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};
8 : use std::ops::{Deref, Range};
9 : use std::sync::Arc;
10 :
11 : use super::layer_manager::LayerManager;
12 : use super::{
13 : CompactFlags, CreateImageLayersError, DurationRecorder, ImageLayerCreationMode,
14 : RecordedDuration, Timeline,
15 : };
16 :
17 : use anyhow::{anyhow, bail, Context};
18 : use bytes::Bytes;
19 : use enumset::EnumSet;
20 : use fail::fail_point;
21 : use itertools::Itertools;
22 : use pageserver_api::key::KEY_SIZE;
23 : use pageserver_api::keyspace::ShardedRange;
24 : use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
25 : use serde::Serialize;
26 : use tokio_util::sync::CancellationToken;
27 : use tracing::{debug, info, info_span, trace, warn, Instrument};
28 : use utils::id::TimelineId;
29 :
30 : use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
31 : use crate::page_cache;
32 : use crate::statvfs::Statvfs;
33 : use crate::tenant::checks::check_valid_layermap;
34 : use crate::tenant::remote_timeline_client::WaitCompletionError;
35 : use crate::tenant::storage_layer::batch_split_writer::{
36 : BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
37 : };
38 : use crate::tenant::storage_layer::filter_iterator::FilterIterator;
39 : use crate::tenant::storage_layer::merge_iterator::MergeIterator;
40 : use crate::tenant::storage_layer::{
41 : AsLayerDesc, PersistentLayerDesc, PersistentLayerKey, ValueReconstructState,
42 : };
43 : use crate::tenant::timeline::ImageLayerCreationOutcome;
44 : use crate::tenant::timeline::{drop_rlock, DeltaLayerWriter, ImageLayerWriter};
45 : use crate::tenant::timeline::{Layer, ResidentLayer};
46 : use crate::tenant::{DeltaLayer, MaybeOffloaded};
47 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
48 : use pageserver_api::config::tenant_conf_defaults::{
49 : DEFAULT_CHECKPOINT_DISTANCE, DEFAULT_COMPACTION_THRESHOLD,
50 : };
51 :
52 : use pageserver_api::key::Key;
53 : use pageserver_api::keyspace::KeySpace;
54 : use pageserver_api::record::NeonWalRecord;
55 : use pageserver_api::value::Value;
56 :
57 : use utils::lsn::Lsn;
58 :
59 : use pageserver_compaction::helpers::{fully_contains, overlaps_with};
60 : use pageserver_compaction::interface::*;
61 :
62 : use super::CompactionError;
63 :
64 : /// Maximum number of deltas before generating an image layer in bottom-most compaction.
65 : const COMPACTION_DELTA_THRESHOLD: usize = 5;
66 :
67 : pub struct GcCompactionJobDescription {
68 : /// All layers to read in the compaction job
69 : selected_layers: Vec<Layer>,
70 : /// GC cutoff of the job
71 : gc_cutoff: Lsn,
72 : /// LSNs to retain for the job
73 : retain_lsns_below_horizon: Vec<Lsn>,
74 : /// Maximum layer LSN processed in this compaction
75 : max_layer_lsn: Lsn,
76 : /// Only compact layers overlapping with this range
77 : compaction_key_range: Range<Key>,
78 : /// When partial compaction is enabled, these layers need to be rewritten to ensure no overlap.
79 : /// This field is here solely for debugging. The field will not be read once the compaction
80 : /// description is generated.
81 : rewrite_layers: Vec<Arc<PersistentLayerDesc>>,
82 : }
83 :
84 : /// The result of bottom-most compaction for a single key at each LSN.
85 : #[derive(Debug)]
86 : #[cfg_attr(test, derive(PartialEq))]
87 : pub struct KeyLogAtLsn(pub Vec<(Lsn, Value)>);
88 :
89 : /// The result of bottom-most compaction.
90 : #[derive(Debug)]
91 : #[cfg_attr(test, derive(PartialEq))]
92 : pub(crate) struct KeyHistoryRetention {
93 : /// Stores logs to reconstruct the value at the given LSN, that is to say, logs <= LSN or image == LSN.
94 : pub(crate) below_horizon: Vec<(Lsn, KeyLogAtLsn)>,
95 : /// Stores logs to reconstruct the value at any LSN above the horizon, that is to say, log > LSN.
96 : pub(crate) above_horizon: KeyLogAtLsn,
97 : }
98 :
99 : impl KeyHistoryRetention {
100 : /// Hack: skip delta layer if we need to produce a layer of a same key-lsn.
101 : ///
102 : /// This can happen if we have removed some deltas in "the middle" of some existing layer's key-lsn-range.
103 : /// For example, consider the case where a single delta with range [0x10,0x50) exists.
104 : /// And we have branches at LSN 0x10, 0x20, 0x30.
105 : /// Then we delete branch @ 0x20.
106 : /// Bottom-most compaction may now delete the delta [0x20,0x30).
107 : /// And that wouldnt' change the shape of the layer.
108 : ///
109 : /// Note that bottom-most-gc-compaction never _adds_ new data in that case, only removes.
110 : ///
111 : /// `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.
112 58 : async fn discard_key(key: &PersistentLayerKey, tline: &Arc<Timeline>, dry_run: bool) -> bool {
113 58 : if dry_run {
114 0 : return true;
115 58 : }
116 58 : let guard = tline.layers.read().await;
117 58 : if !guard.contains_key(key) {
118 38 : return false;
119 20 : }
120 20 : let layer_generation = guard.get_from_key(key).metadata().generation;
121 20 : drop(guard);
122 20 : if layer_generation == tline.generation {
123 20 : info!(
124 : key=%key,
125 : ?layer_generation,
126 0 : "discard layer due to duplicated layer key in the same generation",
127 : );
128 20 : true
129 : } else {
130 0 : false
131 : }
132 58 : }
133 :
134 : /// Pipe a history of a single key to the writers.
135 : ///
136 : /// If `image_writer` is none, the images will be placed into the delta layers.
137 : /// The delta writer will contain all images and deltas (below and above the horizon) except the bottom-most images.
138 : #[allow(clippy::too_many_arguments)]
139 530 : async fn pipe_to(
140 530 : self,
141 530 : key: Key,
142 530 : delta_writer: &mut SplitDeltaLayerWriter,
143 530 : mut image_writer: Option<&mut SplitImageLayerWriter>,
144 530 : stat: &mut CompactionStatistics,
145 530 : ctx: &RequestContext,
146 530 : ) -> anyhow::Result<()> {
147 530 : let mut first_batch = true;
148 1658 : for (cutoff_lsn, KeyLogAtLsn(logs)) in self.below_horizon {
149 1128 : if first_batch {
150 530 : if logs.len() == 1 && logs[0].1.is_image() {
151 516 : let Value::Image(img) = &logs[0].1 else {
152 0 : unreachable!()
153 : };
154 516 : stat.produce_image_key(img);
155 516 : if let Some(image_writer) = image_writer.as_mut() {
156 516 : image_writer.put_image(key, img.clone(), ctx).await?;
157 : } else {
158 0 : delta_writer
159 0 : .put_value(key, cutoff_lsn, Value::Image(img.clone()), ctx)
160 0 : .await?;
161 : }
162 : } else {
163 28 : for (lsn, val) in logs {
164 14 : stat.produce_key(&val);
165 14 : delta_writer.put_value(key, lsn, val, ctx).await?;
166 : }
167 : }
168 530 : first_batch = false;
169 : } else {
170 684 : for (lsn, val) in logs {
171 86 : stat.produce_key(&val);
172 86 : delta_writer.put_value(key, lsn, val, ctx).await?;
173 : }
174 : }
175 : }
176 530 : let KeyLogAtLsn(above_horizon_logs) = self.above_horizon;
177 568 : for (lsn, val) in above_horizon_logs {
178 38 : stat.produce_key(&val);
179 38 : delta_writer.put_value(key, lsn, val, ctx).await?;
180 : }
181 530 : Ok(())
182 530 : }
183 : }
184 :
185 : #[derive(Debug, Serialize, Default)]
186 : struct CompactionStatisticsNumSize {
187 : num: u64,
188 : size: u64,
189 : }
190 :
191 : #[derive(Debug, Serialize, Default)]
192 : pub struct CompactionStatistics {
193 : delta_layer_visited: CompactionStatisticsNumSize,
194 : image_layer_visited: CompactionStatisticsNumSize,
195 : delta_layer_produced: CompactionStatisticsNumSize,
196 : image_layer_produced: CompactionStatisticsNumSize,
197 : num_delta_layer_discarded: usize,
198 : num_image_layer_discarded: usize,
199 : num_unique_keys_visited: usize,
200 : wal_keys_visited: CompactionStatisticsNumSize,
201 : image_keys_visited: CompactionStatisticsNumSize,
202 : wal_produced: CompactionStatisticsNumSize,
203 : image_produced: CompactionStatisticsNumSize,
204 : }
205 :
206 : impl CompactionStatistics {
207 846 : fn estimated_size_of_value(val: &Value) -> usize {
208 304 : match val {
209 542 : Value::Image(img) => img.len(),
210 0 : Value::WalRecord(NeonWalRecord::Postgres { rec, .. }) => rec.len(),
211 304 : _ => std::mem::size_of::<NeonWalRecord>(),
212 : }
213 846 : }
214 1372 : fn estimated_size_of_key() -> usize {
215 1372 : KEY_SIZE // TODO: distinguish image layer and delta layer (count LSN in delta layer)
216 1372 : }
217 62 : fn visit_delta_layer(&mut self, size: u64) {
218 62 : self.delta_layer_visited.num += 1;
219 62 : self.delta_layer_visited.size += size;
220 62 : }
221 58 : fn visit_image_layer(&mut self, size: u64) {
222 58 : self.image_layer_visited.num += 1;
223 58 : self.image_layer_visited.size += size;
224 58 : }
225 530 : fn on_unique_key_visited(&mut self) {
226 530 : self.num_unique_keys_visited += 1;
227 530 : }
228 176 : fn visit_wal_key(&mut self, val: &Value) {
229 176 : self.wal_keys_visited.num += 1;
230 176 : self.wal_keys_visited.size +=
231 176 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
232 176 : }
233 542 : fn visit_image_key(&mut self, val: &Value) {
234 542 : self.image_keys_visited.num += 1;
235 542 : self.image_keys_visited.size +=
236 542 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
237 542 : }
238 138 : fn produce_key(&mut self, val: &Value) {
239 138 : match val {
240 10 : Value::Image(img) => self.produce_image_key(img),
241 128 : Value::WalRecord(_) => self.produce_wal_key(val),
242 : }
243 138 : }
244 128 : fn produce_wal_key(&mut self, val: &Value) {
245 128 : self.wal_produced.num += 1;
246 128 : self.wal_produced.size +=
247 128 : Self::estimated_size_of_value(val) as u64 + Self::estimated_size_of_key() as u64;
248 128 : }
249 526 : fn produce_image_key(&mut self, val: &Bytes) {
250 526 : self.image_produced.num += 1;
251 526 : self.image_produced.size += val.len() as u64 + Self::estimated_size_of_key() as u64;
252 526 : }
253 12 : fn discard_delta_layer(&mut self) {
254 12 : self.num_delta_layer_discarded += 1;
255 12 : }
256 8 : fn discard_image_layer(&mut self) {
257 8 : self.num_image_layer_discarded += 1;
258 8 : }
259 12 : fn produce_delta_layer(&mut self, size: u64) {
260 12 : self.delta_layer_produced.num += 1;
261 12 : self.delta_layer_produced.size += size;
262 12 : }
263 26 : fn produce_image_layer(&mut self, size: u64) {
264 26 : self.image_layer_produced.num += 1;
265 26 : self.image_layer_produced.size += size;
266 26 : }
267 : }
268 :
269 : impl Timeline {
270 : /// TODO: cancellation
271 : ///
272 : /// Returns whether the compaction has pending tasks.
273 364 : pub(crate) async fn compact_legacy(
274 364 : self: &Arc<Self>,
275 364 : cancel: &CancellationToken,
276 364 : flags: EnumSet<CompactFlags>,
277 364 : ctx: &RequestContext,
278 364 : ) -> Result<bool, CompactionError> {
279 364 : if flags.contains(CompactFlags::EnhancedGcBottomMostCompaction) {
280 0 : self.compact_with_gc(cancel, flags, ctx)
281 0 : .await
282 0 : .map_err(CompactionError::Other)?;
283 0 : return Ok(false);
284 364 : }
285 364 :
286 364 : if flags.contains(CompactFlags::DryRun) {
287 0 : return Err(CompactionError::Other(anyhow!(
288 0 : "dry-run mode is not supported for legacy compaction for now"
289 0 : )));
290 364 : }
291 364 :
292 364 : // High level strategy for compaction / image creation:
293 364 : //
294 364 : // 1. First, calculate the desired "partitioning" of the
295 364 : // currently in-use key space. The goal is to partition the
296 364 : // key space into roughly fixed-size chunks, but also take into
297 364 : // account any existing image layers, and try to align the
298 364 : // chunk boundaries with the existing image layers to avoid
299 364 : // too much churn. Also try to align chunk boundaries with
300 364 : // relation boundaries. In principle, we don't know about
301 364 : // relation boundaries here, we just deal with key-value
302 364 : // pairs, and the code in pgdatadir_mapping.rs knows how to
303 364 : // map relations into key-value pairs. But in practice we know
304 364 : // that 'field6' is the block number, and the fields 1-5
305 364 : // identify a relation. This is just an optimization,
306 364 : // though.
307 364 : //
308 364 : // 2. Once we know the partitioning, for each partition,
309 364 : // decide if it's time to create a new image layer. The
310 364 : // criteria is: there has been too much "churn" since the last
311 364 : // image layer? The "churn" is fuzzy concept, it's a
312 364 : // combination of too many delta files, or too much WAL in
313 364 : // total in the delta file. Or perhaps: if creating an image
314 364 : // file would allow to delete some older files.
315 364 : //
316 364 : // 3. After that, we compact all level0 delta files if there
317 364 : // are too many of them. While compacting, we also garbage
318 364 : // collect any page versions that are no longer needed because
319 364 : // of the new image layers we created in step 2.
320 364 : //
321 364 : // TODO: This high level strategy hasn't been implemented yet.
322 364 : // Below are functions compact_level0() and create_image_layers()
323 364 : // but they are a bit ad hoc and don't quite work like it's explained
324 364 : // above. Rewrite it.
325 364 :
326 364 : // Is the timeline being deleted?
