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 : "fmt"
10 : "math"
11 : "sort"
12 : "strings"
13 :
14 : "github.com/cockroachdb/errors"
15 : "github.com/cockroachdb/pebble/internal/base"
16 : "github.com/cockroachdb/pebble/internal/humanize"
17 : "github.com/cockroachdb/pebble/internal/invariants"
18 : "github.com/cockroachdb/pebble/internal/manifest"
19 : )
20 :
21 : // The minimum count for an intra-L0 compaction. This matches the RocksDB
22 : // heuristic.
23 : const minIntraL0Count = 4
24 :
25 : type compactionEnv struct {
26 : // diskAvailBytes holds a statistic on the number of bytes available on
27 : // disk, as reported by the filesystem. It's used to be more restrictive in
28 : // expanding compactions if available disk space is limited.
29 : //
30 : // The cached value (d.diskAvailBytes) is updated whenever a file is deleted
31 : // and whenever a compaction or flush completes. Since file removal is the
32 : // primary means of reclaiming space, there is a rough bound on the
33 : // statistic's staleness when available bytes is growing. Compactions and
34 : // flushes are longer, slower operations and provide a much looser bound
35 : // when available bytes is decreasing.
36 : diskAvailBytes uint64
37 : earliestUnflushedSeqNum base.SeqNum
38 : earliestSnapshotSeqNum base.SeqNum
39 : inProgressCompactions []compactionInfo
40 : readCompactionEnv readCompactionEnv
41 : }
42 :
43 : type compactionPicker interface {
44 : getScores([]compactionInfo) [numLevels]float64
45 : getBaseLevel() int
46 : estimatedCompactionDebt(l0ExtraSize uint64) uint64
47 : pickAuto(env compactionEnv) (pc *pickedCompaction)
48 : pickElisionOnlyCompaction(env compactionEnv) (pc *pickedCompaction)
49 : pickRewriteCompaction(env compactionEnv) (pc *pickedCompaction)
50 : pickReadTriggeredCompaction(env compactionEnv) (pc *pickedCompaction)
51 : forceBaseLevel1()
52 : }
53 :
54 : // readCompactionEnv is used to hold data required to perform read compactions
55 : type readCompactionEnv struct {
56 : rescheduleReadCompaction *bool
57 : readCompactions *readCompactionQueue
58 : flushing bool
59 : }
60 :
61 : // Information about in-progress compactions provided to the compaction picker.
62 : // These are used to constrain the new compactions that will be picked.
63 : type compactionInfo struct {
64 : // versionEditApplied is true if this compaction's version edit has already
65 : // been committed. The compaction may still be in-progress deleting newly
66 : // obsolete files.
67 : versionEditApplied bool
68 : inputs []compactionLevel
69 : outputLevel int
70 : smallest InternalKey
71 : largest InternalKey
72 : }
73 :
74 0 : func (info compactionInfo) String() string {
75 0 : var buf bytes.Buffer
76 0 : var largest int
77 0 : for i, in := range info.inputs {
78 0 : if i > 0 {
79 0 : fmt.Fprintf(&buf, " -> ")
80 0 : }
81 0 : fmt.Fprintf(&buf, "L%d", in.level)
82 0 : in.files.Each(func(m *fileMetadata) {
83 0 : fmt.Fprintf(&buf, " %s", m.FileNum)
84 0 : })
85 0 : if largest < in.level {
86 0 : largest = in.level
87 0 : }
88 : }
89 0 : if largest != info.outputLevel || len(info.inputs) == 1 {
90 0 : fmt.Fprintf(&buf, " -> L%d", info.outputLevel)
91 0 : }
92 0 : return buf.String()
93 : }
94 :
95 : type sortCompactionLevelsByPriority []candidateLevelInfo
96 :
97 1 : func (s sortCompactionLevelsByPriority) Len() int {
98 1 : return len(s)
99 1 : }
100 :
101 : // A level should be picked for compaction if the compensatedScoreRatio is >= the
102 : // compactionScoreThreshold.
103 : const compactionScoreThreshold = 1
104 :
105 : // Less should return true if s[i] must be placed earlier than s[j] in the final
106 : // sorted list. The candidateLevelInfo for the level placed earlier is more likely
107 : // to be picked for a compaction.
108 1 : func (s sortCompactionLevelsByPriority) Less(i, j int) bool {
109 1 : iShouldCompact := s[i].compensatedScoreRatio >= compactionScoreThreshold
110 1 : jShouldCompact := s[j].compensatedScoreRatio >= compactionScoreThreshold
111 1 : // Ordering is defined as decreasing on (shouldCompact, uncompensatedScoreRatio)
112 1 : // where shouldCompact is 1 for true and 0 for false.
113 1 : if iShouldCompact && !jShouldCompact {
114 1 : return true
115 1 : }
116 1 : if !iShouldCompact && jShouldCompact {
117 1 : return false
118 1 : }
119 :
120 1 : if s[i].uncompensatedScoreRatio != s[j].uncompensatedScoreRatio {
121 1 : return s[i].uncompensatedScoreRatio > s[j].uncompensatedScoreRatio
122 1 : }
123 1 : return s[i].level < s[j].level
124 : }
125 :
126 1 : func (s sortCompactionLevelsByPriority) Swap(i, j int) {
127 1 : s[i], s[j] = s[j], s[i]
128 1 : }
129 :
130 : // sublevelInfo is used to tag a LevelSlice for an L0 sublevel with the
131 : // sublevel.
132 : type sublevelInfo struct {
133 : manifest.LevelSlice
134 : sublevel manifest.Layer
135 : }
136 :
137 1 : func (cl sublevelInfo) Clone() sublevelInfo {
138 1 : return sublevelInfo{
139 1 : sublevel: cl.sublevel,
140 1 : LevelSlice: cl.LevelSlice,
141 1 : }
142 1 : }
143 0 : func (cl sublevelInfo) String() string {
144 0 : return fmt.Sprintf(`Sublevel %s; Levels %s`, cl.sublevel, cl.LevelSlice)
145 0 : }
146 :
147 : // generateSublevelInfo will generate the level slices for each of the sublevels
148 : // from the level slice for all of L0.
149 1 : func generateSublevelInfo(cmp base.Compare, levelFiles manifest.LevelSlice) []sublevelInfo {
150 1 : sublevelMap := make(map[uint64][]*fileMetadata)
151 1 : it := levelFiles.Iter()
152 1 : for f := it.First(); f != nil; f = it.Next() {
153 1 : sublevelMap[uint64(f.SubLevel)] = append(sublevelMap[uint64(f.SubLevel)], f)
154 1 : }
155 :
156 1 : var sublevels []int
157 1 : for level := range sublevelMap {
158 1 : sublevels = append(sublevels, int(level))
159 1 : }
160 1 : sort.Ints(sublevels)
161 1 :
162 1 : var levelSlices []sublevelInfo
163 1 : for _, sublevel := range sublevels {
164 1 : metas := sublevelMap[uint64(sublevel)]
165 1 : levelSlices = append(
166 1 : levelSlices,
167 1 : sublevelInfo{
168 1 : manifest.NewLevelSliceKeySorted(cmp, metas),
169 1 : manifest.L0Sublevel(sublevel),
170 1 : },
171 1 : )
172 1 : }
173 1 : return levelSlices
174 : }
175 :
176 : // compactionPickerMetrics holds metrics related to the compaction picking process
177 : type compactionPickerMetrics struct {
178 : // scores contains the compensatedScoreRatio from the candidateLevelInfo.
179 : scores []float64
180 : singleLevelOverlappingRatio float64
181 : multiLevelOverlappingRatio float64
182 : }
183 :
184 : // pickedCompaction contains information about a compaction that has already
185 : // been chosen, and is being constructed. Compaction construction info lives in
186 : // this struct, and is copied over into the compaction struct when that's
187 : // created.
188 : type pickedCompaction struct {
189 : cmp Compare
190 : // score of the chosen compaction. This is the same as the
191 : // compensatedScoreRatio in the candidateLevelInfo.
192 : score float64
193 : // kind indicates the kind of compaction.
194 : kind compactionKind
195 : // startLevel is the level that is being compacted. Inputs from startLevel
196 : // and outputLevel will be merged to produce a set of outputLevel files.
197 : startLevel *compactionLevel
198 : // outputLevel is the level that files are being produced in. outputLevel is
199 : // equal to startLevel+1 except when:
200 : // - if startLevel is 0, the output level equals compactionPicker.baseLevel().
201 : // - in multilevel compaction, the output level is the lowest level involved in
202 : // the compaction
203 : outputLevel *compactionLevel
204 : // extraLevels contain additional levels in between the input and output
205 : // levels that get compacted in multi level compactions
206 : extraLevels []*compactionLevel
207 : inputs []compactionLevel
208 : // LBase at the time of compaction picking.
209 : baseLevel int
210 : // L0-specific compaction info. Set to a non-nil value for all compactions
211 : // where startLevel == 0 that were generated by L0Sublevels.
212 : lcf *manifest.L0CompactionFiles
213 : // maxOutputFileSize is the maximum size of an individual table created
214 : // during compaction.
215 : maxOutputFileSize uint64
216 : // maxOverlapBytes is the maximum number of bytes of overlap allowed for a
217 : // single output table with the tables in the grandparent level.
218 : maxOverlapBytes uint64
219 : // maxReadCompactionBytes is the maximum bytes a read compaction is allowed to
220 : // overlap in its output level with. If the overlap is greater than
221 : // maxReadCompaction bytes, then we don't proceed with the compaction.
222 : maxReadCompactionBytes uint64
223 : // The boundaries of the input data.
224 : smallest InternalKey
225 : largest InternalKey
226 : version *version
227 : pickerMetrics compactionPickerMetrics
228 : }
229 :
230 1 : func (pc *pickedCompaction) userKeyBounds() base.UserKeyBounds {
231 1 : return base.UserKeyBoundsFromInternal(pc.smallest, pc.largest)
232 1 : }
233 :
234 1 : func defaultOutputLevel(startLevel, baseLevel int) int {
235 1 : outputLevel := startLevel + 1
236 1 : if startLevel == 0 {
237 1 : outputLevel = baseLevel
238 1 : }
239 1 : if outputLevel >= numLevels-1 {
240 1 : outputLevel = numLevels - 1
241 1 : }
242 1 : return outputLevel
243 : }
244 :
245 : func newPickedCompaction(
246 : opts *Options, cur *version, startLevel, outputLevel, baseLevel int,
247 1 : ) *pickedCompaction {
248 1 : if startLevel > 0 && startLevel < baseLevel {
249 0 : panic(fmt.Sprintf("invalid compaction: start level %d should not be empty (base level %d)",
250 0 : startLevel, baseLevel))
251 : }
252 :
253 1 : adjustedLevel := adjustedOutputLevel(outputLevel, baseLevel)
254 1 : pc := &pickedCompaction{
255 1 : cmp: opts.Comparer.Compare,
256 1 : version: cur,
257 1 : baseLevel: baseLevel,
258 1 : inputs: []compactionLevel{{level: startLevel}, {level: outputLevel}},
259 1 : maxOutputFileSize: uint64(opts.Level(adjustedLevel).TargetFileSize),
260 1 : maxOverlapBytes: maxGrandparentOverlapBytes(opts, adjustedLevel),
261 1 : maxReadCompactionBytes: maxReadCompactionBytes(opts, adjustedLevel),
262 1 : }
263 1 : pc.startLevel = &pc.inputs[0]
264 1 : pc.outputLevel = &pc.inputs[1]
265 1 : return pc
266 : }
267 :
268 : // adjustedOutputLevel is the output level used for the purpose of
269 : // determining the target output file size, overlap bytes, and expanded
270 : // bytes, taking into account the base level.
271 1 : func adjustedOutputLevel(outputLevel int, baseLevel int) int {
272 1 : adjustedOutputLevel := outputLevel
273 1 : if adjustedOutputLevel > 0 {
274 1 : // Output level is in the range [baseLevel, numLevels]. For the purpose of
275 1 : // determining the target output file size, overlap bytes, and expanded
276 1 : // bytes, we want to adjust the range to [1,numLevels].
