Line data Source code
1 : // Copyright 2013 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 : "context"
10 : "fmt"
11 : "io"
12 : "math"
13 : "runtime/pprof"
14 : "slices"
15 : "sort"
16 : "sync/atomic"
17 : "time"
18 :
19 : "github.com/cockroachdb/errors"
20 : "github.com/cockroachdb/pebble/internal/base"
21 : "github.com/cockroachdb/pebble/internal/compact"
22 : "github.com/cockroachdb/pebble/internal/invalidating"
23 : "github.com/cockroachdb/pebble/internal/invariants"
24 : "github.com/cockroachdb/pebble/internal/keyspan"
25 : "github.com/cockroachdb/pebble/internal/keyspan/keyspanimpl"
26 : "github.com/cockroachdb/pebble/internal/manifest"
27 : "github.com/cockroachdb/pebble/internal/private"
28 : "github.com/cockroachdb/pebble/internal/rangedel"
29 : "github.com/cockroachdb/pebble/internal/rangekey"
30 : "github.com/cockroachdb/pebble/objstorage"
31 : "github.com/cockroachdb/pebble/objstorage/objstorageprovider/objiotracing"
32 : "github.com/cockroachdb/pebble/objstorage/remote"
33 : "github.com/cockroachdb/pebble/sstable"
34 : "github.com/cockroachdb/pebble/vfs"
35 : )
36 :
37 : var errEmptyTable = errors.New("pebble: empty table")
38 :
39 : // ErrCancelledCompaction is returned if a compaction is cancelled by a
40 : // concurrent excise or ingest-split operation.
41 : var ErrCancelledCompaction = errors.New("pebble: compaction cancelled by a concurrent operation, will retry compaction")
42 :
43 : var compactLabels = pprof.Labels("pebble", "compact")
44 : var flushLabels = pprof.Labels("pebble", "flush")
45 : var gcLabels = pprof.Labels("pebble", "gc")
46 :
47 : // getInternalWriterProperties accesses a private variable (in the
48 : // internal/private package) initialized by the sstable Writer. This indirection
49 : // is necessary to ensure non-Pebble users constructing sstables for ingestion
50 : // are unable to set internal-only properties.
51 : var getInternalWriterProperties = private.SSTableInternalProperties.(func(*sstable.Writer) *sstable.Properties)
52 :
53 : // expandedCompactionByteSizeLimit is the maximum number of bytes in all
54 : // compacted files. We avoid expanding the lower level file set of a compaction
55 : // if it would make the total compaction cover more than this many bytes.
56 1 : func expandedCompactionByteSizeLimit(opts *Options, level int, availBytes uint64) uint64 {
57 1 : v := uint64(25 * opts.Level(level).TargetFileSize)
58 1 :
59 1 : // Never expand a compaction beyond half the available capacity, divided
60 1 : // by the maximum number of concurrent compactions. Each of the concurrent
61 1 : // compactions may expand up to this limit, so this attempts to limit
62 1 : // compactions to half of available disk space. Note that this will not
63 1 : // prevent compaction picking from pursuing compactions that are larger
64 1 : // than this threshold before expansion.
65 1 : diskMax := (availBytes / 2) / uint64(opts.MaxConcurrentCompactions())
66 1 : if v > diskMax {
67 1 : v = diskMax
68 1 : }
69 1 : return v
70 : }
71 :
72 : // maxGrandparentOverlapBytes is the maximum bytes of overlap with level+1
73 : // before we stop building a single file in a level-1 to level compaction.
74 1 : func maxGrandparentOverlapBytes(opts *Options, level int) uint64 {
75 1 : return uint64(10 * opts.Level(level).TargetFileSize)
76 1 : }
77 :
78 : // maxReadCompactionBytes is used to prevent read compactions which
79 : // are too wide.
80 1 : func maxReadCompactionBytes(opts *Options, level int) uint64 {
81 1 : return uint64(10 * opts.Level(level).TargetFileSize)
82 1 : }
83 :
84 : // noCloseIter wraps around a FragmentIterator, intercepting and eliding
85 : // calls to Close. It is used during compaction to ensure that rangeDelIters
86 : // are not closed prematurely.
87 : type noCloseIter struct {
88 : keyspan.FragmentIterator
89 : }
90 :
91 1 : func (i noCloseIter) Close() error {
92 1 : return nil
93 1 : }
94 :
95 : type compactionLevel struct {
96 : level int
97 : files manifest.LevelSlice
98 : // l0SublevelInfo contains information about L0 sublevels being compacted.
99 : // It's only set for the start level of a compaction starting out of L0 and
100 : // is nil for all other compactions.
101 : l0SublevelInfo []sublevelInfo
102 : }
103 :
104 1 : func (cl compactionLevel) Clone() compactionLevel {
105 1 : newCL := compactionLevel{
106 1 : level: cl.level,
107 1 : files: cl.files.Reslice(func(start, end *manifest.LevelIterator) {}),
108 : }
109 1 : return newCL
110 : }
111 1 : func (cl compactionLevel) String() string {
112 1 : return fmt.Sprintf(`Level %d, Files %s`, cl.level, cl.files)
113 1 : }
114 :
115 : // compactionWritable is a objstorage.Writable wrapper that, on every write,
116 : // updates a metric in `versions` on bytes written by in-progress compactions so
117 : // far. It also increments a per-compaction `written` int.
118 : type compactionWritable struct {
119 : objstorage.Writable
120 :
121 : versions *versionSet
122 : written *int64
123 : }
124 :
125 : // Write is part of the objstorage.Writable interface.
126 1 : func (c *compactionWritable) Write(p []byte) error {
127 1 : if err := c.Writable.Write(p); err != nil {
128 0 : return err
129 0 : }
130 :
131 1 : *c.written += int64(len(p))
132 1 : c.versions.incrementCompactionBytes(int64(len(p)))
133 1 : return nil
134 : }
135 :
136 : type compactionKind int
137 :
138 : const (
139 : compactionKindDefault compactionKind = iota
140 : compactionKindFlush
141 : // compactionKindMove denotes a move compaction where the input file is
142 : // retained and linked in a new level without being obsoleted.
143 : compactionKindMove
144 : // compactionKindCopy denotes a copy compaction where the input file is
145 : // copied byte-by-byte into a new file with a new FileNum in the output level.
146 : compactionKindCopy
147 : // compactionKindDeleteOnly denotes a compaction that only deletes input
148 : // files. It can occur when wide range tombstones completely contain sstables.
149 : compactionKindDeleteOnly
150 : compactionKindElisionOnly
151 : compactionKindRead
152 : compactionKindRewrite
153 : compactionKindIngestedFlushable
154 : )
155 :
156 1 : func (k compactionKind) String() string {
157 1 : switch k {
158 1 : case compactionKindDefault:
159 1 : return "default"
160 0 : case compactionKindFlush:
161 0 : return "flush"
162 1 : case compactionKindMove:
163 1 : return "move"
164 1 : case compactionKindDeleteOnly:
165 1 : return "delete-only"
166 1 : case compactionKindElisionOnly:
167 1 : return "elision-only"
168 1 : case compactionKindRead:
169 1 : return "read"
170 1 : case compactionKindRewrite:
171 1 : return "rewrite"
172 0 : case compactionKindIngestedFlushable:
173 0 : return "ingested-flushable"
174 1 : case compactionKindCopy:
175 1 : return "copy"
176 : }
177 0 : return "?"
178 : }
179 :
180 : // compaction is a table compaction from one level to the next, starting from a
181 : // given version.
182 : type compaction struct {
183 : // cancel is a bool that can be used by other goroutines to signal a compaction
184 : // to cancel, such as if a conflicting excise operation raced it to manifest
185 : // application. Only holders of the manifest lock will write to this atomic.
186 : cancel atomic.Bool
187 :
188 : kind compactionKind
189 : // isDownload is true if this compaction was started as part of a Download
190 : // operation. In this case kind is compactionKindCopy or
191 : // compactionKindRewrite.
192 : isDownload bool
193 :
194 : cmp Compare
195 : equal Equal
196 : comparer *base.Comparer
197 : formatKey base.FormatKey
198 : logger Logger
199 : version *version
200 : stats base.InternalIteratorStats
201 : beganAt time.Time
202 : // versionEditApplied is set to true when a compaction has completed and the
203 : // resulting version has been installed (if successful), but the compaction
204 : // goroutine is still cleaning up (eg, deleting obsolete files).
205 : versionEditApplied bool
206 : bufferPool sstable.BufferPool
207 :
208 : // startLevel is the level that is being compacted. Inputs from startLevel
209 : // and outputLevel will be merged to produce a set of outputLevel files.
210 : startLevel *compactionLevel
211 :
212 : // outputLevel is the level that files are being produced in. outputLevel is
213 : // equal to startLevel+1 except when:
214 : // - if startLevel is 0, the output level equals compactionPicker.baseLevel().
215 : // - in multilevel compaction, the output level is the lowest level involved in
216 : // the compaction
217 : // A compaction's outputLevel is nil for delete-only compactions.
218 : outputLevel *compactionLevel
219 :
220 : // extraLevels point to additional levels in between the input and output
221 : // levels that get compacted in multilevel compactions
222 : extraLevels []*compactionLevel
223 :
224 : inputs []compactionLevel
225 :
226 : // maxOutputFileSize is the maximum size of an individual table created
227 : // during compaction.
228 : maxOutputFileSize uint64
229 : // maxOverlapBytes is the maximum number of bytes of overlap allowed for a
230 : // single output table with the tables in the grandparent level.
231 : maxOverlapBytes uint64
232 :
233 : // flushing contains the flushables (aka memtables) that are being flushed.
234 : flushing flushableList
235 : // bytesWritten contains the number of bytes that have been written to outputs.
236 : bytesWritten int64
237 :
238 : // The boundaries of the input data.
239 : smallest InternalKey
240 : largest InternalKey
241 :
242 : // rangeDelInterlaving is an interleaving iterator for range deletions, that
243 : // interleaves range tombstones among the point keys.
244 : rangeDelInterleaving keyspan.InterleavingIter
245 : // rangeKeyInterleaving is the interleaving iter for range keys.
246 : rangeKeyInterleaving keyspan.InterleavingIter
247 :
248 : // A list of objects to close when the compaction finishes. Used by input
249 : // iteration to keep rangeDelIters open for the lifetime of the compaction,
250 : // and only close them when the compaction finishes.
251 : closers []io.Closer
252 :
253 : // grandparents are the tables in level+2 that overlap with the files being
254 : // compacted. Used to determine output table boundaries. Do not assume that the actual files
255 : // in the grandparent when this compaction finishes will be the same.
256 : grandparents manifest.LevelSlice
257 :
258 : // Boundaries at which flushes to L0 should be split. Determined by
259 : // L0Sublevels. If nil, flushes aren't split.
260 : l0Limits [][]byte
261 :
262 : delElision compact.TombstoneElision
263 : rangeKeyElision compact.TombstoneElision
264 :
265 : // allowedZeroSeqNum is true if seqnums can be zeroed if there are no
266 : // snapshots requiring them to be kept. This determination is made by
267 : // looking for an sstable which overlaps the bounds of the compaction at a
268 : // lower level in the LSM during runCompaction.
269 : allowedZeroSeqNum bool
270 :
271 : metrics map[int]*LevelMetrics
272 :
273 : pickerMetrics compactionPickerMetrics
274 : }
275 :
276 1 : func (c *compaction) makeInfo(jobID JobID) CompactionInfo {
277 1 : info := CompactionInfo{
278 1 : JobID: int(jobID),
279 1 : Reason: c.kind.String(),
280 1 : Input: make([]LevelInfo, 0, len(c.inputs)),
281 1 : Annotations: []string{},
282 1 : }
283 1 : if c.isDownload {
284 1 : info.Reason = "download," + info.Reason
285 1 : }
286 1 : for _, cl := range c.inputs {
287 1 : inputInfo := LevelInfo{Level: cl.level, Tables: nil}
288 1 : iter := cl.files.Iter()
289 1 : for m := iter.First(); m != nil; m = iter.Next() {
290 1 : inputInfo.Tables = append(inputInfo.Tables, m.TableInfo())
291 1 : }
292 1 : info.Input = append(info.Input, inputInfo)
293 : }
294 1 : if c.outputLevel != nil {
295 1 : info.Output.Level = c.outputLevel.level
296 1 :
297 1 : // If there are no inputs from the output level (eg, a move
298 1 : // compaction), add an empty LevelInfo to info.Input.
