Line data Source code
1 : //! New compaction implementation. The algorithm itself is implemented in the
2 : //! compaction crate. This file implements the callbacks and structs that allow
3 : //! the algorithm to drive the process.
4 : //!
5 : //! The old legacy algorithm is implemented directly in `timeline.rs`.
6 :
7 : use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
8 : use std::ops::{Deref, Range};
9 : use std::sync::Arc;
10 :
11 : use super::layer_manager::LayerManager;
12 : use super::{
13 : CompactFlags, CompactOptions, CreateImageLayersError, DurationRecorder, GetVectoredError,
14 : ImageLayerCreationMode, LastImageLayerCreationStatus, RecordedDuration, Timeline,
15 : };
16 :
17 : use anyhow::{anyhow, bail, Context};
18 : use bytes::Bytes;
19 : use enumset::EnumSet;
20 : use fail::fail_point;
21 : use itertools::Itertools;
22 : use pageserver_api::key::KEY_SIZE;
23 : use pageserver_api::keyspace::ShardedRange;
24 : use pageserver_api::models::CompactInfoResponse;
25 : use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
26 : use serde::Serialize;
27 : use tokio_util::sync::CancellationToken;
28 : use tracing::{debug, info, info_span, trace, warn, Instrument};
29 : use utils::critical;
30 : use utils::id::TimelineId;
31 :
32 : use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
33 : use crate::page_cache;
34 : use crate::statvfs::Statvfs;
35 : use crate::tenant::checks::check_valid_layermap;
36 : use crate::tenant::gc_block::GcBlock;
37 : use crate::tenant::layer_map::LayerMap;
38 : use crate::tenant::remote_timeline_client::WaitCompletionError;
39 : use crate::tenant::storage_layer::batch_split_writer::{
40 : BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
41 : };
42 : use crate::tenant::storage_layer::filter_iterator::FilterIterator;
43 : use crate::tenant::storage_layer::merge_iterator::MergeIterator;
44 : use crate::tenant::storage_layer::{
45 : AsLayerDesc, PersistentLayerDesc, PersistentLayerKey, ValueReconstructState,
46 : };
47 : use crate::tenant::timeline::{drop_rlock, DeltaLayerWriter, ImageLayerWriter};
48 : use crate::tenant::timeline::{ImageLayerCreationOutcome, IoConcurrency};
49 : use crate::tenant::timeline::{Layer, ResidentLayer};
50 : use crate::tenant::{gc_block, DeltaLayer, MaybeOffloaded};
51 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
52 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_CHECKPOINT_DISTANCE;
53 :
54 : use pageserver_api::key::Key;
55 : use pageserver_api::keyspace::KeySpace;
56 : use pageserver_api::record::NeonWalRecord;
57 : use pageserver_api::value::Value;
58 :
59 : use utils::lsn::Lsn;
60 :
61 : use pageserver_compaction::helpers::{fully_contains, overlaps_with};
62 : use pageserver_compaction::interface::*;
63 :
64 : use super::CompactionError;
65 :
66 : /// Maximum number of deltas before generating an image layer in bottom-most compaction.
67 : const COMPACTION_DELTA_THRESHOLD: usize = 5;
68 :
69 : #[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
70 : pub struct GcCompactionJobId(pub usize);
71 :
72 : impl std::fmt::Display for GcCompactionJobId {
73 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
74 0 : write!(f, "{}", self.0)
75 0 : }
76 : }
77 :
78 : #[derive(Debug, Clone)]
79 : pub enum GcCompactionQueueItem {
80 : Manual(CompactOptions),
81 : SubCompactionJob(CompactOptions),
82 : #[allow(dead_code)]
83 : UpdateL2Lsn(Lsn),
84 : Notify(GcCompactionJobId),
85 : }
86 :
87 : impl GcCompactionQueueItem {
88 0 : pub fn into_compact_info_resp(
89 0 : self,
90 0 : id: GcCompactionJobId,
91 0 : running: bool,
92 0 : ) -> Option<CompactInfoResponse> {
93 0 : match self {
94 0 : GcCompactionQueueItem::Manual(options) => Some(CompactInfoResponse {
95 0 : compact_key_range: options.compact_key_range,
96 0 : compact_lsn_range: options.compact_lsn_range,
97 0 : sub_compaction: options.sub_compaction,
98 0 : running,
99 0 : job_id: id.0,
100 0 : }),
101 0 : GcCompactionQueueItem::SubCompactionJob(options) => Some(CompactInfoResponse {
102 0 : compact_key_range: options.compact_key_range,
103 0 : compact_lsn_range: options.compact_lsn_range,
104 0 : sub_compaction: options.sub_compaction,
105 0 : running,
106 0 : job_id: id.0,
107 0 : }),
108 0 : GcCompactionQueueItem::UpdateL2Lsn(_) => None,
109 0 : GcCompactionQueueItem::Notify(_) => None,
110 : }
111 0 : }
112 : }
113 :
114 : struct GcCompactionQueueInner {
115 : running: Option<(GcCompactionJobId, GcCompactionQueueItem)>,
116 : queued: VecDeque<(GcCompactionJobId, GcCompactionQueueItem)>,
117 : notify: HashMap<GcCompactionJobId, tokio::sync::oneshot::Sender<()>>,
118 : gc_guards: HashMap<GcCompactionJobId, gc_block::Guard>,
119 : last_id: GcCompactionJobId,
120 : }
121 :
122 : impl GcCompactionQueueInner {
123 0 : fn next_id(&mut self) -> GcCompactionJobId {
124 0 : let id = self.last_id;
125 0 : self.last_id = GcCompactionJobId(id.0 + 1);
126 0 : id
127 0 : }
128 : }
129 :
130 : /// A structure to store gc_compaction jobs.
131 : pub struct GcCompactionQueue {
132 : /// All items in the queue, and the currently-running job.
133 : inner: std::sync::Mutex<GcCompactionQueueInner>,
134 : /// Ensure only one thread is consuming the queue.
135 : consumer_lock: tokio::sync::Mutex<()>,
136 : }
137 :
138 : impl GcCompactionQueue {
139 0 : pub fn new() -> Self {
140 0 : GcCompactionQueue {
141 0 : inner: std::sync::Mutex::new(GcCompactionQueueInner {
142 0 : running: None,
143 0 : queued: VecDeque::new(),
144 0 : notify: HashMap::new(),
145 0 : gc_guards: HashMap::new(),
146 0 : last_id: GcCompactionJobId(0),
147 0 : }),
148 0 : consumer_lock: tokio::sync::Mutex::new(()),
149 0 : }
150 0 : }
151 :
152 0 : pub fn cancel_scheduled(&self) {
153 0 : let mut guard = self.inner.lock().unwrap();
154 0 : guard.queued.clear();
155 0 : guard.notify.clear();
156 0 : guard.gc_guards.clear();
157 0 : }
158 :
159 : /// Schedule a manual compaction job.
160 0 : pub fn schedule_manual_compaction(
161 0 : &self,
162 0 : options: CompactOptions,
163 0 : notify: Option<tokio::sync::oneshot::Sender<()>>,
164 0 : ) -> GcCompactionJobId {
165 0 : let mut guard = self.inner.lock().unwrap();
166 0 : let id = guard.next_id();
167 0 : guard
168 0 : .queued
169 0 : .push_back((id, GcCompactionQueueItem::Manual(options)));
170 0 : if let Some(notify) = notify {
171 0 : guard.notify.insert(id, notify);
172 0 : }
173 0 : info!("scheduled compaction job id={}", id);
174 0 : id
175 0 : }
176 :
177 : /// Trigger an auto compaction.
178 : #[allow(dead_code)]
179 0 : pub fn trigger_auto_compaction(&self, _: &Arc<Timeline>) {}
180 :
181 : /// Notify the caller the job has finished and unblock GC.
182 0 : fn notify_and_unblock(&self, id: GcCompactionJobId) {
183 0 : info!("compaction job id={} finished", id);
184 0 : let mut guard = self.inner.lock().unwrap();
185 0 : if let Some(blocking) = guard.gc_guards.remove(&id) {
186 0 : drop(blocking)
187 0 : }
188 0 : if let Some(tx) = guard.notify.remove(&id) {
189 0 : let _ = tx.send(());
190 0 : }
191 0 : }
192 :
193 0 : async fn handle_sub_compaction(
194 0 : &self,
195 0 : id: GcCompactionJobId,
196 0 : options: CompactOptions,
197 0 : timeline: &Arc<Timeline>,
198 0 : gc_block: &GcBlock,
199 0 : ) -> Result<(), CompactionError> {
200 0 : info!("running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
201 0 : let jobs: Vec<GcCompactJob> = timeline
202 0 : .gc_compaction_split_jobs(
203 0 : GcCompactJob::from_compact_options(options.clone()),
204 0 : options.sub_compaction_max_job_size_mb,
205 0 : )
206 0 : .await
207 0 : .map_err(CompactionError::Other)?;
208 0 : if jobs.is_empty() {
209 0 : info!("no jobs to run, skipping scheduled compaction task");
210 0 : self.notify_and_unblock(id);
211 : } else {
212 0 : let gc_guard = match gc_block.start().await {
213 0 : Ok(guard) => guard,
214 0 : Err(e) => {
215 0 : return Err(CompactionError::Other(anyhow!(
216 0 : "cannot run gc-compaction because gc is blocked: {}",
217 0 : e
218 0 : )));
219 : }
220 : };
221 :
222 0 : let jobs_len = jobs.len();
223 0 : let mut pending_tasks = Vec::new();
224 0 : for job in jobs {
225 : // Unfortunately we need to convert the `GcCompactJob` back to `CompactionOptions`
226 : // until we do further refactors to allow directly call `compact_with_gc`.
227 0 : let mut flags: EnumSet<CompactFlags> = EnumSet::default();
228 0 : flags |= CompactFlags::EnhancedGcBottomMostCompaction;
229 0 : if job.dry_run {
230 0 : flags |= CompactFlags::DryRun;
231 0 : }
232 0 : let options = CompactOptions {
233 0 : flags,
234 0 : sub_compaction: false,
235 0 : compact_key_range: Some(job.compact_key_range.into()),
236 0 : compact_lsn_range: Some(job.compact_lsn_range.into()),
237 0 : sub_compaction_max_job_size_mb: None,
238 0 : };
239 0 : pending_tasks.push(GcCompactionQueueItem::SubCompactionJob(options));
240 : }
241 0 : pending_tasks.push(GcCompactionQueueItem::Notify(id));
242 0 : {
243 0 : let mut guard = self.inner.lock().unwrap();
244 0 : guard.gc_guards.insert(id, gc_guard);
245 0 : let mut tasks = Vec::new();
246 0 : for task in pending_tasks {
247 0 : let id = guard.next_id();
248 0 : tasks.push((id, task));
249 0 : }
250 0 : tasks.reverse();
251 0 : for item in tasks {
252 0 : guard.queued.push_front(item);
253 0 : }
254 : }
255 0 : info!("scheduled enhanced gc bottom-most compaction with sub-compaction, split into {} jobs", jobs_len);
256 : }
257 0 : Ok(())
258 0 : }
259 :
260 : /// Take a job from the queue and process it. Returns if there are still pending tasks.
261 0 : pub async fn iteration(
262 0 : &self,
263 0 : cancel: &CancellationToken,
264 0 : ctx: &RequestContext,
265 0 : gc_block: &GcBlock,
266 0 : timeline: &Arc<Timeline>,
267 0 : ) -> Result<CompactionOutcome, CompactionError> {
268 0 : let _one_op_at_a_time_guard = self.consumer_lock.lock().await;
269 : let has_pending_tasks;
270 0 : let (id, item) = {
271 0 : let mut guard = self.inner.lock().unwrap();
272 0 : let Some((id, item)) = guard.queued.pop_front() else {
273 0 : return Ok(CompactionOutcome::Done);
274 : };
275 0 : guard.running = Some((id, item.clone()));
276 0 : has_pending_tasks = !guard.queued.is_empty();
277 0 : (id, item)
278 0 : };
279 0 :
280 0 : match item {
281 0 : GcCompactionQueueItem::Manual(options) => {
282 0 : if !options
283 0 : .flags
284 0 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
285 : {
286 0 : warn!("ignoring scheduled compaction task: scheduled task must be gc compaction: {:?}", options);
287 0 : } else if options.sub_compaction {
288 0 : self.handle_sub_compaction(id, options, timeline, gc_block)
289 0 : .await?;
290 : } else {
291 0 : let gc_guard = match gc_block.start().await {
292 0 : Ok(guard) => guard,
293 0 : Err(e) => {
294 0 : return Err(CompactionError::Other(anyhow!(
295 0 : "cannot run gc-compaction because gc is blocked: {}",
296 0 : e
297 0 : )));
298 : }
299 : };
300 0 : {
301 0 : let mut guard = self.inner.lock().unwrap();
302 0 : guard.gc_guards.insert(id, gc_guard);
303 0 : }
304 0 : let _ = timeline.compact_with_options(cancel, options, ctx).await?;
305 0 : self.notify_and_unblock(id);
306 : }
307 : }
308 0 : GcCompactionQueueItem::SubCompactionJob(options) => {
309 0 : let _ = timeline.compact_with_options(cancel, options, ctx).await?;
310 : }
311 0 : GcCompactionQueueItem::Notify(id) => {
312 0 : self.notify_and_unblock(id);
313 0 : }
314 : GcCompactionQueueItem::UpdateL2Lsn(_) => {
315 0 : unreachable!()
316 : }
317 : }
318 0 : {
319 0 : let mut guard = self.inner.lock().unwrap();
320 0 : guard.running = None;
321 0 : }
322 0 : Ok(if has_pending_tasks {
323 0 : CompactionOutcome::Pending
324 : } else {
325 0 : CompactionOutcome::Done
326 : })
327 0 : }
328 :
329 : #[allow(clippy::type_complexity)]
330 0 : pub fn remaining_jobs(
331 0 : &self,
332 0 : ) -> (
333 0 : Option<(GcCompactionJobId, GcCompactionQueueItem)>,
334 0 : VecDeque<(GcCompactionJobId, GcCompactionQueueItem)>,
335 0 : ) {
336 0 : let guard = self.inner.lock().unwrap();
337 0 : (guard.running.clone(), guard.queued.clone())
338 0 : }
339 :
340 : #[allow(dead_code)]
341 0 : pub fn remaining_jobs_num(&self) -> usize {
342 0 : let guard = self.inner.lock().unwrap();
343 0 : guard.queued.len() + if guard.running.is_some() { 1 } else { 0 }
344 0 : }
345 : }
346 :
347 : /// A job description for the gc-compaction job. This structure describes the rectangle range that the job will
348 : /// process. The exact layers that need to be compacted/rewritten will be generated when `compact_with_gc` gets
349 : /// called.
350 : #[derive(Debug, Clone)]
351 : pub(crate) struct GcCompactJob {
352 : pub dry_run: bool,
353 : /// The key range to be compacted. The compaction algorithm will only regenerate key-value pairs within this range
354 : /// [left inclusive, right exclusive), and other pairs will be rewritten into new files if necessary.
355 : pub compact_key_range: Range<Key>,
356 : /// The LSN range to be compacted. The compaction algorithm will use this range to determine the layers to be
357 : /// selected for the compaction, and it does not guarantee the generated layers will have exactly the same LSN range
358 : /// as specified here. The true range being compacted is `min_lsn/max_lsn` in [`GcCompactionJobDescription`].
359 : /// min_lsn will always <= the lower bound specified here, and max_lsn will always >= the upper bound specified here.
360 : pub compact_lsn_range: Range<Lsn>,
361 : }
362 :
363 : impl GcCompactJob {
364 108 : pub fn from_compact_options(options: CompactOptions) -> Self {
365 108 : GcCompactJob {
366 108 : dry_run: options.flags.contains(CompactFlags::DryRun),
367 108 : compact_key_range: options
368 108 : .compact_key_range
369 108 : .map(|x| x.into())
370 108 : .unwrap_or(Key::MIN..Key::MAX),
371 108 : compact_lsn_range: options
372 108 : .compact_lsn_range
373 108 : .map(|x| x.into())
374 108 : .unwrap_or(Lsn::INVALID..Lsn::MAX),
375 108 : }
376 108 : }
377 : }
378 :
379 : /// A job description for the gc-compaction job. This structure is generated when `compact_with_gc` is called
380 : /// and contains the exact layers we want to compact.
381 : pub struct GcCompactionJobDescription {
382 : /// All layers to read in the compaction job
383 : selected_layers: Vec<Layer>,
384 : /// GC cutoff of the job. This is the lowest LSN that will be accessed by the read/GC path and we need to
385 : /// keep all deltas <= this LSN or generate an image == this LSN.
386 : gc_cutoff: Lsn,
387 : /// LSNs to retain for the job. Read path will use this LSN so we need to keep deltas <= this LSN or
388 : /// generate an image == this LSN.
389 : retain_lsns_below_horizon: Vec<Lsn>,
390 : /// Maximum layer LSN processed in this compaction, that is max(end_lsn of layers). Exclusive. All data
391 : /// \>= this LSN will be kept and will not be rewritten.
392 : max_layer_lsn: Lsn,
393 : /// Minimum layer LSN processed in this compaction, that is min(start_lsn of layers). Inclusive.
394 : /// All access below (strict lower than `<`) this LSN will be routed through the normal read path instead of
395 : /// k-merge within gc-compaction.
396 : min_layer_lsn: Lsn,
397 : /// Only compact layers overlapping with this range.
398 : compaction_key_range: Range<Key>,
399 : /// When partial compaction is enabled, these layers need to be rewritten to ensure no overlap.
400 : /// This field is here solely for debugging. The field will not be read once the compaction
401 : /// description is generated.
402 : rewrite_layers: Vec<Arc<PersistentLayerDesc>>,
403 : }
404 :
405 : /// The result of bottom-most compaction for a single key at each LSN.
406 : #[derive(Debug)]
407 : #[cfg_attr(test, derive(PartialEq))]
408 : pub struct KeyLogAtLsn(pub Vec<(Lsn, Value)>);
409 :
410 : /// The result of bottom-most compaction.
411 : #[derive(Debug)]
412 : #[cfg_attr(test, derive(PartialEq))]
413 : pub(crate) struct KeyHistoryRetention {
414 : /// Stores logs to reconstruct the value at the given LSN, that is to say, logs <= LSN or image == LSN.
415 : pub(crate) below_horizon: Vec<(Lsn, KeyLogAtLsn)>,
416 : /// Stores logs to reconstruct the value at any LSN above the horizon, that is to say, log > LSN.
417 : pub(crate) above_horizon: KeyLogAtLsn,
418 : }
419 :
420 : impl KeyHistoryRetention {
421 : /// Hack: skip delta layer if we need to produce a layer of a same key-lsn.
422 : ///
423 : /// This can happen if we have removed some deltas in "the middle" of some existing layer's key-lsn-range.
424 : /// For example, consider the case where a single delta with range [0x10,0x50) exists.
425 : /// And we have branches at LSN 0x10, 0x20, 0x30.
426 : /// Then we delete branch @ 0x20.
427 : /// Bottom-most compaction may now delete the delta [0x20,0x30).
428 : /// And that wouldnt' change the shape of the layer.
429 : ///
430 : /// Note that bottom-most-gc-compaction never _adds_ new data in that case, only removes.
