pub(crate) async fn evict_layer(
2398 0 : &self,
2399 0 : layer_file_name: &LayerName,
2400 0 : ) -> anyhow::Result<Option<bool>> {
2401 0 : let _gate = self
2402 0 : .gate
2403 0 : .enter()
2404 0 : .map_err(|_| anyhow::anyhow!("Shutting down"))?;
2405 :
2406 0 : let Some(local_layer) = self.find_layer(layer_file_name).await? else {
2407 0 : return Ok(None);
2408 : };
2409 :
2410 : // curl has this by default
2411 0 : let timeout = std::time::Duration::from_secs(120);
2412 0 :
2413 0 : match local_layer.evict_and_wait(timeout).await {
2414 0 : Ok(()) => Ok(Some(true)),
2415 0 : Err(EvictionError::NotFound) => Ok(Some(false)),
2416 0 : Err(EvictionError::Downloaded) => Ok(Some(false)),
2417 0 : Err(EvictionError::Timeout) => Ok(Some(false)),
2418 : }
2419 0 : }
2420 :
2421 28818060 : fn should_roll(
2422 28818060 : &self,
2423 28818060 : layer_size: u64,
2424 28818060 : projected_layer_size: u64,
2425 28818060 : checkpoint_distance: u64,
2426 28818060 : projected_lsn: Lsn,
2427 28818060 : last_freeze_at: Lsn,
2428 28818060 : opened_at: Instant,
2429 28818060 : ) -> bool {
2430 28818060 : let distance = projected_lsn.widening_sub(last_freeze_at);
2431 28818060 :
2432 28818060 : // Rolling the open layer can be triggered by:
2433 28818060 : // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
2434 28818060 : // the safekeepers need to store. For sharded tenants, we multiply by shard count to
2435 28818060 : // account for how writes are distributed across shards: we expect each node to consume
2436 28818060 : // 1/count of the LSN on average.
2437 28818060 : // 2. The size of the currently open layer.
2438 28818060 : // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
2439 28818060 : // up and suspend activity.
2440 28818060 : if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
2441 0 : info!(
2442 0 : "Will roll layer at {} with layer size {} due to LSN distance ({})",
2443 : projected_lsn, layer_size, distance
2444 : );
2445 :
2446 0 : true
2447 28818060 : } else if projected_layer_size >= checkpoint_distance {
2448 : // NB: this check is relied upon by:
2449 480 : let _ = IndexEntry::validate_checkpoint_distance;
2450 480 : info!(
2451 0 : "Will roll layer at {} with layer size {} due to layer size ({})",
2452 : projected_lsn, layer_size, projected_layer_size
2453 : );
2454 :
2455 480 : true
2456 28817580 : } else if distance > 0 && opened_at.elapsed() >= self.get_checkpoint_timeout() {
2457 0 : info!(
2458 0 : "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
2459 0 : projected_lsn,
2460 0 : layer_size,
2461 0 : opened_at.elapsed()
2462 : );
2463 :
2464 0 : true
2465 : } else {
2466 28817580 : false
2467 : }
2468 28818060 : }
2469 : }
2470 :
2471 : /// Number of times we will compute partition within a checkpoint distance.
2472 : const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
2473 :
2474 : // Private functions
2475 : impl Timeline {
2476 72 : pub(crate) fn get_lsn_lease_length(&self) -> Duration {
2477 72 : let tenant_conf = self.tenant_conf.load();
2478 72 : tenant_conf
2479 72 : .tenant_conf
2480 72 : .lsn_lease_length
2481 72 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2482 72 : }
2483 :
2484 0 : pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
2485 0 : let tenant_conf = self.tenant_conf.load();
2486 0 : tenant_conf
2487 0 : .tenant_conf
2488 0 : .lsn_lease_length_for_ts
2489 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
2490 0 : }
2491 :
2492 0 : pub(crate) fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
2493 0 : let tenant_conf = self.tenant_conf.load();
2494 0 : tenant_conf.is_gc_blocked_by_lsn_lease_deadline()
2495 0 : }
2496 :
2497 0 : pub(crate) fn get_lazy_slru_download(&self) -> bool {
2498 0 : let tenant_conf = self.tenant_conf.load();
2499 0 : tenant_conf
2500 0 : .tenant_conf
2501 0 : .lazy_slru_download
2502 0 : .unwrap_or(self.conf.default_tenant_conf.lazy_slru_download)
2503 0 : }
2504 :
2505 : /// Checks if a get page request should get perf tracing
2506 : ///
2507 : /// The configuration priority is: tenant config override, default tenant config,
2508 : /// pageserver config.
2509 0 : pub(crate) fn is_get_page_request_sampled(&self) -> bool {
2510 0 : let tenant_conf = self.tenant_conf.load();
2511 0 : let ratio = tenant_conf
2512 0 : .tenant_conf
2513 0 : .sampling_ratio
2514 0 : .flatten()
2515 0 : .or(self.conf.default_tenant_conf.sampling_ratio)
2516 0 : .or(self.conf.tracing.as_ref().map(|t| t.sampling_ratio));
2517 0 :
2518 0 : match ratio {
2519 0 : Some(r) => {
2520 0 : if r.numerator == 0 {
2521 0 : false
2522 : } else {
2523 0 : rand::thread_rng().gen_range(0..r.denominator) < r.numerator
2524 : }
2525 : }
2526 0 : None => false,
2527 : }
2528 0 : }
2529 :
2530 28827156 : fn get_checkpoint_distance(&self) -> u64 {
2531 28827156 : let tenant_conf = self.tenant_conf.load();
2532 28827156 : tenant_conf
2533 28827156 : .tenant_conf
2534 28827156 : .checkpoint_distance
2535 28827156 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2536 28827156 : }
2537 :
2538 28817580 : fn get_checkpoint_timeout(&self) -> Duration {
2539 28817580 : let tenant_conf = self.tenant_conf.load();
2540 28817580 : tenant_conf
2541 28817580 : .tenant_conf
2542 28817580 : .checkpoint_timeout
2543 28817580 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2544 28817580 : }
2545 :
2546 15168 : fn get_compaction_period(&self) -> Duration {
2547 15168 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2548 15168 : tenant_conf
2549 15168 : .compaction_period
2550 15168 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2551 15168 : }
2552 :
2553 4068 : fn get_compaction_target_size(&self) -> u64 {
2554 4068 : let tenant_conf = self.tenant_conf.load();
2555 4068 : tenant_conf
2556 4068 : .tenant_conf
2557 4068 : .compaction_target_size
2558 4068 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2559 4068 : }
2560 :
2561 9480 : fn get_compaction_threshold(&self) -> usize {
2562 9480 : let tenant_conf = self.tenant_conf.load();
2563 9480 : tenant_conf
2564 9480 : .tenant_conf
2565 9480 : .compaction_threshold
2566 9480 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2567 9480 : }
2568 :
2569 : /// Returns `true` if the rel_size_v2 config is enabled. NOTE: the write path and read path
2570 : /// should look at `get_rel_size_v2_status()` to get the actual status of the timeline. It is
2571 : /// possible that the index part persists the state while the config doesn't get persisted.
2572 11676 : pub(crate) fn get_rel_size_v2_enabled(&self) -> bool {
2573 11676 : let tenant_conf = self.tenant_conf.load();
2574 11676 : tenant_conf
2575 11676 : .tenant_conf
2576 11676 : .rel_size_v2_enabled
2577 11676 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
2578 11676 : }
2579 :
2580 13188 : pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
2581 13188 : self.rel_size_v2_status
2582 13188 : .load()
2583 13188 : .as_ref()
2584 13188 : .map(|s| s.as_ref().clone())
2585 13188 : .unwrap_or(RelSizeMigration::Legacy)
2586 13188 : }
2587 :
2588 168 : fn get_compaction_upper_limit(&self) -> usize {
2589 168 : let tenant_conf = self.tenant_conf.load();
2590 168 : tenant_conf
2591 168 : .tenant_conf
2592 168 : .compaction_upper_limit
2593 168 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
2594 168 : }
2595 :
2596 0 : pub fn get_compaction_l0_first(&self) -> bool {
2597 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2598 0 : tenant_conf
2599 0 : .compaction_l0_first
2600 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
2601 0 : }
2602 :
2603 0 : pub fn get_compaction_l0_semaphore(&self) -> bool {
2604 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2605 0 : tenant_conf
2606 0 : .compaction_l0_semaphore
2607 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_semaphore)
2608 0 : }
2609 :
2610 7584 : fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
2611 : // By default, delay L0 flushes at 3x the compaction threshold. The compaction threshold
2612 : // defaults to 10, and L0 compaction is generally able to keep L0 counts below 30.
2613 : const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 3;
2614 :
2615 : // If compaction is disabled, don't delay.
2616 7584 : if self.get_compaction_period() == Duration::ZERO {
2617 7572 : return None;
2618 12 : }
2619 12 :
2620 12 : let compaction_threshold = self.get_compaction_threshold();
2621 12 : let tenant_conf = self.tenant_conf.load();
2622 12 : let l0_flush_delay_threshold = tenant_conf
2623 12 : .tenant_conf
2624 12 : .l0_flush_delay_threshold
2625 12 : .or(self.conf.default_tenant_conf.l0_flush_delay_threshold)
2626 12 : .unwrap_or(DEFAULT_L0_FLUSH_DELAY_FACTOR * compaction_threshold);
2627 12 :
2628 12 : // 0 disables backpressure.
2629 12 : if l0_flush_delay_threshold == 0 {
2630 0 : return None;
2631 12 : }
2632 12 :
2633 12 : // Clamp the flush delay threshold to the compaction threshold; it doesn't make sense to
2634 12 : // backpressure flushes below this.
2635 12 : // TODO: the tenant config should have validation to prevent this instead.
2636 12 : debug_assert!(l0_flush_delay_threshold >= compaction_threshold);
2637 12 : Some(max(l0_flush_delay_threshold, compaction_threshold))
2638 7584 : }
2639 :
2640 7584 : fn get_l0_flush_stall_threshold(&self) -> Option<usize> {
2641 : // Disable L0 stalls by default. Stalling can cause unavailability if L0 compaction isn't
2642 : // responsive, and it can e.g. block on other compaction via the compaction semaphore or
2643 : // sibling timelines. We need more confidence before enabling this.
2644 : const DEFAULT_L0_FLUSH_STALL_FACTOR: usize = 0; // TODO: default to e.g. 5
2645 :
2646 : // If compaction is disabled, don't stall.
2647 7584 : if self.get_compaction_period() == Duration::ZERO {
2648 7572 : return None;
2649 12 : }
2650 12 :
2651 12 : // If compaction is failing, don't stall and try to keep the tenant alive. This may not be a
2652 12 : // good idea: read amp can grow unbounded, leading to terrible performance, and we may take
2653 12 : // on unbounded compaction debt that can take a long time to fix once compaction comes back
2654 12 : // online. At least we'll delay flushes, slowing down the growth and buying some time.
2655 12 : if self.compaction_failed.load(AtomicOrdering::Relaxed) {
2656 0 : return None;
2657 12 : }
2658 12 :
2659 12 : let compaction_threshold = self.get_compaction_threshold();
2660 12 : let tenant_conf = self.tenant_conf.load();
2661 12 : let l0_flush_stall_threshold = tenant_conf
2662 12 : .tenant_conf
2663 12 : .l0_flush_stall_threshold
2664 12 : .or(self.conf.default_tenant_conf.l0_flush_stall_threshold);
2665 12 :
2666 12 : // Tests sometimes set compaction_threshold=1 to generate lots of layer files, and don't
2667 12 : // handle the 20-second compaction delay. Some (e.g. `test_backward_compatibility`) can't
2668 12 : // easily adjust the L0 backpressure settings, so just disable stalls in this case.
2669 12 : if cfg!(feature = "testing")
2670 12 : && compaction_threshold == 1
2671 0 : && l0_flush_stall_threshold.is_none()
2672 : {
2673 0 : return None;
2674 12 : }
2675 12 :
2676 12 : let l0_flush_stall_threshold = l0_flush_stall_threshold
2677 12 : .unwrap_or(DEFAULT_L0_FLUSH_STALL_FACTOR * compaction_threshold);
2678 12 :
2679 12 : // 0 disables backpressure.
2680 12 : if l0_flush_stall_threshold == 0 {
2681 12 : return None;
2682 0 : }
2683 0 :
2684 0 : // Clamp the flush stall threshold to the compaction threshold; it doesn't make sense to
2685 0 : // backpressure flushes below this.
2686 0 : // TODO: the tenant config should have validation to prevent this instead.
2687 0 : debug_assert!(l0_flush_stall_threshold >= compaction_threshold);
2688 0 : Some(max(l0_flush_stall_threshold, compaction_threshold))
2689 7584 : }
2690 :
2691 84 : fn get_image_creation_threshold(&self) -> usize {
2692 84 : let tenant_conf = self.tenant_conf.load();
2693 84 : tenant_conf
2694 84 : .tenant_conf
2695 84 : .image_creation_threshold
2696 84 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2697 84 : }
2698 :
2699 2184 : fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
2700 2184 : let tenant_conf = &self.tenant_conf.load();
2701 2184 : tenant_conf
2702 2184 : .tenant_conf
2703 2184 : .compaction_algorithm
2704 2184 : .as_ref()
2705 2184 : .unwrap_or(&self.conf.default_tenant_conf.compaction_algorithm)
2706 2184 : .clone()
2707 2184 : }
2708 :
2709 2184 : pub fn get_compaction_shard_ancestor(&self) -> bool {
2710 2184 : let tenant_conf = self.tenant_conf.load();
2711 2184 : tenant_conf
2712 2184 : .tenant_conf
2713 2184 : .compaction_shard_ancestor
2714 2184 : .unwrap_or(self.conf.default_tenant_conf.compaction_shard_ancestor)
2715 2184 : }
2716 :
2717 0 : fn get_eviction_policy(&self) -> EvictionPolicy {
2718 0 : let tenant_conf = self.tenant_conf.load();
2719 0 : tenant_conf
2720 0 : .tenant_conf
2721 0 : .eviction_policy
2722 0 : .unwrap_or(self.conf.default_tenant_conf.eviction_policy)
2723 0 : }
2724 :
2725 2784 : fn get_evictions_low_residence_duration_metric_threshold(
2726 2784 : tenant_conf: &pageserver_api::models::TenantConfig,
2727 2784 : default_tenant_conf: &pageserver_api::config::TenantConfigToml,
2728 2784 : ) -> Duration {
2729 2784 : tenant_conf
2730 2784 : .evictions_low_residence_duration_metric_threshold
2731 2784 : .unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
2732 2784 : }
2733 :
2734 3480 : fn get_image_layer_creation_check_threshold(&self) -> u8 {
2735 3480 : let tenant_conf = self.tenant_conf.load();
2736 3480 : tenant_conf
2737 3480 : .tenant_conf
2738 3480 : .image_layer_creation_check_threshold
2739 3480 : .unwrap_or(
2740 3480 : self.conf
2741 3480 : .default_tenant_conf
2742 3480 : .image_layer_creation_check_threshold,
2743 3480 : )
2744 3480 : }
2745 :
2746 324 : fn get_gc_compaction_settings(&self) -> GcCompactionCombinedSettings {
2747 324 : let tenant_conf = &self.tenant_conf.load();
2748 324 : let gc_compaction_enabled = tenant_conf
2749 324 : .tenant_conf
2750 324 : .gc_compaction_enabled
2751 324 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_enabled);
2752 324 : let gc_compaction_verification = tenant_conf
2753 324 : .tenant_conf
2754 324 : .gc_compaction_verification
2755 324 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_verification);
2756 324 : let gc_compaction_initial_threshold_kb = tenant_conf
2757 324 : .tenant_conf
2758 324 : .gc_compaction_initial_threshold_kb
2759 324 : .unwrap_or(
2760 324 : self.conf
2761 324 : .default_tenant_conf
2762 324 : .gc_compaction_initial_threshold_kb,
2763 324 : );
2764 324 : let gc_compaction_ratio_percent = tenant_conf
2765 324 : .tenant_conf
2766 324 : .gc_compaction_ratio_percent
2767 324 : .unwrap_or(self.conf.default_tenant_conf.gc_compaction_ratio_percent);
2768 324 : GcCompactionCombinedSettings {
2769 324 : gc_compaction_enabled,
2770 324 : gc_compaction_verification,
2771 324 : gc_compaction_initial_threshold_kb,
2772 324 : gc_compaction_ratio_percent,
2773 324 : }
2774 324 : }
2775 :
2776 0 : fn get_image_creation_preempt_threshold(&self) -> usize {
2777 0 : let tenant_conf = self.tenant_conf.load();
2778 0 : tenant_conf
2779 0 : .tenant_conf
2780 0 : .image_creation_preempt_threshold
2781 0 : .unwrap_or(
2782 0 : self.conf
2783 0 : .default_tenant_conf
2784 0 : .image_creation_preempt_threshold,
2785 0 : )
2786 0 : }
2787 :
2788 : /// Resolve the effective WAL receiver protocol to use for this tenant.
2789 : ///
2790 : /// Priority order is:
2791 : /// 1. Tenant config override
2792 : /// 2. Default value for tenant config override
2793 : /// 3. Pageserver config override
2794 : /// 4. Pageserver config default
2795 0 : pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
2796 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2797 0 : tenant_conf
2798 0 : .wal_receiver_protocol_override
2799 0 : .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
2800 0 : .unwrap_or(self.conf.wal_receiver_protocol)
2801 0 : }
2802 :
2803 0 : pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
2804 0 : // NB: Most tenant conf options are read by background loops, so,
2805 0 : // changes will automatically be picked up.
2806 0 :
2807 0 : // The threshold is embedded in the metric. So, we need to update it.
2808 0 : {
2809 0 : let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
2810 0 : &new_conf.tenant_conf,
2811 0 : &self.conf.default_tenant_conf,
2812 0 : );
2813 0 :
2814 0 : let tenant_id_str = self.tenant_shard_id.tenant_id.to_string();
2815 0 : let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
2816 0 :
2817 0 : let timeline_id_str = self.timeline_id.to_string();
2818 0 :
2819 0 : self.remote_client.update_config(&new_conf.location);
2820 0 :
2821 0 : self.metrics
2822 0 : .evictions_with_low_residence_duration
2823 0 : .write()
2824 0 : .unwrap()
2825 0 : .change_threshold(
2826 0 : &tenant_id_str,
2827 0 : &shard_id_str,
2828 0 : &timeline_id_str,
2829 0 : new_threshold,
2830 0 : );
2831 0 : }
2832 0 : }
2833 :
2834 : /// Open a Timeline handle.
2835 : ///
2836 : /// Loads the metadata for the timeline into memory, but not the layer map.
2837 : #[allow(clippy::too_many_arguments)]
2838 2784 : pub(super) fn new(
2839 2784 : conf: &'static PageServerConf,
2840 2784 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
2841 2784 : metadata: &TimelineMetadata,
2842 2784 : previous_heatmap: Option<PreviousHeatmap>,
2843 2784 : ancestor: Option<Arc<Timeline>>,
2844 2784 : timeline_id: TimelineId,
2845 2784 : tenant_shard_id: TenantShardId,
2846 2784 : generation: Generation,
2847 2784 : shard_identity: ShardIdentity,
2848 2784 : walredo_mgr: Option<Arc<super::WalRedoManager>>,
2849 2784 : resources: TimelineResources,
2850 2784 : pg_version: u32,
2851 2784 : state: TimelineState,
2852 2784 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
2853 2784 : create_idempotency: crate::tenant::CreateTimelineIdempotency,
2854 2784 : gc_compaction_state: Option<GcCompactionState>,
2855 2784 : rel_size_v2_status: Option<RelSizeMigration>,
2856 2784 : cancel: CancellationToken,
2857 2784 : ) -> Arc<Self> {
2858 2784 : let disk_consistent_lsn = metadata.disk_consistent_lsn();
2859 2784 : let (state, _) = watch::channel(state);
2860 2784 :
2861 2784 : let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
2862 2784 : let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
2863 2784 :
2864 2784 : let evictions_low_residence_duration_metric_threshold = {
2865 2784 : let loaded_tenant_conf = tenant_conf.load();
2866 2784 : Self::get_evictions_low_residence_duration_metric_threshold(
2867 2784 : &loaded_tenant_conf.tenant_conf,
2868 2784 : &conf.default_tenant_conf,
2869 2784 : )
2870 : };
2871 :
2872 2784 : if let Some(ancestor) = &ancestor {
2873 1416 : let mut ancestor_gc_info = ancestor.gc_info.write().unwrap();
2874 1416 : // If we construct an explicit timeline object, it's obviously not offloaded
2875 1416 : let is_offloaded = MaybeOffloaded::No;
2876 1416 : ancestor_gc_info.insert_child(timeline_id, metadata.ancestor_lsn(), is_offloaded);
2877 1416 : }
2878 :
2879 2784 : Arc::new_cyclic(|myself| {
2880 2784 : let metrics = Arc::new(TimelineMetrics::new(
2881 2784 : &tenant_shard_id,
2882 2784 : &timeline_id,
2883 2784 : crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
2884 2784 : "mtime",
2885 2784 : evictions_low_residence_duration_metric_threshold,
2886 2784 : ),
2887 2784 : ));
2888 2784 : let aux_file_metrics = metrics.aux_file_size_gauge.clone();
2889 :
2890 2784 : let mut result = Timeline {
2891 2784 : conf,
2892 2784 : tenant_conf,
2893 2784 : myself: myself.clone(),
2894 2784 : timeline_id,
2895 2784 : tenant_shard_id,
2896 2784 : generation,
2897 2784 : shard_identity,
2898 2784 : pg_version,
2899 2784 : layers: Default::default(),
2900 2784 : gc_compaction_layer_update_lock: tokio::sync::RwLock::new(()),
2901 2784 :
2902 2784 : walredo_mgr,
2903 2784 : walreceiver: Mutex::new(None),
2904 2784 :
2905 2784 : remote_client: Arc::new(resources.remote_client),
2906 2784 :
2907 2784 : // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
2908 2784 : last_record_lsn: SeqWait::new(RecordLsn {
2909 2784 : last: disk_consistent_lsn,
2910 2784 : prev: metadata.prev_record_lsn().unwrap_or(Lsn(0)),
2911 2784 : }),
2912 2784 : disk_consistent_lsn: AtomicLsn::new(disk_consistent_lsn.0),
2913 2784 :
2914 2784 : gc_compaction_state: ArcSwap::new(Arc::new(gc_compaction_state)),
2915 2784 :
2916 2784 : last_freeze_at: AtomicLsn::new(disk_consistent_lsn.0),
2917 2784 : last_freeze_ts: RwLock::new(Instant::now()),
2918 2784 :
2919 2784 : loaded_at: (disk_consistent_lsn, SystemTime::now()),
2920 2784 :
2921 2784 : ancestor_timeline: ancestor,
2922 2784 : ancestor_lsn: metadata.ancestor_lsn(),
2923 2784 :
2924 2784 : metrics,
2925 2784 :
2926 2784 : query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new(
2927 2784 : &tenant_shard_id,
2928 2784 : &timeline_id,
2929 2784 : resources.pagestream_throttle_metrics,
2930 2784 : ),
2931 2784 :
2932 22272 : directory_metrics: array::from_fn(|_| AtomicU64::new(0)),
2933 22272 : directory_metrics_inited: array::from_fn(|_| AtomicBool::new(false)),
2934 2784 :
2935 2784 : flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
2936 2784 :
2937 2784 : layer_flush_start_tx,
2938 2784 : layer_flush_done_tx,
2939 2784 :
2940 2784 : write_lock: tokio::sync::Mutex::new(None),
2941 2784 :
2942 2784 : gc_info: std::sync::RwLock::new(GcInfo::default()),
2943 2784 :
2944 2784 : last_image_layer_creation_status: ArcSwap::new(Arc::new(
2945 2784 : LastImageLayerCreationStatus::default(),
2946 2784 : )),
2947 2784 :
2948 2784 : applied_gc_cutoff_lsn: Rcu::new(metadata.latest_gc_cutoff_lsn()),
2949 2784 : initdb_lsn: metadata.initdb_lsn(),
2950 2784 :
2951 2784 : current_logical_size: if disk_consistent_lsn.is_valid() {
2952 : // we're creating timeline data with some layer files existing locally,
2953 : // need to recalculate timeline's logical size based on data in the layers.
2954 1440 : LogicalSize::deferred_initial(disk_consistent_lsn)
2955 : } else {
2956 : // we're creating timeline data without any layers existing locally,
2957 : // initial logical size is 0.
2958 1344 : LogicalSize::empty_initial()
2959 : },
2960 :
2961 2784 : partitioning: GuardArcSwap::new((
2962 2784 : (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()),
2963 2784 : Lsn(0),
2964 2784 : )),
2965 2784 : repartition_threshold: 0,
2966 2784 : last_image_layer_creation_check_at: AtomicLsn::new(0),
2967 2784 : last_image_layer_creation_check_instant: Mutex::new(None),
2968 2784 :
2969 2784 : last_received_wal: Mutex::new(None),
2970 2784 : rel_size_cache: RwLock::new(RelSizeCache {
2971 2784 : complete_as_of: disk_consistent_lsn,
2972 2784 : map: HashMap::new(),
2973 2784 : }),
2974 2784 :
2975 2784 : download_all_remote_layers_task_info: RwLock::new(None),
2976 2784 :
2977 2784 : state,
2978 2784 :
2979 2784 : eviction_task_timeline_state: tokio::sync::Mutex::new(
2980 2784 : EvictionTaskTimelineState::default(),
2981 2784 : ),
2982 2784 : delete_progress: TimelineDeleteProgress::default(),
2983 2784 :
2984 2784 : cancel,
2985 2784 : gate: Gate::default(),
2986 2784 :
2987 2784 : compaction_lock: tokio::sync::Mutex::default(),
2988 2784 : compaction_failed: AtomicBool::default(),
2989 2784 : l0_compaction_trigger: resources.l0_compaction_trigger,
2990 2784 : gc_lock: tokio::sync::Mutex::default(),
2991 2784 :
2992 2784 : standby_horizon: AtomicLsn::new(0),
2993 2784 :
2994 2784 : pagestream_throttle: resources.pagestream_throttle,
2995 2784 :
2996 2784 : aux_file_size_estimator: AuxFileSizeEstimator::new(aux_file_metrics),
2997 2784 :
2998 2784 : #[cfg(test)]
2999 2784 : extra_test_dense_keyspace: ArcSwap::new(Arc::new(KeySpace::default())),
3000 2784 :
3001 2784 : l0_flush_global_state: resources.l0_flush_global_state,
3002 2784 :
3003 2784 : handles: Default::default(),
3004 2784 :
3005 2784 : attach_wal_lag_cooldown,
3006 2784 :
3007 2784 : create_idempotency,
3008 2784 :
3009 2784 : page_trace: Default::default(),
3010 2784 :
3011 2784 : previous_heatmap: ArcSwapOption::from_pointee(previous_heatmap),
3012 2784 :
3013 2784 : heatmap_layers_downloader: Mutex::new(None),
3014 2784 :
3015 2784 : rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
3016 2784 :
3017 2784 : wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
3018 2784 : };
3019 2784 :
3020 2784 : result.repartition_threshold =
3021 2784 : result.get_checkpoint_distance() / REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE;
3022 2784 :
3023 2784 : result
3024 2784 : .metrics
3025 2784 : .last_record_lsn_gauge
3026 2784 : .set(disk_consistent_lsn.0 as i64);
3027 2784 : result
3028 2784 : })
3029 2784 : }
3030 :
3031 4044 : pub(super) fn maybe_spawn_flush_loop(self: &Arc<Self>) {
3032 4044 : let Ok(guard) = self.gate.enter() else {
3033 0 : info!("cannot start flush loop when the timeline gate has already been closed");
3034 0 : return;
3035 : };
3036 4044 : let mut flush_loop_state = self.flush_loop_state.lock().unwrap();
3037 4044 : match *flush_loop_state {
3038 2748 : FlushLoopState::NotStarted => (),
3039 : FlushLoopState::Running { .. } => {
3040 1296 : info!(
3041 0 : "skipping attempt to start flush_loop twice {}/{}",
3042 0 : self.tenant_shard_id, self.timeline_id
3043 : );
3044 1296 : return;
3045 : }
3046 : FlushLoopState::Exited => {
3047 0 : info!(
3048 0 : "ignoring attempt to restart exited flush_loop {}/{}",
3049 0 : self.tenant_shard_id, self.timeline_id
3050 : );
3051 0 : return;
3052 : }
3053 : }
3054 :
3055 2748 : let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
3056 2748 : let self_clone = Arc::clone(self);
3057 2748 :
3058 2748 : debug!("spawning flush loop");
3059 2748 : *flush_loop_state = FlushLoopState::Running {
3060 2748 : #[cfg(test)]
3061 2748 : expect_initdb_optimization: false,
3062 2748 : #[cfg(test)]
3063 2748 : initdb_optimization_count: 0,
3064 2748 : };
3065 2748 : task_mgr::spawn(
3066 2748 : task_mgr::BACKGROUND_RUNTIME.handle(),
3067 2748 : task_mgr::TaskKind::LayerFlushTask,
3068 2748 : self.tenant_shard_id,
3069 2748 : Some(self.timeline_id),
3070 2748 : "layer flush task",
3071 2748 : async move {
3072 2748 : let _guard = guard;
3073 2748 : let background_ctx = RequestContext::todo_child(TaskKind::LayerFlushTask, DownloadBehavior::Error).with_scope_timeline(&self_clone);
3074 2748 : self_clone.flush_loop(layer_flush_start_rx, &background_ctx).await;
3075 60 : let mut flush_loop_state = self_clone.flush_loop_state.lock().unwrap();
3076 60 : assert!(matches!(*flush_loop_state, FlushLoopState::Running{..}));
3077 60 : *flush_loop_state = FlushLoopState::Exited;
3078 60 : Ok(())
3079 60 : }
3080 2748 : .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))
3081 : );
3082 4044 : }
3083 :
3084 0 : pub(crate) fn update_gc_compaction_state(
3085 0 : &self,
3086 0 : gc_compaction_state: GcCompactionState,
3087 0 : ) -> anyhow::Result<()> {
3088 0 : self.gc_compaction_state
3089 0 : .store(Arc::new(Some(gc_compaction_state.clone())));
3090 0 : self.remote_client
3091 0 : .schedule_index_upload_for_gc_compaction_state_update(gc_compaction_state)
3092 0 : }
3093 :
3094 0 : pub(crate) fn update_rel_size_v2_status(
3095 0 : &self,
3096 0 : rel_size_v2_status: RelSizeMigration,
3097 0 : ) -> anyhow::Result<()> {
3098 0 : self.rel_size_v2_status
3099 0 : .store(Some(Arc::new(rel_size_v2_status.clone())));
3100 0 : self.remote_client
3101 0 : .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
3102 0 : }
3103 :
3104 0 : pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
3105 0 : self.gc_compaction_state.load_full().as_ref().clone()
3106 0 : }
3107 :
3108 : /// Creates and starts the wal receiver.
