Line data Source code
1 : // Copyright 2018 The LevelDB-Go and Pebble Authors. All rights reserved. Use
2 : // of this source code is governed by a BSD-style license that can be found in
3 : // the LICENSE file.
4 :
5 : package pebble
6 :
7 : import (
8 : "bytes"
9 : "cmp"
10 : "fmt"
11 : "iter"
12 : "math"
13 : "slices"
14 : "sort"
15 : "strings"
16 :
17 : "github.com/cockroachdb/errors"
18 : "github.com/cockroachdb/pebble/internal/base"
19 : "github.com/cockroachdb/pebble/internal/humanize"
20 : "github.com/cockroachdb/pebble/internal/invariants"
21 : "github.com/cockroachdb/pebble/internal/manifest"
22 : "github.com/cockroachdb/pebble/internal/problemspans"
23 : )
24 :
25 : // The minimum count for an intra-L0 compaction. This matches the RocksDB
26 : // heuristic.
27 : const minIntraL0Count = 4
28 :
29 : type compactionEnv struct {
30 : // diskAvailBytes holds a statistic on the number of bytes available on
31 : // disk, as reported by the filesystem. It's used to be more restrictive in
32 : // expanding compactions if available disk space is limited.
33 : //
34 : // The cached value (d.diskAvailBytes) is updated whenever a file is deleted
35 : // and whenever a compaction or flush completes. Since file removal is the
36 : // primary means of reclaiming space, there is a rough bound on the
37 : // statistic's staleness when available bytes is growing. Compactions and
38 : // flushes are longer, slower operations and provide a much looser bound
39 : // when available bytes is decreasing.
40 : diskAvailBytes uint64
41 : earliestUnflushedSeqNum base.SeqNum
42 : earliestSnapshotSeqNum base.SeqNum
43 : inProgressCompactions []compactionInfo
44 : readCompactionEnv readCompactionEnv
45 : // problemSpans is checked by the compaction picker to avoid compactions that
46 : // overlap an active "problem span". It can be nil when there are no problem
47 : // spans.
48 : problemSpans *problemspans.ByLevel
49 : }
50 :
51 : type compactionPickerMetrics struct {
52 : levels [numLevels]struct {
53 : score float64
54 : fillFactor float64
55 : compensatedFillFactor float64
56 : }
57 : }
58 :
59 : type compactionPicker interface {
60 : getMetrics([]compactionInfo) compactionPickerMetrics
61 : getBaseLevel() int
62 : estimatedCompactionDebt() uint64
63 : pickAutoScore(env compactionEnv) (pc pickedCompaction)
64 : pickAutoNonScore(env compactionEnv) (pc pickedCompaction)
65 : forceBaseLevel1()
66 : }
67 :
68 : // A pickedCompaction describes a potential compaction that the compaction
69 : // picker has selected, based on its heuristics. When a compaction begins to
70 : // execute, it is converted into a compaction struct by ConstructCompaction.
71 : type pickedCompaction interface {
72 : // ManualID returns the ID of the manual compaction, or 0 if the picked
73 : // compaction is not a result of a manual compaction.
74 : ManualID() uint64
75 : // ConstructCompaction creates a compaction from the picked compaction.
76 : ConstructCompaction(*DB, CompactionGrantHandle) compaction
77 : // WaitingCompaction returns a WaitingCompaction description of this
78 : // compaction for consumption by the compaction scheduler.
79 : WaitingCompaction() WaitingCompaction
80 : }
81 :
82 : // readCompactionEnv is used to hold data required to perform read compactions
83 : type readCompactionEnv struct {
84 : rescheduleReadCompaction *bool
85 : readCompactions *readCompactionQueue
86 : flushing bool
87 : }
88 :
89 : // Information about in-progress compactions provided to the compaction picker.
90 : // These are used to constrain the new compactions that will be picked.
91 : type compactionInfo struct {
92 : // versionEditApplied is true if this compaction's version edit has already
93 : // been committed. The compaction may still be in-progress deleting newly
94 : // obsolete files.
95 : versionEditApplied bool
96 : // kind indicates the kind of compaction.
97 : kind compactionKind
98 : inputs []compactionLevel
99 : outputLevel int
100 : // bounds may be nil if the compaction does not involve sstables
101 : // (specifically, a blob file rewrite).
102 : bounds *base.UserKeyBounds
103 : }
104 :
105 0 : func (info compactionInfo) String() string {
106 0 : var buf bytes.Buffer
107 0 : var largest int
108 0 : for i, in := range info.inputs {
109 0 : if i > 0 {
110 0 : fmt.Fprintf(&buf, " -> ")
111 0 : }
112 0 : fmt.Fprintf(&buf, "L%d", in.level)
113 0 : for f := range in.files.All() {
114 0 : fmt.Fprintf(&buf, " %s", f.TableNum)
115 0 : }
116 0 : if largest < in.level {
117 0 : largest = in.level
118 0 : }
119 : }
120 0 : if largest != info.outputLevel || len(info.inputs) == 1 {
121 0 : fmt.Fprintf(&buf, " -> L%d", info.outputLevel)
122 0 : }
123 0 : return buf.String()
124 : }
125 :
126 : // sublevelInfo is used to tag a LevelSlice for an L0 sublevel with the
127 : // sublevel.
128 : type sublevelInfo struct {
129 : manifest.LevelSlice
130 : sublevel manifest.Layer
131 : }
132 :
133 1 : func (cl sublevelInfo) Clone() sublevelInfo {
134 1 : return sublevelInfo{
135 1 : sublevel: cl.sublevel,
136 1 : LevelSlice: cl.LevelSlice,
137 1 : }
138 1 : }
139 0 : func (cl sublevelInfo) String() string {
140 0 : return fmt.Sprintf(`Sublevel %s; Levels %s`, cl.sublevel, cl.LevelSlice)
141 0 : }
142 :
143 : // generateSublevelInfo will generate the level slices for each of the sublevels
144 : // from the level slice for all of L0.
145 1 : func generateSublevelInfo(cmp base.Compare, levelFiles manifest.LevelSlice) []sublevelInfo {
146 1 : sublevelMap := make(map[uint64][]*manifest.TableMetadata)
147 1 : for f := range levelFiles.All() {
148 1 : sublevelMap[uint64(f.SubLevel)] = append(sublevelMap[uint64(f.SubLevel)], f)
149 1 : }
150 :
151 1 : var sublevels []int
152 1 : for level := range sublevelMap {
153 1 : sublevels = append(sublevels, int(level))
154 1 : }
155 1 : sort.Ints(sublevels)
156 1 :
157 1 : var levelSlices []sublevelInfo
158 1 : for _, sublevel := range sublevels {
159 1 : metas := sublevelMap[uint64(sublevel)]
160 1 : levelSlices = append(
161 1 : levelSlices,
162 1 : sublevelInfo{
163 1 : manifest.NewLevelSliceKeySorted(cmp, metas),
164 1 : manifest.L0Sublevel(sublevel),
165 1 : },
166 1 : )
167 1 : }
168 1 : return levelSlices
169 : }
170 :
171 : // pickedCompactionMetrics holds metrics related to the compaction picking process
172 : type pickedCompactionMetrics struct {
173 : // scores contains candidateLevelInfo.scores.
174 : scores []float64
175 : singleLevelOverlappingRatio float64
176 : multiLevelOverlappingRatio float64
177 : }
178 :
179 : // pickedTableCompaction contains information about a compaction of sstables
180 : // that has already been chosen, and is being constructed. Compaction
181 : // construction info lives in this struct, and is copied over into the
182 : // compaction struct in constructCompaction.
183 : type pickedTableCompaction struct {
184 : // score of the chosen compaction (candidateLevelInfo.score).
185 : score float64
186 : // kind indicates the kind of compaction.
187 : kind compactionKind
188 : // manualID > 0 iff this is a manual compaction. It exists solely for
189 : // internal bookkeeping.
190 : manualID uint64
191 : // startLevel is the level that is being compacted. Inputs from startLevel
192 : // and outputLevel will be merged to produce a set of outputLevel files.
193 : startLevel *compactionLevel
194 : // outputLevel is the level that files are being produced in. outputLevel is
195 : // equal to startLevel+1 except when:
196 : // - if startLevel is 0, the output level equals compactionPicker.baseLevel().
197 : // - in multilevel compaction, the output level is the lowest level involved in
198 : // the compaction
199 : outputLevel *compactionLevel
200 : // inputs contain levels involved in the compaction in ascending order
201 : inputs []compactionLevel
202 : // LBase at the time of compaction picking. Might be uninitialized for
203 : // intra-L0 compactions.
204 : baseLevel int
205 : // L0-specific compaction info. Set to a non-nil value for all compactions
206 : // where startLevel == 0 that were generated by L0Sublevels.
207 : lcf *manifest.L0CompactionFiles
208 : // maxOutputFileSize is the maximum size of an individual table created
209 : // during compaction.
210 : maxOutputFileSize uint64
211 : // maxOverlapBytes is the maximum number of bytes of overlap allowed for a
212 : // single output table with the tables in the grandparent level.
213 : maxOverlapBytes uint64
214 : // maxReadCompactionBytes is the maximum bytes a read compaction is allowed to
215 : // overlap in its output level with. If the overlap is greater than
216 : // maxReadCompaction bytes, then we don't proceed with the compaction.
217 : maxReadCompactionBytes uint64
218 :
219 : // The boundaries of the input data.
220 : bounds base.UserKeyBounds
221 : version *manifest.Version
222 : l0Organizer *manifest.L0Organizer
223 : pickerMetrics pickedCompactionMetrics
224 : }
225 :
226 : // Assert that *pickedTableCompaction implements pickedCompaction.
227 : var _ pickedCompaction = (*pickedTableCompaction)(nil)
228 :
229 : // ManualID returns the ID of the manual compaction, or 0 if the picked
230 : // compaction is not a result of a manual compaction.
231 1 : func (pc *pickedTableCompaction) ManualID() uint64 { return pc.manualID }
232 :
233 : // Kind returns the kind of compaction.
234 0 : func (pc *pickedTableCompaction) Kind() compactionKind { return pc.kind }
235 :
236 : // Score returns the score of the level at the time the compaction was picked.
237 0 : func (pc *pickedTableCompaction) Score() float64 { return pc.score }
238 :
239 : // ConstructCompaction creates a compaction struct from the
240 : // pickedTableCompaction.
241 : func (pc *pickedTableCompaction) ConstructCompaction(
242 : d *DB, grantHandle CompactionGrantHandle,
243 1 : ) compaction {
244 1 : return newCompaction(
245 1 : pc,
246 1 : d.opts,
247 1 : d.timeNow(),
248 1 : d.ObjProvider(),
249 1 : grantHandle,
250 1 : d.TableFormat(),
251 1 : d.determineCompactionValueSeparation)
252 1 : }
253 :
254 : // WaitingCompaction returns a WaitingCompaction description of this compaction
255 : // for consumption by the compaction scheduler.
256 1 : func (pc *pickedTableCompaction) WaitingCompaction() WaitingCompaction {
257 1 : if pc.manualID > 0 {
258 1 : return WaitingCompaction{Priority: manualCompactionPriority, Score: pc.score}
259 1 : }
260 1 : entry, ok := scheduledCompactionMap[pc.kind]
261 1 : if !ok {
262 0 : panic(errors.AssertionFailedf("unexpected compactionKind %s", pc.kind))
263 : }
264 1 : return WaitingCompaction{
265 1 : Optional: entry.optional,
266 1 : Priority: entry.priority,
267 1 : Score: pc.score,
268 1 : }
269 : }
270 :
271 1 : func defaultOutputLevel(startLevel, baseLevel int) int {
272 1 : outputLevel := startLevel + 1
273 1 : if startLevel == 0 {
274 1 : outputLevel = baseLevel
275 1 : }
276 1 : if outputLevel >= numLevels-1 {
277 1 : outputLevel = numLevels - 1
278 1 : }
279 1 : return outputLevel
280 : }
281 :
282 : func newPickedTableCompaction(
283 : opts *Options,
284 : cur *manifest.Version,
285 : l0Organizer *manifest.L0Organizer,
286 : startLevel, outputLevel, baseLevel int,
287 1 : ) *pickedTableCompaction {
288 1 : if outputLevel > 0 && baseLevel == 0 {
289 0 : panic("base level cannot be 0")
290 : }
291 1 : if startLevel > 0 && startLevel < baseLevel {
292 0 : panic(fmt.Sprintf("invalid compaction: start level %d should not be empty (base level %d)",
293 0 : startLevel, baseLevel))
294 : }
295 :
296 1 : targetFileSize := opts.TargetFileSize(outputLevel, baseLevel)
297 1 : pc := &pickedTableCompaction{
298 1 : version: cur,
299 1 : l0Organizer: l0Organizer,
300 1 : baseLevel: baseLevel,
301 1 : inputs: []compactionLevel{{level: startLevel}, {level: outputLevel}},
302 1 : maxOutputFileSize: uint64(targetFileSize),
303 1 : maxOverlapBytes: maxGrandparentOverlapBytes(targetFileSize),
304 1 : maxReadCompactionBytes: maxReadCompactionBytes(targetFileSize),
305 1 : }
306 1 : pc.startLevel = &pc.inputs[0]
307 1 : pc.outputLevel = &pc.inputs[1]
308 1 : return pc
309 : }
310 :
311 : // adjustedOutputLevel is the output level used for the purpose of
312 : // determining the target output file size, overlap bytes, and expanded
313 : // bytes, taking into account the base level.