327 364 : if self.is_stopping() {
328 0 : trace!("Dropping out of compaction on timeline shutdown");
329 0 : return Err(CompactionError::ShuttingDown);
330 364 : }
331 364 :
332 364 : let target_file_size = self.get_checkpoint_distance();
333 :
334 : // Define partitioning schema if needed
335 :
336 : // FIXME: the match should only cover repartitioning, not the next steps
337 364 : let (partition_count, has_pending_tasks) = match self
338 364 : .repartition(
339 364 : self.get_last_record_lsn(),
340 364 : self.get_compaction_target_size(),
341 364 : flags,
342 364 : ctx,
343 364 : )
344 15721 : .await
345 : {
346 364 : Ok(((dense_partitioning, sparse_partitioning), lsn)) => {
347 364 : // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
348 364 : let image_ctx = RequestContextBuilder::extend(ctx)
349 364 : .access_stats_behavior(AccessStatsBehavior::Skip)
350 364 : .build();
351 364 :
352 364 : // 2. Compact
353 364 : let timer = self.metrics.compact_time_histo.start_timer();
354 364 : let fully_compacted = self
355 364 : .compact_level0(
356 364 : target_file_size,
357 364 : flags.contains(CompactFlags::ForceL0Compaction),
358 364 : ctx,
359 364 : )
360 9899 : .await?;
361 364 : timer.stop_and_record();
362 364 :
363 364 : let mut partitioning = dense_partitioning;
364 364 : partitioning
365 364 : .parts
366 364 : .extend(sparse_partitioning.into_dense().parts);
367 364 :
368 364 : // 3. Create new image layers for partitions that have been modified
369 364 : // "enough". Skip image layer creation if L0 compaction cannot keep up.
370 364 : if fully_compacted {
371 364 : let image_layers = self
372 364 : .create_image_layers(
373 364 : &partitioning,
374 364 : lsn,
375 364 : if flags.contains(CompactFlags::ForceImageLayerCreation) {
376 14 : ImageLayerCreationMode::Force
377 : } else {
378 350 : ImageLayerCreationMode::Try
379 : },
380 364 : &image_ctx,
381 : )
382 11439 : .await?;
383 :
384 364 : self.upload_new_image_layers(image_layers)?;
385 : } else {
386 0 : info!("skipping image layer generation due to L0 compaction did not include all layers.");
387 : }
388 364 : (partitioning.parts.len(), !fully_compacted)
389 : }
390 0 : Err(err) => {
391 0 : // no partitioning? This is normal, if the timeline was just created
392 0 : // as an empty timeline. Also in unit tests, when we use the timeline
393 0 : // as a simple key-value store, ignoring the datadir layout. Log the
394 0 : // error but continue.
395 0 : //
396 0 : // Suppress error when it's due to cancellation
397 0 : if !self.cancel.is_cancelled() && !err.is_cancelled() {
398 0 : tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
399 0 : }
400 0 : (1, false)
401 : }
402 : };
403 :
404 364 : if self.shard_identity.count >= ShardCount::new(2) {
405 : // Limit the number of layer rewrites to the number of partitions: this means its
406 : // runtime should be comparable to a full round of image layer creations, rather than
407 : // being potentially much longer.
408 0 : let rewrite_max = partition_count;
409 0 :
410 0 : self.compact_shard_ancestors(rewrite_max, ctx).await?;
411 364 : }
412 :
413 364 : Ok(has_pending_tasks)
414 364 : }
415 :
416 : /// Check for layers that are elegible to be rewritten:
417 : /// - Shard splitting: After a shard split, ancestor layers beyond pitr_interval, so that
418 : /// we don't indefinitely retain keys in this shard that aren't needed.
419 : /// - For future use: layers beyond pitr_interval that are in formats we would
420 : /// rather not maintain compatibility with indefinitely.
421 : ///
422 : /// Note: this phase may read and write many gigabytes of data: use rewrite_max to bound
423 : /// how much work it will try to do in each compaction pass.
424 0 : async fn compact_shard_ancestors(
425 0 : self: &Arc<Self>,
426 0 : rewrite_max: usize,
427 0 : ctx: &RequestContext,
428 0 : ) -> Result<(), CompactionError> {
429 0 : let mut drop_layers = Vec::new();
430 0 : let mut layers_to_rewrite: Vec<Layer> = Vec::new();
431 0 :
432 0 : // We will use the Lsn cutoff of the last GC as a threshold for rewriting layers: if a
433 0 : // layer is behind this Lsn, it indicates that the layer is being retained beyond the
434 0 : // pitr_interval, for example because a branchpoint references it.
435 0 : //
436 0 : // Holding this read guard also blocks [`Self::gc_timeline`] from entering while we
437 0 : // are rewriting layers.
438 0 : let latest_gc_cutoff = self.get_latest_gc_cutoff_lsn();
439 0 :
440 0 : tracing::info!(
441 0 : "latest_gc_cutoff: {}, pitr cutoff {}",
442 0 : *latest_gc_cutoff,
443 0 : self.gc_info.read().unwrap().cutoffs.time
444 : );
445 :
446 0 : let layers = self.layers.read().await;
447 0 : for layer_desc in layers.layer_map()?.iter_historic_layers() {
448 0 : let layer = layers.get_from_desc(&layer_desc);
449 0 : if layer.metadata().shard.shard_count == self.shard_identity.count {
450 : // This layer does not belong to a historic ancestor, no need to re-image it.
451 0 : continue;
452 0 : }
453 0 :
454 0 : // This layer was created on an ancestor shard: check if it contains any data for this shard.
455 0 : let sharded_range = ShardedRange::new(layer_desc.get_key_range(), &self.shard_identity);
456 0 : let layer_local_page_count = sharded_range.page_count();
457 0 : let layer_raw_page_count = ShardedRange::raw_size(&layer_desc.get_key_range());
458 0 : if layer_local_page_count == 0 {
459 : // This ancestral layer only covers keys that belong to other shards.
460 : // We include the full metadata in the log: if we had some critical bug that caused
461 : // us to incorrectly drop layers, this would simplify manually debugging + reinstating those layers.
462 0 : info!(%layer, old_metadata=?layer.metadata(),
463 0 : "dropping layer after shard split, contains no keys for this shard.",
464 : );
465 :
466 0 : if cfg!(debug_assertions) {
467 : // Expensive, exhaustive check of keys in this layer: this guards against ShardedRange's calculations being
468 : // wrong. If ShardedRange claims the local page count is zero, then no keys in this layer
469 : // should be !is_key_disposable()
470 0 : let range = layer_desc.get_key_range();
471 0 : let mut key = range.start;
472 0 : while key < range.end {
473 0 : debug_assert!(self.shard_identity.is_key_disposable(&key));
474 0 : key = key.next();
475 : }
476 0 : }
477 :
478 0 : drop_layers.push(layer);
479 0 : continue;
480 0 : } else if layer_local_page_count != u32::MAX
481 0 : && layer_local_page_count == layer_raw_page_count
482 : {
483 0 : debug!(%layer,
484 0 : "layer is entirely shard local ({} keys), no need to filter it",
485 : layer_local_page_count
486 : );
487 0 : continue;
488 0 : }
489 0 :
490 0 : // Don't bother re-writing a layer unless it will at least halve its size
491 0 : if layer_local_page_count != u32::MAX
492 0 : && layer_local_page_count > layer_raw_page_count / 2
493 : {
494 0 : debug!(%layer,
495 0 : "layer is already mostly local ({}/{}), not rewriting",
496 : layer_local_page_count,
497 : layer_raw_page_count
498 : );
499 0 : }
500 :
501 : // Don't bother re-writing a layer if it is within the PITR window: it will age-out eventually
502 : // without incurring the I/O cost of a rewrite.
503 0 : if layer_desc.get_lsn_range().end >= *latest_gc_cutoff {
504 0 : debug!(%layer, "Skipping rewrite of layer still in GC window ({} >= {})",
505 0 : layer_desc.get_lsn_range().end, *latest_gc_cutoff);
506 0 : continue;
507 0 : }
508 0 :
509 0 : if layer_desc.is_delta() {
510 : // We do not yet implement rewrite of delta layers
511 0 : debug!(%layer, "Skipping rewrite of delta layer");
512 0 : continue;
513 0 : }
514 0 :
515 0 : // Only rewrite layers if their generations differ. This guarantees:
516 0 : // - that local rewrite is safe, as local layer paths will differ between existing layer and rewritten one
517 0 : // - that the layer is persistent in remote storage, as we only see old-generation'd layer via loading from remote storage
518 0 : if layer.metadata().generation == self.generation {
519 0 : debug!(%layer, "Skipping rewrite, is not from old generation");
520 0 : continue;
521 0 : }
522 0 :
523 0 : if layers_to_rewrite.len() >= rewrite_max {
524 0 : tracing::info!(%layer, "Will rewrite layer on a future compaction, already rewrote {}",
525 0 : layers_to_rewrite.len()
526 : );
527 0 : continue;
528 0 : }
529 0 :
530 0 : // Fall through: all our conditions for doing a rewrite passed.
531 0 : layers_to_rewrite.push(layer);
532 : }
533 :
534 : // Drop read lock on layer map before we start doing time-consuming I/O
535 0 : drop(layers);
536 0 :
537 0 : let mut replace_image_layers = Vec::new();
538 :
539 0 : for layer in layers_to_rewrite {
540 0 : tracing::info!(layer=%layer, "Rewriting layer after shard split...");
541 0 : let mut image_layer_writer = ImageLayerWriter::new(
542 0 : self.conf,
543 0 : self.timeline_id,
544 0 : self.tenant_shard_id,
545 0 : &layer.layer_desc().key_range,
546 0 : layer.layer_desc().image_layer_lsn(),
547 0 : ctx,
548 0 : )
549 0 : .await
550 0 : .map_err(CompactionError::Other)?;
551 :
552 : // Safety of layer rewrites:
553 : // - We are writing to a different local file path than we are reading from, so the old Layer
554 : // cannot interfere with the new one.
555 : // - In the page cache, contents for a particular VirtualFile are stored with a file_id that
556 : // is different for two layers with the same name (in `ImageLayerInner::new` we always
557 : // acquire a fresh id from [`crate::page_cache::next_file_id`]. So readers do not risk
558 : // reading the index from one layer file, and then data blocks from the rewritten layer file.
559 : // - Any readers that have a reference to the old layer will keep it alive until they are done
560 : // with it. If they are trying to promote from remote storage, that will fail, but this is the same
561 : // as for compaction generally: compaction is allowed to delete layers that readers might be trying to use.
562 : // - We do not run concurrently with other kinds of compaction, so the only layer map writes we race with are:
563 : // - GC, which at worst witnesses us "undelete" a layer that they just deleted.
564 : // - ingestion, which only inserts layers, therefore cannot collide with us.
565 0 : let resident = layer.download_and_keep_resident().await?;
566 :
567 0 : let keys_written = resident
568 0 : .filter(&self.shard_identity, &mut image_layer_writer, ctx)
569 0 : .await?;
570 :
571 0 : if keys_written > 0 {
572 0 : let (desc, path) = image_layer_writer
573 0 : .finish(ctx)
574 0 : .await
575 0 : .map_err(CompactionError::Other)?;
576 0 : let new_layer = Layer::finish_creating(self.conf, self, desc, &path)
577 0 : .map_err(CompactionError::Other)?;
578 0 : tracing::info!(layer=%new_layer, "Rewrote layer, {} -> {} bytes",
579 0 : layer.metadata().file_size,
580 0 : new_layer.metadata().file_size);
581 :
582 0 : replace_image_layers.push((layer, new_layer));
583 0 : } else {
584 0 : // Drop the old layer. Usually for this case we would already have noticed that
585 0 : // the layer has no data for us with the ShardedRange check above, but
586 0 : drop_layers.push(layer);
587 0 : }
588 : }
589 :
590 : // At this point, we have replaced local layer files with their rewritten form, but not yet uploaded
591 : // metadata to reflect that. If we restart here, the replaced layer files will look invalid (size mismatch
592 : // to remote index) and be removed. This is inefficient but safe.
593 0 : fail::fail_point!("compact-shard-ancestors-localonly");
594 0 :
595 0 : // Update the LayerMap so that readers will use the new layers, and enqueue it for writing to remote storage
596 0 : self.rewrite_layers(replace_image_layers, drop_layers)
597 0 : .await?;
598 :
599 0 : fail::fail_point!("compact-shard-ancestors-enqueued");
600 0 :
601 0 : // We wait for all uploads to complete before finishing this compaction stage. This is not
602 0 : // necessary for correctness, but it simplifies testing, and avoids proceeding with another
603 0 : // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
604 0 : // load.
605 0 : match self.remote_client.wait_completion().await {
606 0 : Ok(()) => (),
607 0 : Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
608 : Err(WaitCompletionError::UploadQueueShutDownOrStopped) => {
609 0 : return Err(CompactionError::ShuttingDown)
610 : }
611 : }
612 :
613 0 : fail::fail_point!("compact-shard-ancestors-persistent");
614 0 :
615 0 : Ok(())
616 0 : }
617 :
618 : /// Update the LayerVisibilityHint of layers covered by image layers, based on whether there is
619 : /// an image layer between them and the most recent readable LSN (branch point or tip of timeline). The
620 : /// purpose of the visibility hint is to record which layers need to be available to service reads.
621 : ///
622 : /// The result may be used as an input to eviction and secondary downloads to de-prioritize layers
623 : /// that we know won't be needed for reads.
624 196 : pub(super) async fn update_layer_visibility(
625 196 : &self,
626 196 : ) -> Result<(), super::layer_manager::Shutdown> {
627 196 : let head_lsn = self.get_last_record_lsn();
628 :
629 : // We will sweep through layers in reverse-LSN order. We only do historic layers. L0 deltas
630 : // are implicitly left visible, because LayerVisibilityHint's default is Visible, and we never modify it here.
631 : // Note that L0 deltas _can_ be covered by image layers, but we consider them 'visible' because we anticipate that
632 : // they will be subject to L0->L1 compaction in the near future.
633 196 : let layer_manager = self.layers.read().await;
634 196 : let layer_map = layer_manager.layer_map()?;
635 :
636 196 : let readable_points = {
637 196 : let children = self.gc_info.read().unwrap().retain_lsns.clone();
638 196 :
639 196 : let mut readable_points = Vec::with_capacity(children.len() + 1);
640 196 : for (child_lsn, _child_timeline_id, is_offloaded) in &children {
641 0 : if *is_offloaded == MaybeOffloaded::Yes {
642 0 : continue;
643 0 : }
644 0 : readable_points.push(*child_lsn);
645 : }
646 196 : readable_points.push(head_lsn);
647 196 : readable_points
648 196 : };
649 196 :
650 196 : let (layer_visibility, covered) = layer_map.get_visibility(readable_points);
651 512 : for (layer_desc, visibility) in layer_visibility {
652 316 : // FIXME: a more efficiency bulk zip() through the layers rather than NlogN getting each one
653 316 : let layer = layer_manager.get_from_desc(&layer_desc);
654 316 : layer.set_visibility(visibility);
655 316 : }
656 :
657 : // TODO: publish our covered KeySpace to our parent, so that when they update their visibility, they can
658 : // avoid assuming that everything at a branch point is visible.