431 : ///
432 : /// `discard_key` will only be called when the writer reaches its target (instead of for every key), so it's fine to grab a lock inside.
433 148 : async fn discard_key(key: &PersistentLayerKey, tline: &Arc<Timeline>, dry_run: bool) -> bool {
434 148 : if dry_run {
435 0 : return true;
436 148 : }
437 148 : if LayerMap::is_l0(&key.key_range, key.is_delta) {
438 : // gc-compaction should not produce L0 deltas, otherwise it will break the layer order.
439 : // We should ignore such layers.
440 0 : return true;
441 148 : }
442 : let layer_generation;
443 : {
444 148 : let guard = tline.layers.read().await;
445 148 : if !guard.contains_key(key) {
446 104 : return false;
447 44 : }
448 44 : layer_generation = guard.get_from_key(key).metadata().generation;
449 44 : }
450 44 : if layer_generation == tline.generation {
451 44 : info!(
452 : key=%key,
453 : ?layer_generation,
454 0 : "discard layer due to duplicated layer key in the same generation",
455 : );
456 44 : true
457 : } else {
458 0 : false
459 : }
460 148 : }
461 :
462 : /// Pipe a history of a single key to the writers.
463 : ///
464 : /// If `image_writer` is none, the images will be placed into the delta layers.
465 : /// The delta writer will contain all images and deltas (below and above the horizon) except the bottom-most images.
466 : #[allow(clippy::too_many_arguments)]
467 1244 : async fn pipe_to(
468 1244 : self,
469 1244 : key: Key,
470 1244 : delta_writer: &mut SplitDeltaLayerWriter,
471 1244 : mut image_writer: Option<&mut SplitImageLayerWriter>,
472 1244 : stat: &mut CompactionStatistics,
473 1244 : ctx: &RequestContext,
474 1244 : ) -> anyhow::Result<()> {
475 1244 : let mut first_batch = true;
476 4024 : for (cutoff_lsn, KeyLogAtLsn(logs)) in self.below_horizon {
477 2780 : if first_batch {
478 1244 : if logs.len() == 1 && logs[0].1.is_image() {
479 1168 : let Value::Image(img) = &logs[0].1 else {
480 0 : unreachable!()
481 : };
482 1168 : stat.produce_image_key(img);
483 1168 : if let Some(image_writer) = image_writer.as_mut() {
484 1168 : image_writer.put_image(key, img.clone(), ctx).await?;
485 : } else {
486 0 : delta_writer
487 0 : .put_value(key, cutoff_lsn, Value::Image(img.clone()), ctx)
488 0 : .await?;
489 : }
490 : } else {
491 132 : for (lsn, val) in logs {
492 56 : stat.produce_key(&val);
493 56 : delta_writer.put_value(key, lsn, val, ctx).await?;
494 : }
495 : }
496 1244 : first_batch = false;
497 : } else {
498 1768 : for (lsn, val) in logs {
499 232 : stat.produce_key(&val);
500 232 : delta_writer.put_value(key, lsn, val, ctx).await?;
501 : }
502 : }
503 : }
504 1244 : let KeyLogAtLsn(above_horizon_logs) = self.above_horizon;
505 1360 : for (lsn, val) in above_horizon_logs {
506 116 : stat.produce_key(&val);
507 116 : delta_writer.put_value(key, lsn, val, ctx).await?;
508 : }
509 1244 : Ok(())
510 1244 : }
511 : }
512 :
513 : #[derive(Debug, Serialize, Default)]
514 : struct CompactionStatisticsNumSize {
515 : num: u64,
516 : size: u64,
517 : }
518 :
519 : #[derive(Debug, Serialize, Default)]
520 : pub struct CompactionStatistics {
521 : delta_layer_visited: CompactionStatisticsNumSize,
522 : image_layer_visited: CompactionStatisticsNumSize,
523 : delta_layer_produced: CompactionStatisticsNumSize,
524 : image_layer_produced: CompactionStatisticsNumSize,
525 : num_delta_layer_discarded: usize,
526 : num_image_layer_discarded: usize,
527 : num_unique_keys_visited: usize,
528 : wal_keys_visited: CompactionStatisticsNumSize,
529 : image_keys_visited: CompactionStatisticsNumSize,
530 : wal_produced: CompactionStatisticsNumSize,
531 : image_produced: CompactionStatisticsNumSize,
532 : }
533 :
534 : impl CompactionStatistics {
535 2084 : fn estimated_size_of_value(val: &Value) -> usize {
536 864 : match val {
537 1220 : Value::Image(img) => img.len(),
538 0 : Value::WalRecord(NeonWalRecord::Postgres { rec, .. }) => rec.len(),
539 864 : _ => std::mem::size_of::<NeonWalRecord>(),
540 : }
541 2084 : }
542 3272 : fn estimated_size_of_key() -> usize {
543 3272 : KEY_SIZE // TODO: distinguish image layer and delta layer (count LSN in delta layer)
544 3272 : }
545 172 : fn visit_delta_layer(&mut self, size: u64) {
546 172 : self.delta_layer_visited.num += 1;
547 172 : self.delta_layer_visited.size += size;
548 172 : }
549 132 : fn visit_image_layer(&mut self, size: u64) {
550 132 : self.image_layer_visited.num += 1;
551 132 : self.image_layer_visited.size += size;
552 132 : }
553 1244 : fn on_unique_key_visited(&mut self) {
554 1244 : self.num_unique_keys_visited += 1;
555 1244 : }
556 480 : fn visit_wal_key(&mut self, val: &Value) {
557 480 : self.wal_keys_visited.num += 1;
558 480 : self.wal_keys_visited.size +=
559 480 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
560 480 : }
561 1220 : fn visit_image_key(&mut self, val: &Value) {
562 1220 : self.image_keys_visited.num += 1;
563 1220 : self.image_keys_visited.size +=
564 1220 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
565 1220 : }
566 404 : fn produce_key(&mut self, val: &Value) {
567 404 : match val {
568 20 : Value::Image(img) => self.produce_image_key(img),
569 384 : Value::WalRecord(_) => self.produce_wal_key(val),
570 : }
571 404 : }
572 384 : fn produce_wal_key(&mut self, val: &Value) {
573 384 : self.wal_produced.num += 1;
574 384 : self.wal_produced.size +=
575 384 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
576 384 : }
577 1188 : fn produce_image_key(&mut self, val: &Bytes) {
578 1188 : self.image_produced.num += 1;
579 1188 : self.image_produced.size += val.len() as u64 + Self::estimated_size_of_key() as u64;
580 1188 : }
581 28 : fn discard_delta_layer(&mut self) {
582 28 : self.num_delta_layer_discarded += 1;
583 28 : }
584 16 : fn discard_image_layer(&mut self) {
585 16 : self.num_image_layer_discarded += 1;
586 16 : }
587 44 : fn produce_delta_layer(&mut self, size: u64) {
588 44 : self.delta_layer_produced.num += 1;
589 44 : self.delta_layer_produced.size += size;
590 44 : }
591 60 : fn produce_image_layer(&mut self, size: u64) {
592 60 : self.image_layer_produced.num += 1;
593 60 : self.image_layer_produced.size += size;
594 60 : }
595 : }
596 :
597 : #[derive(Default, Debug, Clone, Copy, PartialEq, Eq)]
598 : pub enum CompactionOutcome {
599 : #[default]
600 : /// No layers need to be compacted after this round. Compaction doesn't need
601 : /// to be immediately scheduled.
602 : Done,
603 : /// Still has pending layers to be compacted after this round. Ideally, the scheduler
604 : /// should immediately schedule another compaction.
605 : Pending,
606 : /// A timeline needs L0 compaction. Yield and schedule an immediate L0 compaction pass (only
607 : /// guaranteed when `compaction_l0_first` is enabled).
608 : YieldForL0,
609 : /// Compaction was skipped, because the timeline is ineligible for compaction.
610 : Skipped,
611 : }
612 :
613 : impl Timeline {
614 : /// TODO: cancellation
615 : ///
616 : /// Returns whether the compaction has pending tasks.
617 728 : pub(crate) async fn compact_legacy(
618 728 : self: &Arc<Self>,
619 728 : cancel: &CancellationToken,
620 728 : options: CompactOptions,
621 728 : ctx: &RequestContext,
622 728 : ) -> Result<CompactionOutcome, CompactionError> {
623 728 : if options
624 728 : .flags
625 728 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
626 : {
627 0 : self.compact_with_gc(cancel, options, ctx)
628 0 : .await
629 0 : .map_err(CompactionError::Other)?;
630 0 : return Ok(CompactionOutcome::Done);
631 728 : }
632 728 :
633 728 : if options.flags.contains(CompactFlags::DryRun) {
634 0 : return Err(CompactionError::Other(anyhow!(
635 0 : "dry-run mode is not supported for legacy compaction for now"
636 0 : )));
637 728 : }
638 728 :
639 728 : if options.compact_key_range.is_some() || options.compact_lsn_range.is_some() {
640 : // maybe useful in the future? could implement this at some point
641 0 : return Err(CompactionError::Other(anyhow!(
642 0 : "compaction range is not supported for legacy compaction for now"
643 0 : )));
644 728 : }
645 728 :
646 728 : // High level strategy for compaction / image creation:
647 728 : //
648 728 : // 1. First, do a L0 compaction to ensure we move the L0
649 728 : // layers into the historic layer map get flat levels of
650 728 : // layers. If we did not compact all L0 layers, we will
651 728 : // prioritize compacting the timeline again and not do
652 728 : // any of the compactions below.
653 728 : //
654 728 : // 2. Then, calculate the desired "partitioning" of the
655 728 : // currently in-use key space. The goal is to partition the
656 728 : // key space into roughly fixed-size chunks, but also take into
657 728 : // account any existing image layers, and try to align the
658 728 : // chunk boundaries with the existing image layers to avoid
659 728 : // too much churn. Also try to align chunk boundaries with
660 728 : // relation boundaries. In principle, we don't know about
661 728 : // relation boundaries here, we just deal with key-value
662 728 : // pairs, and the code in pgdatadir_mapping.rs knows how to
663 728 : // map relations into key-value pairs. But in practice we know
664 728 : // that 'field6' is the block number, and the fields 1-5
665 728 : // identify a relation. This is just an optimization,
666 728 : // though.
667 728 : //
668 728 : // 3. Once we know the partitioning, for each partition,
669 728 : // decide if it's time to create a new image layer. The
670 728 : // criteria is: there has been too much "churn" since the last
671 728 : // image layer? The "churn" is fuzzy concept, it's a
672 728 : // combination of too many delta files, or too much WAL in
673 728 : // total in the delta file. Or perhaps: if creating an image
674 728 : // file would allow to delete some older files.
675 728 : //
676 728 : // 4. In the end, if the tenant gets auto-sharded, we will run
677 728 : // a shard-ancestor compaction.
678 728 :
679 728 : // Is the timeline being deleted?
680 728 : if self.is_stopping() {
681 0 : trace!("Dropping out of compaction on timeline shutdown");
682 0 : return Err(CompactionError::ShuttingDown);
683 728 : }
684 728 :
685 728 : let target_file_size = self.get_checkpoint_distance();
686 :
687 : // Define partitioning schema if needed
688 :
689 : // 1. L0 Compact
690 728 : let l0_outcome = {
691 728 : let timer = self.metrics.compact_time_histo.start_timer();
692 728 : let l0_outcome = self
693 728 : .compact_level0(
694 728 : target_file_size,
695 728 : options.flags.contains(CompactFlags::ForceL0Compaction),
696 728 : ctx,
697 728 : )
698 728 : .await?;
699 728 : timer.stop_and_record();
700 728 : l0_outcome
701 728 : };
702 728 :
703 728 : if options.flags.contains(CompactFlags::OnlyL0Compaction) {
704 0 : return Ok(l0_outcome);
705 728 : }
706 728 :
707 728 : // Yield if we have pending L0 compaction. The scheduler will do another pass.
708 728 : if (l0_outcome == CompactionOutcome::Pending || l0_outcome == CompactionOutcome::YieldForL0)
709 0 : && !options.flags.contains(CompactFlags::NoYield)
710 : {
711 0 : info!("image/ancestor compaction yielding for L0 compaction");
712 0 : return Ok(CompactionOutcome::YieldForL0);
713 728 : }
714 728 :
715 728 : // 2. Repartition and create image layers if necessary
716 728 : match self
717 728 : .repartition(
718 728 : self.get_last_record_lsn(),
719 728 : self.get_compaction_target_size(),
720 728 : options.flags,
721 728 : ctx,
722 728 : )
723 728 : .await
724 : {
725 728 : Ok(((dense_partitioning, sparse_partitioning), lsn)) => {
726 728 : // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
727 728 : let image_ctx = RequestContextBuilder::extend(ctx)
728 728 : .access_stats_behavior(AccessStatsBehavior::Skip)
729 728 : .build();
730 728 :
731 728 : let mut partitioning = dense_partitioning;
732 728 : partitioning
733 728 : .parts
734 728 : .extend(sparse_partitioning.into_dense().parts);
735 :
736 : // 3. Create new image layers for partitions that have been modified "enough".
737 728 : let (image_layers, outcome) = self
738 728 : .create_image_layers(
739 728 : &partitioning,
740 728 : lsn,
741 728 : if options
742 728 : .flags
743 728 : .contains(CompactFlags::ForceImageLayerCreation)
744 : {
745 28 : ImageLayerCreationMode::Force
746 : } else {
747 700 : ImageLayerCreationMode::Try
748 : },
749 728 : &image_ctx,
750 728 : self.last_image_layer_creation_status
751 728 : .load()
752 728 : .as_ref()
753 728 : .clone(),
754 728 : !options.flags.contains(CompactFlags::NoYield),
755 728 : )
756 728 : .await
757 728 : .inspect_err(|err| {
758 : if let CreateImageLayersError::GetVectoredError(
759 : GetVectoredError::MissingKey(_),
760 0 : ) = err
761 : {
762 0 : critical!("missing key during compaction: {err:?}");
763 0 : }
764 728 : })?;
765 :
766 728 : self.last_image_layer_creation_status
767 728 : .store(Arc::new(outcome.clone()));
768 728 :
769 728 : self.upload_new_image_layers(image_layers)?;
770 728 : if let LastImageLayerCreationStatus::Incomplete { .. } = outcome {
771 : // Yield and do not do any other kind of compaction.
772 0 : info!("skipping shard ancestor compaction due to pending image layer generation tasks (preempted by L0 compaction).");
773 0 : return Ok(CompactionOutcome::YieldForL0);
774 728 : }
775 : }
776 0 : Err(err) => {
777 0 : // no partitioning? This is normal, if the timeline was just created
778 0 : // as an empty timeline. Also in unit tests, when we use the timeline
779 0 : // as a simple key-value store, ignoring the datadir layout. Log the
780 0 : // error but continue.
781 0 : //
782 0 : // Suppress error when it's due to cancellation
783 0 : if !self.cancel.is_cancelled() && !err.is_cancelled() {
784 0 : tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
785 0 : }
786 : }
787 : };
788 :
789 728 : let partition_count = self.partitioning.read().0 .0.parts.len();
790 728 :
791 728 : // 4. Shard ancestor compaction
792 728 :
793 728 : if self.shard_identity.count >= ShardCount::new(2) {
794 : // Limit the number of layer rewrites to the number of partitions: this means its
795 : // runtime should be comparable to a full round of image layer creations, rather than
796 : // being potentially much longer.
797 0 : let rewrite_max = partition_count;
798 0 :
799 0 : self.compact_shard_ancestors(rewrite_max, ctx).await?;
800 728 : }
801 :
802 728 : Ok(CompactionOutcome::Done)
803 728 : }
804 :
805 : /// Check for layers that are elegible to be rewritten:
806 : /// - Shard splitting: After a shard split, ancestor layers beyond pitr_interval, so that
807 : /// we don't indefinitely retain keys in this shard that aren't needed.
808 : /// - For future use: layers beyond pitr_interval that are in formats we would
809 : /// rather not maintain compatibility with indefinitely.
810 : ///
811 : /// Note: this phase may read and write many gigabytes of data: use rewrite_max to bound
812 : /// how much work it will try to do in each compaction pass.
813 0 : async fn compact_shard_ancestors(
814 0 : self: &Arc<Self>,
815 0 : rewrite_max: usize,
816 0 : ctx: &RequestContext,
817 0 : ) -> Result<(), CompactionError> {
818 0 : let mut drop_layers = Vec::new();
819 0 : let mut layers_to_rewrite: Vec<Layer> = Vec::new();
820 0 :
821 0 : // We will use the Lsn cutoff of the last GC as a threshold for rewriting layers: if a
822 0 : // layer is behind this Lsn, it indicates that the layer is being retained beyond the
823 0 : // pitr_interval, for example because a branchpoint references it.
824 0 : //
825 0 : // Holding this read guard also blocks [`Self::gc_timeline`] from entering while we
826 0 : // are rewriting layers.
827 0 : let latest_gc_cutoff = self.get_applied_gc_cutoff_lsn();
828 0 :
829 0 : tracing::info!(
830 0 : "latest_gc_cutoff: {}, pitr cutoff {}",
831 0 : *latest_gc_cutoff,
832 0 : self.gc_info.read().unwrap().cutoffs.time
833 : );
834 :
835 0 : let layers = self.layers.read().await;
836 0 : for layer_desc in layers.layer_map()?.iter_historic_layers() {
837 0 : let layer = layers.get_from_desc(&layer_desc);
838 0 : if layer.metadata().shard.shard_count == self.shard_identity.count {
839 : // This layer does not belong to a historic ancestor, no need to re-image it.
840 0 : continue;
841 0 : }
842 0 :
843 0 : // This layer was created on an ancestor shard: check if it contains any data for this shard.
844 0 : let sharded_range = ShardedRange::new(layer_desc.get_key_range(), &self.shard_identity);
845 0 : let layer_local_page_count = sharded_range.page_count();
846 0 : let layer_raw_page_count = ShardedRange::raw_size(&layer_desc.get_key_range());
847 0 : if layer_local_page_count == 0 {
848 : // This ancestral layer only covers keys that belong to other shards.
849 : // We include the full metadata in the log: if we had some critical bug that caused
850 : // us to incorrectly drop layers, this would simplify manually debugging + reinstating those layers.
851 0 : info!(%layer, old_metadata=?layer.metadata(),
852 0 : "dropping layer after shard split, contains no keys for this shard.",
853 : );
854 :
855 0 : if cfg!(debug_assertions) {
856 : // Expensive, exhaustive check of keys in this layer: this guards against ShardedRange's calculations being
857 : // wrong. If ShardedRange claims the local page count is zero, then no keys in this layer
858 : // should be !is_key_disposable()
859 0 : let range = layer_desc.get_key_range();
860 0 : let mut key = range.start;
861 0 : while key < range.end {
862 0 : debug_assert!(self.shard_identity.is_key_disposable(&key));
863 0 : key = key.next();
864 : }
865 0 : }
866 :
867 0 : drop_layers.push(layer);
868 0 : continue;
869 0 : } else if layer_local_page_count != u32::MAX
870 0 : && layer_local_page_count == layer_raw_page_count
871 : {
872 0 : debug!(%layer,
873 0 : "layer is entirely shard local ({} keys), no need to filter it",
874 : layer_local_page_count
875 : );
876 0 : continue;
877 0 : }
878 0 :
879 0 : // Don't bother re-writing a layer unless it will at least halve its size
880 0 : if layer_local_page_count != u32::MAX
881 0 : && layer_local_page_count > layer_raw_page_count / 2
882 : {
883 0 : debug!(%layer,
884 0 : "layer is already mostly local ({}/{}), not rewriting",
885 : layer_local_page_count,
886 : layer_raw_page_count
887 : );
888 0 : }
889 :
890 : // Don't bother re-writing a layer if it is within the PITR window: it will age-out eventually
891 : // without incurring the I/O cost of a rewrite.