3109 : ///
3110 : /// This function is expected to be called at most once per Timeline's lifecycle
3111 : /// when the timeline is activated.
3112 0 : fn launch_wal_receiver(
3113 0 : self: &Arc<Self>,
3114 0 : ctx: &RequestContext,
3115 0 : broker_client: BrokerClientChannel,
3116 0 : ) {
3117 0 : info!(
3118 0 : "launching WAL receiver for timeline {} of tenant {}",
3119 0 : self.timeline_id, self.tenant_shard_id
3120 : );
3121 :
3122 0 : let tenant_conf = self.tenant_conf.load();
3123 0 : let wal_connect_timeout = tenant_conf
3124 0 : .tenant_conf
3125 0 : .walreceiver_connect_timeout
3126 0 : .unwrap_or(self.conf.default_tenant_conf.walreceiver_connect_timeout);
3127 0 : let lagging_wal_timeout = tenant_conf
3128 0 : .tenant_conf
3129 0 : .lagging_wal_timeout
3130 0 : .unwrap_or(self.conf.default_tenant_conf.lagging_wal_timeout);
3131 0 : let max_lsn_wal_lag = tenant_conf
3132 0 : .tenant_conf
3133 0 : .max_lsn_wal_lag
3134 0 : .unwrap_or(self.conf.default_tenant_conf.max_lsn_wal_lag);
3135 0 :
3136 0 : let mut guard = self.walreceiver.lock().unwrap();
3137 0 : assert!(
3138 0 : guard.is_none(),
3139 0 : "multiple launches / re-launches of WAL receiver are not supported"
3140 : );
3141 0 : *guard = Some(WalReceiver::start(
3142 0 : Arc::clone(self),
3143 0 : WalReceiverConf {
3144 0 : protocol: self.resolve_wal_receiver_protocol(),
3145 0 : wal_connect_timeout,
3146 0 : lagging_wal_timeout,
3147 0 : max_lsn_wal_lag,
3148 0 : auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
3149 0 : availability_zone: self.conf.availability_zone.clone(),
3150 0 : ingest_batch_size: self.conf.ingest_batch_size,
3151 0 : validate_wal_contiguity: self.conf.validate_wal_contiguity,
3152 0 : },
3153 0 : broker_client,
3154 0 : ctx,
3155 0 : ));
3156 0 : }
3157 :
3158 : /// Initialize with an empty layer map. Used when creating a new timeline.
3159 2748 : pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
3160 2748 : let mut layers = self.layers.try_write().expect(
3161 2748 : "in the context where we call this function, no other task has access to the object",
3162 2748 : );
3163 2748 : layers
3164 2748 : .open_mut()
3165 2748 : .expect("in this context the LayerManager must still be open")
3166 2748 : .initialize_empty(Lsn(start_lsn.0));
3167 2748 : }
3168 :
3169 : /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
3170 : /// files.
3171 36 : pub(super) async fn load_layer_map(
3172 36 : &self,
3173 36 : disk_consistent_lsn: Lsn,
3174 36 : index_part: IndexPart,
3175 36 : ) -> anyhow::Result<()> {
3176 : use LayerName::*;
3177 : use init::Decision::*;
3178 : use init::{Discovered, DismissedLayer};
3179 :
3180 36 : let mut guard = self.layers.write().await;
3181 :
3182 36 : let timer = self.metrics.load_layer_map_histo.start_timer();
3183 36 :
3184 36 : // Scan timeline directory and create ImageLayerName and DeltaFilename
3185 36 : // structs representing all files on disk
3186 36 : let timeline_path = self
3187 36 : .conf
3188 36 : .timeline_path(&self.tenant_shard_id, &self.timeline_id);
3189 36 : let conf = self.conf;
3190 36 : let span = tracing::Span::current();
3191 36 :
3192 36 : // Copy to move into the task we're about to spawn
3193 36 : let this = self.myself.upgrade().expect("&self method holds the arc");
3194 :
3195 36 : let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
3196 36 : move || {
3197 36 : let _g = span.entered();
3198 36 : let discovered = init::scan_timeline_dir(&timeline_path)?;
3199 36 : let mut discovered_layers = Vec::with_capacity(discovered.len());
3200 36 : let mut unrecognized_files = Vec::new();
3201 36 :
3202 36 : let mut path = timeline_path;
3203 :
3204 132 : for discovered in discovered {
3205 96 : let (name, kind) = match discovered {
3206 96 : Discovered::Layer(layer_file_name, local_metadata) => {
3207 96 : discovered_layers.push((layer_file_name, local_metadata));
3208 96 : continue;
3209 : }
3210 0 : Discovered::IgnoredBackup(path) => {
3211 0 : std::fs::remove_file(path)
3212 0 : .or_else(fs_ext::ignore_not_found)
3213 0 : .fatal_err("Removing .old file");
3214 0 : continue;
3215 : }
3216 0 : Discovered::Unknown(file_name) => {
3217 0 : // we will later error if there are any
3218 0 : unrecognized_files.push(file_name);
3219 0 : continue;
3220 : }
3221 0 : Discovered::Ephemeral(name) => (name, "old ephemeral file"),
3222 0 : Discovered::Temporary(name) => (name, "temporary timeline file"),
3223 0 : Discovered::TemporaryDownload(name) => (name, "temporary download"),
3224 : };
3225 0 : path.push(Utf8Path::new(&name));
3226 0 : init::cleanup(&path, kind)?;
3227 0 : path.pop();
3228 : }
3229 :
3230 36 : if !unrecognized_files.is_empty() {
3231 : // assume that if there are any there are many many.
3232 0 : let n = unrecognized_files.len();
3233 0 : let first = &unrecognized_files[..n.min(10)];
3234 0 : anyhow::bail!(
3235 0 : "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
3236 0 : );
3237 36 : }
3238 36 :
3239 36 : let decided = init::reconcile(discovered_layers, &index_part, disk_consistent_lsn);
3240 36 :
3241 36 : let mut loaded_layers = Vec::new();
3242 36 : let mut needs_cleanup = Vec::new();
3243 36 : let mut total_physical_size = 0;
3244 :
3245 132 : for (name, decision) in decided {
3246 96 : let decision = match decision {
3247 96 : Ok(decision) => decision,
3248 0 : Err(DismissedLayer::Future { local }) => {
3249 0 : if let Some(local) = local {
3250 0 : init::cleanup_future_layer(
3251 0 : &local.local_path,
3252 0 : &name,
3253 0 : disk_consistent_lsn,
3254 0 : )?;
3255 0 : }
3256 0 : needs_cleanup.push(name);
3257 0 : continue;
3258 : }
3259 0 : Err(DismissedLayer::LocalOnly(local)) => {
3260 0 : init::cleanup_local_only_file(&name, &local)?;
3261 : // this file never existed remotely, we will have to do rework
3262 0 : continue;
3263 : }
3264 0 : Err(DismissedLayer::BadMetadata(local)) => {
3265 0 : init::cleanup_local_file_for_remote(&local)?;
3266 : // this file never existed remotely, we will have to do rework
3267 0 : continue;
3268 : }
3269 : };
3270 :
3271 96 : match &name {
3272 72 : Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
3273 24 : Image(i) => assert!(i.lsn <= disk_consistent_lsn),
3274 : }
3275 :
3276 96 : tracing::debug!(layer=%name, ?decision, "applied");
3277 :
3278 96 : let layer = match decision {
3279 96 : Resident { local, remote } => {
3280 96 : total_physical_size += local.file_size;
3281 96 : Layer::for_resident(conf, &this, local.local_path, name, remote)
3282 96 : .drop_eviction_guard()
3283 : }
3284 0 : Evicted(remote) => Layer::for_evicted(conf, &this, name, remote),
3285 : };
3286 :
3287 96 : loaded_layers.push(layer);
3288 : }
3289 36 : Ok((loaded_layers, needs_cleanup, total_physical_size))
3290 36 : }
3291 36 : })
3292 36 : .await
3293 36 : .map_err(anyhow::Error::new)
3294 36 : .and_then(|x| x)?;
3295 :
3296 36 : let num_layers = loaded_layers.len();
3297 36 :
3298 36 : guard
3299 36 : .open_mut()
3300 36 : .expect("layermanager must be open during init")
3301 36 : .initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
3302 36 :
3303 36 : self.remote_client
3304 36 : .schedule_layer_file_deletion(&needs_cleanup)?;
3305 36 : self.remote_client
3306 36 : .schedule_index_upload_for_file_changes()?;
3307 : // This barrier orders above DELETEs before any later operations.
3308 : // This is critical because code executing after the barrier might
3309 : // create again objects with the same key that we just scheduled for deletion.
3310 : // For example, if we just scheduled deletion of an image layer "from the future",
3311 : // later compaction might run again and re-create the same image layer.
3312 : // "from the future" here means an image layer whose LSN is > IndexPart::disk_consistent_lsn.
3313 : // "same" here means same key range and LSN.
3314 : //
3315 : // Without a barrier between above DELETEs and the re-creation's PUTs,
3316 : // the upload queue may execute the PUT first, then the DELETE.
3317 : // In our example, we will end up with an IndexPart referencing a non-existent object.
3318 : //
3319 : // 1. a future image layer is created and uploaded
3320 : // 2. ps restart
3321 : // 3. the future layer from (1) is deleted during load layer map
3322 : // 4. image layer is re-created and uploaded
3323 : // 5. deletion queue would like to delete (1) but actually deletes (4)
3324 : // 6. delete by name works as expected, but it now deletes the wrong (later) version
3325 : //
3326 : // See https://github.com/neondatabase/neon/issues/5878
3327 : //
3328 : // NB: generation numbers naturally protect against this because they disambiguate
3329 : // (1) and (4)
3330 : // TODO: this is basically a no-op now, should we remove it?
3331 36 : self.remote_client.schedule_barrier()?;
3332 : // TenantShard::create_timeline will wait for these uploads to happen before returning, or
3333 : // on retry.
3334 :
3335 : // Now that we have the full layer map, we may calculate the visibility of layers within it (a global scan)
3336 36 : drop(guard); // drop write lock, update_layer_visibility will take a read lock.
3337 36 : self.update_layer_visibility().await?;
3338 :
3339 36 : info!(
3340 0 : "loaded layer map with {} layers at {}, total physical size: {}",
3341 : num_layers, disk_consistent_lsn, total_physical_size
3342 : );
3343 :
3344 36 : timer.stop_and_record();
3345 36 : Ok(())
3346 36 : }
3347 :
3348 : /// Retrieve current logical size of the timeline.
3349 : ///
3350 : /// The size could be lagging behind the actual number, in case
3351 : /// the initial size calculation has not been run (gets triggered on the first size access).
3352 : ///
3353 : /// return size and boolean flag that shows if the size is exact
3354 0 : pub(crate) fn get_current_logical_size(
3355 0 : self: &Arc<Self>,
3356 0 : priority: GetLogicalSizePriority,
3357 0 : ctx: &RequestContext,
3358 0 : ) -> logical_size::CurrentLogicalSize {
3359 0 : if !self.tenant_shard_id.is_shard_zero() {
3360 : // Logical size is only accurately maintained on shard zero: when called elsewhere, for example
3361 : // when HTTP API is serving a GET for timeline zero, return zero
3362 0 : return logical_size::CurrentLogicalSize::Approximate(logical_size::Approximate::zero());
3363 0 : }
3364 0 :
3365 0 : let current_size = self.current_logical_size.current_size();
3366 0 : debug!("Current size: {current_size:?}");
3367 :
3368 0 : match (current_size.accuracy(), priority) {
3369 0 : (logical_size::Accuracy::Exact, _) => (), // nothing to do
3370 0 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::Background) => {
3371 0 : // background task will eventually deliver an exact value, we're in no rush
3372 0 : }
3373 : (logical_size::Accuracy::Approximate, GetLogicalSizePriority::User) => {
3374 : // background task is not ready, but user is asking for it now;
3375 : // => make the background task skip the line
3376 : // (The alternative would be to calculate the size here, but,
3377 : // it can actually take a long time if the user has a lot of rels.
3378 : // And we'll inevitable need it again; So, let the background task do the work.)
3379 0 : match self
3380 0 : .current_logical_size
3381 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
3382 0 : .get()
3383 : {
3384 0 : Some(cancel) => cancel.cancel(),
3385 : None => {
3386 0 : match self.current_state() {
3387 0 : TimelineState::Broken { .. } | TimelineState::Stopping => {
3388 0 : // Can happen when timeline detail endpoint is used when deletion is ongoing (or its broken).
3389 0 : // Don't make noise.
3390 0 : }
3391 : TimelineState::Loading => {
3392 : // Import does not return an activated timeline.
3393 0 : info!(
3394 0 : "discarding priority boost for logical size calculation because timeline is not yet active"
3395 : );
3396 : }
3397 : TimelineState::Active => {
3398 : // activation should be setting the once cell
3399 0 : warn!(
3400 0 : "unexpected: cancel_wait_for_background_loop_concurrency_limit_semaphore not set, priority-boosting of logical size calculation will not work"
3401 : );
3402 0 : debug_assert!(false);
3403 : }
3404 : }
3405 : }
3406 : }
3407 : }
3408 : }
3409 :
3410 0 : if let CurrentLogicalSize::Approximate(_) = ¤t_size {
3411 0 : if ctx.task_kind() == TaskKind::WalReceiverConnectionHandler {
3412 0 : let first = self
3413 0 : .current_logical_size
3414 0 : .did_return_approximate_to_walreceiver
3415 0 : .compare_exchange(
3416 0 : false,
3417 0 : true,
3418 0 : AtomicOrdering::Relaxed,
3419 0 : AtomicOrdering::Relaxed,
3420 0 : )
3421 0 : .is_ok();
3422 0 : if first {
3423 0 : crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE.inc();
3424 0 : }
3425 0 : }
3426 0 : }
3427 :
3428 0 : current_size
3429 0 : }
3430 :
3431 0 : fn spawn_initial_logical_size_computation_task(self: &Arc<Self>, ctx: &RequestContext) {
3432 0 : let Some(initial_part_end) = self.current_logical_size.initial_part_end else {
3433 : // nothing to do for freshly created timelines;
3434 0 : assert_eq!(
3435 0 : self.current_logical_size.current_size().accuracy(),
3436 0 : logical_size::Accuracy::Exact,
3437 0 : );
3438 0 : self.current_logical_size.initialized.add_permits(1);
3439 0 : return;
3440 : };
3441 :
3442 0 : let cancel_wait_for_background_loop_concurrency_limit_semaphore = CancellationToken::new();
3443 0 : let token = cancel_wait_for_background_loop_concurrency_limit_semaphore.clone();
3444 0 : self.current_logical_size
3445 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore.set(token)
3446 0 : .expect("initial logical size calculation task must be spawned exactly once per Timeline object");
3447 0 :
3448 0 : let self_clone = Arc::clone(self);
3449 0 : let background_ctx = ctx.detached_child(
3450 0 : TaskKind::InitialLogicalSizeCalculation,
3451 0 : DownloadBehavior::Download,
3452 0 : );
3453 0 : task_mgr::spawn(
3454 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
3455 0 : task_mgr::TaskKind::InitialLogicalSizeCalculation,
3456 0 : self.tenant_shard_id,
3457 0 : Some(self.timeline_id),
3458 0 : "initial size calculation",
3459 : // NB: don't log errors here, task_mgr will do that.
3460 0 : async move {
3461 0 : self_clone
3462 0 : .initial_logical_size_calculation_task(
3463 0 : initial_part_end,
3464 0 : cancel_wait_for_background_loop_concurrency_limit_semaphore,
3465 0 : background_ctx,
3466 0 : )
3467 0 : .await;
3468 0 : Ok(())
3469 0 : }
3470 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)),
3471 : );
3472 0 : }
3473 :
3474 : /// # Cancellation
3475 : ///
3476 : /// This method is sensitive to `Timeline::cancel`.
3477 : ///
3478 : /// It is _not_ sensitive to task_mgr::shutdown_token().
3479 : ///
3480 : /// # Cancel-Safety
3481 : ///
3482 : /// It does Timeline IO, hence this should be polled to completion because
3483 : /// we could be leaving in-flight IOs behind, which is safe, but annoying
3484 : /// to reason about.
3485 0 : async fn initial_logical_size_calculation_task(
3486 0 : self: Arc<Self>,
3487 0 : initial_part_end: Lsn,
3488 0 : skip_concurrency_limiter: CancellationToken,
3489 0 : background_ctx: RequestContext,
3490 0 : ) {
3491 0 : scopeguard::defer! {
3492 0 : // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
3493 0 : self.current_logical_size.initialized.add_permits(1);
3494 0 : }
3495 0 :
3496 0 : let try_once = |attempt: usize| {
3497 0 : let background_ctx = &background_ctx;
3498 0 : let self_ref = &self;
3499 0 : let skip_concurrency_limiter = &skip_concurrency_limiter;
3500 0 : async move {
3501 0 : let wait_for_permit = super::tasks::acquire_concurrency_permit(
3502 0 : BackgroundLoopKind::InitialLogicalSizeCalculation,
3503 0 : background_ctx,
3504 0 : );
3505 :
3506 : use crate::metrics::initial_logical_size::StartCircumstances;
3507 0 : let (_maybe_permit, circumstances) = tokio::select! {
3508 0 : permit = wait_for_permit => {
3509 0 : (Some(permit), StartCircumstances::AfterBackgroundTasksRateLimit)
3510 : }
3511 0 : _ = self_ref.cancel.cancelled() => {
3512 0 : return Err(CalculateLogicalSizeError::Cancelled);
3513 : }
3514 0 : () = skip_concurrency_limiter.cancelled() => {
3515 : // Some action that is part of a end user interaction requested logical size
3516 : // => break out of the rate limit
3517 : // TODO: ideally we'd not run on BackgroundRuntime but the requester's runtime;
3518 : // but then again what happens if they cancel; also, we should just be using
3519 : // one runtime across the entire process, so, let's leave this for now.
3520 0 : (None, StartCircumstances::SkippedConcurrencyLimiter)
3521 : }
3522 : };
3523 :
3524 0 : let metrics_guard = if attempt == 1 {
3525 0 : crate::metrics::initial_logical_size::START_CALCULATION.first(circumstances)
3526 : } else {
3527 0 : crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
3528 : };
3529 :
3530 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
3531 0 : self_ref.conf,
3532 0 : self_ref
3533 0 : .gate
3534 0 : .enter()
3535 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
3536 : );
3537 :
3538 0 : let calculated_size = self_ref
3539 0 : .logical_size_calculation_task(
3540 0 : initial_part_end,
3541 0 : LogicalSizeCalculationCause::Initial,
3542 0 : background_ctx,
3543 0 : )
3544 0 : .await?;
3545 :
3546 0 : self_ref
3547 0 : .trigger_aux_file_size_computation(
3548 0 : initial_part_end,
3549 0 : background_ctx,
3550 0 : io_concurrency,
3551 0 : )
3552 0 : .await?;
3553 :
3554 : // TODO: add aux file size to logical size
3555 :
3556 0 : Ok((calculated_size, metrics_guard))
3557 0 : }
3558 0 : };
3559 :
3560 0 : let retrying = async {
3561 0 : let mut attempt = 0;
3562 : loop {
3563 0 : attempt += 1;
3564 0 :
3565 0 : match try_once(attempt).await {
3566 0 : Ok(res) => return ControlFlow::Continue(res),
3567 0 : Err(CalculateLogicalSizeError::Cancelled) => return ControlFlow::Break(()),
3568 : Err(
3569 0 : e @ (CalculateLogicalSizeError::Decode(_)
3570 0 : | CalculateLogicalSizeError::PageRead(_)),
3571 0 : ) => {
3572 0 : warn!(attempt, "initial size calculation failed: {e:?}");
3573 : // exponential back-off doesn't make sense at these long intervals;
3574 : // use fixed retry interval with generous jitter instead
3575 0 : let sleep_duration = Duration::from_secs(
3576 0 : u64::try_from(
3577 0 : // 1hour base
3578 0 : (60_i64 * 60_i64)
3579 0 : // 10min jitter
3580 0 : + rand::thread_rng().gen_range(-10 * 60..10 * 60),
3581 0 : )
3582 0 : .expect("10min < 1hour"),
3583 0 : );
3584 0 : tokio::select! {
3585 0 : _ = tokio::time::sleep(sleep_duration) => {}
3586 0 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
3587 : }
3588 : }
3589 : }
3590 : }
3591 0 : };
3592 :
3593 0 : let (calculated_size, metrics_guard) = match retrying.await {
3594 0 : ControlFlow::Continue(calculated_size) => calculated_size,
3595 0 : ControlFlow::Break(()) => return,
3596 : };
3597 :
3598 : // we cannot query current_logical_size.current_size() to know the current
3599 : // *negative* value, only truncated to u64.
3600 0 : let added = self
3601 0 : .current_logical_size
3602 0 : .size_added_after_initial
3603 0 : .load(AtomicOrdering::Relaxed);
3604 0 :
3605 0 : let sum = calculated_size.saturating_add_signed(added);
3606 0 :
3607 0 : // set the gauge value before it can be set in `update_current_logical_size`.
3608 0 : self.metrics.current_logical_size_gauge.set(sum);
3609 0 :
3610 0 : self.current_logical_size
3611 0 : .initial_logical_size
3612 0 : .set((calculated_size, metrics_guard.calculation_result_saved()))
3613 0 : .ok()
3614 0 : .expect("only this task sets it");
3615 0 : }
3616 :
3617 84 : pub(crate) fn spawn_ondemand_logical_size_calculation(
3618 84 : self: &Arc<Self>,
3619 84 : lsn: Lsn,
3620 84 : cause: LogicalSizeCalculationCause,
3621 84 : ctx: RequestContext,
3622 84 : ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
3623 84 : let (sender, receiver) = oneshot::channel();
3624 84 : let self_clone = Arc::clone(self);
3625 84 : // XXX if our caller loses interest, i.e., ctx is cancelled,
3626 84 : // we should stop the size calculation work and return an error.
3627 84 : // That would require restructuring this function's API to
3628 84 : // return the result directly, instead of a Receiver for the result.
3629 84 : let ctx = ctx.detached_child(
3630 84 : TaskKind::OndemandLogicalSizeCalculation,
3631 84 : DownloadBehavior::Download,
3632 84 : );
3633 84 : task_mgr::spawn(
3634 84 : task_mgr::BACKGROUND_RUNTIME.handle(),
3635 84 : task_mgr::TaskKind::OndemandLogicalSizeCalculation,
3636 84 : self.tenant_shard_id,
3637 84 : Some(self.timeline_id),
3638 84 : "ondemand logical size calculation",
3639 84 : async move {
3640 84 : let res = self_clone
3641 84 : .logical_size_calculation_task(lsn, cause, &ctx)
3642 84 : .await;
3643 84 : let _ = sender.send(res).ok();
3644 84 : Ok(()) // Receiver is responsible for handling errors
3645 84 : }
3646 84 : .in_current_span(),
3647 84 : );
3648 84 : receiver
3649 84 : }
3650 :
3651 : #[instrument(skip_all)]
3652 : async fn logical_size_calculation_task(
3653 : self: &Arc<Self>,
3654 : lsn: Lsn,
3655 : cause: LogicalSizeCalculationCause,
3656 : ctx: &RequestContext,
3657 : ) -> Result<u64, CalculateLogicalSizeError> {
3658 : crate::span::debug_assert_current_span_has_tenant_and_timeline_id();
3659 : // We should never be calculating logical sizes on shard !=0, because these shards do not have
3660 : // accurate relation sizes, and they do not emit consumption metrics.
3661 : debug_assert!(self.tenant_shard_id.is_shard_zero());
3662 :
3663 : let guard = self
3664 : .gate
3665 : .enter()
3666 0 : .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
3667 :
3668 : self.calculate_logical_size(lsn, cause, &guard, ctx).await
3669 : }
3670 :
3671 : /// Calculate the logical size of the database at the latest LSN.
3672 : ///
3673 : /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
3674 : /// especially if we need to download remote layers.
3675 84 : async fn calculate_logical_size(
3676 84 : &self,
3677 84 : up_to_lsn: Lsn,
3678 84 : cause: LogicalSizeCalculationCause,
3679 84 : _guard: &GateGuard,
3680 84 : ctx: &RequestContext,
3681 84 : ) -> Result<u64, CalculateLogicalSizeError> {
3682 84 : info!(
3683 0 : "Calculating logical size for timeline {} at {}",
3684 : self.timeline_id, up_to_lsn
3685 : );
3686 :
3687 84 : if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
3688 : {
3689 0 : return Err(CalculateLogicalSizeError::Cancelled);
3690 84 : }
3691 :
3692 : // See if we've already done the work for initial size calculation.
3693 : // This is a short-cut for timelines that are mostly unused.
3694 84 : if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
3695 0 : return Ok(size);
3696 84 : }
3697 84 : let storage_time_metrics = match cause {
3698 : LogicalSizeCalculationCause::Initial
3699 : | LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize
3700 0 : | LogicalSizeCalculationCause::TenantSizeHandler => &self.metrics.logical_size_histo,
3701 : LogicalSizeCalculationCause::EvictionTaskImitation => {
3702 84 : &self.metrics.imitate_logical_size_histo
3703 : }
3704 : };
3705 84 : let timer = storage_time_metrics.start_timer();
3706 84 : let logical_size = self
3707 84 : .get_current_logical_size_non_incremental(up_to_lsn, ctx)
3708 84 : .await?;
3709 84 : debug!("calculated logical size: {logical_size}");
3710 84 : timer.stop_and_record();
3711 84 : Ok(logical_size)
3712 84 : }
3713 :
3714 : /// Update current logical size, adding `delta' to the old value.
3715 1623420 : fn update_current_logical_size(&self, delta: i64) {
3716 1623420 : let logical_size = &self.current_logical_size;
3717 1623420 : logical_size.increment_size(delta);
3718 1623420 :
3719 1623420 : // Also set the value in the prometheus gauge. Note that
3720 1623420 : // there is a race condition here: if this is is called by two
3721 1623420 : // threads concurrently, the prometheus gauge might be set to
3722 1623420 : // one value while current_logical_size is set to the
3723 1623420 : // other.
3724 1623420 : match logical_size.current_size() {
3725 1623420 : CurrentLogicalSize::Exact(ref new_current_size) => self
3726 1623420 : .metrics
3727 1623420 : .current_logical_size_gauge
3728 1623420 : .set(new_current_size.into()),
3729 0 : CurrentLogicalSize::Approximate(_) => {
3730 0 : // don't update the gauge yet, this allows us not to update the gauge back and
3731 0 : // forth between the initial size calculation task.