277 1 : adjustedOutputLevel = 1 + outputLevel - baseLevel
278 1 : }
279 1 : return adjustedOutputLevel
280 : }
281 :
282 : func newPickedCompactionFromL0(
283 : lcf *manifest.L0CompactionFiles, opts *Options, vers *version, baseLevel int, isBase bool,
284 1 : ) *pickedCompaction {
285 1 : outputLevel := baseLevel
286 1 : if !isBase {
287 1 : outputLevel = 0 // Intra L0
288 1 : }
289 :
290 1 : pc := newPickedCompaction(opts, vers, 0, outputLevel, baseLevel)
291 1 : pc.lcf = lcf
292 1 : pc.outputLevel.level = outputLevel
293 1 :
294 1 : // Manually build the compaction as opposed to calling
295 1 : // pickAutoHelper. This is because L0Sublevels has already added
296 1 : // any overlapping L0 SSTables that need to be added, and
297 1 : // because compactions built by L0SSTables do not necessarily
298 1 : // pick contiguous sequences of files in pc.version.Levels[0].
299 1 : files := make([]*manifest.FileMetadata, 0, len(lcf.Files))
300 1 : iter := vers.Levels[0].Iter()
301 1 : for f := iter.First(); f != nil; f = iter.Next() {
302 1 : if lcf.FilesIncluded[f.L0Index] {
303 1 : files = append(files, f)
304 1 : }
305 : }
306 1 : pc.startLevel.files = manifest.NewLevelSliceSeqSorted(files)
307 1 : return pc
308 : }
309 :
310 0 : func (pc *pickedCompaction) String() string {
311 0 : var builder strings.Builder
312 0 : builder.WriteString(fmt.Sprintf(`Score=%f, `, pc.score))
313 0 : builder.WriteString(fmt.Sprintf(`Kind=%s, `, pc.kind))
314 0 : builder.WriteString(fmt.Sprintf(`AdjustedOutputLevel=%d, `, adjustedOutputLevel(pc.outputLevel.level, pc.baseLevel)))
315 0 : builder.WriteString(fmt.Sprintf(`maxOutputFileSize=%d, `, pc.maxOutputFileSize))
316 0 : builder.WriteString(fmt.Sprintf(`maxReadCompactionBytes=%d, `, pc.maxReadCompactionBytes))
317 0 : builder.WriteString(fmt.Sprintf(`smallest=%s, `, pc.smallest))
318 0 : builder.WriteString(fmt.Sprintf(`largest=%s, `, pc.largest))
319 0 : builder.WriteString(fmt.Sprintf(`version=%s, `, pc.version))
320 0 : builder.WriteString(fmt.Sprintf(`inputs=%s, `, pc.inputs))
321 0 : builder.WriteString(fmt.Sprintf(`startlevel=%s, `, pc.startLevel))
322 0 : builder.WriteString(fmt.Sprintf(`outputLevel=%s, `, pc.outputLevel))
323 0 : builder.WriteString(fmt.Sprintf(`extraLevels=%s, `, pc.extraLevels))
324 0 : builder.WriteString(fmt.Sprintf(`l0SublevelInfo=%s, `, pc.startLevel.l0SublevelInfo))
325 0 : builder.WriteString(fmt.Sprintf(`lcf=%s`, pc.lcf))
326 0 : return builder.String()
327 0 : }
328 :
329 : // Clone creates a deep copy of the pickedCompaction
330 1 : func (pc *pickedCompaction) clone() *pickedCompaction {
331 1 :
332 1 : // Quickly copy over fields that do not require special deep copy care, and
333 1 : // set all fields that will require a deep copy to nil.
334 1 : newPC := &pickedCompaction{
335 1 : cmp: pc.cmp,
336 1 : score: pc.score,
337 1 : kind: pc.kind,
338 1 : baseLevel: pc.baseLevel,
339 1 : maxOutputFileSize: pc.maxOutputFileSize,
340 1 : maxOverlapBytes: pc.maxOverlapBytes,
341 1 : maxReadCompactionBytes: pc.maxReadCompactionBytes,
342 1 : smallest: pc.smallest.Clone(),
343 1 : largest: pc.largest.Clone(),
344 1 :
345 1 : // TODO(msbutler): properly clone picker metrics
346 1 : pickerMetrics: pc.pickerMetrics,
347 1 :
348 1 : // Both copies see the same manifest, therefore, it's ok for them to se
349 1 : // share the same pc. version.
350 1 : version: pc.version,
351 1 : }
352 1 :
353 1 : newPC.inputs = make([]compactionLevel, len(pc.inputs))
354 1 : newPC.extraLevels = make([]*compactionLevel, 0, len(pc.extraLevels))
355 1 : for i := range pc.inputs {
356 1 : newPC.inputs[i] = pc.inputs[i].Clone()
357 1 : if i == 0 {
358 1 : newPC.startLevel = &newPC.inputs[i]
359 1 : } else if i == len(pc.inputs)-1 {
360 1 : newPC.outputLevel = &newPC.inputs[i]
361 1 : } else {
362 0 : newPC.extraLevels = append(newPC.extraLevels, &newPC.inputs[i])
363 0 : }
364 : }
365 :
366 1 : if len(pc.startLevel.l0SublevelInfo) > 0 {
367 1 : newPC.startLevel.l0SublevelInfo = make([]sublevelInfo, len(pc.startLevel.l0SublevelInfo))
368 1 : for i := range pc.startLevel.l0SublevelInfo {
369 1 : newPC.startLevel.l0SublevelInfo[i] = pc.startLevel.l0SublevelInfo[i].Clone()
370 1 : }
371 : }
372 1 : if pc.lcf != nil {
373 1 : newPC.lcf = pc.lcf.Clone()
374 1 : }
375 1 : return newPC
376 : }
377 :
378 : // maybeExpandBounds is a helper function for setupInputs which ensures the
379 : // pickedCompaction's smallest and largest internal keys are updated iff
380 : // the candidate keys expand the key span. This avoids a bug for multi-level
381 : // compactions: during the second call to setupInputs, the picked compaction's
382 : // smallest and largest keys should not decrease the key span.
383 1 : func (pc *pickedCompaction) maybeExpandBounds(smallest InternalKey, largest InternalKey) {
384 1 : if len(smallest.UserKey) == 0 && len(largest.UserKey) == 0 {
385 1 : return
386 1 : }
387 1 : if len(pc.smallest.UserKey) == 0 && len(pc.largest.UserKey) == 0 {
388 1 : pc.smallest = smallest
389 1 : pc.largest = largest
390 1 : return
391 1 : }
392 1 : if base.InternalCompare(pc.cmp, pc.smallest, smallest) >= 0 {
393 1 : pc.smallest = smallest
394 1 : }
395 1 : if base.InternalCompare(pc.cmp, pc.largest, largest) <= 0 {
396 1 : pc.largest = largest
397 1 : }
398 : }
399 :
400 : // setupInputs returns true if a compaction has been set up. It returns false if
401 : // a concurrent compaction is occurring on the start or output level files.
402 : func (pc *pickedCompaction) setupInputs(
403 : opts *Options, diskAvailBytes uint64, startLevel *compactionLevel,
404 1 : ) bool {
405 1 : // maxExpandedBytes is the maximum size of an expanded compaction. If
406 1 : // growing a compaction results in a larger size, the original compaction
407 1 : // is used instead.
408 1 : maxExpandedBytes := expandedCompactionByteSizeLimit(
409 1 : opts, adjustedOutputLevel(pc.outputLevel.level, pc.baseLevel), diskAvailBytes,
410 1 : )
411 1 :
412 1 : if anyTablesCompacting(startLevel.files) {
413 1 : return false
414 1 : }
415 :
416 1 : pc.maybeExpandBounds(manifest.KeyRange(pc.cmp, startLevel.files.Iter()))
417 1 :
418 1 : // Determine the sstables in the output level which overlap with the input
419 1 : // sstables. No need to do this for intra-L0 compactions; outputLevel.files is
420 1 : // left empty for those.
421 1 : if startLevel.level != pc.outputLevel.level {
422 1 : pc.outputLevel.files = pc.version.Overlaps(pc.outputLevel.level, pc.userKeyBounds())
423 1 : if anyTablesCompacting(pc.outputLevel.files) {
424 1 : return false
425 1 : }
426 :
427 1 : pc.maybeExpandBounds(manifest.KeyRange(pc.cmp,
428 1 : startLevel.files.Iter(), pc.outputLevel.files.Iter()))
429 : }
430 :
431 : // Grow the sstables in startLevel.level as long as it doesn't affect the number
432 : // of sstables included from pc.outputLevel.level.
433 1 : if pc.lcf != nil && startLevel.level == 0 && pc.outputLevel.level != 0 {
434 1 : // Call the L0-specific compaction extension method. Similar logic as
435 1 : // pc.grow. Additional L0 files are optionally added to the compaction at
436 1 : // this step. Note that the bounds passed in are not the bounds of the
437 1 : // compaction, but rather the smallest and largest internal keys that
438 1 : // the compaction cannot include from L0 without pulling in more Lbase
439 1 : // files. Consider this example:
440 1 : //
441 1 : // L0: c-d e+f g-h
442 1 : // Lbase: a-b e+f i-j
443 1 : // a b c d e f g h i j
444 1 : //
445 1 : // The e-f files have already been chosen in the compaction. As pulling
446 1 : // in more LBase files is undesirable, the logic below will pass in
447 1 : // smallest = b and largest = i to ExtendL0ForBaseCompactionTo, which
448 1 : // will expand the compaction to include c-d and g-h from L0. The
449 1 : // bounds passed in are exclusive; the compaction cannot be expanded
450 1 : // to include files that "touch" it.
451 1 : smallestBaseKey := base.InvalidInternalKey
452 1 : largestBaseKey := base.InvalidInternalKey
453 1 : if pc.outputLevel.files.Empty() {
454 1 : baseIter := pc.version.Levels[pc.outputLevel.level].Iter()
455 1 : if sm := baseIter.SeekLT(pc.cmp, pc.smallest.UserKey); sm != nil {
456 1 : smallestBaseKey = sm.Largest
457 1 : }
458 1 : if la := baseIter.SeekGE(pc.cmp, pc.largest.UserKey); la != nil {
459 1 : largestBaseKey = la.Smallest
460 1 : }
461 1 : } else {
462 1 : // NB: We use Reslice to access the underlying level's files, but
463 1 : // we discard the returned slice. The pc.outputLevel.files slice
464 1 : // is not modified.
465 1 : _ = pc.outputLevel.files.Reslice(func(start, end *manifest.LevelIterator) {
466 1 : if sm := start.Prev(); sm != nil {
467 1 : smallestBaseKey = sm.Largest
468 1 : }
469 1 : if la := end.Next(); la != nil {
470 1 : largestBaseKey = la.Smallest
471 1 : }
472 : })
473 : }
474 1 : oldLcf := pc.lcf.Clone()
475 1 : if pc.version.L0Sublevels.ExtendL0ForBaseCompactionTo(smallestBaseKey, largestBaseKey, pc.lcf) {
476 1 : var newStartLevelFiles []*fileMetadata
477 1 : iter := pc.version.Levels[0].Iter()
478 1 : var sizeSum uint64
479 1 : for j, f := 0, iter.First(); f != nil; j, f = j+1, iter.Next() {
480 1 : if pc.lcf.FilesIncluded[f.L0Index] {
481 1 : newStartLevelFiles = append(newStartLevelFiles, f)
482 1 : sizeSum += f.Size
483 1 : }
484 : }
485 1 : if sizeSum+pc.outputLevel.files.SizeSum() < maxExpandedBytes {
486 1 : startLevel.files = manifest.NewLevelSliceSeqSorted(newStartLevelFiles)
487 1 : pc.smallest, pc.largest = manifest.KeyRange(pc.cmp,
488 1 : startLevel.files.Iter(), pc.outputLevel.files.Iter())
489 1 : } else {
490 1 : *pc.lcf = *oldLcf
491 1 : }
492 : }
493 1 : } else if pc.grow(pc.smallest, pc.largest, maxExpandedBytes, startLevel) {
494 1 : pc.maybeExpandBounds(manifest.KeyRange(pc.cmp,
495 1 : startLevel.files.Iter(), pc.outputLevel.files.Iter()))
496 1 : }
497 :
498 1 : if pc.startLevel.level == 0 {
499 1 : // We don't change the input files for the compaction beyond this point.
500 1 : pc.startLevel.l0SublevelInfo = generateSublevelInfo(pc.cmp, pc.startLevel.files)
501 1 : }
502 :
503 1 : return true
504 : }
505 :
506 : // grow grows the number of inputs at c.level without changing the number of
507 : // c.level+1 files in the compaction, and returns whether the inputs grew. sm
508 : // and la are the smallest and largest InternalKeys in all of the inputs.