659 196 : drop(covered);
660 196 : Ok(())
661 196 : }
662 :
663 : /// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
664 : /// as Level 1 files. Returns whether the L0 layers are fully compacted.
665 364 : async fn compact_level0(
666 364 : self: &Arc<Self>,
667 364 : target_file_size: u64,
668 364 : force_compaction_ignore_threshold: bool,
669 364 : ctx: &RequestContext,
670 364 : ) -> Result<bool, CompactionError> {
671 : let CompactLevel0Phase1Result {
672 364 : new_layers,
673 364 : deltas_to_compact,
674 364 : fully_compacted,
675 : } = {
676 364 : let phase1_span = info_span!("compact_level0_phase1");
677 364 : let ctx = ctx.attached_child();
678 364 : let mut stats = CompactLevel0Phase1StatsBuilder {
679 364 : version: Some(2),
680 364 : tenant_id: Some(self.tenant_shard_id),
681 364 : timeline_id: Some(self.timeline_id),
682 364 : ..Default::default()
683 364 : };
684 364 :
685 364 : let begin = tokio::time::Instant::now();
686 364 : let phase1_layers_locked = self.layers.read().await;
687 364 : let now = tokio::time::Instant::now();
688 364 : stats.read_lock_acquisition_micros =
689 364 : DurationRecorder::Recorded(RecordedDuration(now - begin), now);
690 364 : self.compact_level0_phase1(
691 364 : phase1_layers_locked,
692 364 : stats,
693 364 : target_file_size,
694 364 : force_compaction_ignore_threshold,
695 364 : &ctx,
696 364 : )
697 364 : .instrument(phase1_span)
698 9897 : .await?
699 : };
700 :
701 364 : if new_layers.is_empty() && deltas_to_compact.is_empty() {
702 : // nothing to do
703 336 : return Ok(true);
704 28 : }
705 28 :
706 28 : self.finish_compact_batch(&new_layers, &Vec::new(), &deltas_to_compact)
707 1 : .await?;
708 28 : Ok(fully_compacted)
709 364 : }
710 :
711 : /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
712 364 : async fn compact_level0_phase1<'a>(
713 364 : self: &'a Arc<Self>,
714 364 : guard: tokio::sync::RwLockReadGuard<'a, LayerManager>,
715 364 : mut stats: CompactLevel0Phase1StatsBuilder,
716 364 : target_file_size: u64,
717 364 : force_compaction_ignore_threshold: bool,
718 364 : ctx: &RequestContext,
719 364 : ) -> Result<CompactLevel0Phase1Result, CompactionError> {
720 364 : stats.read_lock_held_spawn_blocking_startup_micros =
721 364 : stats.read_lock_acquisition_micros.till_now(); // set by caller
722 364 : let layers = guard.layer_map()?;
723 364 : let level0_deltas = layers.level0_deltas();
724 364 : stats.level0_deltas_count = Some(level0_deltas.len());
725 364 :
726 364 : // Only compact if enough layers have accumulated.
727 364 : let threshold = self.get_compaction_threshold();
728 364 : if level0_deltas.is_empty() || level0_deltas.len() < threshold {
729 336 : if force_compaction_ignore_threshold {
730 0 : if !level0_deltas.is_empty() {
731 0 : info!(
732 0 : level0_deltas = level0_deltas.len(),
733 0 : threshold, "too few deltas to compact, but forcing compaction"
734 : );
735 : } else {
736 0 : info!(
737 0 : level0_deltas = level0_deltas.len(),
738 0 : threshold, "too few deltas to compact, cannot force compaction"
739 : );
740 0 : return Ok(CompactLevel0Phase1Result::default());
741 : }
742 : } else {
743 336 : debug!(
744 0 : level0_deltas = level0_deltas.len(),
745 0 : threshold, "too few deltas to compact"
746 : );
747 336 : return Ok(CompactLevel0Phase1Result::default());
748 : }
749 28 : }
750 :
751 28 : let mut level0_deltas = level0_deltas
752 28 : .iter()
753 402 : .map(|x| guard.get_from_desc(x))
754 28 : .collect::<Vec<_>>();
755 28 :
756 28 : // Gather the files to compact in this iteration.
757 28 : //
758 28 : // Start with the oldest Level 0 delta file, and collect any other
759 28 : // level 0 files that form a contiguous sequence, such that the end
760 28 : // LSN of previous file matches the start LSN of the next file.
761 28 : //
762 28 : // Note that if the files don't form such a sequence, we might
763 28 : // "compact" just a single file. That's a bit pointless, but it allows
764 28 : // us to get rid of the level 0 file, and compact the other files on
765 28 : // the next iteration. This could probably made smarter, but such
766 28 : // "gaps" in the sequence of level 0 files should only happen in case
767 28 : // of a crash, partial download from cloud storage, or something like
768 28 : // that, so it's not a big deal in practice.
769 748 : level0_deltas.sort_by_key(|l| l.layer_desc().lsn_range.start);
770 28 : let mut level0_deltas_iter = level0_deltas.iter();
771 28 :
772 28 : let first_level0_delta = level0_deltas_iter.next().unwrap();
773 28 : let mut prev_lsn_end = first_level0_delta.layer_desc().lsn_range.end;
774 28 : let mut deltas_to_compact = Vec::with_capacity(level0_deltas.len());
775 28 :
776 28 : // Accumulate the size of layers in `deltas_to_compact`
777 28 : let mut deltas_to_compact_bytes = 0;
778 28 :
779 28 : // Under normal circumstances, we will accumulate up to compaction_interval L0s of size
780 28 : // checkpoint_distance each. To avoid edge cases using extra system resources, bound our
781 28 : // work in this function to only operate on this much delta data at once.
782 28 : //
783 28 : // Take the max of the configured value & the default, so that tests that configure tiny values
784 28 : // can still use a sensible amount of memory, but if a deployed system configures bigger values we
785 28 : // still let them compact a full stack of L0s in one go.
786 28 : let delta_size_limit = std::cmp::max(
787 28 : self.get_compaction_threshold(),
788 28 : DEFAULT_COMPACTION_THRESHOLD,
789 28 : ) as u64
790 28 : * std::cmp::max(self.get_checkpoint_distance(), DEFAULT_CHECKPOINT_DISTANCE);
791 28 :
792 28 : let mut fully_compacted = true;
793 28 :
794 28 : deltas_to_compact.push(first_level0_delta.download_and_keep_resident().await?);
795 402 : for l in level0_deltas_iter {
796 374 : let lsn_range = &l.layer_desc().lsn_range;
797 374 :
798 374 : if lsn_range.start != prev_lsn_end {
799 0 : break;
800 374 : }
801 374 : deltas_to_compact.push(l.download_and_keep_resident().await?);
802 374 : deltas_to_compact_bytes += l.metadata().file_size;
803 374 : prev_lsn_end = lsn_range.end;
804 374 :
805 374 : if deltas_to_compact_bytes >= delta_size_limit {
806 0 : info!(
807 0 : l0_deltas_selected = deltas_to_compact.len(),
808 0 : l0_deltas_total = level0_deltas.len(),
809 0 : "L0 compaction picker hit max delta layer size limit: {}",
810 : delta_size_limit
811 : );
812 0 : fully_compacted = false;
813 0 :
814 0 : // Proceed with compaction, but only a subset of L0s
815 0 : break;
816 374 : }
817 : }
818 28 : let lsn_range = Range {
819 28 : start: deltas_to_compact
820 28 : .first()
821 28 : .unwrap()
822 28 : .layer_desc()
823 28 : .lsn_range
824 28 : .start,
825 28 : end: deltas_to_compact.last().unwrap().layer_desc().lsn_range.end,
826 28 : };
827 28 :
828 28 : info!(
829 0 : "Starting Level0 compaction in LSN range {}-{} for {} layers ({} deltas in total)",
830 0 : lsn_range.start,
831 0 : lsn_range.end,
832 0 : deltas_to_compact.len(),
833 0 : level0_deltas.len()
834 : );
835 :
836 402 : for l in deltas_to_compact.iter() {
837 402 : info!("compact includes {l}");
838 : }
839 :
840 : // We don't need the original list of layers anymore. Drop it so that
841 : // we don't accidentally use it later in the function.
842 28 : drop(level0_deltas);
843 28 :
844 28 : stats.read_lock_held_prerequisites_micros = stats
845 28 : .read_lock_held_spawn_blocking_startup_micros
846 28 : .till_now();
847 :
848 : // TODO: replace with streaming k-merge
849 28 : let all_keys = {
850 28 : let mut all_keys = Vec::new();
851 402 : for l in deltas_to_compact.iter() {
852 402 : if self.cancel.is_cancelled() {
853 0 : return Err(CompactionError::ShuttingDown);
854 402 : }
855 402 : let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
856 402 : let keys = delta
857 402 : .index_entries(ctx)
858 2124 : .await
859 402 : .map_err(CompactionError::Other)?;
860 402 : all_keys.extend(keys);
861 : }
862 : // The current stdlib sorting implementation is designed in a way where it is
863 : // particularly fast where the slice is made up of sorted sub-ranges.
864 4423784 : all_keys.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
865 28 : all_keys
866 28 : };
867 28 :
868 28 : stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
869 :
870 : // Determine N largest holes where N is number of compacted layers. The vec is sorted by key range start.
871 : //
872 : // A hole is a key range for which this compaction doesn't have any WAL records.
873 : // Our goal in this compaction iteration is to avoid creating L1s that, in terms of their key range,
874 : // cover the hole, but actually don't contain any WAL records for that key range.
875 : // The reason is that the mere stack of L1s (`count_deltas`) triggers image layer creation (`create_image_layers`).
876 : // That image layer creation would be useless for a hole range covered by L1s that don't contain any WAL records.
877 : //
878 : // The algorithm chooses holes as follows.
879 : // - Slide a 2-window over the keys in key orde to get the hole range (=distance between two keys).
880 : // - Filter: min threshold on range length
881 : // - Rank: by coverage size (=number of image layers required to reconstruct each key in the range for which we have any data)
882 : //
883 : // For more details, intuition, and some ASCII art see https://github.com/neondatabase/neon/pull/3597#discussion_r1112704451
884 : #[derive(PartialEq, Eq)]
885 : struct Hole {
886 : key_range: Range<Key>,
887 : coverage_size: usize,
888 : }
889 28 : let holes: Vec<Hole> = {
890 : use std::cmp::Ordering;
891 : impl Ord for Hole {
892 0 : fn cmp(&self, other: &Self) -> Ordering {
893 0 : self.coverage_size.cmp(&other.coverage_size).reverse()
894 0 : }
895 : }
896 : impl PartialOrd for Hole {
897 0 : fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
898 0 : Some(self.cmp(other))
899 0 : }
900 : }
901 28 : let max_holes = deltas_to_compact.len();
902 28 : let last_record_lsn = self.get_last_record_lsn();
903 28 : let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
904 28 : let min_hole_coverage_size = 3; // TODO: something more flexible?
905 28 : // min-heap (reserve space for one more element added before eviction)
906 28 : let mut heap: BinaryHeap<Hole> = BinaryHeap::with_capacity(max_holes + 1);
907 28 : let mut prev: Option<Key> = None;
908 :
909 2064038 : for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
910 2064038 : if let Some(prev_key) = prev {
911 : // just first fast filter, do not create hole entries for metadata keys. The last hole in the
912 : // compaction is the gap between data key and metadata keys.
913 2064010 : if next_key.to_i128() - prev_key.to_i128() >= min_hole_range
914 0 : && !Key::is_metadata_key(&prev_key)
915 : {
916 0 : let key_range = prev_key..next_key;
917 0 : // Measuring hole by just subtraction of i128 representation of key range boundaries
918 0 : // has not so much sense, because largest holes will corresponds field1/field2 changes.
919 0 : // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
920 0 : // That is why it is better to measure size of hole as number of covering image layers.
921 0 : let coverage_size =
922 0 : layers.image_coverage(&key_range, last_record_lsn).len();
923 0 : if coverage_size >= min_hole_coverage_size {
924 0 : heap.push(Hole {
925 0 : key_range,
926 0 : coverage_size,
927 0 : });
928 0 : if heap.len() > max_holes {
929 0 : heap.pop(); // remove smallest hole
930 0 : }
931 0 : }
932 2064010 : }
933 28 : }
934 2064038 : prev = Some(next_key.next());
935 : }
936 28 : let mut holes = heap.into_vec();
937 28 : holes.sort_unstable_by_key(|hole| hole.key_range.start);
938 28 : holes
939 28 : };
940 28 : stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
941 28 : drop_rlock(guard);
942 28 :
943 28 : if self.cancel.is_cancelled() {
944 0 : return Err(CompactionError::ShuttingDown);
945 28 : }
946 28 :
947 28 : stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
948 :
949 : // This iterator walks through all key-value pairs from all the layers
950 : // we're compacting, in key, LSN order.
951 : // If there's both a Value::Image and Value::WalRecord for the same (key,lsn),
952 : // then the Value::Image is ordered before Value::WalRecord.
953 28 : let mut all_values_iter = {
954 28 : let mut deltas = Vec::with_capacity(deltas_to_compact.len());
955 402 : for l in deltas_to_compact.iter() {
956 402 : let l = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
957 402 : deltas.push(l);
958 : }
959 28 : MergeIterator::create(&deltas, &[], ctx)
960 28 : };
961 28 :
962 28 : // This iterator walks through all keys and is needed to calculate size used by each key
963 28 : let mut all_keys_iter = all_keys
964 28 : .iter()
965 2064038 : .map(|DeltaEntry { key, lsn, size, .. }| (*key, *lsn, *size))
966 2064010 : .coalesce(|mut prev, cur| {
967 2064010 : // Coalesce keys that belong to the same key pair.
968 2064010 : // This ensures that compaction doesn't put them
969 2064010 : // into different layer files.
970 2064010 : // Still limit this by the target file size,
971 2064010 : // so that we keep the size of the files in
972 2064010 : // check.
973 2064010 : if prev.0 == cur.0 && prev.2 < target_file_size {
974 40038 : prev.2 += cur.2;
975 40038 : Ok(prev)
976 : } else {
977 2023972 : Err((prev, cur))
978 : }
979 2064010 : });
980 28 :
981 28 : // Merge the contents of all the input delta layers into a new set
982 28 : // of delta layers, based on the current partitioning.