3732 0 : }
3733 : }
3734 1623420 : }
3735 :
3736 18144 : pub(crate) fn update_directory_entries_count(&self, kind: DirectoryKind, count: MetricsUpdate) {
3737 18144 : // TODO: this directory metrics is not correct -- we could have multiple reldirs in the system
3738 18144 : // for each of the database, but we only store one value, and therefore each pgdirmodification
3739 18144 : // would overwrite the previous value if they modify different databases.
3740 18144 :
3741 18144 : match count {
3742 6612 : MetricsUpdate::Set(count) => {
3743 6612 : self.directory_metrics[kind.offset()].store(count, AtomicOrdering::Relaxed);
3744 6612 : self.directory_metrics_inited[kind.offset()].store(true, AtomicOrdering::Relaxed);
3745 6612 : }
3746 11520 : MetricsUpdate::Add(count) => {
3747 11520 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3748 11520 : // it's fine.
3749 11520 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3750 11520 : // The metrics has been initialized with `MetricsUpdate::Set` before, so we can add/sub
3751 11520 : // the value reliably.
3752 11520 : self.directory_metrics[kind.offset()].fetch_add(count, AtomicOrdering::Relaxed);
3753 11520 : }
3754 : // Otherwise, ignore this update
3755 : }
3756 12 : MetricsUpdate::Sub(count) => {
3757 12 : // TODO: these operations are not atomic; but we only have one writer to the metrics, so
3758 12 : // it's fine.
3759 12 : if self.directory_metrics_inited[kind.offset()].load(AtomicOrdering::Relaxed) {
3760 12 : // The metrics has been initialized with `MetricsUpdate::Set` before.
3761 12 : // The operation could overflow so we need to normalize the value.
3762 12 : let prev_val =
3763 12 : self.directory_metrics[kind.offset()].load(AtomicOrdering::Relaxed);
3764 12 : let res = prev_val.saturating_sub(count);
3765 12 : self.directory_metrics[kind.offset()].store(res, AtomicOrdering::Relaxed);
3766 12 : }
3767 : // Otherwise, ignore this update
3768 : }
3769 : };
3770 :
3771 : // TODO: remove this, there's no place in the code that updates this aux metrics.
3772 18144 : let aux_metric =
3773 18144 : self.directory_metrics[DirectoryKind::AuxFiles.offset()].load(AtomicOrdering::Relaxed);
3774 18144 :
3775 18144 : let sum_of_entries = self
3776 18144 : .directory_metrics
3777 18144 : .iter()
3778 145152 : .map(|v| v.load(AtomicOrdering::Relaxed))
3779 18144 : .sum();
3780 : // Set a high general threshold and a lower threshold for the auxiliary files,
3781 : // as we can have large numbers of relations in the db directory.
3782 : const SUM_THRESHOLD: u64 = 5000;
3783 : const AUX_THRESHOLD: u64 = 1000;
3784 18144 : if sum_of_entries >= SUM_THRESHOLD || aux_metric >= AUX_THRESHOLD {
3785 0 : self.metrics
3786 0 : .directory_entries_count_gauge
3787 0 : .set(sum_of_entries);
3788 18144 : } else if let Some(metric) = Lazy::get(&self.metrics.directory_entries_count_gauge) {
3789 0 : metric.set(sum_of_entries);
3790 18144 : }
3791 18144 : }
3792 :
3793 0 : async fn find_layer(
3794 0 : &self,
3795 0 : layer_name: &LayerName,
3796 0 : ) -> Result<Option<Layer>, layer_manager::Shutdown> {
3797 0 : let guard = self.layers.read().await;
3798 0 : let layer = guard
3799 0 : .layer_map()?
3800 0 : .iter_historic_layers()
3801 0 : .find(|l| &l.layer_name() == layer_name)
3802 0 : .map(|found| guard.get_from_desc(&found));
3803 0 : Ok(layer)
3804 0 : }
3805 :
3806 0 : pub(super) fn should_keep_previous_heatmap(&self, new_heatmap_end_lsn: Lsn) -> bool {
3807 0 : let crnt = self.previous_heatmap.load();
3808 0 : match crnt.as_deref() {
3809 0 : Some(PreviousHeatmap::Active { end_lsn, .. }) => match end_lsn {
3810 0 : Some(crnt_end_lsn) => *crnt_end_lsn > new_heatmap_end_lsn,
3811 0 : None => true,
3812 : },
3813 0 : Some(PreviousHeatmap::Obsolete) => false,
3814 0 : None => false,
3815 : }
3816 0 : }
3817 :
3818 : /// The timeline heatmap is a hint to secondary locations from the primary location,
3819 : /// indicating which layers are currently on-disk on the primary.
3820 : ///
3821 : /// None is returned if the Timeline is in a state where uploading a heatmap
3822 : /// doesn't make sense, such as shutting down or initializing. The caller
3823 : /// should treat this as a cue to simply skip doing any heatmap uploading
3824 : /// for this timeline.
3825 96 : pub(crate) async fn generate_heatmap(&self) -> Option<HeatMapTimeline> {
3826 96 : if !self.is_active() {
3827 0 : return None;
3828 96 : }
3829 :
3830 96 : let guard = self.layers.read().await;
3831 :
3832 : // Firstly, if there's any heatmap left over from when this location
3833 : // was a secondary, take that into account. Keep layers that are:
3834 : // * present in the layer map
3835 : // * visible
3836 : // * non-resident
3837 : // * not evicted since we read the heatmap
3838 : //
3839 : // Without this, a new cold, attached location would clobber the previous
3840 : // heatamp.
3841 96 : let previous_heatmap = self.previous_heatmap.load();
3842 96 : let visible_non_resident = match previous_heatmap.as_deref() {
3843 : Some(PreviousHeatmap::Active {
3844 72 : heatmap, read_at, ..
3845 276 : }) => Some(heatmap.all_layers().filter_map(|hl| {
3846 276 : let desc: PersistentLayerDesc = hl.name.clone().into();
3847 276 : let layer = guard.try_get_from_key(&desc.key())?;
3848 :
3849 276 : if layer.visibility() == LayerVisibilityHint::Covered {
3850 0 : return None;
3851 276 : }
3852 276 :
3853 276 : if layer.is_likely_resident() {
3854 120 : return None;
3855 156 : }
3856 156 :
3857 156 : if layer.last_evicted_at().happened_after(*read_at) {
3858 36 : return None;
3859 120 : }
3860 120 :
3861 120 : Some((desc, hl.metadata.clone(), hl.access_time, hl.cold))
3862 276 : })),
3863 0 : Some(PreviousHeatmap::Obsolete) => None,
3864 24 : None => None,
3865 : };
3866 :
3867 : // Secondly, all currently visible, resident layers are included.
3868 216 : let resident = guard.likely_resident_layers().filter_map(|layer| {
3869 216 : match layer.visibility() {
3870 : LayerVisibilityHint::Visible => {
3871 : // Layer is visible to one or more read LSNs: elegible for inclusion in layer map
3872 204 : let last_activity_ts = layer.latest_activity();
3873 204 : Some((
3874 204 : layer.layer_desc().clone(),
3875 204 : layer.metadata(),
3876 204 : last_activity_ts,
3877 204 : false, // these layers are not cold
3878 204 : ))
3879 : }
3880 : LayerVisibilityHint::Covered => {
3881 : // Layer is resident but unlikely to be read: not elegible for inclusion in heatmap.
3882 12 : None
3883 : }
3884 : }
3885 216 : });
3886 :
3887 96 : let mut layers = match visible_non_resident {
3888 72 : Some(non_resident) => {
3889 72 : let mut non_resident = non_resident.peekable();
3890 72 : if non_resident.peek().is_none() {
3891 24 : tracing::info!(timeline_id=%self.timeline_id, "Previous heatmap now obsolete");
3892 24 : self.previous_heatmap
3893 24 : .store(Some(PreviousHeatmap::Obsolete.into()));
3894 48 : }
3895 :
3896 72 : non_resident.chain(resident).collect::<Vec<_>>()
3897 : }
3898 24 : None => resident.collect::<Vec<_>>(),
3899 : };
3900 :
3901 : // Sort layers in order of which to download first. For a large set of layers to download, we
3902 : // want to prioritize those layers which are most likely to still be in the resident many minutes
3903 : // or hours later:
3904 : // - Cold layers go last for convenience when a human inspects the heatmap.
3905 : // - Download L0s last, because they churn the fastest: L0s on a fast-writing tenant might
3906 : // only exist for a few minutes before being compacted into L1s.
3907 : // - For L1 & image layers, download most recent LSNs first: the older the LSN, the sooner
3908 : // the layer is likely to be covered by an image layer during compaction.
3909 718 : layers.sort_by_key(|(desc, _meta, _atime, cold)| {
3910 718 : std::cmp::Reverse((
3911 718 : *cold,
3912 718 : !LayerMap::is_l0(&desc.key_range, desc.is_delta),
3913 718 : desc.lsn_range.end,
3914 718 : ))
3915 718 : });
3916 96 :
3917 96 : let layers = layers
3918 96 : .into_iter()
3919 324 : .map(|(desc, meta, atime, cold)| {
3920 324 : HeatMapLayer::new(desc.layer_name(), meta, atime, cold)
3921 324 : })
3922 96 : .collect();
3923 96 :
3924 96 : Some(HeatMapTimeline::new(self.timeline_id, layers))
3925 96 : }
3926 :
3927 0 : pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
3928 0 : let guard = self.layers.read().await;
3929 :
3930 0 : let now = SystemTime::now();
3931 0 : let mut heatmap_layers = Vec::default();
3932 0 : for vl in guard.visible_layers() {
3933 0 : if vl.layer_desc().get_lsn_range().start >= end_lsn {
3934 0 : continue;
3935 0 : }
3936 0 :
3937 0 : let hl = HeatMapLayer {
3938 0 : name: vl.layer_desc().layer_name(),
3939 0 : metadata: vl.metadata(),
3940 0 : access_time: now,
3941 0 : cold: true,
3942 0 : };
3943 0 : heatmap_layers.push(hl);
3944 : }
3945 :
3946 0 : tracing::info!(
3947 0 : "Generating unarchival heatmap with {} layers",
3948 0 : heatmap_layers.len()
3949 : );
3950 :
3951 0 : let heatmap = HeatMapTimeline::new(self.timeline_id, heatmap_layers);
3952 0 : PreviousHeatmap::Active {
3953 0 : heatmap,
3954 0 : read_at: Instant::now(),
3955 0 : end_lsn: Some(end_lsn),
3956 0 : }
3957 0 : }
3958 :
3959 : /// Returns true if the given lsn is or was an ancestor branchpoint.
3960 0 : pub(crate) fn is_ancestor_lsn(&self, lsn: Lsn) -> bool {
3961 0 : // upon timeline detach, we set the ancestor_lsn to Lsn::INVALID and the store the original
3962 0 : // branchpoint in the value in IndexPart::lineage
3963 0 : self.ancestor_lsn == lsn
3964 0 : || (self.ancestor_lsn == Lsn::INVALID
3965 0 : && self.remote_client.is_previous_ancestor_lsn(lsn))
3966 0 : }
3967 : }
3968 :
3969 : #[derive(Clone)]
3970 : /// Type representing a query in the ([`Lsn`], [`Key`]) space.
3971 : /// In other words, a set of segments in a 2D space.
3972 : ///
3973 : /// This representation has the advatange of avoiding hash map
3974 : /// allocations for uniform queries.
3975 : pub(crate) enum VersionedKeySpaceQuery {
3976 : /// Variant for queries at a single [`Lsn`]
3977 : Uniform { keyspace: KeySpace, lsn: Lsn },
3978 : /// Variant for queries at multiple [`Lsn`]s
3979 : Scattered {
3980 : keyspaces_at_lsn: Vec<(Lsn, KeySpace)>,
3981 : },
3982 : }
3983 :
3984 : impl VersionedKeySpaceQuery {
3985 3656468 : pub(crate) fn uniform(keyspace: KeySpace, lsn: Lsn) -> Self {
3986 3656468 : Self::Uniform { keyspace, lsn }
3987 3656468 : }
3988 :
3989 122304 : pub(crate) fn scattered(keyspaces_at_lsn: Vec<(Lsn, KeySpace)>) -> Self {
3990 122304 : Self::Scattered { keyspaces_at_lsn }
3991 122304 : }
3992 :
3993 : /// Returns the most recent (largest) LSN included in the query.
3994 : /// If any of the LSNs included in the query are invalid, returns
3995 : /// an error instead.
3996 7557544 : fn high_watermark_lsn(&self) -> Result<Lsn, GetVectoredError> {
3997 7557544 : match self {
3998 7312936 : Self::Uniform { lsn, .. } => {
3999 7312936 : if !lsn.is_valid() {
4000 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4001 7312936 : }
4002 7312936 :
4003 7312936 : Ok(*lsn)
4004 : }
4005 244608 : Self::Scattered { keyspaces_at_lsn } => {
4006 244608 : let mut max_lsn = None;
4007 506616 : for (lsn, _keyspace) in keyspaces_at_lsn.iter() {
4008 506616 : if !lsn.is_valid() {
4009 0 : return Err(GetVectoredError::InvalidLsn(*lsn));
4010 506616 : }
4011 506616 : max_lsn = std::cmp::max(max_lsn, Some(lsn));
4012 : }
4013 :
4014 244608 : if let Some(computed) = max_lsn {
4015 244608 : Ok(*computed)
4016 : } else {
4017 0 : Err(GetVectoredError::Other(anyhow!("empty input")))
4018 : }
4019 : }
4020 : }
4021 7557544 : }
4022 :
4023 : /// Returns the total keyspace being queried: the result of projecting
4024 : /// everything in the key dimensions onto the key axis.
4025 3910844 : fn total_keyspace(&self) -> KeySpace {
4026 3910844 : match self {
4027 3666236 : Self::Uniform { keyspace, .. } => keyspace.clone(),
4028 244608 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4029 244608 : .iter()
4030 506616 : .map(|(_lsn, keyspace)| keyspace)
4031 506616 : .fold(KeySpace::default(), |mut acc, v| {
4032 506616 : acc.merge(v);
4033 506616 : acc
4034 506616 : }),
4035 : }
4036 3910844 : }
4037 :
4038 : /// Returns LSN for a specific key.
4039 : ///
4040 : /// Invariant: requested key must be part of [`Self::total_keyspace`]
4041 4776788 : pub(super) fn map_key_to_lsn(&self, key: &Key) -> Lsn {
4042 4776788 : match self {
4043 3898484 : Self::Uniform { lsn, .. } => *lsn,
4044 878304 : Self::Scattered { keyspaces_at_lsn } => {
4045 878304 : keyspaces_at_lsn
4046 878304 : .iter()
4047 5040632 : .find(|(_lsn, keyspace)| keyspace.contains(key))
4048 878304 : .expect("Returned key was requested")
4049 878304 : .0
4050 : }
4051 : }
4052 4776788 : }
4053 :
4054 : /// Remove any parts of the query (segments) which overlap with the provided
4055 : /// key space (also segments).
4056 11616244 : fn remove_overlapping_with(&mut self, to_remove: &KeySpace) -> KeySpace {
4057 11616244 : match self {
4058 11371636 : Self::Uniform { keyspace, .. } => keyspace.remove_overlapping_with(to_remove),
4059 244608 : Self::Scattered { keyspaces_at_lsn } => {
4060 244608 : let mut removed_accum = KeySpaceRandomAccum::new();
4061 506616 : keyspaces_at_lsn.iter_mut().for_each(|(_lsn, keyspace)| {
4062 506616 : let removed = keyspace.remove_overlapping_with(to_remove);
4063 506616 : removed_accum.add_keyspace(removed);
4064 506616 : });
4065 244608 :
4066 244608 : removed_accum.to_keyspace()
4067 : }
4068 : }
4069 11616244 : }
4070 :
4071 8910440 : fn is_empty(&self) -> bool {
4072 8910440 : match self {
4073 8665832 : Self::Uniform { keyspace, .. } => keyspace.is_empty(),
4074 244608 : Self::Scattered { keyspaces_at_lsn } => keyspaces_at_lsn
4075 244608 : .iter()
4076 375612 : .all(|(_lsn, keyspace)| keyspace.is_empty()),
4077 : }
4078 8910440 : }
4079 :
4080 : /// "Lower" the query on the LSN dimension
4081 1352908 : fn lower(&mut self, to: Lsn) {
4082 1352908 : match self {
4083 1352908 : Self::Uniform { lsn, .. } => {
4084 1352908 : // If the originally requested LSN is smaller than the starting
4085 1352908 : // LSN of the ancestor we are descending into, we need to respect that.
4086 1352908 : // Hence the min.
4087 1352908 : *lsn = std::cmp::min(*lsn, to);
4088 1352908 : }
4089 0 : Self::Scattered { keyspaces_at_lsn } => {
4090 0 : keyspaces_at_lsn.iter_mut().for_each(|(lsn, _keyspace)| {
4091 0 : *lsn = std::cmp::min(*lsn, to);
4092 0 : });
4093 0 : }
4094 : }
4095 1352908 : }
4096 : }
4097 :
4098 : impl std::fmt::Display for VersionedKeySpaceQuery {
4099 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
4100 0 : write!(f, "[")?;
4101 :
4102 0 : match self {
4103 0 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4104 0 : write!(f, "{keyspace} @ {lsn}")?;
4105 : }
4106 0 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4107 0 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4108 0 : write!(f, "{keyspace} @ {lsn},")?;
4109 : }
4110 : }
4111 : }
4112 :
4113 0 : write!(f, "]")
4114 0 : }
4115 : }
4116 :
4117 : impl Timeline {
4118 : #[allow(clippy::doc_lazy_continuation)]
4119 : /// Get the data needed to reconstruct all keys in the provided keyspace
4120 : ///
4121 : /// The algorithm is as follows:
4122 : /// 1. While some keys are still not done and there's a timeline to visit:
4123 : /// 2. Visit the timeline (see [`Timeline::get_vectored_reconstruct_data_timeline`]:
4124 : /// 2.1: Build the fringe for the current keyspace
4125 : /// 2.2 Visit the newest layer from the fringe to collect all values for the range it
4126 : /// intersects
4127 : /// 2.3. Pop the timeline from the fringe
4128 : /// 2.4. If the fringe is empty, go back to 1
4129 3778772 : async fn get_vectored_reconstruct_data(
4130 3778772 : &self,
4131 3778772 : mut query: VersionedKeySpaceQuery,
4132 3778772 : reconstruct_state: &mut ValuesReconstructState,
4133 3778772 : ctx: &RequestContext,
4134 3778772 : ) -> Result<(), GetVectoredError> {
4135 3778772 : let original_hwm_lsn = query.high_watermark_lsn().unwrap();
4136 3778772 :
4137 3778772 : let mut timeline_owned: Arc<Timeline>;
4138 3778772 : let mut timeline = self;
4139 :
4140 3778760 : let missing_keyspace = loop {
4141 5131668 : if self.cancel.is_cancelled() {
4142 0 : return Err(GetVectoredError::Cancelled);
4143 5131668 : }
4144 :
4145 : let TimelineVisitOutcome {
4146 5131668 : completed_keyspace: completed,
4147 5131668 : image_covered_keyspace,
4148 : } = {
4149 5131668 : let ctx = RequestContextBuilder::from(ctx)
4150 5131668 : .perf_span(|crnt_perf_span| {
4151 0 : info_span!(
4152 : target: PERF_TRACE_TARGET,
4153 0 : parent: crnt_perf_span,
4154 : "PLAN_IO_TIMELINE",
4155 : timeline = %timeline.timeline_id,
4156 0 : high_watermark_lsn = %query.high_watermark_lsn().unwrap(),
4157 : )
4158 5131668 : })
4159 5131668 : .attached_child();
4160 5131668 :
4161 5131668 : Self::get_vectored_reconstruct_data_timeline(
4162 5131668 : timeline,
4163 5131668 : &query,
4164 5131668 : reconstruct_state,
4165 5131668 : &self.cancel,
4166 5131668 : &ctx,
4167 5131668 : )
4168 5131668 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4169 5131668 : .await?
4170 : };
4171 :
4172 5131668 : query.remove_overlapping_with(&completed);
4173 5131668 :
4174 5131668 : // Do not descend into the ancestor timeline for aux files.
4175 5131668 : // We don't return a blanket [`GetVectoredError::MissingKey`] to avoid
4176 5131668 : // stalling compaction.
4177 5131668 : query.remove_overlapping_with(&KeySpace {
4178 5131668 : ranges: vec![NON_INHERITED_RANGE, Key::sparse_non_inherited_keyspace()],
4179 5131668 : });
4180 5131668 :
4181 5131668 : // Keyspace is fully retrieved
4182 5131668 : if query.is_empty() {
4183 3777200 : break None;
4184 1354468 : }
4185 :
4186 1354468 : let Some(ancestor_timeline) = timeline.ancestor_timeline.as_ref() else {
4187 : // Not fully retrieved but no ancestor timeline.
4188 1560 : break Some(query.total_keyspace());
4189 : };
4190 :
4191 : // Now we see if there are keys covered by the image layer but does not exist in the
4192 : // image layer, which means that the key does not exist.
4193 :
4194 : // The block below will stop the vectored search if any of the keys encountered an image layer
4195 : // which did not contain a snapshot for said key. Since we have already removed all completed
4196 : // keys from `keyspace`, we expect there to be no overlap between it and the image covered key
4197 : // space. If that's not the case, we had at least one key encounter a gap in the image layer
4198 : // and stop the search as a result of that.
4199 1352908 : let mut removed = query.remove_overlapping_with(&image_covered_keyspace);
4200 1352908 : // Do not fire missing key error and end early for sparse keys. Note that we hava already removed
4201 1352908 : // non-inherited keyspaces before, so we can safely do a full `SPARSE_RANGE` remove instead of
4202 1352908 : // figuring out what is the inherited key range and do a fine-grained pruning.
4203 1352908 : removed.remove_overlapping_with(&KeySpace {
4204 1352908 : ranges: vec![SPARSE_RANGE],
4205 1352908 : });
4206 1352908 : if !removed.is_empty() {
4207 0 : break Some(removed);
4208 1352908 : }
4209 1352908 :
4210 1352908 : // Each key range in the original query is at some point in the LSN space.
4211 1352908 : // When descending into the ancestor, lower all ranges in the LSN space
4212 1352908 : // such that new changes on the parent timeline are not visible.
4213 1352908 : query.lower(timeline.ancestor_lsn);
4214 1352908 :
4215 1352908 : let ctx = RequestContextBuilder::from(ctx)
4216 1352908 : .perf_span(|crnt_perf_span| {
4217 0 : info_span!(
4218 : target: PERF_TRACE_TARGET,
4219 0 : parent: crnt_perf_span,
4220 : "GET_ANCESTOR",
4221 : timeline = %timeline.timeline_id,
4222 0 : ancestor = %ancestor_timeline.timeline_id,
4223 : ancestor_lsn = %timeline.ancestor_lsn
4224 : )
4225 1352908 : })
4226 1352908 : .attached_child();
4227 :
4228 1352908 : timeline_owned = timeline
4229 1352908 : .get_ready_ancestor_timeline(ancestor_timeline, &ctx)
4230 1352908 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
4231 1352908 : .await?;
4232 1352896 : timeline = &*timeline_owned;
4233 : };
4234 :
4235 : // Remove sparse keys from the keyspace so that it doesn't fire errors.
4236 3778760 : let missing_keyspace = if let Some(missing_keyspace) = missing_keyspace {
4237 1560 : let mut missing_keyspace = missing_keyspace;
4238 1560 : missing_keyspace.remove_overlapping_with(&KeySpace {
4239 1560 : ranges: vec![SPARSE_RANGE],
4240 1560 : });
4241 1560 : if missing_keyspace.is_empty() {
4242 1476 : None
4243 : } else {
4244 84 : Some(missing_keyspace)
4245 : }
4246 : } else {
4247 3777200 : None
4248 : };
4249 :
4250 3778760 : if let Some(missing_keyspace) = missing_keyspace {
4251 84 : return Err(GetVectoredError::MissingKey(Box::new(MissingKeyError {
4252 84 : keyspace: missing_keyspace, /* better if we can store the full keyspace */
4253 84 : shard: self.shard_identity.number,
4254 84 : original_hwm_lsn,
4255 84 : ancestor_lsn: Some(timeline.ancestor_lsn),
4256 84 : backtrace: None,
4257 84 : read_path: std::mem::take(&mut reconstruct_state.read_path),
4258 84 : query: None,
4259 84 : })));
4260 3778676 : }
4261 3778676 :
4262 3778676 : Ok(())
4263 3778772 : }
4264 :
4265 5131668 : async fn get_vectored_init_fringe(
4266 5131668 : &self,
4267 5131668 : query: &VersionedKeySpaceQuery,
4268 5131668 : ) -> Result<LayerFringe, GetVectoredError> {
4269 5131668 : let mut fringe = LayerFringe::new();
4270 5131668 : let guard = self.layers.read().await;
4271 :
4272 5131668 : match query {
4273 5009364 : VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
4274 5009364 : // LSNs requested by the compute or determined by the pageserver
4275 5009364 : // are inclusive. Queries to the layer map use exclusive LSNs.
4276 5009364 : // Hence, bump the value before the query - same in the other
4277 5009364 : // match arm.
4278 5009364 : let cont_lsn = Lsn(lsn.0 + 1);
4279 5009364 : guard.update_search_fringe(keyspace, cont_lsn, &mut fringe)?;
4280 : }
4281 122304 : VersionedKeySpaceQuery::Scattered { keyspaces_at_lsn } => {
4282 253308 : for (lsn, keyspace) in keyspaces_at_lsn.iter() {
4283 253308 : let cont_lsn_for_keyspace = Lsn(lsn.0 + 1);
4284 253308 : guard.update_search_fringe(keyspace, cont_lsn_for_keyspace, &mut fringe)?;
4285 : }
4286 : }
4287 : }
4288 :
4289 5131668 : Ok(fringe)
4290 5131668 : }
4291 :
4292 : /// Collect the reconstruct data for a keyspace from the specified timeline.
4293 : ///
4294 : /// Maintain a fringe [`LayerFringe`] which tracks all the layers that intersect
4295 : /// the current keyspace. The current keyspace of the search at any given timeline
4296 : /// is the original keyspace minus all the keys that have been completed minus
4297 : /// any keys for which we couldn't find an intersecting layer. It's not tracked explicitly,
4298 : /// but if you merge all the keyspaces in the fringe, you get the "current keyspace".
4299 : ///
4300 : /// This is basically a depth-first search visitor implementation where a vertex
4301 : /// is the (layer, lsn range, key space) tuple. The fringe acts as the stack.
4302 : ///
4303 : /// At each iteration pop the top of the fringe (the layer with the highest Lsn)
4304 : /// and get all the required reconstruct data from the layer in one go.
4305 : ///
4306 : /// Returns the completed keyspace and the keyspaces with image coverage. The caller
4307 : /// decides how to deal with these two keyspaces.
4308 5131668 : async fn get_vectored_reconstruct_data_timeline(
4309 5131668 : timeline: &Timeline,
4310 5131668 : query: &VersionedKeySpaceQuery,
4311 5131668 : reconstruct_state: &mut ValuesReconstructState,
4312 5131668 : cancel: &CancellationToken,
4313 5131668 : ctx: &RequestContext,
4314 5131668 : ) -> Result<TimelineVisitOutcome, GetVectoredError> {
4315 5131668 : // Prevent GC from progressing while visiting the current timeline.
4316 5131668 : // If we are GC-ing because a new image layer was added while traversing
4317 5131668 : // the timeline, then it will remove layers that are required for fulfilling
4318 5131668 : // the current get request (read-path cannot "look back" and notice the new
4319 5131668 : // image layer).
4320 5131668 : let _gc_cutoff_holder = timeline.get_applied_gc_cutoff_lsn();
4321 :
4322 : // See `compaction::compact_with_gc` for why we need this.