314 0 : func adjustedOutputLevel(outputLevel int, baseLevel int) int {
315 0 : if outputLevel == 0 {
316 0 : return 0
317 0 : }
318 0 : if baseLevel == 0 {
319 0 : panic("base level cannot be 0")
320 : }
321 : // Output level is in the range [baseLevel, numLevels). For the purpose of
322 : // determining the target output file size, overlap bytes, and expanded
323 : // bytes, we want to adjust the range to [1, numLevels).
324 0 : return 1 + outputLevel - baseLevel
325 : }
326 :
327 : func newPickedCompactionFromL0(
328 : lcf *manifest.L0CompactionFiles,
329 : opts *Options,
330 : vers *manifest.Version,
331 : l0Organizer *manifest.L0Organizer,
332 : baseLevel int,
333 : isBase bool,
334 1 : ) *pickedTableCompaction {
335 1 : outputLevel := baseLevel
336 1 : if !isBase {
337 1 : outputLevel = 0 // Intra L0
338 1 : }
339 :
340 1 : pc := newPickedTableCompaction(opts, vers, l0Organizer, 0, outputLevel, baseLevel)
341 1 : pc.lcf = lcf
342 1 :
343 1 : // Manually build the compaction as opposed to calling
344 1 : // pickAutoHelper. This is because L0Sublevels has already added
345 1 : // any overlapping L0 SSTables that need to be added, and
346 1 : // because compactions built by L0SSTables do not necessarily
347 1 : // pick contiguous sequences of files in pc.version.Levels[0].
348 1 : pc.startLevel.files = manifest.NewLevelSliceSeqSorted(lcf.Files)
349 1 : return pc
350 : }
351 :
352 0 : func (pc *pickedTableCompaction) String() string {
353 0 : var builder strings.Builder
354 0 : builder.WriteString(fmt.Sprintf(`Score=%f, `, pc.score))
355 0 : builder.WriteString(fmt.Sprintf(`Kind=%s, `, pc.kind))
356 0 : builder.WriteString(fmt.Sprintf(`AdjustedOutputLevel=%d, `, adjustedOutputLevel(pc.outputLevel.level, pc.baseLevel)))
357 0 : builder.WriteString(fmt.Sprintf(`maxOutputFileSize=%d, `, pc.maxOutputFileSize))
358 0 : builder.WriteString(fmt.Sprintf(`maxReadCompactionBytes=%d, `, pc.maxReadCompactionBytes))
359 0 : builder.WriteString(fmt.Sprintf(`bounds=%s, `, pc.bounds))
360 0 : builder.WriteString(fmt.Sprintf(`version=%s, `, pc.version))
361 0 : builder.WriteString(fmt.Sprintf(`inputs=%s, `, pc.inputs))
362 0 : builder.WriteString(fmt.Sprintf(`startlevel=%s, `, pc.startLevel))
363 0 : builder.WriteString(fmt.Sprintf(`outputLevel=%s, `, pc.outputLevel))
364 0 : builder.WriteString(fmt.Sprintf(`l0SublevelInfo=%s, `, pc.startLevel.l0SublevelInfo))
365 0 : builder.WriteString(fmt.Sprintf(`lcf=%s`, pc.lcf))
366 0 : return builder.String()
367 0 : }
368 :
369 : // Clone creates a deep copy of the pickedCompaction
370 1 : func (pc *pickedTableCompaction) clone() *pickedTableCompaction {
371 1 :
372 1 : // Quickly copy over fields that do not require special deep copy care, and
373 1 : // set all fields that will require a deep copy to nil.
374 1 : newPC := &pickedTableCompaction{
375 1 : score: pc.score,
376 1 : kind: pc.kind,
377 1 : baseLevel: pc.baseLevel,
378 1 : maxOutputFileSize: pc.maxOutputFileSize,
379 1 : maxOverlapBytes: pc.maxOverlapBytes,
380 1 : maxReadCompactionBytes: pc.maxReadCompactionBytes,
381 1 : bounds: pc.bounds.Clone(),
382 1 :
383 1 : // TODO(msbutler): properly clone picker metrics
384 1 : pickerMetrics: pc.pickerMetrics,
385 1 :
386 1 : // Both copies see the same manifest, therefore, it's ok for them to share
387 1 : // the same pc.version and pc.l0Organizer.
388 1 : version: pc.version,
389 1 : l0Organizer: pc.l0Organizer,
390 1 : }
391 1 :
392 1 : newPC.inputs = make([]compactionLevel, len(pc.inputs))
393 1 : for i := range pc.inputs {
394 1 : newPC.inputs[i] = pc.inputs[i].Clone()
395 1 : if i == 0 {
396 1 : newPC.startLevel = &newPC.inputs[i]
397 1 : } else if i == len(pc.inputs)-1 {
398 1 : newPC.outputLevel = &newPC.inputs[i]
399 1 : }
400 : }
401 :
402 1 : if len(pc.startLevel.l0SublevelInfo) > 0 {
403 1 : newPC.startLevel.l0SublevelInfo = make([]sublevelInfo, len(pc.startLevel.l0SublevelInfo))
404 1 : for i := range pc.startLevel.l0SublevelInfo {
405 1 : newPC.startLevel.l0SublevelInfo[i] = pc.startLevel.l0SublevelInfo[i].Clone()
406 1 : }
407 : }
408 1 : if pc.lcf != nil {
409 1 : newPC.lcf = pc.lcf.Clone()
410 1 : }
411 1 : return newPC
412 : }
413 :
414 : // setupInputs returns true if a compaction has been set up using the provided inputLevel and
415 : // pc.outputLevel. It returns false if a concurrent compaction is occurring on the start or
416 : // output level files. Note that inputLevel is not necessarily pc.startLevel. In multiLevel
417 : // compactions, inputs are set by calling setupInputs once for each adjacent pair of levels.
418 : // This will preserve level invariants when expanding the compaction. pc.bounds will be updated
419 : // to reflect the key range of the inputs.
420 : func (pc *pickedTableCompaction) setupInputs(
421 : opts *Options,
422 : diskAvailBytes uint64,
423 : inProgressCompactions []compactionInfo,
424 : inputLevel *compactionLevel,
425 : problemSpans *problemspans.ByLevel,
426 1 : ) bool {
427 1 : cmp := opts.Comparer.Compare
428 1 : if !canCompactTables(inputLevel.files, inputLevel.level, problemSpans) {
429 1 : return false
430 1 : }
431 1 : pc.bounds = manifest.ExtendKeyRange(cmp, pc.bounds, inputLevel.files.All())
432 1 :
433 1 : // Setup output files and attempt to grow the inputLevel files with
434 1 : // the expanded key range. No need to do this for intra-L0 compactions;
435 1 : // outputLevel.files is left empty for those.
436 1 : if inputLevel.level != pc.outputLevel.level {
437 1 : // Determine the sstables in the output level which overlap with the compaction
438 1 : // key range.
439 1 : pc.outputLevel.files = pc.version.Overlaps(pc.outputLevel.level, pc.bounds)
440 1 : if !canCompactTables(pc.outputLevel.files, pc.outputLevel.level, problemSpans) {
441 1 : return false
442 1 : }
443 1 : pc.bounds = manifest.ExtendKeyRange(cmp, pc.bounds, pc.outputLevel.files.All())
444 1 :
445 1 : // maxExpandedBytes is the maximum size of an expanded compaction. If
446 1 : // growing a compaction results in a larger size, the original compaction
447 1 : // is used instead.
448 1 : targetFileSize := opts.TargetFileSize(pc.outputLevel.level, pc.baseLevel)
449 1 : maxExpandedBytes := expandedCompactionByteSizeLimit(opts, targetFileSize, diskAvailBytes)
450 1 :
451 1 : // Grow the sstables in inputLevel.level as long as it doesn't affect the number
452 1 : // of sstables included from pc.outputLevel.level.
453 1 : if pc.lcf != nil && inputLevel.level == 0 {
454 1 : pc.growL0ForBase(cmp, maxExpandedBytes)
455 1 : } else if pc.grow(cmp, pc.bounds, maxExpandedBytes, inputLevel, problemSpans) {
456 1 : // inputLevel was expanded, adjust key range if necessary.
457 1 : pc.bounds = manifest.ExtendKeyRange(cmp, pc.bounds, inputLevel.files.All())
458 1 : }
459 : }
460 :
461 1 : if inputLevel.level == 0 {
462 1 : // If L0 is involved, it should always be the startLevel of the compaction.
463 1 : pc.startLevel.l0SublevelInfo = generateSublevelInfo(cmp, pc.startLevel.files)
464 1 : }
465 :
466 1 : return !outputKeyRangeAlreadyCompacting(cmp, inProgressCompactions, pc)
467 : }
468 :
469 : // grow grows the number of inputs at startLevel without changing the number of
470 : // pc.outputLevel files in the compaction, and returns whether the inputs grew. sm
471 : // and la are the smallest and largest InternalKeys in all of the inputs.
472 : func (pc *pickedTableCompaction) grow(
473 : cmp base.Compare,
474 : bounds base.UserKeyBounds,
475 : maxExpandedBytes uint64,
476 : inputLevel *compactionLevel,
477 : problemSpans *problemspans.ByLevel,
478 1 : ) bool {
479 1 : if pc.outputLevel.files.Empty() {
480 1 : return false
481 1 : }
482 1 : expandedInputLevel := pc.version.Overlaps(inputLevel.level, bounds)
483 1 : if !canCompactTables(expandedInputLevel, inputLevel.level, problemSpans) {
484 1 : return false
485 1 : }
486 1 : if expandedInputLevel.Len() <= inputLevel.files.Len() {
487 1 : return false
488 1 : }
489 1 : if expandedInputLevel.AggregateSizeSum()+pc.outputLevel.files.AggregateSizeSum() >= maxExpandedBytes {
490 1 : return false
491 1 : }
492 : // Check that expanding the input level does not change the number of overlapping files in output level.
493 : // We need to include the outputLevel iter because without it, in a multiLevel scenario,
494 : // expandedInputLevel's key range not fully cover all files currently in pc.outputLevel,
495 : // since pc.outputLevel was created using the entire key range which includes higher levels.
496 1 : expandedOutputLevel := pc.version.Overlaps(pc.outputLevel.level,
497 1 : manifest.KeyRange(cmp, expandedInputLevel.All(), pc.outputLevel.files.All()))
498 1 : if expandedOutputLevel.Len() != pc.outputLevel.files.Len() {
499 1 : return false
500 1 : }
501 1 : if !canCompactTables(expandedOutputLevel, pc.outputLevel.level, problemSpans) {
502 0 : return false
503 0 : }
504 1 : inputLevel.files = expandedInputLevel
505 1 : return true
506 : }
507 :
508 : // Similar logic as pc.grow. Additional L0 files are optionally added to the
509 : // compaction at this step. Note that the bounds passed in are not the bounds
510 : // of the compaction, but rather the smallest and largest internal keys that
511 : // the compaction cannot include from L0 without pulling in more Lbase
512 : // files. Consider this example:
513 : //
514 : // L0: c-d e+f g-h
515 : // Lbase: a-b e+f i-j
516 : //
517 : // a b c d e f g h i j
518 : //
519 : // The e-f files have already been chosen in the compaction. As pulling
520 : // in more LBase files is undesirable, the logic below will pass in
521 : // smallest = b and largest = i to ExtendL0ForBaseCompactionTo, which
522 : // will expand the compaction to include c-d and g-h from L0. The
523 : // bounds passed in are exclusive; the compaction cannot be expanded
524 : // to include files that "touch" it.
525 1 : func (pc *pickedTableCompaction) growL0ForBase(cmp base.Compare, maxExpandedBytes uint64) bool {
526 1 : if invariants.Enabled {
527 1 : if pc.startLevel.level != 0 {
528 0 : panic(fmt.Sprintf("pc.startLevel.level is %d, expected 0", pc.startLevel.level))
529 : }
530 : }
531 :
532 1 : if pc.outputLevel.files.Empty() {
533 1 : // If there are no overlapping fields in the output level, we do not
534 1 : // attempt to expand the compaction to encourage move compactions.
535 1 : return false
536 1 : }
537 :
538 1 : smallestBaseKey := base.InvalidInternalKey
539 1 : largestBaseKey := base.InvalidInternalKey
540 1 : // NB: We use Reslice to access the underlying level's files, but
541 1 : // we discard the returned slice. The pc.outputLevel.files slice
542 1 : // is not modified.
543 1 : _ = pc.outputLevel.files.Reslice(func(start, end *manifest.LevelIterator) {
544 1 : if sm := start.Prev(); sm != nil {
545 1 : smallestBaseKey = sm.Largest()
546 1 : }
547 1 : if la := end.Next(); la != nil {
548 1 : largestBaseKey = la.Smallest()
549 1 : }
550 : })
551 1 : oldLcf := pc.lcf.Clone()
552 1 : if !pc.l0Organizer.ExtendL0ForBaseCompactionTo(smallestBaseKey, largestBaseKey, pc.lcf) {
553 1 : return false
554 1 : }
555 :
556 1 : var newStartLevelFiles []*manifest.TableMetadata
557 1 : iter := pc.version.Levels[0].Iter()
558 1 : var sizeSum uint64
559 1 : for j, f := 0, iter.First(); f != nil; j, f = j+1, iter.Next() {
560 1 : if pc.lcf.FilesIncluded[f.L0Index] {
561 1 : newStartLevelFiles = append(newStartLevelFiles, f)
562 1 : sizeSum += f.Size
563 1 : }
564 : }
565 :
566 1 : if sizeSum+pc.outputLevel.files.AggregateSizeSum() >= maxExpandedBytes {
567 1 : *pc.lcf = *oldLcf
568 1 : return false
569 1 : }
570 :
571 1 : pc.startLevel.files = manifest.NewLevelSliceSeqSorted(newStartLevelFiles)
572 1 : pc.bounds = manifest.ExtendKeyRange(cmp, pc.bounds,
573 1 : pc.startLevel.files.All(), pc.outputLevel.files.All())
574 1 : return true
575 : }
576 :
577 : // estimatedInputSize returns an estimate of the size of the compaction's
578 : // inputs, including the estimated physical size of input tables' blob
579 : // references.
