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 : use std::time::{Duration, Instant};
11 :
12 : use super::layer_manager::LayerManager;
13 : use super::{
14 : CompactFlags, CompactOptions, CompactionError, CreateImageLayersError, DurationRecorder,
15 : GetVectoredError, ImageLayerCreationMode, LastImageLayerCreationStatus, RecordedDuration,
16 : Timeline,
17 : };
18 :
19 : use crate::tenant::timeline::DeltaEntry;
20 : use crate::walredo::RedoAttemptType;
21 : use anyhow::{Context, anyhow};
22 : use bytes::Bytes;
23 : use enumset::EnumSet;
24 : use fail::fail_point;
25 : use futures::FutureExt;
26 : use itertools::Itertools;
27 : use once_cell::sync::Lazy;
28 : use pageserver_api::config::tenant_conf_defaults::DEFAULT_CHECKPOINT_DISTANCE;
29 : use pageserver_api::key::{KEY_SIZE, Key};
30 : use pageserver_api::keyspace::{KeySpace, ShardedRange};
31 : use pageserver_api::models::{CompactInfoResponse, CompactKeyRange};
32 : use pageserver_api::record::NeonWalRecord;
33 : use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
34 : use pageserver_api::value::Value;
35 : use pageserver_compaction::helpers::{fully_contains, overlaps_with};
36 : use pageserver_compaction::interface::*;
37 : use serde::Serialize;
38 : use tokio::sync::{OwnedSemaphorePermit, Semaphore};
39 : use tokio_util::sync::CancellationToken;
40 : use tracing::{Instrument, debug, error, info, info_span, trace, warn};
41 : use utils::critical;
42 : use utils::id::TimelineId;
43 : use utils::lsn::Lsn;
44 :
45 : use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
46 : use crate::page_cache;
47 : use crate::statvfs::Statvfs;
48 : use crate::tenant::checks::check_valid_layermap;
49 : use crate::tenant::gc_block::GcBlock;
50 : use crate::tenant::layer_map::LayerMap;
51 : use crate::tenant::remote_timeline_client::WaitCompletionError;
52 : use crate::tenant::remote_timeline_client::index::GcCompactionState;
53 : use crate::tenant::storage_layer::batch_split_writer::{
54 : BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
55 : };
56 : use crate::tenant::storage_layer::filter_iterator::FilterIterator;
57 : use crate::tenant::storage_layer::merge_iterator::MergeIterator;
58 : use crate::tenant::storage_layer::{
59 : AsLayerDesc, PersistentLayerDesc, PersistentLayerKey, ValueReconstructState,
60 : };
61 : use crate::tenant::tasks::log_compaction_error;
62 : use crate::tenant::timeline::{
63 : DeltaLayerWriter, ImageLayerCreationOutcome, ImageLayerWriter, IoConcurrency, Layer,
64 : ResidentLayer, drop_rlock,
65 : };
66 : use crate::tenant::{DeltaLayer, MaybeOffloaded};
67 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
68 :
69 : /// Maximum number of deltas before generating an image layer in bottom-most compaction.
70 : const COMPACTION_DELTA_THRESHOLD: usize = 5;
71 :
72 : #[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
73 : pub struct GcCompactionJobId(pub usize);
74 :
75 : impl std::fmt::Display for GcCompactionJobId {
76 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
77 0 : write!(f, "{}", self.0)
78 0 : }
79 : }
80 :
81 : pub struct GcCompactionCombinedSettings {
82 : pub gc_compaction_enabled: bool,
83 : pub gc_compaction_initial_threshold_kb: u64,
84 : pub gc_compaction_ratio_percent: u64,
85 : }
86 :
87 : #[derive(Debug, Clone)]
88 : pub enum GcCompactionQueueItem {
89 : MetaJob {
90 : /// Compaction options
91 : options: CompactOptions,
92 : /// Whether the compaction is triggered automatically (determines whether we need to update L2 LSN)
93 : auto: bool,
94 : },
95 : SubCompactionJob(CompactOptions),
96 : Notify(GcCompactionJobId, Option<Lsn>),
97 : }
98 :
99 : impl GcCompactionQueueItem {
100 0 : pub fn into_compact_info_resp(
101 0 : self,
102 0 : id: GcCompactionJobId,
103 0 : running: bool,
104 0 : ) -> Option<CompactInfoResponse> {
105 0 : match self {
106 0 : GcCompactionQueueItem::MetaJob { options, .. } => Some(CompactInfoResponse {
107 0 : compact_key_range: options.compact_key_range,
108 0 : compact_lsn_range: options.compact_lsn_range,
109 0 : sub_compaction: options.sub_compaction,
110 0 : running,
111 0 : job_id: id.0,
112 0 : }),
113 0 : GcCompactionQueueItem::SubCompactionJob(options) => Some(CompactInfoResponse {
114 0 : compact_key_range: options.compact_key_range,
115 0 : compact_lsn_range: options.compact_lsn_range,
116 0 : sub_compaction: options.sub_compaction,
117 0 : running,
118 0 : job_id: id.0,
119 0 : }),
120 0 : GcCompactionQueueItem::Notify(_, _) => None,
121 : }
122 0 : }
123 : }
124 :
125 : #[derive(Default)]
126 : struct GcCompactionGuardItems {
127 : notify: Option<tokio::sync::oneshot::Sender<()>>,
128 : permit: Option<OwnedSemaphorePermit>,
129 : }
130 :
131 : struct GcCompactionQueueInner {
132 : running: Option<(GcCompactionJobId, GcCompactionQueueItem)>,
133 : queued: VecDeque<(GcCompactionJobId, GcCompactionQueueItem)>,
134 : guards: HashMap<GcCompactionJobId, GcCompactionGuardItems>,
135 : last_id: GcCompactionJobId,
136 : }
137 :
138 : impl GcCompactionQueueInner {
139 0 : fn next_id(&mut self) -> GcCompactionJobId {
140 0 : let id = self.last_id;
141 0 : self.last_id = GcCompactionJobId(id.0 + 1);
142 0 : id
143 0 : }
144 : }
145 :
146 : /// A structure to store gc_compaction jobs.
147 : pub struct GcCompactionQueue {
148 : /// All items in the queue, and the currently-running job.
149 : inner: std::sync::Mutex<GcCompactionQueueInner>,
150 : /// Ensure only one thread is consuming the queue.
151 : consumer_lock: tokio::sync::Mutex<()>,
152 : }
153 :
154 0 : static CONCURRENT_GC_COMPACTION_TASKS: Lazy<Arc<Semaphore>> = Lazy::new(|| {
155 0 : // Only allow two timelines on one pageserver to run gc compaction at a time.
156 0 : Arc::new(Semaphore::new(2))
157 0 : });
158 :
159 : impl GcCompactionQueue {
160 0 : pub fn new() -> Self {
161 0 : GcCompactionQueue {
162 0 : inner: std::sync::Mutex::new(GcCompactionQueueInner {
163 0 : running: None,
164 0 : queued: VecDeque::new(),
165 0 : guards: HashMap::new(),
166 0 : last_id: GcCompactionJobId(0),
167 0 : }),
168 0 : consumer_lock: tokio::sync::Mutex::new(()),
169 0 : }
170 0 : }
171 :
172 0 : pub fn cancel_scheduled(&self) {
173 0 : let mut guard = self.inner.lock().unwrap();
174 0 : guard.queued.clear();
175 0 : // TODO: if there is a running job, we should keep the gc guard. However, currently, the cancel
176 0 : // API is only used for testing purposes, so we can drop everything here.
177 0 : guard.guards.clear();
178 0 : }
179 :
180 : /// Schedule a manual compaction job.
181 0 : pub fn schedule_manual_compaction(
182 0 : &self,
183 0 : options: CompactOptions,
184 0 : notify: Option<tokio::sync::oneshot::Sender<()>>,
185 0 : ) -> GcCompactionJobId {
186 0 : let mut guard = self.inner.lock().unwrap();
187 0 : let id = guard.next_id();
188 0 : guard.queued.push_back((
189 0 : id,
190 0 : GcCompactionQueueItem::MetaJob {
191 0 : options,
192 0 : auto: false,
193 0 : },
194 0 : ));
195 0 : guard.guards.entry(id).or_default().notify = notify;
196 0 : info!("scheduled compaction job id={}", id);
197 0 : id
198 0 : }
199 :
200 : /// Schedule an auto compaction job.
201 0 : fn schedule_auto_compaction(
202 0 : &self,
203 0 : options: CompactOptions,
204 0 : permit: OwnedSemaphorePermit,
205 0 : ) -> GcCompactionJobId {
206 0 : let mut guard = self.inner.lock().unwrap();
207 0 : let id = guard.next_id();
208 0 : guard.queued.push_back((
209 0 : id,
210 0 : GcCompactionQueueItem::MetaJob {
211 0 : options,
212 0 : auto: true,
213 0 : },
214 0 : ));
215 0 : guard.guards.entry(id).or_default().permit = Some(permit);
216 0 : id
217 0 : }
218 :
219 : /// Trigger an auto compaction.
220 0 : pub async fn trigger_auto_compaction(
221 0 : &self,
222 0 : timeline: &Arc<Timeline>,
223 0 : ) -> Result<(), CompactionError> {
224 0 : let GcCompactionCombinedSettings {
225 0 : gc_compaction_enabled,
226 0 : gc_compaction_initial_threshold_kb,
227 0 : gc_compaction_ratio_percent,
228 0 : } = timeline.get_gc_compaction_settings();
229 0 : if !gc_compaction_enabled {
230 0 : return Ok(());
231 0 : }
232 0 : if self.remaining_jobs_num() > 0 {
233 : // Only schedule auto compaction when the queue is empty
234 0 : return Ok(());
235 0 : }
236 0 : if timeline.ancestor_timeline().is_some() {
237 : // Do not trigger auto compaction for child timelines. We haven't tested
238 : // it enough in staging yet.
239 0 : return Ok(());
240 0 : }
241 0 : if timeline.get_gc_compaction_watermark() == Lsn::INVALID {
242 : // If the gc watermark is not set, we don't need to trigger auto compaction.
243 : // This check is the same as in `gc_compaction_split_jobs` but we don't log
244 : // here and we can also skip the computation of the trigger condition earlier.
245 0 : return Ok(());
246 0 : }
247 :
248 0 : let Ok(permit) = CONCURRENT_GC_COMPACTION_TASKS.clone().try_acquire_owned() else {
249 : // Only allow one compaction run at a time. TODO: As we do `try_acquire_owned`, we cannot ensure
250 : // the fairness of the lock across timelines. We should listen for both `acquire` and `l0_compaction_trigger`
251 : // to ensure the fairness while avoid starving other tasks.
252 0 : return Ok(());
253 : };
254 :
255 0 : let gc_compaction_state = timeline.get_gc_compaction_state();
256 0 : let l2_lsn = gc_compaction_state
257 0 : .map(|x| x.last_completed_lsn)
258 0 : .unwrap_or(Lsn::INVALID);
259 :
260 0 : let layers = {
261 0 : let guard = timeline.layers.read().await;
262 0 : let layer_map = guard.layer_map()?;
263 0 : layer_map.iter_historic_layers().collect_vec()
264 0 : };
265 0 : let mut l2_size: u64 = 0;
266 0 : let mut l1_size = 0;
267 0 : let gc_cutoff = *timeline.get_applied_gc_cutoff_lsn();
268 0 : for layer in layers {
269 0 : if layer.lsn_range.start <= l2_lsn {
270 0 : l2_size += layer.file_size();
271 0 : } else if layer.lsn_range.start <= gc_cutoff {
272 0 : l1_size += layer.file_size();
273 0 : }
274 : }
275 :
276 0 : fn trigger_compaction(
277 0 : l1_size: u64,
278 0 : l2_size: u64,
279 0 : gc_compaction_initial_threshold_kb: u64,
280 0 : gc_compaction_ratio_percent: u64,
281 0 : ) -> bool {
282 : const AUTO_TRIGGER_LIMIT: u64 = 150 * 1024 * 1024 * 1024; // 150GB
283 0 : if l1_size + l2_size >= AUTO_TRIGGER_LIMIT {
284 : // Do not auto-trigger when physical size >= 150GB
285 0 : return false;
286 0 : }
287 0 : // initial trigger
288 0 : if l2_size == 0 && l1_size >= gc_compaction_initial_threshold_kb * 1024 {
289 0 : info!(
290 0 : "trigger auto-compaction because l1_size={} >= gc_compaction_initial_threshold_kb={}",
291 : l1_size, gc_compaction_initial_threshold_kb
292 : );
293 0 : return true;
294 0 : }
295 0 : // size ratio trigger
296 0 : if l2_size == 0 {
297 0 : return false;
298 0 : }
299 0 : if l1_size as f64 / l2_size as f64 >= (gc_compaction_ratio_percent as f64 / 100.0) {
300 0 : info!(
301 0 : "trigger auto-compaction because l1_size={} / l2_size={} > gc_compaction_ratio_percent={}",
302 : l1_size, l2_size, gc_compaction_ratio_percent
303 : );
304 0 : return true;
305 0 : }
306 0 : false
307 0 : }
308 :
309 0 : if trigger_compaction(
310 0 : l1_size,
311 0 : l2_size,
312 0 : gc_compaction_initial_threshold_kb,
313 0 : gc_compaction_ratio_percent,
314 0 : ) {
315 0 : self.schedule_auto_compaction(
316 0 : CompactOptions {
317 0 : flags: {
318 0 : let mut flags = EnumSet::new();
319 0 : flags |= CompactFlags::EnhancedGcBottomMostCompaction;
320 0 : if timeline.get_compaction_l0_first() {
321 0 : flags |= CompactFlags::YieldForL0;
322 0 : }
323 0 : flags
324 0 : },
325 0 : sub_compaction: true,
326 0 : // Only auto-trigger gc-compaction over the data keyspace due to concerns in
327 0 : // https://github.com/neondatabase/neon/issues/11318.
328 0 : compact_key_range: Some(CompactKeyRange {
329 0 : start: Key::MIN,
330 0 : end: Key::metadata_key_range().start,
331 0 : }),
332 0 : compact_lsn_range: None,
333 0 : sub_compaction_max_job_size_mb: None,
334 0 : },
335 0 : permit,
336 0 : );
337 0 : info!(
338 0 : "scheduled auto gc-compaction: l1_size={}, l2_size={}, l2_lsn={}, gc_cutoff={}",
339 : l1_size, l2_size, l2_lsn, gc_cutoff
340 : );
341 : } else {
342 0 : debug!(
343 0 : "did not trigger auto gc-compaction: l1_size={}, l2_size={}, l2_lsn={}, gc_cutoff={}",
344 : l1_size, l2_size, l2_lsn, gc_cutoff
345 : );
346 : }
347 0 : Ok(())
348 0 : }
349 :
350 : /// Notify the caller the job has finished and unblock GC.
351 0 : fn notify_and_unblock(&self, id: GcCompactionJobId) {
352 0 : info!("compaction job id={} finished", id);
353 0 : let mut guard = self.inner.lock().unwrap();
354 0 : if let Some(items) = guard.guards.remove(&id) {
355 0 : if let Some(tx) = items.notify {
356 0 : let _ = tx.send(());
357 0 : }
358 0 : }
359 0 : }
360 :
361 0 : fn clear_running_job(&self) {
362 0 : let mut guard = self.inner.lock().unwrap();
363 0 : guard.running = None;
364 0 : }
365 :
366 0 : async fn handle_sub_compaction(
367 0 : &self,
368 0 : id: GcCompactionJobId,
369 0 : options: CompactOptions,
370 0 : timeline: &Arc<Timeline>,
371 0 : auto: bool,
372 0 : ) -> Result<(), CompactionError> {
373 0 : info!(
374 0 : "running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs"
375 : );
376 0 : let res = timeline
377 0 : .gc_compaction_split_jobs(
378 0 : GcCompactJob::from_compact_options(options.clone()),
379 0 : options.sub_compaction_max_job_size_mb,
380 0 : )
381 0 : .await;
382 0 : let jobs = match res {
383 0 : Ok(jobs) => jobs,
384 0 : Err(err) => {
385 0 : warn!("cannot split gc-compaction jobs: {}, unblocked gc", err);
386 0 : self.notify_and_unblock(id);
387 0 : return Err(err);
388 : }
389 : };
390 0 : if jobs.is_empty() {
391 0 : info!("no jobs to run, skipping scheduled compaction task");
392 0 : self.notify_and_unblock(id);
393 : } else {
394 0 : let jobs_len = jobs.len();
395 0 : let mut pending_tasks = Vec::new();
396 0 : // gc-compaction might pick more layers or fewer layers to compact. The L2 LSN does not need to be accurate.
397 0 : // And therefore, we simply assume the maximum LSN of all jobs is the expected L2 LSN.
398 0 : let expected_l2_lsn = jobs.iter().map(|job| job.compact_lsn_range.end).max();
399 0 : for job in jobs {
400 : // Unfortunately we need to convert the `GcCompactJob` back to `CompactionOptions`
401 : // until we do further refactors to allow directly call `compact_with_gc`.
402 0 : let mut flags: EnumSet<CompactFlags> = EnumSet::default();
403 0 : flags |= CompactFlags::EnhancedGcBottomMostCompaction;
404 0 : if job.dry_run {
405 0 : flags |= CompactFlags::DryRun;
406 0 : }
407 0 : if options.flags.contains(CompactFlags::YieldForL0) {
408 0 : flags |= CompactFlags::YieldForL0;
409 0 : }
410 0 : let options = CompactOptions {
411 0 : flags,
412 0 : sub_compaction: false,
413 0 : compact_key_range: Some(job.compact_key_range.into()),
414 0 : compact_lsn_range: Some(job.compact_lsn_range.into()),
415 0 : sub_compaction_max_job_size_mb: None,
416 0 : };
417 0 : pending_tasks.push(GcCompactionQueueItem::SubCompactionJob(options));
418 : }
419 :
420 0 : if !auto {
421 0 : pending_tasks.push(GcCompactionQueueItem::Notify(id, None));
422 0 : } else {
423 0 : pending_tasks.push(GcCompactionQueueItem::Notify(id, expected_l2_lsn));
424 0 : }
425 :
426 : {
427 0 : let mut guard = self.inner.lock().unwrap();
428 0 : let mut tasks = Vec::new();
429 0 : for task in pending_tasks {
430 0 : let id = guard.next_id();
431 0 : tasks.push((id, task));
432 0 : }
433 0 : tasks.reverse();
434 0 : for item in tasks {
435 0 : guard.queued.push_front(item);
436 0 : }
437 : }
438 0 : info!(
439 0 : "scheduled enhanced gc bottom-most compaction with sub-compaction, split into {} jobs",
440 : jobs_len
441 : );
442 : }
443 0 : Ok(())
444 0 : }
445 :
446 : /// Take a job from the queue and process it. Returns if there are still pending tasks.
447 0 : pub async fn iteration(
448 0 : &self,
449 0 : cancel: &CancellationToken,
450 0 : ctx: &RequestContext,
451 0 : gc_block: &GcBlock,
452 0 : timeline: &Arc<Timeline>,
453 0 : ) -> Result<CompactionOutcome, CompactionError> {
454 0 : let res = self.iteration_inner(cancel, ctx, gc_block, timeline).await;
455 0 : if let Err(err) = &res {
456 0 : log_compaction_error(err, None, cancel.is_cancelled(), true);
457 0 : }
458 0 : match res {
459 0 : Ok(res) => Ok(res),
460 0 : Err(CompactionError::ShuttingDown) => Err(CompactionError::ShuttingDown),
461 : Err(_) => {
462 : // There are some cases where traditional gc might collect some layer
463 : // files causing gc-compaction cannot read the full history of the key.
464 : // This needs to be resolved in the long-term by improving the compaction
465 : // process. For now, let's simply avoid such errors triggering the
466 : // circuit breaker.
467 0 : Ok(CompactionOutcome::Skipped)
468 : }
469 : }
470 0 : }
471 :
472 0 : async fn iteration_inner(
473 0 : &self,
474 0 : cancel: &CancellationToken,
475 0 : ctx: &RequestContext,
476 0 : gc_block: &GcBlock,
477 0 : timeline: &Arc<Timeline>,
478 0 : ) -> Result<CompactionOutcome, CompactionError> {
479 0 : let Ok(_one_op_at_a_time_guard) = self.consumer_lock.try_lock() else {
480 0 : return Err(CompactionError::AlreadyRunning(
481 0 : "cannot run gc-compaction because another gc-compaction is running. This should not happen because we only call this function from the gc-compaction queue.",
482 0 : ));
483 : };
484 : let has_pending_tasks;
485 0 : let mut yield_for_l0 = false;
486 0 : let Some((id, item)) = ({
487 0 : let mut guard = self.inner.lock().unwrap();
488 0 : if let Some((id, item)) = guard.queued.pop_front() {
489 0 : guard.running = Some((id, item.clone()));
490 0 : has_pending_tasks = !guard.queued.is_empty();
491 0 : Some((id, item))
492 : } else {
493 0 : has_pending_tasks = false;
494 0 : None
495 : }
496 : }) else {
497 0 : self.trigger_auto_compaction(timeline).await?;
498 : // Always yield after triggering auto-compaction. Gc-compaction is a low-priority task and we
499 : // have not implemented preemption mechanism yet. We always want to yield it to more important
500 : // tasks if there is one.
501 0 : return Ok(CompactionOutcome::Done);
502 : };
503 0 : match item {
504 0 : GcCompactionQueueItem::MetaJob { options, auto } => {
505 0 : if !options
506 0 : .flags
507 0 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
508 : {
509 0 : warn!(
510 0 : "ignoring scheduled compaction task: scheduled task must be gc compaction: {:?}",
511 : options
512 : );
513 0 : } else if options.sub_compaction {
514 0 : info!(
515 0 : "running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs"
516 : );
517 0 : self.handle_sub_compaction(id, options, timeline, auto)
518 0 : .await?;
519 : } else {
520 : // Auto compaction always enables sub-compaction so we don't need to handle update_l2_lsn
521 : // in this branch.