299 1 : if len(c.inputs) > 0 && c.inputs[len(c.inputs)-1].level != c.outputLevel.level {
300 0 : info.Input = append(info.Input, LevelInfo{Level: c.outputLevel.level})
301 0 : }
302 1 : } else {
303 1 : // For a delete-only compaction, set the output level to L6. The
304 1 : // output level is not meaningful here, but complicating the
305 1 : // info.Output interface with a pointer doesn't seem worth the
306 1 : // semantic distinction.
307 1 : info.Output.Level = numLevels - 1
308 1 : }
309 :
310 1 : for i, score := range c.pickerMetrics.scores {
311 1 : info.Input[i].Score = score
312 1 : }
313 1 : info.SingleLevelOverlappingRatio = c.pickerMetrics.singleLevelOverlappingRatio
314 1 : info.MultiLevelOverlappingRatio = c.pickerMetrics.multiLevelOverlappingRatio
315 1 : if len(info.Input) > 2 {
316 1 : info.Annotations = append(info.Annotations, "multilevel")
317 1 : }
318 1 : return info
319 : }
320 :
321 1 : func (c *compaction) userKeyBounds() base.UserKeyBounds {
322 1 : return base.UserKeyBoundsFromInternal(c.smallest, c.largest)
323 1 : }
324 :
325 : func newCompaction(
326 : pc *pickedCompaction, opts *Options, beganAt time.Time, provider objstorage.Provider,
327 1 : ) *compaction {
328 1 : c := &compaction{
329 1 : kind: compactionKindDefault,
330 1 : cmp: pc.cmp,
331 1 : equal: opts.Comparer.Equal,
332 1 : comparer: opts.Comparer,
333 1 : formatKey: opts.Comparer.FormatKey,
334 1 : inputs: pc.inputs,
335 1 : smallest: pc.smallest,
336 1 : largest: pc.largest,
337 1 : logger: opts.Logger,
338 1 : version: pc.version,
339 1 : beganAt: beganAt,
340 1 : maxOutputFileSize: pc.maxOutputFileSize,
341 1 : maxOverlapBytes: pc.maxOverlapBytes,
342 1 : pickerMetrics: pc.pickerMetrics,
343 1 : }
344 1 : c.startLevel = &c.inputs[0]
345 1 : if pc.startLevel.l0SublevelInfo != nil {
346 1 : c.startLevel.l0SublevelInfo = pc.startLevel.l0SublevelInfo
347 1 : }
348 1 : c.outputLevel = &c.inputs[1]
349 1 :
350 1 : if len(pc.extraLevels) > 0 {
351 1 : c.extraLevels = pc.extraLevels
352 1 : c.outputLevel = &c.inputs[len(c.inputs)-1]
353 1 : }
354 : // Compute the set of outputLevel+1 files that overlap this compaction (these
355 : // are the grandparent sstables).
356 1 : if c.outputLevel.level+1 < numLevels {
357 1 : c.grandparents = c.version.Overlaps(c.outputLevel.level+1, c.userKeyBounds())
358 1 : }
359 1 : c.delElision, c.rangeKeyElision = compact.SetupTombstoneElision(
360 1 : c.cmp, c.version, c.outputLevel.level, base.UserKeyBoundsFromInternal(c.smallest, c.largest),
361 1 : )
362 1 : c.kind = pc.kind
363 1 :
364 1 : if c.kind == compactionKindDefault && c.outputLevel.files.Empty() && !c.hasExtraLevelData() &&
365 1 : c.startLevel.files.Len() == 1 && c.grandparents.SizeSum() <= c.maxOverlapBytes {
366 1 : // This compaction can be converted into a move or copy from one level
367 1 : // to the next. We avoid such a move if there is lots of overlapping
368 1 : // grandparent data. Otherwise, the move could create a parent file
369 1 : // that will require a very expensive merge later on.
370 1 : iter := c.startLevel.files.Iter()
371 1 : meta := iter.First()
372 1 : isRemote := false
373 1 : // We should always be passed a provider, except in some unit tests.
374 1 : if provider != nil {
375 1 : isRemote = !objstorage.IsLocalTable(provider, meta.FileBacking.DiskFileNum)
376 1 : }
377 : // Avoid a trivial move or copy if all of these are true, as rewriting a
378 : // new file is better:
379 : //
380 : // 1) The source file is a virtual sstable
381 : // 2) The existing file `meta` is on non-remote storage
382 : // 3) The output level prefers shared storage
383 1 : mustCopy := !isRemote && remote.ShouldCreateShared(opts.Experimental.CreateOnShared, c.outputLevel.level)
384 1 : if mustCopy {
385 1 : // If the source is virtual, it's best to just rewrite the file as all
386 1 : // conditions in the above comment are met.
387 1 : if !meta.Virtual {
388 1 : c.kind = compactionKindCopy
389 1 : }
390 1 : } else {
391 1 : c.kind = compactionKindMove
392 1 : }
393 : }
394 1 : return c
395 : }
396 :
397 : func newDeleteOnlyCompaction(
398 : opts *Options, cur *version, inputs []compactionLevel, beganAt time.Time,
399 1 : ) *compaction {
400 1 : c := &compaction{
401 1 : kind: compactionKindDeleteOnly,
402 1 : cmp: opts.Comparer.Compare,
403 1 : equal: opts.Comparer.Equal,
404 1 : comparer: opts.Comparer,
405 1 : formatKey: opts.Comparer.FormatKey,
406 1 : logger: opts.Logger,
407 1 : version: cur,
408 1 : beganAt: beganAt,
409 1 : inputs: inputs,
410 1 : }
411 1 :
412 1 : // Set c.smallest, c.largest.
413 1 : files := make([]manifest.LevelIterator, 0, len(inputs))
414 1 : for _, in := range inputs {
415 1 : files = append(files, in.files.Iter())
416 1 : }
417 1 : c.smallest, c.largest = manifest.KeyRange(opts.Comparer.Compare, files...)
418 1 : return c
419 : }
420 :
421 1 : func adjustGrandparentOverlapBytesForFlush(c *compaction, flushingBytes uint64) {
422 1 : // Heuristic to place a lower bound on compaction output file size
423 1 : // caused by Lbase. Prior to this heuristic we have observed an L0 in
424 1 : // production with 310K files of which 290K files were < 10KB in size.
425 1 : // Our hypothesis is that it was caused by L1 having 2600 files and
426 1 : // ~10GB, such that each flush got split into many tiny files due to
427 1 : // overlapping with most of the files in Lbase.
428 1 : //
429 1 : // The computation below is general in that it accounts
430 1 : // for flushing different volumes of data (e.g. we may be flushing
431 1 : // many memtables). For illustration, we consider the typical
432 1 : // example of flushing a 64MB memtable. So 12.8MB output,
433 1 : // based on the compression guess below. If the compressed bytes
434 1 : // guess is an over-estimate we will end up with smaller files,
435 1 : // and if an under-estimate we will end up with larger files.
436 1 : // With a 2MB target file size, 7 files. We are willing to accept
437 1 : // 4x the number of files, if it results in better write amplification
438 1 : // when later compacting to Lbase, i.e., ~450KB files (target file
439 1 : // size / 4).
440 1 : //
441 1 : // Note that this is a pessimistic heuristic in that
442 1 : // fileCountUpperBoundDueToGrandparents could be far from the actual
443 1 : // number of files produced due to the grandparent limits. For
444 1 : // example, in the extreme, consider a flush that overlaps with 1000
445 1 : // files in Lbase f0...f999, and the initially calculated value of
446 1 : // maxOverlapBytes will cause splits at f10, f20,..., f990, which
447 1 : // means an upper bound file count of 100 files. Say the input bytes
448 1 : // in the flush are such that acceptableFileCount=10. We will fatten
449 1 : // up maxOverlapBytes by 10x to ensure that the upper bound file count
450 1 : // drops to 10. However, it is possible that in practice, even without
451 1 : // this change, we would have produced no more than 10 files, and that
452 1 : // this change makes the files unnecessarily wide. Say the input bytes
453 1 : // are distributed such that 10% are in f0...f9, 10% in f10...f19, ...
454 1 : // 10% in f80...f89 and 10% in f990...f999. The original value of
455 1 : // maxOverlapBytes would have actually produced only 10 sstables. But
456 1 : // by increasing maxOverlapBytes by 10x, we may produce 1 sstable that
457 1 : // spans f0...f89, i.e., a much wider sstable than necessary.
458 1 : //
459 1 : // We could produce a tighter estimate of
460 1 : // fileCountUpperBoundDueToGrandparents if we had knowledge of the key
461 1 : // distribution of the flush. The 4x multiplier mentioned earlier is
462 1 : // a way to try to compensate for this pessimism.
463 1 : //
464 1 : // TODO(sumeer): we don't have compression info for the data being
465 1 : // flushed, but it is likely that existing files that overlap with
466 1 : // this flush in Lbase are representative wrt compression ratio. We
467 1 : // could store the uncompressed size in FileMetadata and estimate
468 1 : // the compression ratio.
469 1 : const approxCompressionRatio = 0.2
470 1 : approxOutputBytes := approxCompressionRatio * float64(flushingBytes)
471 1 : approxNumFilesBasedOnTargetSize :=
472 1 : int(math.Ceil(approxOutputBytes / float64(c.maxOutputFileSize)))
473 1 : acceptableFileCount := float64(4 * approxNumFilesBasedOnTargetSize)
474 1 : // The byte calculation is linear in numGrandparentFiles, but we will
475 1 : // incur this linear cost in findGrandparentLimit too, so we are also
476 1 : // willing to pay it now. We could approximate this cheaply by using
477 1 : // the mean file size of Lbase.
478 1 : grandparentFileBytes := c.grandparents.SizeSum()
479 1 : fileCountUpperBoundDueToGrandparents :=
480 1 : float64(grandparentFileBytes) / float64(c.maxOverlapBytes)
481 1 : if fileCountUpperBoundDueToGrandparents > acceptableFileCount {
482 1 : c.maxOverlapBytes = uint64(
483 1 : float64(c.maxOverlapBytes) *
484 1 : (fileCountUpperBoundDueToGrandparents / acceptableFileCount))
485 1 : }
486 : }
487 :
488 : func newFlush(
489 : opts *Options, cur *version, baseLevel int, flushing flushableList, beganAt time.Time,
490 1 : ) (*compaction, error) {
491 1 : c := &compaction{
492 1 : kind: compactionKindFlush,
493 1 : cmp: opts.Comparer.Compare,
494 1 : equal: opts.Comparer.Equal,
495 1 : comparer: opts.Comparer,
496 1 : formatKey: opts.Comparer.FormatKey,
497 1 : logger: opts.Logger,
498 1 : version: cur,
499 1 : beganAt: beganAt,
500 1 : inputs: []compactionLevel{{level: -1}, {level: 0}},
501 1 : maxOutputFileSize: math.MaxUint64,
502 1 : maxOverlapBytes: math.MaxUint64,
503 1 : flushing: flushing,
504 1 : }
505 1 : c.startLevel = &c.inputs[0]
506 1 : c.outputLevel = &c.inputs[1]
507 1 :
508 1 : if len(flushing) > 0 {
509 1 : if _, ok := flushing[0].flushable.(*ingestedFlushable); ok {
510 1 : if len(flushing) != 1 {
511 0 : panic("pebble: ingestedFlushable must be flushed one at a time.")
512 : }
513 1 : c.kind = compactionKindIngestedFlushable
514 1 : return c, nil
515 : }
516 : }
517 :
518 : // Make sure there's no ingestedFlushable after the first flushable in the
519 : // list.
520 1 : for _, f := range flushing {
521 1 : if _, ok := f.flushable.(*ingestedFlushable); ok {
522 0 : panic("pebble: flushing shouldn't contain ingestedFlushable flushable")
523 : }
524 : }
525 :
526 1 : if cur.L0Sublevels != nil {
527 1 : c.l0Limits = cur.L0Sublevels.FlushSplitKeys()
528 1 : }
529 :
530 1 : smallestSet, largestSet := false, false
531 1 : updatePointBounds := func(iter internalIterator) {
532 1 : if kv := iter.First(); kv != nil {
533 1 : if !smallestSet ||
534 1 : base.InternalCompare(c.cmp, c.smallest, kv.K) > 0 {
535 1 : smallestSet = true
536 1 : c.smallest = kv.K.Clone()
537 1 : }
538 : }
539 1 : if kv := iter.Last(); kv != nil {
540 1 : if !largestSet ||
541 1 : base.InternalCompare(c.cmp, c.largest, kv.K) < 0 {
542 1 : largestSet = true
543 1 : c.largest = kv.K.Clone()
544 1 : }
545 : }
546 : }
547 :
548 1 : updateRangeBounds := func(iter keyspan.FragmentIterator) error {
549 1 : // File bounds require s != nil && !s.Empty(). We only need to check for
550 1 : // s != nil here, as the memtable's FragmentIterator would never surface
551 1 : // empty spans.