892 0 : if layer_desc.get_lsn_range().end >= *latest_gc_cutoff {
893 0 : debug!(%layer, "Skipping rewrite of layer still in GC window ({} >= {})",
894 0 : layer_desc.get_lsn_range().end, *latest_gc_cutoff);
895 0 : continue;
896 0 : }
897 0 :
898 0 : if layer_desc.is_delta() {
899 : // We do not yet implement rewrite of delta layers
900 0 : debug!(%layer, "Skipping rewrite of delta layer");
901 0 : continue;
902 0 : }
903 0 :
904 0 : // Only rewrite layers if their generations differ. This guarantees:
905 0 : // - that local rewrite is safe, as local layer paths will differ between existing layer and rewritten one
906 0 : // - that the layer is persistent in remote storage, as we only see old-generation'd layer via loading from remote storage
907 0 : if layer.metadata().generation == self.generation {
908 0 : debug!(%layer, "Skipping rewrite, is not from old generation");
909 0 : continue;
910 0 : }
911 0 :
912 0 : if layers_to_rewrite.len() >= rewrite_max {
913 0 : tracing::info!(%layer, "Will rewrite layer on a future compaction, already rewrote {}",
914 0 : layers_to_rewrite.len()
915 : );
916 0 : continue;
917 0 : }
918 0 :
919 0 : // Fall through: all our conditions for doing a rewrite passed.
920 0 : layers_to_rewrite.push(layer);
921 : }
922 :
923 : // Drop read lock on layer map before we start doing time-consuming I/O
924 0 : drop(layers);
925 0 :
926 0 : let mut replace_image_layers = Vec::new();
927 :
928 0 : for layer in layers_to_rewrite {
929 0 : tracing::info!(layer=%layer, "Rewriting layer after shard split...");
930 0 : let mut image_layer_writer = ImageLayerWriter::new(
931 0 : self.conf,
932 0 : self.timeline_id,
933 0 : self.tenant_shard_id,
934 0 : &layer.layer_desc().key_range,
935 0 : layer.layer_desc().image_layer_lsn(),
936 0 : ctx,
937 0 : )
938 0 : .await
939 0 : .map_err(CompactionError::Other)?;
940 :
941 : // Safety of layer rewrites:
942 : // - We are writing to a different local file path than we are reading from, so the old Layer
943 : // cannot interfere with the new one.
944 : // - In the page cache, contents for a particular VirtualFile are stored with a file_id that
945 : // is different for two layers with the same name (in `ImageLayerInner::new` we always
946 : // acquire a fresh id from [`crate::page_cache::next_file_id`]. So readers do not risk
947 : // reading the index from one layer file, and then data blocks from the rewritten layer file.
948 : // - Any readers that have a reference to the old layer will keep it alive until they are done
949 : // with it. If they are trying to promote from remote storage, that will fail, but this is the same
950 : // as for compaction generally: compaction is allowed to delete layers that readers might be trying to use.
951 : // - We do not run concurrently with other kinds of compaction, so the only layer map writes we race with are:
952 : // - GC, which at worst witnesses us "undelete" a layer that they just deleted.
953 : // - ingestion, which only inserts layers, therefore cannot collide with us.
954 0 : let resident = layer.download_and_keep_resident().await?;
955 :
956 0 : let keys_written = resident
957 0 : .filter(&self.shard_identity, &mut image_layer_writer, ctx)
958 0 : .await?;
959 :
960 0 : if keys_written > 0 {
961 0 : let (desc, path) = image_layer_writer
962 0 : .finish(ctx)
963 0 : .await
964 0 : .map_err(CompactionError::Other)?;
965 0 : let new_layer = Layer::finish_creating(self.conf, self, desc, &path)
966 0 : .map_err(CompactionError::Other)?;
967 0 : tracing::info!(layer=%new_layer, "Rewrote layer, {} -> {} bytes",
968 0 : layer.metadata().file_size,
969 0 : new_layer.metadata().file_size);
970 :
971 0 : replace_image_layers.push((layer, new_layer));
972 0 : } else {
973 0 : // Drop the old layer. Usually for this case we would already have noticed that
974 0 : // the layer has no data for us with the ShardedRange check above, but
975 0 : drop_layers.push(layer);
976 0 : }
977 : }
978 :
979 : // At this point, we have replaced local layer files with their rewritten form, but not yet uploaded
980 : // metadata to reflect that. If we restart here, the replaced layer files will look invalid (size mismatch
981 : // to remote index) and be removed. This is inefficient but safe.
982 0 : fail::fail_point!("compact-shard-ancestors-localonly");
983 0 :
984 0 : // Update the LayerMap so that readers will use the new layers, and enqueue it for writing to remote storage
985 0 : self.rewrite_layers(replace_image_layers, drop_layers)
986 0 : .await?;
987 :
988 0 : fail::fail_point!("compact-shard-ancestors-enqueued");
989 0 :
990 0 : // We wait for all uploads to complete before finishing this compaction stage. This is not
991 0 : // necessary for correctness, but it simplifies testing, and avoids proceeding with another
992 0 : // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
993 0 : // load.
994 0 : match self.remote_client.wait_completion().await {
995 0 : Ok(()) => (),
996 0 : Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
997 : Err(WaitCompletionError::UploadQueueShutDownOrStopped) => {
998 0 : return Err(CompactionError::ShuttingDown)
999 : }
1000 : }
1001 :
1002 0 : fail::fail_point!("compact-shard-ancestors-persistent");
1003 0 :
1004 0 : Ok(())
1005 0 : }
1006 :
1007 : /// Update the LayerVisibilityHint of layers covered by image layers, based on whether there is
1008 : /// an image layer between them and the most recent readable LSN (branch point or tip of timeline). The
1009 : /// purpose of the visibility hint is to record which layers need to be available to service reads.
1010 : ///
1011 : /// The result may be used as an input to eviction and secondary downloads to de-prioritize layers
1012 : /// that we know won't be needed for reads.
1013 460 : pub(super) async fn update_layer_visibility(
1014 460 : &self,
1015 460 : ) -> Result<(), super::layer_manager::Shutdown> {
1016 460 : let head_lsn = self.get_last_record_lsn();
1017 :
1018 : // We will sweep through layers in reverse-LSN order. We only do historic layers. L0 deltas
1019 : // are implicitly left visible, because LayerVisibilityHint's default is Visible, and we never modify it here.
1020 : // Note that L0 deltas _can_ be covered by image layers, but we consider them 'visible' because we anticipate that
1021 : // they will be subject to L0->L1 compaction in the near future.
1022 460 : let layer_manager = self.layers.read().await;
1023 460 : let layer_map = layer_manager.layer_map()?;
1024 :
1025 460 : let readable_points = {
1026 460 : let children = self.gc_info.read().unwrap().retain_lsns.clone();
1027 460 :
1028 460 : let mut readable_points = Vec::with_capacity(children.len() + 1);
1029 460 : for (child_lsn, _child_timeline_id, is_offloaded) in &children {
1030 0 : if *is_offloaded == MaybeOffloaded::Yes {
1031 0 : continue;
1032 0 : }
1033 0 : readable_points.push(*child_lsn);
1034 : }
1035 460 : readable_points.push(head_lsn);
1036 460 : readable_points
1037 460 : };
1038 460 :
1039 460 : let (layer_visibility, covered) = layer_map.get_visibility(readable_points);
1040 1168 : for (layer_desc, visibility) in layer_visibility {
1041 708 : // FIXME: a more efficiency bulk zip() through the layers rather than NlogN getting each one
1042 708 : let layer = layer_manager.get_from_desc(&layer_desc);
1043 708 : layer.set_visibility(visibility);
1044 708 : }
1045 :
1046 : // TODO: publish our covered KeySpace to our parent, so that when they update their visibility, they can
1047 : // avoid assuming that everything at a branch point is visible.
1048 460 : drop(covered);
1049 460 : Ok(())
1050 460 : }
1051 :
1052 : /// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
1053 : /// as Level 1 files. Returns whether the L0 layers are fully compacted.
1054 728 : async fn compact_level0(
1055 728 : self: &Arc<Self>,
1056 728 : target_file_size: u64,
1057 728 : force_compaction_ignore_threshold: bool,
1058 728 : ctx: &RequestContext,
1059 728 : ) -> Result<CompactionOutcome, CompactionError> {
1060 : let CompactLevel0Phase1Result {
1061 728 : new_layers,
1062 728 : deltas_to_compact,
1063 728 : outcome,
1064 : } = {
1065 728 : let phase1_span = info_span!("compact_level0_phase1");
1066 728 : let ctx = ctx.attached_child();
1067 728 : let mut stats = CompactLevel0Phase1StatsBuilder {
1068 728 : version: Some(2),
1069 728 : tenant_id: Some(self.tenant_shard_id),
1070 728 : timeline_id: Some(self.timeline_id),
1071 728 : ..Default::default()
1072 728 : };
1073 728 :
1074 728 : let begin = tokio::time::Instant::now();
1075 728 : let phase1_layers_locked = self.layers.read().await;
1076 728 : let now = tokio::time::Instant::now();
1077 728 : stats.read_lock_acquisition_micros =
1078 728 : DurationRecorder::Recorded(RecordedDuration(now - begin), now);
1079 728 : self.compact_level0_phase1(
1080 728 : phase1_layers_locked,
1081 728 : stats,
1082 728 : target_file_size,
1083 728 : force_compaction_ignore_threshold,
1084 728 : &ctx,
1085 728 : )
1086 728 : .instrument(phase1_span)
1087 728 : .await?
1088 : };
1089 :
1090 728 : if new_layers.is_empty() && deltas_to_compact.is_empty() {
1091 : // nothing to do
1092 672 : return Ok(CompactionOutcome::Done);
1093 56 : }
1094 56 :
1095 56 : self.finish_compact_batch(&new_layers, &Vec::new(), &deltas_to_compact)
1096 56 : .await?;
1097 56 : Ok(outcome)
1098 728 : }
1099 :
1100 : /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
1101 728 : async fn compact_level0_phase1<'a>(
1102 728 : self: &'a Arc<Self>,
1103 728 : guard: tokio::sync::RwLockReadGuard<'a, LayerManager>,
1104 728 : mut stats: CompactLevel0Phase1StatsBuilder,
1105 728 : target_file_size: u64,
1106 728 : force_compaction_ignore_threshold: bool,
1107 728 : ctx: &RequestContext,
1108 728 : ) -> Result<CompactLevel0Phase1Result, CompactionError> {
1109 728 : stats.read_lock_held_spawn_blocking_startup_micros =
1110 728 : stats.read_lock_acquisition_micros.till_now(); // set by caller
1111 728 : let layers = guard.layer_map()?;
1112 728 : let level0_deltas = layers.level0_deltas();
1113 728 : stats.level0_deltas_count = Some(level0_deltas.len());
1114 728 :
1115 728 : // Only compact if enough layers have accumulated.
1116 728 : let threshold = self.get_compaction_threshold();
1117 728 : if level0_deltas.is_empty() || level0_deltas.len() < threshold {
1118 672 : if force_compaction_ignore_threshold {
1119 0 : if !level0_deltas.is_empty() {
1120 0 : info!(
1121 0 : level0_deltas = level0_deltas.len(),
1122 0 : threshold, "too few deltas to compact, but forcing compaction"
1123 : );
1124 : } else {
1125 0 : info!(
1126 0 : level0_deltas = level0_deltas.len(),
1127 0 : threshold, "too few deltas to compact, cannot force compaction"
1128 : );
1129 0 : return Ok(CompactLevel0Phase1Result::default());
1130 : }
1131 : } else {
1132 672 : debug!(
1133 0 : level0_deltas = level0_deltas.len(),
1134 0 : threshold, "too few deltas to compact"
1135 : );
1136 672 : return Ok(CompactLevel0Phase1Result::default());
1137 : }
1138 56 : }
1139 :
1140 56 : let mut level0_deltas = level0_deltas
1141 56 : .iter()
1142 804 : .map(|x| guard.get_from_desc(x))
1143 56 : .collect::<Vec<_>>();
1144 56 :
1145 56 : // Gather the files to compact in this iteration.
1146 56 : //
1147 56 : // Start with the oldest Level 0 delta file, and collect any other
1148 56 : // level 0 files that form a contiguous sequence, such that the end
1149 56 : // LSN of previous file matches the start LSN of the next file.
1150 56 : //
1151 56 : // Note that if the files don't form such a sequence, we might
1152 56 : // "compact" just a single file. That's a bit pointless, but it allows
1153 56 : // us to get rid of the level 0 file, and compact the other files on
1154 56 : // the next iteration. This could probably made smarter, but such
1155 56 : // "gaps" in the sequence of level 0 files should only happen in case
1156 56 : // of a crash, partial download from cloud storage, or something like
1157 56 : // that, so it's not a big deal in practice.
1158 1496 : level0_deltas.sort_by_key(|l| l.layer_desc().lsn_range.start);
1159 56 : let mut level0_deltas_iter = level0_deltas.iter();
1160 56 :
1161 56 : let first_level0_delta = level0_deltas_iter.next().unwrap();
1162 56 : let mut prev_lsn_end = first_level0_delta.layer_desc().lsn_range.end;
1163 56 : let mut deltas_to_compact = Vec::with_capacity(level0_deltas.len());
1164 56 :
1165 56 : // Accumulate the size of layers in `deltas_to_compact`
1166 56 : let mut deltas_to_compact_bytes = 0;
1167 56 :
1168 56 : // Under normal circumstances, we will accumulate up to compaction_upper_limit L0s of size
1169 56 : // checkpoint_distance each. To avoid edge cases using extra system resources, bound our
1170 56 : // work in this function to only operate on this much delta data at once.
1171 56 : //
1172 56 : // In general, compaction_threshold should be <= compaction_upper_limit, but in case that
1173 56 : // the constraint is not respected, we use the larger of the two.
1174 56 : let delta_size_limit = std::cmp::max(
1175 56 : self.get_compaction_upper_limit(),
1176 56 : self.get_compaction_threshold(),
1177 56 : ) as u64
1178 56 : * std::cmp::max(self.get_checkpoint_distance(), DEFAULT_CHECKPOINT_DISTANCE);
1179 56 :
1180 56 : let mut fully_compacted = true;
1181 56 :
1182 56 : deltas_to_compact.push(first_level0_delta.download_and_keep_resident().await?);
1183 804 : for l in level0_deltas_iter {
1184 748 : let lsn_range = &l.layer_desc().lsn_range;
1185 748 :
1186 748 : if lsn_range.start != prev_lsn_end {
1187 0 : break;
1188 748 : }
1189 748 : deltas_to_compact.push(l.download_and_keep_resident().await?);
1190 748 : deltas_to_compact_bytes += l.metadata().file_size;
1191 748 : prev_lsn_end = lsn_range.end;
1192 748 :
1193 748 : if deltas_to_compact_bytes >= delta_size_limit {
1194 0 : info!(
1195 0 : l0_deltas_selected = deltas_to_compact.len(),
1196 0 : l0_deltas_total = level0_deltas.len(),
1197 0 : "L0 compaction picker hit max delta layer size limit: {}",
1198 : delta_size_limit
1199 : );
1200 0 : fully_compacted = false;
1201 0 :
1202 0 : // Proceed with compaction, but only a subset of L0s
1203 0 : break;
1204 748 : }
1205 : }
1206 56 : let lsn_range = Range {
1207 56 : start: deltas_to_compact
1208 56 : .first()
1209 56 : .unwrap()
1210 56 : .layer_desc()
1211 56 : .lsn_range
1212 56 : .start,
1213 56 : end: deltas_to_compact.last().unwrap().layer_desc().lsn_range.end,
1214 56 : };
1215 56 :
1216 56 : info!(
1217 0 : "Starting Level0 compaction in LSN range {}-{} for {} layers ({} deltas in total)",
1218 0 : lsn_range.start,
1219 0 : lsn_range.end,
1220 0 : deltas_to_compact.len(),
1221 0 : level0_deltas.len()
1222 : );
1223 :
1224 804 : for l in deltas_to_compact.iter() {
1225 804 : info!("compact includes {l}");
1226 : }
1227 :
1228 : // We don't need the original list of layers anymore. Drop it so that
1229 : // we don't accidentally use it later in the function.
1230 56 : drop(level0_deltas);
1231 56 :
1232 56 : stats.read_lock_held_prerequisites_micros = stats
1233 56 : .read_lock_held_spawn_blocking_startup_micros
1234 56 : .till_now();
1235 :
1236 : // TODO: replace with streaming k-merge
1237 56 : let all_keys = {
1238 56 : let mut all_keys = Vec::new();
1239 804 : for l in deltas_to_compact.iter() {
1240 804 : if self.cancel.is_cancelled() {
1241 0 : return Err(CompactionError::ShuttingDown);
1242 804 : }
1243 804 : let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
1244 804 : let keys = delta
1245 804 : .index_entries(ctx)
1246 804 : .await
1247 804 : .map_err(CompactionError::Other)?;
1248 804 : all_keys.extend(keys);
1249 : }
1250 : // The current stdlib sorting implementation is designed in a way where it is
1251 : // particularly fast where the slice is made up of sorted sub-ranges.
1252 8847562 : all_keys.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
1253 56 : all_keys
1254 56 : };
1255 56 :
1256 56 : stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
1257 :
1258 : // Determine N largest holes where N is number of compacted layers. The vec is sorted by key range start.
1259 : //
1260 : // A hole is a key range for which this compaction doesn't have any WAL records.
1261 : // Our goal in this compaction iteration is to avoid creating L1s that, in terms of their key range,
1262 : // cover the hole, but actually don't contain any WAL records for that key range.
1263 : // The reason is that the mere stack of L1s (`count_deltas`) triggers image layer creation (`create_image_layers`).
1264 : // That image layer creation would be useless for a hole range covered by L1s that don't contain any WAL records.
1265 : //
1266 : // The algorithm chooses holes as follows.
1267 : // - Slide a 2-window over the keys in key orde to get the hole range (=distance between two keys).
1268 : // - Filter: min threshold on range length
1269 : // - Rank: by coverage size (=number of image layers required to reconstruct each key in the range for which we have any data)
1270 : //
1271 : // For more details, intuition, and some ASCII art see https://github.com/neondatabase/neon/pull/3597#discussion_r1112704451
1272 : #[derive(PartialEq, Eq)]
1273 : struct Hole {
1274 : key_range: Range<Key>,
1275 : coverage_size: usize,
1276 : }
1277 56 : let holes: Vec<Hole> = {
1278 : use std::cmp::Ordering;
1279 : impl Ord for Hole {
1280 0 : fn cmp(&self, other: &Self) -> Ordering {
1281 0 : self.coverage_size.cmp(&other.coverage_size).reverse()
1282 0 : }
1283 : }
1284 : impl PartialOrd for Hole {
1285 0 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
1286 0 : Some(self.cmp(other))
1287 0 : }
1288 : }
1289 56 : let max_holes = deltas_to_compact.len();
1290 56 : let last_record_lsn = self.get_last_record_lsn();
1291 56 : let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
1292 56 : let min_hole_coverage_size = 3; // TODO: something more flexible?