983 28 : //
984 28 : // 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.
985 28 : // It's possible that there is a single key with so many page versions that storing all of them in a single layer file
986 28 : // would be too large. In that case, we also split on the LSN dimension.
987 28 : //
988 28 : // LSN
989 28 : // ^
990 28 : // |
991 28 : // | +-----------+ +--+--+--+--+
992 28 : // | | | | | | | |
993 28 : // | +-----------+ | | | | |
994 28 : // | | | | | | | |
995 28 : // | +-----------+ ==> | | | | |
996 28 : // | | | | | | | |
997 28 : // | +-----------+ | | | | |
998 28 : // | | | | | | | |
999 28 : // | +-----------+ +--+--+--+--+
1000 28 : // |
1001 28 : // +--------------> key
1002 28 : //
1003 28 : //
1004 28 : // If one key (X) has a lot of page versions:
1005 28 : //
1006 28 : // LSN
1007 28 : // ^
1008 28 : // | (X)
1009 28 : // | +-----------+ +--+--+--+--+
1010 28 : // | | | | | | | |
1011 28 : // | +-----------+ | | +--+ |
1012 28 : // | | | | | | | |
1013 28 : // | +-----------+ ==> | | | | |
1014 28 : // | | | | | +--+ |
1015 28 : // | +-----------+ | | | | |
1016 28 : // | | | | | | | |
1017 28 : // | +-----------+ +--+--+--+--+
1018 28 : // |
1019 28 : // +--------------> key
1020 28 : // TODO: this actually divides the layers into fixed-size chunks, not
1021 28 : // based on the partitioning.
1022 28 : //
1023 28 : // TODO: we should also opportunistically materialize and
1024 28 : // garbage collect what we can.
1025 28 : let mut new_layers = Vec::new();
1026 28 : let mut prev_key: Option<Key> = None;
1027 28 : let mut writer: Option<DeltaLayerWriter> = None;
1028 28 : let mut key_values_total_size = 0u64;
1029 28 : let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
1030 28 : let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
1031 28 : let mut next_hole = 0; // index of next hole in holes vector
1032 28 :
1033 28 : let mut keys = 0;
1034 :
1035 2064066 : while let Some((key, lsn, value)) = all_values_iter
1036 2064066 : .next()
1037 3425 : .await
1038 2064066 : .map_err(CompactionError::Other)?
1039 : {
1040 2064038 : keys += 1;
1041 2064038 :
1042 2064038 : if keys % 32_768 == 0 && self.cancel.is_cancelled() {
1043 : // avoid hitting the cancellation token on every key. in benches, we end up
1044 : // shuffling an order of million keys per layer, this means we'll check it
1045 : // around tens of times per layer.
1046 0 : return Err(CompactionError::ShuttingDown);
1047 2064038 : }
1048 2064038 :
1049 2064038 : let same_key = prev_key.map_or(false, |prev_key| prev_key == key);
1050 2064038 : // We need to check key boundaries once we reach next key or end of layer with the same key
1051 2064038 : if !same_key || lsn == dup_end_lsn {
1052 2024000 : let mut next_key_size = 0u64;
1053 2024000 : let is_dup_layer = dup_end_lsn.is_valid();
1054 2024000 : dup_start_lsn = Lsn::INVALID;
1055 2024000 : if !same_key {
1056 2024000 : dup_end_lsn = Lsn::INVALID;
1057 2024000 : }
1058 : // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
1059 2024000 : for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
1060 2024000 : next_key_size = next_size;
1061 2024000 : if key != next_key {
1062 2023972 : if dup_end_lsn.is_valid() {
1063 0 : // We are writting segment with duplicates:
1064 0 : // place all remaining values of this key in separate segment
1065 0 : dup_start_lsn = dup_end_lsn; // new segments starts where old stops
1066 0 : dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
1067 2023972 : }
1068 2023972 : break;
1069 28 : }
1070 28 : key_values_total_size += next_size;
1071 28 : // Check if it is time to split segment: if total keys size is larger than target file size.
1072 28 : // We need to avoid generation of empty segments if next_size > target_file_size.
1073 28 : if key_values_total_size > target_file_size && lsn != next_lsn {
1074 : // Split key between multiple layers: such layer can contain only single key
1075 0 : dup_start_lsn = if dup_end_lsn.is_valid() {
1076 0 : dup_end_lsn // new segment with duplicates starts where old one stops
1077 : } else {
1078 0 : lsn // start with the first LSN for this key
1079 : };
1080 0 : dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
1081 0 : break;
1082 28 : }
1083 : }
1084 : // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
1085 2024000 : if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
1086 0 : dup_start_lsn = dup_end_lsn;
1087 0 : dup_end_lsn = lsn_range.end;
1088 2024000 : }
1089 2024000 : if writer.is_some() {
1090 2023972 : let written_size = writer.as_mut().unwrap().size();
1091 2023972 : let contains_hole =
1092 2023972 : next_hole < holes.len() && key >= holes[next_hole].key_range.end;
1093 : // check if key cause layer overflow or contains hole...
1094 2023972 : if is_dup_layer
1095 2023972 : || dup_end_lsn.is_valid()
1096 2023972 : || written_size + key_values_total_size > target_file_size
1097 2023692 : || contains_hole
1098 : {
1099 : // ... if so, flush previous layer and prepare to write new one
1100 280 : let (desc, path) = writer
1101 280 : .take()
1102 280 : .unwrap()
1103 280 : .finish(prev_key.unwrap().next(), ctx)
1104 711 : .await
1105 280 : .map_err(CompactionError::Other)?;
1106 280 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1107 280 : .map_err(CompactionError::Other)?;
1108 :
1109 280 : new_layers.push(new_delta);
1110 280 : writer = None;
1111 280 :
1112 280 : if contains_hole {
1113 0 : // skip hole
1114 0 : next_hole += 1;
1115 280 : }
1116 2023692 : }
1117 28 : }
1118 : // Remember size of key value because at next iteration we will access next item
1119 2024000 : key_values_total_size = next_key_size;
1120 40038 : }
1121 2064038 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
1122 0 : Err(CompactionError::Other(anyhow::anyhow!(
1123 0 : "failpoint delta-layer-writer-fail-before-finish"
1124 0 : )))
1125 2064038 : });
1126 :
1127 2064038 : if !self.shard_identity.is_key_disposable(&key) {
1128 2064038 : if writer.is_none() {
1129 308 : if self.cancel.is_cancelled() {
1130 : // to be somewhat responsive to cancellation, check for each new layer
1131 0 : return Err(CompactionError::ShuttingDown);
1132 308 : }
1133 : // Create writer if not initiaized yet
1134 308 : writer = Some(
1135 : DeltaLayerWriter::new(
1136 308 : self.conf,
1137 308 : self.timeline_id,
1138 308 : self.tenant_shard_id,
1139 308 : key,
1140 308 : if dup_end_lsn.is_valid() {
1141 : // this is a layer containing slice of values of the same key
1142 0 : debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
1143 0 : dup_start_lsn..dup_end_lsn
1144 : } else {
1145 308 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
1146 308 : lsn_range.clone()
1147 : },
1148 308 : ctx,
1149 : )
1150 154 : .await
1151 308 : .map_err(CompactionError::Other)?,
1152 : );
1153 :
1154 308 : keys = 0;
1155 2063730 : }
1156 :
1157 2064038 : writer
1158 2064038 : .as_mut()
1159 2064038 : .unwrap()
1160 2064038 : .put_value(key, lsn, value, ctx)
1161 1225 : .await
1162 2064038 : .map_err(CompactionError::Other)?;
1163 : } else {
1164 0 : debug!(
1165 0 : "Dropping key {} during compaction (it belongs on shard {:?})",
1166 0 : key,
1167 0 : self.shard_identity.get_shard_number(&key)
1168 : );
1169 : }
1170 :
1171 2064038 : if !new_layers.is_empty() {
1172 19786 : fail_point!("after-timeline-compacted-first-L1");
1173 2044252 : }
1174 :
1175 2064038 : prev_key = Some(key);
1176 : }
1177 28 : if let Some(writer) = writer {
1178 28 : let (desc, path) = writer
1179 28 : .finish(prev_key.unwrap().next(), ctx)
1180 1988 : .await
1181 28 : .map_err(CompactionError::Other)?;
1182 28 : let new_delta = Layer::finish_creating(self.conf, self, desc, &path)
1183 28 : .map_err(CompactionError::Other)?;
1184 28 : new_layers.push(new_delta);
1185 0 : }
1186 :
1187 : // Sync layers
1188 28 : if !new_layers.is_empty() {
1189 : // Print a warning if the created layer is larger than double the target size
1190 : // Add two pages for potential overhead. This should in theory be already
1191 : // accounted for in the target calculation, but for very small targets,
1192 : // we still might easily hit the limit otherwise.
1193 28 : let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
1194 308 : for layer in new_layers.iter() {
1195 308 : if layer.layer_desc().file_size > warn_limit {
1196 0 : warn!(
1197 : %layer,
1198 0 : "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
1199 : );
1200 308 : }
1201 : }
1202 :
1203 : // The writer.finish() above already did the fsync of the inodes.
1204 : // We just need to fsync the directory in which these inodes are linked,
1205 : // which we know to be the timeline directory.
1206 : //
1207 : // We use fatal_err() below because the after writer.finish() returns with success,
1208 : // the in-memory state of the filesystem already has the layer file in its final place,
1209 : // and subsequent pageserver code could think it's durable while it really isn't.
1210 28 : let timeline_dir = VirtualFile::open(
1211 28 : &self
1212 28 : .conf
1213 28 : .timeline_path(&self.tenant_shard_id, &self.timeline_id),
1214 28 : ctx,
1215 28 : )
1216 14 : .await
1217 28 : .fatal_err("VirtualFile::open for timeline dir fsync");
1218 28 : timeline_dir
1219 28 : .sync_all()
1220 14 : .await
1221 28 : .fatal_err("VirtualFile::sync_all timeline dir");
1222 0 : }
1223 :
1224 28 : stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
1225 28 : stats.new_deltas_count = Some(new_layers.len());
1226 308 : stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
1227 28 :
1228 28 : match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
1229 28 : .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
1230 : {
1231 28 : Ok(stats_json) => {
1232 28 : info!(
1233 0 : stats_json = stats_json.as_str(),
1234 0 : "compact_level0_phase1 stats available"
1235 : )
1236 : }
1237 0 : Err(e) => {
1238 0 : warn!("compact_level0_phase1 stats failed to serialize: {:#}", e);
1239 : }
1240 : }
1241 :
1242 : // Without this, rustc complains about deltas_to_compact still
1243 : // being borrowed when we `.into_iter()` below.
1244 28 : drop(all_values_iter);
1245 28 :
1246 28 : Ok(CompactLevel0Phase1Result {
1247 28 : new_layers,
1248 28 : deltas_to_compact: deltas_to_compact
1249 28 : .into_iter()
1250 402 : .map(|x| x.drop_eviction_guard())
1251 28 : .collect::<Vec<_>>(),
1252 28 : fully_compacted,
1253 28 : })
1254 364 : }
1255 : }
1256 :
1257 : #[derive(Default)]
1258 : struct CompactLevel0Phase1Result {
1259 : new_layers: Vec<ResidentLayer>,
1260 : deltas_to_compact: Vec<Layer>,
1261 : // Whether we have included all L0 layers, or selected only part of them due to the
1262 : // L0 compaction size limit.
1263 : fully_compacted: bool,
1264 : }
1265 :
1266 : #[derive(Default)]
1267 : struct CompactLevel0Phase1StatsBuilder {
1268 : version: Option<u64>,
1269 : tenant_id: Option<TenantShardId>,
1270 : timeline_id: Option<TimelineId>,
1271 : read_lock_acquisition_micros: DurationRecorder,
1272 : read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
1273 : read_lock_held_key_sort_micros: DurationRecorder,
1274 : read_lock_held_prerequisites_micros: DurationRecorder,
1275 : read_lock_held_compute_holes_micros: DurationRecorder,
1276 : read_lock_drop_micros: DurationRecorder,
1277 : write_layer_files_micros: DurationRecorder,
1278 : level0_deltas_count: Option<usize>,
1279 : new_deltas_count: Option<usize>,
1280 : new_deltas_size: Option<u64>,
1281 : }
1282 :
1283 : #[derive(serde::Serialize)]
1284 : struct CompactLevel0Phase1Stats {
1285 : version: u64,
1286 : tenant_id: TenantShardId,
1287 : timeline_id: TimelineId,
1288 : read_lock_acquisition_micros: RecordedDuration,
1289 : read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
1290 : read_lock_held_key_sort_micros: RecordedDuration,
1291 : read_lock_held_prerequisites_micros: RecordedDuration,
1292 : read_lock_held_compute_holes_micros: RecordedDuration,
1293 : read_lock_drop_micros: RecordedDuration,
1294 : write_layer_files_micros: RecordedDuration,
1295 : level0_deltas_count: usize,
1296 : new_deltas_count: usize,
1297 : new_deltas_size: u64,
1298 : }
1299 :
1300 : impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
1301 : type Error = anyhow::Error;
1302 :
1303 28 : fn try_from(value: CompactLevel0Phase1StatsBuilder) -> Result<Self, Self::Error> {
1304 28 : Ok(Self {
1305 28 : version: value.version.ok_or_else(|| anyhow!("version not set"))?,
1306 28 : tenant_id: value
1307 28 : .tenant_id
1308 28 : .ok_or_else(|| anyhow!("tenant_id not set"))?,
1309 28 : timeline_id: value
1310 28 : .timeline_id
1311 28 : .ok_or_else(|| anyhow!("timeline_id not set"))?,
1312 28 : read_lock_acquisition_micros: value
1313 28 : .read_lock_acquisition_micros
1314 28 : .into_recorded()
1315 28 : .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
1316 28 : read_lock_held_spawn_blocking_startup_micros: value
1317 28 : .read_lock_held_spawn_blocking_startup_micros
1318 28 : .into_recorded()
1319 28 : .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
1320 28 : read_lock_held_key_sort_micros: value
1321 28 : .read_lock_held_key_sort_micros
1322 28 : .into_recorded()
1323 28 : .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
1324 28 : read_lock_held_prerequisites_micros: value
1325 28 : .read_lock_held_prerequisites_micros
1326 28 : .into_recorded()
1327 28 : .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
1328 28 : read_lock_held_compute_holes_micros: value
1329 28 : .read_lock_held_compute_holes_micros
1330 28 : .into_recorded()
1331 28 : .ok_or_else(|| anyhow!("read_lock_held_compute_holes_micros not set"))?,
1332 28 : read_lock_drop_micros: value
1333 28 : .read_lock_drop_micros
1334 28 : .into_recorded()
1335 28 : .ok_or_else(|| anyhow!("read_lock_drop_micros not set"))?,
1336 28 : write_layer_files_micros: value
1337 28 : .write_layer_files_micros
1338 28 : .into_recorded()
1339 28 : .ok_or_else(|| anyhow!("write_layer_files_micros not set"))?,
1340 28 : level0_deltas_count: value
1341 28 : .level0_deltas_count
1342 28 : .ok_or_else(|| anyhow!("level0_deltas_count not set"))?,
1343 28 : new_deltas_count: value
1344 28 : .new_deltas_count
1345 28 : .ok_or_else(|| anyhow!("new_deltas_count not set"))?,
1346 28 : new_deltas_size: value
1347 28 : .new_deltas_size
1348 28 : .ok_or_else(|| anyhow!("new_deltas_size not set"))?,
1349 : })
1350 28 : }
1351 : }
1352 :
1353 : impl Timeline {
1354 : /// Entry point for new tiered compaction algorithm.