4323 5131668 : let _guard = timeline.gc_compaction_layer_update_lock.read().await;
4324 :
4325 : // Initialize the fringe
4326 5131668 : let mut fringe = timeline.get_vectored_init_fringe(query).await?;
4327 :
4328 5131668 : let mut completed_keyspace = KeySpace::default();
4329 5131668 : let mut image_covered_keyspace = KeySpaceRandomAccum::new();
4330 :
4331 10448294 : while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
4332 5316626 : if cancel.is_cancelled() {
4333 0 : return Err(GetVectoredError::Cancelled);
4334 5316626 : }
4335 :
4336 5316626 : if let Some(ref mut read_path) = reconstruct_state.read_path {
4337 5316626 : read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
4338 5316626 : }
4339 :
4340 : // Visit the layer and plan IOs for it
4341 5316626 : let next_cont_lsn = lsn_range.start;
4342 5316626 : layer_to_read
4343 5316626 : .get_values_reconstruct_data(
4344 5316626 : keyspace_to_read.clone(),
4345 5316626 : lsn_range,
4346 5316626 : reconstruct_state,
4347 5316626 : ctx,
4348 5316626 : )
4349 5316626 : .await?;
4350 :
4351 5316626 : let mut unmapped_keyspace = keyspace_to_read;
4352 5316626 : let cont_lsn = next_cont_lsn;
4353 5316626 :
4354 5316626 : reconstruct_state.on_layer_visited(&layer_to_read);
4355 5316626 :
4356 5316626 : let (keys_done_last_step, keys_with_image_coverage) =
4357 5316626 : reconstruct_state.consume_done_keys();
4358 5316626 : unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
4359 5316626 : completed_keyspace.merge(&keys_done_last_step);
4360 5316626 : if let Some(keys_with_image_coverage) = keys_with_image_coverage {
4361 181440 : unmapped_keyspace
4362 181440 : .remove_overlapping_with(&KeySpace::single(keys_with_image_coverage.clone()));
4363 181440 : image_covered_keyspace.add_range(keys_with_image_coverage);
4364 5135186 : }
4365 :
4366 : // Query the layer map for the next layers to read.
4367 : //
4368 : // Do not descent any further if the last layer we visited
4369 : // completed all keys in the keyspace it inspected. This is not
4370 : // required for correctness, but avoids visiting extra layers
4371 : // which turns out to be a perf bottleneck in some cases.
4372 5316626 : if !unmapped_keyspace.is_empty() {
4373 1473978 : let guard = timeline.layers.read().await;
4374 1473978 : guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
4375 :
4376 : // It's safe to drop the layer map lock after planning the next round of reads.
4377 : // The fringe keeps readable handles for the layers which are safe to read even
4378 : // if layers were compacted or flushed.
4379 : //
4380 : // The more interesting consideration is: "Why is the read algorithm still correct
4381 : // if the layer map changes while it is operating?". Doing a vectored read on a
4382 : // timeline boils down to pushing an imaginary lsn boundary downwards for each range
4383 : // covered by the read. The layer map tells us how to move the lsn downwards for a
4384 : // range at *a particular point in time*. It is fine for the answer to be different
4385 : // at two different time points.
4386 1473978 : drop(guard);
4387 3842648 : }
4388 : }
4389 :
4390 5131668 : Ok(TimelineVisitOutcome {
4391 5131668 : completed_keyspace,
4392 5131668 : image_covered_keyspace: image_covered_keyspace.consume_keyspace(),
4393 5131668 : })
4394 5131668 : }
4395 :
4396 1352908 : async fn get_ready_ancestor_timeline(
4397 1352908 : &self,
4398 1352908 : ancestor: &Arc<Timeline>,
4399 1352908 : ctx: &RequestContext,
4400 1352908 : ) -> Result<Arc<Timeline>, GetReadyAncestorError> {
4401 1352908 : // It's possible that the ancestor timeline isn't active yet, or
4402 1352908 : // is active but hasn't yet caught up to the branch point. Wait
4403 1352908 : // for it.
4404 1352908 : //
4405 1352908 : // This cannot happen while the pageserver is running normally,
4406 1352908 : // because you cannot create a branch from a point that isn't
4407 1352908 : // present in the pageserver yet. However, we don't wait for the
4408 1352908 : // branch point to be uploaded to cloud storage before creating
4409 1352908 : // a branch. I.e., the branch LSN need not be remote consistent
4410 1352908 : // for the branching operation to succeed.
4411 1352908 : //
4412 1352908 : // Hence, if we try to load a tenant in such a state where
4413 1352908 : // 1. the existence of the branch was persisted (in IndexPart and/or locally)
4414 1352908 : // 2. but the ancestor state is behind branch_lsn because it was not yet persisted
4415 1352908 : // then we will need to wait for the ancestor timeline to
4416 1352908 : // re-stream WAL up to branch_lsn before we access it.
4417 1352908 : //
4418 1352908 : // How can a tenant get in such a state?
4419 1352908 : // - ungraceful pageserver process exit
4420 1352908 : // - detach+attach => this is a bug, https://github.com/neondatabase/neon/issues/4219
4421 1352908 : //
4422 1352908 : // NB: this could be avoided by requiring
4423 1352908 : // branch_lsn >= remote_consistent_lsn
4424 1352908 : // during branch creation.
4425 1352908 : match ancestor.wait_to_become_active(ctx).await {
4426 1352896 : Ok(()) => {}
4427 : Err(TimelineState::Stopping) => {
4428 : // If an ancestor is stopping, it means the tenant is stopping: handle this the same as if this timeline was stopping.
4429 0 : return Err(GetReadyAncestorError::Cancelled);
4430 : }
4431 12 : Err(state) => {
4432 12 : return Err(GetReadyAncestorError::BadState {
4433 12 : timeline_id: ancestor.timeline_id,
4434 12 : state,
4435 12 : });
4436 : }
4437 : }
4438 1352896 : ancestor
4439 1352896 : .wait_lsn(
4440 1352896 : self.ancestor_lsn,
4441 1352896 : WaitLsnWaiter::Timeline(self),
4442 1352896 : WaitLsnTimeout::Default,
4443 1352896 : ctx,
4444 1352896 : )
4445 1352896 : .await
4446 1352896 : .map_err(|e| match e {
4447 0 : e @ WaitLsnError::Timeout(_) => GetReadyAncestorError::AncestorLsnTimeout(e),
4448 0 : WaitLsnError::Shutdown => GetReadyAncestorError::Cancelled,
4449 0 : WaitLsnError::BadState(state) => GetReadyAncestorError::BadState {
4450 0 : timeline_id: ancestor.timeline_id,
4451 0 : state,
4452 0 : },
4453 1352896 : })?;
4454 :
4455 1352896 : Ok(ancestor.clone())
4456 1352908 : }
4457 :
4458 1782936 : pub(crate) fn get_shard_identity(&self) -> &ShardIdentity {
4459 1782936 : &self.shard_identity
4460 1782936 : }
4461 :
4462 : #[inline(always)]
4463 0 : pub(crate) fn shard_timeline_id(&self) -> ShardTimelineId {
4464 0 : ShardTimelineId {
4465 0 : shard_index: ShardIndex {
4466 0 : shard_number: self.shard_identity.number,
4467 0 : shard_count: self.shard_identity.count,
4468 0 : },
4469 0 : timeline_id: self.timeline_id,
4470 0 : }
4471 0 : }
4472 :
4473 : /// Returns a non-frozen open in-memory layer for ingestion.
4474 : ///
4475 : /// Takes a witness of timeline writer state lock being held, because it makes no sense to call
4476 : /// this function without holding the mutex.
4477 7848 : async fn get_layer_for_write(
4478 7848 : &self,
4479 7848 : lsn: Lsn,
4480 7848 : _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4481 7848 : ctx: &RequestContext,
4482 7848 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
4483 7848 : let mut guard = self.layers.write().await;
4484 :
4485 7848 : let last_record_lsn = self.get_last_record_lsn();
4486 7848 : ensure!(
4487 7848 : lsn > last_record_lsn,
4488 0 : "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})",
4489 : lsn,
4490 : last_record_lsn,
4491 : );
4492 :
4493 7848 : let layer = guard
4494 7848 : .open_mut()?
4495 7848 : .get_layer_for_write(
4496 7848 : lsn,
4497 7848 : self.conf,
4498 7848 : self.timeline_id,
4499 7848 : self.tenant_shard_id,
4500 7848 : &self.gate,
4501 7848 : &self.cancel,
4502 7848 : ctx,
4503 7848 : )
4504 7848 : .await?;
4505 7848 : Ok(layer)
4506 7848 : }
4507 :
4508 31674624 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
4509 31674624 : assert!(new_lsn.is_aligned());
4510 :
4511 31674624 : self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
4512 31674624 : self.last_record_lsn.advance(new_lsn);
4513 31674624 : }
4514 :
4515 : /// Freeze any existing open in-memory layer and unconditionally notify the flush loop.
4516 : ///
4517 : /// Unconditional flush loop notification is given because in sharded cases we will want to
4518 : /// leave an Lsn gap. Unsharded tenants do not have Lsn gaps.
4519 7272 : async fn freeze_inmem_layer_at(
4520 7272 : &self,
4521 7272 : at: Lsn,
4522 7272 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
4523 7272 : ) -> Result<u64, FlushLayerError> {
4524 7272 : let frozen = {
4525 7272 : let mut guard = self.layers.write().await;
4526 7272 : guard
4527 7272 : .open_mut()?
4528 7272 : .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
4529 7272 : .await
4530 : };
4531 :
4532 7272 : if frozen {
4533 7104 : let now = Instant::now();
4534 7104 : *(self.last_freeze_ts.write().unwrap()) = now;
4535 7104 : }
4536 :
4537 : // Increment the flush cycle counter and wake up the flush task.
4538 : // Remember the new value, so that when we listen for the flush
4539 : // to finish, we know when the flush that we initiated has
4540 : // finished, instead of some other flush that was started earlier.
4541 7272 : let mut my_flush_request = 0;
4542 7272 :
4543 7272 : let flush_loop_state = { *self.flush_loop_state.lock().unwrap() };
4544 7272 : if !matches!(flush_loop_state, FlushLoopState::Running { .. }) {
4545 0 : return Err(FlushLayerError::NotRunning(flush_loop_state));
4546 7272 : }
4547 7272 :
4548 7272 : self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
4549 7272 : my_flush_request = *counter + 1;
4550 7272 : *counter = my_flush_request;
4551 7272 : *lsn = std::cmp::max(at, *lsn);
4552 7272 : });
4553 7272 :
4554 7272 : assert_ne!(my_flush_request, 0);
4555 :
4556 7272 : Ok(my_flush_request)
4557 7272 : }
4558 :
4559 : /// Layer flusher task's main loop.
4560 2748 : async fn flush_loop(
4561 2748 : self: &Arc<Self>,
4562 2748 : mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
4563 2748 : ctx: &RequestContext,
4564 2748 : ) {
4565 : // Subscribe to L0 delta layer updates, for compaction backpressure.
4566 2748 : let mut watch_l0 = match self.layers.read().await.layer_map() {
4567 2748 : Ok(lm) => lm.watch_level0_deltas(),
4568 0 : Err(Shutdown) => return,
4569 : };
4570 :
4571 2748 : info!("started flush loop");
4572 : loop {
4573 9870 : tokio::select! {
4574 9870 : _ = self.cancel.cancelled() => {
4575 60 : info!("shutting down layer flush task due to Timeline::cancel");
4576 60 : break;
4577 : },
4578 9870 : _ = layer_flush_start_rx.changed() => {}
4579 7122 : }
4580 7122 : trace!("waking up");
4581 7122 : let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
4582 7122 :
4583 7122 : // The highest LSN to which we flushed in the loop over frozen layers
4584 7122 : let mut flushed_to_lsn = Lsn(0);
4585 :
4586 7122 : let result = loop {
4587 14226 : if self.cancel.is_cancelled() {
4588 0 : info!("dropping out of flush loop for timeline shutdown");
4589 : // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
4590 : // anyone waiting on that will respect self.cancel as well: they will stop
4591 : // waiting at the same time we as drop out of this loop.
4592 0 : return;
4593 14226 : }
4594 14226 :
4595 14226 : // Break to notify potential waiters as soon as we've flushed the requested LSN. If
4596 14226 : // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
4597 14226 : if flushed_to_lsn >= frozen_to_lsn {
4598 6954 : break Ok(());
4599 7272 : }
4600 :
4601 : // Fetch the next layer to flush, if any.
4602 7272 : let (layer, l0_count, frozen_count, frozen_size) = {
4603 7272 : let layers = self.layers.read().await;
4604 7272 : let Ok(lm) = layers.layer_map() else {
4605 0 : info!("dropping out of flush loop for timeline shutdown");
4606 0 : return;
4607 : };
4608 7272 : let l0_count = lm.level0_deltas().len();
4609 7272 : let frozen_count = lm.frozen_layers.len();
4610 7272 : let frozen_size: u64 = lm
4611 7272 : .frozen_layers
4612 7272 : .iter()
4613 7272 : .map(|l| l.estimated_in_mem_size())
4614 7272 : .sum();
4615 7272 : let layer = lm.frozen_layers.front().cloned();
4616 7272 : (layer, l0_count, frozen_count, frozen_size)
4617 7272 : // drop 'layers' lock
4618 7272 : };
4619 7272 : let Some(layer) = layer else {
4620 168 : break Ok(());
4621 : };
4622 :
4623 : // Stall flushes to backpressure if compaction can't keep up. This is propagated up
4624 : // to WAL ingestion by having ephemeral layer rolls wait for flushes.
4625 7104 : if let Some(stall_threshold) = self.get_l0_flush_stall_threshold() {
4626 0 : if l0_count >= stall_threshold {
4627 0 : warn!(
4628 0 : "stalling layer flushes for compaction backpressure at {l0_count} \
4629 0 : L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4630 : );
4631 0 : let stall_timer = self
4632 0 : .metrics
4633 0 : .flush_delay_histo
4634 0 : .start_timer()
4635 0 : .record_on_drop();
4636 0 : tokio::select! {
4637 0 : result = watch_l0.wait_for(|l0| *l0 < stall_threshold) => {
4638 0 : if let Ok(l0) = result.as_deref() {
4639 0 : let delay = stall_timer.elapsed().as_secs_f64();
4640 0 : info!("resuming layer flushes at {l0} L0 layers after {delay:.3}s");
4641 0 : }
4642 : },
4643 0 : _ = self.cancel.cancelled() => {},
4644 : }
4645 0 : continue; // check again
4646 0 : }
4647 7104 : }
4648 :
4649 : // Flush the layer.
4650 7104 : let flush_timer = self.metrics.flush_time_histo.start_timer();
4651 7104 : match self.flush_frozen_layer(layer, ctx).await {
4652 7104 : Ok(layer_lsn) => flushed_to_lsn = max(flushed_to_lsn, layer_lsn),
4653 : Err(FlushLayerError::Cancelled) => {
4654 0 : info!("dropping out of flush loop for timeline shutdown");
4655 0 : return;
4656 : }
4657 0 : err @ Err(
4658 0 : FlushLayerError::NotRunning(_)
4659 0 : | FlushLayerError::Other(_)
4660 0 : | FlushLayerError::CreateImageLayersError(_),
4661 0 : ) => {
4662 0 : error!("could not flush frozen layer: {err:?}");
4663 0 : break err.map(|_| ());
4664 : }
4665 : }
4666 7104 : let flush_duration = flush_timer.stop_and_record();
4667 7104 :
4668 7104 : // Notify the tenant compaction loop if L0 compaction is needed.
4669 7104 : let l0_count = *watch_l0.borrow();
4670 7104 : if l0_count >= self.get_compaction_threshold() {
4671 2868 : self.l0_compaction_trigger.notify_one();
4672 4236 : }
4673 :
4674 : // Delay the next flush to backpressure if compaction can't keep up. We delay by the
4675 : // flush duration such that the flush takes 2x as long. This is propagated up to WAL
4676 : // ingestion by having ephemeral layer rolls wait for flushes.
4677 7104 : if let Some(delay_threshold) = self.get_l0_flush_delay_threshold() {
4678 12 : if l0_count >= delay_threshold {
4679 0 : let delay = flush_duration.as_secs_f64();
4680 0 : info!(
4681 0 : "delaying layer flush by {delay:.3}s for compaction backpressure at \
4682 0 : {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
4683 : );
4684 0 : let _delay_timer = self
4685 0 : .metrics
4686 0 : .flush_delay_histo
4687 0 : .start_timer()
4688 0 : .record_on_drop();
4689 0 : tokio::select! {
4690 0 : _ = tokio::time::sleep(flush_duration) => {},
4691 0 : _ = watch_l0.wait_for(|l0| *l0 < delay_threshold) => {},
4692 0 : _ = self.cancel.cancelled() => {},
4693 : }
4694 12 : }
4695 7092 : }
4696 : };
4697 :
4698 : // Unsharded tenants should never advance their LSN beyond the end of the
4699 : // highest layer they write: such gaps between layer data and the frozen LSN
4700 : // are only legal on sharded tenants.
4701 7122 : debug_assert!(
4702 7122 : self.shard_identity.count.count() > 1
4703 7122 : || flushed_to_lsn >= frozen_to_lsn
4704 168 : || !flushed_to_lsn.is_valid()
4705 : );
4706 :
4707 7122 : if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
4708 : // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
4709 : // to us via layer_flush_start_rx, then advance it here.
4710 : //
4711 : // This path is only taken for tenants with multiple shards: single sharded tenants should
4712 : // never encounter a gap in the wal.
4713 0 : let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
4714 0 : tracing::debug!(
4715 0 : "Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}"
4716 : );
4717 0 : if self.set_disk_consistent_lsn(frozen_to_lsn) {
4718 0 : if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
4719 0 : tracing::warn!(
4720 0 : "Failed to schedule metadata upload after updating disk_consistent_lsn: {e}"
4721 : );
4722 0 : }
4723 0 : }
4724 7122 : }
4725 :
4726 : // Notify any listeners that we're done
4727 7122 : let _ = self
4728 7122 : .layer_flush_done_tx
4729 7122 : .send_replace((flush_counter, result));
4730 : }
4731 60 : }
4732 :
4733 : /// Waits any flush request created by [`Self::freeze_inmem_layer_at`] to complete.
4734 6792 : async fn wait_flush_completion(&self, request: u64) -> Result<(), FlushLayerError> {
4735 6792 : let mut rx = self.layer_flush_done_tx.subscribe();
4736 : loop {
4737 : {
4738 13789 : let (last_result_counter, last_result) = &*rx.borrow();
4739 13789 : if *last_result_counter >= request {
4740 6792 : if let Err(err) = last_result {
4741 : // We already logged the original error in
4742 : // flush_loop. We cannot propagate it to the caller
4743 : // here, because it might not be Cloneable
4744 0 : return Err(err.clone());
4745 : } else {
4746 6792 : return Ok(());
4747 : }
4748 6997 : }
4749 6997 : }
4750 6997 : trace!("waiting for flush to complete");
4751 6997 : tokio::select! {
4752 6997 : rx_e = rx.changed() => {
4753 6997 : rx_e.map_err(|_| FlushLayerError::NotRunning(*self.flush_loop_state.lock().unwrap()))?;
4754 : },
4755 : // Cancellation safety: we are not leaving an I/O in-flight for the flush, we're just ignoring
4756 : // the notification from [`flush_loop`] that it completed.
4757 6997 : _ = self.cancel.cancelled() => {
4758 0 : tracing::info!("Cancelled layer flush due on timeline shutdown");
4759 0 : return Ok(())
4760 : }
4761 : };
4762 6997 : trace!("done")
4763 : }
4764 6792 : }
4765 :
4766 : /// Flush one frozen in-memory layer to disk, as a new delta layer.
4767 : ///
4768 : /// Return value is the last lsn (inclusive) of the layer that was frozen.
4769 : #[instrument(skip_all, fields(layer=%frozen_layer))]
4770 : async fn flush_frozen_layer(
4771 : self: &Arc<Self>,
4772 : frozen_layer: Arc<InMemoryLayer>,
4773 : ctx: &RequestContext,
4774 : ) -> Result<Lsn, FlushLayerError> {
4775 : debug_assert_current_span_has_tenant_and_timeline_id();
4776 :
4777 : // As a special case, when we have just imported an image into the repository,
4778 : // instead of writing out a L0 delta layer, we directly write out image layer
4779 : // files instead. This is possible as long as *all* the data imported into the
4780 : // repository have the same LSN.
4781 : let lsn_range = frozen_layer.get_lsn_range();
4782 :
4783 : // Whether to directly create image layers for this flush, or flush them as delta layers
4784 : let create_image_layer =
4785 : lsn_range.start == self.initdb_lsn && lsn_range.end == Lsn(self.initdb_lsn.0 + 1);
4786 :
4787 : #[cfg(test)]
4788 : {
4789 : match &mut *self.flush_loop_state.lock().unwrap() {
4790 : FlushLoopState::NotStarted | FlushLoopState::Exited => {
4791 : panic!("flush loop not running")
4792 : }
4793 : FlushLoopState::Running {
4794 : expect_initdb_optimization,
4795 : initdb_optimization_count,
4796 : ..
4797 : } => {
4798 : if create_image_layer {
4799 : *initdb_optimization_count += 1;
4800 : } else {
4801 : assert!(!*expect_initdb_optimization, "expected initdb optimization");
4802 : }
4803 : }
4804 : }
4805 : }
4806 :
4807 : let (layers_to_upload, delta_layer_to_add) = if create_image_layer {
4808 : // Note: The 'ctx' in use here has DownloadBehavior::Error. We should not
4809 : // require downloading anything during initial import.
4810 : let ((rel_partition, metadata_partition), _lsn) = self
4811 : .repartition(
4812 : self.initdb_lsn,
4813 : self.get_compaction_target_size(),
4814 : EnumSet::empty(),
4815 : ctx,
4816 : )
4817 : .await
4818 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
4819 :
4820 : if self.cancel.is_cancelled() {
4821 : return Err(FlushLayerError::Cancelled);
4822 : }
4823 :
4824 : // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace.
4825 : // So that the key ranges don't overlap.
4826 : let mut partitions = KeyPartitioning::default();
4827 : partitions.parts.extend(rel_partition.parts);
4828 : if !metadata_partition.parts.is_empty() {
4829 : assert_eq!(
4830 : metadata_partition.parts.len(),
4831 : 1,
4832 : "currently sparse keyspace should only contain a single metadata keyspace"
4833 : );
4834 : // Safety: create_image_layers treat sparse keyspaces differently that it does not scan
4835 : // every single key within the keyspace, and therefore, it's safe to force converting it
4836 : // into a dense keyspace before calling this function.
4837 : partitions
4838 : .parts
4839 : .extend(metadata_partition.into_dense().parts);
4840 : }
4841 :
4842 : let mut layers_to_upload = Vec::new();
4843 : let (generated_image_layers, is_complete) = self
4844 : .create_image_layers(
4845 : &partitions,
4846 : self.initdb_lsn,
4847 : ImageLayerCreationMode::Initial,
4848 : ctx,
4849 : LastImageLayerCreationStatus::Initial,
4850 : false, // don't yield for L0, we're flushing L0
4851 : )
4852 : .await?;
4853 : debug_assert!(
4854 : matches!(is_complete, LastImageLayerCreationStatus::Complete),
4855 : "init image generation mode must fully cover the keyspace"
4856 : );
4857 : layers_to_upload.extend(generated_image_layers);
4858 :
4859 : (layers_to_upload, None)
4860 : } else {
4861 : // Normal case, write out a L0 delta layer file.
4862 : // `create_delta_layer` will not modify the layer map.
4863 : // We will remove frozen layer and add delta layer in one atomic operation later.
4864 : let Some(layer) = self
4865 : .create_delta_layer(&frozen_layer, None, ctx)
4866 : .await
4867 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?
4868 : else {
4869 : panic!("delta layer cannot be empty if no filter is applied");
4870 : };
4871 : (
4872 : // FIXME: even though we have a single image and single delta layer assumption
4873 : // we push them to vec
4874 : vec![layer.clone()],
4875 : Some(layer),
4876 : )
4877 : };
4878 :
4879 : pausable_failpoint!("flush-layer-cancel-after-writing-layer-out-pausable");
4880 :
4881 : if self.cancel.is_cancelled() {
4882 : return Err(FlushLayerError::Cancelled);
4883 : }
4884 :
4885 : let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
4886 :
4887 : // The new on-disk layers are now in the layer map. We can remove the
4888 : // in-memory layer from the map now. The flushed layer is stored in
4889 : // the mapping in `create_delta_layer`.
4890 : {
4891 : let mut guard = self.layers.write().await;
4892 :
4893 : guard.open_mut()?.finish_flush_l0_layer(
4894 : delta_layer_to_add.as_ref(),
4895 : &frozen_layer,
4896 : &self.metrics,
4897 : );
4898 :
4899 : if self.set_disk_consistent_lsn(disk_consistent_lsn) {
4900 : // Schedule remote uploads that will reflect our new disk_consistent_lsn
4901 : self.schedule_uploads(disk_consistent_lsn, layers_to_upload)
4902 0 : .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
4903 : }
4904 : // release lock on 'layers'
4905 : };
4906 :
4907 : // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
4908 : // a compaction can delete the file and then it won't be available for uploads any more.
4909 : // We still schedule the upload, resulting in an error, but ideally we'd somehow avoid this
4910 : // race situation.
4911 : // See https://github.com/neondatabase/neon/issues/4526
4912 : pausable_failpoint!("flush-frozen-pausable");
4913 :
4914 : // This failpoint is used by another test case `test_pageserver_recovery`.
4915 : fail_point!("flush-frozen-exit");
4916 :
4917 : Ok(Lsn(lsn_range.end.0 - 1))
4918 : }
4919 :
4920 : /// Return true if the value changed
4921 : ///
4922 : /// This function must only be used from the layer flush task.
4923 7104 : fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
4924 7104 : let old_value = self.disk_consistent_lsn.fetch_max(new_value);
4925 7104 : assert!(
4926 7104 : new_value >= old_value,
4927 0 : "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}"
4928 : );
4929 :
4930 7104 : self.metrics
4931 7104 : .disk_consistent_lsn_gauge
4932 7104 : .set(new_value.0 as i64);
4933 7104 : new_value != old_value
4934 7104 : }
4935 :
4936 : /// Update metadata file
4937 7404 : fn schedule_uploads(
4938 7404 : &self,
4939 7404 : disk_consistent_lsn: Lsn,
4940 7404 : layers_to_upload: impl IntoIterator<Item = ResidentLayer>,
4941 7404 : ) -> anyhow::Result<()> {
4942 7404 : // We can only save a valid 'prev_record_lsn' value on disk if we
4943 7404 : // flushed *all* in-memory changes to disk. We only track
4944 7404 : // 'prev_record_lsn' in memory for the latest processed record, so we
4945 7404 : // don't remember what the correct value that corresponds to some old
4946 7404 : // LSN is. But if we flush everything, then the value corresponding
4947 7404 : // current 'last_record_lsn' is correct and we can store it on disk.
4948 7404 : let RecordLsn {
4949 7404 : last: last_record_lsn,
4950 7404 : prev: prev_record_lsn,
4951 7404 : } = self.last_record_lsn.load();
4952 7404 : let ondisk_prev_record_lsn = if disk_consistent_lsn == last_record_lsn {
4953 6636 : Some(prev_record_lsn)
4954 : } else {
4955 768 : None
4956 : };
4957 :
4958 7404 : let update = crate::tenant::metadata::MetadataUpdate::new(
4959 7404 : disk_consistent_lsn,
4960 7404 : ondisk_prev_record_lsn,
4961 7404 : *self.applied_gc_cutoff_lsn.read(),
4962 7404 : );
4963 7404 :
4964 7404 : fail_point!("checkpoint-before-saving-metadata", |x| bail!(
4965 0 : "{}",
4966 0 : x.unwrap()
4967 7404 : ));
4968 :
4969 14580 : for layer in layers_to_upload {
4970 7176 : self.remote_client.schedule_layer_file_upload(layer)?;
4971 : }
4972 7404 : self.remote_client
4973 7404 : .schedule_index_upload_for_metadata_update(&update)?;
4974 :
4975 7404 : Ok(())
4976 7404 : }
4977 :
4978 0 : pub(crate) async fn preserve_initdb_archive(&self) -> anyhow::Result<()> {
4979 0 : self.remote_client
4980 0 : .preserve_initdb_archive(
4981 0 : &self.tenant_shard_id.tenant_id,
4982 0 : &self.timeline_id,
4983 0 : &self.cancel,
4984 0 : )
4985 0 : .await
4986 0 : }
4987 :
4988 : // Write out the given frozen in-memory layer as a new L0 delta file. This L0 file will not be tracked
4989 : // in layer map immediately. The caller is responsible to put it into the layer map.
4990 5808 : async fn create_delta_layer(
4991 5808 : self: &Arc<Self>,
4992 5808 : frozen_layer: &Arc<InMemoryLayer>,
4993 5808 : key_range: Option<Range<Key>>,
4994 5808 : ctx: &RequestContext,
4995 5808 : ) -> anyhow::Result<Option<ResidentLayer>> {
4996 5808 : let self_clone = Arc::clone(self);
4997 5808 : let frozen_layer = Arc::clone(frozen_layer);
4998 5808 : let ctx = ctx.attached_child();
4999 5808 : let work = async move {
5000 5808 : let Some((desc, path)) = frozen_layer
5001 5808 : .write_to_disk(
5002 5808 : &ctx,
5003 5808 : key_range,
5004 5808 : self_clone.l0_flush_global_state.inner(),
5005 5808 : &self_clone.gate,
5006 5808 : self_clone.cancel.clone(),
5007 5808 : )
5008 5808 : .await?