522 0 : let _gc_guard = match gc_block.start().await {
523 0 : Ok(guard) => guard,
524 0 : Err(e) => {
525 0 : self.notify_and_unblock(id);
526 0 : self.clear_running_job();
527 0 : return Err(CompactionError::Other(anyhow!(
528 0 : "cannot run gc-compaction because gc is blocked: {}",
529 0 : e
530 0 : )));
531 : }
532 : };
533 0 : let res = timeline.compact_with_options(cancel, options, ctx).await;
534 0 : let compaction_result = match res {
535 0 : Ok(res) => res,
536 0 : Err(err) => {
537 0 : warn!(%err, "failed to run gc-compaction");
538 0 : self.notify_and_unblock(id);
539 0 : self.clear_running_job();
540 0 : return Err(err);
541 : }
542 : };
543 0 : if compaction_result == CompactionOutcome::YieldForL0 {
544 0 : yield_for_l0 = true;
545 0 : }
546 : }
547 : }
548 0 : GcCompactionQueueItem::SubCompactionJob(options) => {
549 : // TODO: error handling, clear the queue if any task fails?
550 0 : let _gc_guard = match gc_block.start().await {
551 0 : Ok(guard) => guard,
552 0 : Err(e) => {
553 0 : self.clear_running_job();
554 0 : return Err(CompactionError::Other(anyhow!(
555 0 : "cannot run gc-compaction because gc is blocked: {}",
556 0 : e
557 0 : )));
558 : }
559 : };
560 0 : let res = timeline.compact_with_options(cancel, options, ctx).await;
561 0 : let compaction_result = match res {
562 0 : Ok(res) => res,
563 0 : Err(err) => {
564 0 : warn!(%err, "failed to run gc-compaction subcompaction job");
565 0 : self.clear_running_job();
566 0 : return Err(err);
567 : }
568 : };
569 0 : if compaction_result == CompactionOutcome::YieldForL0 {
570 0 : // We will permenantly give up a task if we yield for L0 compaction: the preempted subcompaction job won't be running
571 0 : // again. This ensures that we don't keep doing duplicated work within gc-compaction. Not directly returning here because
572 0 : // we need to clean things up before returning from the function.
573 0 : yield_for_l0 = true;
574 0 : }
575 : }
576 0 : GcCompactionQueueItem::Notify(id, l2_lsn) => {
577 0 : self.notify_and_unblock(id);
578 0 : if let Some(l2_lsn) = l2_lsn {
579 0 : let current_l2_lsn = timeline
580 0 : .get_gc_compaction_state()
581 0 : .map(|x| x.last_completed_lsn)
582 0 : .unwrap_or(Lsn::INVALID);
583 0 : if l2_lsn >= current_l2_lsn {
584 0 : info!("l2_lsn updated to {}", l2_lsn);
585 0 : timeline
586 0 : .update_gc_compaction_state(GcCompactionState {
587 0 : last_completed_lsn: l2_lsn,
588 0 : })
589 0 : .map_err(CompactionError::Other)?;
590 : } else {
591 0 : warn!(
592 0 : "l2_lsn updated to {} but it is less than the current l2_lsn {}",
593 : l2_lsn, current_l2_lsn
594 : );
595 : }
596 0 : }
597 : }
598 : }
599 0 : self.clear_running_job();
600 0 : Ok(if yield_for_l0 {
601 0 : tracing::info!("give up gc-compaction: yield for L0 compaction");
602 0 : CompactionOutcome::YieldForL0
603 0 : } else if has_pending_tasks {
604 0 : CompactionOutcome::Pending
605 : } else {
606 0 : CompactionOutcome::Done
607 : })
608 0 : }
609 :
610 : #[allow(clippy::type_complexity)]
611 0 : pub fn remaining_jobs(
612 0 : &self,
613 0 : ) -> (
614 0 : Option<(GcCompactionJobId, GcCompactionQueueItem)>,
615 0 : VecDeque<(GcCompactionJobId, GcCompactionQueueItem)>,
616 0 : ) {
617 0 : let guard = self.inner.lock().unwrap();
618 0 : (guard.running.clone(), guard.queued.clone())
619 0 : }
620 :
621 0 : pub fn remaining_jobs_num(&self) -> usize {
622 0 : let guard = self.inner.lock().unwrap();
623 0 : guard.queued.len() + if guard.running.is_some() { 1 } else { 0 }
624 0 : }
625 : }
626 :
627 : /// A job description for the gc-compaction job. This structure describes the rectangle range that the job will
628 : /// process. The exact layers that need to be compacted/rewritten will be generated when `compact_with_gc` gets
629 : /// called.
630 : #[derive(Debug, Clone)]
631 : pub(crate) struct GcCompactJob {
632 : pub dry_run: bool,
633 : /// The key range to be compacted. The compaction algorithm will only regenerate key-value pairs within this range
634 : /// [left inclusive, right exclusive), and other pairs will be rewritten into new files if necessary.
635 : pub compact_key_range: Range<Key>,
636 : /// The LSN range to be compacted. The compaction algorithm will use this range to determine the layers to be
637 : /// selected for the compaction, and it does not guarantee the generated layers will have exactly the same LSN range
638 : /// as specified here. The true range being compacted is `min_lsn/max_lsn` in [`GcCompactionJobDescription`].
639 : /// min_lsn will always <= the lower bound specified here, and max_lsn will always >= the upper bound specified here.
640 : pub compact_lsn_range: Range<Lsn>,
641 : }
642 :
643 : impl GcCompactJob {
644 112 : pub fn from_compact_options(options: CompactOptions) -> Self {
645 112 : GcCompactJob {
646 112 : dry_run: options.flags.contains(CompactFlags::DryRun),
647 112 : compact_key_range: options
648 112 : .compact_key_range
649 112 : .map(|x| x.into())
650 112 : .unwrap_or(Key::MIN..Key::MAX),
651 112 : compact_lsn_range: options
652 112 : .compact_lsn_range
653 112 : .map(|x| x.into())
654 112 : .unwrap_or(Lsn::INVALID..Lsn::MAX),
655 112 : }
656 112 : }
657 : }
658 :
659 : /// A job description for the gc-compaction job. This structure is generated when `compact_with_gc` is called
660 : /// and contains the exact layers we want to compact.
661 : pub struct GcCompactionJobDescription {
662 : /// All layers to read in the compaction job
663 : selected_layers: Vec<Layer>,
664 : /// GC cutoff of the job. This is the lowest LSN that will be accessed by the read/GC path and we need to
665 : /// keep all deltas <= this LSN or generate an image == this LSN.
666 : gc_cutoff: Lsn,
667 : /// LSNs to retain for the job. Read path will use this LSN so we need to keep deltas <= this LSN or
668 : /// generate an image == this LSN.
669 : retain_lsns_below_horizon: Vec<Lsn>,
670 : /// Maximum layer LSN processed in this compaction, that is max(end_lsn of layers). Exclusive. All data
671 : /// \>= this LSN will be kept and will not be rewritten.
672 : max_layer_lsn: Lsn,
673 : /// Minimum layer LSN processed in this compaction, that is min(start_lsn of layers). Inclusive.
674 : /// All access below (strict lower than `<`) this LSN will be routed through the normal read path instead of
675 : /// k-merge within gc-compaction.
676 : min_layer_lsn: Lsn,
677 : /// Only compact layers overlapping with this range.
678 : compaction_key_range: Range<Key>,
679 : /// When partial compaction is enabled, these layers need to be rewritten to ensure no overlap.
680 : /// This field is here solely for debugging. The field will not be read once the compaction
681 : /// description is generated.
682 : rewrite_layers: Vec<Arc<PersistentLayerDesc>>,
683 : }
684 :
685 : /// The result of bottom-most compaction for a single key at each LSN.
686 : #[derive(Debug)]
687 : #[cfg_attr(test, derive(PartialEq))]
688 : pub struct KeyLogAtLsn(pub Vec<(Lsn, Value)>);
689 :
690 : /// The result of bottom-most compaction.
691 : #[derive(Debug)]
692 : #[cfg_attr(test, derive(PartialEq))]
693 : pub(crate) struct KeyHistoryRetention {
694 : /// Stores logs to reconstruct the value at the given LSN, that is to say, logs <= LSN or image == LSN.
695 : pub(crate) below_horizon: Vec<(Lsn, KeyLogAtLsn)>,
696 : /// Stores logs to reconstruct the value at any LSN above the horizon, that is to say, log > LSN.
697 : pub(crate) above_horizon: KeyLogAtLsn,
698 : }
699 :
700 : impl KeyHistoryRetention {
701 : /// Hack: skip delta layer if we need to produce a layer of a same key-lsn.
702 : ///
703 : /// This can happen if we have removed some deltas in "the middle" of some existing layer's key-lsn-range.
704 : /// For example, consider the case where a single delta with range [0x10,0x50) exists.
705 : /// And we have branches at LSN 0x10, 0x20, 0x30.
706 : /// Then we delete branch @ 0x20.
707 : /// Bottom-most compaction may now delete the delta [0x20,0x30).
708 : /// And that wouldnt' change the shape of the layer.
709 : ///
710 : /// Note that bottom-most-gc-compaction never _adds_ new data in that case, only removes.
711 : ///
712 : /// `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.
713 148 : async fn discard_key(key: &PersistentLayerKey, tline: &Arc<Timeline>, dry_run: bool) -> bool {
714 148 : if dry_run {
715 0 : return true;
716 148 : }
717 148 : if LayerMap::is_l0(&key.key_range, key.is_delta) {
718 : // gc-compaction should not produce L0 deltas, otherwise it will break the layer order.
719 : // We should ignore such layers.
720 0 : return true;
721 148 : }
722 : let layer_generation;
723 : {
724 148 : let guard = tline.layers.read().await;
725 148 : if !guard.contains_key(key) {
726 104 : return false;
727 44 : }
728 44 : layer_generation = guard.get_from_key(key).metadata().generation;
729 44 : }
730 44 : if layer_generation == tline.generation {
731 44 : info!(
732 : key=%key,
733 : ?layer_generation,
734 0 : "discard layer due to duplicated layer key in the same generation",
735 : );
736 44 : true
737 : } else {
738 0 : false
739 : }
740 148 : }
741 :
742 : /// Pipe a history of a single key to the writers.
743 : ///
744 : /// If `image_writer` is none, the images will be placed into the delta layers.
745 : /// The delta writer will contain all images and deltas (below and above the horizon) except the bottom-most images.
746 : #[allow(clippy::too_many_arguments)]
747 1276 : async fn pipe_to(
748 1276 : self,
749 1276 : key: Key,
750 1276 : delta_writer: &mut SplitDeltaLayerWriter,
751 1276 : mut image_writer: Option<&mut SplitImageLayerWriter>,
752 1276 : stat: &mut CompactionStatistics,
753 1276 : ctx: &RequestContext,
754 1276 : ) -> anyhow::Result<()> {
755 1276 : let mut first_batch = true;
756 4088 : for (cutoff_lsn, KeyLogAtLsn(logs)) in self.below_horizon {
757 2812 : if first_batch {
758 1276 : if logs.len() == 1 && logs[0].1.is_image() {
759 1200 : let Value::Image(img) = &logs[0].1 else {
760 0 : unreachable!()
761 : };
762 1200 : stat.produce_image_key(img);
763 1200 : if let Some(image_writer) = image_writer.as_mut() {
764 1200 : image_writer.put_image(key, img.clone(), ctx).await?;
765 : } else {
766 0 : delta_writer
767 0 : .put_value(key, cutoff_lsn, Value::Image(img.clone()), ctx)
768 0 : .await?;
769 : }
770 : } else {
771 132 : for (lsn, val) in logs {
772 56 : stat.produce_key(&val);
773 56 : delta_writer.put_value(key, lsn, val, ctx).await?;
774 : }
775 : }
776 1276 : first_batch = false;
777 : } else {
778 1768 : for (lsn, val) in logs {
779 232 : stat.produce_key(&val);
780 232 : delta_writer.put_value(key, lsn, val, ctx).await?;
781 : }
782 : }
783 : }
784 1276 : let KeyLogAtLsn(above_horizon_logs) = self.above_horizon;
785 1392 : for (lsn, val) in above_horizon_logs {
786 116 : stat.produce_key(&val);
787 116 : delta_writer.put_value(key, lsn, val, ctx).await?;
788 : }
789 1276 : Ok(())
790 1276 : }
791 : }
792 :
793 : #[derive(Debug, Serialize, Default)]
794 : struct CompactionStatisticsNumSize {
795 : num: u64,
796 : size: u64,
797 : }
798 :
799 : #[derive(Debug, Serialize, Default)]
800 : pub struct CompactionStatistics {
801 : /// Delta layer visited (maybe compressed, physical size)
802 : delta_layer_visited: CompactionStatisticsNumSize,
803 : /// Image layer visited (maybe compressed, physical size)
804 : image_layer_visited: CompactionStatisticsNumSize,
805 : /// Delta layer produced (maybe compressed, physical size)
806 : delta_layer_produced: CompactionStatisticsNumSize,
807 : /// Image layer produced (maybe compressed, physical size)
808 : image_layer_produced: CompactionStatisticsNumSize,
809 : /// Delta layer discarded (maybe compressed, physical size of the layer being discarded instead of the original layer)
810 : delta_layer_discarded: CompactionStatisticsNumSize,
811 : /// Image layer discarded (maybe compressed, physical size of the layer being discarded instead of the original layer)
812 : image_layer_discarded: CompactionStatisticsNumSize,
813 : num_unique_keys_visited: usize,
814 : /// Delta visited (uncompressed, original size)
815 : wal_keys_visited: CompactionStatisticsNumSize,
816 : /// Image visited (uncompressed, original size)
817 : image_keys_visited: CompactionStatisticsNumSize,
818 : /// Delta produced (uncompressed, original size)
819 : wal_produced: CompactionStatisticsNumSize,
820 : /// Image produced (uncompressed, original size)
821 : image_produced: CompactionStatisticsNumSize,
822 :
823 : // Time spent in each phase
824 : time_acquire_lock_secs: f64,
825 : time_analyze_secs: f64,
826 : time_download_layer_secs: f64,
827 : time_to_first_kv_pair_secs: f64,
828 : time_main_loop_secs: f64,
829 : time_final_phase_secs: f64,
830 : time_total_secs: f64,
831 :
832 : // Summary
833 : /// Ratio of the key-value size after/before gc-compaction.
834 : uncompressed_retention_ratio: f64,
835 : /// Ratio of the physical size after/before gc-compaction.
836 : compressed_retention_ratio: f64,
837 : }
838 :
839 : impl CompactionStatistics {
840 2136 : fn estimated_size_of_value(val: &Value) -> usize {
841 876 : match val {
842 1260 : Value::Image(img) => img.len(),
843 0 : Value::WalRecord(NeonWalRecord::Postgres { rec, .. }) => rec.len(),
844 876 : _ => std::mem::size_of::<NeonWalRecord>(),
845 : }
846 2136 : }
847 3356 : fn estimated_size_of_key() -> usize {
848 3356 : KEY_SIZE // TODO: distinguish image layer and delta layer (count LSN in delta layer)
849 3356 : }
850 176 : fn visit_delta_layer(&mut self, size: u64) {
851 176 : self.delta_layer_visited.num += 1;
852 176 : self.delta_layer_visited.size += size;
853 176 : }
854 140 : fn visit_image_layer(&mut self, size: u64) {
855 140 : self.image_layer_visited.num += 1;
856 140 : self.image_layer_visited.size += size;
857 140 : }
858 1280 : fn on_unique_key_visited(&mut self) {
859 1280 : self.num_unique_keys_visited += 1;
860 1280 : }
861 492 : fn visit_wal_key(&mut self, val: &Value) {
862 492 : self.wal_keys_visited.num += 1;
863 492 : self.wal_keys_visited.size +=
864 492 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
865 492 : }
866 1260 : fn visit_image_key(&mut self, val: &Value) {
867 1260 : self.image_keys_visited.num += 1;
868 1260 : self.image_keys_visited.size +=
869 1260 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
870 1260 : }
871 404 : fn produce_key(&mut self, val: &Value) {
872 404 : match val {
873 20 : Value::Image(img) => self.produce_image_key(img),
874 384 : Value::WalRecord(_) => self.produce_wal_key(val),
875 : }
876 404 : }
877 384 : fn produce_wal_key(&mut self, val: &Value) {
878 384 : self.wal_produced.num += 1;
879 384 : self.wal_produced.size +=
880 384 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
881 384 : }
882 1220 : fn produce_image_key(&mut self, val: &Bytes) {
883 1220 : self.image_produced.num += 1;
884 1220 : self.image_produced.size += val.len() as u64 + Self::estimated_size_of_key() as u64;
885 1220 : }
886 28 : fn discard_delta_layer(&mut self, original_size: u64) {
887 28 : self.delta_layer_discarded.num += 1;
888 28 : self.delta_layer_discarded.size += original_size;
889 28 : }
890 16 : fn discard_image_layer(&mut self, original_size: u64) {
891 16 : self.image_layer_discarded.num += 1;
892 16 : self.image_layer_discarded.size += original_size;
893 16 : }
894 48 : fn produce_delta_layer(&mut self, size: u64) {
895 48 : self.delta_layer_produced.num += 1;
896 48 : self.delta_layer_produced.size += size;
897 48 : }
898 60 : fn produce_image_layer(&mut self, size: u64) {
899 60 : self.image_layer_produced.num += 1;
900 60 : self.image_layer_produced.size += size;
901 60 : }
902 104 : fn finalize(&mut self) {
903 104 : let original_key_value_size = self.image_keys_visited.size + self.wal_keys_visited.size;
904 104 : let produced_key_value_size = self.image_produced.size + self.wal_produced.size;
905 104 : self.uncompressed_retention_ratio =
906 104 : produced_key_value_size as f64 / (original_key_value_size as f64 + 1.0); // avoid div by 0
907 104 : let original_physical_size = self.image_layer_visited.size + self.delta_layer_visited.size;
908 104 : let produced_physical_size = self.image_layer_produced.size
909 104 : + self.delta_layer_produced.size
910 104 : + self.image_layer_discarded.size
911 104 : + self.delta_layer_discarded.size; // Also include the discarded layers to make the ratio accurate
912 104 : self.compressed_retention_ratio =
913 104 : produced_physical_size as f64 / (original_physical_size as f64 + 1.0); // avoid div by 0
914 104 : }
915 : }
916 :
917 : #[derive(Default, Debug, Clone, Copy, PartialEq, Eq)]
918 : pub enum CompactionOutcome {
919 : #[default]
920 : /// No layers need to be compacted after this round. Compaction doesn't need
921 : /// to be immediately scheduled.
922 : Done,
923 : /// Still has pending layers to be compacted after this round. Ideally, the scheduler
924 : /// should immediately schedule another compaction.
925 : Pending,
926 : /// A timeline needs L0 compaction. Yield and schedule an immediate L0 compaction pass (only
927 : /// guaranteed when `compaction_l0_first` is enabled).
928 : YieldForL0,
929 : /// Compaction was skipped, because the timeline is ineligible for compaction.
930 : Skipped,
931 : }
932 :
933 : impl Timeline {
934 : /// TODO: cancellation
935 : ///
936 : /// Returns whether the compaction has pending tasks.
937 728 : pub(crate) async fn compact_legacy(
938 728 : self: &Arc<Self>,
939 728 : cancel: &CancellationToken,
940 728 : options: CompactOptions,
941 728 : ctx: &RequestContext,
942 728 : ) -> Result<CompactionOutcome, CompactionError> {
943 728 : if options
944 728 : .flags
945 728 : .contains(CompactFlags::EnhancedGcBottomMostCompaction)
946 : {
947 0 : self.compact_with_gc(cancel, options, ctx).await?;
948 0 : return Ok(CompactionOutcome::Done);
949 728 : }
950 728 :
951 728 : if options.flags.contains(CompactFlags::DryRun) {
952 0 : return Err(CompactionError::Other(anyhow!(
953 0 : "dry-run mode is not supported for legacy compaction for now"
954 0 : )));
955 728 : }
956 728 :
957 728 : if options.compact_key_range.is_some() || options.compact_lsn_range.is_some() {
958 : // maybe useful in the future? could implement this at some point
959 0 : return Err(CompactionError::Other(anyhow!(
960 0 : "compaction range is not supported for legacy compaction for now"
961 0 : )));
962 728 : }
963 728 :
964 728 : // High level strategy for compaction / image creation:
965 728 : //
966 728 : // 1. First, do a L0 compaction to ensure we move the L0
967 728 : // layers into the historic layer map get flat levels of
968 728 : // layers. If we did not compact all L0 layers, we will
969 728 : // prioritize compacting the timeline again and not do
970 728 : // any of the compactions below.
971 728 : //
972 728 : // 2. Then, calculate the desired "partitioning" of the
973 728 : // currently in-use key space. The goal is to partition the
974 728 : // key space into roughly fixed-size chunks, but also take into
975 728 : // account any existing image layers, and try to align the
976 728 : // chunk boundaries with the existing image layers to avoid
977 728 : // too much churn. Also try to align chunk boundaries with
978 728 : // relation boundaries. In principle, we don't know about
979 728 : // relation boundaries here, we just deal with key-value
980 728 : // pairs, and the code in pgdatadir_mapping.rs knows how to
981 728 : // map relations into key-value pairs. But in practice we know
982 728 : // that 'field6' is the block number, and the fields 1-5
983 728 : // identify a relation. This is just an optimization,
984 728 : // though.
985 728 : //
986 728 : // 3. Once we know the partitioning, for each partition,
987 728 : // decide if it's time to create a new image layer. The
988 728 : // criteria is: there has been too much "churn" since the last
989 728 : // image layer? The "churn" is fuzzy concept, it's a
990 728 : // combination of too many delta files, or too much WAL in
991 728 : // total in the delta file. Or perhaps: if creating an image
992 728 : // file would allow to delete some older files.
993 728 : //
994 728 : // 4. In the end, if the tenant gets auto-sharded, we will run
995 728 : // a shard-ancestor compaction.
996 728 :
997 728 : // Is the timeline being deleted?