1355 : ///
1356 : /// All the real work is in the implementation in the pageserver_compaction
1357 : /// crate. The code here would apply to any algorithm implemented by the
1358 : /// same interface, but tiered is the only one at the moment.
1359 : ///
1360 : /// TODO: cancellation
1361 0 : pub(crate) async fn compact_tiered(
1362 0 : self: &Arc<Self>,
1363 0 : _cancel: &CancellationToken,
1364 0 : ctx: &RequestContext,
1365 0 : ) -> Result<(), CompactionError> {
1366 0 : let fanout = self.get_compaction_threshold() as u64;
1367 0 : let target_file_size = self.get_checkpoint_distance();
1368 :
1369 : // Find the top of the historical layers
1370 0 : let end_lsn = {
1371 0 : let guard = self.layers.read().await;
1372 0 : let layers = guard.layer_map()?;
1373 :
1374 0 : let l0_deltas = layers.level0_deltas();
1375 0 :
1376 0 : // As an optimization, if we find that there are too few L0 layers,
1377 0 : // bail out early. We know that the compaction algorithm would do
1378 0 : // nothing in that case.
1379 0 : if l0_deltas.len() < fanout as usize {
1380 : // doesn't need compacting
1381 0 : return Ok(());
1382 0 : }
1383 0 : l0_deltas.iter().map(|l| l.lsn_range.end).max().unwrap()
1384 0 : };
1385 0 :
1386 0 : // Is the timeline being deleted?
1387 0 : if self.is_stopping() {
1388 0 : trace!("Dropping out of compaction on timeline shutdown");
1389 0 : return Err(CompactionError::ShuttingDown);
1390 0 : }
1391 :
1392 0 : let (dense_ks, _sparse_ks) = self.collect_keyspace(end_lsn, ctx).await?;
1393 : // TODO(chi): ignore sparse_keyspace for now, compact it in the future.
1394 0 : let mut adaptor = TimelineAdaptor::new(self, (end_lsn, dense_ks));
1395 0 :
1396 0 : pageserver_compaction::compact_tiered::compact_tiered(
1397 0 : &mut adaptor,
1398 0 : end_lsn,
1399 0 : target_file_size,
1400 0 : fanout,
1401 0 : ctx,
1402 0 : )
1403 0 : .await
1404 : // TODO: compact_tiered needs to return CompactionError
1405 0 : .map_err(CompactionError::Other)?;
1406 :
1407 0 : adaptor.flush_updates().await?;
1408 0 : Ok(())
1409 0 : }
1410 :
1411 : /// Take a list of images and deltas, produce images and deltas according to GC horizon and retain_lsns.
1412 : ///
1413 : /// It takes a key, the values of the key within the compaction process, a GC horizon, and all retain_lsns below the horizon.
1414 : /// For now, it requires the `accumulated_values` contains the full history of the key (i.e., the key with the lowest LSN is
1415 : /// an image or a WAL not requiring a base image). This restriction will be removed once we implement gc-compaction on branch.
1416 : ///
1417 : /// The function returns the deltas and the base image that need to be placed at each of the retain LSN. For example, we have:
1418 : ///
1419 : /// A@0x10, +B@0x20, +C@0x30, +D@0x40, +E@0x50, +F@0x60
1420 : /// horizon = 0x50, retain_lsn = 0x20, 0x40, delta_threshold=3
1421 : ///
1422 : /// The function will produce:
1423 : ///
1424 : /// ```plain
1425 : /// 0x20(retain_lsn) -> img=AB@0x20 always produce a single image below the lowest retain LSN
1426 : /// 0x40(retain_lsn) -> deltas=[+C@0x30, +D@0x40] two deltas since the last base image, keeping the deltas
1427 : /// 0x50(horizon) -> deltas=[ABCDE@0x50] three deltas since the last base image, generate an image but put it in the delta
1428 : /// above_horizon -> deltas=[+F@0x60] full history above the horizon
1429 : /// ```
1430 : ///
1431 : /// Note that `accumulated_values` must be sorted by LSN and should belong to a single key.
1432 538 : pub(crate) async fn generate_key_retention(
1433 538 : self: &Arc<Timeline>,
1434 538 : key: Key,
1435 538 : full_history: &[(Key, Lsn, Value)],
1436 538 : horizon: Lsn,
1437 538 : retain_lsn_below_horizon: &[Lsn],
1438 538 : delta_threshold_cnt: usize,
1439 538 : base_img_from_ancestor: Option<(Key, Lsn, Bytes)>,
1440 538 : ) -> anyhow::Result<KeyHistoryRetention> {
1441 538 : // Pre-checks for the invariants
1442 538 : if cfg!(debug_assertions) {
1443 1306 : for (log_key, _, _) in full_history {
1444 768 : assert_eq!(log_key, &key, "mismatched key");
1445 : }
1446 538 : for i in 1..full_history.len() {
1447 230 : assert!(full_history[i - 1].1 <= full_history[i].1, "unordered LSN");
1448 230 : if full_history[i - 1].1 == full_history[i].1 {
1449 0 : assert!(
1450 0 : matches!(full_history[i - 1].2, Value::Image(_)),
1451 0 : "unordered delta/image, or duplicated delta"
1452 : );
1453 230 : }
1454 : }
1455 : // There was an assertion for no base image that checks if the first
1456 : // record in the history is `will_init` before, but it was removed.
1457 : // This is explained in the test cases for generate_key_retention.
1458 : // Search "incomplete history" for more information.
1459 1148 : for lsn in retain_lsn_below_horizon {
1460 610 : assert!(lsn < &horizon, "retain lsn must be below horizon")
1461 : }
1462 538 : for i in 1..retain_lsn_below_horizon.len() {
1463 278 : assert!(
1464 278 : retain_lsn_below_horizon[i - 1] <= retain_lsn_below_horizon[i],
1465 0 : "unordered LSN"
1466 : );
1467 : }
1468 0 : }
1469 538 : let has_ancestor = base_img_from_ancestor.is_some();
1470 : // Step 1: split history into len(retain_lsn_below_horizon) + 2 buckets, where the last bucket is for all deltas above the horizon,
1471 : // and the second-to-last bucket is for the horizon. Each bucket contains lsn_last_bucket < deltas <= lsn_this_bucket.
1472 538 : let (mut split_history, lsn_split_points) = {
1473 538 : let mut split_history = Vec::new();
1474 538 : split_history.resize_with(retain_lsn_below_horizon.len() + 2, Vec::new);
1475 538 : let mut lsn_split_points = Vec::with_capacity(retain_lsn_below_horizon.len() + 1);
1476 1148 : for lsn in retain_lsn_below_horizon {
1477 610 : lsn_split_points.push(*lsn);
1478 610 : }
1479 538 : lsn_split_points.push(horizon);
1480 538 : let mut current_idx = 0;
1481 1306 : for item @ (_, lsn, _) in full_history {
1482 944 : while current_idx < lsn_split_points.len() && *lsn > lsn_split_points[current_idx] {
1483 176 : current_idx += 1;
1484 176 : }
1485 768 : split_history[current_idx].push(item);
1486 : }
1487 538 : (split_history, lsn_split_points)
1488 : };
1489 : // Step 2: filter out duplicated records due to the k-merge of image/delta layers
1490 2224 : for split_for_lsn in &mut split_history {
1491 1686 : let mut prev_lsn = None;
1492 1686 : let mut new_split_for_lsn = Vec::with_capacity(split_for_lsn.len());
1493 1686 : for record @ (_, lsn, _) in std::mem::take(split_for_lsn) {
1494 768 : if let Some(prev_lsn) = &prev_lsn {
1495 106 : if *prev_lsn == lsn {
1496 : // The case that we have an LSN with both data from the delta layer and the image layer. As
1497 : // `ValueWrapper` ensures that an image is ordered before a delta at the same LSN, we simply
1498 : // drop this delta and keep the image.
1499 : //
1500 : // For example, we have delta layer key1@0x10, key1@0x20, and image layer key1@0x10, we will
1501 : // keep the image for key1@0x10 and the delta for key1@0x20. key1@0x10 delta will be simply
1502 : // dropped.
1503 : //
1504 : // TODO: in case we have both delta + images for a given LSN and it does not exceed the delta
1505 : // threshold, we could have kept delta instead to save space. This is an optimization for the future.
1506 0 : continue;
1507 106 : }
1508 662 : }
1509 768 : prev_lsn = Some(lsn);
1510 768 : new_split_for_lsn.push(record);
1511 : }
1512 1686 : *split_for_lsn = new_split_for_lsn;
1513 : }
1514 : // Step 3: generate images when necessary
1515 538 : let mut retention = Vec::with_capacity(split_history.len());
1516 538 : let mut records_since_last_image = 0;
1517 538 : let batch_cnt = split_history.len();
1518 538 : assert!(
1519 538 : batch_cnt >= 2,
1520 0 : "should have at least below + above horizon batches"
1521 : );
1522 538 : let mut replay_history: Vec<(Key, Lsn, Value)> = Vec::new();
1523 538 : if let Some((key, lsn, img)) = base_img_from_ancestor {
1524 18 : replay_history.push((key, lsn, Value::Image(img)));
1525 520 : }
1526 :
1527 : /// Generate debug information for the replay history
1528 0 : fn generate_history_trace(replay_history: &[(Key, Lsn, Value)]) -> String {
1529 : use std::fmt::Write;
1530 0 : let mut output = String::new();
1531 0 : if let Some((key, _, _)) = replay_history.first() {
1532 0 : write!(output, "key={} ", key).unwrap();
1533 0 : let mut cnt = 0;
1534 0 : for (_, lsn, val) in replay_history {
1535 0 : if val.is_image() {
1536 0 : write!(output, "i@{} ", lsn).unwrap();
1537 0 : } else if val.will_init() {
1538 0 : write!(output, "di@{} ", lsn).unwrap();
1539 0 : } else {
1540 0 : write!(output, "d@{} ", lsn).unwrap();
1541 0 : }
1542 0 : cnt += 1;
1543 0 : if cnt >= 128 {
1544 0 : write!(output, "... and more").unwrap();
1545 0 : break;
1546 0 : }
1547 : }
1548 0 : } else {
1549 0 : write!(output, "<no history>").unwrap();
1550 0 : }
1551 0 : output
1552 0 : }
1553 :
1554 0 : fn generate_debug_trace(
1555 0 : replay_history: Option<&[(Key, Lsn, Value)]>,
1556 0 : full_history: &[(Key, Lsn, Value)],
1557 0 : lsns: &[Lsn],
1558 0 : horizon: Lsn,
1559 0 : ) -> String {
1560 : use std::fmt::Write;
1561 0 : let mut output = String::new();
1562 0 : if let Some(replay_history) = replay_history {
1563 0 : writeln!(
1564 0 : output,
1565 0 : "replay_history: {}",
1566 0 : generate_history_trace(replay_history)
1567 0 : )
1568 0 : .unwrap();
1569 0 : } else {
1570 0 : writeln!(output, "replay_history: <disabled>",).unwrap();
1571 0 : }
1572 0 : writeln!(
1573 0 : output,
1574 0 : "full_history: {}",
1575 0 : generate_history_trace(full_history)
1576 0 : )
1577 0 : .unwrap();
1578 0 : writeln!(
1579 0 : output,
1580 0 : "when processing: [{}] horizon={}",
1581 0 : lsns.iter().map(|l| format!("{l}")).join(","),
1582 0 : horizon
1583 0 : )
1584 0 : .unwrap();
1585 0 : output
1586 0 : }
1587 :
1588 1686 : for (i, split_for_lsn) in split_history.into_iter().enumerate() {
1589 : // TODO: there could be image keys inside the splits, and we can compute records_since_last_image accordingly.