552 1 : if s, err := iter.First(); err != nil {
553 0 : return err
554 1 : } else if s != nil {
555 1 : if key := s.SmallestKey(); !smallestSet ||
556 1 : base.InternalCompare(c.cmp, c.smallest, key) > 0 {
557 1 : smallestSet = true
558 1 : c.smallest = key.Clone()
559 1 : }
560 : }
561 1 : if s, err := iter.Last(); err != nil {
562 0 : return err
563 1 : } else if s != nil {
564 1 : if key := s.LargestKey(); !largestSet ||
565 1 : base.InternalCompare(c.cmp, c.largest, key) < 0 {
566 1 : largestSet = true
567 1 : c.largest = key.Clone()
568 1 : }
569 : }
570 1 : return nil
571 : }
572 :
573 1 : var flushingBytes uint64
574 1 : for i := range flushing {
575 1 : f := flushing[i]
576 1 : updatePointBounds(f.newIter(nil))
577 1 : if rangeDelIter := f.newRangeDelIter(nil); rangeDelIter != nil {
578 1 : if err := updateRangeBounds(rangeDelIter); err != nil {
579 0 : return nil, err
580 0 : }
581 : }
582 1 : if rangeKeyIter := f.newRangeKeyIter(nil); rangeKeyIter != nil {
583 1 : if err := updateRangeBounds(rangeKeyIter); err != nil {
584 0 : return nil, err
585 0 : }
586 : }
587 1 : flushingBytes += f.inuseBytes()
588 : }
589 :
590 1 : if opts.FlushSplitBytes > 0 {
591 1 : c.maxOutputFileSize = uint64(opts.Level(0).TargetFileSize)
592 1 : c.maxOverlapBytes = maxGrandparentOverlapBytes(opts, 0)
593 1 : c.grandparents = c.version.Overlaps(baseLevel, c.userKeyBounds())
594 1 : adjustGrandparentOverlapBytesForFlush(c, flushingBytes)
595 1 : }
596 :
597 : // We don't elide tombstones for flushes.
598 1 : c.delElision, c.rangeKeyElision = compact.NoTombstoneElision(), compact.NoTombstoneElision()
599 1 : return c, nil
600 : }
601 :
602 1 : func (c *compaction) hasExtraLevelData() bool {
603 1 : if len(c.extraLevels) == 0 {
604 1 : // not a multi level compaction
605 1 : return false
606 1 : } else if c.extraLevels[0].files.Empty() {
607 1 : // a multi level compaction without data in the intermediate input level;
608 1 : // e.g. for a multi level compaction with levels 4,5, and 6, this could
609 1 : // occur if there is no files to compact in 5, or in 5 and 6 (i.e. a move).
610 1 : return false
611 1 : }
612 1 : return true
613 : }
614 :
615 : // findGrandparentLimit takes the start user key for a table and returns the
616 : // user key to which that table can extend without excessively overlapping
617 : // the grandparent level. If no limit is needed considering the grandparent
618 : // files, this function returns nil. This is done in order to prevent a table
619 : // at level N from overlapping too much data at level N+1. We want to avoid
620 : // such large overlaps because they translate into large compactions. The
621 : // current heuristic stops output of a table if the addition of another key
622 : // would cause the table to overlap more than 10x the target file size at
623 : // level N. See maxGrandparentOverlapBytes.
624 1 : func (c *compaction) findGrandparentLimit(start []byte) []byte {
625 1 : iter := c.grandparents.Iter()
626 1 : var overlappedBytes uint64
627 1 : var greater bool
628 1 : for f := iter.SeekGE(c.cmp, start); f != nil; f = iter.Next() {
629 1 : overlappedBytes += f.Size
630 1 : // To ensure forward progress we always return a larger user
631 1 : // key than where we started. See comments above clients of
632 1 : // this function for how this is used.
633 1 : greater = greater || c.cmp(f.Smallest.UserKey, start) > 0
634 1 : if !greater {
635 1 : continue
636 : }
637 :
638 : // We return the smallest bound of a sstable rather than the
639 : // largest because the smallest is always inclusive, and limits
640 : // are used exlusively when truncating range tombstones. If we
641 : // truncated an output to the largest key while there's a
642 : // pending tombstone, the next output file would also overlap
643 : // the same grandparent f.
644 1 : if overlappedBytes > c.maxOverlapBytes {
645 1 : return f.Smallest.UserKey
646 1 : }
647 : }
648 1 : return nil
649 : }
650 :
651 : // findL0Limit takes the start key for a table and returns the user key to which
652 : // that table can be extended without hitting the next l0Limit. Having flushed
653 : // sstables "bridging across" an l0Limit could lead to increased L0 -> LBase
654 : // compaction sizes as well as elevated read amplification.
655 1 : func (c *compaction) findL0Limit(start []byte) []byte {
656 1 : if c.startLevel.level > -1 || c.outputLevel.level != 0 || len(c.l0Limits) == 0 {
657 1 : return nil
658 1 : }
659 1 : index := sort.Search(len(c.l0Limits), func(i int) bool {
660 1 : return c.cmp(c.l0Limits[i], start) > 0
661 1 : })
662 1 : if index < len(c.l0Limits) {
663 1 : return c.l0Limits[index]
664 1 : }
665 1 : return nil
666 : }
667 :
668 : // errorOnUserKeyOverlap returns an error if the last two written sstables in
669 : // this compaction have revisions of the same user key present in both sstables,
670 : // when it shouldn't (eg. when splitting flushes).
671 1 : func (c *compaction) errorOnUserKeyOverlap(ve *versionEdit) error {
672 1 : if n := len(ve.NewFiles); n > 1 {
673 1 : meta := ve.NewFiles[n-1].Meta
674 1 : prevMeta := ve.NewFiles[n-2].Meta
675 1 : if !prevMeta.Largest.IsExclusiveSentinel() &&
676 1 : c.cmp(prevMeta.Largest.UserKey, meta.Smallest.UserKey) >= 0 {
677 1 : return errors.Errorf("pebble: compaction split user key across two sstables: %s in %s and %s",
678 1 : prevMeta.Largest.Pretty(c.formatKey),
679 1 : prevMeta.FileNum,
680 1 : meta.FileNum)
681 1 : }
682 : }
683 1 : return nil
684 : }
685 :
686 : // allowZeroSeqNum returns true if seqnum's can be zeroed if there are no
687 : // snapshots requiring them to be kept. It performs this determination by
688 : // looking at the TombstoneElision values which are set up based on sstables
689 : // which overlap the bounds of the compaction at a lower level in the LSM.
690 1 : func (c *compaction) allowZeroSeqNum() bool {
691 1 : // TODO(peter): we disable zeroing of seqnums during flushing to match
692 1 : // RocksDB behavior and to avoid generating overlapping sstables during
693 1 : // DB.replayWAL. When replaying WAL files at startup, we flush after each
694 1 : // WAL is replayed building up a single version edit that is
695 1 : // applied. Because we don't apply the version edit after each flush, this
696 1 : // code doesn't know that L0 contains files and zeroing of seqnums should
697 1 : // be disabled. That is fixable, but it seems safer to just match the
698 1 : // RocksDB behavior for now.
699 1 : return len(c.flushing) == 0 && c.delElision.ElidesEverything() && c.rangeKeyElision.ElidesEverything()
700 1 : }
701 :
702 : // newInputIter returns an iterator over all the input tables in a compaction.
703 : func (c *compaction) newInputIter(
704 : newIters tableNewIters, newRangeKeyIter keyspanimpl.TableNewSpanIter,
705 1 : ) (_ internalIterator, retErr error) {
706 1 : // Validate the ordering of compaction input files for defense in depth.
707 1 : if len(c.flushing) == 0 {
708 1 : if c.startLevel.level >= 0 {
709 1 : err := manifest.CheckOrdering(c.cmp, c.formatKey,
710 1 : manifest.Level(c.startLevel.level), c.startLevel.files.Iter())
711 1 : if err != nil {
712 1 : return nil, err
713 1 : }
714 : }
715 1 : err := manifest.CheckOrdering(c.cmp, c.formatKey,
716 1 : manifest.Level(c.outputLevel.level), c.outputLevel.files.Iter())
717 1 : if err != nil {
718 1 : return nil, err
719 1 : }
720 1 : if c.startLevel.level == 0 {
721 1 : if c.startLevel.l0SublevelInfo == nil {
722 0 : panic("l0SublevelInfo not created for compaction out of L0")
723 : }
724 1 : for _, info := range c.startLevel.l0SublevelInfo {
725 1 : err := manifest.CheckOrdering(c.cmp, c.formatKey,
726 1 : info.sublevel, info.Iter())
727 1 : if err != nil {
728 1 : return nil, err
729 1 : }
730 : }
731 : }
732 1 : if len(c.extraLevels) > 0 {
733 1 : if len(c.extraLevels) > 1 {
734 0 : panic("n>2 multi level compaction not implemented yet")
735 : }
736 1 : interLevel := c.extraLevels[0]
737 1 : err := manifest.CheckOrdering(c.cmp, c.formatKey,
738 1 : manifest.Level(interLevel.level), interLevel.files.Iter())
739 1 : if err != nil {
740 0 : return nil, err
741 0 : }
742 : }
743 : }
744 :
745 : // There are three classes of keys that a compaction needs to process: point
746 : // keys, range deletion tombstones and range keys. Collect all iterators for
747 : // all these classes of keys from all the levels. We'll aggregate them
748 : // together farther below.
749 : //
750 : // numInputLevels is an approximation of the number of iterator levels. Due
751 : // to idiosyncrasies in iterator construction, we may (rarely) exceed this
752 : // initial capacity.
753 1 : numInputLevels := max(len(c.flushing), len(c.inputs))
754 1 : iters := make([]internalIterator, 0, numInputLevels)
755 1 : rangeDelIters := make([]keyspan.FragmentIterator, 0, numInputLevels)
756 1 : rangeKeyIters := make([]keyspan.FragmentIterator, 0, numInputLevels)
757 1 :
758 1 : // If construction of the iterator inputs fails, ensure that we close all
759 1 : // the consitutent iterators.
760 1 : defer func() {
761 1 : if retErr != nil {
762 0 : for _, iter := range iters {
763 0 : if iter != nil {
764 0 : iter.Close()
765 0 : }
766 : }
767 0 : for _, rangeDelIter := range rangeDelIters {
768 0 : rangeDelIter.Close()
769 0 : }
770 : }
771 : }()
772 1 : iterOpts := IterOptions{
773 1 : CategoryAndQoS: sstable.CategoryAndQoS{
774 1 : Category: "pebble-compaction",
775 1 : QoSLevel: sstable.NonLatencySensitiveQoSLevel,
776 1 : },
777 1 : logger: c.logger,
778 1 : }
779 1 :
780 1 : // Populate iters, rangeDelIters and rangeKeyIters with the appropriate
781 1 : // constituent iterators. This depends on whether this is a flush or a
782 1 : // compaction.
783 1 : if len(c.flushing) != 0 {
784 1 : // If flushing, we need to build the input iterators over the memtables
785 1 : // stored in c.flushing.
786 1 : for i := range c.flushing {
787 1 : f := c.flushing[i]
788 1 : iters = append(iters, f.newFlushIter(nil))
789 1 : rangeDelIter := f.newRangeDelIter(nil)
790 1 : if rangeDelIter != nil {
791 1 : rangeDelIters = append(rangeDelIters, rangeDelIter)
792 1 : }
793 1 : if rangeKeyIter := f.newRangeKeyIter(nil); rangeKeyIter != nil {
794 1 : rangeKeyIters = append(rangeKeyIters, rangeKeyIter)
795 1 : }
796 : }
797 1 : } else {
798 1 : addItersForLevel := func(level *compactionLevel, l manifest.Level) error {
799 1 : // Add a *levelIter for point iterators. Because we don't call
800 1 : // initRangeDel, the levelIter will close and forget the range
801 1 : // deletion iterator when it steps on to a new file. Surfacing range
802 1 : // deletions to compactions are handled below.
803 1 : iters = append(iters, newLevelIter(context.Background(),
804 1 : iterOpts, c.comparer, newIters, level.files.Iter(), l, internalIterOpts{
805 1 : compaction: true,
806 1 : bufferPool: &c.bufferPool,
807 1 : }))
808 1 : // TODO(jackson): Use keyspanimpl.LevelIter to avoid loading all the range
809 1 : // deletions into memory upfront. (See #2015, which reverted this.) There
810 1 : // will be no user keys that are split between sstables within a level in
811 1 : // Cockroach 23.1, which unblocks this optimization.