998 728 : if self.is_stopping() {
999 0 : trace!("Dropping out of compaction on timeline shutdown");
1000 0 : return Err(CompactionError::ShuttingDown);
1001 728 : }
1002 728 :
1003 728 : let target_file_size = self.get_checkpoint_distance();
1004 :
1005 : // Define partitioning schema if needed
1006 :
1007 : // 1. L0 Compact
1008 728 : let l0_outcome = {
1009 728 : let timer = self.metrics.compact_time_histo.start_timer();
1010 728 : let l0_outcome = self
1011 728 : .compact_level0(
1012 728 : target_file_size,
1013 728 : options.flags.contains(CompactFlags::ForceL0Compaction),
1014 728 : ctx,
1015 728 : )
1016 728 : .await?;
1017 728 : timer.stop_and_record();
1018 728 : l0_outcome
1019 728 : };
1020 728 :
1021 728 : if options.flags.contains(CompactFlags::OnlyL0Compaction) {
1022 0 : return Ok(l0_outcome);
1023 728 : }
1024 728 :
1025 728 : // Yield if we have pending L0 compaction. The scheduler will do another pass.
1026 728 : if (l0_outcome == CompactionOutcome::Pending || l0_outcome == CompactionOutcome::YieldForL0)
1027 0 : && options.flags.contains(CompactFlags::YieldForL0)
1028 : {
1029 0 : info!("image/ancestor compaction yielding for L0 compaction");
1030 0 : return Ok(CompactionOutcome::YieldForL0);
1031 728 : }
1032 728 :
1033 728 : // 2. Repartition and create image layers if necessary
1034 728 : match self
1035 728 : .repartition(
1036 728 : self.get_last_record_lsn(),
1037 728 : self.get_compaction_target_size(),
1038 728 : options.flags,
1039 728 : ctx,
1040 728 : )
1041 728 : .await
1042 : {
1043 728 : Ok(((dense_partitioning, sparse_partitioning), lsn)) => {
1044 728 : // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
1045 728 : let image_ctx = RequestContextBuilder::from(ctx)
1046 728 : .access_stats_behavior(AccessStatsBehavior::Skip)
1047 728 : .attached_child();
1048 728 :
1049 728 : let mut partitioning = dense_partitioning;
1050 728 : partitioning
1051 728 : .parts
1052 728 : .extend(sparse_partitioning.into_dense().parts);
1053 :
1054 : // 3. Create new image layers for partitions that have been modified "enough".
1055 728 : let (image_layers, outcome) = self
1056 728 : .create_image_layers(
1057 728 : &partitioning,
1058 728 : lsn,
1059 728 : if options
1060 728 : .flags
1061 728 : .contains(CompactFlags::ForceImageLayerCreation)
1062 : {
1063 28 : ImageLayerCreationMode::Force
1064 : } else {
1065 700 : ImageLayerCreationMode::Try
1066 : },
1067 728 : &image_ctx,
1068 728 : self.last_image_layer_creation_status
1069 728 : .load()
1070 728 : .as_ref()
1071 728 : .clone(),
1072 728 : options.flags.contains(CompactFlags::YieldForL0),
1073 728 : )
1074 728 : .await
1075 728 : .inspect_err(|err| {
1076 : if let CreateImageLayersError::GetVectoredError(
1077 : GetVectoredError::MissingKey(_),
1078 0 : ) = err
1079 : {
1080 0 : critical!("missing key during compaction: {err:?}");
1081 0 : }
1082 728 : })?;
1083 :
1084 728 : self.last_image_layer_creation_status
1085 728 : .store(Arc::new(outcome.clone()));
1086 728 :
1087 728 : self.upload_new_image_layers(image_layers)?;
1088 728 : if let LastImageLayerCreationStatus::Incomplete { .. } = outcome {
1089 : // Yield and do not do any other kind of compaction.
1090 0 : info!(
1091 0 : "skipping shard ancestor compaction due to pending image layer generation tasks (preempted by L0 compaction)."
1092 : );
1093 0 : return Ok(CompactionOutcome::YieldForL0);
1094 728 : }
1095 : }
1096 :
1097 : // Suppress errors when cancelled.
1098 0 : Err(_) if self.cancel.is_cancelled() => {}
1099 0 : Err(err) if err.is_cancel() => {}
1100 :
1101 : // Alert on critical errors that indicate data corruption.
1102 0 : Err(err) if err.is_critical() => {
1103 0 : critical!("could not compact, repartitioning keyspace failed: {err:?}");
1104 : }
1105 :
1106 : // Log other errors. No partitioning? This is normal, if the timeline was just created
1107 : // as an empty timeline. Also in unit tests, when we use the timeline as a simple
1108 : // key-value store, ignoring the datadir layout. Log the error but continue.
1109 0 : Err(err) => error!("could not compact, repartitioning keyspace failed: {err:?}"),
1110 : };
1111 :
1112 728 : let partition_count = self.partitioning.read().0.0.parts.len();
1113 728 :
1114 728 : // 4. Shard ancestor compaction
1115 728 :
1116 728 : if self.shard_identity.count >= ShardCount::new(2) {
1117 : // Limit the number of layer rewrites to the number of partitions: this means its
1118 : // runtime should be comparable to a full round of image layer creations, rather than
1119 : // being potentially much longer.
1120 0 : let rewrite_max = partition_count;
1121 0 :
1122 0 : self.compact_shard_ancestors(rewrite_max, ctx).await?;
1123 728 : }
1124 :
1125 728 : Ok(CompactionOutcome::Done)
1126 728 : }
1127 :
1128 : /// Check for layers that are elegible to be rewritten:
1129 : /// - Shard splitting: After a shard split, ancestor layers beyond pitr_interval, so that
1130 : /// we don't indefinitely retain keys in this shard that aren't needed.
1131 : /// - For future use: layers beyond pitr_interval that are in formats we would
1132 : /// rather not maintain compatibility with indefinitely.
1133 : ///
1134 : /// Note: this phase may read and write many gigabytes of data: use rewrite_max to bound
1135 : /// how much work it will try to do in each compaction pass.
1136 0 : async fn compact_shard_ancestors(
1137 0 : self: &Arc<Self>,
1138 0 : rewrite_max: usize,
1139 0 : ctx: &RequestContext,
1140 0 : ) -> Result<(), CompactionError> {
1141 0 : let mut drop_layers = Vec::new();
1142 0 : let mut layers_to_rewrite: Vec<Layer> = Vec::new();
1143 0 :
1144 0 : // We will use the Lsn cutoff of the last GC as a threshold for rewriting layers: if a
1145 0 : // layer is behind this Lsn, it indicates that the layer is being retained beyond the
1146 0 : // pitr_interval, for example because a branchpoint references it.
1147 0 : //
1148 0 : // Holding this read guard also blocks [`Self::gc_timeline`] from entering while we
1149 0 : // are rewriting layers.
1150 0 : let latest_gc_cutoff = self.get_applied_gc_cutoff_lsn();
1151 0 :
1152 0 : tracing::info!(
1153 0 : "starting shard ancestor compaction, latest_gc_cutoff: {}, pitr cutoff {}",
1154 0 : *latest_gc_cutoff,
1155 0 : self.gc_info.read().unwrap().cutoffs.time
1156 : );
1157 :
1158 0 : let layers = self.layers.read().await;
1159 0 : for layer_desc in layers.layer_map()?.iter_historic_layers() {
1160 0 : let layer = layers.get_from_desc(&layer_desc);
1161 0 : if layer.metadata().shard.shard_count == self.shard_identity.count {
1162 : // This layer does not belong to a historic ancestor, no need to re-image it.
1163 0 : continue;
1164 0 : }
1165 0 :
1166 0 : // This layer was created on an ancestor shard: check if it contains any data for this shard.
1167 0 : let sharded_range = ShardedRange::new(layer_desc.get_key_range(), &self.shard_identity);
1168 0 : let layer_local_page_count = sharded_range.page_count();
1169 0 : let layer_raw_page_count = ShardedRange::raw_size(&layer_desc.get_key_range());
1170 0 : if layer_local_page_count == 0 {
1171 : // This ancestral layer only covers keys that belong to other shards.
1172 : // We include the full metadata in the log: if we had some critical bug that caused
1173 : // us to incorrectly drop layers, this would simplify manually debugging + reinstating those layers.
1174 0 : info!(%layer, old_metadata=?layer.metadata(),
1175 0 : "dropping layer after shard split, contains no keys for this shard.",
1176 : );
1177 :
1178 0 : if cfg!(debug_assertions) {
1179 : // Expensive, exhaustive check of keys in this layer: this guards against ShardedRange's calculations being
1180 : // wrong. If ShardedRange claims the local page count is zero, then no keys in this layer
1181 : // should be !is_key_disposable()
1182 : // TODO: exclude sparse keyspace from this check, otherwise it will infinitely loop.
1183 0 : let range = layer_desc.get_key_range();
1184 0 : let mut key = range.start;
1185 0 : while key < range.end {
1186 0 : debug_assert!(self.shard_identity.is_key_disposable(&key));
1187 0 : key = key.next();
1188 : }
1189 0 : }
1190 :
1191 0 : drop_layers.push(layer);
1192 0 : continue;
1193 0 : } else if layer_local_page_count != u32::MAX
1194 0 : && layer_local_page_count == layer_raw_page_count
1195 : {
1196 0 : debug!(%layer,
1197 0 : "layer is entirely shard local ({} keys), no need to filter it",
1198 : layer_local_page_count
1199 : );
1200 0 : continue;
1201 0 : }
1202 0 :
1203 0 : // Don't bother re-writing a layer unless it will at least halve its size
1204 0 : if layer_local_page_count != u32::MAX
1205 0 : && layer_local_page_count > layer_raw_page_count / 2
1206 : {
1207 0 : debug!(%layer,
1208 0 : "layer is already mostly local ({}/{}), not rewriting",
1209 : layer_local_page_count,
1210 : layer_raw_page_count
1211 : );
1212 0 : }
1213 :
1214 : // Don't bother re-writing a layer if it is within the PITR window: it will age-out eventually
1215 : // without incurring the I/O cost of a rewrite.
1216 0 : if layer_desc.get_lsn_range().end >= *latest_gc_cutoff {
1217 0 : debug!(%layer, "Skipping rewrite of layer still in GC window ({} >= {})",
1218 0 : layer_desc.get_lsn_range().end, *latest_gc_cutoff);
1219 0 : continue;
1220 0 : }
1221 0 :
1222 0 : if layer_desc.is_delta() {
1223 : // We do not yet implement rewrite of delta layers
1224 0 : debug!(%layer, "Skipping rewrite of delta layer");
1225 0 : continue;
1226 0 : }
1227 0 :
1228 0 : // Only rewrite layers if their generations differ. This guarantees:
1229 0 : // - that local rewrite is safe, as local layer paths will differ between existing layer and rewritten one
1230 0 : // - that the layer is persistent in remote storage, as we only see old-generation'd layer via loading from remote storage
1231 0 : if layer.metadata().generation == self.generation {
1232 0 : debug!(%layer, "Skipping rewrite, is not from old generation");
1233 0 : continue;
1234 0 : }
1235 0 :
1236 0 : if layers_to_rewrite.len() >= rewrite_max {
1237 0 : tracing::info!(%layer, "Will rewrite layer on a future compaction, already rewrote {}",
1238 0 : layers_to_rewrite.len()
1239 : );
1240 0 : continue;
1241 0 : }
1242 0 :
1243 0 : // Fall through: all our conditions for doing a rewrite passed.
1244 0 : layers_to_rewrite.push(layer);
1245 : }
1246 :
1247 : // Drop read lock on layer map before we start doing time-consuming I/O
1248 0 : drop(layers);
1249 0 :
1250 0 : let mut replace_image_layers = Vec::new();
1251 :
1252 0 : for layer in layers_to_rewrite {
1253 0 : if self.cancel.is_cancelled() {
1254 0 : return Err(CompactionError::ShuttingDown);
1255 0 : }
1256 0 :
1257 0 : tracing::info!(layer=%layer, "Rewriting layer after shard split...");
1258 0 : let mut image_layer_writer = ImageLayerWriter::new(
1259 0 : self.conf,
1260 0 : self.timeline_id,
1261 0 : self.tenant_shard_id,
1262 0 : &layer.layer_desc().key_range,
1263 0 : layer.layer_desc().image_layer_lsn(),
1264 0 : ctx,
1265 0 : )
1266 0 : .await
1267 0 : .map_err(CompactionError::Other)?;
1268 :
1269 : // Safety of layer rewrites:
1270 : // - We are writing to a different local file path than we are reading from, so the old Layer
1271 : // cannot interfere with the new one.
1272 : // - In the page cache, contents for a particular VirtualFile are stored with a file_id that
1273 : // is different for two layers with the same name (in `ImageLayerInner::new` we always
1274 : // acquire a fresh id from [`crate::page_cache::next_file_id`]. So readers do not risk
1275 : // reading the index from one layer file, and then data blocks from the rewritten layer file.
1276 : // - Any readers that have a reference to the old layer will keep it alive until they are done
1277 : // with it. If they are trying to promote from remote storage, that will fail, but this is the same
1278 : // as for compaction generally: compaction is allowed to delete layers that readers might be trying to use.
1279 : // - We do not run concurrently with other kinds of compaction, so the only layer map writes we race with are:
1280 : // - GC, which at worst witnesses us "undelete" a layer that they just deleted.
1281 : // - ingestion, which only inserts layers, therefore cannot collide with us.
1282 0 : let resident = layer.download_and_keep_resident(ctx).await?;
1283 :
1284 0 : let keys_written = resident
1285 0 : .filter(&self.shard_identity, &mut image_layer_writer, ctx)
1286 0 : .await?;
1287 :
1288 0 : if keys_written > 0 {
1289 0 : let (desc, path) = image_layer_writer
1290 0 : .finish(ctx)
1291 0 : .await
1292 0 : .map_err(CompactionError::Other)?;
1293 0 : let new_layer = Layer::finish_creating(self.conf, self, desc, &path)
1294 0 : .map_err(CompactionError::Other)?;
1295 0 : tracing::info!(layer=%new_layer, "Rewrote layer, {} -> {} bytes",
1296 0 : layer.metadata().file_size,
1297 0 : new_layer.metadata().file_size);
1298 :
1299 0 : replace_image_layers.push((layer, new_layer));
1300 0 : } else {
1301 0 : // Drop the old layer. Usually for this case we would already have noticed that
1302 0 : // the layer has no data for us with the ShardedRange check above, but
1303 0 : drop_layers.push(layer);
1304 0 : }
1305 : }
1306 :
1307 : // At this point, we have replaced local layer files with their rewritten form, but not yet uploaded
1308 : // metadata to reflect that. If we restart here, the replaced layer files will look invalid (size mismatch
1309 : // to remote index) and be removed. This is inefficient but safe.
1310 0 : fail::fail_point!("compact-shard-ancestors-localonly");
1311 0 :
1312 0 : // Update the LayerMap so that readers will use the new layers, and enqueue it for writing to remote storage
1313 0 : self.rewrite_layers(replace_image_layers, drop_layers)
1314 0 : .await?;
1315 :
1316 0 : fail::fail_point!("compact-shard-ancestors-enqueued");
1317 0 :
1318 0 : // We wait for all uploads to complete before finishing this compaction stage. This is not
1319 0 : // necessary for correctness, but it simplifies testing, and avoids proceeding with another
1320 0 : // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
1321 0 : // load.
1322 0 : match self.remote_client.wait_completion().await {
1323 0 : Ok(()) => (),
1324 0 : Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
1325 : Err(WaitCompletionError::UploadQueueShutDownOrStopped) => {
1326 0 : return Err(CompactionError::ShuttingDown);
1327 : }
1328 : }
1329 :
1330 0 : fail::fail_point!("compact-shard-ancestors-persistent");
1331 0 :
1332 0 : Ok(())
1333 0 : }
1334 :
1335 : /// Update the LayerVisibilityHint of layers covered by image layers, based on whether there is
1336 : /// an image layer between them and the most recent readable LSN (branch point or tip of timeline). The
1337 : /// purpose of the visibility hint is to record which layers need to be available to service reads.
1338 : ///
1339 : /// The result may be used as an input to eviction and secondary downloads to de-prioritize layers
1340 : /// that we know won't be needed for reads.
1341 476 : pub(crate) async fn update_layer_visibility(
1342 476 : &self,
1343 476 : ) -> Result<(), super::layer_manager::Shutdown> {
1344 476 : let head_lsn = self.get_last_record_lsn();
1345 :
1346 : // We will sweep through layers in reverse-LSN order. We only do historic layers. L0 deltas
1347 : // are implicitly left visible, because LayerVisibilityHint's default is Visible, and we never modify it here.
1348 : // Note that L0 deltas _can_ be covered by image layers, but we consider them 'visible' because we anticipate that
1349 : // they will be subject to L0->L1 compaction in the near future.
1350 476 : let layer_manager = self.layers.read().await;
1351 476 : let layer_map = layer_manager.layer_map()?;
1352 :
1353 476 : let readable_points = {
1354 476 : let children = self.gc_info.read().unwrap().retain_lsns.clone();
1355 476 :
1356 476 : let mut readable_points = Vec::with_capacity(children.len() + 1);
1357 476 : for (child_lsn, _child_timeline_id, is_offloaded) in &children {
1358 0 : if *is_offloaded == MaybeOffloaded::Yes {
1359 0 : continue;
1360 0 : }
1361 0 : readable_points.push(*child_lsn);
1362 : }
1363 476 : readable_points.push(head_lsn);
1364 476 : readable_points
1365 476 : };
1366 476 :
1367 476 : let (layer_visibility, covered) = layer_map.get_visibility(readable_points);
1368 1200 : for (layer_desc, visibility) in layer_visibility {
1369 724 : // FIXME: a more efficiency bulk zip() through the layers rather than NlogN getting each one
1370 724 : let layer = layer_manager.get_from_desc(&layer_desc);
1371 724 : layer.set_visibility(visibility);
1372 724 : }
1373 :
1374 : // TODO: publish our covered KeySpace to our parent, so that when they update their visibility, they can
1375 : // avoid assuming that everything at a branch point is visible.
1376 476 : drop(covered);
1377 476 : Ok(())
1378 476 : }
1379 :
1380 : /// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
1381 : /// as Level 1 files. Returns whether the L0 layers are fully compacted.
1382 728 : async fn compact_level0(
1383 728 : self: &Arc<Self>,
1384 728 : target_file_size: u64,
1385 728 : force_compaction_ignore_threshold: bool,
1386 728 : ctx: &RequestContext,
1387 728 : ) -> Result<CompactionOutcome, CompactionError> {
1388 : let CompactLevel0Phase1Result {
1389 728 : new_layers,
1390 728 : deltas_to_compact,
1391 728 : outcome,
1392 : } = {
1393 728 : let phase1_span = info_span!("compact_level0_phase1");
1394 728 : let ctx = ctx.attached_child();
1395 728 : let mut stats = CompactLevel0Phase1StatsBuilder {
1396 728 : version: Some(2),
1397 728 : tenant_id: Some(self.tenant_shard_id),
1398 728 : timeline_id: Some(self.timeline_id),
1399 728 : ..Default::default()
1400 728 : };
1401 728 :
1402 728 : let begin = tokio::time::Instant::now();
1403 728 : let phase1_layers_locked = self.layers.read().await;
1404 728 : let now = tokio::time::Instant::now();
1405 728 : stats.read_lock_acquisition_micros =
1406 728 : DurationRecorder::Recorded(RecordedDuration(now - begin), now);
1407 728 : self.compact_level0_phase1(
1408 728 : phase1_layers_locked,
1409 728 : stats,
1410 728 : target_file_size,
1411 728 : force_compaction_ignore_threshold,
1412 728 : &ctx,
1413 728 : )
1414 728 : .instrument(phase1_span)
1415 728 : .await?
1416 : };
1417 :
1418 728 : if new_layers.is_empty() && deltas_to_compact.is_empty() {
1419 : // nothing to do
1420 672 : return Ok(CompactionOutcome::Done);
1421 56 : }
1422 56 :
1423 56 : self.finish_compact_batch(&new_layers, &Vec::new(), &deltas_to_compact)
1424 56 : .await?;
1425 56 : Ok(outcome)
1426 728 : }
1427 :
1428 : /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
1429 728 : async fn compact_level0_phase1<'a>(
1430 728 : self: &'a Arc<Self>,
1431 728 : guard: tokio::sync::RwLockReadGuard<'a, LayerManager>,
1432 728 : mut stats: CompactLevel0Phase1StatsBuilder,
1433 728 : target_file_size: u64,
1434 728 : force_compaction_ignore_threshold: bool,
1435 728 : ctx: &RequestContext,
1436 728 : ) -> Result<CompactLevel0Phase1Result, CompactionError> {
1437 728 : stats.read_lock_held_spawn_blocking_startup_micros =
1438 728 : stats.read_lock_acquisition_micros.till_now(); // set by caller
1439 728 : let layers = guard.layer_map()?;
1440 728 : let level0_deltas = layers.level0_deltas();
1441 728 : stats.level0_deltas_count = Some(level0_deltas.len());
1442 728 :
1443 728 : // Only compact if enough layers have accumulated.
1444 728 : let threshold = self.get_compaction_threshold();
1445 728 : if level0_deltas.is_empty() || level0_deltas.len() < threshold {
1446 672 : if force_compaction_ignore_threshold {
1447 0 : if !level0_deltas.is_empty() {
1448 0 : info!(
1449 0 : level0_deltas = level0_deltas.len(),
1450 0 : threshold, "too few deltas to compact, but forcing compaction"
1451 : );
1452 : } else {
1453 0 : info!(
1454 0 : level0_deltas = level0_deltas.len(),
1455 0 : threshold, "too few deltas to compact, cannot force compaction"
1456 : );
1457 0 : return Ok(CompactLevel0Phase1Result::default());
1458 : }
1459 : } else {
1460 672 : debug!(
1461 0 : level0_deltas = level0_deltas.len(),
1462 0 : threshold, "too few deltas to compact"
1463 : );
1464 672 : return Ok(CompactLevel0Phase1Result::default());
1465 : }
1466 56 : }
1467 :
1468 56 : let mut level0_deltas = level0_deltas
1469 56 : .iter()
1470 804 : .map(|x| guard.get_from_desc(x))
1471 56 : .collect::<Vec<_>>();
1472 56 :
1473 56 : // Gather the files to compact in this iteration.
1474 56 : //
1475 56 : // Start with the oldest Level 0 delta file, and collect any other
1476 56 : // level 0 files that form a contiguous sequence, such that the end
1477 56 : // LSN of previous file matches the start LSN of the next file.