509 : func (pc *pickedCompaction) grow(
510 : sm, la InternalKey, maxExpandedBytes uint64, startLevel *compactionLevel,
511 1 : ) bool {
512 1 : if pc.outputLevel.files.Empty() {
513 1 : return false
514 1 : }
515 1 : grow0 := pc.version.Overlaps(startLevel.level, base.UserKeyBoundsFromInternal(sm, la))
516 1 : if anyTablesCompacting(grow0) {
517 1 : return false
518 1 : }
519 1 : if grow0.Len() <= startLevel.files.Len() {
520 1 : return false
521 1 : }
522 1 : if grow0.SizeSum()+pc.outputLevel.files.SizeSum() >= maxExpandedBytes {
523 1 : return false
524 1 : }
525 : // We need to include the outputLevel iter because without it, in a multiLevel scenario,
526 : // sm1 and la1 could shift the output level keyspace when pc.outputLevel.files is set to grow1.
527 1 : sm1, la1 := manifest.KeyRange(pc.cmp, grow0.Iter(), pc.outputLevel.files.Iter())
528 1 : grow1 := pc.version.Overlaps(pc.outputLevel.level, base.UserKeyBoundsFromInternal(sm1, la1))
529 1 : if anyTablesCompacting(grow1) {
530 1 : return false
531 1 : }
532 1 : if grow1.Len() != pc.outputLevel.files.Len() {
533 1 : return false
534 1 : }
535 1 : startLevel.files = grow0
536 1 : pc.outputLevel.files = grow1
537 1 : return true
538 : }
539 :
540 1 : func (pc *pickedCompaction) compactionSize() uint64 {
541 1 : var bytesToCompact uint64
542 1 : for i := range pc.inputs {
543 1 : bytesToCompact += pc.inputs[i].files.SizeSum()
544 1 : }
545 1 : return bytesToCompact
546 : }
547 :
548 : // setupMultiLevelCandidated returns true if it successfully added another level
549 : // to the compaction.
550 1 : func (pc *pickedCompaction) setupMultiLevelCandidate(opts *Options, diskAvailBytes uint64) bool {
551 1 : pc.inputs = append(pc.inputs, compactionLevel{level: pc.outputLevel.level + 1})
552 1 :
553 1 : // Recalibrate startLevel and outputLevel:
554 1 : // - startLevel and outputLevel pointers may be obsolete after appending to pc.inputs.
555 1 : // - push outputLevel to extraLevels and move the new level to outputLevel
556 1 : pc.startLevel = &pc.inputs[0]
557 1 : pc.extraLevels = []*compactionLevel{&pc.inputs[1]}
558 1 : pc.outputLevel = &pc.inputs[2]
559 1 : return pc.setupInputs(opts, diskAvailBytes, pc.extraLevels[len(pc.extraLevels)-1])
560 1 : }
561 :
562 : // anyTablesCompacting returns true if any tables in the level slice are
563 : // compacting.
564 1 : func anyTablesCompacting(inputs manifest.LevelSlice) bool {
565 1 : it := inputs.Iter()
566 1 : for f := it.First(); f != nil; f = it.Next() {
567 1 : if f.IsCompacting() {
568 1 : return true
569 1 : }
570 : }
571 1 : return false
572 : }
573 :
574 : // newCompactionPickerByScore creates a compactionPickerByScore associated with
575 : // the newest version. The picker is used under logLock (until a new version is
576 : // installed).
577 : func newCompactionPickerByScore(
578 : v *version,
579 : virtualBackings *manifest.VirtualBackings,
580 : opts *Options,
581 : inProgressCompactions []compactionInfo,
582 1 : ) *compactionPickerByScore {
583 1 : p := &compactionPickerByScore{
584 1 : opts: opts,
585 1 : vers: v,
586 1 : virtualBackings: virtualBackings,
587 1 : }
588 1 : p.initLevelMaxBytes(inProgressCompactions)
589 1 : p.initTombstoneDensityAnnotator(opts)
590 1 : return p
591 1 : }
592 :
593 : // Information about a candidate compaction level that has been identified by
594 : // the compaction picker.
595 : type candidateLevelInfo struct {
596 : // The compensatedScore of the level after adjusting according to the other
597 : // levels' sizes. For L0, the compensatedScoreRatio is equivalent to the
598 : // uncompensatedScoreRatio as we don't account for level size compensation in
599 : // L0.
600 : compensatedScoreRatio float64
601 : // The score of the level after accounting for level size compensation before
602 : // adjusting according to other levels' sizes. For L0, the compensatedScore
603 : // is equivalent to the uncompensatedScore as we don't account for level
604 : // size compensation in L0.
605 : compensatedScore float64
606 : // The score of the level to be compacted, calculated using uncompensated file
607 : // sizes and without any adjustments.
608 : uncompensatedScore float64
609 : // uncompensatedScoreRatio is the uncompensatedScore adjusted according to
610 : // the other levels' sizes.
611 : uncompensatedScoreRatio float64
612 : level int
613 : // The level to compact to.
614 : outputLevel int
615 : // The file in level that will be compacted. Additional files may be
616 : // picked by the compaction, and a pickedCompaction created for the
617 : // compaction.
618 : file manifest.LevelFile
619 : }
620 :
621 1 : func (c *candidateLevelInfo) shouldCompact() bool {
622 1 : return c.compensatedScoreRatio >= compactionScoreThreshold
623 1 : }
624 :
625 1 : func fileCompensation(f *fileMetadata) uint64 {
626 1 : return uint64(f.Stats.PointDeletionsBytesEstimate) + f.Stats.RangeDeletionsBytesEstimate
627 1 : }
628 :
629 : // compensatedSize returns f's file size, inflated according to compaction
630 : // priorities.
631 1 : func compensatedSize(f *fileMetadata) uint64 {
632 1 : // Add in the estimate of disk space that may be reclaimed by compacting the
633 1 : // file's tombstones.
634 1 : return f.Size + fileCompensation(f)
635 1 : }
636 :
637 : // compensatedSizeAnnotator is a manifest.Annotator that annotates B-Tree
638 : // nodes with the sum of the files' compensated sizes. Compensated sizes may
639 : // change once a table's stats are loaded asynchronously, so its values are
640 : // marked as cacheable only if a file's stats have been loaded.
641 1 : var compensatedSizeAnnotator = manifest.SumAnnotator(func(f *fileMetadata) (uint64, bool) {
642 1 : return compensatedSize(f), f.StatsValid()
643 1 : })
644 :
645 : // totalCompensatedSize computes the compensated size over a file metadata
646 : // iterator. Note that this function is linear in the files available to the
647 : // iterator. Use the compensatedSizeAnnotator if querying the total
648 : // compensated size of a level.
649 1 : func totalCompensatedSize(iter manifest.LevelIterator) uint64 {
650 1 : var sz uint64
651 1 : for f := iter.First(); f != nil; f = iter.Next() {
652 1 : sz += compensatedSize(f)
653 1 : }
654 1 : return sz
655 : }
656 :
657 : // compactionPickerByScore holds the state and logic for picking a compaction. A
658 : // compaction picker is associated with a single version. A new compaction
659 : // picker is created and initialized every time a new version is installed.
660 : type compactionPickerByScore struct {
661 : opts *Options
662 : vers *version
663 : virtualBackings *manifest.VirtualBackings
664 : // The level to target for L0 compactions. Levels L1 to baseLevel must be
665 : // empty.
666 : baseLevel int
667 : // levelMaxBytes holds the dynamically adjusted max bytes setting for each
668 : // level.
669 : levelMaxBytes [numLevels]int64
670 : // tombstoneDensityAnnotator holds the annotator for choosing tombstone
671 : // density compactions.
672 : // NB: This is declared here rather than globally because
673 : // options.Experimental.MinTombstoneDenseRatio is not known until runtime.
674 : tombstoneDensityAnnotator *manifest.Annotator[fileMetadata]
675 : }
676 :
677 : var _ compactionPicker = &compactionPickerByScore{}
678 :
679 1 : func (p *compactionPickerByScore) getScores(inProgress []compactionInfo) [numLevels]float64 {
680 1 : var scores [numLevels]float64
681 1 : for _, info := range p.calculateLevelScores(inProgress) {
682 1 : scores[info.level] = info.compensatedScoreRatio
683 1 : }
684 1 : return scores
685 : }
686 :
687 1 : func (p *compactionPickerByScore) getBaseLevel() int {
688 1 : if p == nil {
689 0 : return 1
690 0 : }
691 1 : return p.baseLevel
692 : }
693 :
694 : // estimatedCompactionDebt estimates the number of bytes which need to be
695 : // compacted before the LSM tree becomes stable.
696 1 : func (p *compactionPickerByScore) estimatedCompactionDebt(l0ExtraSize uint64) uint64 {
697 1 : if p == nil {
698 0 : return 0
699 0 : }
700 :
701 : // We assume that all the bytes in L0 need to be compacted to Lbase. This is
702 : // unlike the RocksDB logic that figures out whether L0 needs compaction.
703 1 : bytesAddedToNextLevel := l0ExtraSize + p.vers.Levels[0].Size()
704 1 : lbaseSize := p.vers.Levels[p.baseLevel].Size()
705 1 :
706 1 : var compactionDebt uint64
707 1 : if bytesAddedToNextLevel > 0 && lbaseSize > 0 {
708 1 : // We only incur compaction debt if both L0 and Lbase contain data. If L0
709 1 : // is empty, no compaction is necessary. If Lbase is empty, a move-based
710 1 : // compaction from L0 would occur.
711 1 : compactionDebt += bytesAddedToNextLevel + lbaseSize
712 1 : }
713 :
714 : // loop invariant: At the beginning of the loop, bytesAddedToNextLevel is the
715 : // bytes added to `level` in the loop.
716 1 : for level := p.baseLevel; level < numLevels-1; level++ {
717 1 : levelSize := p.vers.Levels[level].Size() + bytesAddedToNextLevel
718 1 : nextLevelSize := p.vers.Levels[level+1].Size()
719 1 : if levelSize > uint64(p.levelMaxBytes[level]) {
720 1 : bytesAddedToNextLevel = levelSize - uint64(p.levelMaxBytes[level])
721 1 : if nextLevelSize > 0 {
722 1 : // We only incur compaction debt if the next level contains data. If the
723 1 : // next level is empty, a move-based compaction would be used.
724 1 : levelRatio := float64(nextLevelSize) / float64(levelSize)
725 1 : // The current level contributes bytesAddedToNextLevel to compactions.
726 1 : // The next level contributes levelRatio * bytesAddedToNextLevel.
727 1 : compactionDebt += uint64(float64(bytesAddedToNextLevel) * (levelRatio + 1))
728 1 : }
729 1 : } else {
730 1 : // We're not moving any bytes to the next level.
731 1 : bytesAddedToNextLevel = 0
732 1 : }
733 : }
734 1 : return compactionDebt
735 : }
736 :
737 1 : func (p *compactionPickerByScore) initLevelMaxBytes(inProgressCompactions []compactionInfo) {
738 1 : // The levelMaxBytes calculations here differ from RocksDB in two ways:
739 1 : //
740 1 : // 1. The use of dbSize vs maxLevelSize. RocksDB uses the size of the maximum
741 1 : // level in L1-L6, rather than determining the size of the bottom level
742 1 : // based on the total amount of data in the dB. The RocksDB calculation is
743 1 : // problematic if L0 contains a significant fraction of data, or if the
744 1 : // level sizes are roughly equal and thus there is a significant fraction
745 1 : // of data outside of the largest level.
746 1 : //
747 1 : // 2. Not adjusting the size of Lbase based on L0. RocksDB computes
748 1 : // baseBytesMax as the maximum of the configured LBaseMaxBytes and the
749 1 : // size of L0. This is problematic because baseBytesMax is used to compute
750 1 : // the max size of lower levels. A very large baseBytesMax will result in
751 1 : // an overly large value for the size of lower levels which will caused
752 1 : // those levels not to be compacted even when they should be
753 1 : // compacted. This often results in "inverted" LSM shapes where Ln is
754 1 : // larger than Ln+1.
755 1 :
756 1 : // Determine the first non-empty level and the total DB size.