1293 56 : // min-heap (reserve space for one more element added before eviction)
1294 56 : let mut heap: BinaryHeap<Hole> = BinaryHeap::with_capacity(max_holes + 1);
1295 56 : let mut prev: Option<Key> = None;
1296 :
1297 4128076 : for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
1298 4128076 : if let Some(prev_key) = prev {
1299 : // just first fast filter, do not create hole entries for metadata keys. The last hole in the
1300 : // compaction is the gap between data key and metadata keys.
1301 4128020 : if next_key.to_i128() - prev_key.to_i128() >= min_hole_range
1302 0 : && !Key::is_metadata_key(&prev_key)
1303 : {
1304 0 : let key_range = prev_key..next_key;
1305 0 : // Measuring hole by just subtraction of i128 representation of key range boundaries
1306 0 : // has not so much sense, because largest holes will corresponds field1/field2 changes.
1307 0 : // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
1308 0 : // That is why it is better to measure size of hole as number of covering image layers.
1309 0 : let coverage_size =
1310 0 : layers.image_coverage(&key_range, last_record_lsn).len();
1311 0 : if coverage_size >= min_hole_coverage_size {
1312 0 : heap.push(Hole {
1313 0 : key_range,
1314 0 : coverage_size,
1315 0 : });
1316 0 : if heap.len() > max_holes {
1317 0 : heap.pop(); // remove smallest hole
1318 0 : }
1319 0 : }
1320 4128020 : }
1321 56 : }
1322 4128076 : prev = Some(next_key.next());
1323 : }
1324 56 : let mut holes = heap.into_vec();
1325 56 : holes.sort_unstable_by_key(|hole| hole.key_range.start);
1326 56 : holes
1327 56 : };
1328 56 : stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
1329 56 : drop_rlock(guard);
1330 56 :
1331 56 : if self.cancel.is_cancelled() {
1332 0 : return Err(CompactionError::ShuttingDown);
1333 56 : }
1334 56 :
1335 56 : stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
1336 :
1337 : // This iterator walks through all key-value pairs from all the layers
1338 : // we're compacting, in key, LSN order.
1339 : // If there's both a Value::Image and Value::WalRecord for the same (key,lsn),
1340 : // then the Value::Image is ordered before Value::WalRecord.
1341 56 : let mut all_values_iter = {
1342 56 : let mut deltas = Vec::with_capacity(deltas_to_compact.len());
1343 804 : for l in deltas_to_compact.iter() {
1344 804 : let l = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
1345 804 : deltas.push(l);
1346 : }
1347 56 : MergeIterator::create(&deltas, &[], ctx)
1348 56 : };
1349 56 :
1350 56 : // This iterator walks through all keys and is needed to calculate size used by each key
1351 56 : let mut all_keys_iter = all_keys
1352 56 : .iter()
1353 4128076 : .map(|DeltaEntry { key, lsn, size, .. }| (*key, *lsn, *size))
1354 4128020 : .coalesce(|mut prev, cur| {
1355 4128020 : // Coalesce keys that belong to the same key pair.
1356 4128020 : // This ensures that compaction doesn't put them
1357 4128020 : // into different layer files.
1358 4128020 : // Still limit this by the target file size,
1359 4128020 : // so that we keep the size of the files in
1360 4128020 : // check.
1361 4128020 : if prev.0 == cur.0 && prev.2 < target_file_size {
1362 80076 : prev.2 += cur.2;
1363 80076 : Ok(prev)
1364 : } else {
1365 4047944 : Err((prev, cur))
1366 : }
1367 4128020 : });
1368 56 :
1369 56 : // Merge the contents of all the input delta layers into a new set
1370 56 : // of delta layers, based on the current partitioning.
1371 56 : //
1372 56 : // We split the new delta layers on the key dimension. We iterate through the key space, and for each key, check if including the next key to the current output layer we're building would cause the layer to become too large. If so, dump the current output layer and start new one.
1373 56 : // It's possible that there is a single key with so many page versions that storing all of them in a single layer file
1374 56 : // would be too large. In that case, we also split on the LSN dimension.
1375 56 : //
1376 56 : // LSN
1377 56 : // ^
1378 56 : // |
1379 56 : // | +-----------+ +--+--+--+--+
1380 56 : // | | | | | | | |
1381 56 : // | +-----------+ | | | | |
1382 56 : // | | | | | | | |
1383 56 : // | +-----------+ ==> | | | | |
1384 56 : // | | | | | | | |
1385 56 : // | +-----------+ | | | | |
1386 56 : // | | | | | | | |
1387 56 : // | +-----------+ +--+--+--+--+
1388 56 : // |
1389 56 : // +--------------> key
1390 56 : //
1391 56 : //
1392 56 : // If one key (X) has a lot of page versions:
1393 56 : //
1394 56 : // LSN
1395 56 : // ^
1396 56 : // | (X)
1397 56 : // | +-----------+ +--+--+--+--+
1398 56 : // | | | | | | | |
1399 56 : // | +-----------+ | | +--+ |
1400 56 : // | | | | | | | |
1401 56 : // | +-----------+ ==> | | | | |
1402 56 : // | | | | | +--+ |
1403 56 : // | +-----------+ | | | | |
1404 56 : // | | | | | | | |
1405 56 : // | +-----------+ +--+--+--+--+
1406 56 : // |
1407 56 : // +--------------> key
1408 56 : // TODO: this actually divides the layers into fixed-size chunks, not
1409 56 : // based on the partitioning.
1410 56 : //
1411 56 : // TODO: we should also opportunistically materialize and
1412 56 : // garbage collect what we can.
1413 56 : let mut new_layers = Vec::new();
1414 56 : let mut prev_key: Option<Key> = None;
1415 56 : let mut writer: Option<DeltaLayerWriter> = None;
1416 56 : let mut key_values_total_size = 0u64;
1417 56 : let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
1418 56 : let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
1419 56 : let mut next_hole = 0; // index of next hole in holes vector
1420 56 :
1421 56 : let mut keys = 0;
1422 :
1423 4128132 : while let Some((key, lsn, value)) = all_values_iter
1424 4128132 : .next()
1425 4128132 : .await
1426 4128132 : .map_err(CompactionError::Other)?
1427 : {
1428 4128076 : keys += 1;
1429 4128076 :
1430 4128076 : if keys % 32_768 == 0 && self.cancel.is_cancelled() {
1431 : // avoid hitting the cancellation token on every key. in benches, we end up
1432 : // shuffling an order of million keys per layer, this means we'll check it
1433 : // around tens of times per layer.
1434 0 : return Err(CompactionError::ShuttingDown);
1435 4128076 : }
1436 4128076 :
1437 4128076 : let same_key = prev_key == Some(key);
1438 4128076 : // We need to check key boundaries once we reach next key or end of layer with the same key
1439 4128076 : if !same_key || lsn == dup_end_lsn {
1440 4048000 : let mut next_key_size = 0u64;
1441 4048000 : let is_dup_layer = dup_end_lsn.is_valid();
1442 4048000 : dup_start_lsn = Lsn::INVALID;
1443 4048000 : if !same_key {
1444 4048000 : dup_end_lsn = Lsn::INVALID;
1445 4048000 : }
1446 : // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
1447 4048000 : for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
1448 4048000 : next_key_size = next_size;
1449 4048000 : if key != next_key {
1450 4047944 : if dup_end_lsn.is_valid() {
1451 0 : // We are writting segment with duplicates:
1452 0 : // place all remaining values of this key in separate segment
1453 0 : dup_start_lsn = dup_end_lsn; // new segments starts where old stops
1454 0 : dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
1455 4047944 : }
1456 4047944 : break;
1457 56 : }
1458 56 : key_values_total_size += next_size;
1459 56 : // Check if it is time to split segment: if total keys size is larger than target file size.
1460 56 : // We need to avoid generation of empty segments if next_size > target_file_size.
1461 56 : if key_values_total_size > target_file_size && lsn != next_lsn {
1462 : // Split key between multiple layers: such layer can contain only single key
1463 0 : dup_start_lsn = if dup_end_lsn.is_valid() {
1464 0 : dup_end_lsn // new segment with duplicates starts where old one stops
1465 : } else {
1466 0 : lsn // start with the first LSN for this key
1467 : };
1468 0 : dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
1469 0 : break;
1470 56 : }
1471 : }
1472 : // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
1473 4048000 : if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
1474 0 : dup_start_lsn = dup_end_lsn;
1475 0 : dup_end_lsn = lsn_range.end;
1476 4048000 : }
1477 4048000 : if writer.is_some() {
1478 4047944 : let written_size = writer.as_mut().unwrap().size();
1479 4047944 : let contains_hole =
1480 4047944 : next_hole < holes.len() && key >= holes[next_hole].key_range.end;
1481 : // check if key cause layer overflow or contains hole...
1482 4047944 : if is_dup_layer
1483 4047944 : || dup_end_lsn.is_valid()
1484 4047944 : || written_size + key_values_total_size > target_file_size
1485 4047384 : || contains_hole
1486 : {
1487 : // ... if so, flush previous layer and prepare to write new one
1488 560 : let (desc, path) = writer
1489 560 : .take()
1490 560 : .unwrap()
1491 560 : .finish(prev_key.unwrap().next(), ctx)
1492 560 : .await
1493 560 : .map_err(CompactionError::Other)?;
1494 560 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1495 560 : .map_err(CompactionError::Other)?;
1496 :
1497 560 : new_layers.push(new_delta);
1498 560 : writer = None;
1499 560 :
1500 560 : if contains_hole {
1501 0 : // skip hole
1502 0 : next_hole += 1;
1503 560 : }
1504 4047384 : }
1505 56 : }
1506 : // Remember size of key value because at next iteration we will access next item
1507 4048000 : key_values_total_size = next_key_size;
1508 80076 : }
1509 4128076 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
1510 0 : Err(CompactionError::Other(anyhow::anyhow!(
1511 0 : "failpoint delta-layer-writer-fail-before-finish"
1512 0 : )))
1513 4128076 : });
1514 :
1515 4128076 : if !self.shard_identity.is_key_disposable(&key) {
1516 4128076 : if writer.is_none() {
1517 616 : if self.cancel.is_cancelled() {
1518 : // to be somewhat responsive to cancellation, check for each new layer
1519 0 : return Err(CompactionError::ShuttingDown);
1520 616 : }
1521 : // Create writer if not initiaized yet
1522 616 : writer = Some(
1523 : DeltaLayerWriter::new(
1524 616 : self.conf,
1525 616 : self.timeline_id,
1526 616 : self.tenant_shard_id,
1527 616 : key,
1528 616 : if dup_end_lsn.is_valid() {
1529 : // this is a layer containing slice of values of the same key
1530 0 : debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
1531 0 : dup_start_lsn..dup_end_lsn
1532 : } else {
1533 616 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
1534 616 : lsn_range.clone()
1535 : },
1536 616 : ctx,
1537 616 : )
1538 616 : .await
1539 616 : .map_err(CompactionError::Other)?,
1540 : );
1541 :
1542 616 : keys = 0;
1543 4127460 : }
1544 :
1545 4128076 : writer
1546 4128076 : .as_mut()
1547 4128076 : .unwrap()
1548 4128076 : .put_value(key, lsn, value, ctx)
1549 4128076 : .await
1550 4128076 : .map_err(CompactionError::Other)?;
1551 : } else {
1552 0 : let owner = self.shard_identity.get_shard_number(&key);
1553 0 :
1554 0 : // This happens after a shard split, when we're compacting an L0 created by our parent shard
1555 0 : debug!("dropping key {key} during compaction (it belongs on shard {owner})");
1556 : }
1557 :
1558 4128076 : if !new_layers.is_empty() {
1559 39572 : fail_point!("after-timeline-compacted-first-L1");
1560 4088504 : }
1561 :
1562 4128076 : prev_key = Some(key);
1563 : }
1564 56 : if let Some(writer) = writer {
1565 56 : let (desc, path) = writer
1566 56 : .finish(prev_key.unwrap().next(), ctx)
1567 56 : .await
1568 56 : .map_err(CompactionError::Other)?;
1569 56 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1570 56 : .map_err(CompactionError::Other)?;
1571 56 : new_layers.push(new_delta);
1572 0 : }
1573 :
1574 : // Sync layers
1575 56 : if !new_layers.is_empty() {
1576 : // Print a warning if the created layer is larger than double the target size
1577 : // Add two pages for potential overhead. This should in theory be already
1578 : // accounted for in the target calculation, but for very small targets,
1579 : // we still might easily hit the limit otherwise.
1580 56 : let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
1581 616 : for layer in new_layers.iter() {
1582 616 : if layer.layer_desc().file_size > warn_limit {
1583 0 : warn!(
1584 : %layer,
1585 0 : "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
1586 : );
1587 616 : }
1588 : }
1589 :
1590 : // The writer.finish() above already did the fsync of the inodes.
1591 : // We just need to fsync the directory in which these inodes are linked,
1592 : // which we know to be the timeline directory.
1593 : //
1594 : // We use fatal_err() below because the after writer.finish() returns with success,
1595 : // the in-memory state of the filesystem already has the layer file in its final place,
1596 : // and subsequent pageserver code could think it's durable while it really isn't.
1597 56 : let timeline_dir = VirtualFile::open(
1598 56 : &self
1599 56 : .conf
1600 56 : .timeline_path(&self.tenant_shard_id, &self.timeline_id),
1601 56 : ctx,
1602 56 : )
1603 56 : .await
1604 56 : .fatal_err("VirtualFile::open for timeline dir fsync");
1605 56 : timeline_dir
1606 56 : .sync_all()
1607 56 : .await
1608 56 : .fatal_err("VirtualFile::sync_all timeline dir");
1609 0 : }
1610 :
1611 56 : stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
1612 56 : stats.new_deltas_count = Some(new_layers.len());
1613 616 : stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
1614 56 :
1615 56 : match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
1616 56 : .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
1617 : {
1618 56 : Ok(stats_json) => {
1619 56 : info!(
1620 0 : stats_json = stats_json.as_str(),
1621 0 : "compact_level0_phase1 stats available"
1622 : )
1623 : }
1624 0 : Err(e) => {
1625 0 : warn!("compact_level0_phase1 stats failed to serialize: {:#}", e);
1626 : }
1627 : }
1628 :
1629 : // Without this, rustc complains about deltas_to_compact still
1630 : // being borrowed when we `.into_iter()` below.
1631 56 : drop(all_values_iter);
1632 56 :
1633 56 : Ok(CompactLevel0Phase1Result {
1634 56 : new_layers,
1635 56 : deltas_to_compact: deltas_to_compact
1636 56 : .into_iter()
1637 804 : .map(|x| x.drop_eviction_guard())
1638 56 : .collect::<Vec<_>>(),
1639 56 : outcome: if fully_compacted {
1640 56 : CompactionOutcome::Done
1641 : } else {
1642 0 : CompactionOutcome::Pending
1643 : },
1644 : })
1645 728 : }
1646 : }
1647 :
1648 : #[derive(Default)]
1649 : struct CompactLevel0Phase1Result {
1650 : new_layers: Vec<ResidentLayer>,
1651 : deltas_to_compact: Vec<Layer>,
1652 : // Whether we have included all L0 layers, or selected only part of them due to the
1653 : // L0 compaction size limit.
1654 : outcome: CompactionOutcome,
1655 : }
1656 :
1657 : #[derive(Default)]
1658 : struct CompactLevel0Phase1StatsBuilder {
1659 : version: Option<u64>,
1660 : tenant_id: Option<TenantShardId>,
1661 : timeline_id: Option<TimelineId>,
1662 : read_lock_acquisition_micros: DurationRecorder,
1663 : read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
1664 : read_lock_held_key_sort_micros: DurationRecorder,
1665 : read_lock_held_prerequisites_micros: DurationRecorder,
1666 : read_lock_held_compute_holes_micros: DurationRecorder,
1667 : read_lock_drop_micros: DurationRecorder,
1668 : write_layer_files_micros: DurationRecorder,
1669 : level0_deltas_count: Option<usize>,
1670 : new_deltas_count: Option<usize>,
1671 : new_deltas_size: Option<u64>,
1672 : }
1673 :
1674 : #[derive(serde::Serialize)]
1675 : struct CompactLevel0Phase1Stats {
1676 : version: u64,
1677 : tenant_id: TenantShardId,
1678 : timeline_id: TimelineId,
1679 : read_lock_acquisition_micros: RecordedDuration,
1680 : read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
1681 : read_lock_held_key_sort_micros: RecordedDuration,
1682 : read_lock_held_prerequisites_micros: RecordedDuration,
1683 : read_lock_held_compute_holes_micros: RecordedDuration,
1684 : read_lock_drop_micros: RecordedDuration,
1685 : write_layer_files_micros: RecordedDuration,
1686 : level0_deltas_count: usize,
1687 : new_deltas_count: usize,
1688 : new_deltas_size: u64,
1689 : }
1690 :
1691 : impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
1692 : type Error = anyhow::Error;
1693 :
1694 56 : fn try_from(value: CompactLevel0Phase1StatsBuilder) -> Result<Self, Self::Error> {
1695 56 : Ok(Self {
1696 56 : version: value.version.ok_or_else(|| anyhow!("version not set"))?,
1697 56 : tenant_id: value
1698 56 : .tenant_id
1699 56 : .ok_or_else(|| anyhow!("tenant_id not set"))?,
1700 56 : timeline_id: value
1701 56 : .timeline_id
1702 56 : .ok_or_else(|| anyhow!("timeline_id not set"))?,
1703 56 : read_lock_acquisition_micros: value
1704 56 : .read_lock_acquisition_micros
1705 56 : .into_recorded()
1706 56 : .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
1707 56 : read_lock_held_spawn_blocking_startup_micros: value
1708 56 : .read_lock_held_spawn_blocking_startup_micros
1709 56 : .into_recorded()
1710 56 : .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
1711 56 : read_lock_held_key_sort_micros: value
1712 56 : .read_lock_held_key_sort_micros
1713 56 : .into_recorded()
1714 56 : .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
1715 56 : read_lock_held_prerequisites_micros: value
1716 56 : .read_lock_held_prerequisites_micros
1717 56 : .into_recorded()
1718 56 : .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
1719 56 : read_lock_held_compute_holes_micros: value
1720 56 : .read_lock_held_compute_holes_micros
1721 56 : .into_recorded()
1722 56 : .ok_or_else(|| anyhow!("read_lock_held_compute_holes_micros not set"))?,
1723 56 : read_lock_drop_micros: value
1724 56 : .read_lock_drop_micros
1725 56 : .into_recorded()
1726 56 : .ok_or_else(|| anyhow!("read_lock_drop_micros not set"))?,
1727 56 : write_layer_files_micros: value
1728 56 : .write_layer_files_micros
1729 56 : .into_recorded()
1730 56 : .ok_or_else(|| anyhow!("write_layer_files_micros not set"))?,
1731 56 : level0_deltas_count: value
1732 56 : .level0_deltas_count
1733 56 : .ok_or_else(|| anyhow!("level0_deltas_count not set"))?,
1734 56 : new_deltas_count: value
1735 56 : .new_deltas_count
1736 56 : .ok_or_else(|| anyhow!("new_deltas_count not set"))?,
1737 56 : new_deltas_size: value
1738 56 : .new_deltas_size
1739 56 : .ok_or_else(|| anyhow!("new_deltas_size not set"))?,
1740 : })
1741 56 : }
1742 : }
1743 :
1744 : impl Timeline {
1745 : /// Entry point for new tiered compaction algorithm.