812 1 :
813 1 : // Add the range deletion iterator for each file as an independent level
814 1 : // in mergingIter, as opposed to making a levelIter out of those. This
815 1 : // is safer as levelIter expects all keys coming from underlying
816 1 : // iterators to be in order. Due to compaction / tombstone writing
817 1 : // logic in finishOutput(), it is possible for range tombstones to not
818 1 : // be strictly ordered across all files in one level.
819 1 : //
820 1 : // Consider this example from the metamorphic tests (also repeated in
821 1 : // finishOutput()), consisting of three L3 files with their bounds
822 1 : // specified in square brackets next to the file name:
823 1 : //
824 1 : // ./000240.sst [tmgc#391,MERGE-tmgc#391,MERGE]
825 1 : // tmgc#391,MERGE [786e627a]
826 1 : // tmgc-udkatvs#331,RANGEDEL
827 1 : //
828 1 : // ./000241.sst [tmgc#384,MERGE-tmgc#384,MERGE]
829 1 : // tmgc#384,MERGE [666c7070]
830 1 : // tmgc-tvsalezade#383,RANGEDEL
831 1 : // tmgc-tvsalezade#331,RANGEDEL
832 1 : //
833 1 : // ./000242.sst [tmgc#383,RANGEDEL-tvsalezade#72057594037927935,RANGEDEL]
834 1 : // tmgc-tvsalezade#383,RANGEDEL
835 1 : // tmgc#375,SET [72646c78766965616c72776865676e79]
836 1 : // tmgc-tvsalezade#356,RANGEDEL
837 1 : //
838 1 : // Here, the range tombstone in 000240.sst falls "after" one in
839 1 : // 000241.sst, despite 000240.sst being ordered "before" 000241.sst for
840 1 : // levelIter's purposes. While each file is still consistent before its
841 1 : // bounds, it's safer to have all rangedel iterators be visible to
842 1 : // mergingIter.
843 1 : iter := level.files.Iter()
844 1 : for f := iter.First(); f != nil; f = iter.Next() {
845 1 : rangeDelIter, closer, err := c.newRangeDelIter(
846 1 : newIters, iter.Take(), iterOpts, l)
847 1 : if err != nil {
848 0 : // The error will already be annotated with the BackingFileNum, so
849 0 : // we annotate it with the FileNum.
850 0 : return errors.Wrapf(err, "pebble: could not open table %s", errors.Safe(f.FileNum))
851 0 : }
852 1 : if rangeDelIter == nil {
853 1 : continue
854 : }
855 1 : rangeDelIters = append(rangeDelIters, rangeDelIter)
856 1 : c.closers = append(c.closers, closer)
857 : }
858 :
859 : // Check if this level has any range keys.
860 1 : hasRangeKeys := false
861 1 : for f := iter.First(); f != nil; f = iter.Next() {
862 1 : if f.HasRangeKeys {
863 1 : hasRangeKeys = true
864 1 : break
865 : }
866 : }
867 1 : if hasRangeKeys {
868 1 : li := &keyspanimpl.LevelIter{}
869 1 : newRangeKeyIterWrapper := func(file *manifest.FileMetadata, iterOptions keyspan.SpanIterOptions) (keyspan.FragmentIterator, error) {
870 1 : iter, err := newRangeKeyIter(file, iterOptions)
871 1 : if err != nil {
872 0 : return nil, err
873 1 : } else if iter == nil {
874 0 : return emptyKeyspanIter, nil
875 0 : }
876 : // Ensure that the range key iter is not closed until the compaction is
877 : // finished. This is necessary because range key processing
878 : // requires the range keys to be held in memory for up to the
879 : // lifetime of the compaction.
880 1 : c.closers = append(c.closers, iter)
881 1 : iter = noCloseIter{iter}
882 1 :
883 1 : // We do not need to truncate range keys to sstable boundaries, or
884 1 : // only read within the file's atomic compaction units, unlike with
885 1 : // range tombstones. This is because range keys were added after we
886 1 : // stopped splitting user keys across sstables, so all the range keys
887 1 : // in this sstable must wholly lie within the file's bounds.
888 1 : return iter, err
889 : }
890 1 : li.Init(keyspan.SpanIterOptions{}, c.cmp, newRangeKeyIterWrapper, level.files.Iter(), l, manifest.KeyTypeRange)
891 1 : rangeKeyIters = append(rangeKeyIters, li)
892 : }
893 1 : return nil
894 : }
895 :
896 1 : for i := range c.inputs {
897 1 : // If the level is annotated with l0SublevelInfo, expand it into one
898 1 : // level per sublevel.
899 1 : // TODO(jackson): Perform this expansion even earlier when we pick the
900 1 : // compaction?
901 1 : if len(c.inputs[i].l0SublevelInfo) > 0 {
902 1 : for _, info := range c.startLevel.l0SublevelInfo {
903 1 : sublevelCompactionLevel := &compactionLevel{0, info.LevelSlice, nil}
904 1 : if err := addItersForLevel(sublevelCompactionLevel, info.sublevel); err != nil {
905 0 : return nil, err
906 0 : }
907 : }
908 1 : continue
909 : }
910 1 : if err := addItersForLevel(&c.inputs[i], manifest.Level(c.inputs[i].level)); err != nil {
911 0 : return nil, err
912 0 : }
913 : }
914 : }
915 :
916 : // If there's only one constituent point iterator, we can avoid the overhead
917 : // of a *mergingIter. This is possible, for example, when performing a flush
918 : // of a single memtable. Otherwise, combine all the iterators into a merging
919 : // iter.
920 1 : iter := iters[0]
921 1 : if len(iters) > 1 {
922 1 : iter = newMergingIter(c.logger, &c.stats, c.cmp, nil, iters...)
923 1 : }
924 :
925 : // In normal operation, levelIter iterates over the point operations in a
926 : // level, and initializes a rangeDelIter pointer for the range deletions in
927 : // each table. During compaction, we want to iterate over the merged view of
928 : // point operations and range deletions. In order to do this we create one
929 : // levelIter per level to iterate over the point operations, and collect up
930 : // all the range deletion files.
931 : //
932 : // The range deletion levels are combined with a keyspanimpl.MergingIter. The
933 : // resulting merged rangedel iterator is then included using an
934 : // InterleavingIter.
935 : // TODO(jackson): Consider using a defragmenting iterator to stitch together
936 : // logical range deletions that were fragmented due to previous file
937 : // boundaries.
938 1 : if len(rangeDelIters) > 0 {
939 1 : mi := &keyspanimpl.MergingIter{}
940 1 : mi.Init(c.cmp, keyspan.NoopTransform, new(keyspanimpl.MergingBuffers), rangeDelIters...)
941 1 : c.rangeDelInterleaving.Init(c.comparer, iter, mi, keyspan.InterleavingIterOpts{})
942 1 : iter = &c.rangeDelInterleaving
943 1 : }
944 :
945 : // If there are range key iterators, we need to combine them using
946 : // keyspanimpl.MergingIter, and then interleave them among the points.
947 1 : if len(rangeKeyIters) > 0 {
948 1 : mi := &keyspanimpl.MergingIter{}
949 1 : mi.Init(c.cmp, keyspan.NoopTransform, new(keyspanimpl.MergingBuffers), rangeKeyIters...)
950 1 : di := &keyspan.DefragmentingIter{}
951 1 : di.Init(c.comparer, mi, keyspan.DefragmentInternal, keyspan.StaticDefragmentReducer, new(keyspan.DefragmentingBuffers))
952 1 : c.rangeKeyInterleaving.Init(c.comparer, iter, di, keyspan.InterleavingIterOpts{})
953 1 : iter = &c.rangeKeyInterleaving
954 1 : }
955 1 : return iter, nil
956 : }
957 :
958 : func (c *compaction) newRangeDelIter(
959 : newIters tableNewIters, f manifest.LevelFile, opts IterOptions, l manifest.Level,
960 1 : ) (keyspan.FragmentIterator, io.Closer, error) {
961 1 : opts.level = l
962 1 : iterSet, err := newIters(context.Background(), f.FileMetadata, &opts,
963 1 : internalIterOpts{
964 1 : compaction: true,
965 1 : bufferPool: &c.bufferPool,
966 1 : }, iterRangeDeletions)
967 1 : if err != nil {
968 0 : return nil, nil, err
969 1 : } else if iterSet.rangeDeletion == nil {
970 1 : // The file doesn't contain any range deletions.
971 1 : return nil, nil, nil
972 1 : }
973 : // Ensure that rangeDelIter is not closed until the compaction is
974 : // finished. This is necessary because range tombstone processing
975 : // requires the range tombstones to be held in memory for up to the
976 : // lifetime of the compaction.
977 1 : return noCloseIter{iterSet.rangeDeletion}, iterSet.rangeDeletion, nil
978 : }
979 :
980 1 : func (c *compaction) String() string {
981 1 : if len(c.flushing) != 0 {
982 0 : return "flush\n"
983 0 : }
984 :
985 1 : var buf bytes.Buffer
986 1 : for level := c.startLevel.level; level <= c.outputLevel.level; level++ {
987 1 : i := level - c.startLevel.level
988 1 : fmt.Fprintf(&buf, "%d:", level)
989 1 : iter := c.inputs[i].files.Iter()
990 1 : for f := iter.First(); f != nil; f = iter.Next() {
991 1 : fmt.Fprintf(&buf, " %s:%s-%s", f.FileNum, f.Smallest, f.Largest)
992 1 : }
993 1 : fmt.Fprintf(&buf, "\n")
994 : }
995 1 : return buf.String()
996 : }
997 :
998 : type manualCompaction struct {
999 : // Count of the retries either due to too many concurrent compactions, or a
1000 : // concurrent compaction to overlapping levels.
1001 : retries int
1002 : level int
1003 : outputLevel int
1004 : done chan error
1005 : start []byte
1006 : end []byte
1007 : split bool
1008 : }
1009 :
1010 : type readCompaction struct {
1011 : level int
1012 : // [start, end] key ranges are used for de-duping.
1013 : start []byte
1014 : end []byte
1015 :
1016 : // The file associated with the compaction.
1017 : // If the file no longer belongs in the same
1018 : // level, then we skip the compaction.
1019 : fileNum base.FileNum
1020 : }
1021 :
1022 1 : func (d *DB) addInProgressCompaction(c *compaction) {
1023 1 : d.mu.compact.inProgress[c] = struct{}{}
1024 1 : var isBase, isIntraL0 bool
1025 1 : for _, cl := range c.inputs {
1026 1 : iter := cl.files.Iter()
1027 1 : for f := iter.First(); f != nil; f = iter.Next() {
1028 1 : if f.IsCompacting() {
1029 0 : d.opts.Logger.Fatalf("L%d->L%d: %s already being compacted", c.startLevel.level, c.outputLevel.level, f.FileNum)
1030 0 : }
1031 1 : f.SetCompactionState(manifest.CompactionStateCompacting)
1032 1 : if c.startLevel != nil && c.outputLevel != nil && c.startLevel.level == 0 {
1033 1 : if c.outputLevel.level == 0 {
1034 1 : f.IsIntraL0Compacting = true
1035 1 : isIntraL0 = true
1036 1 : } else {
1037 1 : isBase = true
1038 1 : }
1039 : }
1040 : }
1041 : }
1042 :
1043 1 : if (isIntraL0 || isBase) && c.version.L0Sublevels != nil {
1044 1 : l0Inputs := []manifest.LevelSlice{c.startLevel.files}
1045 1 : if isIntraL0 {
1046 1 : l0Inputs = append(l0Inputs, c.outputLevel.files)
1047 1 : }
1048 1 : if err := c.version.L0Sublevels.UpdateStateForStartedCompaction(l0Inputs, isBase); err != nil {
1049 0 : d.opts.Logger.Fatalf("could not update state for compaction: %s", err)
1050 0 : }
1051 : }
1052 : }
1053 :
1054 : // Removes compaction markers from files in a compaction. The rollback parameter
1055 : // indicates whether the compaction state should be rolled back to its original
1056 : // state in the case of an unsuccessful compaction.
1057 : //
1058 : // DB.mu must be held when calling this method, however this method can drop and
1059 : // re-acquire that mutex. All writes to the manifest for this compaction should
1060 : // have completed by this point.