757 1 : firstNonEmptyLevel := -1
758 1 : var dbSize uint64
759 1 : for level := 1; level < numLevels; level++ {
760 1 : if p.vers.Levels[level].Size() > 0 {
761 1 : if firstNonEmptyLevel == -1 {
762 1 : firstNonEmptyLevel = level
763 1 : }
764 1 : dbSize += p.vers.Levels[level].Size()
765 : }
766 : }
767 1 : for _, c := range inProgressCompactions {
768 1 : if c.outputLevel == 0 || c.outputLevel == -1 {
769 1 : continue
770 : }
771 1 : if c.inputs[0].level == 0 && (firstNonEmptyLevel == -1 || c.outputLevel < firstNonEmptyLevel) {
772 1 : firstNonEmptyLevel = c.outputLevel
773 1 : }
774 : }
775 :
776 : // Initialize the max-bytes setting for each level to "infinity" which will
777 : // disallow compaction for that level. We'll fill in the actual value below
778 : // for levels we want to allow compactions from.
779 1 : for level := 0; level < numLevels; level++ {
780 1 : p.levelMaxBytes[level] = math.MaxInt64
781 1 : }
782 :
783 1 : if dbSize == 0 {
784 1 : // No levels for L1 and up contain any data. Target L0 compactions for the
785 1 : // last level or to the level to which there is an ongoing L0 compaction.
786 1 : p.baseLevel = numLevels - 1
787 1 : if firstNonEmptyLevel >= 0 {
788 1 : p.baseLevel = firstNonEmptyLevel
789 1 : }
790 1 : return
791 : }
792 :
793 1 : dbSize += p.vers.Levels[0].Size()
794 1 : bottomLevelSize := dbSize - dbSize/uint64(p.opts.Experimental.LevelMultiplier)
795 1 :
796 1 : curLevelSize := bottomLevelSize
797 1 : for level := numLevels - 2; level >= firstNonEmptyLevel; level-- {
798 1 : curLevelSize = uint64(float64(curLevelSize) / float64(p.opts.Experimental.LevelMultiplier))
799 1 : }
800 :
801 : // Compute base level (where L0 data is compacted to).
802 1 : baseBytesMax := uint64(p.opts.LBaseMaxBytes)
803 1 : p.baseLevel = firstNonEmptyLevel
804 1 : for p.baseLevel > 1 && curLevelSize > baseBytesMax {
805 1 : p.baseLevel--
806 1 : curLevelSize = uint64(float64(curLevelSize) / float64(p.opts.Experimental.LevelMultiplier))
807 1 : }
808 :
809 1 : smoothedLevelMultiplier := 1.0
810 1 : if p.baseLevel < numLevels-1 {
811 1 : smoothedLevelMultiplier = math.Pow(
812 1 : float64(bottomLevelSize)/float64(baseBytesMax),
813 1 : 1.0/float64(numLevels-p.baseLevel-1))
814 1 : }
815 :
816 1 : levelSize := float64(baseBytesMax)
817 1 : for level := p.baseLevel; level < numLevels; level++ {
818 1 : if level > p.baseLevel && levelSize > 0 {
819 1 : levelSize *= smoothedLevelMultiplier
820 1 : }
821 : // Round the result since test cases use small target level sizes, which
822 : // can be impacted by floating-point imprecision + integer truncation.
823 1 : roundedLevelSize := math.Round(levelSize)
824 1 : if roundedLevelSize > float64(math.MaxInt64) {
825 0 : p.levelMaxBytes[level] = math.MaxInt64
826 1 : } else {
827 1 : p.levelMaxBytes[level] = int64(roundedLevelSize)
828 1 : }
829 : }
830 : }
831 :
832 : type levelSizeAdjust struct {
833 : incomingActualBytes uint64
834 : outgoingActualBytes uint64
835 : outgoingCompensatedBytes uint64
836 : }
837 :
838 1 : func (a levelSizeAdjust) compensated() uint64 {
839 1 : return a.incomingActualBytes - a.outgoingCompensatedBytes
840 1 : }
841 :
842 1 : func (a levelSizeAdjust) actual() uint64 {
843 1 : return a.incomingActualBytes - a.outgoingActualBytes
844 1 : }
845 :
846 1 : func calculateSizeAdjust(inProgressCompactions []compactionInfo) [numLevels]levelSizeAdjust {
847 1 : // Compute size adjustments for each level based on the in-progress
848 1 : // compactions. We sum the file sizes of all files leaving and entering each
849 1 : // level in in-progress compactions. For outgoing files, we also sum a
850 1 : // separate sum of 'compensated file sizes', which are inflated according
851 1 : // to deletion estimates.
852 1 : //
853 1 : // When we adjust a level's size according to these values during score
854 1 : // calculation, we subtract the compensated size of start level inputs to
855 1 : // account for the fact that score calculation uses compensated sizes.
856 1 : //
857 1 : // Since compensated file sizes may be compensated because they reclaim
858 1 : // space from the output level's files, we only add the real file size to
859 1 : // the output level.
860 1 : //
861 1 : // This is slightly different from RocksDB's behavior, which simply elides
862 1 : // compacting files from the level size calculation.
863 1 : var sizeAdjust [numLevels]levelSizeAdjust
864 1 : for i := range inProgressCompactions {
865 1 : c := &inProgressCompactions[i]
866 1 : // If this compaction's version edit has already been applied, there's
867 1 : // no need to adjust: The LSM we'll examine will already reflect the
868 1 : // new LSM state.
869 1 : if c.versionEditApplied {
870 1 : continue
871 : }
872 :
873 1 : for _, input := range c.inputs {
874 1 : actualSize := input.files.SizeSum()
875 1 : compensatedSize := totalCompensatedSize(input.files.Iter())
876 1 :
877 1 : if input.level != c.outputLevel {
878 1 : sizeAdjust[input.level].outgoingCompensatedBytes += compensatedSize
879 1 : sizeAdjust[input.level].outgoingActualBytes += actualSize
880 1 : if c.outputLevel != -1 {
881 1 : sizeAdjust[c.outputLevel].incomingActualBytes += actualSize
882 1 : }
883 : }
884 : }
885 : }
886 1 : return sizeAdjust
887 : }
888 :
889 : func (p *compactionPickerByScore) calculateLevelScores(
890 : inProgressCompactions []compactionInfo,
891 1 : ) [numLevels]candidateLevelInfo {
892 1 : var scores [numLevels]candidateLevelInfo
893 1 : for i := range scores {
894 1 : scores[i].level = i
895 1 : scores[i].outputLevel = i + 1
896 1 : }
897 1 : l0UncompensatedScore := calculateL0UncompensatedScore(p.vers, p.opts, inProgressCompactions)
898 1 : scores[0] = candidateLevelInfo{
899 1 : outputLevel: p.baseLevel,
900 1 : uncompensatedScore: l0UncompensatedScore,
901 1 : compensatedScore: l0UncompensatedScore, /* No level size compensation for L0 */
902 1 : }
903 1 : sizeAdjust := calculateSizeAdjust(inProgressCompactions)
904 1 : for level := 1; level < numLevels; level++ {
905 1 : compensatedLevelSize := *compensatedSizeAnnotator.LevelAnnotation(p.vers.Levels[level]) + sizeAdjust[level].compensated()
906 1 : scores[level].compensatedScore = float64(compensatedLevelSize) / float64(p.levelMaxBytes[level])
907 1 : scores[level].uncompensatedScore = float64(p.vers.Levels[level].Size()+sizeAdjust[level].actual()) / float64(p.levelMaxBytes[level])
908 1 : }
909 :
910 : // Adjust each level's {compensated, uncompensated}Score by the uncompensatedScore
911 : // of the next level to get a {compensated, uncompensated}ScoreRatio. If the
912 : // next level has a high uncompensatedScore, and is thus a priority for compaction,
913 : // this reduces the priority for compacting the current level. If the next level
914 : // has a low uncompensatedScore (i.e. it is below its target size), this increases
915 : // the priority for compacting the current level.
916 : //
917 : // The effect of this adjustment is to help prioritize compactions in lower
918 : // levels. The following example shows the compensatedScoreRatio and the
919 : // compensatedScore. In this scenario, L0 has 68 sublevels. L3 (a.k.a. Lbase)
920 : // is significantly above its target size. The original score prioritizes
921 : // compactions from those two levels, but doing so ends up causing a future
922 : // problem: data piles up in the higher levels, starving L5->L6 compactions,
923 : // and to a lesser degree starving L4->L5 compactions.
924 : //
925 : // Note that in the example shown there is no level size compensation so the
926 : // compensatedScore and the uncompensatedScore is the same for each level.
927 : //
928 : // compensatedScoreRatio compensatedScore uncompensatedScore size max-size
929 : // L0 3.2 68.0 68.0 2.2 G -
930 : // L3 3.2 21.1 21.1 1.3 G 64 M
931 : // L4 3.4 6.7 6.7 3.1 G 467 M
932 : // L5 3.4 2.0 2.0 6.6 G 3.3 G
933 : // L6 0.6 0.6 0.6 14 G 24 G
934 1 : var prevLevel int
935 1 : for level := p.baseLevel; level < numLevels; level++ {
936 1 : // The compensated scores, and uncompensated scores will be turned into
937 1 : // ratios as they're adjusted according to other levels' sizes.
938 1 : scores[prevLevel].compensatedScoreRatio = scores[prevLevel].compensatedScore
939 1 : scores[prevLevel].uncompensatedScoreRatio = scores[prevLevel].uncompensatedScore
940 1 :
941 1 : // Avoid absurdly large scores by placing a floor on the score that we'll
942 1 : // adjust a level by. The value of 0.01 was chosen somewhat arbitrarily.
943 1 : const minScore = 0.01
944 1 : if scores[prevLevel].compensatedScoreRatio >= compactionScoreThreshold {
945 1 : if scores[level].uncompensatedScore >= minScore {
946 1 : scores[prevLevel].compensatedScoreRatio /= scores[level].uncompensatedScore
947 1 : } else {
948 1 : scores[prevLevel].compensatedScoreRatio /= minScore
949 1 : }
950 : }
951 1 : if scores[prevLevel].uncompensatedScoreRatio >= compactionScoreThreshold {
952 1 : if scores[level].uncompensatedScore >= minScore {
953 1 : scores[prevLevel].uncompensatedScoreRatio /= scores[level].uncompensatedScore
954 1 : } else {
955 1 : scores[prevLevel].uncompensatedScoreRatio /= minScore
956 1 : }
957 : }
958 1 : prevLevel = level
959 : }
960 : // Set the score ratios for the lowest level.
961 : // INVARIANT: prevLevel == numLevels-1
962 1 : scores[prevLevel].compensatedScoreRatio = scores[prevLevel].compensatedScore
963 1 : scores[prevLevel].uncompensatedScoreRatio = scores[prevLevel].uncompensatedScore
964 1 :
965 1 : sort.Sort(sortCompactionLevelsByPriority(scores[:]))
966 1 : return scores
967 : }
968 :
969 : // calculateL0UncompensatedScore calculates a float score representing the
970 : // relative priority of compacting L0. Level L0 is special in that files within
971 : // L0 may overlap one another, so a different set of heuristics that take into
972 : // account read amplification apply.
973 : func calculateL0UncompensatedScore(
974 : vers *version, opts *Options, inProgressCompactions []compactionInfo,
975 1 : ) float64 {
976 1 : // Use the sublevel count to calculate the score. The base vs intra-L0
977 1 : // compaction determination happens in pickAuto, not here.
978 1 : score := float64(2*vers.L0Sublevels.MaxDepthAfterOngoingCompactions()) /
979 1 : float64(opts.L0CompactionThreshold)
980 1 :
981 1 : // Also calculate a score based on the file count but use it only if it
982 1 : // produces a higher score than the sublevel-based one. This heuristic is
983 1 : // designed to accommodate cases where L0 is accumulating non-overlapping
984 1 : // files in L0. Letting too many non-overlapping files accumulate in few
985 1 : // sublevels is undesirable, because:
986 1 : // 1) we can produce a massive backlog to compact once files do overlap.
987 1 : // 2) constructing L0 sublevels has a runtime that grows superlinearly with
988 1 : // the number of files in L0 and must be done while holding D.mu.
989 1 : noncompactingFiles := vers.Levels[0].Len()
990 1 : for _, c := range inProgressCompactions {
991 1 : for _, cl := range c.inputs {
992 1 : if cl.level == 0 {
993 1 : noncompactingFiles -= cl.files.Len()
994 1 : }
995 : }
996 : }
997 1 : fileScore := float64(noncompactingFiles) / float64(opts.L0CompactionFileThreshold)
998 1 : if score < fileScore {
999 1 : score = fileScore
1000 1 : }
1001 1 : return score
1002 : }
1003 :
1004 : // pickCompactionSeedFile picks a file from `level` in the `vers` to build a
1005 : // compaction around. Currently, this function implements a heuristic similar to
1006 : // RocksDB's kMinOverlappingRatio, seeking to minimize write amplification. This
1007 : // function is linear with respect to the number of files in `level` and
1008 : // `outputLevel`.