5009 : else {
5010 0 : return Ok(None);
5011 : };
5012 5808 : let new_delta = Layer::finish_creating(self_clone.conf, &self_clone, desc, &path)?;
5013 :
5014 : // The write_to_disk() above calls writer.finish() which already did the fsync of the inodes.
5015 : // We just need to fsync the directory in which these inodes are linked,
5016 : // which we know to be the timeline directory.
5017 : //
5018 : // We use fatal_err() below because the after write_to_disk returns with success,
5019 : // the in-memory state of the filesystem already has the layer file in its final place,
5020 : // and subsequent pageserver code could think it's durable while it really isn't.
5021 5808 : let timeline_dir = VirtualFile::open(
5022 5808 : &self_clone
5023 5808 : .conf
5024 5808 : .timeline_path(&self_clone.tenant_shard_id, &self_clone.timeline_id),
5025 5808 : &ctx,
5026 5808 : )
5027 5808 : .await
5028 5808 : .fatal_err("VirtualFile::open for timeline dir fsync");
5029 5808 : timeline_dir
5030 5808 : .sync_all()
5031 5808 : .await
5032 5808 : .fatal_err("VirtualFile::sync_all timeline dir");
5033 5808 : anyhow::Ok(Some(new_delta))
5034 5808 : };
5035 : // Before tokio-epoll-uring, we ran write_to_disk & the sync_all inside spawn_blocking.
5036 : // Preserve that behavior to maintain the same behavior for `virtual_file_io_engine=std-fs`.
5037 : use crate::virtual_file::io_engine::IoEngine;
5038 5808 : match crate::virtual_file::io_engine::get() {
5039 0 : IoEngine::NotSet => panic!("io engine not set"),
5040 : IoEngine::StdFs => {
5041 2904 : let span = tracing::info_span!("blocking");
5042 2904 : tokio::task::spawn_blocking({
5043 2904 : move || Handle::current().block_on(work.instrument(span))
5044 2904 : })
5045 2904 : .await
5046 2904 : .context("spawn_blocking")
5047 2904 : .and_then(|x| x)
5048 : }
5049 : #[cfg(target_os = "linux")]
5050 2904 : IoEngine::TokioEpollUring => work.await,
5051 : }
5052 5808 : }
5053 :
5054 3480 : async fn repartition(
5055 3480 : &self,
5056 3480 : lsn: Lsn,
5057 3480 : partition_size: u64,
5058 3480 : flags: EnumSet<CompactFlags>,
5059 3480 : ctx: &RequestContext,
5060 3480 : ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
5061 3480 : let Ok(mut guard) = self.partitioning.try_write_guard() else {
5062 : // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
5063 : // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
5064 : // and hence before the compaction task starts.
5065 0 : return Err(CompactionError::Other(anyhow!(
5066 0 : "repartition() called concurrently"
5067 0 : )));
5068 : };
5069 3480 : let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
5070 3480 : if lsn < *partition_lsn {
5071 0 : return Err(CompactionError::Other(anyhow!(
5072 0 : "repartition() called with LSN going backwards, this should not happen"
5073 0 : )));
5074 3480 : }
5075 3480 :
5076 3480 : let distance = lsn.0 - partition_lsn.0;
5077 3480 : if *partition_lsn != Lsn(0)
5078 1572 : && distance <= self.repartition_threshold
5079 1572 : && !flags.contains(CompactFlags::ForceRepartition)
5080 : {
5081 1488 : debug!(
5082 : distance,
5083 : threshold = self.repartition_threshold,
5084 0 : "no repartitioning needed"
5085 : );
5086 1488 : return Ok((
5087 1488 : (dense_partition.clone(), sparse_partition.clone()),
5088 1488 : *partition_lsn,
5089 1488 : ));
5090 1992 : }
5091 :
5092 1992 : let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
5093 1992 : let dense_partitioning = dense_ks.partition(&self.shard_identity, partition_size);
5094 1992 : let sparse_partitioning = SparseKeyPartitioning {
5095 1992 : parts: vec![sparse_ks],
5096 1992 : }; // no partitioning for metadata keys for now
5097 1992 : let result = ((dense_partitioning, sparse_partitioning), lsn);
5098 1992 : guard.write(result.clone());
5099 1992 : Ok(result)
5100 3480 : }
5101 :
5102 : // Is it time to create a new image layer for the given partition? True if we want to generate.
5103 84 : async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
5104 84 : let threshold = self.get_image_creation_threshold();
5105 :
5106 84 : let guard = self.layers.read().await;
5107 84 : let Ok(layers) = guard.layer_map() else {
5108 0 : return false;
5109 : };
5110 :
5111 84 : let mut max_deltas = 0;
5112 168 : for part_range in &partition.ranges {
5113 84 : let image_coverage = layers.image_coverage(part_range, lsn);
5114 168 : for (img_range, last_img) in image_coverage {
5115 84 : let img_lsn = if let Some(last_img) = last_img {
5116 0 : last_img.get_lsn_range().end
5117 : } else {
5118 84 : Lsn(0)
5119 : };
5120 : // Let's consider an example:
5121 : //
5122 : // delta layer with LSN range 71-81
5123 : // delta layer with LSN range 81-91
5124 : // delta layer with LSN range 91-101
5125 : // image layer at LSN 100
5126 : //
5127 : // If 'lsn' is still 100, i.e. no new WAL has been processed since the last image layer,
5128 : // there's no need to create a new one. We check this case explicitly, to avoid passing
5129 : // a bogus range to count_deltas below, with start > end. It's even possible that there
5130 : // are some delta layers *later* than current 'lsn', if more WAL was processed and flushed
5131 : // after we read last_record_lsn, which is passed here in the 'lsn' argument.
5132 84 : if img_lsn < lsn {
5133 84 : let num_deltas =
5134 84 : layers.count_deltas(&img_range, &(img_lsn..lsn), Some(threshold));
5135 84 :
5136 84 : max_deltas = max_deltas.max(num_deltas);
5137 84 : if num_deltas >= threshold {
5138 0 : debug!(
5139 0 : "key range {}-{}, has {} deltas on this timeline in LSN range {}..{}",
5140 : img_range.start, img_range.end, num_deltas, img_lsn, lsn
5141 : );
5142 0 : return true;
5143 84 : }
5144 0 : }
5145 : }
5146 : }
5147 :
5148 84 : debug!(
5149 : max_deltas,
5150 0 : "none of the partitioned ranges had >= {threshold} deltas"
5151 : );
5152 84 : false
5153 84 : }
5154 :
5155 : /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
5156 : /// so that at most one image layer will be produced from this function.
5157 : #[allow(clippy::too_many_arguments)]
5158 1452 : async fn create_image_layer_for_rel_blocks(
5159 1452 : self: &Arc<Self>,
5160 1452 : partition: &KeySpace,
5161 1452 : mut image_layer_writer: ImageLayerWriter,
5162 1452 : lsn: Lsn,
5163 1452 : ctx: &RequestContext,
5164 1452 : img_range: Range<Key>,
5165 1452 : io_concurrency: IoConcurrency,
5166 1452 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5167 1452 : let mut wrote_keys = false;
5168 1452 :
5169 1452 : let mut key_request_accum = KeySpaceAccum::new();
5170 9636 : for range in &partition.ranges {
5171 8184 : let mut key = range.start;
5172 17736 : while key < range.end {
5173 : // Decide whether to retain this key: usually we do, but sharded tenants may
5174 : // need to drop keys that don't belong to them. If we retain the key, add it
5175 : // to `key_request_accum` for later issuing a vectored get
5176 9552 : if self.shard_identity.is_key_disposable(&key) {
5177 0 : debug!(
5178 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
5179 0 : key,
5180 0 : self.shard_identity.get_shard_number(&key)
5181 : );
5182 9552 : } else {
5183 9552 : key_request_accum.add_key(key);
5184 9552 : }
5185 :
5186 9552 : let last_key_in_range = key.next() == range.end;
5187 9552 : key = key.next();
5188 9552 :
5189 9552 : // Maybe flush `key_rest_accum`
5190 9552 : if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
5191 9552 : || (last_key_in_range && key_request_accum.raw_size() > 0)
5192 : {
5193 8184 : let query =
5194 8184 : VersionedKeySpaceQuery::uniform(key_request_accum.consume_keyspace(), lsn);
5195 :
5196 8184 : let results = self
5197 8184 : .get_vectored(query, io_concurrency.clone(), ctx)
5198 8184 : .await?;
5199 :
5200 8184 : if self.cancel.is_cancelled() {
5201 0 : return Err(CreateImageLayersError::Cancelled);
5202 8184 : }
5203 :
5204 17736 : for (img_key, img) in results {
5205 9552 : let img = match img {
5206 9552 : Ok(img) => img,
5207 0 : Err(err) => {
5208 0 : // If we fail to reconstruct a VM or FSM page, we can zero the
5209 0 : // page without losing any actual user data. That seems better
5210 0 : // than failing repeatedly and getting stuck.
5211 0 : //
5212 0 : // We had a bug at one point, where we truncated the FSM and VM
5213 0 : // in the pageserver, but the Postgres didn't know about that
5214 0 : // and continued to generate incremental WAL records for pages
5215 0 : // that didn't exist in the pageserver. Trying to replay those
5216 0 : // WAL records failed to find the previous image of the page.
5217 0 : // This special case allows us to recover from that situation.
5218 0 : // See https://github.com/neondatabase/neon/issues/2601.
5219 0 : //
5220 0 : // Unfortunately we cannot do this for the main fork, or for
5221 0 : // any metadata keys, keys, as that would lead to actual data
5222 0 : // loss.
5223 0 : if img_key.is_rel_fsm_block_key() || img_key.is_rel_vm_block_key() {
5224 0 : warn!(
5225 0 : "could not reconstruct FSM or VM key {img_key}, filling with zeros: {err:?}"
5226 : );
5227 0 : ZERO_PAGE.clone()
5228 : } else {
5229 0 : return Err(CreateImageLayersError::from(err));
5230 : }
5231 : }
5232 : };
5233 :
5234 : // Write all the keys we just read into our new image layer.
5235 9552 : image_layer_writer.put_image(img_key, img, ctx).await?;
5236 9552 : wrote_keys = true;
5237 : }
5238 1368 : }
5239 : }
5240 : }
5241 :
5242 1452 : if wrote_keys {
5243 : // Normal path: we have written some data into the new image layer for this
5244 : // partition, so flush it to disk.
5245 1452 : info!(
5246 0 : "produced image layer for rel {}",
5247 0 : ImageLayerName {
5248 0 : key_range: img_range.clone(),
5249 0 : lsn
5250 0 : },
5251 : );
5252 1452 : Ok(ImageLayerCreationOutcome::Generated {
5253 1452 : unfinished_image_layer: image_layer_writer,
5254 1452 : })
5255 : } else {
5256 0 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5257 0 : Ok(ImageLayerCreationOutcome::Empty)
5258 : }
5259 1452 : }
5260 :
5261 : /// Create an image layer for metadata keys. This function produces one image layer for all metadata
5262 : /// keys for now. Because metadata keys cannot exceed basebackup size limit, the image layer for it
5263 : /// would not be too large to fit in a single image layer.
5264 : ///
5265 : /// Creating image layers for metadata keys are different from relational keys. Firstly, instead of
5266 : /// iterating each key and get an image for each of them, we do a `vectored_get` scan over the sparse
5267 : /// keyspace to get all images in one run. Secondly, we use a different image layer generation metrics
5268 : /// for metadata keys than relational keys, which is the number of delta files visited during the scan.
5269 : #[allow(clippy::too_many_arguments)]
5270 1392 : async fn create_image_layer_for_metadata_keys(
5271 1392 : self: &Arc<Self>,
5272 1392 : partition: &KeySpace,
5273 1392 : mut image_layer_writer: ImageLayerWriter,
5274 1392 : lsn: Lsn,
5275 1392 : ctx: &RequestContext,
5276 1392 : img_range: Range<Key>,
5277 1392 : mode: ImageLayerCreationMode,
5278 1392 : io_concurrency: IoConcurrency,
5279 1392 : ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
5280 1392 : // Metadata keys image layer creation.
5281 1392 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
5282 1392 : let begin = Instant::now();
5283 : // Directly use `get_vectored_impl` to skip the max_vectored_read_key limit check. Note that the keyspace should
5284 : // not contain too many keys, otherwise this takes a lot of memory.
5285 1392 : let data = self
5286 1392 : .get_vectored_impl(
5287 1392 : VersionedKeySpaceQuery::uniform(partition.clone(), lsn),
5288 1392 : &mut reconstruct_state,
5289 1392 : ctx,
5290 1392 : )
5291 1392 : .await?;
5292 1392 : let (data, total_kb_retrieved, total_keys_retrieved) = {
5293 1392 : let mut new_data = BTreeMap::new();
5294 1392 : let mut total_kb_retrieved = 0;
5295 1392 : let mut total_keys_retrieved = 0;
5296 61464 : for (k, v) in data {
5297 60072 : let v = v?;
5298 60072 : total_kb_retrieved += KEY_SIZE + v.len();
5299 60072 : total_keys_retrieved += 1;
5300 60072 : new_data.insert(k, v);
5301 : }
5302 1392 : (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
5303 1392 : };
5304 1392 : let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
5305 1392 : let elapsed = begin.elapsed();
5306 1392 :
5307 1392 : let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
5308 1392 : info!(
5309 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",
5310 0 : elapsed.as_secs_f64()
5311 : );
5312 :
5313 1392 : if !trigger_generation && mode == ImageLayerCreationMode::Try {
5314 12 : return Ok(ImageLayerCreationOutcome::Skip);
5315 1380 : }
5316 1380 : if self.cancel.is_cancelled() {
5317 0 : return Err(CreateImageLayersError::Cancelled);
5318 1380 : }
5319 1380 : let mut wrote_any_image = false;
5320 61452 : for (k, v) in data {
5321 60072 : if v.is_empty() {
5322 : // the key has been deleted, it does not need an image
5323 : // in metadata keyspace, an empty image == tombstone
5324 48 : continue;
5325 60024 : }
5326 60024 : wrote_any_image = true;
5327 60024 :
5328 60024 : // No need to handle sharding b/c metadata keys are always on the 0-th shard.
5329 60024 :
5330 60024 : // TODO: split image layers to avoid too large layer files. Too large image files are not handled
5331 60024 : // on the normal data path either.
5332 60024 : image_layer_writer.put_image(k, v, ctx).await?;
5333 : }
5334 :
5335 1380 : if wrote_any_image {
5336 : // Normal path: we have written some data into the new image layer for this
5337 : // partition, so flush it to disk.
5338 72 : info!(
5339 0 : "created image layer for metadata {}",
5340 0 : ImageLayerName {
5341 0 : key_range: img_range.clone(),
5342 0 : lsn
5343 0 : }
5344 : );
5345 72 : Ok(ImageLayerCreationOutcome::Generated {
5346 72 : unfinished_image_layer: image_layer_writer,
5347 72 : })
5348 : } else {
5349 1308 : tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
5350 1308 : Ok(ImageLayerCreationOutcome::Empty)
5351 : }
5352 1392 : }
5353 :
5354 : /// Predicate function which indicates whether we should check if new image layers
5355 : /// are required. Since checking if new image layers are required is expensive in
5356 : /// terms of CPU, we only do it in the following cases:
5357 : /// 1. If the timeline has ingested sufficient WAL to justify the cost
5358 : /// 2. If enough time has passed since the last check:
5359 : /// 1. For large tenants, we wish to perform the check more often since they
5360 : /// suffer from the lack of image layers
5361 : /// 2. For small tenants (that can mostly fit in RAM), we use a much longer interval
5362 3480 : fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
5363 : const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
5364 :
5365 3480 : let last_checks_at = self.last_image_layer_creation_check_at.load();
5366 3480 : let distance = lsn
5367 3480 : .checked_sub(last_checks_at)
5368 3480 : .expect("Attempt to compact with LSN going backwards");
5369 3480 : let min_distance =
5370 3480 : self.get_image_layer_creation_check_threshold() as u64 * self.get_checkpoint_distance();
5371 3480 :
5372 3480 : let distance_based_decision = distance.0 >= min_distance;
5373 3480 :
5374 3480 : let mut time_based_decision = false;
5375 3480 : let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
5376 3480 : if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
5377 2868 : let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
5378 : {
5379 0 : self.get_checkpoint_timeout()
5380 : } else {
5381 2868 : Duration::from_secs(3600 * 48)
5382 : };
5383 :
5384 2868 : time_based_decision = match *last_check_instant {
5385 1572 : Some(last_check) => {
5386 1572 : let elapsed = last_check.elapsed();
5387 1572 : elapsed >= check_required_after
5388 : }
5389 1296 : None => true,
5390 : };
5391 612 : }
5392 :
5393 : // Do the expensive delta layer counting only if this timeline has ingested sufficient
5394 : // WAL since the last check or a checkpoint timeout interval has elapsed since the last
5395 : // check.
5396 3480 : let decision = distance_based_decision || time_based_decision;
5397 :
5398 3480 : if decision {
5399 1308 : self.last_image_layer_creation_check_at.store(lsn);
5400 1308 : *last_check_instant = Some(Instant::now());
5401 2172 : }
5402 :
5403 3480 : decision
5404 3480 : }
5405 :
5406 : /// Returns the image layers generated and an enum indicating whether the process is fully completed.
5407 : /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
5408 : #[tracing::instrument(skip_all, fields(%lsn, %mode))]
5409 : async fn create_image_layers(
5410 : self: &Arc<Timeline>,
5411 : partitioning: &KeyPartitioning,
5412 : lsn: Lsn,
5413 : mode: ImageLayerCreationMode,
5414 : ctx: &RequestContext,
5415 : last_status: LastImageLayerCreationStatus,
5416 : yield_for_l0: bool,
5417 : ) -> Result<(Vec<ResidentLayer>, LastImageLayerCreationStatus), CreateImageLayersError> {
5418 : let timer = self.metrics.create_images_time_histo.start_timer();
5419 :
5420 : if partitioning.parts.is_empty() {
5421 : warn!("no partitions to create image layers for");
5422 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5423 : }
5424 :
5425 : // We need to avoid holes between generated image layers.
5426 : // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
5427 : // image layer with hole between them. In this case such layer can not be utilized by GC.
5428 : //
5429 : // How such hole between partitions can appear?
5430 : // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
5431 : // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
5432 : // If there is delta layer <100000000..300000000> then it never be garbage collected because
5433 : // image layers <100000000..100000099> and <200000000..200000199> are not completely covering it.
5434 : let mut start = Key::MIN;
5435 :
5436 : let check_for_image_layers =
5437 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5438 : info!(
5439 : "resuming image layer creation: last_status=incomplete, continue from {}",
5440 : last_key
5441 : );
5442 : true
5443 : } else {
5444 : self.should_check_if_image_layers_required(lsn)
5445 : };
5446 :
5447 : let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
5448 :
5449 : let mut all_generated = true;
5450 :
5451 : let mut partition_processed = 0;
5452 : let mut total_partitions = partitioning.parts.len();
5453 : let mut last_partition_processed = None;
5454 : let mut partition_parts = partitioning.parts.clone();
5455 :
5456 : if let LastImageLayerCreationStatus::Incomplete { last_key } = last_status {
5457 : // We need to skip the partitions that have already been processed.
5458 : let mut found = false;
5459 : for (i, partition) in partition_parts.iter().enumerate() {
5460 : if last_key <= partition.end().unwrap() {
5461 : // ```plain
5462 : // |------|--------|----------|------|
5463 : // ^last_key
5464 : // ^start from this partition
5465 : // ```
5466 : // Why `i+1` instead of `i`?
5467 : // It is possible that the user did some writes after the previous image layer creation attempt so that
5468 : // a relation grows in size, and the last_key is now in the middle of the partition. In this case, we
5469 : // still want to skip this partition, so that we can make progress and avoid generating image layers over
5470 : // the same partition. Doing a mod to ensure we don't end up with an empty vec.
5471 : if i + 1 >= total_partitions {
5472 : // In general, this case should not happen -- if last_key is on the last partition, the previous
5473 : // iteration of image layer creation should return a complete status.
5474 : break; // with found=false
5475 : }
5476 : partition_parts = partition_parts.split_off(i + 1); // Remove the first i + 1 elements
5477 : total_partitions = partition_parts.len();
5478 : // Update the start key to the partition start.
5479 : start = partition_parts[0].start().unwrap();
5480 : found = true;
5481 : break;
5482 : }
5483 : }
5484 : if !found {
5485 : // Last key is within the last partition, or larger than all partitions.
5486 : return Ok((vec![], LastImageLayerCreationStatus::Complete));
5487 : }
5488 : }
5489 :
5490 : for partition in partition_parts.iter() {
5491 : if self.cancel.is_cancelled() {
5492 : return Err(CreateImageLayersError::Cancelled);
5493 : }
5494 : partition_processed += 1;
5495 : let img_range = start..partition.ranges.last().unwrap().end;
5496 : let compact_metadata = partition.overlaps(&Key::metadata_key_range());
5497 : if compact_metadata {
5498 : for range in &partition.ranges {
5499 : assert!(
5500 : range.start.field1 >= METADATA_KEY_BEGIN_PREFIX
5501 : && range.end.field1 <= METADATA_KEY_END_PREFIX,
5502 : "metadata keys must be partitioned separately"
5503 : );
5504 : }
5505 : if mode == ImageLayerCreationMode::Try && !check_for_image_layers {
5506 : // Skip compaction if there are not enough updates. Metadata compaction will do a scan and
5507 : // might mess up with evictions.
5508 : start = img_range.end;
5509 : continue;
5510 : }
5511 : // For initial and force modes, we always generate image layers for metadata keys.
5512 : } else if let ImageLayerCreationMode::Try = mode {
5513 : // check_for_image_layers = false -> skip
5514 : // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
5515 : if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
5516 : start = img_range.end;
5517 : continue;
5518 : }
5519 : }
5520 : if let ImageLayerCreationMode::Force = mode {
5521 : // When forced to create image layers, we might try and create them where they already
5522 : // exist. This mode is only used in tests/debug.
5523 : let layers = self.layers.read().await;
5524 : if layers.contains_key(&PersistentLayerKey {
5525 : key_range: img_range.clone(),
5526 : lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
5527 : is_delta: false,
5528 : }) {
5529 : // TODO: this can be processed with the BatchLayerWriter::finish_with_discard
5530 : // in the future.
5531 : tracing::info!(
5532 : "Skipping image layer at {lsn} {}..{}, already exists",
5533 : img_range.start,
5534 : img_range.end
5535 : );
5536 : start = img_range.end;
5537 : continue;
5538 : }
5539 : }
5540 :
5541 : let image_layer_writer = ImageLayerWriter::new(
5542 : self.conf,
5543 : self.timeline_id,
5544 : self.tenant_shard_id,
5545 : &img_range,
5546 : lsn,
5547 : &self.gate,
5548 : self.cancel.clone(),
5549 : ctx,
5550 : )
5551 : .await?;
5552 :
5553 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
5554 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
5555 0 : "failpoint image-layer-writer-fail-before-finish"
5556 0 : )))
5557 0 : });
5558 :
5559 : let io_concurrency = IoConcurrency::spawn_from_conf(
5560 : self.conf,
5561 : self.gate
5562 : .enter()
5563 0 : .map_err(|_| CreateImageLayersError::Cancelled)?,
5564 : );
5565 :
5566 : let outcome = if !compact_metadata {
5567 : self.create_image_layer_for_rel_blocks(
5568 : partition,
5569 : image_layer_writer,
5570 : lsn,
5571 : ctx,
5572 : img_range.clone(),
5573 : io_concurrency,
5574 : )
5575 : .await?
5576 : } else {
5577 : self.create_image_layer_for_metadata_keys(
5578 : partition,
5579 : image_layer_writer,
5580 : lsn,
5581 : ctx,
5582 : img_range.clone(),
5583 : mode,
5584 : io_concurrency,
5585 : )
5586 : .await?
5587 : };
5588 : match outcome {
5589 : ImageLayerCreationOutcome::Empty => {
5590 : // No data in this partition, so we don't need to create an image layer (for now).
5591 : // The next image layer should cover this key range, so we don't advance the `start`
5592 : // key.
5593 : }
5594 : ImageLayerCreationOutcome::Generated {
5595 : unfinished_image_layer,
5596 : } => {
5597 : batch_image_writer.add_unfinished_image_writer(
5598 : unfinished_image_layer,
5599 : img_range.clone(),
5600 : lsn,
5601 : );
5602 : // The next image layer should be generated right after this one.
5603 : start = img_range.end;
5604 : }
5605 : ImageLayerCreationOutcome::Skip => {
5606 : // We don't need to create an image layer for this partition.
5607 : // The next image layer should NOT cover this range, otherwise
5608 : // the keyspace becomes empty (reads don't go past image layers).
5609 : start = img_range.end;
5610 : }
5611 : }
5612 :
5613 : if let ImageLayerCreationMode::Try = mode {
5614 : // We have at least made some progress
5615 : if yield_for_l0 && batch_image_writer.pending_layer_num() >= 1 {
5616 : // The `Try` mode is currently only used on the compaction path. We want to avoid
5617 : // image layer generation taking too long time and blocking L0 compaction. So in this
5618 : // mode, we also inspect the current number of L0 layers and skip image layer generation
5619 : // if there are too many of them.
5620 : let image_preempt_threshold = self.get_image_creation_preempt_threshold()
5621 : * self.get_compaction_threshold();
5622 : // TODO: currently we do not respect `get_image_creation_preempt_threshold` and always yield
5623 : // when there is a single timeline with more than L0 threshold L0 layers. As long as the
5624 : // `get_image_creation_preempt_threshold` is set to a value greater than 0, we will yield for L0 compaction.
5625 : if image_preempt_threshold != 0 {
5626 : let should_yield = self
5627 : .l0_compaction_trigger
5628 : .notified()
5629 : .now_or_never()
5630 : .is_some();
5631 : if should_yield {
5632 : tracing::info!(
5633 : "preempt image layer generation at {lsn} when processing partition {}..{}: too many L0 layers",
5634 : partition.start().unwrap(),
5635 : partition.end().unwrap()
5636 : );
5637 : last_partition_processed = Some(partition.clone());
5638 : all_generated = false;
5639 : break;
5640 : }
5641 : }
5642 : }
5643 : }
5644 : }
5645 :
5646 : let image_layers = batch_image_writer.finish(self, ctx).await?;
5647 :
5648 : let mut guard = self.layers.write().await;
5649 :
5650 : // FIXME: we could add the images to be uploaded *before* returning from here, but right
5651 : // now they are being scheduled outside of write lock; current way is inconsistent with
5652 : // compaction lock order.
5653 : guard
5654 : .open_mut()?
5655 : .track_new_image_layers(&image_layers, &self.metrics);
5656 : drop_wlock(guard);
5657 : let duration = timer.stop_and_record();
5658 :
5659 : // Creating image layers may have caused some previously visible layers to be covered
5660 : if !image_layers.is_empty() {
5661 : self.update_layer_visibility().await?;
5662 : }
5663 :
5664 : let total_layer_size = image_layers
5665 : .iter()
5666 1524 : .map(|l| l.metadata().file_size)
5667 : .sum::<u64>();
5668 :
5669 : if !image_layers.is_empty() {
5670 : info!(
5671 : "created {} image layers ({} bytes) in {}s, processed {} out of {} partitions",
5672 : image_layers.len(),
5673 : total_layer_size,
5674 : duration.as_secs_f64(),
5675 : partition_processed,
5676 : total_partitions
5677 : );
5678 : }
5679 :
5680 : Ok((
5681 : image_layers,
5682 : if all_generated {
5683 : LastImageLayerCreationStatus::Complete
5684 : } else {
5685 : LastImageLayerCreationStatus::Incomplete {
5686 : last_key: if let Some(last_partition_processed) = last_partition_processed {
5687 : last_partition_processed.end().unwrap_or(Key::MIN)
5688 : } else {
5689 : // This branch should be unreachable, but in case it happens, we can just return the start key.
5690 : Key::MIN
5691 : },
5692 : }
5693 : },
5694 : ))
5695 : }
5696 :
5697 : /// Wait until the background initial logical size calculation is complete, or
5698 : /// this Timeline is shut down. Calling this function will cause the initial
5699 : /// logical size calculation to skip waiting for the background jobs barrier.
5700 0 : pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
5701 0 : if !self.shard_identity.is_shard_zero() {
5702 : // We don't populate logical size on shard >0: skip waiting for it.