580 1 : func (pc *pickedTableCompaction) estimatedInputSize() uint64 {
581 1 : var bytesToCompact uint64
582 1 : for i := range pc.inputs {
583 1 : bytesToCompact += pc.inputs[i].files.AggregateSizeSum()
584 1 : }
585 1 : return bytesToCompact
586 : }
587 :
588 : // setupMultiLevelCandidate returns true if it successfully added another level
589 : // to the compaction.
590 1 : func (pc *pickedTableCompaction) setupMultiLevelCandidate(opts *Options, env compactionEnv) bool {
591 1 : pc.inputs = append(pc.inputs, compactionLevel{level: pc.outputLevel.level + 1})
592 1 :
593 1 : // Recalibrate startLevel and outputLevel:
594 1 : // - startLevel and outputLevel pointers may be obsolete after appending to pc.inputs.
595 1 : // - push outputLevel to extraLevels and move the new level to outputLevel
596 1 : pc.startLevel = &pc.inputs[0]
597 1 : pc.outputLevel = &pc.inputs[2]
598 1 : return pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, &pc.inputs[1], nil /* TODO(radu) */)
599 1 : }
600 :
601 : // canCompactTables returns true if the tables in the level slice are not
602 : // compacting already and don't intersect any problem spans.
603 : func canCompactTables(
604 : inputs manifest.LevelSlice, level int, problemSpans *problemspans.ByLevel,
605 1 : ) bool {
606 1 : for f := range inputs.All() {
607 1 : if f.IsCompacting() {
608 1 : return false
609 1 : }
610 1 : if problemSpans != nil && problemSpans.Overlaps(level, f.UserKeyBounds()) {
611 0 : return false
612 0 : }
613 : }
614 1 : return true
615 : }
616 :
617 : // newCompactionPickerByScore creates a compactionPickerByScore associated with
618 : // the newest version. The picker is used under logLock (until a new version is
619 : // installed).
620 : func newCompactionPickerByScore(
621 : v *manifest.Version,
622 : lvs *latestVersionState,
623 : opts *Options,
624 : inProgressCompactions []compactionInfo,
625 1 : ) *compactionPickerByScore {
626 1 : p := &compactionPickerByScore{
627 1 : opts: opts,
628 1 : vers: v,
629 1 : latestVersionState: lvs,
630 1 : }
631 1 : p.initLevelMaxBytes(inProgressCompactions)
632 1 : return p
633 1 : }
634 :
635 : // Information about a candidate compaction level that has been identified by
636 : // the compaction picker.
637 : type candidateLevelInfo struct {
638 : // The fill factor of the level, calculated using uncompensated file sizes and
639 : // without any adjustments. A factor > 1 means that the level has more data
640 : // than the ideal size for that level.
641 : //
642 : // For L0, the fill factor is calculated based on the number of sublevels
643 : // (see calculateL0FillFactor).
644 : //
645 : // For L1+, the fill factor is the ratio between the total uncompensated file
646 : // size and the ideal size of the level (based on the total size of the DB).
647 : fillFactor float64
648 :
649 : // The score of the level, used to rank levels.
650 : //
651 : // If the level doesn't require compaction, the score is 0. Otherwise:
652 : // - for L6 the score is equal to the fillFactor;
653 : // - for L0-L5:
654 : // - if the fillFactor is < 1: the score is equal to the fillFactor;
655 : // - if the fillFactor is >= 1: the score is the ratio between the
656 : // fillFactor and the next level's fillFactor.
657 : score float64
658 :
659 : // The fill factor of the level after accounting for level size compensation.
660 : //
661 : // For L0, the compensatedFillFactor is equal to the fillFactor as we don't
662 : // account for level size compensation in L0.
663 : //
664 : // For l1+, the compensatedFillFactor takes into account the estimated
665 : // savings in the lower levels because of deletions.
666 : //
667 : // The compensated fill factor is used to determine if the level should be
668 : // compacted (see calculateLevelScores).
669 : compensatedFillFactor float64
670 :
671 : level int
672 : // The level to compact to.
673 : outputLevel int
674 : // The file in level that will be compacted. Additional files may be
675 : // picked by the compaction, and a pickedCompaction created for the
676 : // compaction.
677 : file manifest.LevelFile
678 : }
679 :
680 1 : func (c *candidateLevelInfo) shouldCompact() bool {
681 1 : return c.score > 0
682 1 : }
683 :
684 1 : func tableTombstoneCompensation(t *manifest.TableMetadata) uint64 {
685 1 : return t.Stats.PointDeletionsBytesEstimate + t.Stats.RangeDeletionsBytesEstimate
686 1 : }
687 :
688 : // tableCompensatedSize returns t's size, including an estimate of the physical
689 : // size of its external references, and inflated according to compaction
690 : // priorities.
691 1 : func tableCompensatedSize(t *manifest.TableMetadata) uint64 {
692 1 : // Add in the estimate of disk space that may be reclaimed by compacting the
693 1 : // table's tombstones.
694 1 : return t.Size + t.EstimatedReferenceSize() + tableTombstoneCompensation(t)
695 1 : }
696 :
697 : // totalCompensatedSize computes the compensated size over a table metadata
698 : // iterator. Note that this function is linear in the files available to the
699 : // iterator. Use the compensatedSizeAnnotator if querying the total
700 : // compensated size of a level.
701 1 : func totalCompensatedSize(iter iter.Seq[*manifest.TableMetadata]) uint64 {
702 1 : var sz uint64
703 1 : for f := range iter {
704 1 : sz += tableCompensatedSize(f)
705 1 : }
706 1 : return sz
707 : }
708 :
709 : // compactionPickerByScore holds the state and logic for picking a compaction. A
710 : // compaction picker is associated with a single version. A new compaction
711 : // picker is created and initialized every time a new version is installed.
712 : type compactionPickerByScore struct {
713 : opts *Options
714 : vers *manifest.Version
715 : // Unlike vers, which is immutable and the latest version when this picker
716 : // is created, latestVersionState represents the mutable state of the latest
717 : // version. This means that at some point in the future a
718 : // compactionPickerByScore created in the past will have mutually
719 : // inconsistent state in vers and latestVersionState. This is not a problem
720 : // since (a) a new picker is created in UpdateVersionLocked when a new
721 : // version is installed, and (b) only the latest picker is used for picking
722 : // compactions. This is ensured by holding versionSet.logLock for both (a)
723 : // and (b).
724 : latestVersionState *latestVersionState
725 : // The level to target for L0 compactions. Levels L1 to baseLevel must be
726 : // empty.
727 : baseLevel int
728 : // levelMaxBytes holds the dynamically adjusted max bytes setting for each
729 : // level.
730 : levelMaxBytes [numLevels]int64
731 : dbSizeBytes uint64
732 : }
733 :
734 : var _ compactionPicker = &compactionPickerByScore{}
735 :
736 1 : func (p *compactionPickerByScore) getMetrics(inProgress []compactionInfo) compactionPickerMetrics {
737 1 : var m compactionPickerMetrics
738 1 : for _, info := range p.calculateLevelScores(inProgress) {
739 1 : m.levels[info.level].score = info.score
740 1 : m.levels[info.level].fillFactor = info.fillFactor
741 1 : m.levels[info.level].compensatedFillFactor = info.compensatedFillFactor
742 1 : }
743 1 : return m
744 : }
745 :
746 1 : func (p *compactionPickerByScore) getBaseLevel() int {
747 1 : if p == nil {
748 0 : return 1
749 0 : }
750 1 : return p.baseLevel
751 : }
752 :
753 : // estimatedCompactionDebt estimates the number of bytes which need to be
754 : // compacted before the LSM tree becomes stable.
755 1 : func (p *compactionPickerByScore) estimatedCompactionDebt() uint64 {
756 1 : if p == nil {
757 0 : return 0
758 0 : }
759 :
760 : // We assume that all the bytes in L0 need to be compacted to Lbase. This is
761 : // unlike the RocksDB logic that figures out whether L0 needs compaction.
762 1 : bytesAddedToNextLevel := p.vers.Levels[0].AggregateSize()
763 1 : lbaseSize := p.vers.Levels[p.baseLevel].AggregateSize()
764 1 :
765 1 : var compactionDebt uint64
766 1 : if bytesAddedToNextLevel > 0 && lbaseSize > 0 {
767 1 : // We only incur compaction debt if both L0 and Lbase contain data. If L0
768 1 : // is empty, no compaction is necessary. If Lbase is empty, a move-based
769 1 : // compaction from L0 would occur.
770 1 : compactionDebt += bytesAddedToNextLevel + lbaseSize
771 1 : }
772 :
773 : // loop invariant: At the beginning of the loop, bytesAddedToNextLevel is the
774 : // bytes added to `level` in the loop.
775 1 : for level := p.baseLevel; level < numLevels-1; level++ {
776 1 : levelSize := p.vers.Levels[level].AggregateSize() + bytesAddedToNextLevel
777 1 : nextLevelSize := p.vers.Levels[level+1].AggregateSize()
778 1 : if levelSize > uint64(p.levelMaxBytes[level]) {
779 1 : bytesAddedToNextLevel = levelSize - uint64(p.levelMaxBytes[level])
780 1 : if nextLevelSize > 0 {
781 1 : // We only incur compaction debt if the next level contains data. If the
782 1 : // next level is empty, a move-based compaction would be used.
783 1 : levelRatio := float64(nextLevelSize) / float64(levelSize)
784 1 : // The current level contributes bytesAddedToNextLevel to compactions.
785 1 : // The next level contributes levelRatio * bytesAddedToNextLevel.
786 1 : compactionDebt += uint64(float64(bytesAddedToNextLevel) * (levelRatio + 1))
787 1 : }
788 1 : } else {
789 1 : // We're not moving any bytes to the next level.
790 1 : bytesAddedToNextLevel = 0
791 1 : }
792 : }
793 1 : return compactionDebt
794 : }
795 :
796 1 : func (p *compactionPickerByScore) initLevelMaxBytes(inProgressCompactions []compactionInfo) {
797 1 : // The levelMaxBytes calculations here differ from RocksDB in two ways:
798 1 : //
799 1 : // 1. The use of dbSize vs maxLevelSize. RocksDB uses the size of the maximum
800 1 : // level in L1-L6, rather than determining the size of the bottom level
801 1 : // based on the total amount of data in the dB. The RocksDB calculation is
802 1 : // problematic if L0 contains a significant fraction of data, or if the
803 1 : // level sizes are roughly equal and thus there is a significant fraction
804 1 : // of data outside of the largest level.
805 1 : //
806 1 : // 2. Not adjusting the size of Lbase based on L0. RocksDB computes
807 1 : // baseBytesMax as the maximum of the configured LBaseMaxBytes and the
808 1 : // size of L0. This is problematic because baseBytesMax is used to compute
809 1 : // the max size of lower levels. A very large baseBytesMax will result in
810 1 : // an overly large value for the size of lower levels which will caused
811 1 : // those levels not to be compacted even when they should be
812 1 : // compacted. This often results in "inverted" LSM shapes where Ln is
813 1 : // larger than Ln+1.
814 1 :
815 1 : // Determine the first non-empty level and the total DB size.
816 1 : firstNonEmptyLevel := -1
817 1 : var dbSize uint64
818 1 : for level := 1; level < numLevels; level++ {
819 1 : if p.vers.Levels[level].AggregateSize() > 0 {
820 1 : if firstNonEmptyLevel == -1 {
821 1 : firstNonEmptyLevel = level
822 1 : }
823 1 : dbSize += p.vers.Levels[level].AggregateSize()
824 : }
825 : }
826 1 : for _, c := range inProgressCompactions {
827 1 : if c.outputLevel == 0 || c.outputLevel == -1 {
828 1 : continue
829 : }
830 1 : if c.inputs[0].level == 0 && (firstNonEmptyLevel == -1 || c.outputLevel < firstNonEmptyLevel) {
831 1 : firstNonEmptyLevel = c.outputLevel
832 1 : }
833 : }
834 :
835 : // Initialize the max-bytes setting for each level to "infinity" which will
836 : // disallow compaction for that level. We'll fill in the actual value below
837 : // for levels we want to allow compactions from.
838 1 : for level := 0; level < numLevels; level++ {
839 1 : p.levelMaxBytes[level] = math.MaxInt64
840 1 : }
841 :
842 1 : dbSizeBelowL0 := dbSize
843 1 : dbSize += p.vers.Levels[0].AggregateSize()
844 1 : p.dbSizeBytes = dbSize
845 1 : if dbSizeBelowL0 == 0 {
846 1 : // No levels for L1 and up contain any data. Target L0 compactions for the
847 1 : // last level or to the level to which there is an ongoing L0 compaction.
848 1 : p.baseLevel = numLevels - 1
849 1 : if firstNonEmptyLevel >= 0 {
850 1 : p.baseLevel = firstNonEmptyLevel
851 1 : }
852 1 : return
853 : }
854 :
855 1 : bottomLevelSize := dbSize - dbSize/uint64(p.opts.Experimental.LevelMultiplier)
856 1 :
857 1 : curLevelSize := bottomLevelSize
858 1 : for level := numLevels - 2; level >= firstNonEmptyLevel; level-- {
859 1 : curLevelSize = uint64(float64(curLevelSize) / float64(p.opts.Experimental.LevelMultiplier))
860 1 : }
861 :
862 : // Compute base level (where L0 data is compacted to).