1009 : func pickCompactionSeedFile(
1010 : vers *version,
1011 : virtualBackings *manifest.VirtualBackings,
1012 : opts *Options,
1013 : level, outputLevel int,
1014 : earliestSnapshotSeqNum base.SeqNum,
1015 1 : ) (manifest.LevelFile, bool) {
1016 1 : // Select the file within the level to compact. We want to minimize write
1017 1 : // amplification, but also ensure that (a) deletes are propagated to the
1018 1 : // bottom level in a timely fashion, and (b) virtual sstables that are
1019 1 : // pinning backing sstables where most of the data is garbage are compacted
1020 1 : // away. Doing (a) and (b) reclaims disk space. A table's smallest sequence
1021 1 : // number provides a measure of its age. The ratio of overlapping-bytes /
1022 1 : // table-size gives an indication of write amplification (a smaller ratio is
1023 1 : // preferrable).
1024 1 : //
1025 1 : // The current heuristic is based off the RocksDB kMinOverlappingRatio
1026 1 : // heuristic. It chooses the file with the minimum overlapping ratio with
1027 1 : // the target level, which minimizes write amplification.
1028 1 : //
1029 1 : // The heuristic uses a "compensated size" for the denominator, which is the
1030 1 : // file size inflated by (a) an estimate of the space that may be reclaimed
1031 1 : // through compaction, and (b) a fraction of the amount of garbage in the
1032 1 : // backing sstable pinned by this (virtual) sstable.
1033 1 : //
1034 1 : // TODO(peter): For concurrent compactions, we may want to try harder to
1035 1 : // pick a seed file whose resulting compaction bounds do not overlap with
1036 1 : // an in-progress compaction.
1037 1 :
1038 1 : cmp := opts.Comparer.Compare
1039 1 : startIter := vers.Levels[level].Iter()
1040 1 : outputIter := vers.Levels[outputLevel].Iter()
1041 1 :
1042 1 : var file manifest.LevelFile
1043 1 : smallestRatio := uint64(math.MaxUint64)
1044 1 :
1045 1 : outputFile := outputIter.First()
1046 1 :
1047 1 : for f := startIter.First(); f != nil; f = startIter.Next() {
1048 1 : var overlappingBytes uint64
1049 1 : compacting := f.IsCompacting()
1050 1 : if compacting {
1051 1 : // Move on if this file is already being compacted. We'll likely
1052 1 : // still need to move past the overlapping output files regardless,
1053 1 : // but in cases where all start-level files are compacting we won't.
1054 1 : continue
1055 : }
1056 :
1057 : // Trim any output-level files smaller than f.
1058 1 : for outputFile != nil && sstableKeyCompare(cmp, outputFile.Largest, f.Smallest) < 0 {
1059 1 : outputFile = outputIter.Next()
1060 1 : }
1061 :
1062 1 : for outputFile != nil && sstableKeyCompare(cmp, outputFile.Smallest, f.Largest) <= 0 && !compacting {
1063 1 : overlappingBytes += outputFile.Size
1064 1 : compacting = compacting || outputFile.IsCompacting()
1065 1 :
1066 1 : // For files in the bottommost level of the LSM, the
1067 1 : // Stats.RangeDeletionsBytesEstimate field is set to the estimate
1068 1 : // of bytes /within/ the file itself that may be dropped by
1069 1 : // recompacting the file. These bytes from obsolete keys would not
1070 1 : // need to be rewritten if we compacted `f` into `outputFile`, so
1071 1 : // they don't contribute to write amplification. Subtracting them
1072 1 : // out of the overlapping bytes helps prioritize these compactions
1073 1 : // that are cheaper than their file sizes suggest.
1074 1 : if outputLevel == numLevels-1 && outputFile.LargestSeqNum < earliestSnapshotSeqNum {
1075 1 : overlappingBytes -= outputFile.Stats.RangeDeletionsBytesEstimate
1076 1 : }
1077 :
1078 : // If the file in the next level extends beyond f's largest key,
1079 : // break out and don't advance outputIter because f's successor
1080 : // might also overlap.
1081 : //
1082 : // Note, we stop as soon as we encounter an output-level file with a
1083 : // largest key beyond the input-level file's largest bound. We
1084 : // perform a simple user key comparison here using sstableKeyCompare
1085 : // which handles the potential for exclusive largest key bounds.
1086 : // There's some subtlety when the bounds are equal (eg, equal and
1087 : // inclusive, or equal and exclusive). Current Pebble doesn't split
1088 : // user keys across sstables within a level (and in format versions
1089 : // FormatSplitUserKeysMarkedCompacted and later we guarantee no
1090 : // split user keys exist within the entire LSM). In that case, we're
1091 : // assured that neither the input level nor the output level's next
1092 : // file shares the same user key, so compaction expansion will not
1093 : // include them in any compaction compacting `f`.
1094 : //
1095 : // NB: If we /did/ allow split user keys, or we're running on an
1096 : // old database with an earlier format major version where there are
1097 : // existing split user keys, this logic would be incorrect. Consider
1098 : // L1: [a#120,a#100] [a#80,a#60]
1099 : // L2: [a#55,a#45] [a#35,a#25] [a#15,a#5]
1100 : // While considering the first file in L1, [a#120,a#100], we'd skip
1101 : // past all of the files in L2. When considering the second file in
1102 : // L1, we'd improperly conclude that the second file overlaps
1103 : // nothing in the second level and is cheap to compact, when in
1104 : // reality we'd need to expand the compaction to include all 5
1105 : // files.
1106 1 : if sstableKeyCompare(cmp, outputFile.Largest, f.Largest) > 0 {
1107 1 : break
1108 : }
1109 1 : outputFile = outputIter.Next()
1110 : }
1111 :
1112 : // If the input level file or one of the overlapping files is
1113 : // compacting, we're not going to be able to compact this file
1114 : // anyways, so skip it.
1115 1 : if compacting {
1116 1 : continue
1117 : }
1118 :
1119 1 : compSz := compensatedSize(f) + responsibleForGarbageBytes(virtualBackings, f)
1120 1 : scaledRatio := overlappingBytes * 1024 / compSz
1121 1 : if scaledRatio < smallestRatio {
1122 1 : smallestRatio = scaledRatio
1123 1 : file = startIter.Take()
1124 1 : }
1125 : }
1126 1 : return file, file.FileMetadata != nil
1127 : }
1128 :
1129 : // responsibleForGarbageBytes returns the amount of garbage in the backing
1130 : // sstable that we consider the responsibility of this virtual sstable. For
1131 : // non-virtual sstables, this is of course 0. For virtual sstables, we equally
1132 : // distribute the responsibility of the garbage across all the virtual
1133 : // sstables that are referencing the same backing sstable. One could
1134 : // alternatively distribute this in proportion to the virtual sst sizes, but
1135 : // it isn't clear that more sophisticated heuristics are worth it, given that
1136 : // the garbage cannot be reclaimed until all the referencing virtual sstables
1137 : // are compacted.
1138 1 : func responsibleForGarbageBytes(virtualBackings *manifest.VirtualBackings, m *fileMetadata) uint64 {
1139 1 : if !m.Virtual {
1140 1 : return 0
1141 1 : }
1142 1 : useCount, virtualizedSize := virtualBackings.Usage(m.FileBacking.DiskFileNum)
1143 1 : // Since virtualizedSize is the sum of the estimated size of all virtual
1144 1 : // ssts, we allow for the possibility that virtualizedSize could exceed
1145 1 : // m.FileBacking.Size.
1146 1 : totalGarbage := int64(m.FileBacking.Size) - int64(virtualizedSize)
1147 1 : if totalGarbage <= 0 {
1148 1 : return 0
1149 1 : }
1150 1 : if useCount == 0 {
1151 0 : // This cannot happen if m exists in the latest version. The call to
1152 0 : // ResponsibleForGarbageBytes during compaction picking ensures that m
1153 0 : // exists in the latest version by holding versionSet.logLock.
1154 0 : panic(errors.AssertionFailedf("%s has zero useCount", m.String()))
1155 : }
1156 1 : return uint64(totalGarbage) / uint64(useCount)
1157 : }
1158 :
1159 : // pickAuto picks the best compaction, if any.
1160 : //
1161 : // On each call, pickAuto computes per-level size adjustments based on
1162 : // in-progress compactions, and computes a per-level score. The levels are
1163 : // iterated over in decreasing score order trying to find a valid compaction
1164 : // anchored at that level.
1165 : //
1166 : // If a score-based compaction cannot be found, pickAuto falls back to looking
1167 : // for an elision-only compaction to remove obsolete keys.
1168 1 : func (p *compactionPickerByScore) pickAuto(env compactionEnv) (pc *pickedCompaction) {
1169 1 : // Compaction concurrency is controlled by L0 read-amp. We allow one
1170 1 : // additional compaction per L0CompactionConcurrency sublevels, as well as
1171 1 : // one additional compaction per CompactionDebtConcurrency bytes of
1172 1 : // compaction debt. Compaction concurrency is tied to L0 sublevels as that
1173 1 : // signal is independent of the database size. We tack on the compaction
1174 1 : // debt as a second signal to prevent compaction concurrency from dropping
1175 1 : // significantly right after a base compaction finishes, and before those
1176 1 : // bytes have been compacted further down the LSM.
1177 1 : if n := len(env.inProgressCompactions); n > 0 {
1178 1 : l0ReadAmp := p.vers.L0Sublevels.MaxDepthAfterOngoingCompactions()
1179 1 : compactionDebt := p.estimatedCompactionDebt(0)
1180 1 : ccSignal1 := n * p.opts.Experimental.L0CompactionConcurrency
1181 1 : ccSignal2 := uint64(n) * p.opts.Experimental.CompactionDebtConcurrency
1182 1 : if l0ReadAmp < ccSignal1 && compactionDebt < ccSignal2 {
1183 1 : return nil
1184 1 : }
1185 : }
1186 :
1187 1 : scores := p.calculateLevelScores(env.inProgressCompactions)
1188 1 :
1189 1 : // TODO(bananabrick): Either remove, or change this into an event sent to the
1190 1 : // EventListener.
1191 1 : logCompaction := func(pc *pickedCompaction) {
1192 0 : var buf bytes.Buffer
1193 0 : for i := 0; i < numLevels; i++ {
1194 0 : if i != 0 && i < p.baseLevel {
1195 0 : continue
1196 : }
1197 :
1198 0 : var info *candidateLevelInfo
1199 0 : for j := range scores {
1200 0 : if scores[j].level == i {
1201 0 : info = &scores[j]
1202 0 : break
1203 : }
1204 : }
1205 :
1206 0 : marker := " "
1207 0 : if pc.startLevel.level == info.level {
1208 0 : marker = "*"
1209 0 : }
1210 0 : fmt.Fprintf(&buf, " %sL%d: %5.1f %5.1f %5.1f %5.1f %8s %8s",
1211 0 : marker, info.level, info.compensatedScoreRatio, info.compensatedScore,
1212 0 : info.uncompensatedScoreRatio, info.uncompensatedScore,
1213 0 : humanize.Bytes.Int64(int64(totalCompensatedSize(
1214 0 : p.vers.Levels[info.level].Iter(),
1215 0 : ))),
1216 0 : humanize.Bytes.Int64(p.levelMaxBytes[info.level]),
1217 0 : )
1218 0 :
1219 0 : count := 0
1220 0 : for i := range env.inProgressCompactions {
1221 0 : c := &env.inProgressCompactions[i]
1222 0 : if c.inputs[0].level != info.level {
1223 0 : continue
1224 : }
1225 0 : count++
1226 0 : if count == 1 {
1227 0 : fmt.Fprintf(&buf, " [")
1228 0 : } else {
1229 0 : fmt.Fprintf(&buf, " ")
1230 0 : }
1231 0 : fmt.Fprintf(&buf, "L%d->L%d", c.inputs[0].level, c.outputLevel)
1232 : }
1233 0 : if count > 0 {
1234 0 : fmt.Fprintf(&buf, "]")
1235 0 : }
1236 0 : fmt.Fprintf(&buf, "\n")
1237 : }
1238 0 : p.opts.Logger.Infof("pickAuto: L%d->L%d\n%s",
1239 0 : pc.startLevel.level, pc.outputLevel.level, buf.String())
1240 : }
1241 :
1242 : // Check for a score-based compaction. candidateLevelInfos are first sorted
1243 : // by whether they should be compacted, so if we find a level which shouldn't
1244 : // be compacted, we can break early.