1746 : ///
1747 : /// All the real work is in the implementation in the pageserver_compaction
1748 : /// crate. The code here would apply to any algorithm implemented by the
1749 : /// same interface, but tiered is the only one at the moment.
1750 : ///
1751 : /// TODO: cancellation
1752 0 : pub(crate) async fn compact_tiered(
1753 0 : self: &Arc<Self>,
1754 0 : _cancel: &CancellationToken,
1755 0 : ctx: &RequestContext,
1756 0 : ) -> Result<(), CompactionError> {
1757 0 : let fanout = self.get_compaction_threshold() as u64;
1758 0 : let target_file_size = self.get_checkpoint_distance();
1759 :
1760 : // Find the top of the historical layers
1761 0 : let end_lsn = {
1762 0 : let guard = self.layers.read().await;
1763 0 : let layers = guard.layer_map()?;
1764 :
1765 0 : let l0_deltas = layers.level0_deltas();
1766 0 :
1767 0 : // As an optimization, if we find that there are too few L0 layers,
1768 0 : // bail out early. We know that the compaction algorithm would do
1769 0 : // nothing in that case.
1770 0 : if l0_deltas.len() < fanout as usize {
1771 : // doesn't need compacting
1772 0 : return Ok(());
1773 0 : }
1774 0 : l0_deltas.iter().map(|l| l.lsn_range.end).max().unwrap()
1775 0 : };
1776 0 :
1777 0 : // Is the timeline being deleted?
1778 0 : if self.is_stopping() {
1779 0 : trace!("Dropping out of compaction on timeline shutdown");
1780 0 : return Err(CompactionError::ShuttingDown);
1781 0 : }
1782 :
1783 0 : let (dense_ks, _sparse_ks) = self.collect_keyspace(end_lsn, ctx).await?;
1784 : // TODO(chi): ignore sparse_keyspace for now, compact it in the future.
1785 0 : let mut adaptor = TimelineAdaptor::new(self, (end_lsn, dense_ks));
1786 0 :
1787 0 : pageserver_compaction::compact_tiered::compact_tiered(
1788 0 : &mut adaptor,
1789 0 : end_lsn,
1790 0 : target_file_size,
1791 0 : fanout,
1792 0 : ctx,
1793 0 : )
1794 0 : .await
1795 : // TODO: compact_tiered needs to return CompactionError
1796 0 : .map_err(CompactionError::Other)?;
1797 :
1798 0 : adaptor.flush_updates().await?;
1799 0 : Ok(())
1800 0 : }
1801 :
1802 : /// Take a list of images and deltas, produce images and deltas according to GC horizon and retain_lsns.
1803 : ///
1804 : /// It takes a key, the values of the key within the compaction process, a GC horizon, and all retain_lsns below the horizon.
1805 : /// For now, it requires the `accumulated_values` contains the full history of the key (i.e., the key with the lowest LSN is
1806 : /// an image or a WAL not requiring a base image). This restriction will be removed once we implement gc-compaction on branch.
1807 : ///
1808 : /// The function returns the deltas and the base image that need to be placed at each of the retain LSN. For example, we have:
1809 : ///
1810 : /// A@0x10, +B@0x20, +C@0x30, +D@0x40, +E@0x50, +F@0x60
1811 : /// horizon = 0x50, retain_lsn = 0x20, 0x40, delta_threshold=3
1812 : ///
1813 : /// The function will produce:
1814 : ///
1815 : /// ```plain
1816 : /// 0x20(retain_lsn) -> img=AB@0x20 always produce a single image below the lowest retain LSN
1817 : /// 0x40(retain_lsn) -> deltas=[+C@0x30, +D@0x40] two deltas since the last base image, keeping the deltas
1818 : /// 0x50(horizon) -> deltas=[ABCDE@0x50] three deltas since the last base image, generate an image but put it in the delta
1819 : /// above_horizon -> deltas=[+F@0x60] full history above the horizon
1820 : /// ```
1821 : ///
1822 : /// Note that `accumulated_values` must be sorted by LSN and should belong to a single key.
1823 1260 : pub(crate) async fn generate_key_retention(
1824 1260 : self: &Arc<Timeline>,
1825 1260 : key: Key,
1826 1260 : full_history: &[(Key, Lsn, Value)],
1827 1260 : horizon: Lsn,
1828 1260 : retain_lsn_below_horizon: &[Lsn],
1829 1260 : delta_threshold_cnt: usize,
1830 1260 : base_img_from_ancestor: Option<(Key, Lsn, Bytes)>,
1831 1260 : ) -> anyhow::Result<KeyHistoryRetention> {
1832 : // Pre-checks for the invariants
1833 :
1834 1260 : let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
1835 :
1836 1260 : if debug_mode {
1837 3060 : for (log_key, _, _) in full_history {
1838 1800 : assert_eq!(log_key, &key, "mismatched key");
1839 : }
1840 1260 : for i in 1..full_history.len() {
1841 540 : assert!(full_history[i - 1].1 <= full_history[i].1, "unordered LSN");
1842 540 : if full_history[i - 1].1 == full_history[i].1 {
1843 0 : assert!(
1844 0 : matches!(full_history[i - 1].2, Value::Image(_)),
1845 0 : "unordered delta/image, or duplicated delta"
1846 : );
1847 540 : }
1848 : }
1849 : // There was an assertion for no base image that checks if the first
1850 : // record in the history is `will_init` before, but it was removed.
1851 : // This is explained in the test cases for generate_key_retention.
1852 : // Search "incomplete history" for more information.
1853 2820 : for lsn in retain_lsn_below_horizon {
1854 1560 : assert!(lsn < &horizon, "retain lsn must be below horizon")
1855 : }
1856 1260 : for i in 1..retain_lsn_below_horizon.len() {
1857 712 : assert!(
1858 712 : retain_lsn_below_horizon[i - 1] <= retain_lsn_below_horizon[i],
1859 0 : "unordered LSN"
1860 : );
1861 : }
1862 0 : }
1863 1260 : let has_ancestor = base_img_from_ancestor.is_some();
1864 : // Step 1: split history into len(retain_lsn_below_horizon) + 2 buckets, where the last bucket is for all deltas above the horizon,
1865 : // and the second-to-last bucket is for the horizon. Each bucket contains lsn_last_bucket < deltas <= lsn_this_bucket.
1866 1260 : let (mut split_history, lsn_split_points) = {
1867 1260 : let mut split_history = Vec::new();
1868 1260 : split_history.resize_with(retain_lsn_below_horizon.len() + 2, Vec::new);
1869 1260 : let mut lsn_split_points = Vec::with_capacity(retain_lsn_below_horizon.len() + 1);
1870 2820 : for lsn in retain_lsn_below_horizon {
1871 1560 : lsn_split_points.push(*lsn);
1872 1560 : }
1873 1260 : lsn_split_points.push(horizon);
1874 1260 : let mut current_idx = 0;
1875 3060 : for item @ (_, lsn, _) in full_history {
1876 2288 : while current_idx < lsn_split_points.len() && *lsn > lsn_split_points[current_idx] {
1877 488 : current_idx += 1;
1878 488 : }
1879 1800 : split_history[current_idx].push(item);
1880 : }
1881 1260 : (split_history, lsn_split_points)
1882 : };
1883 : // Step 2: filter out duplicated records due to the k-merge of image/delta layers
1884 5340 : for split_for_lsn in &mut split_history {
1885 4080 : let mut prev_lsn = None;
1886 4080 : let mut new_split_for_lsn = Vec::with_capacity(split_for_lsn.len());
1887 4080 : for record @ (_, lsn, _) in std::mem::take(split_for_lsn) {
1888 1800 : if let Some(prev_lsn) = &prev_lsn {
1889 236 : if *prev_lsn == lsn {
1890 : // The case that we have an LSN with both data from the delta layer and the image layer. As
1891 : // `ValueWrapper` ensures that an image is ordered before a delta at the same LSN, we simply
1892 : // drop this delta and keep the image.
1893 : //
1894 : // For example, we have delta layer key1@0x10, key1@0x20, and image layer key1@0x10, we will
1895 : // keep the image for key1@0x10 and the delta for key1@0x20. key1@0x10 delta will be simply
1896 : // dropped.
1897 : //
1898 : // TODO: in case we have both delta + images for a given LSN and it does not exceed the delta
1899 : // threshold, we could have kept delta instead to save space. This is an optimization for the future.
1900 0 : continue;
1901 236 : }
1902 1564 : }
1903 1800 : prev_lsn = Some(lsn);
1904 1800 : new_split_for_lsn.push(record);
1905 : }
1906 4080 : *split_for_lsn = new_split_for_lsn;
1907 : }
1908 : // Step 3: generate images when necessary
1909 1260 : let mut retention = Vec::with_capacity(split_history.len());
1910 1260 : let mut records_since_last_image = 0;
1911 1260 : let batch_cnt = split_history.len();
1912 1260 : assert!(
1913 1260 : batch_cnt >= 2,
1914 0 : "should have at least below + above horizon batches"
1915 : );
1916 1260 : let mut replay_history: Vec<(Key, Lsn, Value)> = Vec::new();
1917 1260 : if let Some((key, lsn, img)) = base_img_from_ancestor {
1918 84 : replay_history.push((key, lsn, Value::Image(img)));
1919 1176 : }
1920 :
1921 : /// Generate debug information for the replay history
1922 0 : fn generate_history_trace(replay_history: &[(Key, Lsn, Value)]) -> String {
1923 : use std::fmt::Write;
1924 0 : let mut output = String::new();
1925 0 : if let Some((key, _, _)) = replay_history.first() {
1926 0 : write!(output, "key={} ", key).unwrap();
1927 0 : let mut cnt = 0;
1928 0 : for (_, lsn, val) in replay_history {
1929 0 : if val.is_image() {
1930 0 : write!(output, "i@{} ", lsn).unwrap();
1931 0 : } else if val.will_init() {
1932 0 : write!(output, "di@{} ", lsn).unwrap();
1933 0 : } else {
1934 0 : write!(output, "d@{} ", lsn).unwrap();
1935 0 : }
1936 0 : cnt += 1;
1937 0 : if cnt >= 128 {
1938 0 : write!(output, "... and more").unwrap();
1939 0 : break;
1940 0 : }
1941 : }
1942 0 : } else {
1943 0 : write!(output, "<no history>").unwrap();
1944 0 : }
1945 0 : output
1946 0 : }
1947 :
1948 0 : fn generate_debug_trace(
1949 0 : replay_history: Option<&[(Key, Lsn, Value)]>,
1950 0 : full_history: &[(Key, Lsn, Value)],
1951 0 : lsns: &[Lsn],
1952 0 : horizon: Lsn,
1953 0 : ) -> String {
1954 : use std::fmt::Write;
1955 0 : let mut output = String::new();
1956 0 : if let Some(replay_history) = replay_history {
1957 0 : writeln!(
1958 0 : output,
1959 0 : "replay_history: {}",
1960 0 : generate_history_trace(replay_history)
1961 0 : )
1962 0 : .unwrap();
1963 0 : } else {
1964 0 : writeln!(output, "replay_history: <disabled>",).unwrap();
1965 0 : }
1966 0 : writeln!(
1967 0 : output,
1968 0 : "full_history: {}",
1969 0 : generate_history_trace(full_history)
1970 0 : )
1971 0 : .unwrap();
1972 0 : writeln!(
1973 0 : output,
1974 0 : "when processing: [{}] horizon={}",
1975 0 : lsns.iter().map(|l| format!("{l}")).join(","),
1976 0 : horizon
1977 0 : )
1978 0 : .unwrap();
1979 0 : output
1980 0 : }
1981 :
1982 1260 : let mut key_exists = false;
1983 4080 : for (i, split_for_lsn) in split_history.into_iter().enumerate() {
1984 : // TODO: there could be image keys inside the splits, and we can compute records_since_last_image accordingly.
1985 4080 : records_since_last_image += split_for_lsn.len();
1986 : // Whether to produce an image into the final layer files
1987 4080 : let produce_image = if i == 0 && !has_ancestor {
1988 : // We always generate images for the first batch (below horizon / lowest retain_lsn)
1989 1176 : true
1990 2904 : } else if i == batch_cnt - 1 {
1991 : // Do not generate images for the last batch (above horizon)
1992 1260 : false
1993 1644 : } else if records_since_last_image == 0 {
1994 1288 : false
1995 356 : } else if records_since_last_image >= delta_threshold_cnt {
1996 : // Generate images when there are too many records
1997 12 : true
1998 : } else {
1999 344 : false
2000 : };
2001 4080 : replay_history.extend(split_for_lsn.iter().map(|x| (*x).clone()));
2002 : // Only retain the items after the last image record
2003 5028 : for idx in (0..replay_history.len()).rev() {
2004 5028 : if replay_history[idx].2.will_init() {
2005 4080 : replay_history = replay_history[idx..].to_vec();
2006 4080 : break;
2007 948 : }
2008 : }
2009 4080 : if replay_history.is_empty() && !key_exists {
2010 : // The key does not exist at earlier LSN, we can skip this iteration.
2011 0 : retention.push(Vec::new());
2012 0 : continue;
2013 4080 : } else {
2014 4080 : key_exists = true;
2015 4080 : }
2016 4080 : let Some((_, _, val)) = replay_history.first() else {
2017 0 : unreachable!("replay history should not be empty once it exists")
2018 : };
2019 4080 : if !val.will_init() {
2020 0 : return Err(anyhow::anyhow!("invalid history, no base image")).with_context(|| {
2021 0 : generate_debug_trace(
2022 0 : Some(&replay_history),
2023 0 : full_history,
2024 0 : retain_lsn_below_horizon,
2025 0 : horizon,
2026 0 : )
2027 0 : });
2028 4080 : }
2029 : // Whether to reconstruct the image. In debug mode, we will generate an image
2030 : // at every retain_lsn to ensure data is not corrupted, but we won't put the
2031 : // image into the final layer.
2032 4080 : let generate_image = produce_image || debug_mode;
2033 4080 : if produce_image {
2034 1188 : records_since_last_image = 0;
2035 2892 : }
2036 4080 : let img_and_lsn = if generate_image {
2037 4080 : let replay_history_for_debug = if debug_mode {
2038 4080 : Some(replay_history.clone())
2039 : } else {
2040 0 : None
2041 : };
2042 4080 : let replay_history_for_debug_ref = replay_history_for_debug.as_deref();
2043 4080 : let history = if produce_image {
2044 1188 : std::mem::take(&mut replay_history)
2045 : } else {
2046 2892 : replay_history.clone()
2047 : };
2048 4080 : let mut img = None;
2049 4080 : let mut records = Vec::with_capacity(history.len());
2050 4080 : if let (_, lsn, Value::Image(val)) = history.first().as_ref().unwrap() {
2051 4036 : img = Some((*lsn, val.clone()));
2052 4036 : for (_, lsn, val) in history.into_iter().skip(1) {
2053 920 : let Value::WalRecord(rec) = val else {
2054 0 : return Err(anyhow::anyhow!(
2055 0 : "invalid record, first record is image, expect walrecords"
2056 0 : ))
2057 0 : .with_context(|| {
2058 0 : generate_debug_trace(
2059 0 : replay_history_for_debug_ref,
2060 0 : full_history,
2061 0 : retain_lsn_below_horizon,
2062 0 : horizon,
2063 0 : )
2064 0 : });
2065 : };
2066 920 : records.push((lsn, rec));
2067 : }
2068 : } else {
2069 72 : for (_, lsn, val) in history.into_iter() {
2070 72 : let Value::WalRecord(rec) = val else {
2071 0 : return Err(anyhow::anyhow!("invalid record, first record is walrecord, expect rest are walrecord"))
2072 0 : .with_context(|| generate_debug_trace(
2073 0 : replay_history_for_debug_ref,
2074 0 : full_history,
2075 0 : retain_lsn_below_horizon,
2076 0 : horizon,
2077 0 : ));
2078 : };
2079 72 : records.push((lsn, rec));
2080 : }
2081 : }
2082 4080 : records.reverse();
2083 4080 : let state = ValueReconstructState { img, records };
2084 : // last batch does not generate image so i is always in range, unless we force generate
2085 : // an image during testing
2086 4080 : let request_lsn = if i >= lsn_split_points.len() {
2087 1260 : Lsn::MAX
2088 : } else {
2089 2820 : lsn_split_points[i]
2090 : };
2091 4080 : let img = self.reconstruct_value(key, request_lsn, state).await?;
2092 4080 : Some((request_lsn, img))
2093 : } else {
2094 0 : None
2095 : };
2096 4080 : if produce_image {
2097 1188 : let (request_lsn, img) = img_and_lsn.unwrap();
2098 1188 : replay_history.push((key, request_lsn, Value::Image(img.clone())));
2099 1188 : retention.push(vec![(request_lsn, Value::Image(img))]);
2100 2892 : } else {
2101 2892 : let deltas = split_for_lsn
2102 2892 : .iter()
2103 2892 : .map(|(_, lsn, value)| (*lsn, value.clone()))
2104 2892 : .collect_vec();
2105 2892 : retention.push(deltas);
2106 2892 : }
2107 : }
2108 1260 : let mut result = Vec::with_capacity(retention.len());
2109 1260 : assert_eq!(retention.len(), lsn_split_points.len() + 1);
2110 4080 : for (idx, logs) in retention.into_iter().enumerate() {
2111 4080 : if idx == lsn_split_points.len() {
2112 1260 : return Ok(KeyHistoryRetention {
2113 1260 : below_horizon: result,
2114 1260 : above_horizon: KeyLogAtLsn(logs),
2115 1260 : });
2116 2820 : } else {
2117 2820 : result.push((lsn_split_points[idx], KeyLogAtLsn(logs)));
2118 2820 : }
2119 : }
2120 0 : unreachable!("key retention is empty")
2121 1260 : }
2122 :
2123 : /// Check how much space is left on the disk
2124 104 : async fn check_available_space(self: &Arc<Self>) -> anyhow::Result<u64> {
2125 104 : let tenants_dir = self.conf.tenants_path();
2126 :
2127 104 : let stat = Statvfs::get(&tenants_dir, None)
2128 104 : .context("statvfs failed, presumably directory got unlinked")?;
2129 :
2130 104 : let (avail_bytes, _) = stat.get_avail_total_bytes();
2131 104 :
2132 104 : Ok(avail_bytes)
2133 104 : }
2134 :
2135 : /// Check if the compaction can proceed safely without running out of space. We assume the size
2136 : /// upper bound of the produced files of a compaction job is the same as all layers involved in
2137 : /// the compaction. Therefore, we need `2 * layers_to_be_compacted_size` at least to do a
2138 : /// compaction.