1590 1686 : records_since_last_image += split_for_lsn.len();
1591 1686 : let generate_image = if i == 0 && !has_ancestor {
1592 : // We always generate images for the first batch (below horizon / lowest retain_lsn)
1593 520 : true
1594 1166 : } else if i == batch_cnt - 1 {
1595 : // Do not generate images for the last batch (above horizon)
1596 538 : false
1597 628 : } else if records_since_last_image >= delta_threshold_cnt {
1598 : // Generate images when there are too many records
1599 6 : true
1600 : } else {
1601 622 : false
1602 : };
1603 1686 : replay_history.extend(split_for_lsn.iter().map(|x| (*x).clone()));
1604 : // Only retain the items after the last image record
1605 2030 : for idx in (0..replay_history.len()).rev() {
1606 2030 : if replay_history[idx].2.will_init() {
1607 1686 : replay_history = replay_history[idx..].to_vec();
1608 1686 : break;
1609 344 : }
1610 : }
1611 1686 : if let Some((_, _, val)) = replay_history.first() {
1612 1686 : if !val.will_init() {
1613 0 : return Err(anyhow::anyhow!("invalid history, no base image")).with_context(
1614 0 : || {
1615 0 : generate_debug_trace(
1616 0 : Some(&replay_history),
1617 0 : full_history,
1618 0 : retain_lsn_below_horizon,
1619 0 : horizon,
1620 0 : )
1621 0 : },
1622 0 : );
1623 1686 : }
1624 0 : }
1625 1686 : if generate_image && records_since_last_image > 0 {
1626 526 : records_since_last_image = 0;
1627 526 : let replay_history_for_debug = if cfg!(debug_assertions) {
1628 526 : Some(replay_history.clone())
1629 : } else {
1630 0 : None
1631 : };
1632 526 : let replay_history_for_debug_ref = replay_history_for_debug.as_deref();
1633 526 : let history = std::mem::take(&mut replay_history);
1634 526 : let mut img = None;
1635 526 : let mut records = Vec::with_capacity(history.len());
1636 526 : if let (_, lsn, Value::Image(val)) = history.first().as_ref().unwrap() {
1637 504 : img = Some((*lsn, val.clone()));
1638 504 : for (_, lsn, val) in history.into_iter().skip(1) {
1639 34 : let Value::WalRecord(rec) = val else {
1640 0 : return Err(anyhow::anyhow!(
1641 0 : "invalid record, first record is image, expect walrecords"
1642 0 : ))
1643 0 : .with_context(|| {
1644 0 : generate_debug_trace(
1645 0 : replay_history_for_debug_ref,
1646 0 : full_history,
1647 0 : retain_lsn_below_horizon,
1648 0 : horizon,
1649 0 : )
1650 0 : });
1651 : };
1652 34 : records.push((lsn, rec));
1653 : }
1654 : } else {
1655 36 : for (_, lsn, val) in history.into_iter() {
1656 36 : let Value::WalRecord(rec) = val else {
1657 0 : return Err(anyhow::anyhow!("invalid record, first record is walrecord, expect rest are walrecord"))
1658 0 : .with_context(|| generate_debug_trace(
1659 0 : replay_history_for_debug_ref,
1660 0 : full_history,
1661 0 : retain_lsn_below_horizon,
1662 0 : horizon,
1663 0 : ));
1664 : };
1665 36 : records.push((lsn, rec));
1666 : }
1667 : }
1668 526 : records.reverse();
1669 526 : let state = ValueReconstructState { img, records };
1670 526 : let request_lsn = lsn_split_points[i]; // last batch does not generate image so i is always in range
1671 526 : let img = self.reconstruct_value(key, request_lsn, state).await?;
1672 526 : replay_history.push((key, request_lsn, Value::Image(img.clone())));
1673 526 : retention.push(vec![(request_lsn, Value::Image(img))]);
1674 1160 : } else {
1675 1160 : let deltas = split_for_lsn
1676 1160 : .iter()
1677 1160 : .map(|(_, lsn, value)| (*lsn, value.clone()))
1678 1160 : .collect_vec();
1679 1160 : retention.push(deltas);
1680 1160 : }
1681 : }
1682 538 : let mut result = Vec::with_capacity(retention.len());
1683 538 : assert_eq!(retention.len(), lsn_split_points.len() + 1);
1684 1686 : for (idx, logs) in retention.into_iter().enumerate() {
1685 1686 : if idx == lsn_split_points.len() {
1686 538 : return Ok(KeyHistoryRetention {
1687 538 : below_horizon: result,
1688 538 : above_horizon: KeyLogAtLsn(logs),
1689 538 : });
1690 1148 : } else {
1691 1148 : result.push((lsn_split_points[idx], KeyLogAtLsn(logs)));
1692 1148 : }
1693 : }
1694 0 : unreachable!("key retention is empty")
1695 538 : }
1696 :
1697 : /// Check how much space is left on the disk
1698 40 : async fn check_available_space(self: &Arc<Self>) -> anyhow::Result<u64> {
1699 40 : let tenants_dir = self.conf.tenants_path();
1700 :
1701 40 : let stat = Statvfs::get(&tenants_dir, None)
1702 40 : .context("statvfs failed, presumably directory got unlinked")?;
1703 :
1704 40 : let (avail_bytes, _) = stat.get_avail_total_bytes();
1705 40 :
1706 40 : Ok(avail_bytes)
1707 40 : }
1708 :
1709 : /// Check if the compaction can proceed safely without running out of space. We assume the size
1710 : /// upper bound of the produced files of a compaction job is the same as all layers involved in
1711 : /// the compaction. Therefore, we need `2 * layers_to_be_compacted_size` at least to do a
1712 : /// compaction.
1713 40 : async fn check_compaction_space(
1714 40 : self: &Arc<Self>,
1715 40 : layer_selection: &[Layer],
1716 40 : ) -> anyhow::Result<()> {
1717 40 : let available_space = self.check_available_space().await?;
1718 40 : let mut remote_layer_size = 0;
1719 40 : let mut all_layer_size = 0;
1720 160 : for layer in layer_selection {
1721 120 : let needs_download = layer.needs_download().await?;
1722 120 : if needs_download.is_some() {
1723 0 : remote_layer_size += layer.layer_desc().file_size;
1724 120 : }
1725 120 : all_layer_size += layer.layer_desc().file_size;
1726 : }
1727 40 : let allocated_space = (available_space as f64 * 0.8) as u64; /* reserve 20% space for other tasks */
1728 40 : if all_layer_size /* space needed for newly-generated file */ + remote_layer_size /* space for downloading layers */ > allocated_space
1729 : {
1730 0 : return Err(anyhow!("not enough space for compaction: available_space={}, allocated_space={}, all_layer_size={}, remote_layer_size={}, required_space={}",
1731 0 : available_space, allocated_space, all_layer_size, remote_layer_size, all_layer_size + remote_layer_size));
1732 40 : }
1733 40 : Ok(())
1734 40 : }
1735 :
1736 30 : pub(crate) async fn compact_with_gc(
1737 30 : self: &Arc<Self>,
1738 30 : cancel: &CancellationToken,
1739 30 : flags: EnumSet<CompactFlags>,
1740 30 : ctx: &RequestContext,
1741 30 : ) -> anyhow::Result<()> {
1742 30 : self.partial_compact_with_gc(Key::MIN..Key::MAX, cancel, flags, ctx)
1743 708 : .await
1744 30 : }
1745 :
1746 : /// An experimental compaction building block that combines compaction with garbage collection.
1747 : ///
1748 : /// The current implementation picks all delta + image layers that are below or intersecting with
1749 : /// the GC horizon without considering retain_lsns. Then, it does a full compaction over all these delta
1750 : /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
1751 : /// and create delta layers with all deltas >= gc horizon.
1752 : ///
1753 : /// If `key_range` is provided, it will only compact the keys within the range, aka partial compaction.
1754 : /// Partial compaction will read and process all layers overlapping with the key range, even if it might
1755 : /// contain extra keys. After the gc-compaction phase completes, delta layers that are not fully contained
1756 : /// within the key range will be rewritten to ensure they do not overlap with the delta layers. Providing
1757 : /// Key::MIN..Key..MAX to the function indicates a full compaction, though technically, `Key::MAX` is not
1758 : /// part of the range.
1759 40 : pub(crate) async fn partial_compact_with_gc(
1760 40 : self: &Arc<Self>,
1761 40 : compaction_key_range: Range<Key>,
1762 40 : cancel: &CancellationToken,
1763 40 : flags: EnumSet<CompactFlags>,
1764 40 : ctx: &RequestContext,
1765 40 : ) -> anyhow::Result<()> {
1766 40 : // Block other compaction/GC tasks from running for now. GC-compaction could run along
1767 40 : // with legacy compaction tasks in the future. Always ensure the lock order is compaction -> gc.
1768 40 : // Note that we already acquired the compaction lock when the outer `compact` function gets called.
1769 40 :
1770 40 : let gc_lock = async {
1771 40 : tokio::select! {
1772 40 : guard = self.gc_lock.lock() => Ok(guard),
1773 : // TODO: refactor to CompactionError to correctly pass cancelled error
1774 40 : _ = cancel.cancelled() => Err(anyhow!("cancelled")),
1775 : }
1776 40 : };
1777 :
1778 40 : let gc_lock = crate::timed(
1779 40 : gc_lock,
1780 40 : "acquires gc lock",
1781 40 : std::time::Duration::from_secs(5),
1782 40 : )
1783 1 : .await?;
1784 :
1785 40 : let dry_run = flags.contains(CompactFlags::DryRun);
1786 40 :
1787 40 : if compaction_key_range == (Key::MIN..Key::MAX) {
1788 30 : info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compaction_key_range={}..{}", compaction_key_range.start, compaction_key_range.end);
1789 : } else {
1790 10 : info!("running enhanced gc bottom-most compaction, dry_run={dry_run}");
1791 : }
1792 :
1793 40 : scopeguard::defer! {
1794 40 : info!("done enhanced gc bottom-most compaction");
1795 40 : };
1796 40 :
1797 40 : let mut stat = CompactionStatistics::default();
1798 :
1799 : // Step 0: pick all delta layers + image layers below/intersect with the GC horizon.
1800 : // The layer selection has the following properties:
1801 : // 1. If a layer is in the selection, all layers below it are in the selection.
1802 : // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
1803 40 : let job_desc = {
1804 40 : let guard = self.layers.read().await;
1805 40 : let layers = guard.layer_map()?;
1806 40 : let gc_info = self.gc_info.read().unwrap();
1807 40 : let mut retain_lsns_below_horizon = Vec::new();
1808 40 : let gc_cutoff = gc_info.cutoffs.select_min();
1809 44 : for (lsn, _timeline_id, _is_offloaded) in &gc_info.retain_lsns {
1810 44 : if lsn < &gc_cutoff {
1811 44 : retain_lsns_below_horizon.push(*lsn);
1812 44 : }
1813 : }
1814 40 : for lsn in gc_info.leases.keys() {
1815 0 : if lsn < &gc_cutoff {
1816 0 : retain_lsns_below_horizon.push(*lsn);
1817 0 : }
1818 : }
1819 40 : let mut selected_layers: Vec<Layer> = Vec::new();
1820 40 : drop(gc_info);
1821 : // Pick all the layers intersect or below the gc_cutoff, get the largest LSN in the selected layers.
1822 40 : let Some(max_layer_lsn) = layers
1823 40 : .iter_historic_layers()
1824 182 : .filter(|desc| desc.get_lsn_range().start <= gc_cutoff)
1825 150 : .map(|desc| desc.get_lsn_range().end)
1826 40 : .max()
1827 : else {
1828 0 : info!("no layers to compact with gc");
1829 0 : return Ok(());
1830 : };
1831 : // Then, pick all the layers that are below the max_layer_lsn. This is to ensure we can pick all single-key
1832 : // layers to compact.
1833 40 : let mut rewrite_layers = Vec::new();
1834 182 : for desc in layers.iter_historic_layers() {
1835 182 : if desc.get_lsn_range().end <= max_layer_lsn
1836 150 : && overlaps_with(&desc.get_key_range(), &compaction_key_range)
1837 : {
1838 : // If the layer overlaps with the compaction key range, we need to read it to obtain all keys within the range,
1839 : // even if it might contain extra keys
1840 120 : selected_layers.push(guard.get_from_desc(&desc));
1841 120 : // 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
1842 120 : // to overlap image layers)
1843 120 : if desc.is_delta()
1844 62 : && !fully_contains(&compaction_key_range, &desc.get_key_range())
1845 2 : {
1846 2 : rewrite_layers.push(desc);
1847 118 : }
1848 62 : }
1849 : }
1850 40 : if selected_layers.is_empty() {
1851 0 : info!("no layers to compact with gc");
1852 0 : return Ok(());
1853 40 : }
1854 40 : retain_lsns_below_horizon.sort();
1855 40 : GcCompactionJobDescription {
1856 40 : selected_layers,
1857 40 : gc_cutoff,
1858 40 : retain_lsns_below_horizon,
1859 40 : max_layer_lsn,
1860 40 : compaction_key_range,
1861 40 : rewrite_layers,
1862 40 : }
1863 : };
1864 40 : let lowest_retain_lsn = if self.ancestor_timeline.is_some() {
1865 2 : Lsn(self.ancestor_lsn.0 + 1)
1866 : } else {
1867 38 : let res = job_desc
1868 38 : .retain_lsns_below_horizon
1869 38 : .first()
1870 38 : .copied()
1871 38 : .unwrap_or(job_desc.gc_cutoff);
1872 38 : if cfg!(debug_assertions) {
1873 38 : assert_eq!(
1874 38 : res,
1875 38 : job_desc
1876 38 : .retain_lsns_below_horizon
1877 38 : .iter()
1878 38 : .min()
1879 38 : .copied()
1880 38 : .unwrap_or(job_desc.gc_cutoff)
1881 38 : );
1882 0 : }
1883 38 : res
1884 : };
1885 40 : info!(
1886 0 : "picked {} layers for compaction ({} layers need rewriting) with max_layer_lsn={} gc_cutoff={} lowest_retain_lsn={}, key_range={}..{}",
1887 0 : job_desc.selected_layers.len(),
1888 0 : job_desc.rewrite_layers.len(),
1889 : job_desc.max_layer_lsn,
1890 : job_desc.gc_cutoff,
1891 : lowest_retain_lsn,
1892 : job_desc.compaction_key_range.start,
1893 : job_desc.compaction_key_range.end
1894 : );
1895 :
1896 160 : for layer in &job_desc.selected_layers {
1897 120 : debug!("read layer: {}", layer.layer_desc().key());
1898 : }
1899 42 : for layer in &job_desc.rewrite_layers {
1900 2 : debug!("rewrite layer: {}", layer.key());
1901 : }
1902 :
1903 40 : self.check_compaction_space(&job_desc.selected_layers)
1904 91 : .await?;
1905 :
1906 : // Generate statistics for the compaction
1907 160 : for layer in &job_desc.selected_layers {
1908 120 : let desc = layer.layer_desc();
1909 120 : if desc.is_delta() {
1910 62 : stat.visit_delta_layer(desc.file_size());
1911 62 : } else {
1912 58 : stat.visit_image_layer(desc.file_size());
1913 58 : }
1914 : }
1915 :
1916 : // Step 1: construct a k-merge iterator over all layers.
1917 : // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
1918 40 : let layer_names = job_desc
1919 40 : .selected_layers
1920 40 : .iter()
1921 120 : .map(|layer| layer.layer_desc().layer_name())
1922 40 : .collect_vec();
1923 40 : if let Some(err) = check_valid_layermap(&layer_names) {
1924 0 : warn!("gc-compaction layer map check failed because {}, this is normal if partial compaction is not finished yet", err);
1925 40 : }
1926 : // The maximum LSN we are processing in this compaction loop
1927 40 : let end_lsn = job_desc
1928 40 : .selected_layers
1929 40 : .iter()
1930 120 : .map(|l| l.layer_desc().lsn_range.end)
1931 40 : .max()
1932 40 : .unwrap();
1933 40 : let mut delta_layers = Vec::new();
1934 40 : let mut image_layers = Vec::new();
1935 40 : let mut downloaded_layers = Vec::new();
1936 160 : for layer in &job_desc.selected_layers {
1937 120 : let resident_layer = layer.download_and_keep_resident().await?;
1938 120 : downloaded_layers.push(resident_layer);
1939 : }
1940 160 : for resident_layer in &downloaded_layers {
1941 120 : if resident_layer.layer_desc().is_delta() {
1942 62 : let layer = resident_layer.get_as_delta(ctx).await?;
1943 62 : delta_layers.push(layer);
1944 : } else {
1945 58 : let layer = resident_layer.get_as_image(ctx).await?;
1946 58 : image_layers.push(layer);
1947 : }
1948 : }
1949 40 : let (dense_ks, sparse_ks) = self.collect_gc_compaction_keyspace().await?;
1950 40 : let mut merge_iter = FilterIterator::create(
1951 40 : MergeIterator::create(&delta_layers, &image_layers, ctx),
1952 40 : dense_ks,
1953 40 : sparse_ks,
1954 40 : )?;
1955 :
1956 : // Step 2: Produce images+deltas.