1061 1 : func (d *DB) clearCompactingState(c *compaction, rollback bool) {
1062 1 : c.versionEditApplied = true
1063 1 : for _, cl := range c.inputs {
1064 1 : iter := cl.files.Iter()
1065 1 : for f := iter.First(); f != nil; f = iter.Next() {
1066 1 : if !f.IsCompacting() {
1067 0 : d.opts.Logger.Fatalf("L%d->L%d: %s not being compacted", c.startLevel.level, c.outputLevel.level, f.FileNum)
1068 0 : }
1069 1 : if !rollback {
1070 1 : // On success all compactions other than move-compactions transition the
1071 1 : // file into the Compacted state. Move-compacted files become eligible
1072 1 : // for compaction again and transition back to NotCompacting.
1073 1 : if c.kind != compactionKindMove {
1074 1 : f.SetCompactionState(manifest.CompactionStateCompacted)
1075 1 : } else {
1076 1 : f.SetCompactionState(manifest.CompactionStateNotCompacting)
1077 1 : }
1078 1 : } else {
1079 1 : // Else, on rollback, all input files unconditionally transition back to
1080 1 : // NotCompacting.
1081 1 : f.SetCompactionState(manifest.CompactionStateNotCompacting)
1082 1 : }
1083 1 : f.IsIntraL0Compacting = false
1084 : }
1085 : }
1086 1 : l0InProgress := inProgressL0Compactions(d.getInProgressCompactionInfoLocked(c))
1087 1 : func() {
1088 1 : // InitCompactingFileInfo requires that no other manifest writes be
1089 1 : // happening in parallel with it, i.e. we're not in the midst of installing
1090 1 : // another version. Otherwise, it's possible that we've created another
1091 1 : // L0Sublevels instance, but not added it to the versions list, causing
1092 1 : // all the indices in FileMetadata to be inaccurate. To ensure this,
1093 1 : // grab the manifest lock.
1094 1 : d.mu.versions.logLock()
1095 1 : defer d.mu.versions.logUnlock()
1096 1 : d.mu.versions.currentVersion().L0Sublevels.InitCompactingFileInfo(l0InProgress)
1097 1 : }()
1098 : }
1099 :
1100 1 : func (d *DB) calculateDiskAvailableBytes() uint64 {
1101 1 : if space, err := d.opts.FS.GetDiskUsage(d.dirname); err == nil {
1102 1 : d.diskAvailBytes.Store(space.AvailBytes)
1103 1 : return space.AvailBytes
1104 1 : } else if !errors.Is(err, vfs.ErrUnsupported) {
1105 1 : d.opts.EventListener.BackgroundError(err)
1106 1 : }
1107 1 : return d.diskAvailBytes.Load()
1108 : }
1109 :
1110 : // maybeScheduleFlush schedules a flush if necessary.
1111 : //
1112 : // d.mu must be held when calling this.
1113 1 : func (d *DB) maybeScheduleFlush() {
1114 1 : if d.mu.compact.flushing || d.closed.Load() != nil || d.opts.ReadOnly {
1115 1 : return
1116 1 : }
1117 1 : if len(d.mu.mem.queue) <= 1 {
1118 1 : return
1119 1 : }
1120 :
1121 1 : if !d.passedFlushThreshold() {
1122 1 : return
1123 1 : }
1124 :
1125 1 : d.mu.compact.flushing = true
1126 1 : go d.flush()
1127 : }
1128 :
1129 1 : func (d *DB) passedFlushThreshold() bool {
1130 1 : var n int
1131 1 : var size uint64
1132 1 : for ; n < len(d.mu.mem.queue)-1; n++ {
1133 1 : if !d.mu.mem.queue[n].readyForFlush() {
1134 1 : break
1135 : }
1136 1 : if d.mu.mem.queue[n].flushForced {
1137 1 : // A flush was forced. Pretend the memtable size is the configured
1138 1 : // size. See minFlushSize below.
1139 1 : size += d.opts.MemTableSize
1140 1 : } else {
1141 1 : size += d.mu.mem.queue[n].totalBytes()
1142 1 : }
1143 : }
1144 1 : if n == 0 {
1145 1 : // None of the immutable memtables are ready for flushing.
1146 1 : return false
1147 1 : }
1148 :
1149 : // Only flush once the sum of the queued memtable sizes exceeds half the
1150 : // configured memtable size. This prevents flushing of memtables at startup
1151 : // while we're undergoing the ramp period on the memtable size. See
1152 : // DB.newMemTable().
1153 1 : minFlushSize := d.opts.MemTableSize / 2
1154 1 : return size >= minFlushSize
1155 : }
1156 :
1157 1 : func (d *DB) maybeScheduleDelayedFlush(tbl *memTable, dur time.Duration) {
1158 1 : var mem *flushableEntry
1159 1 : for _, m := range d.mu.mem.queue {
1160 1 : if m.flushable == tbl {
1161 1 : mem = m
1162 1 : break
1163 : }
1164 : }
1165 1 : if mem == nil || mem.flushForced {
1166 1 : return
1167 1 : }
1168 1 : deadline := d.timeNow().Add(dur)
1169 1 : if !mem.delayedFlushForcedAt.IsZero() && deadline.After(mem.delayedFlushForcedAt) {
1170 1 : // Already scheduled to flush sooner than within `dur`.
1171 1 : return
1172 1 : }
1173 1 : mem.delayedFlushForcedAt = deadline
1174 1 : go func() {
1175 1 : timer := time.NewTimer(dur)
1176 1 : defer timer.Stop()
1177 1 :
1178 1 : select {
1179 1 : case <-d.closedCh:
1180 1 : return
1181 1 : case <-mem.flushed:
1182 1 : return
1183 1 : case <-timer.C:
1184 1 : d.commit.mu.Lock()
1185 1 : defer d.commit.mu.Unlock()
1186 1 : d.mu.Lock()
1187 1 : defer d.mu.Unlock()
1188 1 :
1189 1 : // NB: The timer may fire concurrently with a call to Close. If a
1190 1 : // Close call beat us to acquiring d.mu, d.closed holds ErrClosed,
1191 1 : // and it's too late to flush anything. Otherwise, the Close call
1192 1 : // will block on locking d.mu until we've finished scheduling the
1193 1 : // flush and set `d.mu.compact.flushing` to true. Close will wait
1194 1 : // for the current flush to complete.
1195 1 : if d.closed.Load() != nil {
1196 1 : return
1197 1 : }
1198 :
1199 1 : if d.mu.mem.mutable == tbl {
1200 1 : d.makeRoomForWrite(nil)
1201 1 : } else {
1202 1 : mem.flushForced = true
1203 1 : }
1204 1 : d.maybeScheduleFlush()
1205 : }
1206 : }()
1207 : }
1208 :
1209 1 : func (d *DB) flush() {
1210 1 : pprof.Do(context.Background(), flushLabels, func(context.Context) {
1211 1 : flushingWorkStart := time.Now()
1212 1 : d.mu.Lock()
1213 1 : defer d.mu.Unlock()
1214 1 : idleDuration := flushingWorkStart.Sub(d.mu.compact.noOngoingFlushStartTime)
1215 1 : var bytesFlushed uint64
1216 1 : var err error
1217 1 : if bytesFlushed, err = d.flush1(); err != nil {
1218 1 : // TODO(peter): count consecutive flush errors and backoff.
1219 1 : d.opts.EventListener.BackgroundError(err)
1220 1 : }
1221 1 : d.mu.compact.flushing = false
1222 1 : d.mu.compact.noOngoingFlushStartTime = time.Now()
1223 1 : workDuration := d.mu.compact.noOngoingFlushStartTime.Sub(flushingWorkStart)
1224 1 : d.mu.compact.flushWriteThroughput.Bytes += int64(bytesFlushed)
1225 1 : d.mu.compact.flushWriteThroughput.WorkDuration += workDuration
1226 1 : d.mu.compact.flushWriteThroughput.IdleDuration += idleDuration
1227 1 : // More flush work may have arrived while we were flushing, so schedule
1228 1 : // another flush if needed.
1229 1 : d.maybeScheduleFlush()
1230 1 : // The flush may have produced too many files in a level, so schedule a
1231 1 : // compaction if needed.
1232 1 : d.maybeScheduleCompaction()
1233 1 : d.mu.compact.cond.Broadcast()
1234 : })
1235 : }
1236 :
1237 : // runIngestFlush is used to generate a flush version edit for sstables which
1238 : // were ingested as flushables. Both DB.mu and the manifest lock must be held
1239 : // while runIngestFlush is called.
1240 1 : func (d *DB) runIngestFlush(c *compaction) (*manifest.VersionEdit, error) {
1241 1 : if len(c.flushing) != 1 {
1242 0 : panic("pebble: ingestedFlushable must be flushed one at a time.")
1243 : }
1244 :
1245 : // Construct the VersionEdit, levelMetrics etc.
1246 1 : c.metrics = make(map[int]*LevelMetrics, numLevels)
1247 1 : // Finding the target level for ingestion must use the latest version
1248 1 : // after the logLock has been acquired.
1249 1 : c.version = d.mu.versions.currentVersion()
1250 1 :
1251 1 : baseLevel := d.mu.versions.picker.getBaseLevel()
1252 1 : iterOpts := IterOptions{logger: d.opts.Logger}
1253 1 : ve := &versionEdit{}
1254 1 : var ingestSplitFiles []ingestSplitFile
1255 1 : ingestFlushable := c.flushing[0].flushable.(*ingestedFlushable)
1256 1 :
1257 1 : updateLevelMetricsOnExcise := func(m *fileMetadata, level int, added []newFileEntry) {
1258 1 : levelMetrics := c.metrics[level]
1259 1 : if levelMetrics == nil {
1260 1 : levelMetrics = &LevelMetrics{}
1261 1 : c.metrics[level] = levelMetrics
1262 1 : }
1263 1 : levelMetrics.NumFiles--
1264 1 : levelMetrics.Size -= int64(m.Size)
1265 1 : for i := range added {
1266 1 : levelMetrics.NumFiles++
1267 1 : levelMetrics.Size += int64(added[i].Meta.Size)
1268 1 : }
1269 : }
1270 :
1271 1 : suggestSplit := d.opts.Experimental.IngestSplit != nil && d.opts.Experimental.IngestSplit() &&
1272 1 : d.FormatMajorVersion() >= FormatVirtualSSTables
1273 1 :
1274 1 : if suggestSplit || ingestFlushable.exciseSpan.Valid() {
1275 1 : // We could add deleted files to ve.
1276 1 : ve.DeletedFiles = make(map[manifest.DeletedFileEntry]*manifest.FileMetadata)
1277 1 : }
1278 :
1279 1 : replacedFiles := make(map[base.FileNum][]newFileEntry)
1280 1 : for _, file := range ingestFlushable.files {
1281 1 : var fileToSplit *fileMetadata
1282 1 : var level int
1283 1 :
1284 1 : // This file fits perfectly within the excise span, so we can slot it at L6.
1285 1 : if ingestFlushable.exciseSpan.Valid() &&
1286 1 : ingestFlushable.exciseSpan.Contains(d.cmp, file.FileMetadata.Smallest) &&
1287 1 : ingestFlushable.exciseSpan.Contains(d.cmp, file.FileMetadata.Largest) {
1288 1 : level = 6
1289 1 : } else {
1290 1 : var err error
1291 1 : level, fileToSplit, err = ingestTargetLevel(
1292 1 : d.newIters, d.tableNewRangeKeyIter, iterOpts, d.opts.Comparer,
1293 1 : c.version, baseLevel, d.mu.compact.inProgress, file.FileMetadata,
1294 1 : suggestSplit,
1295 1 : )
1296 1 : if err != nil {
1297 0 : return nil, err
1298 0 : }
1299 : }
1300 :
1301 : // Add the current flushableIngest file to the version.
1302 1 : ve.NewFiles = append(ve.NewFiles, newFileEntry{Level: level, Meta: file.FileMetadata})
1303 1 : if fileToSplit != nil {
1304 0 : ingestSplitFiles = append(ingestSplitFiles, ingestSplitFile{
1305 0 : ingestFile: file.FileMetadata,
1306 0 : splitFile: fileToSplit,
1307 0 : level: level,
1308 0 : })
1309 0 : }
1310 1 : levelMetrics := c.metrics[level]
1311 1 : if levelMetrics == nil {
1312 1 : levelMetrics = &LevelMetrics{}
1313 1 : c.metrics[level] = levelMetrics
1314 1 : }
1315 1 : levelMetrics.BytesIngested += file.Size
1316 1 : levelMetrics.TablesIngested++
1317 : }
1318 1 : if ingestFlushable.exciseSpan.Valid() {
1319 1 : // Iterate through all levels and find files that intersect with exciseSpan.