1245 1 : for i := range scores {
1246 1 : info := &scores[i]
1247 1 : if !info.shouldCompact() {
1248 1 : break
1249 : }
1250 1 : if info.level == numLevels-1 {
1251 1 : continue
1252 : }
1253 :
1254 1 : if info.level == 0 {
1255 1 : pc = pickL0(env, p.opts, p.vers, p.baseLevel)
1256 1 : // Fail-safe to protect against compacting the same sstable
1257 1 : // concurrently.
1258 1 : if pc != nil && !inputRangeAlreadyCompacting(env, pc) {
1259 1 : p.addScoresToPickedCompactionMetrics(pc, scores)
1260 1 : pc.score = info.compensatedScoreRatio
1261 1 : // TODO(bananabrick): Create an EventListener for logCompaction.
1262 1 : if false {
1263 0 : logCompaction(pc)
1264 0 : }
1265 1 : return pc
1266 : }
1267 1 : continue
1268 : }
1269 :
1270 : // info.level > 0
1271 1 : var ok bool
1272 1 : info.file, ok = pickCompactionSeedFile(p.vers, p.virtualBackings, p.opts, info.level, info.outputLevel, env.earliestSnapshotSeqNum)
1273 1 : if !ok {
1274 1 : continue
1275 : }
1276 :
1277 1 : pc := pickAutoLPositive(env, p.opts, p.vers, *info, p.baseLevel)
1278 1 : // Fail-safe to protect against compacting the same sstable concurrently.
1279 1 : if pc != nil && !inputRangeAlreadyCompacting(env, pc) {
1280 1 : p.addScoresToPickedCompactionMetrics(pc, scores)
1281 1 : pc.score = info.compensatedScoreRatio
1282 1 : // TODO(bananabrick): Create an EventListener for logCompaction.
1283 1 : if false {
1284 0 : logCompaction(pc)
1285 0 : }
1286 1 : return pc
1287 : }
1288 : }
1289 :
1290 : // Check for files which contain excessive point tombstones that could slow
1291 : // down reads. Unlike elision-only compactions, these compactions may select
1292 : // a file at any level rather than only the lowest level.
1293 1 : if pc := p.pickTombstoneDensityCompaction(env); pc != nil {
1294 1 : return pc
1295 1 : }
1296 :
1297 : // Check for L6 files with tombstones that may be elided. These files may
1298 : // exist if a snapshot prevented the elision of a tombstone or because of
1299 : // a move compaction. These are low-priority compactions because they
1300 : // don't help us keep up with writes, just reclaim disk space.
1301 1 : if pc := p.pickElisionOnlyCompaction(env); pc != nil {
1302 1 : return pc
1303 1 : }
1304 :
1305 1 : if pc := p.pickReadTriggeredCompaction(env); pc != nil {
1306 0 : return pc
1307 0 : }
1308 :
1309 : // NB: This should only be run if a read compaction wasn't
1310 : // scheduled.
1311 : //
1312 : // We won't be scheduling a read compaction right now, and in
1313 : // read heavy workloads, compactions won't be scheduled frequently
1314 : // because flushes aren't frequent. So we need to signal to the
1315 : // iterator to schedule a compaction when it adds compactions to
1316 : // the read compaction queue.
1317 : //
1318 : // We need the nil check here because without it, we have some
1319 : // tests which don't set that variable fail. Since there's a
1320 : // chance that one of those tests wouldn't want extra compactions
1321 : // to be scheduled, I added this check here, instead of
1322 : // setting rescheduleReadCompaction in those tests.
1323 1 : if env.readCompactionEnv.rescheduleReadCompaction != nil {
1324 1 : *env.readCompactionEnv.rescheduleReadCompaction = true
1325 1 : }
1326 :
1327 : // At the lowest possible compaction-picking priority, look for files marked
1328 : // for compaction. Pebble will mark files for compaction if they have atomic
1329 : // compaction units that span multiple files. While current Pebble code does
1330 : // not construct such sstables, RocksDB and earlier versions of Pebble may
1331 : // have created them. These split user keys form sets of files that must be
1332 : // compacted together for correctness (referred to as "atomic compaction
1333 : // units" within the code). Rewrite them in-place.
1334 : //
1335 : // It's also possible that a file may have been marked for compaction by
1336 : // even earlier versions of Pebble code, since FileMetadata's
1337 : // MarkedForCompaction field is persisted in the manifest. That's okay. We
1338 : // previously would've ignored the designation, whereas now we'll re-compact
1339 : // the file in place.
1340 1 : if p.vers.Stats.MarkedForCompaction > 0 {
1341 0 : if pc := p.pickRewriteCompaction(env); pc != nil {
1342 0 : return pc
1343 0 : }
1344 : }
1345 :
1346 1 : return nil
1347 : }
1348 :
1349 : func (p *compactionPickerByScore) addScoresToPickedCompactionMetrics(
1350 : pc *pickedCompaction, candInfo [numLevels]candidateLevelInfo,
1351 1 : ) {
1352 1 :
1353 1 : // candInfo is sorted by score, not by compaction level.
1354 1 : infoByLevel := [numLevels]candidateLevelInfo{}
1355 1 : for i := range candInfo {
1356 1 : level := candInfo[i].level
1357 1 : infoByLevel[level] = candInfo[i]
1358 1 : }
1359 : // Gather the compaction scores for the levels participating in the compaction.
1360 1 : pc.pickerMetrics.scores = make([]float64, len(pc.inputs))
1361 1 : inputIdx := 0
1362 1 : for i := range infoByLevel {
1363 1 : if pc.inputs[inputIdx].level == infoByLevel[i].level {
1364 1 : pc.pickerMetrics.scores[inputIdx] = infoByLevel[i].compensatedScoreRatio
1365 1 : inputIdx++
1366 1 : }
1367 1 : if inputIdx == len(pc.inputs) {
1368 1 : break
1369 : }
1370 : }
1371 : }
1372 :
1373 : // elisionOnlyAnnotator is a manifest.Annotator that annotates B-Tree
1374 : // nodes with the *fileMetadata of a file meeting the obsolete keys criteria
1375 : // for an elision-only compaction within the subtree. If multiple files meet
1376 : // the criteria, it chooses whichever file has the lowest LargestSeqNum. The
1377 : // lowest LargestSeqNum file will be the first eligible for an elision-only
1378 : // compaction once snapshots less than or equal to its LargestSeqNum are closed.
1379 : var elisionOnlyAnnotator = &manifest.Annotator[fileMetadata]{
1380 : Aggregator: manifest.PickFileAggregator{
1381 1 : Filter: func(f *fileMetadata) (eligible bool, cacheOK bool) {
1382 1 : if f.IsCompacting() {
1383 1 : return false, true
1384 1 : }
1385 1 : if !f.StatsValid() {
1386 1 : return false, false
1387 1 : }
1388 : // Bottommost files are large and not worthwhile to compact just
1389 : // to remove a few tombstones. Consider a file eligible only if
1390 : // either its own range deletions delete at least 10% of its data or
1391 : // its deletion tombstones make at least 10% of its entries.
1392 : //
1393 : // TODO(jackson): This does not account for duplicate user keys
1394 : // which may be collapsed. Ideally, we would have 'obsolete keys'
1395 : // statistics that would include tombstones, the keys that are
1396 : // dropped by tombstones and duplicated user keys. See #847.
1397 : //
1398 : // Note that tables that contain exclusively range keys (i.e. no point keys,
1399 : // `NumEntries` and `RangeDeletionsBytesEstimate` are both zero) are excluded
1400 : // from elision-only compactions.
1401 : // TODO(travers): Consider an alternative heuristic for elision of range-keys.
1402 1 : return f.Stats.RangeDeletionsBytesEstimate*10 >= f.Size || f.Stats.NumDeletions*10 > f.Stats.NumEntries, true
1403 : },
1404 1 : Compare: func(f1 *fileMetadata, f2 *fileMetadata) bool {
1405 1 : return f1.LargestSeqNum < f2.LargestSeqNum
1406 1 : },
1407 : },
1408 : }
1409 :
1410 : // markedForCompactionAnnotator is a manifest.Annotator that annotates B-Tree
1411 : // nodes with the *fileMetadata of a file that is marked for compaction
1412 : // within the subtree. If multiple files meet the criteria, it chooses
1413 : // whichever file has the lowest LargestSeqNum.
1414 : var markedForCompactionAnnotator = &manifest.Annotator[fileMetadata]{
1415 : Aggregator: manifest.PickFileAggregator{
1416 0 : Filter: func(f *fileMetadata) (eligible bool, cacheOK bool) {
1417 0 : return f.MarkedForCompaction, true
1418 0 : },
1419 0 : Compare: func(f1 *fileMetadata, f2 *fileMetadata) bool {
1420 0 : return f1.LargestSeqNum < f2.LargestSeqNum
1421 0 : },
1422 : },
1423 : }
1424 :
1425 : // pickedCompactionFromCandidateFile creates a pickedCompaction from a *fileMetadata
1426 : // with various checks to ensure that the file still exists in the expected level
1427 : // and isn't already being compacted.
1428 : func (p *compactionPickerByScore) pickedCompactionFromCandidateFile(
1429 : candidate *fileMetadata, env compactionEnv, startLevel int, outputLevel int, kind compactionKind,
1430 1 : ) *pickedCompaction {
1431 1 : if candidate == nil || candidate.IsCompacting() {
1432 1 : return nil
1433 1 : }
1434 :
1435 1 : var inputs manifest.LevelSlice
1436 1 : if startLevel == 0 {
1437 1 : // Overlapping L0 files must also be compacted alongside the candidate.
1438 1 : inputs = p.vers.Overlaps(0, candidate.UserKeyBounds())
1439 1 : } else {
1440 1 : inputs = p.vers.Levels[startLevel].Find(p.opts.Comparer.Compare, candidate)
1441 1 : }
1442 1 : if invariants.Enabled {
1443 1 : found := false
1444 1 : inputs.Each(func(f *fileMetadata) {
1445 1 : if f.FileNum == candidate.FileNum {
1446 1 : found = true
1447 1 : }
1448 : })
1449 1 : if !found {
1450 0 : panic(fmt.Sprintf("file %s not found in level %d as expected", candidate.FileNum, startLevel))
1451 : }
1452 : }
1453 :
1454 1 : pc := newPickedCompaction(p.opts, p.vers, startLevel, outputLevel, p.baseLevel)
1455 1 : pc.kind = kind
1456 1 : pc.startLevel.files = inputs
1457 1 : pc.smallest, pc.largest = manifest.KeyRange(pc.cmp, pc.startLevel.files.Iter())
1458 1 :
1459 1 : // Fail-safe to protect against compacting the same sstable concurrently.
1460 1 : if inputRangeAlreadyCompacting(env, pc) {
1461 1 : return nil
1462 1 : }
1463 :
1464 1 : if !pc.setupInputs(p.opts, env.diskAvailBytes, pc.startLevel) {
1465 1 : return nil
1466 1 : }
1467 :
1468 1 : return pc
1469 : }
1470 :
1471 : // pickElisionOnlyCompaction looks for compactions of sstables in the
1472 : // bottommost level containing obsolete records that may now be dropped.
1473 : func (p *compactionPickerByScore) pickElisionOnlyCompaction(
1474 : env compactionEnv,
1475 1 : ) (pc *pickedCompaction) {
1476 1 : if p.opts.private.disableElisionOnlyCompactions {
1477 1 : return nil
1478 1 : }
1479 1 : candidate := elisionOnlyAnnotator.LevelAnnotation(p.vers.Levels[numLevels-1])
1480 1 : if candidate == nil {
1481 1 : return nil
1482 1 : }
1483 1 : if candidate.LargestSeqNum >= env.earliestSnapshotSeqNum {
1484 1 : return nil
1485 1 : }
1486 1 : return p.pickedCompactionFromCandidateFile(candidate, env, numLevels-1, numLevels-1, compactionKindElisionOnly)
1487 : }
1488 :
1489 : // pickRewriteCompaction attempts to construct a compaction that
1490 : // rewrites a file marked for compaction. pickRewriteCompaction will
1491 : // pull in adjacent files in the file's atomic compaction unit if
1492 : // necessary. A rewrite compaction outputs files to the same level as
1493 : // the input level.
1494 0 : func (p *compactionPickerByScore) pickRewriteCompaction(env compactionEnv) (pc *pickedCompaction) {
1495 0 : for l := numLevels - 1; l >= 0; l-- {
1496 0 : candidate := markedForCompactionAnnotator.LevelAnnotation(p.vers.Levels[l])
1497 0 : if candidate == nil {
1498 0 : // Try the next level.