863 1 : baseBytesMax := uint64(p.opts.LBaseMaxBytes)
864 1 : p.baseLevel = firstNonEmptyLevel
865 1 : for p.baseLevel > 1 && curLevelSize > baseBytesMax {
866 1 : p.baseLevel--
867 1 : curLevelSize = uint64(float64(curLevelSize) / float64(p.opts.Experimental.LevelMultiplier))
868 1 : }
869 :
870 1 : smoothedLevelMultiplier := 1.0
871 1 : if p.baseLevel < numLevels-1 {
872 1 : smoothedLevelMultiplier = math.Pow(
873 1 : float64(bottomLevelSize)/float64(baseBytesMax),
874 1 : 1.0/float64(numLevels-p.baseLevel-1))
875 1 : }
876 :
877 1 : levelSize := float64(baseBytesMax)
878 1 : for level := p.baseLevel; level < numLevels; level++ {
879 1 : if level > p.baseLevel && levelSize > 0 {
880 1 : levelSize *= smoothedLevelMultiplier
881 1 : }
882 : // Round the result since test cases use small target level sizes, which
883 : // can be impacted by floating-point imprecision + integer truncation.
884 1 : roundedLevelSize := math.Round(levelSize)
885 1 : if roundedLevelSize > float64(math.MaxInt64) {
886 0 : p.levelMaxBytes[level] = math.MaxInt64
887 1 : } else {
888 1 : p.levelMaxBytes[level] = int64(roundedLevelSize)
889 1 : }
890 : }
891 : }
892 :
893 : type levelSizeAdjust struct {
894 : incomingActualBytes uint64
895 : outgoingActualBytes uint64
896 : outgoingCompensatedBytes uint64
897 : }
898 :
899 1 : func (a levelSizeAdjust) compensated() uint64 {
900 1 : return a.incomingActualBytes - a.outgoingCompensatedBytes
901 1 : }
902 :
903 1 : func (a levelSizeAdjust) actual() uint64 {
904 1 : return a.incomingActualBytes - a.outgoingActualBytes
905 1 : }
906 :
907 1 : func calculateSizeAdjust(inProgressCompactions []compactionInfo) [numLevels]levelSizeAdjust {
908 1 : // Compute size adjustments for each level based on the in-progress
909 1 : // compactions. We sum the file sizes of all files leaving and entering each
910 1 : // level in in-progress compactions. For outgoing files, we also sum a
911 1 : // separate sum of 'compensated file sizes', which are inflated according
912 1 : // to deletion estimates.
913 1 : //
914 1 : // When we adjust a level's size according to these values during score
915 1 : // calculation, we subtract the compensated size of start level inputs to
916 1 : // account for the fact that score calculation uses compensated sizes.
917 1 : //
918 1 : // Since compensated file sizes may be compensated because they reclaim
919 1 : // space from the output level's files, we only add the real file size to
920 1 : // the output level.
921 1 : //
922 1 : // This is slightly different from RocksDB's behavior, which simply elides
923 1 : // compacting files from the level size calculation.
924 1 : var sizeAdjust [numLevels]levelSizeAdjust
925 1 : for i := range inProgressCompactions {
926 1 : c := &inProgressCompactions[i]
927 1 : // If this compaction's version edit has already been applied, there's
928 1 : // no need to adjust: The LSM we'll examine will already reflect the
929 1 : // new LSM state.
930 1 : if c.versionEditApplied {
931 1 : continue
932 : }
933 :
934 1 : for _, input := range c.inputs {
935 1 : actualSize := input.files.AggregateSizeSum()
936 1 : compensatedSize := totalCompensatedSize(input.files.All())
937 1 :
938 1 : if input.level != c.outputLevel {
939 1 : sizeAdjust[input.level].outgoingCompensatedBytes += compensatedSize
940 1 : sizeAdjust[input.level].outgoingActualBytes += actualSize
941 1 : if c.outputLevel != -1 {
942 1 : sizeAdjust[c.outputLevel].incomingActualBytes += actualSize
943 1 : }
944 : }
945 : }
946 : }
947 1 : return sizeAdjust
948 : }
949 :
950 : // calculateLevelScores calculates the candidateLevelInfo for all levels and
951 : // returns them in decreasing score order.
952 : func (p *compactionPickerByScore) calculateLevelScores(
953 : inProgressCompactions []compactionInfo,
954 1 : ) [numLevels]candidateLevelInfo {
955 1 : var scores [numLevels]candidateLevelInfo
956 1 : for i := range scores {
957 1 : scores[i].level = i
958 1 : scores[i].outputLevel = i + 1
959 1 : }
960 1 : l0FillFactor := calculateL0FillFactor(p.vers, p.latestVersionState.l0Organizer, p.opts, inProgressCompactions)
961 1 : scores[0] = candidateLevelInfo{
962 1 : outputLevel: p.baseLevel,
963 1 : fillFactor: l0FillFactor,
964 1 : compensatedFillFactor: l0FillFactor, // No compensation for L0.
965 1 : }
966 1 : sizeAdjust := calculateSizeAdjust(inProgressCompactions)
967 1 : for level := 1; level < numLevels; level++ {
968 1 : compensatedLevelSize :=
969 1 : // Actual file size.
970 1 : p.vers.Levels[level].AggregateSize() +
971 1 : // Point deletions.
972 1 : *pointDeletionsBytesEstimateAnnotator.LevelAnnotation(p.vers.Levels[level]) +
973 1 : // Range deletions.
974 1 : *rangeDeletionsBytesEstimateAnnotator.LevelAnnotation(p.vers.Levels[level]) +
975 1 : // Adjustments for in-progress compactions.
976 1 : sizeAdjust[level].compensated()
977 1 : scores[level].compensatedFillFactor = float64(compensatedLevelSize) / float64(p.levelMaxBytes[level])
978 1 : scores[level].fillFactor = float64(p.vers.Levels[level].AggregateSize()+sizeAdjust[level].actual()) / float64(p.levelMaxBytes[level])
979 1 : }
980 :
981 : // Adjust each level's fill factor by the fill factor of the next level to get
982 : // an (uncompensated) score; and each level's compensated fill factor by the
983 : // fill factor of the next level to get a compensated score.
984 : //
985 : // The compensated score is used to determine if the level should be compacted
986 : // at all. The (uncompensated) score is used as the value used to rank levels.
987 : //
988 : // If the next level has a high fill factor, and is thus a priority for
989 : // compaction, this reduces the priority for compacting the current level. If
990 : // the next level has a low fill factor (i.e. it is below its target size),
991 : // this increases the priority for compacting the current level.
992 : //
993 : // The effect of this adjustment is to help prioritize compactions in lower
994 : // levels. The following example shows the scores and the fill factors. In this
995 : // scenario, L0 has 68 sublevels. L3 (a.k.a. Lbase) is significantly above its
996 : // target size. The original score prioritizes compactions from those two
997 : // levels, but doing so ends up causing a future problem: data piles up in the
998 : // higher levels, starving L5->L6 compactions, and to a lesser degree starving
999 : // L4->L5 compactions.
1000 : //
1001 : // Note that in the example shown there is no level size compensation so the
1002 : // compensatedFillFactor and fillFactor are the same for each level.
1003 : //
1004 : // score fillFactor compensatedFillFactor size max-size
1005 : // L0 3.2 68.0 68.0 2.2 G -
1006 : // L3 3.2 21.1 21.1 1.3 G 64 M
1007 : // L4 3.4 6.7 6.7 3.1 G 467 M
1008 : // L5 3.4 2.0 2.0 6.6 G 3.3 G
1009 : // L6 0 0.6 0.6 14 G 24 G
1010 : //
1011 : // TODO(radu): the way compensation works needs some rethinking. For example,
1012 : // if compacting L5 can free up a lot of space in L6, the score of L5 should
1013 : // go *up* with the fill factor of L6, not the other way around.
1014 1 : for level := 0; level < numLevels; level++ {
1015 1 : if level > 0 && level < p.baseLevel {
1016 1 : continue
1017 : }
1018 1 : const compensatedFillFactorThreshold = 1.0
1019 1 : if scores[level].compensatedFillFactor < compensatedFillFactorThreshold {
1020 1 : // No need to compact this level; score stays 0.
1021 1 : continue
1022 : }
1023 1 : score := scores[level].fillFactor
1024 1 : compensatedScore := scores[level].compensatedFillFactor
1025 1 : if level < numLevels-1 {
1026 1 : nextLevel := scores[level].outputLevel
1027 1 : // Avoid absurdly large scores by placing a floor on the factor that we'll
1028 1 : // adjust a level by. The value of 0.01 was chosen somewhat arbitrarily.
1029 1 : denominator := max(0.01, scores[nextLevel].fillFactor)
1030 1 : score /= denominator
1031 1 : compensatedScore /= denominator
1032 1 : }
1033 : // The level requires compaction iff both compensatedFillFactor and
1034 : // compensatedScore are >= 1.0.
1035 : //
1036 : // TODO(radu): this seems ad-hoc. In principle, the state of other levels
1037 : // should not come into play when we're determining this level's eligibility
1038 : // for compaction. The score should take care of correctly prioritizing the
1039 : // levels.
1040 1 : const compensatedScoreThreshold = 1.0
1041 1 : if compensatedScore < compensatedScoreThreshold {
1042 1 : // No need to compact this level; score stays 0.
1043 1 : continue
1044 : }
1045 1 : scores[level].score = score
1046 : }
1047 : // Sort by score (decreasing) and break ties by level (increasing).
1048 1 : slices.SortFunc(scores[:], func(a, b candidateLevelInfo) int {
1049 1 : if a.score != b.score {
1050 1 : return cmp.Compare(b.score, a.score)
1051 1 : }
1052 1 : return cmp.Compare(a.level, b.level)
1053 : })
1054 1 : return scores
1055 : }
1056 :
1057 : // calculateL0FillFactor calculates a float value representing the relative
1058 : // priority of compacting L0. A value less than 1 indicates that L0 does not
1059 : // need any compactions.
1060 : //
1061 : // L0 is special in that files within L0 may overlap one another, so a different
1062 : // set of heuristics that take into account read amplification apply.
1063 : func calculateL0FillFactor(
1064 : vers *manifest.Version,
1065 : l0Organizer *manifest.L0Organizer,
1066 : opts *Options,
1067 : inProgressCompactions []compactionInfo,
1068 1 : ) float64 {
1069 1 : // Use the sublevel count to calculate the score. The base vs intra-L0
1070 1 : // compaction determination happens in pickAuto, not here.
1071 1 : score := float64(2*l0Organizer.MaxDepthAfterOngoingCompactions()) /
1072 1 : float64(opts.L0CompactionThreshold)
1073 1 :
1074 1 : // Also calculate a score based on the file count but use it only if it
1075 1 : // produces a higher score than the sublevel-based one. This heuristic is
1076 1 : // designed to accommodate cases where L0 is accumulating non-overlapping
1077 1 : // files in L0. Letting too many non-overlapping files accumulate in few
1078 1 : // sublevels is undesirable, because:
1079 1 : // 1) we can produce a massive backlog to compact once files do overlap.
1080 1 : // 2) constructing L0 sublevels has a runtime that grows superlinearly with
1081 1 : // the number of files in L0 and must be done while holding D.mu.
1082 1 : noncompactingFiles := vers.Levels[0].Len()
1083 1 : for _, c := range inProgressCompactions {
1084 1 : for _, cl := range c.inputs {
1085 1 : if cl.level == 0 {
1086 1 : noncompactingFiles -= cl.files.Len()
1087 1 : }
1088 : }
1089 : }
1090 1 : fileScore := float64(noncompactingFiles) / float64(opts.L0CompactionFileThreshold)
1091 1 : if score < fileScore {
1092 1 : score = fileScore
1093 1 : }
1094 1 : return score
1095 : }
1096 :
1097 : // pickCompactionSeedFile picks a file from `level` in the `vers` to build a
1098 : // compaction around. Currently, this function implements a heuristic similar to
1099 : // RocksDB's kMinOverlappingRatio, seeking to minimize write amplification. This
1100 : // function is linear with respect to the number of files in `level` and
1101 : // `outputLevel`.
1102 : func pickCompactionSeedFile(
1103 : vers *manifest.Version,
1104 : virtualBackings *manifest.VirtualBackings,
1105 : opts *Options,
1106 : level, outputLevel int,
1107 : earliestSnapshotSeqNum base.SeqNum,
1108 : problemSpans *problemspans.ByLevel,
1109 1 : ) (manifest.LevelFile, bool) {
1110 1 : // Select the file within the level to compact. We want to minimize write
1111 1 : // amplification, but also ensure that (a) deletes are propagated to the
1112 1 : // bottom level in a timely fashion, and (b) virtual sstables that are
1113 1 : // pinning backing sstables where most of the data is garbage are compacted
1114 1 : // away. Doing (a) and (b) reclaims disk space. A table's smallest sequence
1115 1 : // number provides a measure of its age. The ratio of overlapping-bytes /
1116 1 : // table-size gives an indication of write amplification (a smaller ratio is
1117 1 : // preferrable).
1118 1 : //
1119 1 : // The current heuristic is based off the RocksDB kMinOverlappingRatio
1120 1 : // heuristic. It chooses the file with the minimum overlapping ratio with
1121 1 : // the target level, which minimizes write amplification.