1320 1 : for l := range c.version.Levels {
1321 1 : overlaps := c.version.Overlaps(l, base.UserKeyBoundsEndExclusive(ingestFlushable.exciseSpan.Start, ingestFlushable.exciseSpan.End))
1322 1 : iter := overlaps.Iter()
1323 1 :
1324 1 : for m := iter.First(); m != nil; m = iter.Next() {
1325 1 : newFiles, err := d.excise(ingestFlushable.exciseSpan.UserKeyBounds(), m, ve, l)
1326 1 : if err != nil {
1327 0 : return nil, err
1328 0 : }
1329 :
1330 1 : if _, ok := ve.DeletedFiles[deletedFileEntry{
1331 1 : Level: l,
1332 1 : FileNum: m.FileNum,
1333 1 : }]; !ok {
1334 1 : // We did not excise this file.
1335 1 : continue
1336 : }
1337 1 : replacedFiles[m.FileNum] = newFiles
1338 1 : updateLevelMetricsOnExcise(m, l, newFiles)
1339 : }
1340 : }
1341 : }
1342 :
1343 1 : if len(ingestSplitFiles) > 0 {
1344 0 : if err := d.ingestSplit(ve, updateLevelMetricsOnExcise, ingestSplitFiles, replacedFiles); err != nil {
1345 0 : return nil, err
1346 0 : }
1347 : }
1348 :
1349 1 : return ve, nil
1350 : }
1351 :
1352 : // flush runs a compaction that copies the immutable memtables from memory to
1353 : // disk.
1354 : //
1355 : // d.mu must be held when calling this, but the mutex may be dropped and
1356 : // re-acquired during the course of this method.
1357 1 : func (d *DB) flush1() (bytesFlushed uint64, err error) {
1358 1 : // NB: The flushable queue can contain flushables of type ingestedFlushable.
1359 1 : // The sstables in ingestedFlushable.files must be placed into the appropriate
1360 1 : // level in the lsm. Let's say the flushable queue contains a prefix of
1361 1 : // regular immutable memtables, then an ingestedFlushable, and then the
1362 1 : // mutable memtable. When the flush of the ingestedFlushable is performed,
1363 1 : // it needs an updated view of the lsm. That is, the prefix of immutable
1364 1 : // memtables must have already been flushed. Similarly, if there are two
1365 1 : // contiguous ingestedFlushables in the queue, then the first flushable must
1366 1 : // be flushed, so that the second flushable can see an updated view of the
1367 1 : // lsm.
1368 1 : //
1369 1 : // Given the above, we restrict flushes to either some prefix of regular
1370 1 : // memtables, or a single flushable of type ingestedFlushable. The DB.flush
1371 1 : // function will call DB.maybeScheduleFlush again, so a new flush to finish
1372 1 : // the remaining flush work should be scheduled right away.
1373 1 : //
1374 1 : // NB: Large batches placed in the flushable queue share the WAL with the
1375 1 : // previous memtable in the queue. We must ensure the property that both the
1376 1 : // large batch and the memtable with which it shares a WAL are flushed
1377 1 : // together. The property ensures that the minimum unflushed log number
1378 1 : // isn't incremented incorrectly. Since a flushableBatch.readyToFlush always
1379 1 : // returns true, and since the large batch will always be placed right after
1380 1 : // the memtable with which it shares a WAL, the property is naturally
1381 1 : // ensured. The large batch will always be placed after the memtable with
1382 1 : // which it shares a WAL because we ensure it in DB.commitWrite by holding
1383 1 : // the commitPipeline.mu and then holding DB.mu. As an extra defensive
1384 1 : // measure, if we try to flush the memtable without also flushing the
1385 1 : // flushable batch in the same flush, since the memtable and flushableBatch
1386 1 : // have the same logNum, the logNum invariant check below will trigger.
1387 1 : var n, inputs int
1388 1 : var inputBytes uint64
1389 1 : var ingest bool
1390 1 : for ; n < len(d.mu.mem.queue)-1; n++ {
1391 1 : if f, ok := d.mu.mem.queue[n].flushable.(*ingestedFlushable); ok {
1392 1 : if n == 0 {
1393 1 : // The first flushable is of type ingestedFlushable. Since these
1394 1 : // must be flushed individually, we perform a flush for just
1395 1 : // this.
1396 1 : if !f.readyForFlush() {
1397 0 : // This check is almost unnecessary, but we guard against it
1398 0 : // just in case this invariant changes in the future.
1399 0 : panic("pebble: ingestedFlushable should always be ready to flush.")
1400 : }
1401 : // By setting n = 1, we ensure that the first flushable(n == 0)
1402 : // is scheduled for a flush. The number of tables added is equal to the
1403 : // number of files in the ingest operation.
1404 1 : n = 1
1405 1 : inputs = len(f.files)
1406 1 : ingest = true
1407 1 : break
1408 1 : } else {
1409 1 : // There was some prefix of flushables which weren't of type
1410 1 : // ingestedFlushable. So, perform a flush for those.
1411 1 : break
1412 : }
1413 : }
1414 1 : if !d.mu.mem.queue[n].readyForFlush() {
1415 1 : break
1416 : }
1417 1 : inputBytes += d.mu.mem.queue[n].inuseBytes()
1418 : }
1419 1 : if n == 0 {
1420 0 : // None of the immutable memtables are ready for flushing.
1421 0 : return 0, nil
1422 0 : }
1423 1 : if !ingest {
1424 1 : // Flushes of memtables add the prefix of n memtables from the flushable
1425 1 : // queue.
1426 1 : inputs = n
1427 1 : }
1428 :
1429 : // Require that every memtable being flushed has a log number less than the
1430 : // new minimum unflushed log number.
1431 1 : minUnflushedLogNum := d.mu.mem.queue[n].logNum
1432 1 : if !d.opts.DisableWAL {
1433 1 : for i := 0; i < n; i++ {
1434 1 : if logNum := d.mu.mem.queue[i].logNum; logNum >= minUnflushedLogNum {
1435 0 : panic(errors.AssertionFailedf("logNum invariant violated: flushing %d items; %d:type=%T,logNum=%d; %d:type=%T,logNum=%d",
1436 0 : n,
1437 0 : i, d.mu.mem.queue[i].flushable, logNum,
1438 0 : n, d.mu.mem.queue[n].flushable, minUnflushedLogNum))
1439 : }
1440 : }
1441 : }
1442 :
1443 1 : c, err := newFlush(d.opts, d.mu.versions.currentVersion(),
1444 1 : d.mu.versions.picker.getBaseLevel(), d.mu.mem.queue[:n], d.timeNow())
1445 1 : if err != nil {
1446 0 : return 0, err
1447 0 : }
1448 1 : d.addInProgressCompaction(c)
1449 1 :
1450 1 : jobID := d.newJobIDLocked()
1451 1 : d.opts.EventListener.FlushBegin(FlushInfo{
1452 1 : JobID: int(jobID),
1453 1 : Input: inputs,
1454 1 : InputBytes: inputBytes,
1455 1 : Ingest: ingest,
1456 1 : })
1457 1 : startTime := d.timeNow()
1458 1 :
1459 1 : var ve *manifest.VersionEdit
1460 1 : var pendingOutputs []compactionOutput
1461 1 : var stats compactStats
1462 1 : // To determine the target level of the files in the ingestedFlushable, we
1463 1 : // need to acquire the logLock, and not release it for that duration. Since,
1464 1 : // we need to acquire the logLock below to perform the logAndApply step
1465 1 : // anyway, we create the VersionEdit for ingestedFlushable outside of
1466 1 : // runCompaction. For all other flush cases, we construct the VersionEdit
1467 1 : // inside runCompaction.
1468 1 : if c.kind != compactionKindIngestedFlushable {
1469 1 : ve, pendingOutputs, stats, err = d.runCompaction(jobID, c)
1470 1 : }
1471 :
1472 : // Acquire logLock. This will be released either on an error, by way of
1473 : // logUnlock, or through a call to logAndApply if there is no error.
1474 1 : d.mu.versions.logLock()
1475 1 :
1476 1 : if c.kind == compactionKindIngestedFlushable {
1477 1 : ve, err = d.runIngestFlush(c)
1478 1 : }
1479 :
1480 1 : info := FlushInfo{
1481 1 : JobID: int(jobID),
1482 1 : Input: inputs,
1483 1 : InputBytes: inputBytes,
1484 1 : Duration: d.timeNow().Sub(startTime),
1485 1 : Done: true,
1486 1 : Ingest: ingest,
1487 1 : Err: err,
1488 1 : }
1489 1 : if err == nil {
1490 1 : validateVersionEdit(ve, d.opts.Experimental.KeyValidationFunc, d.opts.Comparer.FormatKey, d.opts.Logger)
1491 1 : for i := range ve.NewFiles {
1492 1 : e := &ve.NewFiles[i]
1493 1 : info.Output = append(info.Output, e.Meta.TableInfo())
1494 1 : // Ingested tables are not necessarily flushed to L0. Record the level of
1495 1 : // each ingested file explicitly.
1496 1 : if ingest {
1497 1 : info.IngestLevels = append(info.IngestLevels, e.Level)
1498 1 : }
1499 : }
1500 1 : if len(ve.NewFiles) == 0 {
1501 1 : info.Err = errEmptyTable
1502 1 : }
1503 :
1504 : // The flush succeeded or it produced an empty sstable. In either case we
1505 : // want to bump the minimum unflushed log number to the log number of the
1506 : // oldest unflushed memtable.
1507 1 : ve.MinUnflushedLogNum = minUnflushedLogNum
1508 1 : if c.kind != compactionKindIngestedFlushable {
1509 1 : metrics := c.metrics[0]
1510 1 : if d.opts.DisableWAL {
1511 1 : // If the WAL is disabled, every flushable has a zero [logSize],
1512 1 : // resulting in zero bytes in. Instead, use the number of bytes we
1513 1 : // flushed as the BytesIn. This ensures we get a reasonable w-amp
1514 1 : // calculation even when the WAL is disabled.
1515 1 : metrics.BytesIn = metrics.BytesFlushed
1516 1 : } else {
1517 1 : metrics := c.metrics[0]
1518 1 : for i := 0; i < n; i++ {
1519 1 : metrics.BytesIn += d.mu.mem.queue[i].logSize
1520 1 : }
1521 : }
1522 1 : } else {
1523 1 : // c.kind == compactionKindIngestedFlushable && we could have deleted files due
1524 1 : // to ingest-time splits or excises.
1525 1 : ingestFlushable := c.flushing[0].flushable.(*ingestedFlushable)
1526 1 : for c2 := range d.mu.compact.inProgress {
1527 1 : // Check if this compaction overlaps with the excise span. Note that just
1528 1 : // checking if the inputs individually overlap with the excise span
1529 1 : // isn't sufficient; for instance, a compaction could have [a,b] and [e,f]
1530 1 : // as inputs and write it all out as [a,b,e,f] in one sstable. If we're
1531 1 : // doing a [c,d) excise at the same time as this compaction, we will have
1532 1 : // to error out the whole compaction as we can't guarantee it hasn't/won't
1533 1 : // write a file overlapping with the excise span.
1534 1 : if ingestFlushable.exciseSpan.OverlapsInternalKeyRange(d.cmp, c2.smallest, c2.largest) {
1535 1 : c2.cancel.Store(true)
1536 1 : continue
1537 : }
1538 : }
1539 :
1540 1 : if len(ve.DeletedFiles) > 0 {
1541 1 : // Iterate through all other compactions, and check if their inputs have
1542 1 : // been replaced due to an ingest-time split or excise. In that case,
1543 1 : // cancel the compaction.
1544 1 : for c2 := range d.mu.compact.inProgress {
1545 1 : for i := range c2.inputs {
1546 1 : iter := c2.inputs[i].files.Iter()
1547 1 : for f := iter.First(); f != nil; f = iter.Next() {
1548 1 : if _, ok := ve.DeletedFiles[deletedFileEntry{FileNum: f.FileNum, Level: c2.inputs[i].level}]; ok {
1549 1 : c2.cancel.Store(true)
1550 1 : break
1551 : }
1552 : }
1553 : }
1554 : }
1555 : }
1556 : }
1557 1 : err = d.mu.versions.logAndApply(jobID, ve, c.metrics, false, /* forceRotation */
1558 1 : func() []compactionInfo { return d.getInProgressCompactionInfoLocked(c) })
1559 1 : if err != nil {
1560 1 : info.Err = err
1561 1 : // TODO(peter): untested.
1562 1 : for _, f := range pendingOutputs {
1563 1 : // Note that the FileBacking for the file metadata might not have
1564 1 : // been set yet. So, we directly use the FileNum. Since these
1565 1 : // files were generated as compaction outputs, these must be
1566 1 : // physical files on disk. This property might not hold once
1567 1 : // https://github.com/cockroachdb/pebble/issues/389 is
1568 1 : // implemented if #389 creates virtual sstables as output files.