5703 0 : return;
5704 0 : }
5705 0 :
5706 0 : if self.remote_client.is_deleting() {
5707 : // The timeline was created in a deletion-resume state, we don't expect logical size to be populated
5708 0 : return;
5709 0 : }
5710 0 :
5711 0 : if self.current_logical_size.current_size().is_exact() {
5712 : // root timelines are initialized with exact count, but never start the background
5713 : // calculation
5714 0 : return;
5715 0 : }
5716 0 :
5717 0 : if self.cancel.is_cancelled() {
5718 : // We already requested stopping the tenant, so we cannot wait for the logical size
5719 : // calculation to complete given the task might have been already cancelled.
5720 0 : return;
5721 0 : }
5722 :
5723 0 : if let Some(await_bg_cancel) = self
5724 0 : .current_logical_size
5725 0 : .cancel_wait_for_background_loop_concurrency_limit_semaphore
5726 0 : .get()
5727 0 : {
5728 0 : await_bg_cancel.cancel();
5729 0 : } else {
5730 : // We should not wait if we were not able to explicitly instruct
5731 : // the logical size cancellation to skip the concurrency limit semaphore.
5732 : // TODO: this is an unexpected case. We should restructure so that it
5733 : // can't happen.
5734 0 : tracing::warn!(
5735 0 : "await_initial_logical_size: can't get semaphore cancel token, skipping"
5736 : );
5737 0 : debug_assert!(false);
5738 : }
5739 :
5740 0 : tokio::select!(
5741 0 : _ = self.current_logical_size.initialized.acquire() => {},
5742 0 : _ = self.cancel.cancelled() => {}
5743 : )
5744 0 : }
5745 :
5746 : /// Detach this timeline from its ancestor by copying all of ancestors layers as this
5747 : /// Timelines layers up to the ancestor_lsn.
5748 : ///
5749 : /// Requires a timeline that:
5750 : /// - has an ancestor to detach from
5751 : /// - the ancestor does not have an ancestor -- follows from the original RFC limitations, not
5752 : /// a technical requirement
5753 : ///
5754 : /// After the operation has been started, it cannot be canceled. Upon restart it needs to be
5755 : /// polled again until completion.
5756 : ///
5757 : /// During the operation all timelines sharing the data with this timeline will be reparented
5758 : /// from our ancestor to be branches of this timeline.
5759 0 : pub(crate) async fn prepare_to_detach_from_ancestor(
5760 0 : self: &Arc<Timeline>,
5761 0 : tenant: &crate::tenant::TenantShard,
5762 0 : options: detach_ancestor::Options,
5763 0 : behavior: DetachBehavior,
5764 0 : ctx: &RequestContext,
5765 0 : ) -> Result<detach_ancestor::Progress, detach_ancestor::Error> {
5766 0 : detach_ancestor::prepare(self, tenant, behavior, options, ctx).await
5767 0 : }
5768 :
5769 : /// Second step of detach from ancestor; detaches the `self` from it's current ancestor and
5770 : /// reparents any reparentable children of previous ancestor.
5771 : ///
5772 : /// This method is to be called while holding the TenantManager's tenant slot, so during this
5773 : /// method we cannot be deleted nor can any timeline be deleted. After this method returns
5774 : /// successfully, tenant must be reloaded.
5775 : ///
5776 : /// Final step will be to [`Self::complete_detaching_timeline_ancestor`] after optionally
5777 : /// resetting the tenant.
5778 0 : pub(crate) async fn detach_from_ancestor_and_reparent(
5779 0 : self: &Arc<Timeline>,
5780 0 : tenant: &crate::tenant::TenantShard,
5781 0 : prepared: detach_ancestor::PreparedTimelineDetach,
5782 0 : ancestor_timeline_id: TimelineId,
5783 0 : ancestor_lsn: Lsn,
5784 0 : behavior: DetachBehavior,
5785 0 : ctx: &RequestContext,
5786 0 : ) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
5787 0 : detach_ancestor::detach_and_reparent(
5788 0 : self,
5789 0 : tenant,
5790 0 : prepared,
5791 0 : ancestor_timeline_id,
5792 0 : ancestor_lsn,
5793 0 : behavior,
5794 0 : ctx,
5795 0 : )
5796 0 : .await
5797 0 : }
5798 :
5799 : /// Final step which unblocks the GC.
5800 : ///
5801 : /// The tenant must've been reset if ancestry was modified previously (in tenant manager).
5802 0 : pub(crate) async fn complete_detaching_timeline_ancestor(
5803 0 : self: &Arc<Timeline>,
5804 0 : tenant: &crate::tenant::TenantShard,
5805 0 : attempt: detach_ancestor::Attempt,
5806 0 : ctx: &RequestContext,
5807 0 : ) -> Result<(), detach_ancestor::Error> {
5808 0 : detach_ancestor::complete(self, tenant, attempt, ctx).await
5809 0 : }
5810 : }
5811 :
5812 : impl Drop for Timeline {
5813 60 : fn drop(&mut self) {
5814 60 : if let Some(ancestor) = &self.ancestor_timeline {
5815 : // This lock should never be poisoned, but in case it is we do a .map() instead of
5816 : // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
5817 24 : if let Ok(mut gc_info) = ancestor.gc_info.write() {
5818 24 : if !gc_info.remove_child_not_offloaded(self.timeline_id) {
5819 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
5820 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
5821 24 : }
5822 0 : }
5823 36 : }
5824 60 : info!(
5825 0 : "Timeline {} for tenant {} is being dropped",
5826 : self.timeline_id, self.tenant_shard_id.tenant_id
5827 : );
5828 60 : }
5829 : }
5830 :
5831 : /// Top-level failure to compact.
5832 : #[derive(Debug, thiserror::Error)]
5833 : pub(crate) enum CompactionError {
5834 : #[error("The timeline or pageserver is shutting down")]
5835 : ShuttingDown,
5836 : /// Compaction tried to offload a timeline and failed
5837 : #[error("Failed to offload timeline: {0}")]
5838 : Offload(OffloadError),
5839 : /// Compaction cannot be done right now; page reconstruction and so on.
5840 : #[error("Failed to collect keyspace: {0}")]
5841 : CollectKeySpaceError(#[from] CollectKeySpaceError),
5842 : #[error(transparent)]
5843 : Other(anyhow::Error),
5844 : #[error("Compaction already running: {0}")]
5845 : AlreadyRunning(&'static str),
5846 : }
5847 :
5848 : impl CompactionError {
5849 : /// Errors that can be ignored, i.e., cancel and shutdown.
5850 0 : pub fn is_cancel(&self) -> bool {
5851 0 : matches!(
5852 0 : self,
5853 : Self::ShuttingDown
5854 : | Self::AlreadyRunning(_)
5855 : | Self::CollectKeySpaceError(CollectKeySpaceError::Cancelled)
5856 : | Self::CollectKeySpaceError(CollectKeySpaceError::PageRead(
5857 : PageReconstructError::Cancelled
5858 : ))
5859 : | Self::Offload(OffloadError::Cancelled)
5860 : )
5861 0 : }
5862 :
5863 : /// Critical errors that indicate data corruption.
5864 0 : pub fn is_critical(&self) -> bool {
5865 0 : matches!(
5866 0 : self,
5867 : Self::CollectKeySpaceError(
5868 : CollectKeySpaceError::Decode(_)
5869 : | CollectKeySpaceError::PageRead(
5870 : PageReconstructError::MissingKey(_) | PageReconstructError::WalRedo(_),
5871 : )
5872 : )
5873 : )
5874 0 : }
5875 : }
5876 :
5877 : impl From<OffloadError> for CompactionError {
5878 0 : fn from(e: OffloadError) -> Self {
5879 0 : match e {
5880 0 : OffloadError::Cancelled => Self::ShuttingDown,
5881 0 : _ => Self::Offload(e),
5882 : }
5883 0 : }
5884 : }
5885 :
5886 : impl From<super::upload_queue::NotInitialized> for CompactionError {
5887 0 : fn from(value: super::upload_queue::NotInitialized) -> Self {
5888 0 : match value {
5889 : super::upload_queue::NotInitialized::Uninitialized => {
5890 0 : CompactionError::Other(anyhow::anyhow!(value))
5891 : }
5892 : super::upload_queue::NotInitialized::ShuttingDown
5893 0 : | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
5894 : }
5895 0 : }
5896 : }
5897 :
5898 : impl From<super::storage_layer::layer::DownloadError> for CompactionError {
5899 0 : fn from(e: super::storage_layer::layer::DownloadError) -> Self {
5900 0 : match e {
5901 : super::storage_layer::layer::DownloadError::TimelineShutdown
5902 : | super::storage_layer::layer::DownloadError::DownloadCancelled => {
5903 0 : CompactionError::ShuttingDown
5904 : }
5905 : super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
5906 : | super::storage_layer::layer::DownloadError::DownloadRequired
5907 : | super::storage_layer::layer::DownloadError::NotFile(_)
5908 : | super::storage_layer::layer::DownloadError::DownloadFailed
5909 : | super::storage_layer::layer::DownloadError::PreStatFailed(_) => {
5910 0 : CompactionError::Other(anyhow::anyhow!(e))
5911 : }
5912 : #[cfg(test)]
5913 : super::storage_layer::layer::DownloadError::Failpoint(_) => {
5914 0 : CompactionError::Other(anyhow::anyhow!(e))
5915 : }
5916 : }
5917 0 : }
5918 : }
5919 :
5920 : impl From<layer_manager::Shutdown> for CompactionError {
5921 0 : fn from(_: layer_manager::Shutdown) -> Self {
5922 0 : CompactionError::ShuttingDown
5923 0 : }
5924 : }
5925 :
5926 : #[serde_as]
5927 1176 : #[derive(serde::Serialize)]
5928 : struct RecordedDuration(#[serde_as(as = "serde_with::DurationMicroSeconds")] Duration);
5929 :
5930 : #[derive(Default)]
5931 : enum DurationRecorder {
5932 : #[default]
5933 : NotStarted,
5934 : Recorded(RecordedDuration, tokio::time::Instant),
5935 : }
5936 :
5937 : impl DurationRecorder {
5938 3024 : fn till_now(&self) -> DurationRecorder {
5939 3024 : match self {
5940 : DurationRecorder::NotStarted => {
5941 0 : panic!("must only call on recorded measurements")
5942 : }
5943 3024 : DurationRecorder::Recorded(_, ended) => {
5944 3024 : let now = tokio::time::Instant::now();
5945 3024 : DurationRecorder::Recorded(RecordedDuration(now - *ended), now)
5946 3024 : }
5947 3024 : }
5948 3024 : }
5949 1176 : fn into_recorded(self) -> Option<RecordedDuration> {
5950 1176 : match self {
5951 0 : DurationRecorder::NotStarted => None,
5952 1176 : DurationRecorder::Recorded(recorded, _) => Some(recorded),
5953 : }
5954 1176 : }
5955 : }
5956 :
5957 : /// Descriptor for a delta layer used in testing infra. The start/end key/lsn range of the
5958 : /// delta layer might be different from the min/max key/lsn in the delta layer. Therefore,
5959 : /// the layer descriptor requires the user to provide the ranges, which should cover all
5960 : /// keys specified in the `data` field.
5961 : #[cfg(test)]
5962 : #[derive(Clone)]
5963 : pub struct DeltaLayerTestDesc {
5964 : pub lsn_range: Range<Lsn>,
5965 : pub key_range: Range<Key>,
5966 : pub data: Vec<(Key, Lsn, Value)>,
5967 : }
5968 :
5969 : #[cfg(test)]
5970 : #[derive(Clone)]
5971 : pub struct InMemoryLayerTestDesc {
5972 : pub lsn_range: Range<Lsn>,
5973 : pub data: Vec<(Key, Lsn, Value)>,
5974 : pub is_open: bool,
5975 : }
5976 :
5977 : #[cfg(test)]
5978 : impl DeltaLayerTestDesc {
5979 24 : pub fn new(lsn_range: Range<Lsn>, key_range: Range<Key>, data: Vec<(Key, Lsn, Value)>) -> Self {
5980 24 : Self {
5981 24 : lsn_range,
5982 24 : key_range,
5983 24 : data,
5984 24 : }
5985 24 : }
5986 :
5987 540 : pub fn new_with_inferred_key_range(
5988 540 : lsn_range: Range<Lsn>,
5989 540 : data: Vec<(Key, Lsn, Value)>,
5990 540 : ) -> Self {
5991 1392 : let key_min = data.iter().map(|(key, _, _)| key).min().unwrap();
5992 1392 : let key_max = data.iter().map(|(key, _, _)| key).max().unwrap();
5993 540 : Self {
5994 540 : key_range: (*key_min)..(key_max.next()),
5995 540 : lsn_range,
5996 540 : data,
5997 540 : }
5998 540 : }
5999 :
6000 60 : pub(crate) fn layer_name(&self) -> LayerName {
6001 60 : LayerName::Delta(super::storage_layer::DeltaLayerName {
6002 60 : key_range: self.key_range.clone(),
6003 60 : lsn_range: self.lsn_range.clone(),
6004 60 : })
6005 60 : }
6006 : }
6007 :
6008 : impl Timeline {
6009 168 : async fn finish_compact_batch(
6010 168 : self: &Arc<Self>,
6011 168 : new_deltas: &[ResidentLayer],
6012 168 : new_images: &[ResidentLayer],
6013 168 : layers_to_remove: &[Layer],
6014 168 : ) -> Result<(), CompactionError> {
6015 168 : let mut guard = tokio::select! {
6016 168 : guard = self.layers.write() => guard,
6017 168 : _ = self.cancel.cancelled() => {
6018 0 : return Err(CompactionError::ShuttingDown);
6019 : }
6020 : };
6021 :
6022 168 : let mut duplicated_layers = HashSet::new();
6023 168 :
6024 168 : let mut insert_layers = Vec::with_capacity(new_deltas.len());
6025 :
6026 2016 : for l in new_deltas {
6027 1848 : if guard.contains(l.as_ref()) {
6028 : // expected in tests
6029 0 : tracing::error!(layer=%l, "duplicated L1 layer");
6030 :
6031 : // good ways to cause a duplicate: we repeatedly error after taking the writelock
6032 : // `guard` on self.layers. as of writing this, there are no error returns except
6033 : // for compact_level0_phase1 creating an L0, which does not happen in practice
6034 : // because we have not implemented L0 => L0 compaction.
6035 0 : duplicated_layers.insert(l.layer_desc().key());
6036 1848 : } else if LayerMap::is_l0(&l.layer_desc().key_range, l.layer_desc().is_delta) {
6037 0 : return Err(CompactionError::Other(anyhow::anyhow!(
6038 0 : "compaction generates a L0 layer file as output, which will cause infinite compaction."
6039 0 : )));
6040 1848 : } else {
6041 1848 : insert_layers.push(l.clone());
6042 1848 : }
6043 : }
6044 :
6045 : // only remove those inputs which were not outputs
6046 168 : let remove_layers: Vec<Layer> = layers_to_remove
6047 168 : .iter()
6048 2412 : .filter(|l| !duplicated_layers.contains(&l.layer_desc().key()))
6049 168 : .cloned()
6050 168 : .collect();
6051 168 :
6052 168 : if !new_images.is_empty() {
6053 0 : guard
6054 0 : .open_mut()?
6055 0 : .track_new_image_layers(new_images, &self.metrics);
6056 168 : }
6057 :
6058 168 : guard
6059 168 : .open_mut()?
6060 168 : .finish_compact_l0(&remove_layers, &insert_layers, &self.metrics);
6061 168 :
6062 168 : self.remote_client
6063 168 : .schedule_compaction_update(&remove_layers, new_deltas)?;
6064 :
6065 168 : drop_wlock(guard);
6066 168 :
6067 168 : Ok(())
6068 168 : }
6069 :
6070 0 : async fn rewrite_layers(
6071 0 : self: &Arc<Self>,
6072 0 : mut replace_layers: Vec<(Layer, ResidentLayer)>,
6073 0 : mut drop_layers: Vec<Layer>,
6074 0 : ) -> Result<(), CompactionError> {
6075 0 : let mut guard = self.layers.write().await;
6076 :
6077 : // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
6078 : // to avoid double-removing, and avoid rewriting something that was removed.
6079 0 : replace_layers.retain(|(l, _)| guard.contains(l));
6080 0 : drop_layers.retain(|l| guard.contains(l));
6081 0 :
6082 0 : guard
6083 0 : .open_mut()?
6084 0 : .rewrite_layers(&replace_layers, &drop_layers, &self.metrics);
6085 0 :
6086 0 : let upload_layers: Vec<_> = replace_layers.into_iter().map(|r| r.1).collect();
6087 0 :
6088 0 : self.remote_client
6089 0 : .schedule_compaction_update(&drop_layers, &upload_layers)?;
6090 :
6091 0 : Ok(())
6092 0 : }
6093 :
6094 : /// Schedules the uploads of the given image layers
6095 2184 : fn upload_new_image_layers(
6096 2184 : self: &Arc<Self>,
6097 2184 : new_images: impl IntoIterator<Item = ResidentLayer>,
6098 2184 : ) -> Result<(), super::upload_queue::NotInitialized> {
6099 2340 : for layer in new_images {
6100 156 : self.remote_client.schedule_layer_file_upload(layer)?;
6101 : }
6102 : // should any new image layer been created, not uploading index_part will
6103 : // result in a mismatch between remote_physical_size and layermap calculated
6104 : // size, which will fail some tests, but should not be an issue otherwise.
6105 2184 : self.remote_client
6106 2184 : .schedule_index_upload_for_file_changes()?;
6107 2184 : Ok(())
6108 2184 : }
6109 :
6110 0 : async fn find_gc_time_cutoff(
6111 0 : &self,
6112 0 : now: SystemTime,
6113 0 : pitr: Duration,
6114 0 : cancel: &CancellationToken,
6115 0 : ctx: &RequestContext,
6116 0 : ) -> Result<Option<Lsn>, PageReconstructError> {
6117 0 : debug_assert_current_span_has_tenant_and_timeline_id();
6118 0 : if self.shard_identity.is_shard_zero() {
6119 : // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
6120 0 : let time_range = if pitr == Duration::ZERO {
6121 0 : humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
6122 : } else {
6123 0 : pitr
6124 : };
6125 :
6126 : // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
6127 0 : let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
6128 0 : let timestamp = to_pg_timestamp(time_cutoff);
6129 :
6130 0 : let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
6131 0 : LsnForTimestamp::Present(lsn) => Some(lsn),
6132 0 : LsnForTimestamp::Future(lsn) => {
6133 0 : // The timestamp is in the future. That sounds impossible,
6134 0 : // but what it really means is that there hasn't been
6135 0 : // any commits since the cutoff timestamp.
6136 0 : //
6137 0 : // In this case we should use the LSN of the most recent commit,
6138 0 : // which is implicitly the last LSN in the log.
6139 0 : debug!("future({})", lsn);
6140 0 : Some(self.get_last_record_lsn())
6141 : }
6142 0 : LsnForTimestamp::Past(lsn) => {
6143 0 : debug!("past({})", lsn);
6144 0 : None
6145 : }
6146 0 : LsnForTimestamp::NoData(lsn) => {
6147 0 : debug!("nodata({})", lsn);
6148 0 : None
6149 : }
6150 : };
6151 0 : Ok(time_cutoff)
6152 : } else {
6153 : // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
6154 : // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that
6155 : // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
6156 : // space cutoff that we would also have respected ourselves.
6157 0 : match self
6158 0 : .remote_client
6159 0 : .download_foreign_index(ShardNumber(0), cancel)
6160 0 : .await
6161 : {
6162 0 : Ok((index_part, index_generation, _index_mtime)) => {
6163 0 : tracing::info!(
6164 0 : "GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
6165 0 : index_part.metadata.latest_gc_cutoff_lsn()
6166 : );
6167 0 : Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
6168 : }
6169 : Err(DownloadError::NotFound) => {
6170 : // This is unexpected, because during timeline creations shard zero persists to remote
6171 : // storage before other shards are called, and during timeline deletion non-zeroth shards are
6172 : // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this
6173 : // state, then shard zero should _eventually_ write an index when it GCs.
6174 0 : tracing::warn!("GC couldn't find shard zero's index for timeline");
6175 0 : Ok(None)
6176 : }
6177 0 : Err(e) => {
6178 0 : // TODO: this function should return a different error type than page reconstruct error
6179 0 : Err(PageReconstructError::Other(anyhow::anyhow!(e)))
6180 : }
6181 : }
6182 :
6183 : // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
6184 : // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
6185 : // is going through a migration if we read the old location's index and it has GC'd ahead of the
6186 : // new location. This is legal in principle, but problematic in practice because it might result
6187 : // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
6188 : // because they have GC'd past the branch point.
6189 : }
6190 0 : }
6191 :
6192 : /// Find the Lsns above which layer files need to be retained on
6193 : /// garbage collection.
6194 : ///
6195 : /// We calculate two cutoffs, one based on time and one based on WAL size. `pitr`
6196 : /// controls the time cutoff (or ZERO to disable time-based retention), and `space_cutoff` controls
6197 : /// the space-based retention.
6198 : ///
6199 : /// This function doesn't simply to calculate time & space based retention: it treats time-based
6200 : /// retention as authoritative if enabled, and falls back to space-based retention if calculating
6201 : /// the LSN for a time point isn't possible. Therefore the GcCutoffs::horizon in the response might
6202 : /// be different to the `space_cutoff` input. Callers should treat the min() of the two cutoffs
6203 : /// in the response as the GC cutoff point for the timeline.
6204 : #[instrument(skip_all, fields(timeline_id=%self.timeline_id))]
6205 : pub(super) async fn find_gc_cutoffs(
6206 : &self,
6207 : now: SystemTime,
6208 : space_cutoff: Lsn,
6209 : pitr: Duration,
6210 : cancel: &CancellationToken,
6211 : ctx: &RequestContext,
6212 : ) -> Result<GcCutoffs, PageReconstructError> {
6213 : let _timer = self
6214 : .metrics
6215 : .find_gc_cutoffs_histo
6216 : .start_timer()
6217 : .record_on_drop();
6218 :
6219 : pausable_failpoint!("Timeline::find_gc_cutoffs-pausable");
6220 :
6221 : if cfg!(test) {
6222 : // Unit tests which specify zero PITR interval expect to avoid doing any I/O for timestamp lookup
6223 : if pitr == Duration::ZERO {
6224 : return Ok(GcCutoffs {
6225 : time: self.get_last_record_lsn(),
6226 : space: space_cutoff,
6227 : });
6228 : }
6229 : }
6230 :
6231 : // Calculate a time-based limit on how much to retain:
6232 : // - if PITR interval is set, then this is our cutoff.
6233 : // - if PITR interval is not set, then we do a lookup
6234 : // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
6235 : let time_cutoff = self.find_gc_time_cutoff(now, pitr, cancel, ctx).await?;
6236 :
6237 : Ok(match (pitr, time_cutoff) {
6238 : (Duration::ZERO, Some(time_cutoff)) => {
6239 : // PITR is not set. Retain the size-based limit, or the default time retention,
6240 : // whichever requires less data.
6241 : GcCutoffs {
6242 : time: self.get_last_record_lsn(),
6243 : space: std::cmp::max(time_cutoff, space_cutoff),
6244 : }
6245 : }
6246 : (Duration::ZERO, None) => {
6247 : // PITR is not set, and time lookup failed
6248 : GcCutoffs {
6249 : time: self.get_last_record_lsn(),
6250 : space: space_cutoff,
6251 : }
6252 : }
6253 : (_, None) => {
6254 : // PITR interval is set & we didn't look up a timestamp successfully. Conservatively assume PITR
6255 : // cannot advance beyond what was already GC'd, and respect space-based retention
6256 : GcCutoffs {
6257 : time: *self.get_applied_gc_cutoff_lsn(),
6258 : space: space_cutoff,
6259 : }
6260 : }
6261 : (_, Some(time_cutoff)) => {
6262 : // PITR interval is set and we looked up timestamp successfully. Ignore
6263 : // size based retention and make time cutoff authoritative
6264 : GcCutoffs {
6265 : time: time_cutoff,
6266 : space: time_cutoff,
6267 : }
6268 : }
6269 : })
6270 : }
6271 :
6272 : /// Garbage collect layer files on a timeline that are no longer needed.
6273 : ///
6274 : /// Currently, we don't make any attempt at removing unneeded page versions
6275 : /// within a layer file. We can only remove the whole file if it's fully
6276 : /// obsolete.
6277 24 : pub(super) async fn gc(&self) -> Result<GcResult, GcError> {
6278 : // this is most likely the background tasks, but it might be the spawned task from
6279 : // immediate_gc
6280 24 : let _g = tokio::select! {
6281 24 : guard = self.gc_lock.lock() => guard,
6282 24 : _ = self.cancel.cancelled() => return Ok(GcResult::default()),
6283 : };
6284 24 : let timer = self.metrics.garbage_collect_histo.start_timer();
6285 24 :
6286 24 : fail_point!("before-timeline-gc");
6287 24 :
6288 24 : // Is the timeline being deleted?
6289 24 : if self.is_stopping() {
6290 0 : return Err(GcError::TimelineCancelled);
6291 24 : }
6292 24 :
6293 24 : let (space_cutoff, time_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
6294 24 : let gc_info = self.gc_info.read().unwrap();
6295 24 :
6296 24 : let space_cutoff = min(gc_info.cutoffs.space, self.get_disk_consistent_lsn());
6297 24 : let time_cutoff = gc_info.cutoffs.time;
6298 24 : let retain_lsns = gc_info
6299 24 : .retain_lsns
6300 24 : .iter()
6301 24 : .map(|(lsn, _child_id, _is_offloaded)| *lsn)
6302 24 : .collect();
6303 24 :
6304 24 : // Gets the maximum LSN that holds the valid lease.
6305 24 : //
6306 24 : // Caveat: `refresh_gc_info` is in charged of updating the lease map.
6307 24 : // Here, we do not check for stale leases again.
6308 24 : let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
6309 24 :
6310 24 : (
6311 24 : space_cutoff,
6312 24 : time_cutoff,
6313 24 : retain_lsns,
6314 24 : max_lsn_with_valid_lease,
6315 24 : )
6316 24 : };
6317 24 :
6318 24 : let mut new_gc_cutoff = Lsn::min(space_cutoff, time_cutoff);
6319 24 : let standby_horizon = self.standby_horizon.load();
6320 24 : // Hold GC for the standby, but as a safety guard do it only within some
6321 24 : // reasonable lag.
6322 24 : if standby_horizon != Lsn::INVALID {
6323 0 : if let Some(standby_lag) = new_gc_cutoff.checked_sub(standby_horizon) {
6324 : const MAX_ALLOWED_STANDBY_LAG: u64 = 10u64 << 30; // 10 GB
6325 0 : if standby_lag.0 < MAX_ALLOWED_STANDBY_LAG {
6326 0 : new_gc_cutoff = Lsn::min(standby_horizon, new_gc_cutoff);
6327 0 : trace!("holding off GC for standby apply LSN {}", standby_horizon);
6328 : } else {
6329 0 : warn!(
6330 0 : "standby is lagging for more than {}MB, not holding gc for it",
6331 0 : MAX_ALLOWED_STANDBY_LAG / 1024 / 1024
6332 : )
6333 : }
6334 0 : }
6335 24 : }
6336 :
6337 : // Reset standby horizon to ignore it if it is not updated till next GC.
6338 : // It is an easy way to unset it when standby disappears without adding
6339 : // more conf options.
6340 24 : self.standby_horizon.store(Lsn::INVALID);
6341 24 : self.metrics
6342 24 : .standby_horizon_gauge
6343 24 : .set(Lsn::INVALID.0 as i64);
6344 :
6345 24 : let res = self
6346 24 : .gc_timeline(
6347 24 : space_cutoff,
6348 24 : time_cutoff,
6349 24 : retain_lsns,
6350 24 : max_lsn_with_valid_lease,
6351 24 : new_gc_cutoff,
6352 24 : )
6353 24 : .instrument(
6354 24 : info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
6355 : )
6356 24 : .await?;
6357 :
6358 : // only record successes
6359 24 : timer.stop_and_record();
6360 24 :
6361 24 : Ok(res)
6362 24 : }
6363 :
6364 24 : async fn gc_timeline(
6365 24 : &self,
6366 24 : space_cutoff: Lsn,
6367 24 : time_cutoff: Lsn,
6368 24 : retain_lsns: Vec<Lsn>,
6369 24 : max_lsn_with_valid_lease: Option<Lsn>,
6370 24 : new_gc_cutoff: Lsn,
6371 24 : ) -> Result<GcResult, GcError> {
6372 24 : // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
6373 24 :
6374 24 : let now = SystemTime::now();
6375 24 : let mut result: GcResult = GcResult::default();
6376 24 :
6377 24 : // Nothing to GC. Return early.