2139 104 : async fn check_compaction_space(
2140 104 : self: &Arc<Self>,
2141 104 : layer_selection: &[Layer],
2142 104 : ) -> anyhow::Result<()> {
2143 104 : let available_space = self.check_available_space().await?;
2144 104 : let mut remote_layer_size = 0;
2145 104 : let mut all_layer_size = 0;
2146 408 : for layer in layer_selection {
2147 304 : let needs_download = layer.needs_download().await?;
2148 304 : if needs_download.is_some() {
2149 0 : remote_layer_size += layer.layer_desc().file_size;
2150 304 : }
2151 304 : all_layer_size += layer.layer_desc().file_size;
2152 : }
2153 104 : let allocated_space = (available_space as f64 * 0.8) as u64; /* reserve 20% space for other tasks */
2154 104 : if all_layer_size /* space needed for newly-generated file */ + remote_layer_size /* space for downloading layers */ > allocated_space
2155 : {
2156 0 : return Err(anyhow!("not enough space for compaction: available_space={}, allocated_space={}, all_layer_size={}, remote_layer_size={}, required_space={}",
2157 0 : available_space, allocated_space, all_layer_size, remote_layer_size, all_layer_size + remote_layer_size));
2158 104 : }
2159 104 : Ok(())
2160 104 : }
2161 :
2162 : /// Get a watermark for gc-compaction, that is the lowest LSN that we can use as the `gc_horizon` for
2163 : /// the compaction algorithm. It is min(space_cutoff, time_cutoff, latest_gc_cutoff, standby_horizon).
2164 : /// Leases and retain_lsns are considered in the gc-compaction job itself so we don't need to account for them
2165 : /// here.
2166 108 : pub(crate) fn get_gc_compaction_watermark(self: &Arc<Self>) -> Lsn {
2167 108 : let gc_cutoff_lsn = {
2168 108 : let gc_info = self.gc_info.read().unwrap();
2169 108 : gc_info.min_cutoff()
2170 108 : };
2171 108 :
2172 108 : // TODO: standby horizon should use leases so we don't really need to consider it here.
2173 108 : // let watermark = watermark.min(self.standby_horizon.load());
2174 108 :
2175 108 : // TODO: ensure the child branches will not use anything below the watermark, or consider
2176 108 : // them when computing the watermark.
2177 108 : gc_cutoff_lsn.min(*self.get_applied_gc_cutoff_lsn())
2178 108 : }
2179 :
2180 : /// Split a gc-compaction job into multiple compaction jobs. The split is based on the key range and the estimated size of the compaction job.
2181 : /// The function returns a list of compaction jobs that can be executed separately. If the upper bound of the compact LSN
2182 : /// range is not specified, we will use the latest gc_cutoff as the upper bound, so that all jobs in the jobset acts
2183 : /// like a full compaction of the specified keyspace.
2184 0 : pub(crate) async fn gc_compaction_split_jobs(
2185 0 : self: &Arc<Self>,
2186 0 : job: GcCompactJob,
2187 0 : sub_compaction_max_job_size_mb: Option<u64>,
2188 0 : ) -> anyhow::Result<Vec<GcCompactJob>> {
2189 0 : let compact_below_lsn = if job.compact_lsn_range.end != Lsn::MAX {
2190 0 : job.compact_lsn_range.end
2191 : } else {
2192 0 : self.get_gc_compaction_watermark()
2193 : };
2194 :
2195 0 : if compact_below_lsn == Lsn::INVALID {
2196 0 : tracing::warn!("no layers to compact with gc: gc_cutoff not generated yet, skipping gc bottom-most compaction");
2197 0 : return Ok(vec![]);
2198 0 : }
2199 :
2200 : // Split compaction job to about 4GB each
2201 : const GC_COMPACT_MAX_SIZE_MB: u64 = 4 * 1024;
2202 0 : let sub_compaction_max_job_size_mb =
2203 0 : sub_compaction_max_job_size_mb.unwrap_or(GC_COMPACT_MAX_SIZE_MB);
2204 0 :
2205 0 : let mut compact_jobs = Vec::new();
2206 0 : // For now, we simply use the key partitioning information; we should do a more fine-grained partitioning
2207 0 : // by estimating the amount of files read for a compaction job. We should also partition on LSN.
2208 0 : let ((dense_ks, sparse_ks), _) = self.partitioning.read().as_ref().clone();
2209 : // Truncate the key range to be within user specified compaction range.
2210 0 : fn truncate_to(
2211 0 : source_start: &Key,
2212 0 : source_end: &Key,
2213 0 : target_start: &Key,
2214 0 : target_end: &Key,
2215 0 : ) -> Option<(Key, Key)> {
2216 0 : let start = source_start.max(target_start);
2217 0 : let end = source_end.min(target_end);
2218 0 : if start < end {
2219 0 : Some((*start, *end))
2220 : } else {
2221 0 : None
2222 : }
2223 0 : }
2224 0 : let mut split_key_ranges = Vec::new();
2225 0 : let ranges = dense_ks
2226 0 : .parts
2227 0 : .iter()
2228 0 : .map(|partition| partition.ranges.iter())
2229 0 : .chain(sparse_ks.parts.iter().map(|x| x.0.ranges.iter()))
2230 0 : .flatten()
2231 0 : .cloned()
2232 0 : .collect_vec();
2233 0 : for range in ranges.iter() {
2234 0 : let Some((start, end)) = truncate_to(
2235 0 : &range.start,
2236 0 : &range.end,
2237 0 : &job.compact_key_range.start,
2238 0 : &job.compact_key_range.end,
2239 0 : ) else {
2240 0 : continue;
2241 : };
2242 0 : split_key_ranges.push((start, end));
2243 : }
2244 0 : split_key_ranges.sort();
2245 0 : let all_layers = {
2246 0 : let guard = self.layers.read().await;
2247 0 : let layer_map = guard.layer_map()?;
2248 0 : layer_map.iter_historic_layers().collect_vec()
2249 0 : };
2250 0 : let mut current_start = None;
2251 0 : let ranges_num = split_key_ranges.len();
2252 0 : for (idx, (start, end)) in split_key_ranges.into_iter().enumerate() {
2253 0 : if current_start.is_none() {
2254 0 : current_start = Some(start);
2255 0 : }
2256 0 : let start = current_start.unwrap();
2257 0 : if start >= end {
2258 : // We have already processed this partition.
2259 0 : continue;
2260 0 : }
2261 0 : let overlapping_layers = {
2262 0 : let mut desc = Vec::new();
2263 0 : for layer in all_layers.iter() {
2264 0 : if overlaps_with(&layer.get_key_range(), &(start..end))
2265 0 : && layer.get_lsn_range().start <= compact_below_lsn
2266 0 : {
2267 0 : desc.push(layer.clone());
2268 0 : }
2269 : }
2270 0 : desc
2271 0 : };
2272 0 : let total_size = overlapping_layers.iter().map(|x| x.file_size).sum::<u64>();
2273 0 : if total_size > sub_compaction_max_job_size_mb * 1024 * 1024 || ranges_num == idx + 1 {
2274 : // Try to extend the compaction range so that we include at least one full layer file.
2275 0 : let extended_end = overlapping_layers
2276 0 : .iter()
2277 0 : .map(|layer| layer.key_range.end)
2278 0 : .min();
2279 : // It is possible that the search range does not contain any layer files when we reach the end of the loop.
2280 : // In this case, we simply use the specified key range end.
2281 0 : let end = if let Some(extended_end) = extended_end {
2282 0 : extended_end.max(end)
2283 : } else {
2284 0 : end
2285 : };
2286 0 : let end = if ranges_num == idx + 1 {
2287 : // extend the compaction range to the end of the key range if it's the last partition
2288 0 : end.max(job.compact_key_range.end)
2289 : } else {
2290 0 : end
2291 : };
2292 0 : info!(
2293 0 : "splitting compaction job: {}..{}, estimated_size={}",
2294 : start, end, total_size
2295 : );
2296 0 : compact_jobs.push(GcCompactJob {
2297 0 : dry_run: job.dry_run,
2298 0 : compact_key_range: start..end,
2299 0 : compact_lsn_range: job.compact_lsn_range.start..compact_below_lsn,
2300 0 : });
2301 0 : current_start = Some(end);
2302 0 : }
2303 : }
2304 0 : Ok(compact_jobs)
2305 0 : }
2306 :
2307 : /// An experimental compaction building block that combines compaction with garbage collection.
2308 : ///
2309 : /// The current implementation picks all delta + image layers that are below or intersecting with
2310 : /// the GC horizon without considering retain_lsns. Then, it does a full compaction over all these delta
2311 : /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
2312 : /// and create delta layers with all deltas >= gc horizon.
2313 : ///
2314 : /// If `options.compact_range` is provided, it will only compact the keys within the range, aka partial compaction.
2315 : /// Partial compaction will read and process all layers overlapping with the key range, even if it might
2316 : /// contain extra keys. After the gc-compaction phase completes, delta layers that are not fully contained
2317 : /// within the key range will be rewritten to ensure they do not overlap with the delta layers. Providing
2318 : /// Key::MIN..Key..MAX to the function indicates a full compaction, though technically, `Key::MAX` is not
2319 : /// part of the range.
2320 : ///
2321 : /// If `options.compact_lsn_range.end` is provided, the compaction will only compact layers below or intersect with
2322 : /// the LSN. Otherwise, it will use the gc cutoff by default.
2323 108 : pub(crate) async fn compact_with_gc(
2324 108 : self: &Arc<Self>,
2325 108 : cancel: &CancellationToken,
2326 108 : options: CompactOptions,
2327 108 : ctx: &RequestContext,
2328 108 : ) -> anyhow::Result<()> {
2329 108 : let sub_compaction = options.sub_compaction;
2330 108 : let job = GcCompactJob::from_compact_options(options.clone());
2331 108 : if sub_compaction {
2332 0 : info!("running enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
2333 0 : let jobs = self
2334 0 : .gc_compaction_split_jobs(job, options.sub_compaction_max_job_size_mb)
2335 0 : .await?;
2336 0 : let jobs_len = jobs.len();
2337 0 : for (idx, job) in jobs.into_iter().enumerate() {
2338 0 : info!(
2339 0 : "running enhanced gc bottom-most compaction, sub-compaction {}/{}",
2340 0 : idx + 1,
2341 : jobs_len
2342 : );
2343 0 : self.compact_with_gc_inner(cancel, job, ctx).await?;
2344 : }
2345 0 : if jobs_len == 0 {
2346 0 : info!("no jobs to run, skipping gc bottom-most compaction");
2347 0 : }
2348 0 : return Ok(());
2349 108 : }
2350 108 : self.compact_with_gc_inner(cancel, job, ctx).await
2351 108 : }
2352 :
2353 108 : async fn compact_with_gc_inner(
2354 108 : self: &Arc<Self>,
2355 108 : cancel: &CancellationToken,
2356 108 : job: GcCompactJob,
2357 108 : ctx: &RequestContext,
2358 108 : ) -> anyhow::Result<()> {
2359 108 : // Block other compaction/GC tasks from running for now. GC-compaction could run along
2360 108 : // with legacy compaction tasks in the future. Always ensure the lock order is compaction -> gc.
2361 108 : // Note that we already acquired the compaction lock when the outer `compact` function gets called.
2362 108 :
2363 108 : let gc_lock = async {
2364 108 : tokio::select! {
2365 108 : guard = self.gc_lock.lock() => Ok(guard),
2366 : // TODO: refactor to CompactionError to correctly pass cancelled error
2367 108 : _ = cancel.cancelled() => Err(anyhow!("cancelled")),
2368 : }
2369 108 : };
2370 :
2371 108 : let gc_lock = crate::timed(
2372 108 : gc_lock,
2373 108 : "acquires gc lock",
2374 108 : std::time::Duration::from_secs(5),
2375 108 : )
2376 108 : .await?;
2377 :
2378 108 : let dry_run = job.dry_run;
2379 108 : let compact_key_range = job.compact_key_range;
2380 108 : let compact_lsn_range = job.compact_lsn_range;
2381 :
2382 108 : let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
2383 :
2384 108 : info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compact_key_range={}..{}, compact_lsn_range={}..{}", compact_key_range.start, compact_key_range.end, compact_lsn_range.start, compact_lsn_range.end);
2385 :
2386 108 : scopeguard::defer! {
2387 108 : info!("done enhanced gc bottom-most compaction");
2388 108 : };
2389 108 :
2390 108 : let mut stat = CompactionStatistics::default();
2391 :
2392 : // Step 0: pick all delta layers + image layers below/intersect with the GC horizon.
2393 : // The layer selection has the following properties:
2394 : // 1. If a layer is in the selection, all layers below it are in the selection.
2395 : // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
2396 104 : let job_desc = {
2397 108 : let guard = self.layers.read().await;
2398 108 : let layers = guard.layer_map()?;
2399 108 : let gc_info = self.gc_info.read().unwrap();
2400 108 : let mut retain_lsns_below_horizon = Vec::new();
2401 108 : let gc_cutoff = {
2402 : // Currently, gc-compaction only kicks in after the legacy gc has updated the gc_cutoff.
2403 : // Therefore, it can only clean up data that cannot be cleaned up with legacy gc, instead of
2404 : // cleaning everything that theoritically it could. In the future, it should use `self.gc_info`
2405 : // to get the truth data.
2406 108 : let real_gc_cutoff = self.get_gc_compaction_watermark();
2407 : // The compaction algorithm will keep all keys above the gc_cutoff while keeping only necessary keys below the gc_cutoff for
2408 : // each of the retain_lsn. Therefore, if the user-provided `compact_lsn_range.end` is larger than the real gc cutoff, we will use
2409 : // the real cutoff.
2410 108 : let mut gc_cutoff = if compact_lsn_range.end == Lsn::MAX {
2411 96 : if real_gc_cutoff == Lsn::INVALID {
2412 : // If the gc_cutoff is not generated yet, we should not compact anything.
2413 0 : tracing::warn!("no layers to compact with gc: gc_cutoff not generated yet, skipping gc bottom-most compaction");
2414 0 : return Ok(());
2415 96 : }
2416 96 : real_gc_cutoff
2417 : } else {
2418 12 : compact_lsn_range.end
2419 : };
2420 108 : if gc_cutoff > real_gc_cutoff {
2421 8 : warn!("provided compact_lsn_range.end={} is larger than the real_gc_cutoff={}, using the real gc cutoff", gc_cutoff, real_gc_cutoff);
2422 8 : gc_cutoff = real_gc_cutoff;
2423 100 : }
2424 108 : gc_cutoff
2425 : };
2426 140 : for (lsn, _timeline_id, _is_offloaded) in &gc_info.retain_lsns {
2427 140 : if lsn < &gc_cutoff {
2428 140 : retain_lsns_below_horizon.push(*lsn);
2429 140 : }
2430 : }
2431 108 : for lsn in gc_info.leases.keys() {
2432 0 : if lsn < &gc_cutoff {
2433 0 : retain_lsns_below_horizon.push(*lsn);
2434 0 : }
2435 : }
2436 108 : let mut selected_layers: Vec<Layer> = Vec::new();
2437 108 : drop(gc_info);
2438 : // Firstly, pick all the layers intersect or below the gc_cutoff, get the largest LSN in the selected layers.
2439 108 : let Some(max_layer_lsn) = layers
2440 108 : .iter_historic_layers()
2441 488 : .filter(|desc| desc.get_lsn_range().start <= gc_cutoff)
2442 416 : .map(|desc| desc.get_lsn_range().end)
2443 108 : .max()
2444 : else {
2445 0 : info!("no layers to compact with gc: no historic layers below gc_cutoff, gc_cutoff={}", gc_cutoff);
2446 0 : return Ok(());
2447 : };
2448 : // Next, if the user specifies compact_lsn_range.start, we need to filter some layers out. All the layers (strictly) below
2449 : // the min_layer_lsn computed as below will be filtered out and the data will be accessed using the normal read path, as if
2450 : // it is a branch.
2451 108 : let Some(min_layer_lsn) = layers
2452 108 : .iter_historic_layers()
2453 488 : .filter(|desc| {
2454 488 : if compact_lsn_range.start == Lsn::INVALID {
2455 396 : true // select all layers below if start == Lsn(0)
2456 : } else {
2457 92 : desc.get_lsn_range().end > compact_lsn_range.start // strictly larger than compact_above_lsn
2458 : }
2459 488 : })
2460 452 : .map(|desc| desc.get_lsn_range().start)
2461 108 : .min()
2462 : else {
2463 0 : info!("no layers to compact with gc: no historic layers above compact_above_lsn, compact_above_lsn={}", compact_lsn_range.end);
2464 0 : return Ok(());
2465 : };
2466 : // Then, pick all the layers that are below the max_layer_lsn. This is to ensure we can pick all single-key
2467 : // layers to compact.
2468 108 : let mut rewrite_layers = Vec::new();
2469 488 : for desc in layers.iter_historic_layers() {
2470 488 : if desc.get_lsn_range().end <= max_layer_lsn
2471 416 : && desc.get_lsn_range().start >= min_layer_lsn
2472 380 : && overlaps_with(&desc.get_key_range(), &compact_key_range)
2473 : {
2474 : // If the layer overlaps with the compaction key range, we need to read it to obtain all keys within the range,
2475 : // even if it might contain extra keys
2476 304 : selected_layers.push(guard.get_from_desc(&desc));
2477 304 : // If the layer is not fully contained within the key range, we need to rewrite it if it's a delta layer (it's fine
2478 304 : // to overlap image layers)
2479 304 : if desc.is_delta() && !fully_contains(&compact_key_range, &desc.get_key_range())
2480 4 : {
2481 4 : rewrite_layers.push(desc);
2482 300 : }
2483 184 : }
2484 : }
2485 108 : if selected_layers.is_empty() {
2486 4 : info!("no layers to compact with gc: no layers within the key range, gc_cutoff={}, key_range={}..{}", gc_cutoff, compact_key_range.start, compact_key_range.end);
2487 4 : return Ok(());
2488 104 : }
2489 104 : retain_lsns_below_horizon.sort();
2490 104 : GcCompactionJobDescription {
2491 104 : selected_layers,
2492 104 : gc_cutoff,
2493 104 : retain_lsns_below_horizon,
2494 104 : min_layer_lsn,
2495 104 : max_layer_lsn,
2496 104 : compaction_key_range: compact_key_range,
2497 104 : rewrite_layers,
2498 104 : }
2499 : };
2500 104 : let (has_data_below, lowest_retain_lsn) = if compact_lsn_range.start != Lsn::INVALID {
2501 : // If we only compact above some LSN, we should get the history from the current branch below the specified LSN.
2502 : // We use job_desc.min_layer_lsn as if it's the lowest branch point.
2503 16 : (true, job_desc.min_layer_lsn)
2504 88 : } else if self.ancestor_timeline.is_some() {
2505 : // In theory, we can also use min_layer_lsn here, but using ancestor LSN makes sure the delta layers cover the
2506 : // LSN ranges all the way to the ancestor timeline.