1122 1 : //
1123 1 : // The heuristic uses a "compensated size" for the denominator, which is the
1124 1 : // file size inflated by (a) an estimate of the space that may be reclaimed
1125 1 : // through compaction, and (b) a fraction of the amount of garbage in the
1126 1 : // backing sstable pinned by this (virtual) sstable.
1127 1 : //
1128 1 : // TODO(peter): For concurrent compactions, we may want to try harder to
1129 1 : // pick a seed file whose resulting compaction bounds do not overlap with
1130 1 : // an in-progress compaction.
1131 1 :
1132 1 : cmp := opts.Comparer.Compare
1133 1 : startIter := vers.Levels[level].Iter()
1134 1 : outputIter := vers.Levels[outputLevel].Iter()
1135 1 :
1136 1 : var file manifest.LevelFile
1137 1 : smallestRatio := uint64(math.MaxUint64)
1138 1 :
1139 1 : outputFile := outputIter.First()
1140 1 :
1141 1 : for f := startIter.First(); f != nil; f = startIter.Next() {
1142 1 : var overlappingBytes uint64
1143 1 : if f.IsCompacting() {
1144 1 : // Move on if this file is already being compacted. We'll likely
1145 1 : // still need to move past the overlapping output files regardless,
1146 1 : // but in cases where all start-level files are compacting we won't.
1147 1 : continue
1148 : }
1149 1 : if problemSpans != nil && problemSpans.Overlaps(level, f.UserKeyBounds()) {
1150 0 : // File touches problem span which temporarily disallows auto compactions.
1151 0 : continue
1152 : }
1153 :
1154 : // Trim any output-level files smaller than f.
1155 1 : for outputFile != nil && sstableKeyCompare(cmp, outputFile.Largest(), f.Smallest()) < 0 {
1156 1 : outputFile = outputIter.Next()
1157 1 : }
1158 :
1159 1 : skip := false
1160 1 : for outputFile != nil && sstableKeyCompare(cmp, outputFile.Smallest(), f.Largest()) <= 0 {
1161 1 : overlappingBytes += outputFile.Size
1162 1 : if outputFile.IsCompacting() {
1163 1 : // If one of the overlapping files is compacting, we're not going to be
1164 1 : // able to compact f anyway, so skip it.
1165 1 : skip = true
1166 1 : break
1167 : }
1168 1 : if problemSpans != nil && problemSpans.Overlaps(outputLevel, outputFile.UserKeyBounds()) {
1169 0 : // Overlapping file touches problem span which temporarily disallows auto compactions.
1170 0 : skip = true
1171 0 : break
1172 : }
1173 :
1174 : // For files in the bottommost level of the LSM, the
1175 : // Stats.RangeDeletionsBytesEstimate field is set to the estimate
1176 : // of bytes /within/ the file itself that may be dropped by
1177 : // recompacting the file. These bytes from obsolete keys would not
1178 : // need to be rewritten if we compacted `f` into `outputFile`, so
1179 : // they don't contribute to write amplification. Subtracting them
1180 : // out of the overlapping bytes helps prioritize these compactions
1181 : // that are cheaper than their file sizes suggest.
1182 1 : if outputLevel == numLevels-1 && outputFile.LargestSeqNum < earliestSnapshotSeqNum {
1183 1 : overlappingBytes -= outputFile.Stats.RangeDeletionsBytesEstimate
1184 1 : }
1185 :
1186 : // If the file in the next level extends beyond f's largest key,
1187 : // break out and don't advance outputIter because f's successor
1188 : // might also overlap.
1189 : //
1190 : // Note, we stop as soon as we encounter an output-level file with a
1191 : // largest key beyond the input-level file's largest bound. We
1192 : // perform a simple user key comparison here using sstableKeyCompare
1193 : // which handles the potential for exclusive largest key bounds.
1194 : // There's some subtlety when the bounds are equal (eg, equal and
1195 : // inclusive, or equal and exclusive). Current Pebble doesn't split
1196 : // user keys across sstables within a level (and in format versions
1197 : // FormatSplitUserKeysMarkedCompacted and later we guarantee no
1198 : // split user keys exist within the entire LSM). In that case, we're
1199 : // assured that neither the input level nor the output level's next
1200 : // file shares the same user key, so compaction expansion will not
1201 : // include them in any compaction compacting `f`.
1202 : //
1203 : // NB: If we /did/ allow split user keys, or we're running on an
1204 : // old database with an earlier format major version where there are
1205 : // existing split user keys, this logic would be incorrect. Consider
1206 : // L1: [a#120,a#100] [a#80,a#60]
1207 : // L2: [a#55,a#45] [a#35,a#25] [a#15,a#5]
1208 : // While considering the first file in L1, [a#120,a#100], we'd skip
1209 : // past all of the files in L2. When considering the second file in
1210 : // L1, we'd improperly conclude that the second file overlaps
1211 : // nothing in the second level and is cheap to compact, when in
1212 : // reality we'd need to expand the compaction to include all 5
1213 : // files.
1214 1 : if sstableKeyCompare(cmp, outputFile.Largest(), f.Largest()) > 0 {
1215 1 : break
1216 : }
1217 1 : outputFile = outputIter.Next()
1218 : }
1219 1 : if skip {
1220 1 : continue
1221 : }
1222 :
1223 1 : compSz := tableCompensatedSize(f) + responsibleForGarbageBytes(virtualBackings, f)
1224 1 : scaledRatio := overlappingBytes * 1024 / compSz
1225 1 : if scaledRatio < smallestRatio {
1226 1 : smallestRatio = scaledRatio
1227 1 : file = startIter.Take()
1228 1 : }
1229 : }
1230 1 : return file, file.TableMetadata != nil
1231 : }
1232 :
1233 : // responsibleForGarbageBytes returns the amount of garbage in the backing
1234 : // sstable that we consider the responsibility of this virtual sstable. For
1235 : // non-virtual sstables, this is of course 0. For virtual sstables, we equally
1236 : // distribute the responsibility of the garbage across all the virtual
1237 : // sstables that are referencing the same backing sstable. One could
1238 : // alternatively distribute this in proportion to the virtual sst sizes, but
1239 : // it isn't clear that more sophisticated heuristics are worth it, given that
1240 : // the garbage cannot be reclaimed until all the referencing virtual sstables
1241 : // are compacted.
1242 : func responsibleForGarbageBytes(
1243 : virtualBackings *manifest.VirtualBackings, m *manifest.TableMetadata,
1244 1 : ) uint64 {
1245 1 : if !m.Virtual {
1246 1 : return 0
1247 1 : }
1248 1 : useCount, virtualizedSize := virtualBackings.Usage(m.TableBacking.DiskFileNum)
1249 1 : // Since virtualizedSize is the sum of the estimated size of all virtual
1250 1 : // ssts, we allow for the possibility that virtualizedSize could exceed
1251 1 : // m.TableBacking.Size.
1252 1 : totalGarbage := int64(m.TableBacking.Size) - int64(virtualizedSize)
1253 1 : if totalGarbage <= 0 {
1254 1 : return 0
1255 1 : }
1256 1 : if useCount == 0 {
1257 0 : // This cannot happen if m exists in the latest version. The call to
1258 0 : // ResponsibleForGarbageBytes during compaction picking ensures that m
1259 0 : // exists in the latest version by holding versionSet.logLock.
1260 0 : panic(errors.AssertionFailedf("%s has zero useCount", m.String()))
1261 : }
1262 1 : return uint64(totalGarbage) / uint64(useCount)
1263 : }
1264 :
1265 1 : func (p *compactionPickerByScore) getCompactionConcurrency() int {
1266 1 : lower, upper := p.opts.CompactionConcurrencyRange()
1267 1 : if lower >= upper {
1268 1 : return upper
1269 1 : }
1270 : // Compaction concurrency is controlled by L0 read-amp. We allow one
1271 : // additional compaction per L0CompactionConcurrency sublevels, as well as
1272 : // one additional compaction per CompactionDebtConcurrency bytes of
1273 : // compaction debt. Compaction concurrency is tied to L0 sublevels as that
1274 : // signal is independent of the database size. We tack on the compaction
1275 : // debt as a second signal to prevent compaction concurrency from dropping
1276 : // significantly right after a base compaction finishes, and before those
1277 : // bytes have been compacted further down the LSM.
1278 : //
1279 : // Let n be the number of in-progress compactions.
1280 : //
1281 : // l0ReadAmp >= ccSignal1 then can run another compaction, where
1282 : // ccSignal1 = n * p.opts.Experimental.L0CompactionConcurrency
1283 : // Rearranging,
1284 : // n <= l0ReadAmp / p.opts.Experimental.L0CompactionConcurrency.
1285 : // So we can run up to
1286 : // l0ReadAmp / p.opts.Experimental.L0CompactionConcurrency extra compactions.
1287 1 : l0ReadAmpCompactions := 0
1288 1 : if p.opts.Experimental.L0CompactionConcurrency > 0 {
1289 1 : l0ReadAmp := p.latestVersionState.l0Organizer.MaxDepthAfterOngoingCompactions()
1290 1 : l0ReadAmpCompactions = (l0ReadAmp / p.opts.Experimental.L0CompactionConcurrency)
1291 1 : }
1292 : // compactionDebt >= ccSignal2 then can run another compaction, where
1293 : // ccSignal2 = uint64(n) * p.opts.Experimental.CompactionDebtConcurrency
1294 : // Rearranging,
1295 : // n <= compactionDebt / p.opts.Experimental.CompactionDebtConcurrency
1296 : // So we can run up to
1297 : // compactionDebt / p.opts.Experimental.CompactionDebtConcurrency extra
1298 : // compactions.
1299 1 : compactionDebtCompactions := 0
1300 1 : if p.opts.Experimental.CompactionDebtConcurrency > 0 {
1301 1 : compactionDebt := p.estimatedCompactionDebt()
1302 1 : compactionDebtCompactions = int(compactionDebt / p.opts.Experimental.CompactionDebtConcurrency)
1303 1 : }
1304 :
1305 1 : compactableGarbageCompactions := 0
1306 1 : garbageFractionLimit := p.opts.Experimental.CompactionGarbageFractionForMaxConcurrency()
1307 1 : if garbageFractionLimit > 0 && p.dbSizeBytes > 0 {
1308 1 : compactableGarbageBytes :=
1309 1 : *pointDeletionsBytesEstimateAnnotator.MultiLevelAnnotation(p.vers.Levels[:]) +
1310 1 : *rangeDeletionsBytesEstimateAnnotator.MultiLevelAnnotation(p.vers.Levels[:])
1311 1 : garbageFraction := float64(compactableGarbageBytes) / float64(p.dbSizeBytes)
1312 1 : compactableGarbageCompactions =
1313 1 : int((garbageFraction / garbageFractionLimit) * float64(upper-lower))
1314 1 : }
1315 :
1316 1 : extraCompactions := max(l0ReadAmpCompactions, compactionDebtCompactions, compactableGarbageCompactions, 0)
1317 1 :
1318 1 : return min(lower+extraCompactions, upper)
1319 : }
1320 :
1321 : // TODO(sumeer): remove unless someone actually finds this useful.
1322 : func (p *compactionPickerByScore) logCompactionForTesting(
1323 : env compactionEnv, scores [numLevels]candidateLevelInfo, pc *pickedTableCompaction,
1324 0 : ) {
1325 0 : var buf bytes.Buffer
1326 0 : for i := 0; i < numLevels; i++ {
1327 0 : if i != 0 && i < p.baseLevel {
1328 0 : continue
1329 : }
1330 :
1331 0 : var info *candidateLevelInfo
1332 0 : for j := range scores {
1333 0 : if scores[j].level == i {
1334 0 : info = &scores[j]
1335 0 : break
1336 : }
1337 : }
1338 :
1339 0 : marker := " "
1340 0 : if pc.startLevel.level == info.level {
1341 0 : marker = "*"
1342 0 : }
1343 0 : fmt.Fprintf(&buf, " %sL%d: score:%5.1f fillFactor:%5.1f compensatedFillFactor:%5.1f %8s %8s",
1344 0 : marker, info.level, info.score, info.fillFactor, info.compensatedFillFactor,
1345 0 : humanize.Bytes.Int64(int64(totalCompensatedSize(
1346 0 : p.vers.Levels[info.level].All(),
1347 0 : ))),
1348 0 : humanize.Bytes.Int64(p.levelMaxBytes[info.level]),
1349 0 : )
1350 0 :
1351 0 : count := 0
1352 0 : for i := range env.inProgressCompactions {
1353 0 : c := &env.inProgressCompactions[i]
1354 0 : if c.inputs[0].level != info.level {
1355 0 : continue
1356 : }
1357 0 : count++
1358 0 : if count == 1 {
1359 0 : fmt.Fprintf(&buf, " [")
1360 0 : } else {
1361 0 : fmt.Fprintf(&buf, " ")
1362 0 : }
1363 0 : fmt.Fprintf(&buf, "L%d->L%d", c.inputs[0].level, c.outputLevel)
1364 : }
1365 0 : if count > 0 {
1366 0 : fmt.Fprintf(&buf, "]")
1367 0 : }
1368 0 : fmt.Fprintf(&buf, "\n")
1369 : }
1370 0 : p.opts.Logger.Infof("pickAuto: L%d->L%d\n%s",
1371 0 : pc.startLevel.level, pc.outputLevel.level, buf.String())
1372 : }
1373 :
1374 : // pickAutoScore picks the best score-based compaction, if any.
1375 : //
1376 : // On each call, pickAutoScore computes per-level size adjustments based on
1377 : // in-progress compactions, and computes a per-level score. The levels are
1378 : // iterated over in decreasing score order trying to find a valid compaction
1379 : // anchored at that level.