1478 56 : //
1479 56 : // Note that if the files don't form such a sequence, we might
1480 56 : // "compact" just a single file. That's a bit pointless, but it allows
1481 56 : // us to get rid of the level 0 file, and compact the other files on
1482 56 : // the next iteration. This could probably made smarter, but such
1483 56 : // "gaps" in the sequence of level 0 files should only happen in case
1484 56 : // of a crash, partial download from cloud storage, or something like
1485 56 : // that, so it's not a big deal in practice.
1486 1496 : level0_deltas.sort_by_key(|l| l.layer_desc().lsn_range.start);
1487 56 : let mut level0_deltas_iter = level0_deltas.iter();
1488 56 :
1489 56 : let first_level0_delta = level0_deltas_iter.next().unwrap();
1490 56 : let mut prev_lsn_end = first_level0_delta.layer_desc().lsn_range.end;
1491 56 : let mut deltas_to_compact = Vec::with_capacity(level0_deltas.len());
1492 56 :
1493 56 : // Accumulate the size of layers in `deltas_to_compact`
1494 56 : let mut deltas_to_compact_bytes = 0;
1495 56 :
1496 56 : // Under normal circumstances, we will accumulate up to compaction_upper_limit L0s of size
1497 56 : // checkpoint_distance each. To avoid edge cases using extra system resources, bound our
1498 56 : // work in this function to only operate on this much delta data at once.
1499 56 : //
1500 56 : // In general, compaction_threshold should be <= compaction_upper_limit, but in case that
1501 56 : // the constraint is not respected, we use the larger of the two.
1502 56 : let delta_size_limit = std::cmp::max(
1503 56 : self.get_compaction_upper_limit(),
1504 56 : self.get_compaction_threshold(),
1505 56 : ) as u64
1506 56 : * std::cmp::max(self.get_checkpoint_distance(), DEFAULT_CHECKPOINT_DISTANCE);
1507 56 :
1508 56 : let mut fully_compacted = true;
1509 56 :
1510 56 : deltas_to_compact.push(first_level0_delta.download_and_keep_resident(ctx).await?);
1511 804 : for l in level0_deltas_iter {
1512 748 : let lsn_range = &l.layer_desc().lsn_range;
1513 748 :
1514 748 : if lsn_range.start != prev_lsn_end {
1515 0 : break;
1516 748 : }
1517 748 : deltas_to_compact.push(l.download_and_keep_resident(ctx).await?);
1518 748 : deltas_to_compact_bytes += l.metadata().file_size;
1519 748 : prev_lsn_end = lsn_range.end;
1520 748 :
1521 748 : if deltas_to_compact_bytes >= delta_size_limit {
1522 0 : info!(
1523 0 : l0_deltas_selected = deltas_to_compact.len(),
1524 0 : l0_deltas_total = level0_deltas.len(),
1525 0 : "L0 compaction picker hit max delta layer size limit: {}",
1526 : delta_size_limit
1527 : );
1528 0 : fully_compacted = false;
1529 0 :
1530 0 : // Proceed with compaction, but only a subset of L0s
1531 0 : break;
1532 748 : }
1533 : }
1534 56 : let lsn_range = Range {
1535 56 : start: deltas_to_compact
1536 56 : .first()
1537 56 : .unwrap()
1538 56 : .layer_desc()
1539 56 : .lsn_range
1540 56 : .start,
1541 56 : end: deltas_to_compact.last().unwrap().layer_desc().lsn_range.end,
1542 56 : };
1543 56 :
1544 56 : info!(
1545 0 : "Starting Level0 compaction in LSN range {}-{} for {} layers ({} deltas in total)",
1546 0 : lsn_range.start,
1547 0 : lsn_range.end,
1548 0 : deltas_to_compact.len(),
1549 0 : level0_deltas.len()
1550 : );
1551 :
1552 804 : for l in deltas_to_compact.iter() {
1553 804 : info!("compact includes {l}");
1554 : }
1555 :
1556 : // We don't need the original list of layers anymore. Drop it so that
1557 : // we don't accidentally use it later in the function.
1558 56 : drop(level0_deltas);
1559 56 :
1560 56 : stats.read_lock_held_prerequisites_micros = stats
1561 56 : .read_lock_held_spawn_blocking_startup_micros
1562 56 : .till_now();
1563 :
1564 : // TODO: replace with streaming k-merge
1565 56 : let all_keys = {
1566 56 : let mut all_keys = Vec::new();
1567 804 : for l in deltas_to_compact.iter() {
1568 804 : if self.cancel.is_cancelled() {
1569 0 : return Err(CompactionError::ShuttingDown);
1570 804 : }
1571 804 : let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
1572 804 : let keys = delta
1573 804 : .index_entries(ctx)
1574 804 : .await
1575 804 : .map_err(CompactionError::Other)?;
1576 804 : all_keys.extend(keys);
1577 : }
1578 : // The current stdlib sorting implementation is designed in a way where it is
1579 : // particularly fast where the slice is made up of sorted sub-ranges.
1580 8847530 : all_keys.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
1581 56 : all_keys
1582 56 : };
1583 56 :
1584 56 : stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
1585 :
1586 : // Determine N largest holes where N is number of compacted layers. The vec is sorted by key range start.
1587 : //
1588 : // A hole is a key range for which this compaction doesn't have any WAL records.
1589 : // Our goal in this compaction iteration is to avoid creating L1s that, in terms of their key range,
1590 : // cover the hole, but actually don't contain any WAL records for that key range.
1591 : // The reason is that the mere stack of L1s (`count_deltas`) triggers image layer creation (`create_image_layers`).
1592 : // That image layer creation would be useless for a hole range covered by L1s that don't contain any WAL records.
1593 : //
1594 : // The algorithm chooses holes as follows.
1595 : // - Slide a 2-window over the keys in key orde to get the hole range (=distance between two keys).
1596 : // - Filter: min threshold on range length
1597 : // - Rank: by coverage size (=number of image layers required to reconstruct each key in the range for which we have any data)
1598 : //
1599 : // For more details, intuition, and some ASCII art see https://github.com/neondatabase/neon/pull/3597#discussion_r1112704451
1600 : #[derive(PartialEq, Eq)]
1601 : struct Hole {
1602 : key_range: Range<Key>,
1603 : coverage_size: usize,
1604 : }
1605 56 : let holes: Vec<Hole> = {
1606 : use std::cmp::Ordering;
1607 : impl Ord for Hole {
1608 0 : fn cmp(&self, other: &Self) -> Ordering {
1609 0 : self.coverage_size.cmp(&other.coverage_size).reverse()
1610 0 : }
1611 : }
1612 : impl PartialOrd for Hole {
1613 0 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
1614 0 : Some(self.cmp(other))
1615 0 : }
1616 : }
1617 56 : let max_holes = deltas_to_compact.len();
1618 56 : let last_record_lsn = self.get_last_record_lsn();
1619 56 : let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
1620 56 : let min_hole_coverage_size = 3; // TODO: something more flexible?
1621 56 : // min-heap (reserve space for one more element added before eviction)
1622 56 : let mut heap: BinaryHeap<Hole> = BinaryHeap::with_capacity(max_holes + 1);
1623 56 : let mut prev: Option<Key> = None;
1624 :
1625 4128076 : for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
1626 4128076 : if let Some(prev_key) = prev {
1627 : // just first fast filter, do not create hole entries for metadata keys. The last hole in the
1628 : // compaction is the gap between data key and metadata keys.
1629 4128020 : if next_key.to_i128() - prev_key.to_i128() >= min_hole_range
1630 0 : && !Key::is_metadata_key(&prev_key)
1631 : {
1632 0 : let key_range = prev_key..next_key;
1633 0 : // Measuring hole by just subtraction of i128 representation of key range boundaries
1634 0 : // has not so much sense, because largest holes will corresponds field1/field2 changes.
1635 0 : // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
1636 0 : // That is why it is better to measure size of hole as number of covering image layers.
1637 0 : let coverage_size =
1638 0 : layers.image_coverage(&key_range, last_record_lsn).len();
1639 0 : if coverage_size >= min_hole_coverage_size {
1640 0 : heap.push(Hole {
1641 0 : key_range,
1642 0 : coverage_size,
1643 0 : });
1644 0 : if heap.len() > max_holes {
1645 0 : heap.pop(); // remove smallest hole
1646 0 : }
1647 0 : }
1648 4128020 : }
1649 56 : }
1650 4128076 : prev = Some(next_key.next());
1651 : }
1652 56 : let mut holes = heap.into_vec();
1653 56 : holes.sort_unstable_by_key(|hole| hole.key_range.start);
1654 56 : holes
1655 56 : };
1656 56 : stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
1657 56 : drop_rlock(guard);
1658 56 :
1659 56 : if self.cancel.is_cancelled() {
1660 0 : return Err(CompactionError::ShuttingDown);
1661 56 : }
1662 56 :
1663 56 : stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
1664 :
1665 : // This iterator walks through all key-value pairs from all the layers
1666 : // we're compacting, in key, LSN order.
1667 : // If there's both a Value::Image and Value::WalRecord for the same (key,lsn),
1668 : // then the Value::Image is ordered before Value::WalRecord.
1669 56 : let mut all_values_iter = {
1670 56 : let mut deltas = Vec::with_capacity(deltas_to_compact.len());
1671 804 : for l in deltas_to_compact.iter() {
1672 804 : let l = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
1673 804 : deltas.push(l);
1674 : }
1675 56 : MergeIterator::create(&deltas, &[], ctx)
1676 56 : };
1677 56 :
1678 56 : // This iterator walks through all keys and is needed to calculate size used by each key
1679 56 : let mut all_keys_iter = all_keys
1680 56 : .iter()
1681 4128076 : .map(|DeltaEntry { key, lsn, size, .. }| (*key, *lsn, *size))
1682 4128020 : .coalesce(|mut prev, cur| {
1683 4128020 : // Coalesce keys that belong to the same key pair.
1684 4128020 : // This ensures that compaction doesn't put them
1685 4128020 : // into different layer files.
1686 4128020 : // Still limit this by the target file size,
1687 4128020 : // so that we keep the size of the files in
1688 4128020 : // check.
1689 4128020 : if prev.0 == cur.0 && prev.2 < target_file_size {
1690 80076 : prev.2 += cur.2;
1691 80076 : Ok(prev)
1692 : } else {
1693 4047944 : Err((prev, cur))
1694 : }
1695 4128020 : });
1696 56 :
1697 56 : // Merge the contents of all the input delta layers into a new set
1698 56 : // of delta layers, based on the current partitioning.
1699 56 : //
1700 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.
1701 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
1702 56 : // would be too large. In that case, we also split on the LSN dimension.
1703 56 : //
1704 56 : // LSN
1705 56 : // ^
1706 56 : // |
1707 56 : // | +-----------+ +--+--+--+--+
1708 56 : // | | | | | | | |
1709 56 : // | +-----------+ | | | | |
1710 56 : // | | | | | | | |
1711 56 : // | +-----------+ ==> | | | | |
1712 56 : // | | | | | | | |
1713 56 : // | +-----------+ | | | | |
1714 56 : // | | | | | | | |
1715 56 : // | +-----------+ +--+--+--+--+
1716 56 : // |
1717 56 : // +--------------> key
1718 56 : //
1719 56 : //
1720 56 : // If one key (X) has a lot of page versions:
1721 56 : //
1722 56 : // LSN
1723 56 : // ^
1724 56 : // | (X)
1725 56 : // | +-----------+ +--+--+--+--+
1726 56 : // | | | | | | | |
1727 56 : // | +-----------+ | | +--+ |
1728 56 : // | | | | | | | |
1729 56 : // | +-----------+ ==> | | | | |
1730 56 : // | | | | | +--+ |
1731 56 : // | +-----------+ | | | | |
1732 56 : // | | | | | | | |
1733 56 : // | +-----------+ +--+--+--+--+
1734 56 : // |
1735 56 : // +--------------> key
1736 56 : // TODO: this actually divides the layers into fixed-size chunks, not
1737 56 : // based on the partitioning.
1738 56 : //
1739 56 : // TODO: we should also opportunistically materialize and
1740 56 : // garbage collect what we can.
1741 56 : let mut new_layers = Vec::new();
1742 56 : let mut prev_key: Option<Key> = None;
1743 56 : let mut writer: Option<DeltaLayerWriter> = None;
1744 56 : let mut key_values_total_size = 0u64;
1745 56 : let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
1746 56 : let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
1747 56 : let mut next_hole = 0; // index of next hole in holes vector
1748 56 :
1749 56 : let mut keys = 0;
1750 :
1751 4128132 : while let Some((key, lsn, value)) = all_values_iter
1752 4128132 : .next()
1753 4128132 : .await
1754 4128132 : .map_err(CompactionError::Other)?
1755 : {
1756 4128076 : keys += 1;
1757 4128076 :
1758 4128076 : if keys % 32_768 == 0 && self.cancel.is_cancelled() {
1759 : // avoid hitting the cancellation token on every key. in benches, we end up
1760 : // shuffling an order of million keys per layer, this means we'll check it
1761 : // around tens of times per layer.
1762 0 : return Err(CompactionError::ShuttingDown);
1763 4128076 : }
1764 4128076 :
1765 4128076 : let same_key = prev_key == Some(key);
1766 4128076 : // We need to check key boundaries once we reach next key or end of layer with the same key
1767 4128076 : if !same_key || lsn == dup_end_lsn {
1768 4048000 : let mut next_key_size = 0u64;
1769 4048000 : let is_dup_layer = dup_end_lsn.is_valid();
1770 4048000 : dup_start_lsn = Lsn::INVALID;
1771 4048000 : if !same_key {
1772 4048000 : dup_end_lsn = Lsn::INVALID;
1773 4048000 : }
1774 : // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
1775 4048000 : for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
1776 4048000 : next_key_size = next_size;
1777 4048000 : if key != next_key {
1778 4047944 : if dup_end_lsn.is_valid() {
1779 0 : // We are writting segment with duplicates:
1780 0 : // place all remaining values of this key in separate segment
1781 0 : dup_start_lsn = dup_end_lsn; // new segments starts where old stops
1782 0 : dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
1783 4047944 : }
1784 4047944 : break;
1785 56 : }
1786 56 : key_values_total_size += next_size;
1787 56 : // Check if it is time to split segment: if total keys size is larger than target file size.
1788 56 : // We need to avoid generation of empty segments if next_size > target_file_size.
1789 56 : if key_values_total_size > target_file_size && lsn != next_lsn {
1790 : // Split key between multiple layers: such layer can contain only single key
1791 0 : dup_start_lsn = if dup_end_lsn.is_valid() {
1792 0 : dup_end_lsn // new segment with duplicates starts where old one stops
1793 : } else {
1794 0 : lsn // start with the first LSN for this key
1795 : };
1796 0 : dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
1797 0 : break;
1798 56 : }
1799 : }
1800 : // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
1801 4048000 : if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
1802 0 : dup_start_lsn = dup_end_lsn;
1803 0 : dup_end_lsn = lsn_range.end;
1804 4048000 : }
1805 4048000 : if writer.is_some() {
1806 4047944 : let written_size = writer.as_mut().unwrap().size();
1807 4047944 : let contains_hole =
1808 4047944 : next_hole < holes.len() && key >= holes[next_hole].key_range.end;
1809 : // check if key cause layer overflow or contains hole...
1810 4047944 : if is_dup_layer
1811 4047944 : || dup_end_lsn.is_valid()
1812 4047944 : || written_size + key_values_total_size > target_file_size
1813 4047384 : || contains_hole
1814 : {
1815 : // ... if so, flush previous layer and prepare to write new one
1816 560 : let (desc, path) = writer
1817 560 : .take()
1818 560 : .unwrap()
1819 560 : .finish(prev_key.unwrap().next(), ctx)
1820 560 : .await
1821 560 : .map_err(CompactionError::Other)?;
1822 560 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1823 560 : .map_err(CompactionError::Other)?;
1824 :
1825 560 : new_layers.push(new_delta);
1826 560 : writer = None;
1827 560 :
1828 560 : if contains_hole {
1829 0 : // skip hole
1830 0 : next_hole += 1;
1831 560 : }
1832 4047384 : }
1833 56 : }
1834 : // Remember size of key value because at next iteration we will access next item
1835 4048000 : key_values_total_size = next_key_size;
1836 80076 : }
1837 4128076 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
1838 0 : Err(CompactionError::Other(anyhow::anyhow!(
1839 0 : "failpoint delta-layer-writer-fail-before-finish"
1840 0 : )))
1841 4128076 : });
1842 :
1843 4128076 : if !self.shard_identity.is_key_disposable(&key) {
1844 4128076 : if writer.is_none() {
1845 616 : if self.cancel.is_cancelled() {
1846 : // to be somewhat responsive to cancellation, check for each new layer
1847 0 : return Err(CompactionError::ShuttingDown);
1848 616 : }
1849 : // Create writer if not initiaized yet
1850 616 : writer = Some(
1851 : DeltaLayerWriter::new(
1852 616 : self.conf,
1853 616 : self.timeline_id,
1854 616 : self.tenant_shard_id,
1855 616 : key,
1856 616 : if dup_end_lsn.is_valid() {
1857 : // this is a layer containing slice of values of the same key
1858 0 : debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
1859 0 : dup_start_lsn..dup_end_lsn
1860 : } else {
1861 616 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
1862 616 : lsn_range.clone()
1863 : },
1864 616 : ctx,
1865 616 : )
1866 616 : .await
1867 616 : .map_err(CompactionError::Other)?,
1868 : );
1869 :
1870 616 : keys = 0;
1871 4127460 : }
1872 :
1873 4128076 : writer
1874 4128076 : .as_mut()
1875 4128076 : .unwrap()
1876 4128076 : .put_value(key, lsn, value, ctx)
1877 4128076 : .await
1878 4128076 : .map_err(CompactionError::Other)?;
1879 : } else {
1880 0 : let owner = self.shard_identity.get_shard_number(&key);
1881 0 :
1882 0 : // This happens after a shard split, when we're compacting an L0 created by our parent shard
1883 0 : debug!("dropping key {key} during compaction (it belongs on shard {owner})");
1884 : }
1885 :
1886 4128076 : if !new_layers.is_empty() {
1887 39572 : fail_point!("after-timeline-compacted-first-L1");
1888 4088504 : }
1889 :
1890 4128076 : prev_key = Some(key);
1891 : }
1892 56 : if let Some(writer) = writer {
1893 56 : let (desc, path) = writer
1894 56 : .finish(prev_key.unwrap().next(), ctx)
1895 56 : .await
1896 56 : .map_err(CompactionError::Other)?;
1897 56 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1898 56 : .map_err(CompactionError::Other)?;
1899 56 : new_layers.push(new_delta);
1900 0 : }
1901 :
1902 : // Sync layers
1903 56 : if !new_layers.is_empty() {
1904 : // Print a warning if the created layer is larger than double the target size
1905 : // Add two pages for potential overhead. This should in theory be already
1906 : // accounted for in the target calculation, but for very small targets,
1907 : // we still might easily hit the limit otherwise.
1908 56 : let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
1909 616 : for layer in new_layers.iter() {
1910 616 : if layer.layer_desc().file_size > warn_limit {
1911 0 : warn!(
1912 : %layer,
1913 0 : "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
1914 : );
1915 616 : }
1916 : }
1917 :
1918 : // The writer.finish() above already did the fsync of the inodes.
1919 : // We just need to fsync the directory in which these inodes are linked,
1920 : // which we know to be the timeline directory.
1921 : //
1922 : // We use fatal_err() below because the after writer.finish() returns with success,
1923 : // the in-memory state of the filesystem already has the layer file in its final place,
1924 : // and subsequent pageserver code could think it's durable while it really isn't.
1925 56 : let timeline_dir = VirtualFile::open(
1926 56 : &self
1927 56 : .conf
1928 56 : .timeline_path(&self.tenant_shard_id, &self.timeline_id),
1929 56 : ctx,
1930 56 : )
1931 56 : .await
1932 56 : .fatal_err("VirtualFile::open for timeline dir fsync");
1933 56 : timeline_dir
1934 56 : .sync_all()
1935 56 : .await
1936 56 : .fatal_err("VirtualFile::sync_all timeline dir");
1937 0 : }
1938 :
1939 56 : stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
1940 56 : stats.new_deltas_count = Some(new_layers.len());
1941 616 : stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
1942 56 :
1943 56 : match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
1944 56 : .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
1945 : {
1946 56 : Ok(stats_json) => {
1947 56 : info!(
1948 0 : stats_json = stats_json.as_str(),
1949 0 : "compact_level0_phase1 stats available"
1950 : )
1951 : }
1952 0 : Err(e) => {
1953 0 : warn!("compact_level0_phase1 stats failed to serialize: {:#}", e);
1954 : }
1955 : }
1956 :
1957 : // Without this, rustc complains about deltas_to_compact still
1958 : // being borrowed when we `.into_iter()` below.
1959 56 : drop(all_values_iter);
1960 56 :
1961 56 : Ok(CompactLevel0Phase1Result {
1962 56 : new_layers,
1963 56 : deltas_to_compact: deltas_to_compact
1964 56 : .into_iter()
1965 804 : .map(|x| x.drop_eviction_guard())
1966 56 : .collect::<Vec<_>>(),
1967 56 : outcome: if fully_compacted {
1968 56 : CompactionOutcome::Done
1969 : } else {
1970 0 : CompactionOutcome::Pending
1971 : },
1972 : })
1973 728 : }
1974 : }
1975 :
1976 : #[derive(Default)]
1977 : struct CompactLevel0Phase1Result {
1978 : new_layers: Vec<ResidentLayer>,
1979 : deltas_to_compact: Vec<Layer>,
1980 : // Whether we have included all L0 layers, or selected only part of them due to the
1981 : // L0 compaction size limit.