6378 24 : let latest_gc_cutoff = *self.get_applied_gc_cutoff_lsn();
6379 24 : if latest_gc_cutoff >= new_gc_cutoff {
6380 0 : info!(
6381 0 : "Nothing to GC: new_gc_cutoff_lsn {new_gc_cutoff}, latest_gc_cutoff_lsn {latest_gc_cutoff}",
6382 : );
6383 0 : return Ok(result);
6384 24 : }
6385 :
6386 : // We need to ensure that no one tries to read page versions or create
6387 : // branches at a point before latest_gc_cutoff_lsn. See branch_timeline()
6388 : // for details. This will block until the old value is no longer in use.
6389 : //
6390 : // The GC cutoff should only ever move forwards.
6391 24 : let waitlist = {
6392 24 : let write_guard = self.applied_gc_cutoff_lsn.lock_for_write();
6393 24 : if *write_guard > new_gc_cutoff {
6394 0 : return Err(GcError::BadLsn {
6395 0 : why: format!(
6396 0 : "Cannot move GC cutoff LSN backwards (was {}, new {})",
6397 0 : *write_guard, new_gc_cutoff
6398 0 : ),
6399 0 : });
6400 24 : }
6401 24 :
6402 24 : write_guard.store_and_unlock(new_gc_cutoff)
6403 24 : };
6404 24 : waitlist.wait().await;
6405 :
6406 24 : info!("GC starting");
6407 :
6408 24 : debug!("retain_lsns: {:?}", retain_lsns);
6409 :
6410 24 : let mut layers_to_remove = Vec::new();
6411 :
6412 : // Scan all layers in the timeline (remote or on-disk).
6413 : //
6414 : // Garbage collect the layer if all conditions are satisfied:
6415 : // 1. it is older than cutoff LSN;
6416 : // 2. it is older than PITR interval;
6417 : // 3. it doesn't need to be retained for 'retain_lsns';
6418 : // 4. it does not need to be kept for LSNs holding valid leases.
6419 : // 5. newer on-disk image layers cover the layer's whole key range
6420 : //
6421 : // TODO holding a write lock is too agressive and avoidable
6422 24 : let mut guard = self.layers.write().await;
6423 24 : let layers = guard.layer_map()?;
6424 144 : 'outer: for l in layers.iter_historic_layers() {
6425 144 : result.layers_total += 1;
6426 144 :
6427 144 : // 1. Is it newer than GC horizon cutoff point?
6428 144 : if l.get_lsn_range().end > space_cutoff {
6429 12 : info!(
6430 0 : "keeping {} because it's newer than space_cutoff {}",
6431 0 : l.layer_name(),
6432 : space_cutoff,
6433 : );
6434 12 : result.layers_needed_by_cutoff += 1;
6435 12 : continue 'outer;
6436 132 : }
6437 132 :
6438 132 : // 2. It is newer than PiTR cutoff point?
6439 132 : if l.get_lsn_range().end > time_cutoff {
6440 0 : info!(
6441 0 : "keeping {} because it's newer than time_cutoff {}",
6442 0 : l.layer_name(),
6443 : time_cutoff,
6444 : );
6445 0 : result.layers_needed_by_pitr += 1;
6446 0 : continue 'outer;
6447 132 : }
6448 :
6449 : // 3. Is it needed by a child branch?
6450 : // NOTE With that we would keep data that
6451 : // might be referenced by child branches forever.
6452 : // We can track this in child timeline GC and delete parent layers when
6453 : // they are no longer needed. This might be complicated with long inheritance chains.
6454 : //
6455 : // TODO Vec is not a great choice for `retain_lsns`
6456 132 : for retain_lsn in &retain_lsns {
6457 : // start_lsn is inclusive
6458 0 : if &l.get_lsn_range().start <= retain_lsn {
6459 0 : info!(
6460 0 : "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
6461 0 : l.layer_name(),
6462 0 : retain_lsn,
6463 0 : l.is_incremental(),
6464 : );
6465 0 : result.layers_needed_by_branches += 1;
6466 0 : continue 'outer;
6467 0 : }
6468 : }
6469 :
6470 : // 4. Is there a valid lease that requires us to keep this layer?
6471 132 : if let Some(lsn) = &max_lsn_with_valid_lease {
6472 : // keep if layer start <= any of the lease
6473 108 : if &l.get_lsn_range().start <= lsn {
6474 84 : info!(
6475 0 : "keeping {} because there is a valid lease preventing GC at {}",
6476 0 : l.layer_name(),
6477 : lsn,
6478 : );
6479 84 : result.layers_needed_by_leases += 1;
6480 84 : continue 'outer;
6481 24 : }
6482 24 : }
6483 :
6484 : // 5. Is there a later on-disk layer for this relation?
6485 : //
6486 : // The end-LSN is exclusive, while disk_consistent_lsn is
6487 : // inclusive. For example, if disk_consistent_lsn is 100, it is
6488 : // OK for a delta layer to have end LSN 101, but if the end LSN
6489 : // is 102, then it might not have been fully flushed to disk
6490 : // before crash.
6491 : //
6492 : // For example, imagine that the following layers exist:
6493 : //
6494 : // 1000 - image (A)
6495 : // 1000-2000 - delta (B)
6496 : // 2000 - image (C)
6497 : // 2000-3000 - delta (D)
6498 : // 3000 - image (E)
6499 : //
6500 : // If GC horizon is at 2500, we can remove layers A and B, but
6501 : // we cannot remove C, even though it's older than 2500, because
6502 : // the delta layer 2000-3000 depends on it.
6503 48 : if !layers
6504 48 : .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
6505 : {
6506 36 : info!("keeping {} because it is the latest layer", l.layer_name());
6507 36 : result.layers_not_updated += 1;
6508 36 : continue 'outer;
6509 12 : }
6510 12 :
6511 12 : // We didn't find any reason to keep this file, so remove it.
6512 12 : info!(
6513 0 : "garbage collecting {} is_dropped: xx is_incremental: {}",
6514 0 : l.layer_name(),
6515 0 : l.is_incremental(),
6516 : );
6517 12 : layers_to_remove.push(l);
6518 : }
6519 :
6520 24 : if !layers_to_remove.is_empty() {
6521 : // Persist the new GC cutoff value before we actually remove anything.
6522 : // This unconditionally schedules also an index_part.json update, even though, we will
6523 : // be doing one a bit later with the unlinked gc'd layers.
6524 12 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
6525 12 : self.schedule_uploads(disk_consistent_lsn, None)
6526 12 : .map_err(|e| {
6527 0 : if self.cancel.is_cancelled() {
6528 0 : GcError::TimelineCancelled
6529 : } else {
6530 0 : GcError::Remote(e)
6531 : }
6532 12 : })?;
6533 :
6534 12 : let gc_layers = layers_to_remove
6535 12 : .iter()
6536 12 : .map(|x| guard.get_from_desc(x))
6537 12 : .collect::<Vec<Layer>>();
6538 12 :
6539 12 : result.layers_removed = gc_layers.len() as u64;
6540 12 :
6541 12 : self.remote_client.schedule_gc_update(&gc_layers)?;
6542 :
6543 12 : guard.open_mut()?.finish_gc_timeline(&gc_layers);
6544 12 :
6545 12 : #[cfg(feature = "testing")]
6546 12 : {
6547 12 : result.doomed_layers = gc_layers;
6548 12 : }
6549 12 : }
6550 :
6551 24 : info!(
6552 0 : "GC completed removing {} layers, cutoff {}",
6553 : result.layers_removed, new_gc_cutoff
6554 : );
6555 :
6556 24 : result.elapsed = now.elapsed().unwrap_or(Duration::ZERO);
6557 24 : Ok(result)
6558 24 : }
6559 :
6560 : /// Reconstruct a value, using the given base image and WAL records in 'data'.
6561 4408784 : async fn reconstruct_value(
6562 4408784 : &self,
6563 4408784 : key: Key,
6564 4408784 : request_lsn: Lsn,
6565 4408784 : mut data: ValueReconstructState,
6566 4408784 : redo_attempt_type: RedoAttemptType,
6567 4408784 : ) -> Result<Bytes, PageReconstructError> {
6568 4408784 : // Perform WAL redo if needed
6569 4408784 : data.records.reverse();
6570 :
6571 4408784 : let fire_critical_error = match redo_attempt_type {
6572 4392788 : RedoAttemptType::ReadPage => true,
6573 0 : RedoAttemptType::LegacyCompaction => true,
6574 15996 : RedoAttemptType::GcCompaction => false,
6575 : };
6576 :
6577 : // If we have a page image, and no WAL, we're all set
6578 4408784 : if data.records.is_empty() {
6579 4087496 : if let Some((img_lsn, img)) = &data.img {
6580 4087496 : trace!(
6581 0 : "found page image for key {} at {}, no WAL redo required, req LSN {}",
6582 : key, img_lsn, request_lsn,
6583 : );
6584 4087496 : Ok(img.clone())
6585 : } else {
6586 0 : Err(PageReconstructError::from(anyhow!(
6587 0 : "base image for {key} at {request_lsn} not found"
6588 0 : )))
6589 : }
6590 : } else {
6591 : // We need to do WAL redo.
6592 : //
6593 : // If we don't have a base image, then the oldest WAL record better initialize
6594 : // the page
6595 321288 : if data.img.is_none() && !data.records.first().unwrap().1.will_init() {
6596 0 : Err(PageReconstructError::from(anyhow!(
6597 0 : "Base image for {} at {} not found, but got {} WAL records",
6598 0 : key,
6599 0 : request_lsn,
6600 0 : data.records.len()
6601 0 : )))
6602 : } else {
6603 321288 : if data.img.is_some() {
6604 156348 : trace!(
6605 0 : "found {} WAL records and a base image for {} at {}, performing WAL redo",
6606 0 : data.records.len(),
6607 : key,
6608 : request_lsn
6609 : );
6610 : } else {
6611 164940 : trace!(
6612 0 : "found {} WAL records that will init the page for {} at {}, performing WAL redo",
6613 0 : data.records.len(),
6614 : key,
6615 : request_lsn
6616 : );
6617 : };
6618 321288 : let res = self
6619 321288 : .walredo_mgr
6620 321288 : .as_ref()
6621 321288 : .context("timeline has no walredo manager")
6622 321288 : .map_err(PageReconstructError::WalRedo)?
6623 321288 : .request_redo(
6624 321288 : key,
6625 321288 : request_lsn,
6626 321288 : data.img,
6627 321288 : data.records,
6628 321288 : self.pg_version,
6629 321288 : redo_attempt_type,
6630 321288 : )
6631 321288 : .await;
6632 321276 : let img = match res {
6633 321276 : Ok(img) => img,
6634 0 : Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
6635 12 : Err(walredo::Error::Other(err)) => {
6636 12 : if fire_critical_error {
6637 0 : critical!("walredo failure during page reconstruction: {err:?}");
6638 12 : }
6639 12 : return Err(PageReconstructError::WalRedo(
6640 12 : err.context("reconstruct a page image"),
6641 12 : ));
6642 : }
6643 : };
6644 321276 : Ok(img)
6645 : }
6646 : }
6647 4408784 : }
6648 :
6649 0 : pub(crate) async fn spawn_download_all_remote_layers(
6650 0 : self: Arc<Self>,
6651 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6652 0 : ctx: &RequestContext,
6653 0 : ) -> Result<DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskInfo> {
6654 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6655 :
6656 : // this is not really needed anymore; it has tests which really check the return value from
6657 : // http api. it would be better not to maintain this anymore.
6658 :
6659 0 : let mut status_guard = self.download_all_remote_layers_task_info.write().unwrap();
6660 0 : if let Some(st) = &*status_guard {
6661 0 : match &st.state {
6662 : DownloadRemoteLayersTaskState::Running => {
6663 0 : return Err(st.clone());
6664 : }
6665 : DownloadRemoteLayersTaskState::ShutDown
6666 0 : | DownloadRemoteLayersTaskState::Completed => {
6667 0 : *status_guard = None;
6668 0 : }
6669 : }
6670 0 : }
6671 :
6672 0 : let self_clone = Arc::clone(&self);
6673 0 : let task_ctx = ctx.detached_child(
6674 0 : TaskKind::DownloadAllRemoteLayers,
6675 0 : DownloadBehavior::Download,
6676 0 : );
6677 0 : let task_id = task_mgr::spawn(
6678 0 : task_mgr::BACKGROUND_RUNTIME.handle(),
6679 0 : task_mgr::TaskKind::DownloadAllRemoteLayers,
6680 0 : self.tenant_shard_id,
6681 0 : Some(self.timeline_id),
6682 0 : "download all remote layers task",
6683 0 : async move {
6684 0 : self_clone.download_all_remote_layers(request, &task_ctx).await;
6685 0 : let mut status_guard = self_clone.download_all_remote_layers_task_info.write().unwrap();
6686 0 : match &mut *status_guard {
6687 : None => {
6688 0 : warn!("tasks status is supposed to be Some(), since we are running");
6689 : }
6690 0 : Some(st) => {
6691 0 : let exp_task_id = format!("{}", task_mgr::current_task_id().unwrap());
6692 0 : if st.task_id != exp_task_id {
6693 0 : warn!("task id changed while we were still running, expecting {} but have {}", exp_task_id, st.task_id);
6694 0 : } else {
6695 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6696 0 : }
6697 : }
6698 : };
6699 0 : Ok(())
6700 0 : }
6701 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))
6702 : );
6703 :
6704 0 : let initial_info = DownloadRemoteLayersTaskInfo {
6705 0 : task_id: format!("{task_id}"),
6706 0 : state: DownloadRemoteLayersTaskState::Running,
6707 0 : total_layer_count: 0,
6708 0 : successful_download_count: 0,
6709 0 : failed_download_count: 0,
6710 0 : };
6711 0 : *status_guard = Some(initial_info.clone());
6712 0 :
6713 0 : Ok(initial_info)
6714 0 : }
6715 :
6716 0 : async fn download_all_remote_layers(
6717 0 : self: &Arc<Self>,
6718 0 : request: DownloadRemoteLayersTaskSpawnRequest,
6719 0 : ctx: &RequestContext,
6720 0 : ) {
6721 : use pageserver_api::models::DownloadRemoteLayersTaskState;
6722 :
6723 0 : let remaining = {
6724 0 : let guard = self.layers.read().await;
6725 0 : let Ok(lm) = guard.layer_map() else {
6726 : // technically here we could look into iterating accessible layers, but downloading
6727 : // all layers of a shutdown timeline makes no sense regardless.
6728 0 : tracing::info!("attempted to download all layers of shutdown timeline");
6729 0 : return;
6730 : };
6731 0 : lm.iter_historic_layers()
6732 0 : .map(|desc| guard.get_from_desc(&desc))
6733 0 : .collect::<Vec<_>>()
6734 0 : };
6735 0 : let total_layer_count = remaining.len();
6736 :
6737 : macro_rules! lock_status {
6738 : ($st:ident) => {
6739 : let mut st = self.download_all_remote_layers_task_info.write().unwrap();
6740 : let st = st
6741 : .as_mut()
6742 : .expect("this function is only called after the task has been spawned");
6743 : assert_eq!(
6744 : st.task_id,
6745 : format!(
6746 : "{}",
6747 : task_mgr::current_task_id().expect("we run inside a task_mgr task")
6748 : )
6749 : );
6750 : let $st = st;
6751 : };
6752 : }
6753 :
6754 : {
6755 0 : lock_status!(st);
6756 0 : st.total_layer_count = total_layer_count as u64;
6757 0 : }
6758 0 :
6759 0 : let mut remaining = remaining.into_iter();
6760 0 : let mut have_remaining = true;
6761 0 : let mut js = tokio::task::JoinSet::new();
6762 0 :
6763 0 : let cancel = task_mgr::shutdown_token();
6764 0 :
6765 0 : let limit = request.max_concurrent_downloads;
6766 :
6767 : loop {
6768 0 : while js.len() < limit.get() && have_remaining && !cancel.is_cancelled() {
6769 0 : let Some(next) = remaining.next() else {
6770 0 : have_remaining = false;
6771 0 : break;
6772 : };
6773 :
6774 0 : let span = tracing::info_span!("download", layer = %next);
6775 :
6776 0 : let ctx = ctx.attached_child();
6777 0 : js.spawn(
6778 0 : async move {
6779 0 : let res = next.download(&ctx).await;
6780 0 : (next, res)
6781 0 : }
6782 0 : .instrument(span),
6783 0 : );
6784 0 : }
6785 :
6786 0 : while let Some(res) = js.join_next().await {
6787 0 : match res {
6788 : Ok((_, Ok(_))) => {
6789 0 : lock_status!(st);
6790 0 : st.successful_download_count += 1;
6791 : }
6792 0 : Ok((layer, Err(e))) => {
6793 0 : tracing::error!(%layer, "download failed: {e:#}");
6794 0 : lock_status!(st);
6795 0 : st.failed_download_count += 1;
6796 : }
6797 0 : Err(je) if je.is_cancelled() => unreachable!("not used here"),
6798 0 : Err(je) if je.is_panic() => {
6799 0 : lock_status!(st);
6800 0 : st.failed_download_count += 1;
6801 : }
6802 0 : Err(je) => tracing::warn!("unknown joinerror: {je:?}"),
6803 : }
6804 : }
6805 :
6806 0 : if js.is_empty() && (!have_remaining || cancel.is_cancelled()) {
6807 0 : break;
6808 0 : }
6809 : }
6810 :
6811 : {
6812 0 : lock_status!(st);
6813 0 : st.state = DownloadRemoteLayersTaskState::Completed;
6814 : }
6815 0 : }
6816 :
6817 0 : pub(crate) fn get_download_all_remote_layers_task_info(
6818 0 : &self,
6819 0 : ) -> Option<DownloadRemoteLayersTaskInfo> {
6820 0 : self.download_all_remote_layers_task_info
6821 0 : .read()
6822 0 : .unwrap()
6823 0 : .clone()
6824 0 : }
6825 : }
6826 :
6827 : impl Timeline {
6828 : /// Returns non-remote layers for eviction.
6829 0 : pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
6830 0 : let guard = self.layers.read().await;
6831 0 : let mut max_layer_size: Option<u64> = None;
6832 0 :
6833 0 : let resident_layers = guard
6834 0 : .likely_resident_layers()
6835 0 : .map(|layer| {
6836 0 : let file_size = layer.layer_desc().file_size;
6837 0 : max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
6838 0 :
6839 0 : let last_activity_ts = layer.latest_activity();
6840 0 :
6841 0 : EvictionCandidate {
6842 0 : layer: layer.to_owned().into(),
6843 0 : last_activity_ts,
6844 0 : relative_last_activity: finite_f32::FiniteF32::ZERO,
6845 0 : visibility: layer.visibility(),
6846 0 : }
6847 0 : })
6848 0 : .collect();
6849 0 :
6850 0 : DiskUsageEvictionInfo {
6851 0 : max_layer_size,
6852 0 : resident_layers,
6853 0 : }
6854 0 : }
6855 :
6856 11376 : pub(crate) fn get_shard_index(&self) -> ShardIndex {
6857 11376 : ShardIndex {
6858 11376 : shard_number: self.tenant_shard_id.shard_number,
6859 11376 : shard_count: self.tenant_shard_id.shard_count,
6860 11376 : }
6861 11376 : }
6862 :
6863 : /// Persistently blocks gc for `Manual` reason.
6864 : ///
6865 : /// Returns true if no such block existed before, false otherwise.
6866 0 : pub(crate) async fn block_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<bool> {
6867 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6868 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6869 0 : tenant.gc_block.insert(self, GcBlockingReason::Manual).await
6870 0 : }
6871 :
6872 : /// Persistently unblocks gc for `Manual` reason.
6873 0 : pub(crate) async fn unblock_gc(&self, tenant: &super::TenantShard) -> anyhow::Result<()> {
6874 : use crate::tenant::remote_timeline_client::index::GcBlockingReason;
6875 0 : assert_eq!(self.tenant_shard_id, tenant.tenant_shard_id);
6876 0 : tenant.gc_block.remove(self, GcBlockingReason::Manual).await
6877 0 : }
6878 :
6879 : #[cfg(test)]
6880 372 : pub(super) fn force_advance_lsn(self: &Arc<Timeline>, new_lsn: Lsn) {
6881 372 : self.last_record_lsn.advance(new_lsn);
6882 372 : }
6883 :
6884 : #[cfg(test)]
6885 12 : pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
6886 12 : self.disk_consistent_lsn.store(new_value);
6887 12 : }
6888 :
6889 : /// Force create an image layer and place it into the layer map.
6890 : ///
6891 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
6892 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
6893 : /// placed into the layer map in one run AND be validated.
6894 : #[cfg(test)]
6895 432 : pub(super) async fn force_create_image_layer(
6896 432 : self: &Arc<Timeline>,
6897 432 : lsn: Lsn,
6898 432 : mut images: Vec<(Key, Bytes)>,
6899 432 : check_start_lsn: Option<Lsn>,
6900 432 : ctx: &RequestContext,
6901 432 : ) -> anyhow::Result<()> {
6902 432 : let last_record_lsn = self.get_last_record_lsn();
6903 432 : assert!(
6904 432 : lsn <= last_record_lsn,
6905 0 : "advance last record lsn before inserting a layer, lsn={lsn}, last_record_lsn={last_record_lsn}"
6906 : );
6907 432 : if let Some(check_start_lsn) = check_start_lsn {
6908 432 : assert!(lsn >= check_start_lsn);
6909 0 : }
6910 2880 : images.sort_unstable_by(|(ka, _), (kb, _)| ka.cmp(kb));
6911 432 : let min_key = *images.first().map(|(k, _)| k).unwrap();
6912 432 : let end_key = images.last().map(|(k, _)| k).unwrap().next();
6913 432 : let mut image_layer_writer = ImageLayerWriter::new(
6914 432 : self.conf,
6915 432 : self.timeline_id,
6916 432 : self.tenant_shard_id,
6917 432 : &(min_key..end_key),
6918 432 : lsn,
6919 432 : &self.gate,
6920 432 : self.cancel.clone(),
6921 432 : ctx,
6922 432 : )
6923 432 : .await?;
6924 3744 : for (key, img) in images {
6925 3312 : image_layer_writer.put_image(key, img, ctx).await?;
6926 : }
6927 432 : let (desc, path) = image_layer_writer.finish(ctx).await?;
6928 432 : let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6929 432 : info!("force created image layer {}", image_layer.local_path());
6930 : {
6931 432 : let mut guard = self.layers.write().await;
6932 432 : guard
6933 432 : .open_mut()
6934 432 : .unwrap()
6935 432 : .force_insert_layer(image_layer.clone());
6936 432 : }
6937 432 :
6938 432 : // Update remote_timeline_client state to reflect existence of this layer
6939 432 : self.remote_client
6940 432 : .schedule_layer_file_upload(image_layer)
6941 432 : .unwrap();
6942 432 :
6943 432 : Ok(())
6944 432 : }
6945 :
6946 : /// Force create a delta layer and place it into the layer map.
6947 : ///
6948 : /// DO NOT use this function directly. Use [`TenantShard::branch_timeline_test_with_layers`]
6949 : /// or [`TenantShard::create_test_timeline_with_layers`] to ensure all these layers are
6950 : /// placed into the layer map in one run AND be validated.
6951 : #[cfg(test)]
6952 600 : pub(super) async fn force_create_delta_layer(
6953 600 : self: &Arc<Timeline>,
6954 600 : mut deltas: DeltaLayerTestDesc,
6955 600 : check_start_lsn: Option<Lsn>,
6956 600 : ctx: &RequestContext,
6957 600 : ) -> anyhow::Result<()> {
6958 600 : let last_record_lsn = self.get_last_record_lsn();
6959 600 : deltas
6960 600 : .data
6961 1492368 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
6962 600 : assert!(deltas.data.first().unwrap().0 >= deltas.key_range.start);
6963 600 : assert!(deltas.data.last().unwrap().0 < deltas.key_range.end);
6964 125568 : for (_, lsn, _) in &deltas.data {
6965 124968 : assert!(deltas.lsn_range.start <= *lsn && *lsn < deltas.lsn_range.end);
6966 : }
6967 600 : assert!(
6968 600 : deltas.lsn_range.end <= last_record_lsn,
6969 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
6970 : deltas.lsn_range.end,
6971 : last_record_lsn
6972 : );
6973 600 : if let Some(check_start_lsn) = check_start_lsn {
6974 600 : assert!(deltas.lsn_range.start >= check_start_lsn);
6975 0 : }
6976 600 : let mut delta_layer_writer = DeltaLayerWriter::new(
6977 600 : self.conf,
6978 600 : self.timeline_id,
6979 600 : self.tenant_shard_id,
6980 600 : deltas.key_range.start,
6981 600 : deltas.lsn_range,
6982 600 : &self.gate,
6983 600 : self.cancel.clone(),
6984 600 : ctx,
6985 600 : )
6986 600 : .await?;
6987 125568 : for (key, lsn, val) in deltas.data {
6988 124968 : delta_layer_writer.put_value(key, lsn, val, ctx).await?;
6989 : }
6990 600 : let (desc, path) = delta_layer_writer.finish(deltas.key_range.end, ctx).await?;
6991 600 : let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
6992 600 : info!("force created delta layer {}", delta_layer.local_path());
6993 : {
6994 600 : let mut guard = self.layers.write().await;
6995 600 : guard
6996 600 : .open_mut()
6997 600 : .unwrap()
6998 600 : .force_insert_layer(delta_layer.clone());
6999 600 : }
7000 600 :
7001 600 : // Update remote_timeline_client state to reflect existence of this layer
7002 600 : self.remote_client
7003 600 : .schedule_layer_file_upload(delta_layer)
7004 600 : .unwrap();
7005 600 :
7006 600 : Ok(())
7007 600 : }
7008 :
7009 : /// Force create an in-memory layer and place them into the layer map.
7010 : #[cfg(test)]
7011 48 : pub(super) async fn force_create_in_memory_layer(
7012 48 : self: &Arc<Timeline>,
7013 48 : mut in_memory: InMemoryLayerTestDesc,
7014 48 : check_start_lsn: Option<Lsn>,
7015 48 : ctx: &RequestContext,
7016 48 : ) -> anyhow::Result<()> {
7017 : use utils::bin_ser::BeSer;
7018 :
7019 : // Validate LSNs
7020 48 : if let Some(check_start_lsn) = check_start_lsn {
7021 48 : assert!(in_memory.lsn_range.start >= check_start_lsn);
7022 0 : }
7023 :
7024 48 : let last_record_lsn = self.get_last_record_lsn();
7025 48 : let layer_end_lsn = if in_memory.is_open {
7026 12 : in_memory
7027 12 : .data
7028 12 : .iter()
7029 120 : .map(|(_key, lsn, _value)| lsn)
7030 12 : .max()
7031 12 : .cloned()
7032 : } else {
7033 36 : Some(in_memory.lsn_range.end)
7034 : };
7035 :
7036 48 : if let Some(end) = layer_end_lsn {
7037 48 : assert!(
7038 48 : end <= last_record_lsn,
7039 0 : "advance last record lsn before inserting a layer, end_lsn={}, last_record_lsn={}",
7040 : end,
7041 : last_record_lsn,
7042 : );
7043 0 : }
7044 :
7045 237840 : in_memory.data.iter().for_each(|(_key, lsn, _value)| {
7046 237840 : assert!(*lsn >= in_memory.lsn_range.start);
7047 237840 : assert!(*lsn < in_memory.lsn_range.end);
7048 237840 : });
7049 48 :
7050 48 : // Build the batch
7051 48 : in_memory
7052 48 : .data
7053 3280608 : .sort_unstable_by(|(ka, la, _), (kb, lb, _)| (ka, la).cmp(&(kb, lb)));
7054 48 :
7055 48 : let data = in_memory
7056 48 : .data
7057 48 : .into_iter()
7058 237840 : .map(|(key, lsn, value)| {
7059 237840 : let value_size = value.serialized_size().unwrap() as usize;
7060 237840 : (key.to_compact(), lsn, value_size, value)
7061 237840 : })
7062 48 : .collect::<Vec<_>>();
7063 48 :
7064 48 : let batch = SerializedValueBatch::from_values(data);
7065 :
7066 : // Create the in-memory layer and write the batch into it
7067 48 : let layer = InMemoryLayer::create(
7068 48 : self.conf,
7069 48 : self.timeline_id,
7070 48 : self.tenant_shard_id,
7071 48 : in_memory.lsn_range.start,
7072 48 : &self.gate,
7073 48 : // TODO: if we ever use this function in production code, we need to pass the real cancellation token
7074 48 : &CancellationToken::new(),
7075 48 : ctx,
7076 48 : )
7077 48 : .await
7078 48 : .unwrap();
7079 48 :
7080 48 : layer.put_batch(batch, ctx).await.unwrap();
7081 48 : if !in_memory.is_open {
7082 36 : layer.freeze(in_memory.lsn_range.end).await;
7083 12 : }
7084 :
7085 48 : info!("force created in-memory layer {:?}", in_memory.lsn_range);
7086 :
7087 : // Link the layer to the layer map
7088 : {
7089 48 : let mut guard = self.layers.write().await;
7090 48 : let layer_map = guard.open_mut().unwrap();
7091 48 : layer_map.force_insert_in_memory_layer(Arc::new(layer));
7092 48 : }
7093 48 :
7094 48 : Ok(())
7095 48 : }
7096 :
7097 : /// Return all keys at the LSN in the image layers
7098 : #[cfg(test)]
7099 36 : pub(crate) async fn inspect_image_layers(
7100 36 : self: &Arc<Timeline>,
7101 36 : lsn: Lsn,
7102 36 : ctx: &RequestContext,
7103 36 : io_concurrency: IoConcurrency,
7104 36 : ) -> anyhow::Result<Vec<(Key, Bytes)>> {
7105 36 : let mut all_data = Vec::new();
7106 36 : let guard = self.layers.read().await;
7107 204 : for layer in guard.layer_map()?.iter_historic_layers() {
7108 204 : if !layer.is_delta() && layer.image_layer_lsn() == lsn {
7109 48 : let layer = guard.get_from_desc(&layer);
7110 48 : let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
7111 48 : layer
7112 48 : .get_values_reconstruct_data(
7113 48 : KeySpace::single(Key::MIN..Key::MAX),
7114 48 : lsn..Lsn(lsn.0 + 1),
7115 48 : &mut reconstruct_data,
7116 48 : ctx,
7117 48 : )
7118 48 : .await?;
7119 396 : for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
7120 396 : let v = v.collect_pending_ios().await?;
7121 396 : all_data.push((k, v.img.unwrap().1));
7122 : }
7123 156 : }
7124 : }
7125 36 : all_data.sort();
7126 36 : Ok(all_data)
7127 36 : }
7128 :
7129 : /// Get all historic layer descriptors in the layer map
7130 : #[cfg(test)]
7131 144 : pub(crate) async fn inspect_historic_layers(
7132 144 : self: &Arc<Timeline>,
7133 144 : ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
7134 144 : let mut layers = Vec::new();
7135 144 : let guard = self.layers.read().await;
7136 684 : for layer in guard.layer_map()?.iter_historic_layers() {
7137 684 : layers.push(layer.key());
7138 684 : }
7139 144 : Ok(layers)
7140 144 : }
7141 :
7142 : #[cfg(test)]
7143 60 : pub(crate) fn add_extra_test_dense_keyspace(&self, ks: KeySpace) {
7144 60 : let mut keyspace = self.extra_test_dense_keyspace.load().as_ref().clone();
7145 60 : keyspace.merge(&ks);
7146 60 : self.extra_test_dense_keyspace.store(Arc::new(keyspace));
7147 60 : }
7148 : }
7149 :
7150 : /// Tracking writes ingestion does to a particular in-memory layer.