1569 1 : d.mu.versions.obsoleteTables = append(d.mu.versions.obsoleteTables, tableInfo{
1570 1 : fileInfo: fileInfo{
1571 1 : FileNum: base.PhysicalTableDiskFileNum(f.meta.FileNum),
1572 1 : FileSize: f.meta.Size,
1573 1 : },
1574 1 : isLocal: f.isLocal,
1575 1 : })
1576 1 : }
1577 1 : d.mu.versions.updateObsoleteTableMetricsLocked()
1578 : }
1579 1 : } else {
1580 1 : // We won't be performing the logAndApply step because of the error,
1581 1 : // so logUnlock.
1582 1 : d.mu.versions.logUnlock()
1583 1 : }
1584 :
1585 : // If err != nil, then the flush will be retried, and we will recalculate
1586 : // these metrics.
1587 1 : if err == nil {
1588 1 : d.mu.snapshots.cumulativePinnedCount += stats.cumulativePinnedKeys
1589 1 : d.mu.snapshots.cumulativePinnedSize += stats.cumulativePinnedSize
1590 1 : d.mu.versions.metrics.Keys.MissizedTombstonesCount += stats.countMissizedDels
1591 1 : d.maybeUpdateDeleteCompactionHints(c)
1592 1 : }
1593 :
1594 1 : d.clearCompactingState(c, err != nil)
1595 1 : delete(d.mu.compact.inProgress, c)
1596 1 : d.mu.versions.incrementCompactions(c.kind, c.extraLevels, c.pickerMetrics)
1597 1 :
1598 1 : var flushed flushableList
1599 1 : if err == nil {
1600 1 : flushed = d.mu.mem.queue[:n]
1601 1 : d.mu.mem.queue = d.mu.mem.queue[n:]
1602 1 : d.updateReadStateLocked(d.opts.DebugCheck)
1603 1 : d.updateTableStatsLocked(ve.NewFiles)
1604 1 : if ingest {
1605 1 : d.mu.versions.metrics.Flush.AsIngestCount++
1606 1 : for _, l := range c.metrics {
1607 1 : d.mu.versions.metrics.Flush.AsIngestBytes += l.BytesIngested
1608 1 : d.mu.versions.metrics.Flush.AsIngestTableCount += l.TablesIngested
1609 1 : }
1610 : }
1611 1 : d.maybeTransitionSnapshotsToFileOnlyLocked()
1612 :
1613 : }
1614 : // Signal FlushEnd after installing the new readState. This helps for unit
1615 : // tests that use the callback to trigger a read using an iterator with
1616 : // IterOptions.OnlyReadGuaranteedDurable.
1617 1 : info.TotalDuration = d.timeNow().Sub(startTime)
1618 1 : d.opts.EventListener.FlushEnd(info)
1619 1 :
1620 1 : // The order of these operations matters here for ease of testing.
1621 1 : // Removing the reader reference first allows tests to be guaranteed that
1622 1 : // the memtable reservation has been released by the time a synchronous
1623 1 : // flush returns. readerUnrefLocked may also produce obsolete files so the
1624 1 : // call to deleteObsoleteFiles must happen after it.
1625 1 : for i := range flushed {
1626 1 : flushed[i].readerUnrefLocked(true)
1627 1 : }
1628 :
1629 1 : d.deleteObsoleteFiles(jobID)
1630 1 :
1631 1 : // Mark all the memtables we flushed as flushed.
1632 1 : for i := range flushed {
1633 1 : close(flushed[i].flushed)
1634 1 : }
1635 :
1636 1 : return inputBytes, err
1637 : }
1638 :
1639 : // maybeTransitionSnapshotsToFileOnlyLocked transitions any "eventually
1640 : // file-only" snapshots to be file-only if all their visible state has been
1641 : // flushed to sstables.
1642 : //
1643 : // REQUIRES: d.mu.
1644 1 : func (d *DB) maybeTransitionSnapshotsToFileOnlyLocked() {
1645 1 : earliestUnflushedSeqNum := d.getEarliestUnflushedSeqNumLocked()
1646 1 : currentVersion := d.mu.versions.currentVersion()
1647 1 : for s := d.mu.snapshots.root.next; s != &d.mu.snapshots.root; {
1648 1 : if s.efos == nil {
1649 1 : s = s.next
1650 1 : continue
1651 : }
1652 1 : overlapsFlushable := false
1653 1 : if base.Visible(earliestUnflushedSeqNum, s.efos.seqNum, InternalKeySeqNumMax) {
1654 1 : // There are some unflushed keys that are still visible to the EFOS.
1655 1 : // Check if any memtables older than the EFOS contain keys within a
1656 1 : // protected range of the EFOS. If no, we can transition.
1657 1 : protectedRanges := make([]bounded, len(s.efos.protectedRanges))
1658 1 : for i := range s.efos.protectedRanges {
1659 1 : protectedRanges[i] = s.efos.protectedRanges[i]
1660 1 : }
1661 1 : for i := range d.mu.mem.queue {
1662 1 : if !base.Visible(d.mu.mem.queue[i].logSeqNum, s.efos.seqNum, InternalKeySeqNumMax) {
1663 0 : // All keys in this memtable are newer than the EFOS. Skip this
1664 0 : // memtable.
1665 0 : continue
1666 : }
1667 : // NB: computePossibleOverlaps could have false positives, such as if
1668 : // the flushable is a flushable ingest and not a memtable. In that
1669 : // case we don't open the sstables to check; we just pessimistically
1670 : // assume an overlap.
1671 1 : d.mu.mem.queue[i].computePossibleOverlaps(func(b bounded) shouldContinue {
1672 1 : overlapsFlushable = true
1673 1 : return stopIteration
1674 1 : }, protectedRanges...)
1675 1 : if overlapsFlushable {
1676 1 : break
1677 : }
1678 : }
1679 : }
1680 1 : if overlapsFlushable {
1681 1 : s = s.next
1682 1 : continue
1683 : }
1684 1 : currentVersion.Ref()
1685 1 :
1686 1 : // NB: s.efos.transitionToFileOnlySnapshot could close s, in which
1687 1 : // case s.next would be nil. Save it before calling it.
1688 1 : next := s.next
1689 1 : _ = s.efos.transitionToFileOnlySnapshot(currentVersion)
1690 1 : s = next
1691 : }
1692 : }
1693 :
1694 : // maybeScheduleCompactionAsync should be used when
1695 : // we want to possibly schedule a compaction, but don't
1696 : // want to eat the cost of running maybeScheduleCompaction.
1697 : // This method should be launched in a separate goroutine.
1698 : // d.mu must not be held when this is called.
1699 0 : func (d *DB) maybeScheduleCompactionAsync() {
1700 0 : defer d.compactionSchedulers.Done()
1701 0 :
1702 0 : d.mu.Lock()
1703 0 : d.maybeScheduleCompaction()
1704 0 : d.mu.Unlock()
1705 0 : }
1706 :
1707 : // maybeScheduleCompaction schedules a compaction if necessary.
1708 : //
1709 : // d.mu must be held when calling this.
1710 1 : func (d *DB) maybeScheduleCompaction() {
1711 1 : d.maybeScheduleCompactionPicker(pickAuto)
1712 1 : }
1713 :
1714 1 : func pickAuto(picker compactionPicker, env compactionEnv) *pickedCompaction {
1715 1 : return picker.pickAuto(env)
1716 1 : }
1717 :
1718 1 : func pickElisionOnly(picker compactionPicker, env compactionEnv) *pickedCompaction {
1719 1 : return picker.pickElisionOnlyCompaction(env)
1720 1 : }
1721 :
1722 : // tryScheduleDownloadCompaction tries to start a download compaction.
1723 : //
1724 : // Returns true if we started a download compaction (or completed it
1725 : // immediately because it is a no-op or we hit an error).
1726 : //
1727 : // Requires d.mu to be held. Updates d.mu.compact.downloads.
1728 1 : func (d *DB) tryScheduleDownloadCompaction(env compactionEnv, maxConcurrentDownloads int) bool {
1729 1 : vers := d.mu.versions.currentVersion()
1730 1 : for i := 0; i < len(d.mu.compact.downloads); {
1731 1 : download := d.mu.compact.downloads[i]
1732 1 : switch d.tryLaunchDownloadCompaction(download, vers, env, maxConcurrentDownloads) {
1733 1 : case launchedCompaction:
1734 1 : return true
1735 0 : case didNotLaunchCompaction:
1736 0 : // See if we can launch a compaction for another download task.
1737 0 : i++
1738 1 : case downloadTaskCompleted:
1739 1 : // Task is completed and must be removed.
1740 1 : d.mu.compact.downloads = slices.Delete(d.mu.compact.downloads, i, i+1)
1741 : }
1742 : }
1743 1 : return false
1744 : }
1745 :
1746 : // maybeScheduleCompactionPicker schedules a compaction if necessary,
1747 : // calling `pickFunc` to pick automatic compactions.
1748 : //
1749 : // Requires d.mu to be held.
1750 : func (d *DB) maybeScheduleCompactionPicker(
1751 : pickFunc func(compactionPicker, compactionEnv) *pickedCompaction,
1752 1 : ) {
1753 1 : if d.closed.Load() != nil || d.opts.ReadOnly {
1754 1 : return
1755 1 : }
1756 1 : maxCompactions := d.opts.MaxConcurrentCompactions()
1757 1 : maxDownloads := d.opts.MaxConcurrentDownloads()
1758 1 :
1759 1 : if d.mu.compact.compactingCount >= maxCompactions &&
1760 1 : (len(d.mu.compact.downloads) == 0 || d.mu.compact.downloadingCount >= maxDownloads) {
1761 1 : if len(d.mu.compact.manual) > 0 {
1762 1 : // Inability to run head blocks later manual compactions.
1763 1 : d.mu.compact.manual[0].retries++
1764 1 : }
1765 1 : return
1766 : }
1767 :
1768 : // Compaction picking needs a coherent view of a Version. In particular, we
1769 : // need to exclude concurrent ingestions from making a decision on which level
1770 : // to ingest into that conflicts with our compaction
1771 : // decision. versionSet.logLock provides the necessary mutual exclusion.
1772 1 : d.mu.versions.logLock()
1773 1 : defer d.mu.versions.logUnlock()
1774 1 :
1775 1 : // Check for the closed flag again, in case the DB was closed while we were
1776 1 : // waiting for logLock().
1777 1 : if d.closed.Load() != nil {
1778 0 : return
1779 0 : }
1780 :
1781 1 : env := compactionEnv{
1782 1 : diskAvailBytes: d.diskAvailBytes.Load(),
1783 1 : earliestSnapshotSeqNum: d.mu.snapshots.earliest(),
1784 1 : earliestUnflushedSeqNum: d.getEarliestUnflushedSeqNumLocked(),
1785 1 : }
1786 1 :
1787 1 : if d.mu.compact.compactingCount < maxCompactions {
1788 1 : // Check for delete-only compactions first, because they're expected to be
1789 1 : // cheap and reduce future compaction work.
1790 1 : if !d.opts.private.disableDeleteOnlyCompactions &&
1791 1 : !d.opts.DisableAutomaticCompactions &&
1792 1 : len(d.mu.compact.deletionHints) > 0 {
1793 1 : d.tryScheduleDeleteOnlyCompaction()
1794 1 : }
1795 :
1796 1 : for len(d.mu.compact.manual) > 0 && d.mu.compact.compactingCount < maxCompactions {
1797 1 : if manual := d.mu.compact.manual[0]; !d.tryScheduleManualCompaction(env, manual) {
1798 1 : // Inability to run head blocks later manual compactions.
1799 1 : manual.retries++
1800 1 : break
1801 : }
1802 1 : d.mu.compact.manual = d.mu.compact.manual[1:]
1803 : }
1804 :
1805 1 : for !d.opts.DisableAutomaticCompactions && d.mu.compact.compactingCount < maxCompactions &&
1806 1 : d.tryScheduleAutoCompaction(env, pickFunc) {
1807 1 : }
1808 : }
1809 :
1810 1 : for len(d.mu.compact.downloads) > 0 && d.mu.compact.downloadingCount < maxDownloads &&
1811 1 : d.tryScheduleDownloadCompaction(env, maxDownloads) {
1812 1 : }
1813 : }
1814 :
1815 : // tryScheduleDeleteOnlyCompaction tries to kick off a delete-only compaction
1816 : // for all files that can be deleted as suggested by deletionHints.