1957 40 : let mut accumulated_values = Vec::new();
1958 40 : let mut last_key: Option<Key> = None;
1959 :
1960 : // Only create image layers when there is no ancestor branches. TODO: create covering image layer
1961 : // when some condition meet.
1962 40 : let mut image_layer_writer = if self.ancestor_timeline.is_none() {
1963 : Some(
1964 38 : SplitImageLayerWriter::new(
1965 38 : self.conf,
1966 38 : self.timeline_id,
1967 38 : self.tenant_shard_id,
1968 38 : job_desc.compaction_key_range.start,
1969 38 : lowest_retain_lsn,
1970 38 : self.get_compaction_target_size(),
1971 38 : ctx,
1972 38 : )
1973 19 : .await?,
1974 : )
1975 : } else {
1976 2 : None
1977 : };
1978 :
1979 40 : let mut delta_layer_writer = SplitDeltaLayerWriter::new(
1980 40 : self.conf,
1981 40 : self.timeline_id,
1982 40 : self.tenant_shard_id,
1983 40 : lowest_retain_lsn..end_lsn,
1984 40 : self.get_compaction_target_size(),
1985 40 : )
1986 0 : .await?;
1987 :
1988 : #[derive(Default)]
1989 : struct RewritingLayers {
1990 : before: Option<DeltaLayerWriter>,
1991 : after: Option<DeltaLayerWriter>,
1992 : }
1993 40 : let mut delta_layer_rewriters = HashMap::<Arc<PersistentLayerKey>, RewritingLayers>::new();
1994 :
1995 : /// Returns None if there is no ancestor branch. Throw an error when the key is not found.
1996 : ///
1997 : /// Currently, we always get the ancestor image for each key in the child branch no matter whether the image
1998 : /// is needed for reconstruction. This should be fixed in the future.
1999 : ///
2000 : /// Furthermore, we should do vectored get instead of a single get, or better, use k-merge for ancestor
2001 : /// images.
2002 530 : async fn get_ancestor_image(
2003 530 : tline: &Arc<Timeline>,
2004 530 : key: Key,
2005 530 : ctx: &RequestContext,
2006 530 : ) -> anyhow::Result<Option<(Key, Lsn, Bytes)>> {
2007 530 : if tline.ancestor_timeline.is_none() {
2008 516 : return Ok(None);
2009 14 : };
2010 : // This function is implemented as a get of the current timeline at ancestor LSN, therefore reusing
2011 : // as much existing code as possible.
2012 14 : let img = tline.get(key, tline.ancestor_lsn, ctx).await?;
2013 14 : Ok(Some((key, tline.ancestor_lsn, img)))
2014 530 : }
2015 :
2016 : // Actually, we can decide not to write to the image layer at all at this point because
2017 : // the key and LSN range are determined. However, to keep things simple here, we still
2018 : // create this writer, and discard the writer in the end.
2019 :
2020 822 : while let Some(((key, lsn, val), desc)) = merge_iter.next_with_trace().await? {
2021 782 : if cancel.is_cancelled() {
2022 0 : return Err(anyhow!("cancelled")); // TODO: refactor to CompactionError and pass cancel error
2023 782 : }
2024 782 : if !job_desc.compaction_key_range.contains(&key) {
2025 64 : if !desc.is_delta {
2026 60 : continue;
2027 4 : }
2028 4 : let rewriter = delta_layer_rewriters.entry(desc.clone()).or_default();
2029 4 : let rewriter = if key < job_desc.compaction_key_range.start {
2030 0 : if rewriter.before.is_none() {
2031 0 : rewriter.before = Some(
2032 0 : DeltaLayerWriter::new(
2033 0 : self.conf,
2034 0 : self.timeline_id,
2035 0 : self.tenant_shard_id,
2036 0 : desc.key_range.start,
2037 0 : desc.lsn_range.clone(),
2038 0 : ctx,
2039 0 : )
2040 0 : .await?,
2041 : );
2042 0 : }
2043 0 : rewriter.before.as_mut().unwrap()
2044 4 : } else if key >= job_desc.compaction_key_range.end {
2045 4 : if rewriter.after.is_none() {
2046 2 : rewriter.after = Some(
2047 2 : DeltaLayerWriter::new(
2048 2 : self.conf,
2049 2 : self.timeline_id,
2050 2 : self.tenant_shard_id,
2051 2 : job_desc.compaction_key_range.end,
2052 2 : desc.lsn_range.clone(),
2053 2 : ctx,
2054 2 : )
2055 1 : .await?,
2056 : );
2057 2 : }
2058 4 : rewriter.after.as_mut().unwrap()
2059 : } else {
2060 0 : unreachable!()
2061 : };
2062 4 : rewriter.put_value(key, lsn, val, ctx).await?;
2063 4 : continue;
2064 718 : }
2065 718 : match val {
2066 542 : Value::Image(_) => stat.visit_image_key(&val),
2067 176 : Value::WalRecord(_) => stat.visit_wal_key(&val),
2068 : }
2069 718 : if last_key.is_none() || last_key.as_ref() == Some(&key) {
2070 228 : if last_key.is_none() {
2071 40 : last_key = Some(key);
2072 188 : }
2073 228 : accumulated_values.push((key, lsn, val));
2074 : } else {
2075 490 : let last_key: &mut Key = last_key.as_mut().unwrap();
2076 490 : stat.on_unique_key_visited(); // TODO: adjust statistics for partial compaction
2077 490 : let retention = self
2078 490 : .generate_key_retention(
2079 490 : *last_key,
2080 490 : &accumulated_values,
2081 490 : job_desc.gc_cutoff,
2082 490 : &job_desc.retain_lsns_below_horizon,
2083 490 : COMPACTION_DELTA_THRESHOLD,
2084 490 : get_ancestor_image(self, *last_key, ctx).await?,
2085 : )
2086 0 : .await?;
2087 490 : retention
2088 490 : .pipe_to(
2089 490 : *last_key,
2090 490 : &mut delta_layer_writer,
2091 490 : image_layer_writer.as_mut(),
2092 490 : &mut stat,
2093 490 : ctx,
2094 490 : )
2095 492 : .await?;
2096 490 : accumulated_values.clear();
2097 490 : *last_key = key;
2098 490 : accumulated_values.push((key, lsn, val));
2099 : }
2100 : }
2101 :
2102 : // TODO: move the below part to the loop body
2103 40 : let last_key = last_key.expect("no keys produced during compaction");
2104 40 : stat.on_unique_key_visited();
2105 :
2106 40 : let retention = self
2107 40 : .generate_key_retention(
2108 40 : last_key,
2109 40 : &accumulated_values,
2110 40 : job_desc.gc_cutoff,
2111 40 : &job_desc.retain_lsns_below_horizon,
2112 40 : COMPACTION_DELTA_THRESHOLD,
2113 40 : get_ancestor_image(self, last_key, ctx).await?,
2114 : )
2115 0 : .await?;
2116 40 : retention
2117 40 : .pipe_to(
2118 40 : last_key,
2119 40 : &mut delta_layer_writer,
2120 40 : image_layer_writer.as_mut(),
2121 40 : &mut stat,
2122 40 : ctx,
2123 40 : )
2124 38 : .await?;
2125 : // end: move the above part to the loop body
2126 :
2127 40 : let mut rewrote_delta_layers = Vec::new();
2128 42 : for (key, writers) in delta_layer_rewriters {
2129 2 : if let Some(delta_writer_before) = writers.before {
2130 0 : let (desc, path) = delta_writer_before
2131 0 : .finish(job_desc.compaction_key_range.start, ctx)
2132 0 : .await?;
2133 0 : let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
2134 0 : rewrote_delta_layers.push(layer);
2135 2 : }
2136 2 : if let Some(delta_writer_after) = writers.after {
2137 5 : let (desc, path) = delta_writer_after.finish(key.key_range.end, ctx).await?;
2138 2 : let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
2139 2 : rewrote_delta_layers.push(layer);
2140 0 : }
2141 : }
2142 :
2143 58 : let discard = |key: &PersistentLayerKey| {
2144 58 : let key = key.clone();
2145 58 : async move { KeyHistoryRetention::discard_key(&key, self, dry_run).await }
2146 58 : };
2147 :
2148 40 : let produced_image_layers = if let Some(writer) = image_layer_writer {
2149 38 : if !dry_run {
2150 34 : let end_key = job_desc.compaction_key_range.end;
2151 34 : writer
2152 34 : .finish_with_discard_fn(self, ctx, end_key, discard)
2153 52 : .await?
2154 : } else {
2155 4 : drop(writer);
2156 4 : Vec::new()
2157 : }
2158 : } else {
2159 2 : Vec::new()
2160 : };
2161 :
2162 40 : let produced_delta_layers = if !dry_run {
2163 36 : delta_layer_writer
2164 36 : .finish_with_discard_fn(self, ctx, discard)
2165 30 : .await?
2166 : } else {
2167 4 : drop(delta_layer_writer);
2168 4 : Vec::new()
2169 : };
2170 :
2171 : // TODO: make image/delta/rewrote_delta layers generation atomic. At this point, we already generated resident layers, and if
2172 : // compaction is cancelled at this point, we might have some layers that are not cleaned up.
2173 40 : let mut compact_to = Vec::new();
2174 40 : let mut keep_layers = HashSet::new();
2175 40 : let produced_delta_layers_len = produced_delta_layers.len();
2176 40 : let produced_image_layers_len = produced_image_layers.len();
2177 64 : for action in produced_delta_layers {
2178 24 : match action {
2179 12 : BatchWriterResult::Produced(layer) => {
2180 12 : if cfg!(debug_assertions) {
2181 12 : info!("produced delta layer: {}", layer.layer_desc().key());
2182 0 : }
2183 12 : stat.produce_delta_layer(layer.layer_desc().file_size());
2184 12 : compact_to.push(layer);
2185 : }
2186 12 : BatchWriterResult::Discarded(l) => {
2187 12 : if cfg!(debug_assertions) {
2188 12 : info!("discarded delta layer: {}", l);
2189 0 : }
2190 12 : keep_layers.insert(l);
2191 12 : stat.discard_delta_layer();
2192 : }
2193 : }
2194 : }
2195 42 : for layer in &rewrote_delta_layers {
2196 2 : debug!(
2197 0 : "produced rewritten delta layer: {}",
2198 0 : layer.layer_desc().key()
2199 : );
2200 : }
2201 40 : compact_to.extend(rewrote_delta_layers);
2202 74 : for action in produced_image_layers {
2203 34 : match action {
2204 26 : BatchWriterResult::Produced(layer) => {
2205 26 : debug!("produced image layer: {}", layer.layer_desc().key());
2206 26 : stat.produce_image_layer(layer.layer_desc().file_size());
2207 26 : compact_to.push(layer);
2208 : }
2209 8 : BatchWriterResult::Discarded(l) => {
2210 8 : debug!("discarded image layer: {}", l);
2211 8 : keep_layers.insert(l);
2212 8 : stat.discard_image_layer();
2213 : }
2214 : }
2215 : }
2216 :
2217 40 : let mut layer_selection = job_desc.selected_layers;
2218 :
2219 : // Partial compaction might select more data than it processes, e.g., if
2220 : // the compaction_key_range only partially overlaps:
2221 : //
2222 : // [---compaction_key_range---]
2223 : // [---A----][----B----][----C----][----D----]
2224 : //
2225 : // For delta layers, we will rewrite the layers so that it is cut exactly at
2226 : // the compaction key range, so we can always discard them. However, for image
2227 : // layers, as we do not rewrite them for now, we need to handle them differently.
2228 : // Assume image layers A, B, C, D are all in the `layer_selection`.
2229 : //
2230 : // The created image layers contain whatever is needed from B, C, and from
2231 : // `----]` of A, and from `[---` of D.
2232 : //
2233 : // In contrast, `[---A` and `D----]` have not been processed, so, we must
2234 : // keep that data.
2235 : //
2236 : // The solution for now is to keep A and D completely if they are image layers.
2237 : // (layer_selection is what we'll remove from the layer map, so, retain what
2238 : // is _not_ fully covered by compaction_key_range).
2239 160 : for layer in &layer_selection {
2240 120 : if !layer.layer_desc().is_delta() {
2241 58 : if !overlaps_with(
2242 58 : &layer.layer_desc().key_range,
2243 58 : &job_desc.compaction_key_range,
2244 58 : ) {
2245 0 : bail!("violated constraint: image layer outside of compaction key range");
2246 58 : }
2247 58 : if !fully_contains(
2248 58 : &job_desc.compaction_key_range,
2249 58 : &layer.layer_desc().key_range,
2250 58 : ) {
2251 8 : keep_layers.insert(layer.layer_desc().key());
2252 50 : }
2253 62 : }
2254 : }
2255 :
2256 120 : layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
2257 40 :
2258 40 : info!(
2259 0 : "gc-compaction statistics: {}",
2260 0 : serde_json::to_string(&stat)?
2261 : );
2262 :
2263 40 : if dry_run {
2264 4 : return Ok(());
2265 36 : }
2266 36 :
2267 36 : info!(
2268 0 : "produced {} delta layers and {} image layers, {} layers are kept",
2269 0 : produced_delta_layers_len,
2270 0 : produced_image_layers_len,
2271 0 : layer_selection.len()
2272 : );
2273 :
2274 : // Step 3: Place back to the layer map.
2275 : {
2276 : // TODO: sanity check if the layer map is valid (i.e., should not have overlaps)
2277 36 : let mut guard = self.layers.write().await;
2278 36 : guard
2279 36 : .open_mut()?