1499 0 : continue
1500 : }
1501 0 : pc := p.pickedCompactionFromCandidateFile(candidate, env, l, l, compactionKindRewrite)
1502 0 : if pc != nil {
1503 0 : return pc
1504 0 : }
1505 : }
1506 0 : return nil
1507 : }
1508 :
1509 1 : func (p *compactionPickerByScore) initTombstoneDensityAnnotator(opts *Options) {
1510 1 : p.tombstoneDensityAnnotator = &manifest.Annotator[fileMetadata]{
1511 1 : Aggregator: manifest.PickFileAggregator{
1512 1 : Filter: func(f *fileMetadata) (eligible bool, cacheOK bool) {
1513 1 : if f.IsCompacting() {
1514 1 : return false, true
1515 1 : }
1516 1 : if !f.StatsValid() {
1517 1 : return false, false
1518 1 : }
1519 1 : return f.Stats.TombstoneDenseBlocksRatio > opts.Experimental.TombstoneDenseCompactionThreshold, true
1520 : },
1521 1 : Compare: func(a, b *fileMetadata) bool {
1522 1 : return a.Stats.TombstoneDenseBlocksRatio > b.Stats.TombstoneDenseBlocksRatio
1523 1 : },
1524 : },
1525 : }
1526 : }
1527 :
1528 : // pickTombstoneDensityCompaction looks for a compaction that eliminates
1529 : // regions of extremely high point tombstone density. For each level, it picks
1530 : // a file where the ratio of tombstone-dense blocks is at least
1531 : // options.Experimental.MinTombstoneDenseRatio, prioritizing compaction of
1532 : // files with higher ratios of tombstone-dense blocks.
1533 : func (p *compactionPickerByScore) pickTombstoneDensityCompaction(
1534 : env compactionEnv,
1535 1 : ) (pc *pickedCompaction) {
1536 1 : if p.opts.Experimental.TombstoneDenseCompactionThreshold <= 0 {
1537 0 : // Tombstone density compactions are disabled.
1538 0 : return nil
1539 0 : }
1540 :
1541 1 : var candidate *fileMetadata
1542 1 : var level int
1543 1 : // NB: we don't consider the lowest level because elision-only compactions
1544 1 : // handle that case.
1545 1 : for l := 0; l < numLevels-1; l++ {
1546 1 : f := p.tombstoneDensityAnnotator.LevelAnnotation(p.vers.Levels[l])
1547 1 : newCandidate := p.tombstoneDensityAnnotator.Aggregator.Merge(f, candidate)
1548 1 : if newCandidate != candidate {
1549 1 : candidate = newCandidate
1550 1 : level = l
1551 1 : }
1552 : }
1553 :
1554 1 : return p.pickedCompactionFromCandidateFile(candidate, env, level, defaultOutputLevel(level, p.baseLevel), compactionKindTombstoneDensity)
1555 : }
1556 :
1557 : // pickAutoLPositive picks an automatic compaction for the candidate
1558 : // file in a positive-numbered level. This function must not be used for
1559 : // L0.
1560 : func pickAutoLPositive(
1561 : env compactionEnv, opts *Options, vers *version, cInfo candidateLevelInfo, baseLevel int,
1562 1 : ) (pc *pickedCompaction) {
1563 1 : if cInfo.level == 0 {
1564 0 : panic("pebble: pickAutoLPositive called for L0")
1565 : }
1566 :
1567 1 : pc = newPickedCompaction(opts, vers, cInfo.level, defaultOutputLevel(cInfo.level, baseLevel), baseLevel)
1568 1 : if pc.outputLevel.level != cInfo.outputLevel {
1569 0 : panic("pebble: compaction picked unexpected output level")
1570 : }
1571 1 : pc.startLevel.files = cInfo.file.Slice()
1572 1 : // Files in level 0 may overlap each other, so pick up all overlapping ones.
1573 1 : if pc.startLevel.level == 0 {
1574 0 : cmp := opts.Comparer.Compare
1575 0 : smallest, largest := manifest.KeyRange(cmp, pc.startLevel.files.Iter())
1576 0 : pc.startLevel.files = vers.Overlaps(0, base.UserKeyBoundsFromInternal(smallest, largest))
1577 0 : if pc.startLevel.files.Empty() {
1578 0 : panic("pebble: empty compaction")
1579 : }
1580 : }
1581 :
1582 1 : if !pc.setupInputs(opts, env.diskAvailBytes, pc.startLevel) {
1583 0 : return nil
1584 0 : }
1585 1 : return pc.maybeAddLevel(opts, env.diskAvailBytes)
1586 : }
1587 :
1588 : // maybeAddLevel maybe adds a level to the picked compaction.
1589 1 : func (pc *pickedCompaction) maybeAddLevel(opts *Options, diskAvailBytes uint64) *pickedCompaction {
1590 1 : pc.pickerMetrics.singleLevelOverlappingRatio = pc.overlappingRatio()
1591 1 : if pc.outputLevel.level == numLevels-1 {
1592 1 : // Don't add a level if the current output level is in L6
1593 1 : return pc
1594 1 : }
1595 1 : if !opts.Experimental.MultiLevelCompactionHeuristic.allowL0() && pc.startLevel.level == 0 {
1596 1 : return pc
1597 1 : }
1598 1 : if pc.compactionSize() > expandedCompactionByteSizeLimit(
1599 1 : opts, adjustedOutputLevel(pc.outputLevel.level, pc.baseLevel), diskAvailBytes) {
1600 1 : // Don't add a level if the current compaction exceeds the compaction size limit
1601 1 : return pc
1602 1 : }
1603 1 : return opts.Experimental.MultiLevelCompactionHeuristic.pick(pc, opts, diskAvailBytes)
1604 : }
1605 :
1606 : // MultiLevelHeuristic evaluates whether to add files from the next level into the compaction.
1607 : type MultiLevelHeuristic interface {
1608 : // Evaluate returns the preferred compaction.
1609 : pick(pc *pickedCompaction, opts *Options, diskAvailBytes uint64) *pickedCompaction
1610 :
1611 : // Returns if the heuristic allows L0 to be involved in ML compaction
1612 : allowL0() bool
1613 :
1614 : // String implements fmt.Stringer.
1615 : String() string
1616 : }
1617 :
1618 : // NoMultiLevel will never add an additional level to the compaction.
1619 : type NoMultiLevel struct{}
1620 :
1621 : var _ MultiLevelHeuristic = (*NoMultiLevel)(nil)
1622 :
1623 : func (nml NoMultiLevel) pick(
1624 : pc *pickedCompaction, opts *Options, diskAvailBytes uint64,
1625 1 : ) *pickedCompaction {
1626 1 : return pc
1627 1 : }
1628 :
1629 1 : func (nml NoMultiLevel) allowL0() bool { return false }
1630 1 : func (nml NoMultiLevel) String() string { return "none" }
1631 :
1632 1 : func (pc *pickedCompaction) predictedWriteAmp() float64 {
1633 1 : var bytesToCompact uint64
1634 1 : var higherLevelBytes uint64
1635 1 : for i := range pc.inputs {
1636 1 : levelSize := pc.inputs[i].files.SizeSum()
1637 1 : bytesToCompact += levelSize
1638 1 : if i != len(pc.inputs)-1 {
1639 1 : higherLevelBytes += levelSize
1640 1 : }
1641 : }
1642 1 : return float64(bytesToCompact) / float64(higherLevelBytes)
1643 : }
1644 :
1645 1 : func (pc *pickedCompaction) overlappingRatio() float64 {
1646 1 : var higherLevelBytes uint64
1647 1 : var lowestLevelBytes uint64
1648 1 : for i := range pc.inputs {
1649 1 : levelSize := pc.inputs[i].files.SizeSum()
1650 1 : if i == len(pc.inputs)-1 {
1651 1 : lowestLevelBytes += levelSize
1652 1 : continue
1653 : }
1654 1 : higherLevelBytes += levelSize
1655 : }
1656 1 : return float64(lowestLevelBytes) / float64(higherLevelBytes)
1657 : }
1658 :
1659 : // WriteAmpHeuristic defines a multi level compaction heuristic which will add
1660 : // an additional level to the picked compaction if it reduces predicted write
1661 : // amp of the compaction + the addPropensity constant.
1662 : type WriteAmpHeuristic struct {
1663 : // addPropensity is a constant that affects the propensity to conduct multilevel
1664 : // compactions. If positive, a multilevel compaction may get picked even if
1665 : // the single level compaction has lower write amp, and vice versa.
1666 : AddPropensity float64
1667 :
1668 : // AllowL0 if true, allow l0 to be involved in a ML compaction.
1669 : AllowL0 bool
1670 : }
1671 :
1672 : var _ MultiLevelHeuristic = (*WriteAmpHeuristic)(nil)
1673 :
1674 : // TODO(msbutler): microbenchmark the extent to which multilevel compaction
1675 : // picking slows down the compaction picking process. This should be as fast as
1676 : // possible since Compaction-picking holds d.mu, which prevents WAL rotations,
1677 : // in-progress flushes and compactions from completing, etc. Consider ways to
1678 : // deduplicate work, given that setupInputs has already been called.
1679 : func (wa WriteAmpHeuristic) pick(
1680 : pcOrig *pickedCompaction, opts *Options, diskAvailBytes uint64,
1681 1 : ) *pickedCompaction {
1682 1 : pcMulti := pcOrig.clone()
1683 1 : if !pcMulti.setupMultiLevelCandidate(opts, diskAvailBytes) {
1684 1 : return pcOrig
1685 1 : }
1686 1 : picked := pcOrig
1687 1 : if pcMulti.predictedWriteAmp() <= pcOrig.predictedWriteAmp()+wa.AddPropensity {
1688 1 : picked = pcMulti
1689 1 : }
1690 : // Regardless of what compaction was picked, log the multilevelOverlapping ratio.
1691 1 : picked.pickerMetrics.multiLevelOverlappingRatio = pcMulti.overlappingRatio()
1692 1 : return picked
1693 : }
1694 :
1695 1 : func (wa WriteAmpHeuristic) allowL0() bool {
1696 1 : return wa.AllowL0
1697 1 : }
1698 :
1699 : // String implements fmt.Stringer.
1700 1 : func (wa WriteAmpHeuristic) String() string {
1701 1 : return fmt.Sprintf("wamp(%.2f, %t)", wa.AddPropensity, wa.AllowL0)
1702 1 : }
1703 :
1704 : // Helper method to pick compactions originating from L0. Uses information about
1705 : // sublevels to generate a compaction.
1706 1 : func pickL0(env compactionEnv, opts *Options, vers *version, baseLevel int) (pc *pickedCompaction) {
1707 1 : // It is important to pass information about Lbase files to L0Sublevels
1708 1 : // so it can pick a compaction that does not conflict with an Lbase => Lbase+1
1709 1 : // compaction. Without this, we observed reduced concurrency of L0=>Lbase
1710 1 : // compactions, and increasing read amplification in L0.
1711 1 : //
1712 1 : // TODO(bilal) Remove the minCompactionDepth parameter once fixing it at 1
1713 1 : // has been shown to not cause a performance regression.
1714 1 : lcf, err := vers.L0Sublevels.PickBaseCompaction(1, vers.Levels[baseLevel].Slice())
1715 1 : if err != nil {
1716 0 : opts.Logger.Errorf("error when picking base compaction: %s", err)
1717 0 : return
1718 0 : }
1719 1 : if lcf != nil {
1720 1 : pc = newPickedCompactionFromL0(lcf, opts, vers, baseLevel, true)
1721 1 : pc.setupInputs(opts, env.diskAvailBytes, pc.startLevel)
1722 1 : if pc.startLevel.files.Empty() {
1723 0 : opts.Logger.Fatalf("empty compaction chosen")
1724 0 : }
1725 1 : return pc.maybeAddLevel(opts, env.diskAvailBytes)
1726 : }
1727 :
1728 : // Couldn't choose a base compaction. Try choosing an intra-L0
1729 : // compaction. Note that we pass in L0CompactionThreshold here as opposed to
1730 : // 1, since choosing a single sublevel intra-L0 compaction is
1731 : // counterproductive.