1982 : outcome: CompactionOutcome,
1983 : }
1984 :
1985 : #[derive(Default)]
1986 : struct CompactLevel0Phase1StatsBuilder {
1987 : version: Option<u64>,
1988 : tenant_id: Option<TenantShardId>,
1989 : timeline_id: Option<TimelineId>,
1990 : read_lock_acquisition_micros: DurationRecorder,
1991 : read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
1992 : read_lock_held_key_sort_micros: DurationRecorder,
1993 : read_lock_held_prerequisites_micros: DurationRecorder,
1994 : read_lock_held_compute_holes_micros: DurationRecorder,
1995 : read_lock_drop_micros: DurationRecorder,
1996 : write_layer_files_micros: DurationRecorder,
1997 : level0_deltas_count: Option<usize>,
1998 : new_deltas_count: Option<usize>,
1999 : new_deltas_size: Option<u64>,
2000 : }
2001 :
2002 : #[derive(serde::Serialize)]
2003 : struct CompactLevel0Phase1Stats {
2004 : version: u64,
2005 : tenant_id: TenantShardId,
2006 : timeline_id: TimelineId,
2007 : read_lock_acquisition_micros: RecordedDuration,
2008 : read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
2009 : read_lock_held_key_sort_micros: RecordedDuration,
2010 : read_lock_held_prerequisites_micros: RecordedDuration,
2011 : read_lock_held_compute_holes_micros: RecordedDuration,
2012 : read_lock_drop_micros: RecordedDuration,
2013 : write_layer_files_micros: RecordedDuration,
2014 : level0_deltas_count: usize,
2015 : new_deltas_count: usize,
2016 : new_deltas_size: u64,
2017 : }
2018 :
2019 : impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
2020 : type Error = anyhow::Error;
2021 :
2022 56 : fn try_from(value: CompactLevel0Phase1StatsBuilder) -> Result<Self, Self::Error> {
2023 56 : Ok(Self {
2024 56 : version: value.version.ok_or_else(|| anyhow!("version not set"))?,
2025 56 : tenant_id: value
2026 56 : .tenant_id
2027 56 : .ok_or_else(|| anyhow!("tenant_id not set"))?,
2028 56 : timeline_id: value
2029 56 : .timeline_id
2030 56 : .ok_or_else(|| anyhow!("timeline_id not set"))?,
2031 56 : read_lock_acquisition_micros: value
2032 56 : .read_lock_acquisition_micros
2033 56 : .into_recorded()
2034 56 : .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
2035 56 : read_lock_held_spawn_blocking_startup_micros: value
2036 56 : .read_lock_held_spawn_blocking_startup_micros
2037 56 : .into_recorded()
2038 56 : .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
2039 56 : read_lock_held_key_sort_micros: value
2040 56 : .read_lock_held_key_sort_micros
2041 56 : .into_recorded()
2042 56 : .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
2043 56 : read_lock_held_prerequisites_micros: value
2044 56 : .read_lock_held_prerequisites_micros
2045 56 : .into_recorded()
2046 56 : .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
2047 56 : read_lock_held_compute_holes_micros: value
2048 56 : .read_lock_held_compute_holes_micros
2049 56 : .into_recorded()
2050 56 : .ok_or_else(|| anyhow!("read_lock_held_compute_holes_micros not set"))?,
2051 56 : read_lock_drop_micros: value
2052 56 : .read_lock_drop_micros
2053 56 : .into_recorded()
2054 56 : .ok_or_else(|| anyhow!("read_lock_drop_micros not set"))?,
2055 56 : write_layer_files_micros: value
2056 56 : .write_layer_files_micros
2057 56 : .into_recorded()
2058 56 : .ok_or_else(|| anyhow!("write_layer_files_micros not set"))?,
2059 56 : level0_deltas_count: value
2060 56 : .level0_deltas_count
2061 56 : .ok_or_else(|| anyhow!("level0_deltas_count not set"))?,
2062 56 : new_deltas_count: value
2063 56 : .new_deltas_count
2064 56 : .ok_or_else(|| anyhow!("new_deltas_count not set"))?,
2065 56 : new_deltas_size: value
2066 56 : .new_deltas_size
2067 56 : .ok_or_else(|| anyhow!("new_deltas_size not set"))?,
2068 : })
2069 56 : }
2070 : }
2071 :
2072 : impl Timeline {
2073 : /// Entry point for new tiered compaction algorithm.
2074 : ///
2075 : /// All the real work is in the implementation in the pageserver_compaction
2076 : /// crate. The code here would apply to any algorithm implemented by the
2077 : /// same interface, but tiered is the only one at the moment.
2078 : ///
2079 : /// TODO: cancellation
2080 0 : pub(crate) async fn compact_tiered(
2081 0 : self: &Arc<Self>,
2082 0 : _cancel: &CancellationToken,
2083 0 : ctx: &RequestContext,
2084 0 : ) -> Result<(), CompactionError> {
2085 0 : let fanout = self.get_compaction_threshold() as u64;
2086 0 : let target_file_size = self.get_checkpoint_distance();
2087 :
2088 : // Find the top of the historical layers
2089 0 : let end_lsn = {
2090 0 : let guard = self.layers.read().await;
2091 0 : let layers = guard.layer_map()?;
2092 :
2093 0 : let l0_deltas = layers.level0_deltas();
2094 0 :
2095 0 : // As an optimization, if we find that there are too few L0 layers,
2096 0 : // bail out early. We know that the compaction algorithm would do
2097 0 : // nothing in that case.
2098 0 : if l0_deltas.len() < fanout as usize {
2099 : // doesn't need compacting
2100 0 : return Ok(());
2101 0 : }
2102 0 : l0_deltas.iter().map(|l| l.lsn_range.end).max().unwrap()
2103 0 : };
2104 0 :
2105 0 : // Is the timeline being deleted?
2106 0 : if self.is_stopping() {
2107 0 : trace!("Dropping out of compaction on timeline shutdown");
2108 0 : return Err(CompactionError::ShuttingDown);
2109 0 : }
2110 :
2111 0 : let (dense_ks, _sparse_ks) = self.collect_keyspace(end_lsn, ctx).await?;
2112 : // TODO(chi): ignore sparse_keyspace for now, compact it in the future.
2113 0 : let mut adaptor = TimelineAdaptor::new(self, (end_lsn, dense_ks));
2114 0 :
2115 0 : pageserver_compaction::compact_tiered::compact_tiered(
2116 0 : &mut adaptor,
2117 0 : end_lsn,
2118 0 : target_file_size,
2119 0 : fanout,
2120 0 : ctx,
2121 0 : )
2122 0 : .await
2123 : // TODO: compact_tiered needs to return CompactionError
2124 0 : .map_err(CompactionError::Other)?;
2125 :
2126 0 : adaptor.flush_updates().await?;
2127 0 : Ok(())
2128 0 : }
2129 :
2130 : /// Take a list of images and deltas, produce images and deltas according to GC horizon and retain_lsns.
2131 : ///
2132 : /// It takes a key, the values of the key within the compaction process, a GC horizon, and all retain_lsns below the horizon.
2133 : /// For now, it requires the `accumulated_values` contains the full history of the key (i.e., the key with the lowest LSN is
2134 : /// an image or a WAL not requiring a base image). This restriction will be removed once we implement gc-compaction on branch.
2135 : ///
2136 : /// The function returns the deltas and the base image that need to be placed at each of the retain LSN. For example, we have:
2137 : ///
2138 : /// A@0x10, +B@0x20, +C@0x30, +D@0x40, +E@0x50, +F@0x60
2139 : /// horizon = 0x50, retain_lsn = 0x20, 0x40, delta_threshold=3
2140 : ///
2141 : /// The function will produce:
2142 : ///
2143 : /// ```plain
2144 : /// 0x20(retain_lsn) -> img=AB@0x20 always produce a single image below the lowest retain LSN
2145 : /// 0x40(retain_lsn) -> deltas=[+C@0x30, +D@0x40] two deltas since the last base image, keeping the deltas
2146 : /// 0x50(horizon) -> deltas=[ABCDE@0x50] three deltas since the last base image, generate an image but put it in the delta
2147 : /// above_horizon -> deltas=[+F@0x60] full history above the horizon
2148 : /// ```
2149 : ///
2150 : /// Note that `accumulated_values` must be sorted by LSN and should belong to a single key.
2151 1296 : pub(crate) async fn generate_key_retention(
2152 1296 : self: &Arc<Timeline>,
2153 1296 : key: Key,
2154 1296 : full_history: &[(Key, Lsn, Value)],
2155 1296 : horizon: Lsn,
2156 1296 : retain_lsn_below_horizon: &[Lsn],
2157 1296 : delta_threshold_cnt: usize,
2158 1296 : base_img_from_ancestor: Option<(Key, Lsn, Bytes)>,
2159 1296 : ) -> anyhow::Result<KeyHistoryRetention> {
2160 : // Pre-checks for the invariants
2161 :
2162 1296 : let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
2163 :
2164 1296 : if debug_mode {
2165 3144 : for (log_key, _, _) in full_history {
2166 1848 : assert_eq!(log_key, &key, "mismatched key");
2167 : }
2168 1296 : for i in 1..full_history.len() {
2169 552 : assert!(full_history[i - 1].1 <= full_history[i].1, "unordered LSN");
2170 552 : if full_history[i - 1].1 == full_history[i].1 {
2171 0 : assert!(
2172 0 : matches!(full_history[i - 1].2, Value::Image(_)),
2173 0 : "unordered delta/image, or duplicated delta"
2174 : );
2175 552 : }
2176 : }
2177 : // There was an assertion for no base image that checks if the first
2178 : // record in the history is `will_init` before, but it was removed.
2179 : // This is explained in the test cases for generate_key_retention.
2180 : // Search "incomplete history" for more information.
2181 2856 : for lsn in retain_lsn_below_horizon {
2182 1560 : assert!(lsn < &horizon, "retain lsn must be below horizon")
2183 : }
2184 1296 : for i in 1..retain_lsn_below_horizon.len() {
2185 712 : assert!(
2186 712 : retain_lsn_below_horizon[i - 1] <= retain_lsn_below_horizon[i],
2187 0 : "unordered LSN"
2188 : );
2189 : }
2190 0 : }
2191 1296 : let has_ancestor = base_img_from_ancestor.is_some();
2192 : // Step 1: split history into len(retain_lsn_below_horizon) + 2 buckets, where the last bucket is for all deltas above the horizon,
2193 : // and the second-to-last bucket is for the horizon. Each bucket contains lsn_last_bucket < deltas <= lsn_this_bucket.
2194 1296 : let (mut split_history, lsn_split_points) = {
2195 1296 : let mut split_history = Vec::new();
2196 1296 : split_history.resize_with(retain_lsn_below_horizon.len() + 2, Vec::new);
2197 1296 : let mut lsn_split_points = Vec::with_capacity(retain_lsn_below_horizon.len() + 1);
2198 2856 : for lsn in retain_lsn_below_horizon {
2199 1560 : lsn_split_points.push(*lsn);
2200 1560 : }
2201 1296 : lsn_split_points.push(horizon);
2202 1296 : let mut current_idx = 0;
2203 3144 : for item @ (_, lsn, _) in full_history {
2204 2336 : while current_idx < lsn_split_points.len() && *lsn > lsn_split_points[current_idx] {
2205 488 : current_idx += 1;
2206 488 : }
2207 1848 : split_history[current_idx].push(item);
2208 : }
2209 1296 : (split_history, lsn_split_points)
2210 : };
2211 : // Step 2: filter out duplicated records due to the k-merge of image/delta layers
2212 5448 : for split_for_lsn in &mut split_history {
2213 4152 : let mut prev_lsn = None;
2214 4152 : let mut new_split_for_lsn = Vec::with_capacity(split_for_lsn.len());
2215 4152 : for record @ (_, lsn, _) in std::mem::take(split_for_lsn) {
2216 1848 : if let Some(prev_lsn) = &prev_lsn {
2217 248 : if *prev_lsn == lsn {
2218 : // The case that we have an LSN with both data from the delta layer and the image layer. As
2219 : // `ValueWrapper` ensures that an image is ordered before a delta at the same LSN, we simply
2220 : // drop this delta and keep the image.
2221 : //
2222 : // For example, we have delta layer key1@0x10, key1@0x20, and image layer key1@0x10, we will
2223 : // keep the image for key1@0x10 and the delta for key1@0x20. key1@0x10 delta will be simply
2224 : // dropped.
2225 : //
2226 : // TODO: in case we have both delta + images for a given LSN and it does not exceed the delta
2227 : // threshold, we could have kept delta instead to save space. This is an optimization for the future.
2228 0 : continue;
2229 248 : }
2230 1600 : }
2231 1848 : prev_lsn = Some(lsn);
2232 1848 : new_split_for_lsn.push(record);
2233 : }
2234 4152 : *split_for_lsn = new_split_for_lsn;
2235 : }
2236 : // Step 3: generate images when necessary
2237 1296 : let mut retention = Vec::with_capacity(split_history.len());
2238 1296 : let mut records_since_last_image = 0;
2239 1296 : let batch_cnt = split_history.len();
2240 1296 : assert!(
2241 1296 : batch_cnt >= 2,
2242 0 : "should have at least below + above horizon batches"
2243 : );
2244 1296 : let mut replay_history: Vec<(Key, Lsn, Value)> = Vec::new();
2245 1296 : if let Some((key, lsn, img)) = base_img_from_ancestor {
2246 84 : replay_history.push((key, lsn, Value::Image(img)));
2247 1212 : }
2248 :
2249 : /// Generate debug information for the replay history
2250 0 : fn generate_history_trace(replay_history: &[(Key, Lsn, Value)]) -> String {
2251 : use std::fmt::Write;
2252 0 : let mut output = String::new();
2253 0 : if let Some((key, _, _)) = replay_history.first() {
2254 0 : write!(output, "key={} ", key).unwrap();
2255 0 : let mut cnt = 0;
2256 0 : for (_, lsn, val) in replay_history {
2257 0 : if val.is_image() {
2258 0 : write!(output, "i@{} ", lsn).unwrap();
2259 0 : } else if val.will_init() {
2260 0 : write!(output, "di@{} ", lsn).unwrap();
2261 0 : } else {
2262 0 : write!(output, "d@{} ", lsn).unwrap();
2263 0 : }
2264 0 : cnt += 1;
2265 0 : if cnt >= 128 {
2266 0 : write!(output, "... and more").unwrap();
2267 0 : break;
2268 0 : }
2269 : }
2270 0 : } else {
2271 0 : write!(output, "<no history>").unwrap();
2272 0 : }
2273 0 : output
2274 0 : }
2275 :
2276 0 : fn generate_debug_trace(
2277 0 : replay_history: Option<&[(Key, Lsn, Value)]>,
2278 0 : full_history: &[(Key, Lsn, Value)],
2279 0 : lsns: &[Lsn],
2280 0 : horizon: Lsn,
2281 0 : ) -> String {
2282 : use std::fmt::Write;
2283 0 : let mut output = String::new();
2284 0 : if let Some(replay_history) = replay_history {
2285 0 : writeln!(
2286 0 : output,
2287 0 : "replay_history: {}",
2288 0 : generate_history_trace(replay_history)
2289 0 : )
2290 0 : .unwrap();
2291 0 : } else {
2292 0 : writeln!(output, "replay_history: <disabled>",).unwrap();
2293 0 : }
2294 0 : writeln!(
2295 0 : output,
2296 0 : "full_history: {}",
2297 0 : generate_history_trace(full_history)
2298 0 : )
2299 0 : .unwrap();
2300 0 : writeln!(
2301 0 : output,
2302 0 : "when processing: [{}] horizon={}",
2303 0 : lsns.iter().map(|l| format!("{l}")).join(","),
2304 0 : horizon
2305 0 : )
2306 0 : .unwrap();
2307 0 : output
2308 0 : }
2309 :
2310 1296 : let mut key_exists = false;
2311 4148 : for (i, split_for_lsn) in split_history.into_iter().enumerate() {
2312 : // TODO: there could be image keys inside the splits, and we can compute records_since_last_image accordingly.
2313 4148 : records_since_last_image += split_for_lsn.len();
2314 : // Whether to produce an image into the final layer files
2315 4148 : let produce_image = if i == 0 && !has_ancestor {
2316 : // We always generate images for the first batch (below horizon / lowest retain_lsn)
2317 1212 : true
2318 2936 : } else if i == batch_cnt - 1 {
2319 : // Do not generate images for the last batch (above horizon)
2320 1292 : false
2321 1644 : } else if records_since_last_image == 0 {
2322 1288 : false
2323 356 : } else if records_since_last_image >= delta_threshold_cnt {
2324 : // Generate images when there are too many records
2325 12 : true
2326 : } else {
2327 344 : false
2328 : };
2329 4148 : replay_history.extend(split_for_lsn.iter().map(|x| (*x).clone()));
2330 : // Only retain the items after the last image record
2331 5108 : for idx in (0..replay_history.len()).rev() {
2332 5108 : if replay_history[idx].2.will_init() {
2333 4148 : replay_history = replay_history[idx..].to_vec();
2334 4148 : break;
2335 960 : }
2336 : }
2337 4148 : if replay_history.is_empty() && !key_exists {
2338 : // The key does not exist at earlier LSN, we can skip this iteration.
2339 0 : retention.push(Vec::new());
2340 0 : continue;
2341 4148 : } else {
2342 4148 : key_exists = true;
2343 4148 : }
2344 4148 : let Some((_, _, val)) = replay_history.first() else {
2345 0 : unreachable!("replay history should not be empty once it exists")
2346 : };
2347 4148 : if !val.will_init() {
2348 0 : return Err(anyhow::anyhow!("invalid history, no base image")).with_context(|| {
2349 0 : generate_debug_trace(
2350 0 : Some(&replay_history),
2351 0 : full_history,
2352 0 : retain_lsn_below_horizon,
2353 0 : horizon,
2354 0 : )
2355 0 : });
2356 4148 : }
2357 : // Whether to reconstruct the image. In debug mode, we will generate an image
2358 : // at every retain_lsn to ensure data is not corrupted, but we won't put the
2359 : // image into the final layer.
2360 4148 : let generate_image = produce_image || debug_mode;
2361 4148 : if produce_image {
2362 1224 : records_since_last_image = 0;
2363 2924 : }
2364 4148 : let img_and_lsn = if generate_image {
2365 4148 : let replay_history_for_debug = if debug_mode {
2366 4148 : Some(replay_history.clone())
2367 : } else {
2368 0 : None
2369 : };
2370 4148 : let replay_history_for_debug_ref = replay_history_for_debug.as_deref();
2371 4148 : let history = if produce_image {
2372 1224 : std::mem::take(&mut replay_history)
2373 : } else {
2374 2924 : replay_history.clone()
2375 : };
2376 4148 : let mut img = None;
2377 4148 : let mut records = Vec::with_capacity(history.len());
2378 4148 : if let (_, lsn, Value::Image(val)) = history.first().as_ref().unwrap() {
2379 4104 : img = Some((*lsn, val.clone()));
2380 4104 : for (_, lsn, val) in history.into_iter().skip(1) {
2381 932 : let Value::WalRecord(rec) = val else {
2382 0 : return Err(anyhow::anyhow!(
2383 0 : "invalid record, first record is image, expect walrecords"
2384 0 : ))
2385 0 : .with_context(|| {
2386 0 : generate_debug_trace(
2387 0 : replay_history_for_debug_ref,
2388 0 : full_history,
2389 0 : retain_lsn_below_horizon,
2390 0 : horizon,
2391 0 : )
2392 0 : });
2393 : };
2394 932 : records.push((lsn, rec));
2395 : }
2396 : } else {
2397 72 : for (_, lsn, val) in history.into_iter() {
2398 72 : let Value::WalRecord(rec) = val else {
2399 0 : return Err(anyhow::anyhow!("invalid record, first record is walrecord, expect rest are walrecord"))
2400 0 : .with_context(|| generate_debug_trace(
2401 0 : replay_history_for_debug_ref,
2402 0 : full_history,
2403 0 : retain_lsn_below_horizon,
2404 0 : horizon,
2405 0 : ));
2406 : };
2407 72 : records.push((lsn, rec));
2408 : }
2409 : }
2410 4148 : records.reverse();
2411 4148 : let state = ValueReconstructState { img, records };
2412 : // last batch does not generate image so i is always in range, unless we force generate
2413 : // an image during testing
2414 4148 : let request_lsn = if i >= lsn_split_points.len() {
2415 1292 : Lsn::MAX
2416 : } else {
2417 2856 : lsn_split_points[i]
2418 : };
2419 4148 : let img = self
2420 4148 : .reconstruct_value(key, request_lsn, state, RedoAttemptType::GcCompaction)
2421 4148 : .await?;
2422 4144 : Some((request_lsn, img))
2423 : } else {
2424 0 : None
2425 : };
2426 4144 : if produce_image {
2427 1220 : let (request_lsn, img) = img_and_lsn.unwrap();
2428 1220 : replay_history.push((key, request_lsn, Value::Image(img.clone())));
2429 1220 : retention.push(vec![(request_lsn, Value::Image(img))]);
2430 2924 : } else {
2431 2924 : let deltas = split_for_lsn
2432 2924 : .iter()
2433 2924 : .map(|(_, lsn, value)| (*lsn, value.clone()))
2434 2924 : .collect_vec();
2435 2924 : retention.push(deltas);
2436 2924 : }
2437 : }
2438 1292 : let mut result = Vec::with_capacity(retention.len());
2439 1292 : assert_eq!(retention.len(), lsn_split_points.len() + 1);
2440 4144 : for (idx, logs) in retention.into_iter().enumerate() {
2441 4144 : if idx == lsn_split_points.len() {
2442 1292 : return Ok(KeyHistoryRetention {
2443 1292 : below_horizon: result,
2444 1292 : above_horizon: KeyLogAtLsn(logs),
2445 1292 : });
2446 2852 : } else {
2447 2852 : result.push((lsn_split_points[idx], KeyLogAtLsn(logs)));
2448 2852 : }
2449 : }
2450 0 : unreachable!("key retention is empty")
2451 1296 : }
2452 :
2453 : /// Check how much space is left on the disk
2454 108 : async fn check_available_space(self: &Arc<Self>) -> anyhow::Result<u64> {
2455 108 : let tenants_dir = self.conf.tenants_path();
2456 :
2457 108 : let stat = Statvfs::get(&tenants_dir, None)
2458 108 : .context("statvfs failed, presumably directory got unlinked")?;
2459 :
2460 108 : let (avail_bytes, _) = stat.get_avail_total_bytes();
2461 108 :
2462 108 : Ok(avail_bytes)
2463 108 : }
2464 :
2465 : /// Check if the compaction can proceed safely without running out of space. We assume the size
2466 : /// upper bound of the produced files of a compaction job is the same as all layers involved in
2467 : /// the compaction. Therefore, we need `2 * layers_to_be_compacted_size` at least to do a
2468 : /// compaction.