2280 36 : .finish_gc_compaction(&layer_selection, &compact_to, &self.metrics)
2281 36 : };
2282 36 : self.remote_client
2283 36 : .schedule_compaction_update(&layer_selection, &compact_to)?;
2284 :
2285 36 : drop(gc_lock);
2286 36 :
2287 36 : Ok(())
2288 40 : }
2289 : }
2290 :
2291 : struct TimelineAdaptor {
2292 : timeline: Arc<Timeline>,
2293 :
2294 : keyspace: (Lsn, KeySpace),
2295 :
2296 : new_deltas: Vec<ResidentLayer>,
2297 : new_images: Vec<ResidentLayer>,
2298 : layers_to_delete: Vec<Arc<PersistentLayerDesc>>,
2299 : }
2300 :
2301 : impl TimelineAdaptor {
2302 0 : pub fn new(timeline: &Arc<Timeline>, keyspace: (Lsn, KeySpace)) -> Self {
2303 0 : Self {
2304 0 : timeline: timeline.clone(),
2305 0 : keyspace,
2306 0 : new_images: Vec::new(),
2307 0 : new_deltas: Vec::new(),
2308 0 : layers_to_delete: Vec::new(),
2309 0 : }
2310 0 : }
2311 :
2312 0 : pub async fn flush_updates(&mut self) -> Result<(), CompactionError> {
2313 0 : let layers_to_delete = {
2314 0 : let guard = self.timeline.layers.read().await;
2315 0 : self.layers_to_delete
2316 0 : .iter()
2317 0 : .map(|x| guard.get_from_desc(x))
2318 0 : .collect::<Vec<Layer>>()
2319 0 : };
2320 0 : self.timeline
2321 0 : .finish_compact_batch(&self.new_deltas, &self.new_images, &layers_to_delete)
2322 0 : .await?;
2323 :
2324 0 : self.timeline
2325 0 : .upload_new_image_layers(std::mem::take(&mut self.new_images))?;
2326 :
2327 0 : self.new_deltas.clear();
2328 0 : self.layers_to_delete.clear();
2329 0 : Ok(())
2330 0 : }
2331 : }
2332 :
2333 : #[derive(Clone)]
2334 : struct ResidentDeltaLayer(ResidentLayer);
2335 : #[derive(Clone)]
2336 : struct ResidentImageLayer(ResidentLayer);
2337 :
2338 : impl CompactionJobExecutor for TimelineAdaptor {
2339 : type Key = pageserver_api::key::Key;
2340 :
2341 : type Layer = OwnArc<PersistentLayerDesc>;
2342 : type DeltaLayer = ResidentDeltaLayer;
2343 : type ImageLayer = ResidentImageLayer;
2344 :
2345 : type RequestContext = crate::context::RequestContext;
2346 :
2347 0 : fn get_shard_identity(&self) -> &ShardIdentity {
2348 0 : self.timeline.get_shard_identity()
2349 0 : }
2350 :
2351 0 : async fn get_layers(
2352 0 : &mut self,
2353 0 : key_range: &Range<Key>,
2354 0 : lsn_range: &Range<Lsn>,
2355 0 : _ctx: &RequestContext,
2356 0 : ) -> anyhow::Result<Vec<OwnArc<PersistentLayerDesc>>> {
2357 0 : self.flush_updates().await?;
2358 :
2359 0 : let guard = self.timeline.layers.read().await;
2360 0 : let layer_map = guard.layer_map()?;
2361 :
2362 0 : let result = layer_map
2363 0 : .iter_historic_layers()
2364 0 : .filter(|l| {
2365 0 : overlaps_with(&l.lsn_range, lsn_range) && overlaps_with(&l.key_range, key_range)
2366 0 : })
2367 0 : .map(OwnArc)
2368 0 : .collect();
2369 0 : Ok(result)
2370 0 : }
2371 :
2372 0 : async fn get_keyspace(
2373 0 : &mut self,
2374 0 : key_range: &Range<Key>,
2375 0 : lsn: Lsn,
2376 0 : _ctx: &RequestContext,
2377 0 : ) -> anyhow::Result<Vec<Range<Key>>> {
2378 0 : if lsn == self.keyspace.0 {
2379 0 : Ok(pageserver_compaction::helpers::intersect_keyspace(
2380 0 : &self.keyspace.1.ranges,
2381 0 : key_range,
2382 0 : ))
2383 : } else {
2384 : // The current compaction implementation only ever requests the key space
2385 : // at the compaction end LSN.
2386 0 : anyhow::bail!("keyspace not available for requested lsn");
2387 : }
2388 0 : }
2389 :
2390 0 : async fn downcast_delta_layer(
2391 0 : &self,
2392 0 : layer: &OwnArc<PersistentLayerDesc>,
2393 0 : ) -> anyhow::Result<Option<ResidentDeltaLayer>> {
2394 0 : // this is a lot more complex than a simple downcast...
2395 0 : if layer.is_delta() {
2396 0 : let l = {
2397 0 : let guard = self.timeline.layers.read().await;
2398 0 : guard.get_from_desc(layer)
2399 : };
2400 0 : let result = l.download_and_keep_resident().await?;
2401 :
2402 0 : Ok(Some(ResidentDeltaLayer(result)))
2403 : } else {
2404 0 : Ok(None)
2405 : }
2406 0 : }
2407 :
2408 0 : async fn create_image(
2409 0 : &mut self,
2410 0 : lsn: Lsn,
2411 0 : key_range: &Range<Key>,
2412 0 : ctx: &RequestContext,
2413 0 : ) -> anyhow::Result<()> {
2414 0 : Ok(self.create_image_impl(lsn, key_range, ctx).await?)
2415 0 : }
2416 :
2417 0 : async fn create_delta(
2418 0 : &mut self,
2419 0 : lsn_range: &Range<Lsn>,
2420 0 : key_range: &Range<Key>,
2421 0 : input_layers: &[ResidentDeltaLayer],
2422 0 : ctx: &RequestContext,
2423 0 : ) -> anyhow::Result<()> {
2424 0 : debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
2425 :
2426 0 : let mut all_entries = Vec::new();
2427 0 : for dl in input_layers.iter() {
2428 0 : all_entries.extend(dl.load_keys(ctx).await?);
2429 : }
2430 :
2431 : // The current stdlib sorting implementation is designed in a way where it is
2432 : // particularly fast where the slice is made up of sorted sub-ranges.
2433 0 : all_entries.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
2434 :
2435 0 : let mut writer = DeltaLayerWriter::new(
2436 0 : self.timeline.conf,
2437 0 : self.timeline.timeline_id,
2438 0 : self.timeline.tenant_shard_id,
2439 0 : key_range.start,
2440 0 : lsn_range.clone(),
2441 0 : ctx,
2442 0 : )
2443 0 : .await?;
2444 :
2445 0 : let mut dup_values = 0;
2446 0 :
2447 0 : // This iterator walks through all key-value pairs from all the layers
2448 0 : // we're compacting, in key, LSN order.
2449 0 : let mut prev: Option<(Key, Lsn)> = None;
2450 : for &DeltaEntry {
2451 0 : key, lsn, ref val, ..
2452 0 : } in all_entries.iter()
2453 : {
2454 0 : if prev == Some((key, lsn)) {
2455 : // This is a duplicate. Skip it.
2456 : //
2457 : // It can happen if compaction is interrupted after writing some
2458 : // layers but not all, and we are compacting the range again.
2459 : // The calculations in the algorithm assume that there are no
2460 : // duplicates, so the math on targeted file size is likely off,
2461 : // and we will create smaller files than expected.
2462 0 : dup_values += 1;
2463 0 : continue;
2464 0 : }
2465 :
2466 0 : let value = val.load(ctx).await?;
2467 :
2468 0 : writer.put_value(key, lsn, value, ctx).await?;
2469 :
2470 0 : prev = Some((key, lsn));
2471 : }
2472 :
2473 0 : if dup_values > 0 {
2474 0 : warn!("delta layer created with {} duplicate values", dup_values);
2475 0 : }
2476 :
2477 0 : fail_point!("delta-layer-writer-fail-before-finish", |_| {
2478 0 : Err(anyhow::anyhow!(
2479 0 : "failpoint delta-layer-writer-fail-before-finish"
2480 0 : ))
2481 0 : });
2482 :
2483 0 : let (desc, path) = writer.finish(prev.unwrap().0.next(), ctx).await?;
2484 0 : let new_delta_layer =
2485 0 : Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
2486 :
2487 0 : self.new_deltas.push(new_delta_layer);
2488 0 : Ok(())
2489 0 : }
2490 :
2491 0 : async fn delete_layer(
2492 0 : &mut self,
2493 0 : layer: &OwnArc<PersistentLayerDesc>,
2494 0 : _ctx: &RequestContext,
2495 0 : ) -> anyhow::Result<()> {
2496 0 : self.layers_to_delete.push(layer.clone().0);
2497 0 : Ok(())
2498 0 : }
2499 : }
2500 :
2501 : impl TimelineAdaptor {
2502 0 : async fn create_image_impl(
2503 0 : &mut self,
2504 0 : lsn: Lsn,
2505 0 : key_range: &Range<Key>,
2506 0 : ctx: &RequestContext,
2507 0 : ) -> Result<(), CreateImageLayersError> {
2508 0 : let timer = self.timeline.metrics.create_images_time_histo.start_timer();
2509 :
2510 0 : let image_layer_writer = ImageLayerWriter::new(
2511 0 : self.timeline.conf,
2512 0 : self.timeline.timeline_id,
2513 0 : self.timeline.tenant_shard_id,
2514 0 : key_range,
2515 0 : lsn,
2516 0 : ctx,
2517 0 : )
2518 0 : .await?;
2519 :
2520 0 : fail_point!("image-layer-writer-fail-before-finish", |_| {
2521 0 : Err(CreateImageLayersError::Other(anyhow::anyhow!(
2522 0 : "failpoint image-layer-writer-fail-before-finish"
2523 0 : )))
2524 0 : });
2525 :
2526 0 : let keyspace = KeySpace {
2527 0 : ranges: self.get_keyspace(key_range, lsn, ctx).await?,
2528 : };
2529 : // TODO set proper (stateful) start. The create_image_layer_for_rel_blocks function mostly
2530 0 : let start = Key::MIN;
2531 : let ImageLayerCreationOutcome {
2532 0 : image,
2533 : next_start_key: _,
2534 0 : } = self
2535 0 : .timeline
2536 0 : .create_image_layer_for_rel_blocks(
2537 0 : &keyspace,
2538 0 : image_layer_writer,
2539 0 : lsn,
2540 0 : ctx,
2541 0 : key_range.clone(),
2542 0 : start,
2543 0 : )
2544 0 : .await?;
2545 :
2546 0 : if let Some(image_layer) = image {
2547 0 : self.new_images.push(image_layer);
2548 0 : }
2549 :
2550 0 : timer.stop_and_record();
2551 0 :
2552 0 : Ok(())
2553 0 : }
2554 : }
2555 :
2556 : impl CompactionRequestContext for crate::context::RequestContext {}
2557 :
2558 : #[derive(Debug, Clone)]
2559 : pub struct OwnArc<T>(pub Arc<T>);
2560 :
2561 : impl<T> Deref for OwnArc<T> {
2562 : type Target = <Arc<T> as Deref>::Target;
2563 0 : fn deref(&self) -> &Self::Target {
2564 0 : &self.0
2565 0 : }
2566 : }
2567 :
2568 : impl<T> AsRef<T> for OwnArc<T> {
2569 0 : fn as_ref(&self) -> &T {
2570 0 : self.0.as_ref()
2571 0 : }
2572 : }
2573 :
2574 : impl CompactionLayer<Key> for OwnArc<PersistentLayerDesc> {
2575 0 : fn key_range(&self) -> &Range<Key> {
2576 0 : &self.key_range
2577 0 : }
2578 0 : fn lsn_range(&self) -> &Range<Lsn> {
2579 0 : &self.lsn_range
2580 0 : }
2581 0 : fn file_size(&self) -> u64 {
2582 0 : self.file_size
2583 0 : }
2584 0 : fn short_id(&self) -> std::string::String {
2585 0 : self.as_ref().short_id().to_string()
2586 0 : }
2587 0 : fn is_delta(&self) -> bool {
2588 0 : self.as_ref().is_delta()
2589 0 : }
2590 : }
2591 :
2592 : impl CompactionLayer<Key> for OwnArc<DeltaLayer> {
2593 0 : fn key_range(&self) -> &Range<Key> {
2594 0 : &self.layer_desc().key_range
2595 0 : }
2596 0 : fn lsn_range(&self) -> &Range<Lsn> {
2597 0 : &self.layer_desc().lsn_range
2598 0 : }
2599 0 : fn file_size(&self) -> u64 {
2600 0 : self.layer_desc().file_size
2601 0 : }
2602 0 : fn short_id(&self) -> std::string::String {
2603 0 : self.layer_desc().short_id().to_string()
2604 0 : }
2605 0 : fn is_delta(&self) -> bool {
2606 0 : true
2607 0 : }
2608 : }
2609 :
2610 : use crate::tenant::timeline::DeltaEntry;
2611 :
2612 : impl CompactionLayer<Key> for ResidentDeltaLayer {
2613 0 : fn key_range(&self) -> &Range<Key> {
2614 0 : &self.0.layer_desc().key_range
2615 0 : }
2616 0 : fn lsn_range(&self) -> &Range<Lsn> {
2617 0 : &self.0.layer_desc().lsn_range
2618 0 : }
2619 0 : fn file_size(&self) -> u64 {
2620 0 : self.0.layer_desc().file_size
2621 0 : }
2622 0 : fn short_id(&self) -> std::string::String {
2623 0 : self.0.layer_desc().short_id().to_string()
2624 0 : }
2625 0 : fn is_delta(&self) -> bool {
2626 0 : true
2627 0 : }
2628 : }
2629 :
2630 : impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
2631 : type DeltaEntry<'a> = DeltaEntry<'a>;
2632 :
2633 0 : async fn load_keys<'a>(&self, ctx: &RequestContext) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
2634 0 : self.0.get_as_delta(ctx).await?.index_entries(ctx).await
2635 0 : }
2636 : }
2637 :
2638 : impl CompactionLayer<Key> for ResidentImageLayer {
2639 0 : fn key_range(&self) -> &Range<Key> {
2640 0 : &self.0.layer_desc().key_range
2641 0 : }
2642 0 : fn lsn_range(&self) -> &Range<Lsn> {
2643 0 : &self.0.layer_desc().lsn_range
2644 0 : }
2645 0 : fn file_size(&self) -> u64 {
2646 0 : self.0.layer_desc().file_size
2647 0 : }
2648 0 : fn short_id(&self) -> std::string::String {
2649 0 : self.0.layer_desc().short_id().to_string()
2650 0 : }
2651 0 : fn is_delta(&self) -> bool {
2652 0 : false
2653 0 : }
2654 : }
2655 : impl CompactionImageLayer<TimelineAdaptor> for ResidentImageLayer {}
|