2507 4 : (true, self.ancestor_lsn)
2508 : } else {
2509 84 : let res = job_desc
2510 84 : .retain_lsns_below_horizon
2511 84 : .first()
2512 84 : .copied()
2513 84 : .unwrap_or(job_desc.gc_cutoff);
2514 84 : if debug_mode {
2515 84 : assert_eq!(
2516 84 : res,
2517 84 : job_desc
2518 84 : .retain_lsns_below_horizon
2519 84 : .iter()
2520 84 : .min()
2521 84 : .copied()
2522 84 : .unwrap_or(job_desc.gc_cutoff)
2523 84 : );
2524 0 : }
2525 84 : (false, res)
2526 : };
2527 104 : info!(
2528 0 : "picked {} layers for compaction ({} layers need rewriting) with max_layer_lsn={} min_layer_lsn={} gc_cutoff={} lowest_retain_lsn={}, key_range={}..{}, has_data_below={}",
2529 0 : job_desc.selected_layers.len(),
2530 0 : job_desc.rewrite_layers.len(),
2531 : job_desc.max_layer_lsn,
2532 : job_desc.min_layer_lsn,
2533 : job_desc.gc_cutoff,
2534 : lowest_retain_lsn,
2535 : job_desc.compaction_key_range.start,
2536 : job_desc.compaction_key_range.end,
2537 : has_data_below,
2538 : );
2539 :
2540 408 : for layer in &job_desc.selected_layers {
2541 304 : debug!("read layer: {}", layer.layer_desc().key());
2542 : }
2543 108 : for layer in &job_desc.rewrite_layers {
2544 4 : debug!("rewrite layer: {}", layer.key());
2545 : }
2546 :
2547 104 : self.check_compaction_space(&job_desc.selected_layers)
2548 104 : .await?;
2549 :
2550 : // Generate statistics for the compaction
2551 408 : for layer in &job_desc.selected_layers {
2552 304 : let desc = layer.layer_desc();
2553 304 : if desc.is_delta() {
2554 172 : stat.visit_delta_layer(desc.file_size());
2555 172 : } else {
2556 132 : stat.visit_image_layer(desc.file_size());
2557 132 : }
2558 : }
2559 :
2560 : // Step 1: construct a k-merge iterator over all layers.
2561 : // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
2562 104 : let layer_names = job_desc
2563 104 : .selected_layers
2564 104 : .iter()
2565 304 : .map(|layer| layer.layer_desc().layer_name())
2566 104 : .collect_vec();
2567 104 : if let Some(err) = check_valid_layermap(&layer_names) {
2568 0 : bail!("gc-compaction layer map check failed because {}, cannot proceed with compaction due to potential data loss", err);
2569 104 : }
2570 104 : // The maximum LSN we are processing in this compaction loop
2571 104 : let end_lsn = job_desc
2572 104 : .selected_layers
2573 104 : .iter()
2574 304 : .map(|l| l.layer_desc().lsn_range.end)
2575 104 : .max()
2576 104 : .unwrap();
2577 104 : let mut delta_layers = Vec::new();
2578 104 : let mut image_layers = Vec::new();
2579 104 : let mut downloaded_layers = Vec::new();
2580 104 : let mut total_downloaded_size = 0;
2581 104 : let mut total_layer_size = 0;
2582 408 : for layer in &job_desc.selected_layers {
2583 304 : if layer.needs_download().await?.is_some() {
2584 0 : total_downloaded_size += layer.layer_desc().file_size;
2585 304 : }
2586 304 : total_layer_size += layer.layer_desc().file_size;
2587 304 : let resident_layer = layer.download_and_keep_resident().await?;
2588 304 : downloaded_layers.push(resident_layer);
2589 : }
2590 104 : info!(
2591 0 : "finish downloading layers, downloaded={}, total={}, ratio={:.2}",
2592 0 : total_downloaded_size,
2593 0 : total_layer_size,
2594 0 : total_downloaded_size as f64 / total_layer_size as f64
2595 : );
2596 408 : for resident_layer in &downloaded_layers {
2597 304 : if resident_layer.layer_desc().is_delta() {
2598 172 : let layer = resident_layer.get_as_delta(ctx).await?;
2599 172 : delta_layers.push(layer);
2600 : } else {
2601 132 : let layer = resident_layer.get_as_image(ctx).await?;
2602 132 : image_layers.push(layer);
2603 : }
2604 : }
2605 104 : let (dense_ks, sparse_ks) = self.collect_gc_compaction_keyspace().await?;
2606 104 : let mut merge_iter = FilterIterator::create(
2607 104 : MergeIterator::create(&delta_layers, &image_layers, ctx),
2608 104 : dense_ks,
2609 104 : sparse_ks,
2610 104 : )?;
2611 :
2612 : // Step 2: Produce images+deltas.
2613 104 : let mut accumulated_values = Vec::new();
2614 104 : let mut last_key: Option<Key> = None;
2615 :
2616 : // Only create image layers when there is no ancestor branches. TODO: create covering image layer
2617 : // when some condition meet.
2618 104 : let mut image_layer_writer = if !has_data_below {
2619 : Some(
2620 84 : SplitImageLayerWriter::new(
2621 84 : self.conf,
2622 84 : self.timeline_id,
2623 84 : self.tenant_shard_id,
2624 84 : job_desc.compaction_key_range.start,
2625 84 : lowest_retain_lsn,
2626 84 : self.get_compaction_target_size(),
2627 84 : ctx,
2628 84 : )
2629 84 : .await?,
2630 : )
2631 : } else {
2632 20 : None
2633 : };
2634 :
2635 104 : let mut delta_layer_writer = SplitDeltaLayerWriter::new(
2636 104 : self.conf,
2637 104 : self.timeline_id,
2638 104 : self.tenant_shard_id,
2639 104 : lowest_retain_lsn..end_lsn,
2640 104 : self.get_compaction_target_size(),
2641 104 : )
2642 104 : .await?;
2643 :
2644 : #[derive(Default)]
2645 : struct RewritingLayers {
2646 : before: Option<DeltaLayerWriter>,
2647 : after: Option<DeltaLayerWriter>,
2648 : }
2649 104 : let mut delta_layer_rewriters = HashMap::<Arc<PersistentLayerKey>, RewritingLayers>::new();
2650 :
2651 : /// When compacting not at a bottom range (=`[0,X)`) of the root branch, we "have data below" (`has_data_below=true`).
2652 : /// The two cases are compaction in ancestor branches and when `compact_lsn_range.start` is set.
2653 : /// In those cases, we need to pull up data from below the LSN range we're compaction.
2654 : ///
2655 : /// This function unifies the cases so that later code doesn't have to think about it.
2656 : ///
2657 : /// Currently, we always get the ancestor image for each key in the child branch no matter whether the image
2658 : /// is needed for reconstruction. This should be fixed in the future.
2659 : ///
2660 : /// Furthermore, we should do vectored get instead of a single get, or better, use k-merge for ancestor
2661 : /// images.
2662 1244 : async fn get_ancestor_image(
2663 1244 : this_tline: &Arc<Timeline>,
2664 1244 : key: Key,
2665 1244 : ctx: &RequestContext,
2666 1244 : has_data_below: bool,
2667 1244 : history_lsn_point: Lsn,
2668 1244 : ) -> anyhow::Result<Option<(Key, Lsn, Bytes)>> {
2669 1244 : if !has_data_below {
2670 1168 : return Ok(None);
2671 76 : };
2672 : // This function is implemented as a get of the current timeline at ancestor LSN, therefore reusing
2673 : // as much existing code as possible.
2674 76 : let img = this_tline.get(key, history_lsn_point, ctx).await?;
2675 76 : Ok(Some((key, history_lsn_point, img)))
2676 1244 : }
2677 :
2678 : // Actually, we can decide not to write to the image layer at all at this point because
2679 : // the key and LSN range are determined. However, to keep things simple here, we still
2680 : // create this writer, and discard the writer in the end.
2681 :
2682 1932 : while let Some(((key, lsn, val), desc)) = merge_iter.next_with_trace().await? {
2683 1828 : if cancel.is_cancelled() {
2684 0 : return Err(anyhow!("cancelled")); // TODO: refactor to CompactionError and pass cancel error
2685 1828 : }
2686 1828 : if self.shard_identity.is_key_disposable(&key) {
2687 : // If this shard does not need to store this key, simply skip it.
2688 : //
2689 : // This is not handled in the filter iterator because shard is determined by hash.
2690 : // Therefore, it does not give us any performance benefit to do things like skip
2691 : // a whole layer file as handling key spaces (ranges).
2692 0 : if cfg!(debug_assertions) {
2693 0 : let shard = self.shard_identity.shard_index();
2694 0 : let owner = self.shard_identity.get_shard_number(&key);
2695 0 : panic!("key {key} does not belong on shard {shard}, owned by {owner}");
2696 0 : }
2697 0 : continue;
2698 1828 : }
2699 1828 : if !job_desc.compaction_key_range.contains(&key) {
2700 128 : if !desc.is_delta {
2701 120 : continue;
2702 8 : }
2703 8 : let rewriter = delta_layer_rewriters.entry(desc.clone()).or_default();
2704 8 : let rewriter = if key < job_desc.compaction_key_range.start {
2705 0 : if rewriter.before.is_none() {
2706 0 : rewriter.before = Some(
2707 0 : DeltaLayerWriter::new(
2708 0 : self.conf,
2709 0 : self.timeline_id,
2710 0 : self.tenant_shard_id,
2711 0 : desc.key_range.start,
2712 0 : desc.lsn_range.clone(),
2713 0 : ctx,
2714 0 : )
2715 0 : .await?,
2716 : );
2717 0 : }
2718 0 : rewriter.before.as_mut().unwrap()
2719 8 : } else if key >= job_desc.compaction_key_range.end {
2720 8 : if rewriter.after.is_none() {
2721 4 : rewriter.after = Some(
2722 4 : DeltaLayerWriter::new(
2723 4 : self.conf,
2724 4 : self.timeline_id,
2725 4 : self.tenant_shard_id,
2726 4 : job_desc.compaction_key_range.end,
2727 4 : desc.lsn_range.clone(),
2728 4 : ctx,
2729 4 : )
2730 4 : .await?,
2731 : );
2732 4 : }
2733 8 : rewriter.after.as_mut().unwrap()
2734 : } else {
2735 0 : unreachable!()
2736 : };
2737 8 : rewriter.put_value(key, lsn, val, ctx).await?;
2738 8 : continue;
2739 1700 : }
2740 1700 : match val {
2741 1220 : Value::Image(_) => stat.visit_image_key(&val),
2742 480 : Value::WalRecord(_) => stat.visit_wal_key(&val),
2743 : }
2744 1700 : if last_key.is_none() || last_key.as_ref() == Some(&key) {
2745 560 : if last_key.is_none() {
2746 104 : last_key = Some(key);
2747 456 : }
2748 560 : accumulated_values.push((key, lsn, val));
2749 : } else {
2750 1140 : let last_key: &mut Key = last_key.as_mut().unwrap();
2751 1140 : stat.on_unique_key_visited(); // TODO: adjust statistics for partial compaction
2752 1140 : let retention = self
2753 1140 : .generate_key_retention(
2754 1140 : *last_key,
2755 1140 : &accumulated_values,
2756 1140 : job_desc.gc_cutoff,
2757 1140 : &job_desc.retain_lsns_below_horizon,
2758 1140 : COMPACTION_DELTA_THRESHOLD,
2759 1140 : get_ancestor_image(self, *last_key, ctx, has_data_below, lowest_retain_lsn)
2760 1140 : .await?,
2761 : )
2762 1140 : .await?;
2763 1140 : retention
2764 1140 : .pipe_to(
2765 1140 : *last_key,
2766 1140 : &mut delta_layer_writer,
2767 1140 : image_layer_writer.as_mut(),
2768 1140 : &mut stat,
2769 1140 : ctx,
2770 1140 : )
2771 1140 : .await?;
2772 1140 : accumulated_values.clear();
2773 1140 : *last_key = key;
2774 1140 : accumulated_values.push((key, lsn, val));
2775 : }
2776 : }
2777 :
2778 : // TODO: move the below part to the loop body
2779 104 : let last_key = last_key.expect("no keys produced during compaction");
2780 104 : stat.on_unique_key_visited();
2781 :
2782 104 : let retention = self
2783 104 : .generate_key_retention(
2784 104 : last_key,
2785 104 : &accumulated_values,
2786 104 : job_desc.gc_cutoff,
2787 104 : &job_desc.retain_lsns_below_horizon,
2788 104 : COMPACTION_DELTA_THRESHOLD,
2789 104 : get_ancestor_image(self, last_key, ctx, has_data_below, lowest_retain_lsn).await?,
2790 : )
2791 104 : .await?;
2792 104 : retention
2793 104 : .pipe_to(
2794 104 : last_key,
2795 104 : &mut delta_layer_writer,
2796 104 : image_layer_writer.as_mut(),
2797 104 : &mut stat,
2798 104 : ctx,
2799 104 : )
2800 104 : .await?;
2801 : // end: move the above part to the loop body
2802 :
2803 104 : let mut rewrote_delta_layers = Vec::new();
2804 108 : for (key, writers) in delta_layer_rewriters {
2805 4 : if let Some(delta_writer_before) = writers.before {
2806 0 : let (desc, path) = delta_writer_before
2807 0 : .finish(job_desc.compaction_key_range.start, ctx)
2808 0 : .await?;
2809 0 : let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
2810 0 : rewrote_delta_layers.push(layer);
2811 4 : }
2812 4 : if let Some(delta_writer_after) = writers.after {
2813 4 : let (desc, path) = delta_writer_after.finish(key.key_range.end, ctx).await?;
2814 4 : let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
2815 4 : rewrote_delta_layers.push(layer);
2816 0 : }
2817 : }
2818 :
2819 148 : let discard = |key: &PersistentLayerKey| {
2820 148 : let key = key.clone();
2821 148 : async move { KeyHistoryRetention::discard_key(&key, self, dry_run).await }
2822 148 : };
2823 :
2824 104 : let produced_image_layers = if let Some(writer) = image_layer_writer {
2825 84 : if !dry_run {
2826 76 : let end_key = job_desc.compaction_key_range.end;
2827 76 : writer
2828 76 : .finish_with_discard_fn(self, ctx, end_key, discard)
2829 76 : .await?
2830 : } else {
2831 8 : drop(writer);
2832 8 : Vec::new()
2833 : }
2834 : } else {
2835 20 : Vec::new()
2836 : };
2837 :
2838 104 : let produced_delta_layers = if !dry_run {
2839 96 : delta_layer_writer
2840 96 : .finish_with_discard_fn(self, ctx, discard)
2841 96 : .await?
2842 : } else {
2843 8 : drop(delta_layer_writer);
2844 8 : Vec::new()
2845 : };
2846 :
2847 : // TODO: make image/delta/rewrote_delta layers generation atomic. At this point, we already generated resident layers, and if
2848 : // compaction is cancelled at this point, we might have some layers that are not cleaned up.
2849 104 : let mut compact_to = Vec::new();
2850 104 : let mut keep_layers = HashSet::new();
2851 104 : let produced_delta_layers_len = produced_delta_layers.len();
2852 104 : let produced_image_layers_len = produced_image_layers.len();
2853 176 : for action in produced_delta_layers {
2854 72 : match action {
2855 44 : BatchWriterResult::Produced(layer) => {
2856 44 : if cfg!(debug_assertions) {
2857 44 : info!("produced delta layer: {}", layer.layer_desc().key());
2858 0 : }
2859 44 : stat.produce_delta_layer(layer.layer_desc().file_size());
2860 44 : compact_to.push(layer);
2861 : }
2862 28 : BatchWriterResult::Discarded(l) => {
2863 28 : if cfg!(debug_assertions) {
2864 28 : info!("discarded delta layer: {}", l);
2865 0 : }
2866 28 : keep_layers.insert(l);
2867 28 : stat.discard_delta_layer();
2868 : }
2869 : }
2870 : }
2871 108 : for layer in &rewrote_delta_layers {
2872 4 : debug!(
2873 0 : "produced rewritten delta layer: {}",
2874 0 : layer.layer_desc().key()
2875 : );
2876 : }
2877 104 : compact_to.extend(rewrote_delta_layers);
2878 180 : for action in produced_image_layers {
2879 76 : match action {
2880 60 : BatchWriterResult::Produced(layer) => {
2881 60 : debug!("produced image layer: {}", layer.layer_desc().key());
2882 60 : stat.produce_image_layer(layer.layer_desc().file_size());
2883 60 : compact_to.push(layer);
2884 : }
2885 16 : BatchWriterResult::Discarded(l) => {
2886 16 : debug!("discarded image layer: {}", l);
2887 16 : keep_layers.insert(l);
2888 16 : stat.discard_image_layer();
2889 : }
2890 : }
2891 : }
2892 :
2893 104 : let mut layer_selection = job_desc.selected_layers;
2894 :
2895 : // Partial compaction might select more data than it processes, e.g., if
2896 : // the compaction_key_range only partially overlaps:
2897 : //
2898 : // [---compaction_key_range---]
2899 : // [---A----][----B----][----C----][----D----]
2900 : //
2901 : // For delta layers, we will rewrite the layers so that it is cut exactly at
2902 : // the compaction key range, so we can always discard them. However, for image
2903 : // layers, as we do not rewrite them for now, we need to handle them differently.
2904 : // Assume image layers A, B, C, D are all in the `layer_selection`.
2905 : //
2906 : // The created image layers contain whatever is needed from B, C, and from
2907 : // `----]` of A, and from `[---` of D.
2908 : //
2909 : // In contrast, `[---A` and `D----]` have not been processed, so, we must
2910 : // keep that data.
2911 : //
2912 : // The solution for now is to keep A and D completely if they are image layers.
2913 : // (layer_selection is what we'll remove from the layer map, so, retain what
2914 : // is _not_ fully covered by compaction_key_range).
2915 408 : for layer in &layer_selection {
2916 304 : if !layer.layer_desc().is_delta() {
2917 132 : if !overlaps_with(
2918 132 : &layer.layer_desc().key_range,
2919 132 : &job_desc.compaction_key_range,
2920 132 : ) {
2921 0 : bail!("violated constraint: image layer outside of compaction key range");
2922 132 : }
2923 132 : if !fully_contains(
2924 132 : &job_desc.compaction_key_range,
2925 132 : &layer.layer_desc().key_range,
2926 132 : ) {
2927 16 : keep_layers.insert(layer.layer_desc().key());
2928 116 : }
2929 172 : }
2930 : }
2931 :
2932 304 : layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
2933 104 :
2934 104 : info!(
2935 0 : "gc-compaction statistics: {}",
2936 0 : serde_json::to_string(&stat)?
2937 : );
2938 :
2939 104 : if dry_run {
2940 8 : return Ok(());
2941 96 : }
2942 96 :
2943 96 : info!(
2944 0 : "produced {} delta layers and {} image layers, {} layers are kept",
2945 0 : produced_delta_layers_len,
2946 0 : produced_image_layers_len,
2947 0 : keep_layers.len()
2948 : );
2949 :
2950 : // Step 3: Place back to the layer map.
2951 :
2952 : // First, do a sanity check to ensure the newly-created layer map does not contain overlaps.
2953 96 : let all_layers = {
2954 96 : let guard = self.layers.read().await;
2955 96 : let layer_map = guard.layer_map()?;
2956 96 : layer_map.iter_historic_layers().collect_vec()
2957 96 : };
2958 96 :
2959 96 : let mut final_layers = all_layers
2960 96 : .iter()
2961 428 : .map(|layer| layer.layer_name())
2962 96 : .collect::<HashSet<_>>();
2963 304 : for layer in &layer_selection {
2964 208 : final_layers.remove(&layer.layer_desc().layer_name());
2965 208 : }
2966 204 : for layer in &compact_to {
2967 108 : final_layers.insert(layer.layer_desc().layer_name());
2968 108 : }
2969 96 : let final_layers = final_layers.into_iter().collect_vec();
2970 :
2971 : // TODO: move this check before we call `finish` on image layer writers. However, this will require us to get the layer name before we finish
2972 : // the writer, so potentially, we will need a function like `ImageLayerBatchWriter::get_all_pending_layer_keys` to get all the keys that are
2973 : // in the writer before finalizing the persistent layers. Now we would leave some dangling layers on the disk if the check fails.