1380 : //
1381 : // If a score-based compaction cannot be found, pickAuto falls back to looking
1382 : // for an elision-only compaction to remove obsolete keys.
1383 1 : func (p *compactionPickerByScore) pickAutoScore(env compactionEnv) pickedCompaction {
1384 1 : scores := p.calculateLevelScores(env.inProgressCompactions)
1385 1 :
1386 1 : // Check for a score-based compaction. candidateLevelInfos are first sorted
1387 1 : // by whether they should be compacted, so if we find a level which shouldn't
1388 1 : // be compacted, we can break early.
1389 1 : for i := range scores {
1390 1 : info := &scores[i]
1391 1 : if !info.shouldCompact() {
1392 1 : break
1393 : }
1394 1 : if info.level == numLevels-1 {
1395 1 : continue
1396 : }
1397 :
1398 1 : if info.level == 0 {
1399 1 : ptc := pickL0(env, p.opts, p.vers, p.latestVersionState.l0Organizer, p.baseLevel)
1400 1 : if ptc != nil {
1401 1 : p.addScoresToPickedCompactionMetrics(ptc, scores)
1402 1 : ptc.score = info.score
1403 1 : if false {
1404 0 : p.logCompactionForTesting(env, scores, ptc)
1405 0 : }
1406 1 : return ptc
1407 : }
1408 1 : continue
1409 : }
1410 :
1411 : // info.level > 0
1412 1 : var ok bool
1413 1 : info.file, ok = pickCompactionSeedFile(p.vers, &p.latestVersionState.virtualBackings, p.opts, info.level, info.outputLevel, env.earliestSnapshotSeqNum, env.problemSpans)
1414 1 : if !ok {
1415 1 : continue
1416 : }
1417 :
1418 1 : pc := pickAutoLPositive(env, p.opts, p.vers, p.latestVersionState.l0Organizer, *info, p.baseLevel)
1419 1 : if pc != nil {
1420 1 : p.addScoresToPickedCompactionMetrics(pc, scores)
1421 1 : pc.score = info.score
1422 1 : if false {
1423 0 : p.logCompactionForTesting(env, scores, pc)
1424 0 : }
1425 1 : return pc
1426 : }
1427 : }
1428 1 : return nil
1429 : }
1430 :
1431 : // pickAutoNonScore picks the best non-score-based compaction, if any.
1432 1 : func (p *compactionPickerByScore) pickAutoNonScore(env compactionEnv) (pc pickedCompaction) {
1433 1 : // Check for files which contain excessive point tombstones that could slow
1434 1 : // down reads. Unlike elision-only compactions, these compactions may select
1435 1 : // a file at any level rather than only the lowest level.
1436 1 : if pc := p.pickTombstoneDensityCompaction(env); pc != nil {
1437 1 : return pc
1438 1 : }
1439 :
1440 : // Check for L6 files with tombstones that may be elided. These files may
1441 : // exist if a snapshot prevented the elision of a tombstone or because of
1442 : // a move compaction. These are low-priority compactions because they
1443 : // don't help us keep up with writes, just reclaim disk space.
1444 1 : if pc := p.pickElisionOnlyCompaction(env); pc != nil {
1445 1 : return pc
1446 1 : }
1447 :
1448 : // Check for blob file rewrites. These are low-priority compactions because
1449 : // they don't help us keep up with writes, just reclaim disk space.
1450 1 : if pc := p.pickBlobFileRewriteCompaction(env); pc != nil {
1451 1 : return pc
1452 1 : }
1453 :
1454 1 : if pc := p.pickReadTriggeredCompaction(env); pc != nil {
1455 0 : return pc
1456 0 : }
1457 :
1458 : // NB: This should only be run if a read compaction wasn't
1459 : // scheduled.
1460 : //
1461 : // We won't be scheduling a read compaction right now, and in
1462 : // read heavy workloads, compactions won't be scheduled frequently
1463 : // because flushes aren't frequent. So we need to signal to the
1464 : // iterator to schedule a compaction when it adds compactions to
1465 : // the read compaction queue.
1466 : //
1467 : // We need the nil check here because without it, we have some
1468 : // tests which don't set that variable fail. Since there's a
1469 : // chance that one of those tests wouldn't want extra compactions
1470 : // to be scheduled, I added this check here, instead of
1471 : // setting rescheduleReadCompaction in those tests.
1472 1 : if env.readCompactionEnv.rescheduleReadCompaction != nil {
1473 1 : *env.readCompactionEnv.rescheduleReadCompaction = true
1474 1 : }
1475 :
1476 : // At the lowest possible compaction-picking priority, look for files marked
1477 : // for compaction. Pebble will mark files for compaction if they have atomic
1478 : // compaction units that span multiple files. While current Pebble code does
1479 : // not construct such sstables, RocksDB and earlier versions of Pebble may
1480 : // have created them. These split user keys form sets of files that must be
1481 : // compacted together for correctness (referred to as "atomic compaction
1482 : // units" within the code). Rewrite them in-place.
1483 : //
1484 : // It's also possible that a file may have been marked for compaction by
1485 : // even earlier versions of Pebble code, since TableMetadata's
1486 : // MarkedForCompaction field is persisted in the manifest. That's okay. We
1487 : // previously would've ignored the designation, whereas now we'll re-compact
1488 : // the file in place.
1489 1 : if p.vers.Stats.MarkedForCompaction > 0 {
1490 0 : if pc := p.pickRewriteCompaction(env); pc != nil {
1491 0 : return pc
1492 0 : }
1493 : }
1494 :
1495 1 : return nil
1496 : }
1497 :
1498 : func (p *compactionPickerByScore) addScoresToPickedCompactionMetrics(
1499 : pc *pickedTableCompaction, candInfo [numLevels]candidateLevelInfo,
1500 1 : ) {
1501 1 :
1502 1 : // candInfo is sorted by score, not by compaction level.
1503 1 : infoByLevel := [numLevels]candidateLevelInfo{}
1504 1 : for i := range candInfo {
1505 1 : level := candInfo[i].level
1506 1 : infoByLevel[level] = candInfo[i]
1507 1 : }
1508 : // Gather the compaction scores for the levels participating in the compaction.
1509 1 : pc.pickerMetrics.scores = make([]float64, len(pc.inputs))
1510 1 : inputIdx := 0
1511 1 : for i := range infoByLevel {
1512 1 : if pc.inputs[inputIdx].level == infoByLevel[i].level {
1513 1 : pc.pickerMetrics.scores[inputIdx] = infoByLevel[i].score
1514 1 : inputIdx++
1515 1 : }
1516 1 : if inputIdx == len(pc.inputs) {
1517 1 : break
1518 : }
1519 : }
1520 : }
1521 :
1522 : // elisionOnlyAnnotator is a manifest.Annotator that annotates B-Tree
1523 : // nodes with the *fileMetadata of a file meeting the obsolete keys criteria
1524 : // for an elision-only compaction within the subtree. If multiple files meet
1525 : // the criteria, it chooses whichever file has the lowest LargestSeqNum. The
1526 : // lowest LargestSeqNum file will be the first eligible for an elision-only
1527 : // compaction once snapshots less than or equal to its LargestSeqNum are closed.
1528 : var elisionOnlyAnnotator = &manifest.Annotator[manifest.TableMetadata]{
1529 : Aggregator: manifest.PickFileAggregator{
1530 1 : Filter: func(f *manifest.TableMetadata) (eligible bool, cacheOK bool) {
1531 1 : if f.IsCompacting() {
1532 1 : return false, true
1533 1 : }
1534 1 : if !f.StatsValid() {
1535 1 : return false, false
1536 1 : }
1537 : // Bottommost files are large and not worthwhile to compact just
1538 : // to remove a few tombstones. Consider a file eligible only if
1539 : // either its own range deletions delete at least 10% of its data or
1540 : // its deletion tombstones make at least 10% of its entries.
1541 : //
1542 : // TODO(jackson): This does not account for duplicate user keys
1543 : // which may be collapsed. Ideally, we would have 'obsolete keys'
1544 : // statistics that would include tombstones, the keys that are
1545 : // dropped by tombstones and duplicated user keys. See #847.
1546 : //
1547 : // Note that tables that contain exclusively range keys (i.e. no point keys,
1548 : // `NumEntries` and `RangeDeletionsBytesEstimate` are both zero) are excluded
1549 : // from elision-only compactions.
1550 : // TODO(travers): Consider an alternative heuristic for elision of range-keys.
1551 1 : return f.Stats.RangeDeletionsBytesEstimate*10 >= f.Size || f.Stats.NumDeletions*10 > f.Stats.NumEntries, true
1552 : },
1553 1 : Compare: func(f1 *manifest.TableMetadata, f2 *manifest.TableMetadata) bool {
1554 1 : return f1.LargestSeqNum < f2.LargestSeqNum
1555 1 : },
1556 : },
1557 : }
1558 :
1559 : // markedForCompactionAnnotator is a manifest.Annotator that annotates B-Tree
1560 : // nodes with the *fileMetadata of a file that is marked for compaction
1561 : // within the subtree. If multiple files meet the criteria, it chooses
1562 : // whichever file has the lowest LargestSeqNum.
1563 : var markedForCompactionAnnotator = &manifest.Annotator[manifest.TableMetadata]{
1564 : Aggregator: manifest.PickFileAggregator{
1565 0 : Filter: func(f *manifest.TableMetadata) (eligible bool, cacheOK bool) {
1566 0 : return f.MarkedForCompaction, true
1567 0 : },
1568 0 : Compare: func(f1 *manifest.TableMetadata, f2 *manifest.TableMetadata) bool {
1569 0 : return f1.LargestSeqNum < f2.LargestSeqNum
1570 0 : },
1571 : },
1572 : }
1573 :
1574 : // pickedCompactionFromCandidateFile creates a pickedCompaction from a *fileMetadata
1575 : // with various checks to ensure that the file still exists in the expected level
1576 : // and isn't already being compacted.
1577 : func (p *compactionPickerByScore) pickedCompactionFromCandidateFile(
1578 : candidate *manifest.TableMetadata,
1579 : env compactionEnv,
1580 : startLevel int,
1581 : outputLevel int,
1582 : kind compactionKind,
1583 1 : ) *pickedTableCompaction {
1584 1 : if candidate == nil || candidate.IsCompacting() {
1585 1 : return nil
1586 1 : }
1587 :
1588 1 : var inputs manifest.LevelSlice
1589 1 : if startLevel == 0 {
1590 1 : // Overlapping L0 files must also be compacted alongside the candidate.
1591 1 : inputs = p.vers.Overlaps(0, candidate.UserKeyBounds())
1592 1 : } else {
1593 1 : inputs = p.vers.Levels[startLevel].Find(p.opts.Comparer.Compare, candidate)
1594 1 : }
1595 1 : if invariants.Enabled {
1596 1 : found := false
1597 1 : for f := range inputs.All() {
1598 1 : if f.TableNum == candidate.TableNum {
1599 1 : found = true
1600 1 : }
1601 : }
1602 1 : if !found {
1603 0 : panic(fmt.Sprintf("file %s not found in level %d as expected", candidate.TableNum, startLevel))
1604 : }
1605 : }
1606 :
1607 1 : pc := newPickedTableCompaction(p.opts, p.vers, p.latestVersionState.l0Organizer,
1608 1 : startLevel, outputLevel, p.baseLevel)
1609 1 : pc.kind = kind
1610 1 : pc.startLevel.files = inputs
1611 1 :
1612 1 : if !pc.setupInputs(p.opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, env.problemSpans) {
1613 1 : return nil
1614 1 : }
1615 1 : return pc
1616 : }
1617 :
1618 : // pickElisionOnlyCompaction looks for compactions of sstables in the
1619 : // bottommost level containing obsolete records that may now be dropped.
1620 : func (p *compactionPickerByScore) pickElisionOnlyCompaction(
1621 : env compactionEnv,
1622 1 : ) (pc *pickedTableCompaction) {
1623 1 : if p.opts.private.disableElisionOnlyCompactions {
1624 1 : return nil
1625 1 : }
1626 1 : candidate := elisionOnlyAnnotator.LevelAnnotation(p.vers.Levels[numLevels-1])
1627 1 : if candidate == nil {
1628 1 : return nil
1629 1 : }
1630 1 : if candidate.LargestSeqNum >= env.earliestSnapshotSeqNum {
1631 1 : return nil
1632 1 : }
1633 1 : return p.pickedCompactionFromCandidateFile(candidate, env, numLevels-1, numLevels-1, compactionKindElisionOnly)
1634 : }
1635 :
1636 : // pickRewriteCompaction attempts to construct a compaction that
1637 : // rewrites a file marked for compaction. pickRewriteCompaction will
1638 : // pull in adjacent files in the file's atomic compaction unit if
1639 : // necessary. A rewrite compaction outputs files to the same level as
1640 : // the input level.
1641 : func (p *compactionPickerByScore) pickRewriteCompaction(
1642 : env compactionEnv,
1643 0 : ) (pc *pickedTableCompaction) {
1644 0 : if p.vers.Stats.MarkedForCompaction == 0 {
1645 0 : return nil
1646 0 : }
1647 0 : for l := numLevels - 1; l >= 0; l-- {
1648 0 : candidate := markedForCompactionAnnotator.LevelAnnotation(p.vers.Levels[l])
1649 0 : if candidate == nil {
1650 0 : // Try the next level.