2469 108 : async fn check_compaction_space(
2470 108 : self: &Arc<Self>,
2471 108 : layer_selection: &[Layer],
2472 108 : ) -> Result<(), CompactionError> {
2473 108 : let available_space = self
2474 108 : .check_available_space()
2475 108 : .await
2476 108 : .map_err(CompactionError::Other)?;
2477 108 : let mut remote_layer_size = 0;
2478 108 : let mut all_layer_size = 0;
2479 424 : for layer in layer_selection {
2480 316 : let needs_download = layer
2481 316 : .needs_download()
2482 316 : .await
2483 316 : .context("failed to check if layer needs download")
2484 316 : .map_err(CompactionError::Other)?;
2485 316 : if needs_download.is_some() {
2486 0 : remote_layer_size += layer.layer_desc().file_size;
2487 316 : }
2488 316 : all_layer_size += layer.layer_desc().file_size;
2489 : }
2490 108 : let allocated_space = (available_space as f64 * 0.8) as u64; /* reserve 20% space for other tasks */
2491 108 : if all_layer_size /* space needed for newly-generated file */ + remote_layer_size /* space for downloading layers */ > allocated_space
2492 : {
2493 0 : return Err(CompactionError::Other(anyhow!(
2494 0 : "not enough space for compaction: available_space={}, allocated_space={}, all_layer_size={}, remote_layer_size={}, required_space={}",
2495 0 : available_space,
2496 0 : allocated_space,
2497 0 : all_layer_size,
2498 0 : remote_layer_size,
2499 0 : all_layer_size + remote_layer_size
2500 0 : )));
2501 108 : }
2502 108 : Ok(())
2503 108 : }
2504 :
2505 : /// Get a watermark for gc-compaction, that is the lowest LSN that we can use as the `gc_horizon` for
2506 : /// the compaction algorithm. It is min(space_cutoff, time_cutoff, latest_gc_cutoff, standby_horizon).
2507 : /// Leases and retain_lsns are considered in the gc-compaction job itself so we don't need to account for them
2508 : /// here.
2509 112 : pub(crate) fn get_gc_compaction_watermark(self: &Arc<Self>) -> Lsn {
2510 112 : let gc_cutoff_lsn = {
2511 112 : let gc_info = self.gc_info.read().unwrap();
2512 112 : gc_info.min_cutoff()
2513 112 : };
2514 112 :
2515 112 : // TODO: standby horizon should use leases so we don't really need to consider it here.
2516 112 : // let watermark = watermark.min(self.standby_horizon.load());
2517 112 :
2518 112 : // TODO: ensure the child branches will not use anything below the watermark, or consider
2519 112 : // them when computing the watermark.
2520 112 : gc_cutoff_lsn.min(*self.get_applied_gc_cutoff_lsn())
2521 112 : }
2522 :
2523 : /// 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.
2524 : /// The function returns a list of compaction jobs that can be executed separately. If the upper bound of the compact LSN
2525 : /// range is not specified, we will use the latest gc_cutoff as the upper bound, so that all jobs in the jobset acts
2526 : /// like a full compaction of the specified keyspace.
2527 0 : pub(crate) async fn gc_compaction_split_jobs(
2528 0 : self: &Arc<Self>,
2529 0 : job: GcCompactJob,
2530 0 : sub_compaction_max_job_size_mb: Option<u64>,
2531 0 : ) -> Result<Vec<GcCompactJob>, CompactionError> {
2532 0 : let compact_below_lsn = if job.compact_lsn_range.end != Lsn::MAX {
2533 0 : job.compact_lsn_range.end
2534 : } else {
2535 0 : self.get_gc_compaction_watermark()
2536 : };
2537 :
2538 0 : if compact_below_lsn == Lsn::INVALID {
2539 0 : tracing::warn!(
2540 0 : "no layers to compact with gc: gc_cutoff not generated yet, skipping gc bottom-most compaction"
2541 : );
2542 0 : return Ok(vec![]);
2543 0 : }
2544 :
2545 : // Split compaction job to about 4GB each
2546 : const GC_COMPACT_MAX_SIZE_MB: u64 = 4 * 1024;
2547 0 : let sub_compaction_max_job_size_mb =
2548 0 : sub_compaction_max_job_size_mb.unwrap_or(GC_COMPACT_MAX_SIZE_MB);
2549 0 :
2550 0 : let mut compact_jobs = Vec::<GcCompactJob>::new();
2551 0 : // For now, we simply use the key partitioning information; we should do a more fine-grained partitioning
2552 0 : // by estimating the amount of files read for a compaction job. We should also partition on LSN.
2553 0 : let ((dense_ks, sparse_ks), _) = self.partitioning.read().as_ref().clone();
2554 : // Truncate the key range to be within user specified compaction range.
2555 0 : fn truncate_to(
2556 0 : source_start: &Key,
2557 0 : source_end: &Key,
2558 0 : target_start: &Key,
2559 0 : target_end: &Key,
2560 0 : ) -> Option<(Key, Key)> {
2561 0 : let start = source_start.max(target_start);
2562 0 : let end = source_end.min(target_end);
2563 0 : if start < end {
2564 0 : Some((*start, *end))
2565 : } else {
2566 0 : None
2567 : }
2568 0 : }
2569 0 : let mut split_key_ranges = Vec::new();
2570 0 : let ranges = dense_ks
2571 0 : .parts
2572 0 : .iter()
2573 0 : .map(|partition| partition.ranges.iter())
2574 0 : .chain(sparse_ks.parts.iter().map(|x| x.0.ranges.iter()))
2575 0 : .flatten()
2576 0 : .cloned()
2577 0 : .collect_vec();
2578 0 : for range in ranges.iter() {
2579 0 : let Some((start, end)) = truncate_to(
2580 0 : &range.start,
2581 0 : &range.end,
2582 0 : &job.compact_key_range.start,
2583 0 : &job.compact_key_range.end,
2584 0 : ) else {
2585 0 : continue;
2586 : };
2587 0 : split_key_ranges.push((start, end));
2588 : }
2589 0 : split_key_ranges.sort();
2590 0 : let all_layers = {
2591 0 : let guard = self.layers.read().await;
2592 0 : let layer_map = guard.layer_map()?;
2593 0 : layer_map.iter_historic_layers().collect_vec()
2594 0 : };
2595 0 : let mut current_start = None;
2596 0 : let ranges_num = split_key_ranges.len();
2597 0 : for (idx, (start, end)) in split_key_ranges.into_iter().enumerate() {
2598 0 : if current_start.is_none() {
2599 0 : current_start = Some(start);
2600 0 : }
2601 0 : let start = current_start.unwrap();
2602 0 : if start >= end {
2603 : // We have already processed this partition.
2604 0 : continue;
2605 0 : }
2606 0 : let overlapping_layers = {
2607 0 : let mut desc = Vec::new();
2608 0 : for layer in all_layers.iter() {
2609 0 : if overlaps_with(&layer.get_key_range(), &(start..end))
2610 0 : && layer.get_lsn_range().start <= compact_below_lsn
2611 0 : {
2612 0 : desc.push(layer.clone());
2613 0 : }
2614 : }
2615 0 : desc
2616 0 : };
2617 0 : let total_size = overlapping_layers.iter().map(|x| x.file_size).sum::<u64>();
2618 0 : if total_size > sub_compaction_max_job_size_mb * 1024 * 1024 || ranges_num == idx + 1 {
2619 : // Try to extend the compaction range so that we include at least one full layer file.
2620 0 : let extended_end = overlapping_layers
2621 0 : .iter()
2622 0 : .map(|layer| layer.key_range.end)
2623 0 : .min();
2624 : // It is possible that the search range does not contain any layer files when we reach the end of the loop.
2625 : // In this case, we simply use the specified key range end.
2626 0 : let end = if let Some(extended_end) = extended_end {
2627 0 : extended_end.max(end)
2628 : } else {
2629 0 : end
2630 : };
2631 0 : let end = if ranges_num == idx + 1 {
2632 : // extend the compaction range to the end of the key range if it's the last partition
2633 0 : end.max(job.compact_key_range.end)
2634 : } else {
2635 0 : end
2636 : };
2637 0 : if total_size == 0 && !compact_jobs.is_empty() {
2638 0 : info!(
2639 0 : "splitting compaction job: {}..{}, estimated_size={}, extending the previous job",
2640 : start, end, total_size
2641 : );
2642 0 : compact_jobs.last_mut().unwrap().compact_key_range.end = end;
2643 0 : current_start = Some(end);
2644 : } else {
2645 0 : info!(
2646 0 : "splitting compaction job: {}..{}, estimated_size={}",
2647 : start, end, total_size
2648 : );
2649 0 : compact_jobs.push(GcCompactJob {
2650 0 : dry_run: job.dry_run,
2651 0 : compact_key_range: start..end,
2652 0 : compact_lsn_range: job.compact_lsn_range.start..compact_below_lsn,
2653 0 : });
2654 0 : current_start = Some(end);
2655 : }
2656 0 : }
2657 : }
2658 0 : Ok(compact_jobs)
2659 0 : }
2660 :
2661 : /// An experimental compaction building block that combines compaction with garbage collection.
2662 : ///
2663 : /// The current implementation picks all delta + image layers that are below or intersecting with
2664 : /// the GC horizon without considering retain_lsns. Then, it does a full compaction over all these delta
2665 : /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
2666 : /// and create delta layers with all deltas >= gc horizon.
2667 : ///
2668 : /// If `options.compact_range` is provided, it will only compact the keys within the range, aka partial compaction.
2669 : /// Partial compaction will read and process all layers overlapping with the key range, even if it might
2670 : /// contain extra keys. After the gc-compaction phase completes, delta layers that are not fully contained
2671 : /// within the key range will be rewritten to ensure they do not overlap with the delta layers. Providing
2672 : /// Key::MIN..Key..MAX to the function indicates a full compaction, though technically, `Key::MAX` is not
2673 : /// part of the range.
2674 : ///
2675 : /// If `options.compact_lsn_range.end` is provided, the compaction will only compact layers below or intersect with
2676 : /// the LSN. Otherwise, it will use the gc cutoff by default.
2677 112 : pub(crate) async fn compact_with_gc(
2678 112 : self: &Arc<Self>,
2679 112 : cancel: &CancellationToken,
2680 112 : options: CompactOptions,
2681 112 : ctx: &RequestContext,
2682 112 : ) -> Result<CompactionOutcome, CompactionError> {
2683 112 : let sub_compaction = options.sub_compaction;
2684 112 : let job = GcCompactJob::from_compact_options(options.clone());
2685 112 : let yield_for_l0 = options.flags.contains(CompactFlags::YieldForL0);
2686 112 : if sub_compaction {
2687 0 : info!(
2688 0 : "running enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs"
2689 : );
2690 0 : let jobs = self
2691 0 : .gc_compaction_split_jobs(job, options.sub_compaction_max_job_size_mb)
2692 0 : .await?;
2693 0 : let jobs_len = jobs.len();
2694 0 : for (idx, job) in jobs.into_iter().enumerate() {
2695 0 : info!(
2696 0 : "running enhanced gc bottom-most compaction, sub-compaction {}/{}",
2697 0 : idx + 1,
2698 : jobs_len
2699 : );
2700 0 : self.compact_with_gc_inner(cancel, job, ctx, yield_for_l0)
2701 0 : .await?;
2702 : }
2703 0 : if jobs_len == 0 {
2704 0 : info!("no jobs to run, skipping gc bottom-most compaction");
2705 0 : }
2706 0 : return Ok(CompactionOutcome::Done);
2707 112 : }
2708 112 : self.compact_with_gc_inner(cancel, job, ctx, yield_for_l0)
2709 112 : .await
2710 112 : }
2711 :
2712 112 : async fn compact_with_gc_inner(
2713 112 : self: &Arc<Self>,
2714 112 : cancel: &CancellationToken,
2715 112 : job: GcCompactJob,
2716 112 : ctx: &RequestContext,
2717 112 : yield_for_l0: bool,
2718 112 : ) -> Result<CompactionOutcome, CompactionError> {
2719 112 : // Block other compaction/GC tasks from running for now. GC-compaction could run along
2720 112 : // with legacy compaction tasks in the future. Always ensure the lock order is compaction -> gc.
2721 112 : // Note that we already acquired the compaction lock when the outer `compact` function gets called.
2722 112 :
2723 112 : let timer = Instant::now();
2724 112 : let begin_timer = timer;
2725 112 :
2726 112 : let gc_lock = async {
2727 112 : tokio::select! {
2728 112 : guard = self.gc_lock.lock() => Ok(guard),
2729 112 : _ = cancel.cancelled() => Err(CompactionError::ShuttingDown),
2730 : }
2731 112 : };
2732 :
2733 112 : let time_acquire_lock = timer.elapsed();
2734 112 : let timer = Instant::now();
2735 :
2736 112 : let gc_lock = crate::timed(
2737 112 : gc_lock,
2738 112 : "acquires gc lock",
2739 112 : std::time::Duration::from_secs(5),
2740 112 : )
2741 112 : .await?;
2742 :
2743 112 : let dry_run = job.dry_run;
2744 112 : let compact_key_range = job.compact_key_range;
2745 112 : let compact_lsn_range = job.compact_lsn_range;
2746 :
2747 112 : let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
2748 :
2749 112 : info!(
2750 0 : "running enhanced gc bottom-most compaction, dry_run={dry_run}, compact_key_range={}..{}, compact_lsn_range={}..{}",
2751 : compact_key_range.start,
2752 : compact_key_range.end,
2753 : compact_lsn_range.start,
2754 : compact_lsn_range.end
2755 : );
2756 :
2757 112 : scopeguard::defer! {
2758 112 : info!("done enhanced gc bottom-most compaction");
2759 112 : };
2760 112 :
2761 112 : let mut stat = CompactionStatistics::default();
2762 :
2763 : // Step 0: pick all delta layers + image layers below/intersect with the GC horizon.
2764 : // The layer selection has the following properties:
2765 : // 1. If a layer is in the selection, all layers below it are in the selection.
2766 : // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
2767 108 : let job_desc = {
2768 112 : let guard = self.layers.read().await;
2769 112 : let layers = guard.layer_map()?;
2770 112 : let gc_info = self.gc_info.read().unwrap();
2771 112 : let mut retain_lsns_below_horizon = Vec::new();
2772 112 : let gc_cutoff = {
2773 : // Currently, gc-compaction only kicks in after the legacy gc has updated the gc_cutoff.
2774 : // Therefore, it can only clean up data that cannot be cleaned up with legacy gc, instead of
2775 : // cleaning everything that theoritically it could. In the future, it should use `self.gc_info`
2776 : // to get the truth data.
2777 112 : let real_gc_cutoff = self.get_gc_compaction_watermark();
2778 : // The compaction algorithm will keep all keys above the gc_cutoff while keeping only necessary keys below the gc_cutoff for
2779 : // each of the retain_lsn. Therefore, if the user-provided `compact_lsn_range.end` is larger than the real gc cutoff, we will use
2780 : // the real cutoff.
2781 112 : let mut gc_cutoff = if compact_lsn_range.end == Lsn::MAX {
2782 100 : if real_gc_cutoff == Lsn::INVALID {
2783 : // If the gc_cutoff is not generated yet, we should not compact anything.
2784 0 : tracing::warn!(
2785 0 : "no layers to compact with gc: gc_cutoff not generated yet, skipping gc bottom-most compaction"
2786 : );
2787 0 : return Ok(CompactionOutcome::Skipped);
2788 100 : }
2789 100 : real_gc_cutoff
2790 : } else {
2791 12 : compact_lsn_range.end
2792 : };
2793 112 : if gc_cutoff > real_gc_cutoff {
2794 8 : warn!(
2795 0 : "provided compact_lsn_range.end={} is larger than the real_gc_cutoff={}, using the real gc cutoff",
2796 : gc_cutoff, real_gc_cutoff
2797 : );
2798 8 : gc_cutoff = real_gc_cutoff;
2799 104 : }
2800 112 : gc_cutoff
2801 : };
2802 140 : for (lsn, _timeline_id, _is_offloaded) in &gc_info.retain_lsns {
2803 140 : if lsn < &gc_cutoff {
2804 140 : retain_lsns_below_horizon.push(*lsn);
2805 140 : }
2806 : }
2807 112 : for lsn in gc_info.leases.keys() {
2808 0 : if lsn < &gc_cutoff {
2809 0 : retain_lsns_below_horizon.push(*lsn);
2810 0 : }
2811 : }
2812 112 : let mut selected_layers: Vec<Layer> = Vec::new();
2813 112 : drop(gc_info);
2814 : // Firstly, pick all the layers intersect or below the gc_cutoff, get the largest LSN in the selected layers.
2815 112 : let Some(max_layer_lsn) = layers
2816 112 : .iter_historic_layers()
2817 500 : .filter(|desc| desc.get_lsn_range().start <= gc_cutoff)
2818 428 : .map(|desc| desc.get_lsn_range().end)
2819 112 : .max()
2820 : else {
2821 0 : info!(
2822 0 : "no layers to compact with gc: no historic layers below gc_cutoff, gc_cutoff={}",
2823 : gc_cutoff
2824 : );
2825 0 : return Ok(CompactionOutcome::Done);
2826 : };
2827 : // Next, if the user specifies compact_lsn_range.start, we need to filter some layers out. All the layers (strictly) below
2828 : // the min_layer_lsn computed as below will be filtered out and the data will be accessed using the normal read path, as if
2829 : // it is a branch.
2830 112 : let Some(min_layer_lsn) = layers
2831 112 : .iter_historic_layers()
2832 500 : .filter(|desc| {
2833 500 : if compact_lsn_range.start == Lsn::INVALID {
2834 408 : true // select all layers below if start == Lsn(0)
2835 : } else {
2836 92 : desc.get_lsn_range().end > compact_lsn_range.start // strictly larger than compact_above_lsn
2837 : }
2838 500 : })
2839 464 : .map(|desc| desc.get_lsn_range().start)
2840 112 : .min()
2841 : else {
2842 0 : info!(
2843 0 : "no layers to compact with gc: no historic layers above compact_above_lsn, compact_above_lsn={}",
2844 : compact_lsn_range.end
2845 : );
2846 0 : return Ok(CompactionOutcome::Done);
2847 : };
2848 : // Then, pick all the layers that are below the max_layer_lsn. This is to ensure we can pick all single-key
2849 : // layers to compact.
2850 112 : let mut rewrite_layers = Vec::new();
2851 500 : for desc in layers.iter_historic_layers() {
2852 500 : if desc.get_lsn_range().end <= max_layer_lsn
2853 428 : && desc.get_lsn_range().start >= min_layer_lsn
2854 392 : && overlaps_with(&desc.get_key_range(), &compact_key_range)
2855 : {
2856 : // If the layer overlaps with the compaction key range, we need to read it to obtain all keys within the range,
2857 : // even if it might contain extra keys
2858 316 : selected_layers.push(guard.get_from_desc(&desc));
2859 316 : // 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
2860 316 : // to overlap image layers)
2861 316 : if desc.is_delta() && !fully_contains(&compact_key_range, &desc.get_key_range())
2862 4 : {
2863 4 : rewrite_layers.push(desc);
2864 312 : }
2865 184 : }
2866 : }
2867 112 : if selected_layers.is_empty() {
2868 4 : info!(
2869 0 : "no layers to compact with gc: no layers within the key range, gc_cutoff={}, key_range={}..{}",
2870 : gc_cutoff, compact_key_range.start, compact_key_range.end
2871 : );
2872 4 : return Ok(CompactionOutcome::Done);
2873 108 : }
2874 108 : retain_lsns_below_horizon.sort();
2875 108 : GcCompactionJobDescription {
2876 108 : selected_layers,
2877 108 : gc_cutoff,
2878 108 : retain_lsns_below_horizon,
2879 108 : min_layer_lsn,
2880 108 : max_layer_lsn,
2881 108 : compaction_key_range: compact_key_range,
2882 108 : rewrite_layers,
2883 108 : }
2884 : };
2885 108 : let (has_data_below, lowest_retain_lsn) = if compact_lsn_range.start != Lsn::INVALID {
2886 : // If we only compact above some LSN, we should get the history from the current branch below the specified LSN.
2887 : // We use job_desc.min_layer_lsn as if it's the lowest branch point.
2888 16 : (true, job_desc.min_layer_lsn)
2889 92 : } else if self.ancestor_timeline.is_some() {
2890 : // In theory, we can also use min_layer_lsn here, but using ancestor LSN makes sure the delta layers cover the
2891 : // LSN ranges all the way to the ancestor timeline.
2892 4 : (true, self.ancestor_lsn)
2893 : } else {
2894 88 : let res = job_desc
2895 88 : .retain_lsns_below_horizon
2896 88 : .first()
2897 88 : .copied()
2898 88 : .unwrap_or(job_desc.gc_cutoff);
2899 88 : if debug_mode {
2900 88 : assert_eq!(
2901 88 : res,
2902 88 : job_desc
2903 88 : .retain_lsns_below_horizon
2904 88 : .iter()
2905 88 : .min()
2906 88 : .copied()
2907 88 : .unwrap_or(job_desc.gc_cutoff)
2908 88 : );
2909 0 : }
2910 88 : (false, res)
2911 : };
2912 108 : info!(
2913 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={}",
2914 0 : job_desc.selected_layers.len(),
2915 0 : job_desc.rewrite_layers.len(),
2916 : job_desc.max_layer_lsn,
2917 : job_desc.min_layer_lsn,
2918 : job_desc.gc_cutoff,
2919 : lowest_retain_lsn,
2920 : job_desc.compaction_key_range.start,
2921 : job_desc.compaction_key_range.end,
2922 : has_data_below,
2923 : );
2924 :
2925 108 : let time_analyze = timer.elapsed();
2926 108 : let timer = Instant::now();
2927 :
2928 424 : for layer in &job_desc.selected_layers {
2929 316 : debug!("read layer: {}", layer.layer_desc().key());
2930 : }
2931 112 : for layer in &job_desc.rewrite_layers {
2932 4 : debug!("rewrite layer: {}", layer.key());
2933 : }
2934 :
2935 108 : self.check_compaction_space(&job_desc.selected_layers)
2936 108 : .await?;
2937 :
2938 : // Generate statistics for the compaction
2939 424 : for layer in &job_desc.selected_layers {
2940 316 : let desc = layer.layer_desc();
2941 316 : if desc.is_delta() {
2942 176 : stat.visit_delta_layer(desc.file_size());
2943 176 : } else {
2944 140 : stat.visit_image_layer(desc.file_size());
2945 140 : }
2946 : }
2947 :
2948 : // Step 1: construct a k-merge iterator over all layers.