2974 96 : if let Some(err) = check_valid_layermap(&final_layers) {
2975 0 : bail!("gc-compaction layer map check failed after compaction because {}, compaction result not applied to the layer map due to potential data loss", err);
2976 96 : }
2977 :
2978 : // Between the sanity check and this compaction update, there could be new layers being flushed, but it should be fine because we only
2979 : // operate on L1 layers.
2980 : {
2981 : // Gc-compaction will rewrite the history of a key. This could happen in two ways:
2982 : //
2983 : // 1. We create an image layer to replace all the deltas below the compact LSN. In this case, assume
2984 : // we have 2 delta layers A and B, both below the compact LSN. We create an image layer I to replace
2985 : // A and B at the compact LSN. If the read path finishes reading A, yields, and now we update the layer
2986 : // map, the read path then cannot find any keys below A, reporting a missing key error, while the key
2987 : // now gets stored in I at the compact LSN.
2988 : //
2989 : // --------------- ---------------
2990 : // delta1@LSN20 image1@LSN20
2991 : // --------------- (read path collects delta@LSN20, => --------------- (read path cannot find anything
2992 : // delta1@LSN10 yields) below LSN 20)
2993 : // ---------------
2994 : //
2995 : // 2. We create a delta layer to replace all the deltas below the compact LSN, and in the delta layers,
2996 : // we combines the history of a key into a single image. For example, we have deltas at LSN 1, 2, 3, 4,
2997 : // Assume one delta layer contains LSN 1, 2, 3 and the other contains LSN 4.
2998 : //
2999 : // We let gc-compaction combine delta 2, 3, 4 into an image at LSN 4, which produces a delta layer that
3000 : // contains the delta at LSN 1, the image at LSN 4. If the read path finishes reading the original delta
3001 : // layer containing 4, yields, and we update the layer map to put the delta layer.
3002 : //
3003 : // --------------- ---------------
3004 : // delta1@LSN4 image1@LSN4
3005 : // --------------- (read path collects delta@LSN4, => --------------- (read path collects LSN4 and LSN1,
3006 : // delta1@LSN1-3 yields) delta1@LSN1 which is an invalid history)
3007 : // --------------- ---------------
3008 : //
3009 : // Therefore, the gc-compaction layer update operation should wait for all ongoing reads, block all pending reads,
3010 : // and only allow reads to continue after the update is finished.
3011 :
3012 96 : let update_guard = self.gc_compaction_layer_update_lock.write().await;
3013 : // Acquiring the update guard ensures current read operations end and new read operations are blocked.
3014 : // TODO: can we use `latest_gc_cutoff` Rcu to achieve the same effect?
3015 96 : let mut guard = self.layers.write().await;
3016 96 : guard
3017 96 : .open_mut()?
3018 96 : .finish_gc_compaction(&layer_selection, &compact_to, &self.metrics);
3019 96 : drop(update_guard); // Allow new reads to start ONLY after we finished updating the layer map.
3020 96 : };
3021 96 :
3022 96 : // Schedule an index-only upload to update the `latest_gc_cutoff` in the index_part.json.
3023 96 : // Otherwise, after restart, the index_part only contains the old `latest_gc_cutoff` and
3024 96 : // find_gc_cutoffs will try accessing things below the cutoff. TODO: ideally, this should
3025 96 : // be batched into `schedule_compaction_update`.
3026 96 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
3027 96 : self.schedule_uploads(disk_consistent_lsn, None)?;
3028 : // If a layer gets rewritten throughout gc-compaction, we need to keep that layer only in `compact_to` instead
3029 : // of `compact_from`.
3030 96 : let compact_from = {
3031 96 : let mut compact_from = Vec::new();
3032 96 : let mut compact_to_set = HashMap::new();
3033 204 : for layer in &compact_to {
3034 108 : compact_to_set.insert(layer.layer_desc().key(), layer);
3035 108 : }
3036 304 : for layer in &layer_selection {
3037 208 : if let Some(to) = compact_to_set.get(&layer.layer_desc().key()) {
3038 0 : tracing::info!(
3039 0 : "skipping delete {} because found same layer key at different generation {}",
3040 : layer, to
3041 : );
3042 208 : } else {
3043 208 : compact_from.push(layer.clone());
3044 208 : }
3045 : }
3046 96 : compact_from
3047 96 : };
3048 96 : self.remote_client
3049 96 : .schedule_compaction_update(&compact_from, &compact_to)?;
3050 :
3051 96 : drop(gc_lock);
3052 96 :
3053 96 : Ok(())
3054 108 : }
3055 : }
3056 :
3057 : struct TimelineAdaptor {
3058 : timeline: Arc<Timeline>,
3059 :
3060 : keyspace: (Lsn, KeySpace),
3061 :
3062 : new_deltas: Vec<ResidentLayer>,
3063 : new_images: Vec<ResidentLayer>,
3064 : layers_to_delete: Vec<Arc<PersistentLayerDesc>>,
3065 : }
3066 :
3067 : impl TimelineAdaptor {
3068 0 : pub fn new(timeline: &Arc<Timeline>, keyspace: (Lsn, KeySpace)) -> Self {
3069 0 : Self {
3070 0 : timeline: timeline.clone(),
3071 0 : keyspace,
3072 0 : new_images: Vec::new(),
3073 0 : new_deltas: Vec::new(),
3074 0 : layers_to_delete: Vec::new(),
3075 0 : }
3076 0 : }
3077 :
3078 0 : pub async fn flush_updates(&mut self) -> Result<(), CompactionError> {
3079 0 : let layers_to_delete = {
3080 0 : let guard = self.timeline.layers.read().await;
3081 0 : self.layers_to_delete
3082 0 : .iter()
3083 0 : .map(|x| guard.get_from_desc(x))
3084 0 : .collect::<Vec<Layer>>()
3085 0 : };
3086 0 : self.timeline
3087 0 : .finish_compact_batch(&self.new_deltas, &self.new_images, &layers_to_delete)
3088 0 : .await?;
3089 :
3090 0 : self.timeline
3091 0 : .upload_new_image_layers(std::mem::take(&mut self.new_images))?;
3092 :
3093 0 : self.new_deltas.clear();
3094 0 : self.layers_to_delete.clear();
3095 0 : Ok(())
3096 0 : }
3097 : }
3098 :
3099 : #[derive(Clone)]
3100 : struct ResidentDeltaLayer(ResidentLayer);
3101 : #[derive(Clone)]
3102 : struct ResidentImageLayer(ResidentLayer);
3103 :
3104 : impl CompactionJobExecutor for TimelineAdaptor {
3105 : type Key = pageserver_api::key::Key;
3106 :
3107 : type Layer = OwnArc<PersistentLayerDesc>;
3108 : type DeltaLayer = ResidentDeltaLayer;
3109 : type ImageLayer = ResidentImageLayer;
3110 :
3111 : type RequestContext = crate::context::RequestContext;
3112 :
3113 0 : fn get_shard_identity(&self) -> &ShardIdentity {
3114 0 : self.timeline.get_shard_identity()
3115 0 : }
3116 :
3117 0 : async fn get_layers(
3118 0 : &mut self,
3119 0 : key_range: &Range<Key>,
3120 0 : lsn_range: &Range<Lsn>,
3121 0 : _ctx: &RequestContext,
3122 0 : ) -> anyhow::Result<Vec<OwnArc<PersistentLayerDesc>>> {
3123 0 : self.flush_updates().await?;
3124 :
3125 0 : let guard = self.timeline.layers.read().await;
3126 0 : let layer_map = guard.layer_map()?;
3127 :
3128 0 : let result = layer_map
3129 0 : .iter_historic_layers()
3130 0 : .filter(|l| {
3131 0 : overlaps_with(&l.lsn_range, lsn_range) && overlaps_with(&l.key_range, key_range)
3132 0 : })
3133 0 : .map(OwnArc)
3134 0 : .collect();
3135 0 : Ok(result)
3136 0 : }
3137 :
3138 0 : async fn get_keyspace(
3139 0 : &mut self,
3140 0 : key_range: &Range<Key>,
3141 0 : lsn: Lsn,
3142 0 : _ctx: &RequestContext,
3143 0 : ) -> anyhow::Result<Vec<Range<Key>>> {
3144 0 : if lsn == self.keyspace.0 {
3145 0 : Ok(pageserver_compaction::helpers::intersect_keyspace(
3146 0 : &self.keyspace.1.ranges,
3147 0 : key_range,
3148 0 : ))
3149 : } else {
3150 : // The current compaction implementation only ever requests the key space
3151 : // at the compaction end LSN.
3152 0 : anyhow::bail!("keyspace not available for requested lsn");
3153 : }
3154 0 : }
3155 :
3156 0 : async fn downcast_delta_layer(
3157 0 : &self,
3158 0 : layer: &OwnArc<PersistentLayerDesc>,
3159 0 : ) -> anyhow::Result<Option<ResidentDeltaLayer>> {
3160 0 : // this is a lot more complex than a simple downcast...
3161 0 : if layer.is_delta() {
3162 0 : let l = {
3163 0 : let guard = self.timeline.layers.read().await;
3164 0 : guard.get_from_desc(layer)
3165 : };
3166 0 : let result = l.download_and_keep_resident().await?;
3167 :
3168 0 : Ok(Some(ResidentDeltaLayer(result)))
3169 : } else {
3170 0 : Ok(None)
3171 : }
3172 0 : }
3173 :
3174 0 : async fn create_image(
3175 0 : &mut self,
3176 0 : lsn: Lsn,
3177 0 : key_range: &Range<Key>,
3178 0 : ctx: &RequestContext,
3179 0 : ) -> anyhow::Result<()> {
3180 0 : Ok(self.create_image_impl(lsn, key_range, ctx).await?)
3181 0 : }
3182 :
3183 0 : async fn create_delta(
3184 0 : &mut self,
3185 0 : lsn_range: &Range<Lsn>,
3186 0 : key_range: &Range<Key>,
3187 0 : input_layers: &[ResidentDeltaLayer],
3188 0 : ctx: &RequestContext,
3189 0 : ) -> anyhow::Result<()> {
3190 0 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
3191 :
3192 0 : let mut all_entries = Vec::new();
3193 0 : for dl in input_layers.iter() {
3194 0 : all_entries.extend(dl.load_keys(ctx).await?);
3195 : }
3196 :
3197 : // The current stdlib sorting implementation is designed in a way where it is
3198 : // particularly fast where the slice is made up of sorted sub-ranges.
3199 0 : all_entries.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
3200 :
3201 0 : let mut writer = DeltaLayerWriter::new(
3202 0 : self.timeline.conf,
3203 0 : self.timeline.timeline_id,
3204 0 : self.timeline.tenant_shard_id,
3205 0 : key_range.start,
3206 0 : lsn_range.clone(),
3207 0 : ctx,
3208 0 : )
3209 0 : .await?;
3210 :
3211 0 : let mut dup_values = 0;
3212 0 :
3213 0 : // This iterator walks through all key-value pairs from all the layers
3214 0 : // we're compacting, in key, LSN order.
3215 0 : let mut prev: Option<(Key, Lsn)> = None;
3216 : for &DeltaEntry {
3217 0 : key, lsn, ref val, ..
3218 0 : } in all_entries.iter()
3219 : {
3220 0 : if prev == Some((key, lsn)) {
3221 : // This is a duplicate. Skip it.
3222 : //
3223 : // It can happen if compaction is interrupted after writing some
3224 : // layers but not all, and we are compacting the range again.
3225 : // The calculations in the algorithm assume that there are no
3226 : // duplicates, so the math on targeted file size is likely off,
3227 : // and we will create smaller files than expected.
3228 0 : dup_values += 1;
3229 0 : continue;
3230 0 : }
3231 :
3232 0 : let value = val.load(ctx).await?;
3233 :
3234 0 : writer.put_value(key, lsn, value, ctx).await?;
3235 :
3236 0 : prev = Some((key, lsn));
3237 : }
3238 :
3239 0 : if dup_values > 0 {
3240 0 : warn!("delta layer created with {} duplicate values", dup_values);
3241 0 : }
3242 :
3243 0 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
3244 0 : Err(anyhow::anyhow!(
3245 0 : "failpoint delta-layer-writer-fail-before-finish"
3246 0 : ))
3247 0 : });
3248 :
3249 0 : let (desc, path) = writer.finish(prev.unwrap().0.next(), ctx).await?;
3250 0 : let new_delta_layer =
3251 0 : Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
3252 :
3253 0 : self.new_deltas.push(new_delta_layer);
3254 0 : Ok(())
3255 0 : }
3256 :
3257 0 : async fn delete_layer(
3258 0 : &mut self,
3259 0 : layer: &OwnArc<PersistentLayerDesc>,
3260 0 : _ctx: &RequestContext,
3261 0 : ) -> anyhow::Result<()> {
3262 0 : self.layers_to_delete.push(layer.clone().0);
3263 0 : Ok(())
3264 0 : }
3265 : }
3266 :
3267 : impl TimelineAdaptor {
3268 0 : async fn create_image_impl(
3269 0 : &mut self,
3270 0 : lsn: Lsn,
3271 0 : key_range: &Range<Key>,
3272 0 : ctx: &RequestContext,
3273 0 : ) -> Result<(), CreateImageLayersError> {
3274 0 : let timer = self.timeline.metrics.create_images_time_histo.start_timer();
3275 :
3276 0 : let image_layer_writer = ImageLayerWriter::new(
3277 0 : self.timeline.conf,
3278 0 : self.timeline.timeline_id,
3279 0 : self.timeline.tenant_shard_id,
3280 0 : key_range,
3281 0 : lsn,
3282 0 : ctx,
3283 0 : )
3284 0 : .await?;
3285 :
3286 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
3287 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
3288 0 : "failpoint image-layer-writer-fail-before-finish"
3289 0 : )))
3290 0 : });
3291 :
3292 0 : let keyspace = KeySpace {
3293 0 : ranges: self.get_keyspace(key_range, lsn, ctx).await?,
3294 : };
3295 : // TODO set proper (stateful) start. The create_image_layer_for_rel_blocks function mostly
3296 0 : let outcome = self
3297 0 : .timeline
3298 0 : .create_image_layer_for_rel_blocks(
3299 0 : &keyspace,
3300 0 : image_layer_writer,
3301 0 : lsn,
3302 0 : ctx,
3303 0 : key_range.clone(),
3304 0 : IoConcurrency::sequential(),
3305 0 : )
3306 0 : .await?;
3307 :
3308 : if let ImageLayerCreationOutcome::Generated {
3309 0 : unfinished_image_layer,
3310 0 : } = outcome
3311 : {
3312 0 : let (desc, path) = unfinished_image_layer.finish(ctx).await?;
3313 0 : let image_layer =
3314 0 : Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
3315 0 : self.new_images.push(image_layer);
3316 0 : }
3317 :
3318 0 : timer.stop_and_record();
3319 0 :
3320 0 : Ok(())
3321 0 : }
3322 : }
3323 :
3324 : impl CompactionRequestContext for crate::context::RequestContext {}
3325 :
3326 : #[derive(Debug, Clone)]
3327 : pub struct OwnArc<T>(pub Arc<T>);
3328 :
3329 : impl<T> Deref for OwnArc<T> {
3330 : type Target = <Arc<T> as Deref>::Target;
3331 0 : fn deref(&self) -> &Self::Target {
3332 0 : &self.0
3333 0 : }
3334 : }
3335 :
3336 : impl<T> AsRef<T> for OwnArc<T> {
3337 0 : fn as_ref(&self) -> &T {
3338 0 : self.0.as_ref()
3339 0 : }
3340 : }
3341 :
3342 : impl CompactionLayer<Key> for OwnArc<PersistentLayerDesc> {
3343 0 : fn key_range(&self) -> &Range<Key> {
3344 0 : &self.key_range
3345 0 : }
3346 0 : fn lsn_range(&self) -> &Range<Lsn> {
3347 0 : &self.lsn_range
3348 0 : }
3349 0 : fn file_size(&self) -> u64 {
3350 0 : self.file_size
3351 0 : }
3352 0 : fn short_id(&self) -> std::string::String {
3353 0 : self.as_ref().short_id().to_string()
3354 0 : }
3355 0 : fn is_delta(&self) -> bool {
3356 0 : self.as_ref().is_delta()
3357 0 : }
3358 : }
3359 :
3360 : impl CompactionLayer<Key> for OwnArc<DeltaLayer> {
3361 0 : fn key_range(&self) -> &Range<Key> {
3362 0 : &self.layer_desc().key_range
3363 0 : }
3364 0 : fn lsn_range(&self) -> &Range<Lsn> {
3365 0 : &self.layer_desc().lsn_range
3366 0 : }
3367 0 : fn file_size(&self) -> u64 {
3368 0 : self.layer_desc().file_size
3369 0 : }
3370 0 : fn short_id(&self) -> std::string::String {
3371 0 : self.layer_desc().short_id().to_string()
3372 0 : }
3373 0 : fn is_delta(&self) -> bool {
3374 0 : true
3375 0 : }
3376 : }
3377 :
3378 : use crate::tenant::timeline::DeltaEntry;
3379 :
3380 : impl CompactionLayer<Key> for ResidentDeltaLayer {
3381 0 : fn key_range(&self) -> &Range<Key> {
3382 0 : &self.0.layer_desc().key_range
3383 0 : }
3384 0 : fn lsn_range(&self) -> &Range<Lsn> {
3385 0 : &self.0.layer_desc().lsn_range
3386 0 : }
3387 0 : fn file_size(&self) -> u64 {
3388 0 : self.0.layer_desc().file_size
3389 0 : }
3390 0 : fn short_id(&self) -> std::string::String {
3391 0 : self.0.layer_desc().short_id().to_string()
3392 0 : }
3393 0 : fn is_delta(&self) -> bool {
3394 0 : true
3395 0 : }
3396 : }
3397 :
3398 : impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
3399 : type DeltaEntry<'a> = DeltaEntry<'a>;
3400 :
3401 0 : async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
3402 0 : self.0.get_as_delta(ctx).await?.index_entries(ctx).await
3403 0 : }
3404 : }
3405 :
3406 : impl CompactionLayer<Key> for ResidentImageLayer {
3407 0 : fn key_range(&self) -> &Range<Key> {
3408 0 : &self.0.layer_desc().key_range
3409 0 : }
3410 0 : fn lsn_range(&self) -> &Range<Lsn> {
3411 0 : &self.0.layer_desc().lsn_range
3412 0 : }
3413 0 : fn file_size(&self) -> u64 {
3414 0 : self.0.layer_desc().file_size
3415 0 : }
3416 0 : fn short_id(&self) -> std::string::String {
3417 0 : self.0.layer_desc().short_id().to_string()
3418 0 : }
3419 0 : fn is_delta(&self) -> bool {
3420 0 : false
3421 0 : }
3422 : }
3423 : impl CompactionImageLayer<TimelineAdaptor> for ResidentImageLayer {}
|