1651 0 : continue
1652 : }
1653 0 : pc := p.pickedCompactionFromCandidateFile(candidate, env, l, l, compactionKindRewrite)
1654 0 : if pc != nil {
1655 0 : return pc
1656 0 : }
1657 : }
1658 0 : return nil
1659 : }
1660 :
1661 : // pickBlobFileRewriteCompaction looks for compactions of blob files that
1662 : // can be rewritten to reclaim disk space.
1663 : func (p *compactionPickerByScore) pickBlobFileRewriteCompaction(
1664 : env compactionEnv,
1665 1 : ) (pc *pickedBlobFileCompaction) {
1666 1 : aggregateStats, heuristicStats := p.latestVersionState.blobFiles.Stats()
1667 1 : if heuristicStats.CountFilesEligible == 0 && heuristicStats.CountFilesTooRecent == 0 {
1668 1 : // No blob files with any garbage to rewrite.
1669 1 : return nil
1670 1 : }
1671 1 : policy := p.opts.Experimental.ValueSeparationPolicy()
1672 1 : if policy.TargetGarbageRatio >= 1.0 {
1673 1 : // Blob file rewrite compactions are disabled.
1674 1 : return nil
1675 1 : }
1676 1 : garbagePct := float64(aggregateStats.ValueSize-aggregateStats.ReferencedValueSize) /
1677 1 : float64(aggregateStats.ValueSize)
1678 1 : if garbagePct <= policy.TargetGarbageRatio {
1679 1 : // Not enough garbage to warrant a rewrite compaction.
1680 1 : return nil
1681 1 : }
1682 :
1683 : // Check if there is an ongoing blob file rewrite compaction. If there is,
1684 : // don't schedule a new one.
1685 1 : for _, c := range env.inProgressCompactions {
1686 1 : if c.kind == compactionKindBlobFileRewrite {
1687 1 : return nil
1688 1 : }
1689 : }
1690 :
1691 1 : candidate, ok := p.latestVersionState.blobFiles.ReplacementCandidate()
1692 1 : if !ok {
1693 1 : // None meet the heuristic.
1694 1 : return nil
1695 1 : }
1696 1 : return &pickedBlobFileCompaction{
1697 1 : vers: p.vers,
1698 1 : file: candidate,
1699 1 : referencingTables: p.latestVersionState.blobFiles.ReferencingTables(candidate.FileID),
1700 1 : }
1701 : }
1702 :
1703 : // pickTombstoneDensityCompaction looks for a compaction that eliminates
1704 : // regions of extremely high point tombstone density. For each level, it picks
1705 : // a file where the ratio of tombstone-dense blocks is at least
1706 : // options.Experimental.MinTombstoneDenseRatio, prioritizing compaction of
1707 : // files with higher ratios of tombstone-dense blocks.
1708 : func (p *compactionPickerByScore) pickTombstoneDensityCompaction(
1709 : env compactionEnv,
1710 1 : ) (pc *pickedTableCompaction) {
1711 1 : if p.opts.Experimental.TombstoneDenseCompactionThreshold <= 0 {
1712 0 : // Tombstone density compactions are disabled.
1713 0 : return nil
1714 0 : }
1715 :
1716 1 : var candidate *manifest.TableMetadata
1717 1 : var level int
1718 1 : // If a candidate file has a very high overlapping ratio, point tombstones
1719 1 : // in it are likely sparse in keyspace even if the sstable itself is tombstone
1720 1 : // dense. These tombstones likely wouldn't be slow to iterate over, so we exclude
1721 1 : // these files from tombstone density compactions. The threshold of 40.0 is
1722 1 : // chosen somewhat arbitrarily, after some observations around excessively large
1723 1 : // tombstone density compactions.
1724 1 : const maxOverlappingRatio = 40.0
1725 1 : // NB: we don't consider the lowest level because elision-only compactions
1726 1 : // handle that case.
1727 1 : lastNonEmptyLevel := numLevels - 1
1728 1 : for l := numLevels - 2; l >= 0; l-- {
1729 1 : iter := p.vers.Levels[l].Iter()
1730 1 : for f := iter.First(); f != nil; f = iter.Next() {
1731 1 : if f.IsCompacting() || !f.StatsValid() || f.Size == 0 {
1732 1 : continue
1733 : }
1734 1 : if f.Stats.TombstoneDenseBlocksRatio < p.opts.Experimental.TombstoneDenseCompactionThreshold {
1735 1 : continue
1736 : }
1737 1 : overlaps := p.vers.Overlaps(lastNonEmptyLevel, f.UserKeyBounds())
1738 1 : if float64(overlaps.AggregateSizeSum())/float64(f.Size) > maxOverlappingRatio {
1739 0 : continue
1740 : }
1741 1 : if candidate == nil || candidate.Stats.TombstoneDenseBlocksRatio < f.Stats.TombstoneDenseBlocksRatio {
1742 1 : candidate = f
1743 1 : level = l
1744 1 : }
1745 : }
1746 : // We prefer lower level (ie. L5) candidates over higher level (ie. L4) ones.
1747 1 : if candidate != nil {
1748 1 : break
1749 : }
1750 1 : if !p.vers.Levels[l].Empty() {
1751 1 : lastNonEmptyLevel = l
1752 1 : }
1753 : }
1754 :
1755 1 : return p.pickedCompactionFromCandidateFile(candidate, env, level, defaultOutputLevel(level, p.baseLevel), compactionKindTombstoneDensity)
1756 : }
1757 :
1758 : // pickAutoLPositive picks an automatic compaction for the candidate
1759 : // file in a positive-numbered level. This function must not be used for
1760 : // L0.
1761 : func pickAutoLPositive(
1762 : env compactionEnv,
1763 : opts *Options,
1764 : vers *manifest.Version,
1765 : l0Organizer *manifest.L0Organizer,
1766 : cInfo candidateLevelInfo,
1767 : baseLevel int,
1768 1 : ) (pc *pickedTableCompaction) {
1769 1 : if cInfo.level == 0 {
1770 0 : panic("pebble: pickAutoLPositive called for L0")
1771 : }
1772 :
1773 1 : pc = newPickedTableCompaction(opts, vers, l0Organizer, cInfo.level, defaultOutputLevel(cInfo.level, baseLevel), baseLevel)
1774 1 : if pc.outputLevel.level != cInfo.outputLevel {
1775 0 : panic("pebble: compaction picked unexpected output level")
1776 : }
1777 1 : pc.startLevel.files = cInfo.file.Slice()
1778 1 :
1779 1 : if !pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, env.problemSpans) {
1780 1 : return nil
1781 1 : }
1782 1 : return pc.maybeAddLevel(opts, env)
1783 : }
1784 :
1785 : // maybeAddLevel maybe adds a level to the picked compaction.
1786 : func (pc *pickedTableCompaction) maybeAddLevel(
1787 : opts *Options, env compactionEnv,
1788 1 : ) *pickedTableCompaction {
1789 1 : pc.pickerMetrics.singleLevelOverlappingRatio = pc.overlappingRatio()
1790 1 : if pc.outputLevel.level == numLevels-1 {
1791 1 : // Don't add a level if the current output level is in L6.
1792 1 : return pc
1793 1 : }
1794 1 : if !opts.Experimental.MultiLevelCompactionHeuristic.allowL0() && pc.startLevel.level == 0 {
1795 1 : return pc
1796 1 : }
1797 1 : targetFileSize := opts.TargetFileSize(pc.outputLevel.level, pc.baseLevel)
1798 1 : if pc.estimatedInputSize() > expandedCompactionByteSizeLimit(opts, targetFileSize, env.diskAvailBytes) {
1799 0 : // Don't add a level if the current compaction exceeds the compaction size limit
1800 0 : return pc
1801 0 : }
1802 1 : return opts.Experimental.MultiLevelCompactionHeuristic.pick(pc, opts, env)
1803 : }
1804 :
1805 : // MultiLevelHeuristic evaluates whether to add files from the next level into the compaction.
1806 : type MultiLevelHeuristic interface {
1807 : // Evaluate returns the preferred compaction.
1808 : pick(pc *pickedTableCompaction, opts *Options, env compactionEnv) *pickedTableCompaction
1809 :
1810 : // Returns if the heuristic allows L0 to be involved in ML compaction
1811 : allowL0() bool
1812 :
1813 : // String implements fmt.Stringer.
1814 : String() string
1815 : }
1816 :
1817 : // NoMultiLevel will never add an additional level to the compaction.
1818 : type NoMultiLevel struct{}
1819 :
1820 : var _ MultiLevelHeuristic = (*NoMultiLevel)(nil)
1821 :
1822 : func (nml NoMultiLevel) pick(
1823 : pc *pickedTableCompaction, opts *Options, env compactionEnv,
1824 1 : ) *pickedTableCompaction {
1825 1 : return pc
1826 1 : }
1827 :
1828 1 : func (nml NoMultiLevel) allowL0() bool { return false }
1829 1 : func (nml NoMultiLevel) String() string { return "none" }
1830 :
1831 1 : func (pc *pickedTableCompaction) predictedWriteAmp() float64 {
1832 1 : var bytesToCompact uint64
1833 1 : var higherLevelBytes uint64
1834 1 : for i := range pc.inputs {
1835 1 : levelSize := pc.inputs[i].files.AggregateSizeSum()
1836 1 : bytesToCompact += levelSize
1837 1 : if i != len(pc.inputs)-1 {
1838 1 : higherLevelBytes += levelSize
1839 1 : }
1840 : }
1841 1 : return float64(bytesToCompact) / float64(higherLevelBytes)
1842 : }
1843 :
1844 1 : func (pc *pickedTableCompaction) overlappingRatio() float64 {
1845 1 : var higherLevelBytes uint64
1846 1 : var lowestLevelBytes uint64
1847 1 : for i := range pc.inputs {
1848 1 : levelSize := pc.inputs[i].files.AggregateSizeSum()
1849 1 : if i == len(pc.inputs)-1 {
1850 1 : lowestLevelBytes += levelSize
1851 1 : continue
1852 : }
1853 1 : higherLevelBytes += levelSize
1854 : }
1855 1 : return float64(lowestLevelBytes) / float64(higherLevelBytes)
1856 : }
1857 :
1858 : // WriteAmpHeuristic defines a multi level compaction heuristic which will add
1859 : // an additional level to the picked compaction if it reduces predicted write
1860 : // amp of the compaction + the addPropensity constant.
1861 : type WriteAmpHeuristic struct {
1862 : // addPropensity is a constant that affects the propensity to conduct multilevel
1863 : // compactions. If positive, a multilevel compaction may get picked even if
1864 : // the single level compaction has lower write amp, and vice versa.
1865 : AddPropensity float64
1866 :
1867 : // AllowL0 if true, allow l0 to be involved in a ML compaction.
1868 : AllowL0 bool
1869 : }
1870 :
1871 : var _ MultiLevelHeuristic = (*WriteAmpHeuristic)(nil)
1872 :
1873 : // TODO(msbutler): microbenchmark the extent to which multilevel compaction
1874 : // picking slows down the compaction picking process. This should be as fast as
1875 : // possible since Compaction-picking holds d.mu, which prevents WAL rotations,
1876 : // in-progress flushes and compactions from completing, etc. Consider ways to
1877 : // deduplicate work, given that setupInputs has already been called.
1878 : func (wa WriteAmpHeuristic) pick(
1879 : pcOrig *pickedTableCompaction, opts *Options, env compactionEnv,
1880 1 : ) *pickedTableCompaction {
1881 1 : pcMulti := pcOrig.clone()
1882 1 : if !pcMulti.setupMultiLevelCandidate(opts, env) {
1883 1 : return pcOrig
1884 1 : }
1885 : // We consider the addition of a level as an "expansion" of the compaction.
1886 : // If pcMulti is past the expanded compaction byte size limit already,
1887 : // we don't consider it.
1888 1 : targetFileSize := opts.TargetFileSize(pcMulti.outputLevel.level, pcMulti.baseLevel)
1889 1 : if pcMulti.estimatedInputSize() >= expandedCompactionByteSizeLimit(opts, targetFileSize, env.diskAvailBytes) {
1890 0 : return pcOrig
1891 0 : }
1892 1 : picked := pcOrig
1893 1 : if pcMulti.predictedWriteAmp() <= pcOrig.predictedWriteAmp()+wa.AddPropensity {
1894 1 : picked = pcMulti
1895 1 : }
1896 : // Regardless of what compaction was picked, log the multilevelOverlapping ratio.
1897 1 : picked.pickerMetrics.multiLevelOverlappingRatio = pcMulti.overlappingRatio()
1898 1 : return picked
1899 : }
1900 :
1901 1 : func (wa WriteAmpHeuristic) allowL0() bool {
1902 1 : return wa.AllowL0
1903 1 : }
1904 :
1905 : // String implements fmt.Stringer.
1906 1 : func (wa WriteAmpHeuristic) String() string {
1907 1 : return fmt.Sprintf("wamp(%.2f, %t)", wa.AddPropensity, wa.AllowL0)
1908 1 : }
1909 :
1910 : // Helper method to pick compactions originating from L0. Uses information about
1911 : // sublevels to generate a compaction.
1912 : func pickL0(
1913 : env compactionEnv,
1914 : opts *Options,
1915 : vers *manifest.Version,
1916 : l0Organizer *manifest.L0Organizer,
1917 : baseLevel int,
1918 1 : ) *pickedTableCompaction {
1919 1 : // It is important to pass information about Lbase files to L0Sublevels
1920 1 : // so it can pick a compaction that does not conflict with an Lbase => Lbase+1
1921 1 : // compaction. Without this, we observed reduced concurrency of L0=>Lbase
1922 1 : // compactions, and increasing read amplification in L0.