2949 : // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
2950 108 : let layer_names = job_desc
2951 108 : .selected_layers
2952 108 : .iter()
2953 316 : .map(|layer| layer.layer_desc().layer_name())
2954 108 : .collect_vec();
2955 108 : if let Some(err) = check_valid_layermap(&layer_names) {
2956 0 : return Err(CompactionError::Other(anyhow!(
2957 0 : "gc-compaction layer map check failed because {}, cannot proceed with compaction due to potential data loss",
2958 0 : err
2959 0 : )));
2960 108 : }
2961 108 : // The maximum LSN we are processing in this compaction loop
2962 108 : let end_lsn = job_desc
2963 108 : .selected_layers
2964 108 : .iter()
2965 316 : .map(|l| l.layer_desc().lsn_range.end)
2966 108 : .max()
2967 108 : .unwrap();
2968 108 : let mut delta_layers = Vec::new();
2969 108 : let mut image_layers = Vec::new();
2970 108 : let mut downloaded_layers = Vec::new();
2971 108 : let mut total_downloaded_size = 0;
2972 108 : let mut total_layer_size = 0;
2973 424 : for layer in &job_desc.selected_layers {
2974 316 : if layer
2975 316 : .needs_download()
2976 316 : .await
2977 316 : .context("failed to check if layer needs download")
2978 316 : .map_err(CompactionError::Other)?
2979 316 : .is_some()
2980 0 : {
2981 0 : total_downloaded_size += layer.layer_desc().file_size;
2982 316 : }
2983 316 : total_layer_size += layer.layer_desc().file_size;
2984 316 : if cancel.is_cancelled() {
2985 0 : return Err(CompactionError::ShuttingDown);
2986 316 : }
2987 316 : let should_yield = yield_for_l0
2988 0 : && self
2989 0 : .l0_compaction_trigger
2990 0 : .notified()
2991 0 : .now_or_never()
2992 0 : .is_some();
2993 316 : if should_yield {
2994 0 : tracing::info!("preempt gc-compaction when downloading layers: too many L0 layers");
2995 0 : return Ok(CompactionOutcome::YieldForL0);
2996 316 : }
2997 316 : let resident_layer = layer
2998 316 : .download_and_keep_resident(ctx)
2999 316 : .await
3000 316 : .context("failed to download and keep resident layer")
3001 316 : .map_err(CompactionError::Other)?;
3002 316 : downloaded_layers.push(resident_layer);
3003 : }
3004 108 : info!(
3005 0 : "finish downloading layers, downloaded={}, total={}, ratio={:.2}",
3006 0 : total_downloaded_size,
3007 0 : total_layer_size,
3008 0 : total_downloaded_size as f64 / total_layer_size as f64
3009 : );
3010 424 : for resident_layer in &downloaded_layers {
3011 316 : if resident_layer.layer_desc().is_delta() {
3012 176 : let layer = resident_layer
3013 176 : .get_as_delta(ctx)
3014 176 : .await
3015 176 : .context("failed to get delta layer")
3016 176 : .map_err(CompactionError::Other)?;
3017 176 : delta_layers.push(layer);
3018 : } else {
3019 140 : let layer = resident_layer
3020 140 : .get_as_image(ctx)
3021 140 : .await
3022 140 : .context("failed to get image layer")
3023 140 : .map_err(CompactionError::Other)?;
3024 140 : image_layers.push(layer);
3025 : }
3026 : }
3027 108 : let (dense_ks, sparse_ks) = self
3028 108 : .collect_gc_compaction_keyspace()
3029 108 : .await
3030 108 : .context("failed to collect gc compaction keyspace")
3031 108 : .map_err(CompactionError::Other)?;
3032 108 : let mut merge_iter = FilterIterator::create(
3033 108 : MergeIterator::create(&delta_layers, &image_layers, ctx),
3034 108 : dense_ks,
3035 108 : sparse_ks,
3036 108 : )
3037 108 : .context("failed to create filter iterator")
3038 108 : .map_err(CompactionError::Other)?;
3039 :
3040 108 : let time_download_layer = timer.elapsed();
3041 108 : let mut timer = Instant::now();
3042 108 :
3043 108 : // Step 2: Produce images+deltas.
3044 108 : let mut accumulated_values = Vec::new();
3045 108 : let mut last_key: Option<Key> = None;
3046 :
3047 : // Only create image layers when there is no ancestor branches. TODO: create covering image layer
3048 : // when some condition meet.
3049 108 : let mut image_layer_writer = if !has_data_below {
3050 : Some(
3051 88 : SplitImageLayerWriter::new(
3052 88 : self.conf,
3053 88 : self.timeline_id,
3054 88 : self.tenant_shard_id,
3055 88 : job_desc.compaction_key_range.start,
3056 88 : lowest_retain_lsn,
3057 88 : self.get_compaction_target_size(),
3058 88 : ctx,
3059 88 : )
3060 88 : .await
3061 88 : .context("failed to create image layer writer")
3062 88 : .map_err(CompactionError::Other)?,
3063 : )
3064 : } else {
3065 20 : None
3066 : };
3067 :
3068 108 : let mut delta_layer_writer = SplitDeltaLayerWriter::new(
3069 108 : self.conf,
3070 108 : self.timeline_id,
3071 108 : self.tenant_shard_id,
3072 108 : lowest_retain_lsn..end_lsn,
3073 108 : self.get_compaction_target_size(),
3074 108 : )
3075 108 : .await
3076 108 : .context("failed to create delta layer writer")
3077 108 : .map_err(CompactionError::Other)?;
3078 :
3079 : #[derive(Default)]
3080 : struct RewritingLayers {
3081 : before: Option<DeltaLayerWriter>,
3082 : after: Option<DeltaLayerWriter>,
3083 : }
3084 108 : let mut delta_layer_rewriters = HashMap::<Arc<PersistentLayerKey>, RewritingLayers>::new();
3085 :
3086 : /// When compacting not at a bottom range (=`[0,X)`) of the root branch, we "have data below" (`has_data_below=true`).
3087 : /// The two cases are compaction in ancestor branches and when `compact_lsn_range.start` is set.
3088 : /// In those cases, we need to pull up data from below the LSN range we're compaction.
3089 : ///
3090 : /// This function unifies the cases so that later code doesn't have to think about it.
3091 : ///
3092 : /// Currently, we always get the ancestor image for each key in the child branch no matter whether the image
3093 : /// is needed for reconstruction. This should be fixed in the future.
3094 : ///
3095 : /// Furthermore, we should do vectored get instead of a single get, or better, use k-merge for ancestor
3096 : /// images.
3097 1280 : async fn get_ancestor_image(
3098 1280 : this_tline: &Arc<Timeline>,
3099 1280 : key: Key,
3100 1280 : ctx: &RequestContext,
3101 1280 : has_data_below: bool,
3102 1280 : history_lsn_point: Lsn,
3103 1280 : ) -> anyhow::Result<Option<(Key, Lsn, Bytes)>> {
3104 1280 : if !has_data_below {
3105 1204 : return Ok(None);
3106 76 : };
3107 : // This function is implemented as a get of the current timeline at ancestor LSN, therefore reusing
3108 : // as much existing code as possible.
3109 76 : let img = this_tline.get(key, history_lsn_point, ctx).await?;
3110 76 : Ok(Some((key, history_lsn_point, img)))
3111 1280 : }
3112 :
3113 : // Actually, we can decide not to write to the image layer at all at this point because
3114 : // the key and LSN range are determined. However, to keep things simple here, we still
3115 : // create this writer, and discard the writer in the end.
3116 108 : let mut time_to_first_kv_pair = None;
3117 :
3118 1984 : while let Some(((key, lsn, val), desc)) = merge_iter
3119 1984 : .next_with_trace()
3120 1984 : .await
3121 1984 : .context("failed to get next key-value pair")
3122 1984 : .map_err(CompactionError::Other)?
3123 : {
3124 1880 : if time_to_first_kv_pair.is_none() {
3125 108 : time_to_first_kv_pair = Some(timer.elapsed());
3126 108 : timer = Instant::now();
3127 1772 : }
3128 :
3129 1880 : if cancel.is_cancelled() {
3130 0 : return Err(CompactionError::ShuttingDown);
3131 1880 : }
3132 :
3133 1880 : let should_yield = yield_for_l0
3134 0 : && self
3135 0 : .l0_compaction_trigger
3136 0 : .notified()
3137 0 : .now_or_never()
3138 0 : .is_some();
3139 1880 : if should_yield {
3140 0 : tracing::info!("preempt gc-compaction in the main loop: too many L0 layers");
3141 0 : return Ok(CompactionOutcome::YieldForL0);
3142 1880 : }
3143 1880 : if self.shard_identity.is_key_disposable(&key) {
3144 : // If this shard does not need to store this key, simply skip it.
3145 : //
3146 : // This is not handled in the filter iterator because shard is determined by hash.
3147 : // Therefore, it does not give us any performance benefit to do things like skip
3148 : // a whole layer file as handling key spaces (ranges).
3149 0 : if cfg!(debug_assertions) {
3150 0 : let shard = self.shard_identity.shard_index();
3151 0 : let owner = self.shard_identity.get_shard_number(&key);
3152 0 : panic!("key {key} does not belong on shard {shard}, owned by {owner}");
3153 0 : }
3154 0 : continue;
3155 1880 : }
3156 1880 : if !job_desc.compaction_key_range.contains(&key) {
3157 128 : if !desc.is_delta {
3158 120 : continue;
3159 8 : }
3160 8 : let rewriter = delta_layer_rewriters.entry(desc.clone()).or_default();
3161 8 : let rewriter = if key < job_desc.compaction_key_range.start {
3162 0 : if rewriter.before.is_none() {
3163 0 : rewriter.before = Some(
3164 0 : DeltaLayerWriter::new(
3165 0 : self.conf,
3166 0 : self.timeline_id,
3167 0 : self.tenant_shard_id,
3168 0 : desc.key_range.start,
3169 0 : desc.lsn_range.clone(),
3170 0 : ctx,
3171 0 : )
3172 0 : .await
3173 0 : .context("failed to create delta layer writer")
3174 0 : .map_err(CompactionError::Other)?,
3175 : );
3176 0 : }
3177 0 : rewriter.before.as_mut().unwrap()
3178 8 : } else if key >= job_desc.compaction_key_range.end {
3179 8 : if rewriter.after.is_none() {
3180 4 : rewriter.after = Some(
3181 4 : DeltaLayerWriter::new(
3182 4 : self.conf,
3183 4 : self.timeline_id,
3184 4 : self.tenant_shard_id,
3185 4 : job_desc.compaction_key_range.end,
3186 4 : desc.lsn_range.clone(),
3187 4 : ctx,
3188 4 : )
3189 4 : .await
3190 4 : .context("failed to create delta layer writer")
3191 4 : .map_err(CompactionError::Other)?,
3192 : );
3193 4 : }
3194 8 : rewriter.after.as_mut().unwrap()
3195 : } else {
3196 0 : unreachable!()
3197 : };
3198 8 : rewriter
3199 8 : .put_value(key, lsn, val, ctx)
3200 8 : .await
3201 8 : .context("failed to put value")
3202 8 : .map_err(CompactionError::Other)?;
3203 8 : continue;
3204 1752 : }
3205 1752 : match val {
3206 1260 : Value::Image(_) => stat.visit_image_key(&val),
3207 492 : Value::WalRecord(_) => stat.visit_wal_key(&val),
3208 : }
3209 1752 : if last_key.is_none() || last_key.as_ref() == Some(&key) {
3210 576 : if last_key.is_none() {
3211 108 : last_key = Some(key);
3212 468 : }
3213 576 : accumulated_values.push((key, lsn, val));
3214 : } else {
3215 1176 : let last_key: &mut Key = last_key.as_mut().unwrap();
3216 1176 : stat.on_unique_key_visited(); // TODO: adjust statistics for partial compaction
3217 1176 : let retention = self
3218 1176 : .generate_key_retention(
3219 1176 : *last_key,
3220 1176 : &accumulated_values,
3221 1176 : job_desc.gc_cutoff,
3222 1176 : &job_desc.retain_lsns_below_horizon,
3223 1176 : COMPACTION_DELTA_THRESHOLD,
3224 1176 : get_ancestor_image(self, *last_key, ctx, has_data_below, lowest_retain_lsn)
3225 1176 : .await
3226 1176 : .context("failed to get ancestor image")
3227 1176 : .map_err(CompactionError::Other)?,
3228 : )
3229 1176 : .await
3230 1176 : .context("failed to generate key retention")
3231 1176 : .map_err(CompactionError::Other)?;
3232 1172 : retention
3233 1172 : .pipe_to(
3234 1172 : *last_key,
3235 1172 : &mut delta_layer_writer,
3236 1172 : image_layer_writer.as_mut(),
3237 1172 : &mut stat,
3238 1172 : ctx,
3239 1172 : )
3240 1172 : .await
3241 1172 : .context("failed to pipe to delta layer writer")
3242 1172 : .map_err(CompactionError::Other)?;
3243 1172 : accumulated_values.clear();
3244 1172 : *last_key = key;
3245 1172 : accumulated_values.push((key, lsn, val));
3246 : }
3247 : }
3248 :
3249 : // TODO: move the below part to the loop body
3250 104 : let Some(last_key) = last_key else {
3251 0 : return Err(CompactionError::Other(anyhow!(
3252 0 : "no keys produced during compaction"
3253 0 : )));
3254 : };
3255 104 : stat.on_unique_key_visited();
3256 :
3257 104 : let retention = self
3258 104 : .generate_key_retention(
3259 104 : last_key,
3260 104 : &accumulated_values,
3261 104 : job_desc.gc_cutoff,
3262 104 : &job_desc.retain_lsns_below_horizon,
3263 104 : COMPACTION_DELTA_THRESHOLD,
3264 104 : get_ancestor_image(self, last_key, ctx, has_data_below, lowest_retain_lsn)
3265 104 : .await
3266 104 : .context("failed to get ancestor image")
3267 104 : .map_err(CompactionError::Other)?,
3268 : )
3269 104 : .await
3270 104 : .context("failed to generate key retention")
3271 104 : .map_err(CompactionError::Other)?;
3272 104 : retention
3273 104 : .pipe_to(
3274 104 : last_key,
3275 104 : &mut delta_layer_writer,
3276 104 : image_layer_writer.as_mut(),
3277 104 : &mut stat,
3278 104 : ctx,
3279 104 : )
3280 104 : .await
3281 104 : .context("failed to pipe to delta layer writer")
3282 104 : .map_err(CompactionError::Other)?;
3283 : // end: move the above part to the loop body
3284 :
3285 104 : let time_main_loop = timer.elapsed();
3286 104 : let timer = Instant::now();
3287 104 :
3288 104 : let mut rewrote_delta_layers = Vec::new();
3289 108 : for (key, writers) in delta_layer_rewriters {
3290 4 : if let Some(delta_writer_before) = writers.before {
3291 0 : let (desc, path) = delta_writer_before
3292 0 : .finish(job_desc.compaction_key_range.start, ctx)
3293 0 : .await
3294 0 : .context("failed to finish delta layer writer")
3295 0 : .map_err(CompactionError::Other)?;
3296 0 : let layer = Layer::finish_creating(self.conf, self, desc, &path)
3297 0 : .context("failed to finish creating delta layer")
3298 0 : .map_err(CompactionError::Other)?;
3299 0 : rewrote_delta_layers.push(layer);
3300 4 : }
3301 4 : if let Some(delta_writer_after) = writers.after {
3302 4 : let (desc, path) = delta_writer_after
3303 4 : .finish(key.key_range.end, ctx)
3304 4 : .await
3305 4 : .context("failed to finish delta layer writer")
3306 4 : .map_err(CompactionError::Other)?;
3307 4 : let layer = Layer::finish_creating(self.conf, self, desc, &path)
3308 4 : .context("failed to finish creating delta layer")
3309 4 : .map_err(CompactionError::Other)?;
3310 4 : rewrote_delta_layers.push(layer);
3311 0 : }
3312 : }
3313 :
3314 148 : let discard = |key: &PersistentLayerKey| {
3315 148 : let key = key.clone();
3316 148 : async move { KeyHistoryRetention::discard_key(&key, self, dry_run).await }
3317 148 : };
3318 :
3319 104 : let produced_image_layers = if let Some(writer) = image_layer_writer {
3320 84 : if !dry_run {
3321 76 : let end_key = job_desc.compaction_key_range.end;
3322 76 : writer
3323 76 : .finish_with_discard_fn(self, ctx, end_key, discard)
3324 76 : .await
3325 76 : .context("failed to finish image layer writer")
3326 76 : .map_err(CompactionError::Other)?
3327 : } else {
3328 8 : drop(writer);
3329 8 : Vec::new()
3330 : }
3331 : } else {
3332 20 : Vec::new()
3333 : };
3334 :
3335 104 : let produced_delta_layers = if !dry_run {
3336 96 : delta_layer_writer
3337 96 : .finish_with_discard_fn(self, ctx, discard)
3338 96 : .await
3339 96 : .context("failed to finish delta layer writer")
3340 96 : .map_err(CompactionError::Other)?
3341 : } else {
3342 8 : drop(delta_layer_writer);
3343 8 : Vec::new()
3344 : };
3345 :
3346 : // TODO: make image/delta/rewrote_delta layers generation atomic. At this point, we already generated resident layers, and if
3347 : // compaction is cancelled at this point, we might have some layers that are not cleaned up.
3348 104 : let mut compact_to = Vec::new();
3349 104 : let mut keep_layers = HashSet::new();
3350 104 : let produced_delta_layers_len = produced_delta_layers.len();
3351 104 : let produced_image_layers_len = produced_image_layers.len();
3352 104 :
3353 104 : let layer_selection_by_key = job_desc
3354 104 : .selected_layers
3355 104 : .iter()
3356 304 : .map(|l| (l.layer_desc().key(), l.layer_desc().clone()))
3357 104 : .collect::<HashMap<_, _>>();
3358 :
3359 176 : for action in produced_delta_layers {
3360 72 : match action {
3361 44 : BatchWriterResult::Produced(layer) => {
3362 44 : if cfg!(debug_assertions) {
3363 44 : info!("produced delta layer: {}", layer.layer_desc().key());
3364 0 : }
3365 44 : stat.produce_delta_layer(layer.layer_desc().file_size());
3366 44 : compact_to.push(layer);
3367 : }
3368 28 : BatchWriterResult::Discarded(l) => {
3369 28 : if cfg!(debug_assertions) {
3370 28 : info!("discarded delta layer: {}", l);
3371 0 : }
3372 28 : if let Some(layer_desc) = layer_selection_by_key.get(&l) {
3373 28 : stat.discard_delta_layer(layer_desc.file_size());
3374 28 : } else {
3375 0 : tracing::warn!(
3376 0 : "discarded delta layer not in layer_selection: {}, produced a layer outside of the compaction key range?",
3377 : l
3378 : );
3379 0 : stat.discard_delta_layer(0);
3380 : }
3381 28 : keep_layers.insert(l);
3382 : }
3383 : }
3384 : }
3385 108 : for layer in &rewrote_delta_layers {
3386 4 : debug!(
3387 0 : "produced rewritten delta layer: {}",
3388 0 : layer.layer_desc().key()
3389 : );
3390 : // For now, we include rewritten delta layer size in the "produce_delta_layer". We could
3391 : // make it a separate statistics in the future.
3392 4 : stat.produce_delta_layer(layer.layer_desc().file_size());
3393 : }
3394 104 : compact_to.extend(rewrote_delta_layers);
3395 180 : for action in produced_image_layers {
3396 76 : match action {
3397 60 : BatchWriterResult::Produced(layer) => {
3398 60 : debug!("produced image layer: {}", layer.layer_desc().key());
3399 60 : stat.produce_image_layer(layer.layer_desc().file_size());
3400 60 : compact_to.push(layer);
3401 : }
3402 16 : BatchWriterResult::Discarded(l) => {
3403 16 : debug!("discarded image layer: {}", l);
3404 16 : if let Some(layer_desc) = layer_selection_by_key.get(&l) {
3405 16 : stat.discard_image_layer(layer_desc.file_size());
3406 16 : } else {
3407 0 : tracing::warn!(
3408 0 : "discarded image layer not in layer_selection: {}, produced a layer outside of the compaction key range?",
3409 : l
3410 : );
3411 0 : stat.discard_image_layer(0);
3412 : }
3413 16 : keep_layers.insert(l);
3414 : }
3415 : }
3416 : }
3417 :
3418 104 : let mut layer_selection = job_desc.selected_layers;
3419 :
3420 : // Partial compaction might select more data than it processes, e.g., if
3421 : // the compaction_key_range only partially overlaps:
3422 : //
3423 : // [---compaction_key_range---]
3424 : // [---A----][----B----][----C----][----D----]
3425 : //
3426 : // For delta layers, we will rewrite the layers so that it is cut exactly at
3427 : // the compaction key range, so we can always discard them. However, for image
3428 : // layers, as we do not rewrite them for now, we need to handle them differently.
3429 : // Assume image layers A, B, C, D are all in the `layer_selection`.
3430 : //
3431 : // The created image layers contain whatever is needed from B, C, and from
3432 : // `----]` of A, and from `[---` of D.
3433 : //
3434 : // In contrast, `[---A` and `D----]` have not been processed, so, we must
3435 : // keep that data.
3436 : //
3437 : // The solution for now is to keep A and D completely if they are image layers.
3438 : // (layer_selection is what we'll remove from the layer map, so, retain what
3439 : // is _not_ fully covered by compaction_key_range).