7151 : ///
7152 : /// Cleared upon freezing a layer.
7153 : pub(crate) struct TimelineWriterState {
7154 : open_layer: Arc<InMemoryLayer>,
7155 : current_size: u64,
7156 : // Previous Lsn which passed through
7157 : prev_lsn: Option<Lsn>,
7158 : // Largest Lsn which passed through the current writer
7159 : max_lsn: Option<Lsn>,
7160 : // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
7161 : cached_last_freeze_at: Lsn,
7162 : }
7163 :
7164 : impl TimelineWriterState {
7165 7848 : fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
7166 7848 : Self {
7167 7848 : open_layer,
7168 7848 : current_size,
7169 7848 : prev_lsn: None,
7170 7848 : max_lsn: None,
7171 7848 : cached_last_freeze_at: last_freeze_at,
7172 7848 : }
7173 7848 : }
7174 : }
7175 :
7176 : /// Various functions to mutate the timeline.
7177 : // TODO Currently, Deref is used to allow easy access to read methods from this trait.
7178 : // This is probably considered a bad practice in Rust and should be fixed eventually,
7179 : // but will cause large code changes.
7180 : pub(crate) struct TimelineWriter<'a> {
7181 : tl: &'a Timeline,
7182 : write_guard: tokio::sync::MutexGuard<'a, Option<TimelineWriterState>>,
7183 : }
7184 :
7185 : impl Deref for TimelineWriter<'_> {
7186 : type Target = Timeline;
7187 :
7188 59389992 : fn deref(&self) -> &Self::Target {
7189 59389992 : self.tl
7190 59389992 : }
7191 : }
7192 :
7193 : #[derive(PartialEq)]
7194 : enum OpenLayerAction {
7195 : Roll,
7196 : Open,
7197 : None,
7198 : }
7199 :
7200 : impl TimelineWriter<'_> {
7201 28825464 : async fn handle_open_layer_action(
7202 28825464 : &mut self,
7203 28825464 : at: Lsn,
7204 28825464 : action: OpenLayerAction,
7205 28825464 : ctx: &RequestContext,
7206 28825464 : ) -> anyhow::Result<&Arc<InMemoryLayer>> {
7207 28825464 : match action {
7208 : OpenLayerAction::Roll => {
7209 480 : let freeze_at = self.write_guard.as_ref().unwrap().max_lsn.unwrap();
7210 480 : self.roll_layer(freeze_at).await?;
7211 480 : self.open_layer(at, ctx).await?;
7212 : }
7213 7368 : OpenLayerAction::Open => self.open_layer(at, ctx).await?,
7214 : OpenLayerAction::None => {
7215 28817616 : assert!(self.write_guard.is_some());
7216 : }
7217 : }
7218 :
7219 28825464 : Ok(&self.write_guard.as_ref().unwrap().open_layer)
7220 28825464 : }
7221 :
7222 7848 : async fn open_layer(&mut self, at: Lsn, ctx: &RequestContext) -> anyhow::Result<()> {
7223 7848 : let layer = self
7224 7848 : .tl
7225 7848 : .get_layer_for_write(at, &self.write_guard, ctx)
7226 7848 : .await?;
7227 7848 : let initial_size = layer.size().await?;
7228 :
7229 7848 : let last_freeze_at = self.last_freeze_at.load();
7230 7848 : self.write_guard.replace(TimelineWriterState::new(
7231 7848 : layer,
7232 7848 : initial_size,
7233 7848 : last_freeze_at,
7234 7848 : ));
7235 7848 :
7236 7848 : Ok(())
7237 7848 : }
7238 :
7239 480 : async fn roll_layer(&mut self, freeze_at: Lsn) -> Result<(), FlushLayerError> {
7240 480 : let current_size = self.write_guard.as_ref().unwrap().current_size;
7241 :
7242 : // If layer flushes are backpressured due to compaction not keeping up, wait for the flush
7243 : // to propagate the backpressure up into WAL ingestion.
7244 480 : let l0_count = self
7245 480 : .tl
7246 480 : .layers
7247 480 : .read()
7248 480 : .await
7249 480 : .layer_map()?
7250 480 : .level0_deltas()
7251 480 : .len();
7252 480 : let wait_thresholds = [
7253 480 : self.get_l0_flush_delay_threshold(),
7254 480 : self.get_l0_flush_stall_threshold(),
7255 480 : ];
7256 480 : let wait_threshold = wait_thresholds.into_iter().flatten().min();
7257 :
7258 : // self.write_guard will be taken by the freezing
7259 480 : let flush_id = self
7260 480 : .tl
7261 480 : .freeze_inmem_layer_at(freeze_at, &mut self.write_guard)
7262 480 : .await?;
7263 :
7264 480 : assert!(self.write_guard.is_none());
7265 :
7266 480 : if let Some(wait_threshold) = wait_threshold {
7267 0 : if l0_count >= wait_threshold {
7268 0 : debug!(
7269 0 : "layer roll waiting for flush due to compaction backpressure at {l0_count} L0 layers"
7270 : );
7271 0 : self.tl.wait_flush_completion(flush_id).await?;
7272 0 : }
7273 480 : }
7274 :
7275 480 : if current_size >= self.get_checkpoint_distance() * 2 {
7276 0 : warn!("Flushed oversized open layer with size {}", current_size)
7277 480 : }
7278 :
7279 480 : Ok(())
7280 480 : }
7281 :
7282 28825464 : fn get_open_layer_action(&self, lsn: Lsn, new_value_size: u64) -> OpenLayerAction {
7283 28825464 : let state = &*self.write_guard;
7284 28825464 : let Some(state) = &state else {
7285 7368 : return OpenLayerAction::Open;
7286 : };
7287 :
7288 : #[cfg(feature = "testing")]
7289 28818096 : if state.cached_last_freeze_at < self.tl.last_freeze_at.load() {
7290 : // this check and assertion are not really needed because
7291 : // LayerManager::try_freeze_in_memory_layer will always clear out the
7292 : // TimelineWriterState if something is frozen. however, we can advance last_freeze_at when there
7293 : // is no TimelineWriterState.
7294 0 : assert!(
7295 0 : state.open_layer.end_lsn.get().is_some(),
7296 0 : "our open_layer must be outdated"
7297 : );
7298 :
7299 : // this would be a memory leak waiting to happen because the in-memory layer always has
7300 : // an index
7301 0 : panic!("BUG: TimelineWriterState held on to frozen in-memory layer.");
7302 28818096 : }
7303 28818096 :
7304 28818096 : if state.prev_lsn == Some(lsn) {
7305 : // Rolling mid LSN is not supported by [downstream code].
7306 : // Hence, only roll at LSN boundaries.
7307 : //
7308 : // [downstream code]: https://github.com/neondatabase/neon/pull/7993#discussion_r1633345422
7309 36 : return OpenLayerAction::None;
7310 28818060 : }
7311 28818060 :
7312 28818060 : if state.current_size == 0 {
7313 : // Don't roll empty layers
7314 0 : return OpenLayerAction::None;
7315 28818060 : }
7316 28818060 :
7317 28818060 : if self.tl.should_roll(
7318 28818060 : state.current_size,
7319 28818060 : state.current_size + new_value_size,
7320 28818060 : self.get_checkpoint_distance(),
7321 28818060 : lsn,
7322 28818060 : state.cached_last_freeze_at,
7323 28818060 : state.open_layer.get_opened_at(),
7324 28818060 : ) {
7325 480 : OpenLayerAction::Roll
7326 : } else {
7327 28817580 : OpenLayerAction::None
7328 : }
7329 28825464 : }
7330 :
7331 : /// Put a batch of keys at the specified Lsns.
7332 28825452 : pub(crate) async fn put_batch(
7333 28825452 : &mut self,
7334 28825452 : batch: SerializedValueBatch,
7335 28825452 : ctx: &RequestContext,
7336 28825452 : ) -> anyhow::Result<()> {
7337 28825452 : if !batch.has_data() {
7338 0 : return Ok(());
7339 28825452 : }
7340 28825452 :
7341 28825452 : // In debug builds, assert that we don't write any keys that don't belong to this shard.
7342 28825452 : // We don't assert this in release builds, since key ownership policies may change over
7343 28825452 : // time. Stray keys will be removed during compaction.
7344 28825452 : if cfg!(debug_assertions) {
7345 59369952 : for metadata in &batch.metadata {
7346 30544500 : if let ValueMeta::Serialized(metadata) = metadata {
7347 30544500 : let key = Key::from_compact(metadata.key);
7348 30544500 : assert!(
7349 30544500 : self.shard_identity.is_key_local(&key)
7350 0 : || self.shard_identity.is_key_global(&key),
7351 0 : "key {key} does not belong on shard {}",
7352 0 : self.shard_identity.shard_index()
7353 : );
7354 0 : }
7355 : }
7356 0 : }
7357 :
7358 28825452 : let batch_max_lsn = batch.max_lsn;
7359 28825452 : let buf_size: u64 = batch.buffer_size() as u64;
7360 28825452 :
7361 28825452 : let action = self.get_open_layer_action(batch_max_lsn, buf_size);
7362 28825452 : let layer = self
7363 28825452 : .handle_open_layer_action(batch_max_lsn, action, ctx)
7364 28825452 : .await?;
7365 :
7366 28825452 : let res = layer.put_batch(batch, ctx).await;
7367 :
7368 28825452 : if res.is_ok() {
7369 28825452 : // Update the current size only when the entire write was ok.
7370 28825452 : // In case of failures, we may have had partial writes which
7371 28825452 : // render the size tracking out of sync. That's ok because
7372 28825452 : // the checkpoint distance should be significantly smaller
7373 28825452 : // than the S3 single shot upload limit of 5GiB.
7374 28825452 : let state = self.write_guard.as_mut().unwrap();
7375 28825452 :
7376 28825452 : state.current_size += buf_size;
7377 28825452 : state.prev_lsn = Some(batch_max_lsn);
7378 28825452 : state.max_lsn = std::cmp::max(state.max_lsn, Some(batch_max_lsn));
7379 28825452 : }
7380 :
7381 28825452 : res
7382 28825452 : }
7383 :
7384 : #[cfg(test)]
7385 : /// Test helper, for tests that would like to poke individual values without composing a batch
7386 26340924 : pub(crate) async fn put(
7387 26340924 : &mut self,
7388 26340924 : key: Key,
7389 26340924 : lsn: Lsn,
7390 26340924 : value: &Value,
7391 26340924 : ctx: &RequestContext,
7392 26340924 : ) -> anyhow::Result<()> {
7393 : use utils::bin_ser::BeSer;
7394 26340924 : if !key.is_valid_key_on_write_path() {
7395 0 : bail!(
7396 0 : "the request contains data not supported by pageserver at TimelineWriter::put: {}",
7397 0 : key
7398 0 : );
7399 26340924 : }
7400 26340924 : let val_ser_size = value.serialized_size().unwrap() as usize;
7401 26340924 : let batch = SerializedValueBatch::from_values(vec![(
7402 26340924 : key.to_compact(),
7403 26340924 : lsn,
7404 26340924 : val_ser_size,
7405 26340924 : value.clone(),
7406 26340924 : )]);
7407 26340924 :
7408 26340924 : self.put_batch(batch, ctx).await
7409 26340924 : }
7410 :
7411 12 : pub(crate) async fn delete_batch(
7412 12 : &mut self,
7413 12 : batch: &[(Range<Key>, Lsn)],
7414 12 : ctx: &RequestContext,
7415 12 : ) -> anyhow::Result<()> {
7416 12 : if let Some((_, lsn)) = batch.first() {
7417 12 : let action = self.get_open_layer_action(*lsn, 0);
7418 12 : let layer = self.handle_open_layer_action(*lsn, action, ctx).await?;
7419 12 : layer.put_tombstones(batch).await?;
7420 0 : }
7421 :
7422 12 : Ok(())
7423 12 : }
7424 :
7425 : /// Track the end of the latest digested WAL record.
7426 : /// Remember the (end of) last valid WAL record remembered in the timeline.
7427 : ///
7428 : /// Call this after you have finished writing all the WAL up to 'lsn'.
7429 : ///
7430 : /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
7431 : /// the 'lsn' or anything older. The previous last record LSN is stored alongside
7432 : /// the latest and can be read.
7433 31674624 : pub(crate) fn finish_write(&self, new_lsn: Lsn) {
7434 31674624 : self.tl.finish_write(new_lsn);
7435 31674624 : }
7436 :
7437 1623420 : pub(crate) fn update_current_logical_size(&self, delta: i64) {
7438 1623420 : self.tl.update_current_logical_size(delta)
7439 1623420 : }
7440 : }
7441 :
7442 : // We need TimelineWriter to be send in upcoming conversion of
7443 : // Timeline::layers to tokio::sync::RwLock.
7444 : #[test]
7445 12 : fn is_send() {
7446 12 : fn _assert_send<T: Send>() {}
7447 12 : _assert_send::<TimelineWriter<'_>>();
7448 12 : }
7449 :
7450 : #[cfg(test)]
7451 : mod tests {
7452 : use std::sync::Arc;
7453 :
7454 : use pageserver_api::key::Key;
7455 : use pageserver_api::value::Value;
7456 : use std::iter::Iterator;
7457 : use tracing::Instrument;
7458 : use utils::id::TimelineId;
7459 : use utils::lsn::Lsn;
7460 :
7461 : use super::HeatMapTimeline;
7462 : use crate::context::RequestContextBuilder;
7463 : use crate::tenant::harness::{TenantHarness, test_img};
7464 : use crate::tenant::layer_map::LayerMap;
7465 : use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
7466 : use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
7467 : use crate::tenant::{PreviousHeatmap, Timeline};
7468 :
7469 60 : fn assert_heatmaps_have_same_layers(lhs: &HeatMapTimeline, rhs: &HeatMapTimeline) {
7470 60 : assert_eq!(lhs.all_layers().count(), rhs.all_layers().count());
7471 60 : let lhs_rhs = lhs.all_layers().zip(rhs.all_layers());
7472 300 : for (l, r) in lhs_rhs {
7473 240 : assert_eq!(l.name, r.name);
7474 240 : assert_eq!(l.metadata, r.metadata);
7475 : }
7476 60 : }
7477 :
7478 : #[tokio::test]
7479 12 : async fn test_heatmap_generation() {
7480 12 : let harness = TenantHarness::create("heatmap_generation").await.unwrap();
7481 12 :
7482 12 : let covered_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7483 12 : Lsn(0x10)..Lsn(0x20),
7484 12 : vec![(
7485 12 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7486 12 : Lsn(0x11),
7487 12 : Value::Image(test_img("foo")),
7488 12 : )],
7489 12 : );
7490 12 : let visible_delta = DeltaLayerTestDesc::new_with_inferred_key_range(
7491 12 : Lsn(0x10)..Lsn(0x20),
7492 12 : vec![(
7493 12 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7494 12 : Lsn(0x11),
7495 12 : Value::Image(test_img("foo")),
7496 12 : )],
7497 12 : );
7498 12 : let l0_delta = DeltaLayerTestDesc::new(
7499 12 : Lsn(0x20)..Lsn(0x30),
7500 12 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7501 12 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7502 12 : vec![(
7503 12 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7504 12 : Lsn(0x25),
7505 12 : Value::Image(test_img("foo")),
7506 12 : )],
7507 12 : );
7508 12 : let delta_layers = vec![
7509 12 : covered_delta.clone(),
7510 12 : visible_delta.clone(),
7511 12 : l0_delta.clone(),
7512 12 : ];
7513 12 :
7514 12 : let image_layer = (
7515 12 : Lsn(0x40),
7516 12 : vec![(
7517 12 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7518 12 : test_img("bar"),
7519 12 : )],
7520 12 : );
7521 12 : let image_layers = vec![image_layer];
7522 12 :
7523 12 : let (tenant, ctx) = harness.load().await;
7524 12 : let timeline = tenant
7525 12 : .create_test_timeline_with_layers(
7526 12 : TimelineId::generate(),
7527 12 : Lsn(0x10),
7528 12 : 14,
7529 12 : &ctx,
7530 12 : Vec::new(), // in-memory layers
7531 12 : delta_layers,
7532 12 : image_layers,
7533 12 : Lsn(0x100),
7534 12 : )
7535 12 : .await
7536 12 : .unwrap();
7537 12 : let ctx = &ctx.with_scope_timeline(&timeline);
7538 12 :
7539 12 : // Layer visibility is an input to heatmap generation, so refresh it first
7540 12 : timeline.update_layer_visibility().await.unwrap();
7541 12 :
7542 12 : let heatmap = timeline
7543 12 : .generate_heatmap()
7544 12 : .await
7545 12 : .expect("Infallible while timeline is not shut down");
7546 12 :
7547 12 : assert_eq!(heatmap.timeline_id, timeline.timeline_id);
7548 12 :
7549 12 : // L0 should come last
7550 12 : let heatmap_layers = heatmap.all_layers().collect::<Vec<_>>();
7551 12 : assert_eq!(heatmap_layers.last().unwrap().name, l0_delta.layer_name());
7552 12 :
7553 12 : let mut last_lsn = Lsn::MAX;
7554 60 : for layer in heatmap_layers {
7555 12 : // Covered layer should be omitted
7556 48 : assert!(layer.name != covered_delta.layer_name());
7557 12 :
7558 48 : let layer_lsn = match &layer.name {
7559 24 : LayerName::Delta(d) => d.lsn_range.end,
7560 24 : LayerName::Image(i) => i.lsn,
7561 12 : };
7562 12 :
7563 12 : // Apart from L0s, newest Layers should come first
7564 48 : if !LayerMap::is_l0(layer.name.key_range(), layer.name.is_delta()) {
7565 36 : assert!(layer_lsn <= last_lsn);
7566 36 : last_lsn = layer_lsn;
7567 12 : }
7568 12 : }
7569 12 :
7570 12 : // Evict all the layers and stash the old heatmap in the timeline.
7571 12 : // This simulates a migration to a cold secondary location.
7572 12 :
7573 12 : let guard = timeline.layers.read().await;
7574 12 : let mut all_layers = Vec::new();
7575 12 : let forever = std::time::Duration::from_secs(120);
7576 60 : for layer in guard.likely_resident_layers() {
7577 60 : all_layers.push(layer.clone());
7578 60 : layer.evict_and_wait(forever).await.unwrap();
7579 12 : }
7580 12 : drop(guard);
7581 12 :
7582 12 : timeline
7583 12 : .previous_heatmap
7584 12 : .store(Some(Arc::new(PreviousHeatmap::Active {
7585 12 : heatmap: heatmap.clone(),
7586 12 : read_at: std::time::Instant::now(),
7587 12 : end_lsn: None,
7588 12 : })));
7589 12 :
7590 12 : // Generate a new heatmap and assert that it contains the same layers as the old one.
7591 12 : let post_migration_heatmap = timeline.generate_heatmap().await.unwrap();
7592 12 : assert_heatmaps_have_same_layers(&heatmap, &post_migration_heatmap);
7593 12 :
7594 12 : // Download each layer one by one. Generate the heatmap at each step and check
7595 12 : // that it's stable.
7596 72 : for layer in all_layers {
7597 60 : if layer.visibility() == LayerVisibilityHint::Covered {
7598 12 : continue;
7599 48 : }
7600 48 :
7601 48 : eprintln!("Downloading {layer} and re-generating heatmap");
7602 48 :
7603 48 : let ctx = &RequestContextBuilder::from(ctx)
7604 48 : .download_behavior(crate::context::DownloadBehavior::Download)
7605 48 : .attached_child();
7606 12 :
7607 48 : let _resident = layer
7608 48 : .download_and_keep_resident(ctx)
7609 48 : .instrument(tracing::info_span!(
7610 48 : parent: None,
7611 12 : "download_layer",
7612 12 : tenant_id = %timeline.tenant_shard_id.tenant_id,
7613 0 : shard_id = %timeline.tenant_shard_id.shard_slug(),
7614 0 : timeline_id = %timeline.timeline_id
7615 12 : ))
7616 48 : .await
7617 48 : .unwrap();
7618 12 :
7619 48 : let post_download_heatmap = timeline.generate_heatmap().await.unwrap();
7620 48 : assert_heatmaps_have_same_layers(&heatmap, &post_download_heatmap);
7621 12 : }
7622 12 :
7623 12 : // Everything from the post-migration heatmap is now resident.
7624 12 : // Check that we drop it from memory.
7625 12 : assert!(matches!(
7626 12 : timeline.previous_heatmap.load().as_deref(),
7627 12 : Some(PreviousHeatmap::Obsolete)
7628 12 : ));
7629 12 : }
7630 :
7631 : #[tokio::test]
7632 12 : async fn test_previous_heatmap_obsoletion() {
7633 12 : let harness = TenantHarness::create("heatmap_previous_heatmap_obsoletion")
7634 12 : .await
7635 12 : .unwrap();
7636 12 :
7637 12 : let l0_delta = DeltaLayerTestDesc::new(
7638 12 : Lsn(0x20)..Lsn(0x30),
7639 12 : Key::from_hex("000000000000000000000000000000000000").unwrap()
7640 12 : ..Key::from_hex("FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF").unwrap(),
7641 12 : vec![(
7642 12 : Key::from_hex("720000000033333333444444445500000000").unwrap(),
7643 12 : Lsn(0x25),
7644 12 : Value::Image(test_img("foo")),
7645 12 : )],
7646 12 : );
7647 12 :
7648 12 : let image_layer = (
7649 12 : Lsn(0x40),
7650 12 : vec![(
7651 12 : Key::from_hex("620000000033333333444444445500000000").unwrap(),
7652 12 : test_img("bar"),
7653 12 : )],
7654 12 : );
7655 12 :
7656 12 : let delta_layers = vec![l0_delta];
7657 12 : let image_layers = vec![image_layer];
7658 12 :
7659 12 : let (tenant, ctx) = harness.load().await;
7660 12 : let timeline = tenant
7661 12 : .create_test_timeline_with_layers(
7662 12 : TimelineId::generate(),
7663 12 : Lsn(0x10),
7664 12 : 14,
7665 12 : &ctx,
7666 12 : Vec::new(), // in-memory layers
7667 12 : delta_layers,
7668 12 : image_layers,
7669 12 : Lsn(0x100),
7670 12 : )
7671 12 : .await
7672 12 : .unwrap();
7673 12 :
7674 12 : // Layer visibility is an input to heatmap generation, so refresh it first
7675 12 : timeline.update_layer_visibility().await.unwrap();
7676 12 :
7677 12 : let heatmap = timeline
7678 12 : .generate_heatmap()
7679 12 : .await
7680 12 : .expect("Infallible while timeline is not shut down");
7681 12 :
7682 12 : // Both layers should be in the heatmap
7683 12 : assert!(heatmap.all_layers().count() > 0);
7684 12 :
7685 12 : // Now simulate a migration.
7686 12 : timeline
7687 12 : .previous_heatmap
7688 12 : .store(Some(Arc::new(PreviousHeatmap::Active {
7689 12 : heatmap: heatmap.clone(),
7690 12 : read_at: std::time::Instant::now(),
7691 12 : end_lsn: None,
7692 12 : })));
7693 12 :
7694 12 : // Evict all the layers in the previous heatmap
7695 12 : let guard = timeline.layers.read().await;
7696 12 : let forever = std::time::Duration::from_secs(120);
7697 36 : for layer in guard.likely_resident_layers() {
7698 36 : layer.evict_and_wait(forever).await.unwrap();
7699 12 : }
7700 12 : drop(guard);
7701 12 :
7702 12 : // Generate a new heatmap and check that the previous heatmap
7703 12 : // has been marked obsolete.
7704 12 : let post_eviction_heatmap = timeline
7705 12 : .generate_heatmap()
7706 12 : .await
7707 12 : .expect("Infallible while timeline is not shut down");
7708 12 :
7709 12 : assert_eq!(post_eviction_heatmap.all_layers().count(), 0);
7710 12 : assert!(matches!(
7711 12 : timeline.previous_heatmap.load().as_deref(),
7712 12 : Some(PreviousHeatmap::Obsolete)
7713 12 : ));
7714 12 : }
7715 :
7716 : #[tokio::test]
7717 12 : async fn two_layer_eviction_attempts_at_the_same_time() {
7718 12 : let harness = TenantHarness::create("two_layer_eviction_attempts_at_the_same_time")
7719 12 : .await
7720 12 : .unwrap();
7721 12 :
7722 12 : let (tenant, ctx) = harness.load().await;
7723 12 : let timeline = tenant
7724 12 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
7725 12 : .await
7726 12 : .unwrap();
7727 12 :
7728 12 : let layer = find_some_layer(&timeline).await;
7729 12 : let layer = layer
7730 12 : .keep_resident()
7731 12 : .await
7732 12 : .expect("no download => no downloading errors")
7733 12 : .drop_eviction_guard();
7734 12 :
7735 12 : let forever = std::time::Duration::from_secs(120);
7736 12 :
7737 12 : let first = layer.evict_and_wait(forever);
7738 12 : let second = layer.evict_and_wait(forever);
7739 12 :
7740 12 : let (first, second) = tokio::join!(first, second);
7741 12 :
7742 12 : let res = layer.keep_resident().await;
7743 12 : assert!(res.is_none(), "{res:?}");
7744 12 :
7745 12 : match (first, second) {
7746 12 : (Ok(()), Ok(())) => {
7747 12 : // because there are no more timeline locks being taken on eviction path, we can
7748 12 : // witness all three outcomes here.
7749 12 : }
7750 12 : (Ok(()), Err(EvictionError::NotFound)) | (Err(EvictionError::NotFound), Ok(())) => {
7751 0 : // if one completes before the other, this is fine just as well.
7752 0 : }
7753 12 : other => unreachable!("unexpected {:?}", other),
7754 12 : }
7755 12 : }
7756 :
7757 12 : async fn find_some_layer(timeline: &Timeline) -> Layer {
7758 12 : let layers = timeline.layers.read().await;
7759 12 : let desc = layers
7760 12 : .layer_map()
7761 12 : .unwrap()
7762 12 : .iter_historic_layers()
7763 12 : .next()
7764 12 : .expect("must find one layer to evict");
7765 12 :
7766 12 : layers.get_from_desc(&desc)
7767 12 : }
7768 : }
|