1732 1 : lcf, err = vers.L0Sublevels.PickIntraL0Compaction(env.earliestUnflushedSeqNum, minIntraL0Count)
1733 1 : if err != nil {
1734 0 : opts.Logger.Errorf("error when picking intra-L0 compaction: %s", err)
1735 0 : return
1736 0 : }
1737 1 : if lcf != nil {
1738 1 : pc = newPickedCompactionFromL0(lcf, opts, vers, 0, false)
1739 1 : if !pc.setupInputs(opts, env.diskAvailBytes, pc.startLevel) {
1740 0 : return nil
1741 0 : }
1742 1 : if pc.startLevel.files.Empty() {
1743 0 : opts.Logger.Fatalf("empty compaction chosen")
1744 0 : }
1745 1 : {
1746 1 : iter := pc.startLevel.files.Iter()
1747 1 : if iter.First() == nil || iter.Next() == nil {
1748 0 : // A single-file intra-L0 compaction is unproductive.
1749 0 : return nil
1750 0 : }
1751 : }
1752 :
1753 1 : pc.smallest, pc.largest = manifest.KeyRange(pc.cmp, pc.startLevel.files.Iter())
1754 : }
1755 1 : return pc
1756 : }
1757 :
1758 : func pickManualCompaction(
1759 : vers *version, opts *Options, env compactionEnv, baseLevel int, manual *manualCompaction,
1760 1 : ) (pc *pickedCompaction, retryLater bool) {
1761 1 : outputLevel := manual.level + 1
1762 1 : if manual.level == 0 {
1763 1 : outputLevel = baseLevel
1764 1 : } else if manual.level < baseLevel {
1765 1 : // The start level for a compaction must be >= Lbase. A manual
1766 1 : // compaction could have been created adhering to that condition, and
1767 1 : // then an automatic compaction came in and compacted all of the
1768 1 : // sstables in Lbase to Lbase+1 which caused Lbase to change. Simply
1769 1 : // ignore this manual compaction as there is nothing to do (manual.level
1770 1 : // points to an empty level).
1771 1 : return nil, false
1772 1 : }
1773 : // This conflictsWithInProgress call is necessary for the manual compaction to
1774 : // be retried when it conflicts with an ongoing automatic compaction. Without
1775 : // it, the compaction is dropped due to pc.setupInputs returning false since
1776 : // the input/output range is already being compacted, and the manual
1777 : // compaction ends with a non-compacted LSM.
1778 1 : if conflictsWithInProgress(manual, outputLevel, env.inProgressCompactions, opts.Comparer.Compare) {
1779 1 : return nil, true
1780 1 : }
1781 1 : pc = newPickedCompaction(opts, vers, manual.level, defaultOutputLevel(manual.level, baseLevel), baseLevel)
1782 1 : manual.outputLevel = pc.outputLevel.level
1783 1 : pc.startLevel.files = vers.Overlaps(manual.level, base.UserKeyBoundsInclusive(manual.start, manual.end))
1784 1 : if pc.startLevel.files.Empty() {
1785 1 : // Nothing to do
1786 1 : return nil, false
1787 1 : }
1788 1 : if !pc.setupInputs(opts, env.diskAvailBytes, pc.startLevel) {
1789 1 : // setupInputs returned false indicating there's a conflicting
1790 1 : // concurrent compaction.
1791 1 : return nil, true
1792 1 : }
1793 1 : if pc = pc.maybeAddLevel(opts, env.diskAvailBytes); pc == nil {
1794 0 : return nil, false
1795 0 : }
1796 1 : if pc.outputLevel.level != outputLevel {
1797 1 : if len(pc.extraLevels) > 0 {
1798 1 : // multilevel compactions relax this invariant
1799 1 : } else {
1800 0 : panic("pebble: compaction picked unexpected output level")
1801 : }
1802 : }
1803 : // Fail-safe to protect against compacting the same sstable concurrently.
1804 1 : if inputRangeAlreadyCompacting(env, pc) {
1805 0 : return nil, true
1806 0 : }
1807 1 : return pc, false
1808 : }
1809 :
1810 : // pickDownloadCompaction picks a download compaction for the downloadSpan,
1811 : // which could be specified as being performed either by a copy compaction of
1812 : // the backing file or a rewrite compaction.
1813 : func pickDownloadCompaction(
1814 : vers *version,
1815 : opts *Options,
1816 : env compactionEnv,
1817 : baseLevel int,
1818 : kind compactionKind,
1819 : level int,
1820 : file *fileMetadata,
1821 1 : ) (pc *pickedCompaction) {
1822 1 : // Check if the file is compacting already.
1823 1 : if file.CompactionState == manifest.CompactionStateCompacting {
1824 0 : return nil
1825 0 : }
1826 1 : if kind != compactionKindCopy && kind != compactionKindRewrite {
1827 0 : panic("invalid download/rewrite compaction kind")
1828 : }
1829 1 : pc = newPickedCompaction(opts, vers, level, level, baseLevel)
1830 1 : pc.kind = kind
1831 1 : pc.startLevel.files = manifest.NewLevelSliceKeySorted(opts.Comparer.Compare, []*fileMetadata{file})
1832 1 : if !pc.setupInputs(opts, env.diskAvailBytes, pc.startLevel) {
1833 0 : // setupInputs returned false indicating there's a conflicting
1834 0 : // concurrent compaction.
1835 0 : return nil
1836 0 : }
1837 1 : if pc.outputLevel.level != level {
1838 0 : panic("pebble: download compaction picked unexpected output level")
1839 : }
1840 : // Fail-safe to protect against compacting the same sstable concurrently.
1841 1 : if inputRangeAlreadyCompacting(env, pc) {
1842 0 : return nil
1843 0 : }
1844 1 : return pc
1845 : }
1846 :
1847 : func (p *compactionPickerByScore) pickReadTriggeredCompaction(
1848 : env compactionEnv,
1849 1 : ) (pc *pickedCompaction) {
1850 1 : // If a flush is in-progress or expected to happen soon, it means more writes are taking place. We would
1851 1 : // soon be scheduling more write focussed compactions. In this case, skip read compactions as they are
1852 1 : // lower priority.
1853 1 : if env.readCompactionEnv.flushing || env.readCompactionEnv.readCompactions == nil {
1854 1 : return nil
1855 1 : }
1856 1 : for env.readCompactionEnv.readCompactions.size > 0 {
1857 0 : rc := env.readCompactionEnv.readCompactions.remove()
1858 0 : if pc = pickReadTriggeredCompactionHelper(p, rc, env); pc != nil {
1859 0 : break
1860 : }
1861 : }
1862 1 : return pc
1863 : }
1864 :
1865 : func pickReadTriggeredCompactionHelper(
1866 : p *compactionPickerByScore, rc *readCompaction, env compactionEnv,
1867 0 : ) (pc *pickedCompaction) {
1868 0 : overlapSlice := p.vers.Overlaps(rc.level, base.UserKeyBoundsInclusive(rc.start, rc.end))
1869 0 : if overlapSlice.Empty() {
1870 0 : // If there is no overlap, then the file with the key range
1871 0 : // must have been compacted away. So, we don't proceed to
1872 0 : // compact the same key range again.
1873 0 : return nil
1874 0 : }
1875 :
1876 0 : iter := overlapSlice.Iter()
1877 0 : var fileMatches bool
1878 0 : for f := iter.First(); f != nil; f = iter.Next() {
1879 0 : if f.FileNum == rc.fileNum {
1880 0 : fileMatches = true
1881 0 : break
1882 : }
1883 : }
1884 0 : if !fileMatches {
1885 0 : return nil
1886 0 : }
1887 :
1888 0 : pc = newPickedCompaction(p.opts, p.vers, rc.level, defaultOutputLevel(rc.level, p.baseLevel), p.baseLevel)
1889 0 :
1890 0 : pc.startLevel.files = overlapSlice
1891 0 : if !pc.setupInputs(p.opts, env.diskAvailBytes, pc.startLevel) {
1892 0 : return nil
1893 0 : }
1894 0 : if inputRangeAlreadyCompacting(env, pc) {
1895 0 : return nil
1896 0 : }
1897 0 : pc.kind = compactionKindRead
1898 0 :
1899 0 : // Prevent read compactions which are too wide.
1900 0 : outputOverlaps := pc.version.Overlaps(pc.outputLevel.level, pc.userKeyBounds())
1901 0 : if outputOverlaps.SizeSum() > pc.maxReadCompactionBytes {
1902 0 : return nil
1903 0 : }
1904 :
1905 : // Prevent compactions which start with a small seed file X, but overlap
1906 : // with over allowedCompactionWidth * X file sizes in the output layer.
1907 0 : const allowedCompactionWidth = 35
1908 0 : if outputOverlaps.SizeSum() > overlapSlice.SizeSum()*allowedCompactionWidth {
1909 0 : return nil
1910 0 : }
1911 :
1912 0 : return pc
1913 : }
1914 :
1915 0 : func (p *compactionPickerByScore) forceBaseLevel1() {
1916 0 : p.baseLevel = 1
1917 0 : }
1918 :
1919 1 : func inputRangeAlreadyCompacting(env compactionEnv, pc *pickedCompaction) bool {
1920 1 : for _, cl := range pc.inputs {
1921 1 : iter := cl.files.Iter()
1922 1 : for f := iter.First(); f != nil; f = iter.Next() {
1923 1 : if f.IsCompacting() {
1924 1 : return true
1925 1 : }
1926 : }
1927 : }
1928 :
1929 : // Look for active compactions outputting to the same region of the key
1930 : // space in the same output level. Two potential compactions may conflict
1931 : // without sharing input files if there are no files in the output level
1932 : // that overlap with the intersection of the compactions' key spaces.
1933 : //
1934 : // Consider an active L0->Lbase compaction compacting two L0 files one
1935 : // [a-f] and the other [t-z] into Lbase.
1936 : //
1937 : // L0
1938 : // ↦ 000100 ↤ ↦ 000101 ↤
1939 : // L1
1940 : // ↦ 000004 ↤
1941 : // 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
1942 : //
1943 : // If a new file 000102 [j-p] is flushed while the existing compaction is
1944 : // still ongoing, new file would not be in any compacting sublevel
1945 : // intervals and would not overlap with any Lbase files that are also
1946 : // compacting. However, this compaction cannot be picked because the
1947 : // compaction's output key space [j-p] would overlap the existing
1948 : // compaction's output key space [a-z].
1949 : //
1950 : // L0
1951 : // ↦ 000100* ↤ ↦ 000102 ↤ ↦ 000101* ↤
1952 : // L1
1953 : // ↦ 000004* ↤
1954 : // 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
1955 : //
1956 : // * - currently compacting
1957 1 : if pc.outputLevel != nil && pc.outputLevel.level != 0 {
1958 1 : for _, c := range env.inProgressCompactions {
1959 1 : if pc.outputLevel.level != c.outputLevel {
1960 1 : continue
1961 : }
1962 1 : if base.InternalCompare(pc.cmp, c.largest, pc.smallest) < 0 ||
1963 1 : base.InternalCompare(pc.cmp, c.smallest, pc.largest) > 0 {
1964 1 : continue
1965 : }
1966 :
1967 : // The picked compaction and the in-progress compaction c are
1968 : // outputting to the same region of the key space of the same
1969 : // level.
1970 1 : return true
1971 : }
1972 : }
1973 1 : return false
1974 : }
1975 :
1976 : // conflictsWithInProgress checks if there are any in-progress compactions with overlapping keyspace.
1977 : func conflictsWithInProgress(
1978 : manual *manualCompaction, outputLevel int, inProgressCompactions []compactionInfo, cmp Compare,
1979 1 : ) bool {
1980 1 : for _, c := range inProgressCompactions {
1981 1 : if (c.outputLevel == manual.level || c.outputLevel == outputLevel) &&
1982 1 : isUserKeysOverlapping(manual.start, manual.end, c.smallest.UserKey, c.largest.UserKey, cmp) {
1983 1 : return true
1984 1 : }
1985 1 : for _, in := range c.inputs {
1986 1 : if in.files.Empty() {
1987 1 : continue
1988 : }
1989 1 : iter := in.files.Iter()
1990 1 : smallest := iter.First().Smallest.UserKey
1991 1 : largest := iter.Last().Largest.UserKey
1992 1 : if (in.level == manual.level || in.level == outputLevel) &&
1993 1 : isUserKeysOverlapping(manual.start, manual.end, smallest, largest, cmp) {
1994 1 : return true
1995 1 : }
1996 : }
1997 : }
1998 1 : return false
1999 : }
2000 :
2001 1 : func isUserKeysOverlapping(x1, x2, y1, y2 []byte, cmp Compare) bool {
2002 1 : return cmp(x1, y2) <= 0 && cmp(y1, x2) <= 0
2003 1 : }
|