3440 408 : for layer in &layer_selection {
3441 304 : if !layer.layer_desc().is_delta() {
3442 132 : if !overlaps_with(
3443 132 : &layer.layer_desc().key_range,
3444 132 : &job_desc.compaction_key_range,
3445 132 : ) {
3446 0 : return Err(CompactionError::Other(anyhow!(
3447 0 : "violated constraint: image layer outside of compaction key range"
3448 0 : )));
3449 132 : }
3450 132 : if !fully_contains(
3451 132 : &job_desc.compaction_key_range,
3452 132 : &layer.layer_desc().key_range,
3453 132 : ) {
3454 16 : keep_layers.insert(layer.layer_desc().key());
3455 116 : }
3456 172 : }
3457 : }
3458 :
3459 304 : layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
3460 104 :
3461 104 : let time_final_phase = timer.elapsed();
3462 104 :
3463 104 : stat.time_final_phase_secs = time_final_phase.as_secs_f64();
3464 104 : stat.time_to_first_kv_pair_secs = time_to_first_kv_pair
3465 104 : .unwrap_or(Duration::ZERO)
3466 104 : .as_secs_f64();
3467 104 : stat.time_main_loop_secs = time_main_loop.as_secs_f64();
3468 104 : stat.time_acquire_lock_secs = time_acquire_lock.as_secs_f64();
3469 104 : stat.time_download_layer_secs = time_download_layer.as_secs_f64();
3470 104 : stat.time_analyze_secs = time_analyze.as_secs_f64();
3471 104 : stat.time_total_secs = begin_timer.elapsed().as_secs_f64();
3472 104 : stat.finalize();
3473 104 :
3474 104 : info!(
3475 0 : "gc-compaction statistics: {}",
3476 0 : serde_json::to_string(&stat)
3477 0 : .context("failed to serialize gc-compaction statistics")
3478 0 : .map_err(CompactionError::Other)?
3479 : );
3480 :
3481 104 : if dry_run {
3482 8 : return Ok(CompactionOutcome::Done);
3483 96 : }
3484 96 :
3485 96 : info!(
3486 0 : "produced {} delta layers and {} image layers, {} layers are kept",
3487 0 : produced_delta_layers_len,
3488 0 : produced_image_layers_len,
3489 0 : keep_layers.len()
3490 : );
3491 :
3492 : // Step 3: Place back to the layer map.
3493 :
3494 : // First, do a sanity check to ensure the newly-created layer map does not contain overlaps.
3495 96 : let all_layers = {
3496 96 : let guard = self.layers.read().await;
3497 96 : let layer_map = guard.layer_map()?;
3498 96 : layer_map.iter_historic_layers().collect_vec()
3499 96 : };
3500 96 :
3501 96 : let mut final_layers = all_layers
3502 96 : .iter()
3503 428 : .map(|layer| layer.layer_name())
3504 96 : .collect::<HashSet<_>>();
3505 304 : for layer in &layer_selection {
3506 208 : final_layers.remove(&layer.layer_desc().layer_name());
3507 208 : }
3508 204 : for layer in &compact_to {
3509 108 : final_layers.insert(layer.layer_desc().layer_name());
3510 108 : }
3511 96 : let final_layers = final_layers.into_iter().collect_vec();
3512 :
3513 : // 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
3514 : // the writer, so potentially, we will need a function like `ImageLayerBatchWriter::get_all_pending_layer_keys` to get all the keys that are
3515 : // in the writer before finalizing the persistent layers. Now we would leave some dangling layers on the disk if the check fails.
3516 96 : if let Some(err) = check_valid_layermap(&final_layers) {
3517 0 : return Err(CompactionError::Other(anyhow!(
3518 0 : "gc-compaction layer map check failed after compaction because {}, compaction result not applied to the layer map due to potential data loss",
3519 0 : err
3520 0 : )));
3521 96 : }
3522 :
3523 : // Between the sanity check and this compaction update, there could be new layers being flushed, but it should be fine because we only
3524 : // operate on L1 layers.
3525 : {
3526 : // Gc-compaction will rewrite the history of a key. This could happen in two ways:
3527 : //
3528 : // 1. We create an image layer to replace all the deltas below the compact LSN. In this case, assume
3529 : // we have 2 delta layers A and B, both below the compact LSN. We create an image layer I to replace
3530 : // A and B at the compact LSN. If the read path finishes reading A, yields, and now we update the layer
3531 : // map, the read path then cannot find any keys below A, reporting a missing key error, while the key
3532 : // now gets stored in I at the compact LSN.
3533 : //
3534 : // --------------- ---------------
3535 : // delta1@LSN20 image1@LSN20
3536 : // --------------- (read path collects delta@LSN20, => --------------- (read path cannot find anything
3537 : // delta1@LSN10 yields) below LSN 20)
3538 : // ---------------
3539 : //
3540 : // 2. We create a delta layer to replace all the deltas below the compact LSN, and in the delta layers,
3541 : // we combines the history of a key into a single image. For example, we have deltas at LSN 1, 2, 3, 4,
3542 : // Assume one delta layer contains LSN 1, 2, 3 and the other contains LSN 4.
3543 : //
3544 : // We let gc-compaction combine delta 2, 3, 4 into an image at LSN 4, which produces a delta layer that
3545 : // contains the delta at LSN 1, the image at LSN 4. If the read path finishes reading the original delta
3546 : // layer containing 4, yields, and we update the layer map to put the delta layer.
3547 : //
3548 : // --------------- ---------------
3549 : // delta1@LSN4 image1@LSN4
3550 : // --------------- (read path collects delta@LSN4, => --------------- (read path collects LSN4 and LSN1,
3551 : // delta1@LSN1-3 yields) delta1@LSN1 which is an invalid history)
3552 : // --------------- ---------------
3553 : //
3554 : // Therefore, the gc-compaction layer update operation should wait for all ongoing reads, block all pending reads,
3555 : // and only allow reads to continue after the update is finished.
3556 :
3557 96 : let update_guard = self.gc_compaction_layer_update_lock.write().await;
3558 : // Acquiring the update guard ensures current read operations end and new read operations are blocked.
3559 : // TODO: can we use `latest_gc_cutoff` Rcu to achieve the same effect?
3560 96 : let mut guard = self.layers.write().await;
3561 96 : guard
3562 96 : .open_mut()?
3563 96 : .finish_gc_compaction(&layer_selection, &compact_to, &self.metrics);
3564 96 : drop(update_guard); // Allow new reads to start ONLY after we finished updating the layer map.
3565 96 : };
3566 96 :
3567 96 : // Schedule an index-only upload to update the `latest_gc_cutoff` in the index_part.json.
3568 96 : // Otherwise, after restart, the index_part only contains the old `latest_gc_cutoff` and
3569 96 : // find_gc_cutoffs will try accessing things below the cutoff. TODO: ideally, this should
3570 96 : // be batched into `schedule_compaction_update`.
3571 96 : let disk_consistent_lsn = self.disk_consistent_lsn.load();
3572 96 : self.schedule_uploads(disk_consistent_lsn, None)
3573 96 : .context("failed to schedule uploads")
3574 96 : .map_err(CompactionError::Other)?;
3575 : // If a layer gets rewritten throughout gc-compaction, we need to keep that layer only in `compact_to` instead
3576 : // of `compact_from`.
3577 96 : let compact_from = {
3578 96 : let mut compact_from = Vec::new();
3579 96 : let mut compact_to_set = HashMap::new();
3580 204 : for layer in &compact_to {
3581 108 : compact_to_set.insert(layer.layer_desc().key(), layer);
3582 108 : }
3583 304 : for layer in &layer_selection {
3584 208 : if let Some(to) = compact_to_set.get(&layer.layer_desc().key()) {
3585 0 : tracing::info!(
3586 0 : "skipping delete {} because found same layer key at different generation {}",
3587 : layer,
3588 : to
3589 : );
3590 208 : } else {
3591 208 : compact_from.push(layer.clone());
3592 208 : }
3593 : }
3594 96 : compact_from
3595 96 : };
3596 96 : self.remote_client
3597 96 : .schedule_compaction_update(&compact_from, &compact_to)?;
3598 :
3599 96 : drop(gc_lock);
3600 96 :
3601 96 : Ok(CompactionOutcome::Done)
3602 112 : }
3603 : }
3604 :
3605 : struct TimelineAdaptor {
3606 : timeline: Arc<Timeline>,
3607 :
3608 : keyspace: (Lsn, KeySpace),
3609 :
3610 : new_deltas: Vec<ResidentLayer>,
3611 : new_images: Vec<ResidentLayer>,
3612 : layers_to_delete: Vec<Arc<PersistentLayerDesc>>,
3613 : }
3614 :
3615 : impl TimelineAdaptor {
3616 0 : pub fn new(timeline: &Arc<Timeline>, keyspace: (Lsn, KeySpace)) -> Self {
3617 0 : Self {
3618 0 : timeline: timeline.clone(),
3619 0 : keyspace,
3620 0 : new_images: Vec::new(),
3621 0 : new_deltas: Vec::new(),
3622 0 : layers_to_delete: Vec::new(),
3623 0 : }
3624 0 : }
3625 :
3626 0 : pub async fn flush_updates(&mut self) -> Result<(), CompactionError> {
3627 0 : let layers_to_delete = {
3628 0 : let guard = self.timeline.layers.read().await;
3629 0 : self.layers_to_delete
3630 0 : .iter()
3631 0 : .map(|x| guard.get_from_desc(x))
3632 0 : .collect::<Vec<Layer>>()
3633 0 : };
3634 0 : self.timeline
3635 0 : .finish_compact_batch(&self.new_deltas, &self.new_images, &layers_to_delete)
3636 0 : .await?;
3637 :
3638 0 : self.timeline
3639 0 : .upload_new_image_layers(std::mem::take(&mut self.new_images))?;
3640 :
3641 0 : self.new_deltas.clear();
3642 0 : self.layers_to_delete.clear();
3643 0 : Ok(())
3644 0 : }
3645 : }
3646 :
3647 : #[derive(Clone)]
3648 : struct ResidentDeltaLayer(ResidentLayer);
3649 : #[derive(Clone)]
3650 : struct ResidentImageLayer(ResidentLayer);
3651 :
3652 : impl CompactionJobExecutor for TimelineAdaptor {
3653 : type Key = pageserver_api::key::Key;
3654 :
3655 : type Layer = OwnArc<PersistentLayerDesc>;
3656 : type DeltaLayer = ResidentDeltaLayer;
3657 : type ImageLayer = ResidentImageLayer;
3658 :
3659 : type RequestContext = crate::context::RequestContext;
3660 :
3661 0 : fn get_shard_identity(&self) -> &ShardIdentity {
3662 0 : self.timeline.get_shard_identity()
3663 0 : }
3664 :
3665 0 : async fn get_layers(
3666 0 : &mut self,
3667 0 : key_range: &Range<Key>,
3668 0 : lsn_range: &Range<Lsn>,
3669 0 : _ctx: &RequestContext,
3670 0 : ) -> anyhow::Result<Vec<OwnArc<PersistentLayerDesc>>> {
3671 0 : self.flush_updates().await?;
3672 :
3673 0 : let guard = self.timeline.layers.read().await;
3674 0 : let layer_map = guard.layer_map()?;
3675 :
3676 0 : let result = layer_map
3677 0 : .iter_historic_layers()
3678 0 : .filter(|l| {
3679 0 : overlaps_with(&l.lsn_range, lsn_range) && overlaps_with(&l.key_range, key_range)
3680 0 : })
3681 0 : .map(OwnArc)
3682 0 : .collect();
3683 0 : Ok(result)
3684 0 : }
3685 :
3686 0 : async fn get_keyspace(
3687 0 : &mut self,
3688 0 : key_range: &Range<Key>,
3689 0 : lsn: Lsn,
3690 0 : _ctx: &RequestContext,
3691 0 : ) -> anyhow::Result<Vec<Range<Key>>> {
3692 0 : if lsn == self.keyspace.0 {
3693 0 : Ok(pageserver_compaction::helpers::intersect_keyspace(
3694 0 : &self.keyspace.1.ranges,
3695 0 : key_range,
3696 0 : ))
3697 : } else {
3698 : // The current compaction implementation only ever requests the key space
3699 : // at the compaction end LSN.
3700 0 : anyhow::bail!("keyspace not available for requested lsn");
3701 : }
3702 0 : }
3703 :
3704 0 : async fn downcast_delta_layer(
3705 0 : &self,
3706 0 : layer: &OwnArc<PersistentLayerDesc>,
3707 0 : ctx: &RequestContext,
3708 0 : ) -> anyhow::Result<Option<ResidentDeltaLayer>> {
3709 0 : // this is a lot more complex than a simple downcast...
3710 0 : if layer.is_delta() {
3711 0 : let l = {
3712 0 : let guard = self.timeline.layers.read().await;
3713 0 : guard.get_from_desc(layer)
3714 : };
3715 0 : let result = l.download_and_keep_resident(ctx).await?;
3716 :
3717 0 : Ok(Some(ResidentDeltaLayer(result)))
3718 : } else {
3719 0 : Ok(None)
3720 : }
3721 0 : }
3722 :
3723 0 : async fn create_image(
3724 0 : &mut self,
3725 0 : lsn: Lsn,
3726 0 : key_range: &Range<Key>,
3727 0 : ctx: &RequestContext,
3728 0 : ) -> anyhow::Result<()> {
3729 0 : Ok(self.create_image_impl(lsn, key_range, ctx).await?)
3730 0 : }
3731 :
3732 0 : async fn create_delta(
3733 0 : &mut self,
3734 0 : lsn_range: &Range<Lsn>,
3735 0 : key_range: &Range<Key>,
3736 0 : input_layers: &[ResidentDeltaLayer],
3737 0 : ctx: &RequestContext,
3738 0 : ) -> anyhow::Result<()> {
3739 0 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
3740 :
3741 0 : let mut all_entries = Vec::new();
3742 0 : for dl in input_layers.iter() {
3743 0 : all_entries.extend(dl.load_keys(ctx).await?);
3744 : }
3745 :
3746 : // The current stdlib sorting implementation is designed in a way where it is
3747 : // particularly fast where the slice is made up of sorted sub-ranges.
3748 0 : all_entries.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
3749 :
3750 0 : let mut writer = DeltaLayerWriter::new(
3751 0 : self.timeline.conf,
3752 0 : self.timeline.timeline_id,
3753 0 : self.timeline.tenant_shard_id,
3754 0 : key_range.start,
3755 0 : lsn_range.clone(),
3756 0 : ctx,
3757 0 : )
3758 0 : .await?;
3759 :
3760 0 : let mut dup_values = 0;
3761 0 :
3762 0 : // This iterator walks through all key-value pairs from all the layers
3763 0 : // we're compacting, in key, LSN order.
3764 0 : let mut prev: Option<(Key, Lsn)> = None;
3765 : for &DeltaEntry {
3766 0 : key, lsn, ref val, ..
3767 0 : } in all_entries.iter()
3768 : {
3769 0 : if prev == Some((key, lsn)) {
3770 : // This is a duplicate. Skip it.
3771 : //
3772 : // It can happen if compaction is interrupted after writing some
3773 : // layers but not all, and we are compacting the range again.
3774 : // The calculations in the algorithm assume that there are no
3775 : // duplicates, so the math on targeted file size is likely off,
3776 : // and we will create smaller files than expected.
3777 0 : dup_values += 1;
3778 0 : continue;
3779 0 : }
3780 :
3781 0 : let value = val.load(ctx).await?;
3782 :
3783 0 : writer.put_value(key, lsn, value, ctx).await?;
3784 :
3785 0 : prev = Some((key, lsn));
3786 : }
3787 :
3788 0 : if dup_values > 0 {
3789 0 : warn!("delta layer created with {} duplicate values", dup_values);
3790 0 : }
3791 :
3792 0 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
3793 0 : Err(anyhow::anyhow!(
3794 0 : "failpoint delta-layer-writer-fail-before-finish"
3795 0 : ))
3796 0 : });
3797 :
3798 0 : let (desc, path) = writer.finish(prev.unwrap().0.next(), ctx).await?;
3799 0 : let new_delta_layer =
3800 0 : Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
3801 :
3802 0 : self.new_deltas.push(new_delta_layer);
3803 0 : Ok(())
3804 0 : }
3805 :
3806 0 : async fn delete_layer(
3807 0 : &mut self,
3808 0 : layer: &OwnArc<PersistentLayerDesc>,
3809 0 : _ctx: &RequestContext,
3810 0 : ) -> anyhow::Result<()> {
3811 0 : self.layers_to_delete.push(layer.clone().0);
3812 0 : Ok(())
3813 0 : }
3814 : }
3815 :
3816 : impl TimelineAdaptor {
3817 0 : async fn create_image_impl(
3818 0 : &mut self,
3819 0 : lsn: Lsn,
3820 0 : key_range: &Range<Key>,
3821 0 : ctx: &RequestContext,
3822 0 : ) -> Result<(), CreateImageLayersError> {
3823 0 : let timer = self.timeline.metrics.create_images_time_histo.start_timer();
3824 :
3825 0 : let image_layer_writer = ImageLayerWriter::new(
3826 0 : self.timeline.conf,
3827 0 : self.timeline.timeline_id,
3828 0 : self.timeline.tenant_shard_id,
3829 0 : key_range,
3830 0 : lsn,
3831 0 : ctx,
3832 0 : )
3833 0 : .await?;
3834 :
3835 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
3836 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
3837 0 : "failpoint image-layer-writer-fail-before-finish"
3838 0 : )))
3839 0 : });
3840 :
3841 0 : let keyspace = KeySpace {
3842 0 : ranges: self.get_keyspace(key_range, lsn, ctx).await?,
3843 : };
3844 : // TODO set proper (stateful) start. The create_image_layer_for_rel_blocks function mostly
3845 0 : let outcome = self
3846 0 : .timeline
3847 0 : .create_image_layer_for_rel_blocks(
3848 0 : &keyspace,
3849 0 : image_layer_writer,
3850 0 : lsn,
3851 0 : ctx,
3852 0 : key_range.clone(),
3853 0 : IoConcurrency::sequential(),
3854 0 : )
3855 0 : .await?;
3856 :
3857 : if let ImageLayerCreationOutcome::Generated {
3858 0 : unfinished_image_layer,
3859 0 : } = outcome
3860 : {
3861 0 : let (desc, path) = unfinished_image_layer.finish(ctx).await?;
3862 0 : let image_layer =
3863 0 : Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
3864 0 : self.new_images.push(image_layer);
3865 0 : }
3866 :
3867 0 : timer.stop_and_record();
3868 0 :
3869 0 : Ok(())
3870 0 : }
3871 : }
3872 :
3873 : impl CompactionRequestContext for crate::context::RequestContext {}
3874 :
3875 : #[derive(Debug, Clone)]
3876 : pub struct OwnArc<T>(pub Arc<T>);
3877 :
3878 : impl<T> Deref for OwnArc<T> {
3879 : type Target = <Arc<T> as Deref>::Target;
3880 0 : fn deref(&self) -> &Self::Target {
3881 0 : &self.0
3882 0 : }
3883 : }
3884 :
3885 : impl<T> AsRef<T> for OwnArc<T> {
3886 0 : fn as_ref(&self) -> &T {
3887 0 : self.0.as_ref()
3888 0 : }
3889 : }
3890 :
3891 : impl CompactionLayer<Key> for OwnArc<PersistentLayerDesc> {
3892 0 : fn key_range(&self) -> &Range<Key> {
3893 0 : &self.key_range
3894 0 : }
3895 0 : fn lsn_range(&self) -> &Range<Lsn> {
3896 0 : &self.lsn_range
3897 0 : }
3898 0 : fn file_size(&self) -> u64 {
3899 0 : self.file_size
3900 0 : }
3901 0 : fn short_id(&self) -> std::string::String {
3902 0 : self.as_ref().short_id().to_string()
3903 0 : }
3904 0 : fn is_delta(&self) -> bool {
3905 0 : self.as_ref().is_delta()
3906 0 : }
3907 : }
3908 :
3909 : impl CompactionLayer<Key> for OwnArc<DeltaLayer> {
3910 0 : fn key_range(&self) -> &Range<Key> {
3911 0 : &self.layer_desc().key_range
3912 0 : }
3913 0 : fn lsn_range(&self) -> &Range<Lsn> {
3914 0 : &self.layer_desc().lsn_range
3915 0 : }
3916 0 : fn file_size(&self) -> u64 {
3917 0 : self.layer_desc().file_size
3918 0 : }
3919 0 : fn short_id(&self) -> std::string::String {
3920 0 : self.layer_desc().short_id().to_string()
3921 0 : }
3922 0 : fn is_delta(&self) -> bool {
3923 0 : true
3924 0 : }
3925 : }
3926 :
3927 : impl CompactionLayer<Key> for ResidentDeltaLayer {
3928 0 : fn key_range(&self) -> &Range<Key> {
3929 0 : &self.0.layer_desc().key_range
3930 0 : }
3931 0 : fn lsn_range(&self) -> &Range<Lsn> {
3932 0 : &self.0.layer_desc().lsn_range
3933 0 : }
3934 0 : fn file_size(&self) -> u64 {
3935 0 : self.0.layer_desc().file_size
3936 0 : }
3937 0 : fn short_id(&self) -> std::string::String {
3938 0 : self.0.layer_desc().short_id().to_string()
3939 0 : }
3940 0 : fn is_delta(&self) -> bool {
3941 0 : true
3942 0 : }
3943 : }
3944 :
3945 : impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
3946 : type DeltaEntry<'a> = DeltaEntry<'a>;
3947 :
3948 0 : async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
3949 0 : self.0.get_as_delta(ctx).await?.index_entries(ctx).await
3950 0 : }
3951 : }
3952 :
3953 : impl CompactionLayer<Key> for ResidentImageLayer {
3954 0 : fn key_range(&self) -> &Range<Key> {
3955 0 : &self.0.layer_desc().key_range
3956 0 : }
3957 0 : fn lsn_range(&self) -> &Range<Lsn> {
3958 0 : &self.0.layer_desc().lsn_range
3959 0 : }
3960 0 : fn file_size(&self) -> u64 {
3961 0 : self.0.layer_desc().file_size
3962 0 : }
3963 0 : fn short_id(&self) -> std::string::String {
3964 0 : self.0.layer_desc().short_id().to_string()
3965 0 : }
3966 0 : fn is_delta(&self) -> bool {
3967 0 : false
3968 0 : }
3969 : }
3970 : impl CompactionImageLayer<TimelineAdaptor> for ResidentImageLayer {}
|