1923 1 : //
1924 1 : // TODO(bilal) Remove the minCompactionDepth parameter once fixing it at 1
1925 1 : // has been shown to not cause a performance regression.
1926 1 : lcf := l0Organizer.PickBaseCompaction(opts.Logger, 1, vers.Levels[baseLevel].Slice(), baseLevel, env.problemSpans)
1927 1 : if lcf != nil {
1928 1 : pc := newPickedCompactionFromL0(lcf, opts, vers, l0Organizer, baseLevel, true)
1929 1 : if pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, env.problemSpans) {
1930 1 : if pc.startLevel.files.Empty() {
1931 0 : opts.Logger.Errorf("%v", base.AssertionFailedf("empty compaction chosen"))
1932 0 : }
1933 1 : return pc.maybeAddLevel(opts, env)
1934 : }
1935 : // TODO(radu): investigate why this happens.
1936 : // opts.Logger.Errorf("%v", base.AssertionFailedf("setupInputs failed"))
1937 : }
1938 :
1939 : // Couldn't choose a base compaction. Try choosing an intra-L0
1940 : // compaction. Note that we pass in L0CompactionThreshold here as opposed to
1941 : // 1, since choosing a single sublevel intra-L0 compaction is
1942 : // counterproductive.
1943 1 : lcf = l0Organizer.PickIntraL0Compaction(env.earliestUnflushedSeqNum, minIntraL0Count, env.problemSpans)
1944 1 : if lcf != nil {
1945 1 : pc := newPickedCompactionFromL0(lcf, opts, vers, l0Organizer, baseLevel, false)
1946 1 : if pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, env.problemSpans) {
1947 1 : if pc.startLevel.files.Empty() {
1948 0 : opts.Logger.Fatalf("empty compaction chosen")
1949 0 : }
1950 : // A single-file intra-L0 compaction is unproductive.
1951 1 : if iter := pc.startLevel.files.Iter(); iter.First() != nil && iter.Next() != nil {
1952 1 : pc.bounds = manifest.KeyRange(opts.Comparer.Compare, pc.startLevel.files.All())
1953 1 : return pc
1954 1 : }
1955 0 : } else {
1956 0 : // TODO(radu): investigate why this happens.
1957 0 : // opts.Logger.Errorf("%v", base.AssertionFailedf("setupInputs failed"))
1958 0 : }
1959 : }
1960 1 : return nil
1961 : }
1962 :
1963 : func newPickedManualCompaction(
1964 : vers *manifest.Version,
1965 : l0Organizer *manifest.L0Organizer,
1966 : opts *Options,
1967 : env compactionEnv,
1968 : baseLevel int,
1969 : manual *manualCompaction,
1970 1 : ) (pc *pickedTableCompaction, retryLater bool) {
1971 1 : outputLevel := manual.level + 1
1972 1 : if manual.level == 0 {
1973 1 : outputLevel = baseLevel
1974 1 : } else if manual.level < baseLevel {
1975 1 : // The start level for a compaction must be >= Lbase. A manual
1976 1 : // compaction could have been created adhering to that condition, and
1977 1 : // then an automatic compaction came in and compacted all of the
1978 1 : // sstables in Lbase to Lbase+1 which caused Lbase to change. Simply
1979 1 : // ignore this manual compaction as there is nothing to do (manual.level
1980 1 : // points to an empty level).
1981 1 : return nil, false
1982 1 : }
1983 : // This conflictsWithInProgress call is necessary for the manual compaction to
1984 : // be retried when it conflicts with an ongoing automatic compaction. Without
1985 : // it, the compaction is dropped due to pc.setupInputs returning false since
1986 : // the input/output range is already being compacted, and the manual
1987 : // compaction ends with a non-compacted LSM.
1988 1 : if conflictsWithInProgress(manual, outputLevel, env.inProgressCompactions, opts.Comparer.Compare) {
1989 1 : return nil, true
1990 1 : }
1991 1 : pc = newPickedTableCompaction(opts, vers, l0Organizer, manual.level, defaultOutputLevel(manual.level, baseLevel), baseLevel)
1992 1 : pc.manualID = manual.id
1993 1 : manual.outputLevel = pc.outputLevel.level
1994 1 : pc.startLevel.files = vers.Overlaps(manual.level, base.UserKeyBoundsInclusive(manual.start, manual.end))
1995 1 : if pc.startLevel.files.Empty() {
1996 1 : // Nothing to do
1997 1 : return nil, false
1998 1 : }
1999 : // We use nil problemSpans because we don't want problem spans to prevent
2000 : // manual compactions.
2001 1 : if !pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, nil /* problemSpans */) {
2002 1 : // setupInputs returned false indicating there's a conflicting
2003 1 : // concurrent compaction.
2004 1 : return nil, true
2005 1 : }
2006 1 : if pc = pc.maybeAddLevel(opts, env); pc == nil {
2007 0 : return nil, false
2008 0 : }
2009 1 : if pc.outputLevel.level != outputLevel {
2010 1 : if len(pc.inputs) > 2 {
2011 1 : // Multilevel compactions relax this invariant.
2012 1 : } else {
2013 0 : panic("pebble: compaction picked unexpected output level")
2014 : }
2015 : }
2016 1 : return pc, false
2017 : }
2018 :
2019 : // pickDownloadCompaction picks a download compaction for the downloadSpan,
2020 : // which could be specified as being performed either by a copy compaction of
2021 : // the backing file or a rewrite compaction.
2022 : func pickDownloadCompaction(
2023 : vers *manifest.Version,
2024 : l0Organizer *manifest.L0Organizer,
2025 : opts *Options,
2026 : env compactionEnv,
2027 : baseLevel int,
2028 : kind compactionKind,
2029 : level int,
2030 : file *manifest.TableMetadata,
2031 1 : ) (pc *pickedTableCompaction) {
2032 1 : // Check if the file is compacting already.
2033 1 : if file.CompactionState == manifest.CompactionStateCompacting {
2034 0 : return nil
2035 0 : }
2036 1 : if kind != compactionKindCopy && kind != compactionKindRewrite {
2037 0 : panic("invalid download/rewrite compaction kind")
2038 : }
2039 1 : pc = newPickedTableCompaction(opts, vers, l0Organizer, level, level, baseLevel)
2040 1 : pc.kind = kind
2041 1 : pc.startLevel.files = manifest.NewLevelSliceKeySorted(opts.Comparer.Compare, []*manifest.TableMetadata{file})
2042 1 : if !pc.setupInputs(opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, nil /* problemSpans */) {
2043 0 : // setupInputs returned false indicating there's a conflicting
2044 0 : // concurrent compaction.
2045 0 : return nil
2046 0 : }
2047 1 : if pc.outputLevel.level != level {
2048 0 : panic("pebble: download compaction picked unexpected output level")
2049 : }
2050 1 : return pc
2051 : }
2052 :
2053 : func (p *compactionPickerByScore) pickReadTriggeredCompaction(
2054 : env compactionEnv,
2055 1 : ) (pc *pickedTableCompaction) {
2056 1 : // If a flush is in-progress or expected to happen soon, it means more writes are taking place. We would
2057 1 : // soon be scheduling more write focussed compactions. In this case, skip read compactions as they are
2058 1 : // lower priority.
2059 1 : if env.readCompactionEnv.flushing || env.readCompactionEnv.readCompactions == nil {
2060 1 : return nil
2061 1 : }
2062 1 : for env.readCompactionEnv.readCompactions.size > 0 {
2063 0 : rc := env.readCompactionEnv.readCompactions.remove()
2064 0 : if pc = pickReadTriggeredCompactionHelper(p, rc, env); pc != nil {
2065 0 : break
2066 : }
2067 : }
2068 1 : return pc
2069 : }
2070 :
2071 : func pickReadTriggeredCompactionHelper(
2072 : p *compactionPickerByScore, rc *readCompaction, env compactionEnv,
2073 0 : ) (pc *pickedTableCompaction) {
2074 0 : overlapSlice := p.vers.Overlaps(rc.level, base.UserKeyBoundsInclusive(rc.start, rc.end))
2075 0 : var fileMatches bool
2076 0 : for f := range overlapSlice.All() {
2077 0 : if f.TableNum == rc.tableNum {
2078 0 : fileMatches = true
2079 0 : break
2080 : }
2081 : }
2082 0 : if !fileMatches {
2083 0 : return nil
2084 0 : }
2085 :
2086 0 : pc = newPickedTableCompaction(p.opts, p.vers, p.latestVersionState.l0Organizer,
2087 0 : rc.level, defaultOutputLevel(rc.level, p.baseLevel), p.baseLevel)
2088 0 :
2089 0 : pc.startLevel.files = overlapSlice
2090 0 : if !pc.setupInputs(p.opts, env.diskAvailBytes, env.inProgressCompactions, pc.startLevel, env.problemSpans) {
2091 0 : return nil
2092 0 : }
2093 0 : pc.kind = compactionKindRead
2094 0 :
2095 0 : // Prevent read compactions which are too wide.
2096 0 : outputOverlaps := pc.version.Overlaps(pc.outputLevel.level, pc.bounds)
2097 0 : if outputOverlaps.AggregateSizeSum() > pc.maxReadCompactionBytes {
2098 0 : return nil
2099 0 : }
2100 :
2101 : // Prevent compactions which start with a small seed file X, but overlap
2102 : // with over allowedCompactionWidth * X file sizes in the output layer.
2103 0 : const allowedCompactionWidth = 35
2104 0 : if outputOverlaps.AggregateSizeSum() > overlapSlice.AggregateSizeSum()*allowedCompactionWidth {
2105 0 : return nil
2106 0 : }
2107 :
2108 0 : return pc
2109 : }
2110 :
2111 0 : func (p *compactionPickerByScore) forceBaseLevel1() {
2112 0 : p.baseLevel = 1
2113 0 : }
2114 :
2115 : // outputKeyRangeAlreadyCompacting checks if the input range of the picked
2116 : // compaction is already being written to by an in-progress compaction.
2117 : func outputKeyRangeAlreadyCompacting(
2118 : cmp base.Compare, inProgressCompactions []compactionInfo, pc *pickedTableCompaction,
2119 1 : ) bool {
2120 1 : // Look for active compactions outputting to the same region of the key
2121 1 : // space in the same output level. Two potential compactions may conflict
2122 1 : // without sharing input files if there are no files in the output level
2123 1 : // that overlap with the intersection of the compactions' key spaces.
2124 1 : //
2125 1 : // Consider an active L0->Lbase compaction compacting two L0 files one
2126 1 : // [a-f] and the other [t-z] into Lbase.
2127 1 : //
2128 1 : // L0
2129 1 : // ↦ 000100 ↤ ↦ 000101 ↤
2130 1 : // L1
2131 1 : // ↦ 000004 ↤
2132 1 : // a b c d e f g h i j k l m n o p q r s t u v w x y z
2133 1 : //
2134 1 : // If a new file 000102 [j-p] is flushed while the existing compaction is
2135 1 : // still ongoing, new file would not be in any compacting sublevel
2136 1 : // intervals and would not overlap with any Lbase files that are also
2137 1 : // compacting. However, this compaction cannot be picked because the
2138 1 : // compaction's output key space [j-p] would overlap the existing
2139 1 : // compaction's output key space [a-z].
2140 1 : //
2141 1 : // L0
2142 1 : // ↦ 000100* ↤ ↦ 000102 ↤ ↦ 000101* ↤
2143 1 : // L1
2144 1 : // ↦ 000004* ↤
2145 1 : // a b c d e f g h i j k l m n o p q r s t u v w x y z
2146 1 : //
2147 1 : // * - currently compacting
2148 1 : if pc.outputLevel != nil && pc.outputLevel.level != 0 {
2149 1 : for _, c := range inProgressCompactions {
2150 1 : if pc.outputLevel.level != c.outputLevel {
2151 1 : continue
2152 : }
2153 1 : if !c.bounds.Overlaps(cmp, &pc.bounds) {
2154 1 : continue
2155 : }
2156 : // The picked compaction and the in-progress compaction c are
2157 : // outputting to the same region of the key space of the same
2158 : // level.
2159 1 : return true
2160 : }
2161 : }
2162 1 : return false
2163 : }
2164 :
2165 : // conflictsWithInProgress checks if there are any in-progress compactions with overlapping keyspace.
2166 : func conflictsWithInProgress(
2167 : manual *manualCompaction, outputLevel int, inProgressCompactions []compactionInfo, cmp Compare,
2168 1 : ) bool {
2169 1 : for _, c := range inProgressCompactions {
2170 1 : if (c.outputLevel == manual.level || c.outputLevel == outputLevel) &&
2171 1 : areUserKeysOverlapping(manual.start, manual.end, c.bounds.Start, c.bounds.End.Key, cmp) {
2172 1 : return true
2173 1 : }
2174 1 : for _, in := range c.inputs {
2175 1 : if in.files.Empty() {
2176 1 : continue
2177 : }
2178 1 : iter := in.files.Iter()
2179 1 : smallest := iter.First().Smallest().UserKey
2180 1 : largest := iter.Last().Largest().UserKey
2181 1 : if (in.level == manual.level || in.level == outputLevel) &&
2182 1 : areUserKeysOverlapping(manual.start, manual.end, smallest, largest, cmp) {
2183 1 : return true
2184 1 : }
2185 : }
2186 : }
2187 1 : return false
2188 : }
2189 :
2190 1 : func areUserKeysOverlapping(x1, x2, y1, y2 []byte, cmp Compare) bool {
2191 1 : return cmp(x1, y2) <= 0 && cmp(y1, x2) <= 0
2192 1 : }
|