1817 : //
1818 : // Requires d.mu to be held. Updates d.mu.compact.deletionHints.
1819 1 : func (d *DB) tryScheduleDeleteOnlyCompaction() {
1820 1 : v := d.mu.versions.currentVersion()
1821 1 : snapshots := d.mu.snapshots.toSlice()
1822 1 : inputs, unresolvedHints := checkDeleteCompactionHints(d.cmp, v, d.mu.compact.deletionHints, snapshots)
1823 1 : d.mu.compact.deletionHints = unresolvedHints
1824 1 :
1825 1 : if len(inputs) > 0 {
1826 1 : c := newDeleteOnlyCompaction(d.opts, v, inputs, d.timeNow())
1827 1 : d.mu.compact.compactingCount++
1828 1 : d.addInProgressCompaction(c)
1829 1 : go d.compact(c, nil)
1830 1 : }
1831 : }
1832 :
1833 : // tryScheduleManualCompaction tries to kick off the given manual compaction.
1834 : //
1835 : // Returns false if we are not able to run this compaction at this time.
1836 : //
1837 : // Requires d.mu to be held.
1838 1 : func (d *DB) tryScheduleManualCompaction(env compactionEnv, manual *manualCompaction) bool {
1839 1 : v := d.mu.versions.currentVersion()
1840 1 : env.inProgressCompactions = d.getInProgressCompactionInfoLocked(nil)
1841 1 : pc, retryLater := pickManualCompaction(v, d.opts, env, d.mu.versions.picker.getBaseLevel(), manual)
1842 1 : if pc == nil {
1843 1 : if !retryLater {
1844 1 : // Manual compaction is a no-op. Signal completion and exit.
1845 1 : manual.done <- nil
1846 1 : return true
1847 1 : }
1848 : // We are not able to run this manual compaction at this time.
1849 1 : return false
1850 : }
1851 :
1852 1 : c := newCompaction(pc, d.opts, d.timeNow(), d.ObjProvider())
1853 1 : d.mu.compact.compactingCount++
1854 1 : d.addInProgressCompaction(c)
1855 1 : go d.compact(c, manual.done)
1856 1 : return true
1857 : }
1858 :
1859 : // tryScheduleAutoCompaction tries to kick off an automatic compaction.
1860 : //
1861 : // Returns false if no automatic compactions are necessary or able to run at
1862 : // this time.
1863 : //
1864 : // Requires d.mu to be held.
1865 : func (d *DB) tryScheduleAutoCompaction(
1866 : env compactionEnv, pickFunc func(compactionPicker, compactionEnv) *pickedCompaction,
1867 1 : ) bool {
1868 1 : env.inProgressCompactions = d.getInProgressCompactionInfoLocked(nil)
1869 1 : env.readCompactionEnv = readCompactionEnv{
1870 1 : readCompactions: &d.mu.compact.readCompactions,
1871 1 : flushing: d.mu.compact.flushing || d.passedFlushThreshold(),
1872 1 : rescheduleReadCompaction: &d.mu.compact.rescheduleReadCompaction,
1873 1 : }
1874 1 : pc := pickFunc(d.mu.versions.picker, env)
1875 1 : if pc == nil {
1876 1 : return false
1877 1 : }
1878 1 : c := newCompaction(pc, d.opts, d.timeNow(), d.ObjProvider())
1879 1 : d.mu.compact.compactingCount++
1880 1 : d.addInProgressCompaction(c)
1881 1 : go d.compact(c, nil)
1882 1 : return true
1883 : }
1884 :
1885 : // deleteCompactionHintType indicates whether the deleteCompactionHint was
1886 : // generated from a span containing a range del (point key only), a range key
1887 : // delete (range key only), or both a point and range key.
1888 : type deleteCompactionHintType uint8
1889 :
1890 : const (
1891 : // NOTE: While these are primarily used as enumeration types, they are also
1892 : // used for some bitwise operations. Care should be taken when updating.
1893 : deleteCompactionHintTypeUnknown deleteCompactionHintType = iota
1894 : deleteCompactionHintTypePointKeyOnly
1895 : deleteCompactionHintTypeRangeKeyOnly
1896 : deleteCompactionHintTypePointAndRangeKey
1897 : )
1898 :
1899 : // String implements fmt.Stringer.
1900 1 : func (h deleteCompactionHintType) String() string {
1901 1 : switch h {
1902 0 : case deleteCompactionHintTypeUnknown:
1903 0 : return "unknown"
1904 1 : case deleteCompactionHintTypePointKeyOnly:
1905 1 : return "point-key-only"
1906 1 : case deleteCompactionHintTypeRangeKeyOnly:
1907 1 : return "range-key-only"
1908 1 : case deleteCompactionHintTypePointAndRangeKey:
1909 1 : return "point-and-range-key"
1910 0 : default:
1911 0 : panic(fmt.Sprintf("unknown hint type: %d", h))
1912 : }
1913 : }
1914 :
1915 : // compactionHintFromKeys returns a deleteCompactionHintType given a slice of
1916 : // keyspan.Keys.
1917 1 : func compactionHintFromKeys(keys []keyspan.Key) deleteCompactionHintType {
1918 1 : var hintType deleteCompactionHintType
1919 1 : for _, k := range keys {
1920 1 : switch k.Kind() {
1921 1 : case base.InternalKeyKindRangeDelete:
1922 1 : hintType |= deleteCompactionHintTypePointKeyOnly
1923 1 : case base.InternalKeyKindRangeKeyDelete:
1924 1 : hintType |= deleteCompactionHintTypeRangeKeyOnly
1925 0 : default:
1926 0 : panic(fmt.Sprintf("unsupported key kind: %s", k.Kind()))
1927 : }
1928 : }
1929 1 : return hintType
1930 : }
1931 :
1932 : // A deleteCompactionHint records a user key and sequence number span that has been
1933 : // deleted by a range tombstone. A hint is recorded if at least one sstable
1934 : // falls completely within both the user key and sequence number spans.
1935 : // Once the tombstones and the observed completely-contained sstables fall
1936 : // into the same snapshot stripe, a delete-only compaction may delete any
1937 : // sstables within the range.
1938 : type deleteCompactionHint struct {
1939 : // The type of key span that generated this hint (point key, range key, or
1940 : // both).
1941 : hintType deleteCompactionHintType
1942 : // start and end are user keys specifying a key range [start, end) of
1943 : // deleted keys.
1944 : start []byte
1945 : end []byte
1946 : // The level of the file containing the range tombstone(s) when the hint
1947 : // was created. Only lower levels need to be searched for files that may
1948 : // be deleted.
1949 : tombstoneLevel int
1950 : // The file containing the range tombstone(s) that created the hint.
1951 : tombstoneFile *fileMetadata
1952 : // The smallest and largest sequence numbers of the abutting tombstones
1953 : // merged to form this hint. All of a tables' keys must be less than the
1954 : // tombstone smallest sequence number to be deleted. All of a tables'
1955 : // sequence numbers must fall into the same snapshot stripe as the
1956 : // tombstone largest sequence number to be deleted.
1957 : tombstoneLargestSeqNum uint64
1958 : tombstoneSmallestSeqNum uint64
1959 : // The smallest sequence number of a sstable that was found to be covered
1960 : // by this hint. The hint cannot be resolved until this sequence number is
1961 : // in the same snapshot stripe as the largest tombstone sequence number.
1962 : // This is set when a hint is created, so the LSM may look different and
1963 : // notably no longer contain the sstable that contained the key at this
1964 : // sequence number.
1965 : fileSmallestSeqNum uint64
1966 : }
1967 :
1968 1 : func (h deleteCompactionHint) String() string {
1969 1 : return fmt.Sprintf(
1970 1 : "L%d.%s %s-%s seqnums(tombstone=%d-%d, file-smallest=%d, type=%s)",
1971 1 : h.tombstoneLevel, h.tombstoneFile.FileNum, h.start, h.end,
1972 1 : h.tombstoneSmallestSeqNum, h.tombstoneLargestSeqNum, h.fileSmallestSeqNum,
1973 1 : h.hintType,
1974 1 : )
1975 1 : }
1976 :
1977 : func (h *deleteCompactionHint) canDelete(
1978 : cmp Compare, m *fileMetadata, snapshots compact.Snapshots,
1979 1 : ) bool {
1980 1 : // The file can only be deleted if all of its keys are older than the
1981 1 : // earliest tombstone aggregated into the hint.
1982 1 : if m.LargestSeqNum >= h.tombstoneSmallestSeqNum || m.SmallestSeqNum < h.fileSmallestSeqNum {
1983 1 : return false
1984 1 : }
1985 :
1986 : // The file's oldest key must be in the same snapshot stripe as the
1987 : // newest tombstone. NB: We already checked the hint's sequence numbers,
1988 : // but this file's oldest sequence number might be lower than the hint's
1989 : // smallest sequence number despite the file falling within the key range
1990 : // if this file was constructed after the hint by a compaction.
1991 1 : if snapshots.Index(h.tombstoneLargestSeqNum) != snapshots.Index(m.SmallestSeqNum) {
1992 0 : return false
1993 0 : }
1994 :
1995 1 : switch h.hintType {
1996 1 : case deleteCompactionHintTypePointKeyOnly:
1997 1 : // A hint generated by a range del span cannot delete tables that contain
1998 1 : // range keys.
1999 1 : if m.HasRangeKeys {
2000 0 : return false
2001 0 : }
2002 1 : case deleteCompactionHintTypeRangeKeyOnly:
2003 1 : // A hint generated by a range key del span cannot delete tables that
2004 1 : // contain point keys.
2005 1 : if m.HasPointKeys {
2006 1 : return false
2007 1 : }
2008 1 : case deleteCompactionHintTypePointAndRangeKey:
2009 : // A hint from a span that contains both range dels *and* range keys can
2010 : // only be deleted if both bounds fall within the hint. The next check takes
2011 : // care of this.
2012 0 : default:
2013 0 : panic(fmt.Sprintf("pebble: unknown delete compaction hint type: %d", h.hintType))
2014 : }
2015 :
2016 : // The file's keys must be completely contained within the hint range.
2017 1 : return cmp(h.start, m.Smallest.UserKey) <= 0 && cmp(m.Largest.UserKey, h.end) < 0
2018 : }
2019 :
2020 1 : func (d *DB) maybeUpdateDeleteCompactionHints(c *compaction) {
2021 1 : // Compactions that zero sequence numbers can interfere with compaction
2022 1 : // deletion hints. Deletion hints apply to tables containing keys older
2023 1 : // than a threshold. If a key more recent than the threshold is zeroed in
2024 1 : // a compaction, a delete-only compaction may mistake it as meeting the
2025 1 : // threshold and drop a table containing live data.
2026 1 : //
2027 1 : // To avoid this scenario, compactions that zero sequence numbers remove
2028 1 : // any conflicting deletion hints. A deletion hint is conflicting if both
2029 1 : // of the following conditions apply:
2030 1 : // * its key space overlaps with the compaction
2031 1 : // * at least one of its inputs contains a key as recent as one of the
2032 1 : // hint's tombstones.
2033 1 : //
2034 1 : if !c.allowedZeroSeqNum {
2035 1 : return
2036 1 : }
2037 :
2038 1 : updatedHints := d.mu.compact.deletionHints[:0]
2039 1 : for _, h := range d.mu.compact.deletionHints {
2040 1 : // If the compaction's key space is disjoint from the hint's key
2041 1 : // space, the zeroing of sequence numbers won't affect the hint. Keep
2042 1 : // the hint.
2043 1 : keysDisjoint := d.cmp(h.end, c.smallest.UserKey) < 0 || d.cmp(h.start, c.largest.UserKey) > 0
2044 1 : if keysDisjoint {
2045 0 : updatedHints = append(updatedHints, h)
2046 0 : continue
2047 : }
2048 :
2049 : // All of the compaction's inputs must be older than the hint's
2050 : // tombstones.
2051 1 : inputsOlder := true
2052 1 : for _, in := range c.inputs {
2053 1 : iter := in.files.Iter()
2054 1 : for f := iter.First(); f != nil; f = iter.Next() {
2055 1 : inputsOlder = inputsOlder && f.LargestSeqNum < h.tombstoneSmallestSeqNum
2056 1 : }
2057 : }
2058 1 : if inputsOlder {
2059 0 : updatedHints = append(updatedHints, h)
2060 0 : continue
2061 : }
2062 :
2063 : // Drop h, because the compaction c may have zeroed sequence numbers
2064 : // of keys more recent than some of h's tombstones.
2065 : }
2066 1 : d.mu.compact.deletionHints = updatedHints
2067 : }
2068 :
2069 : func checkDeleteCompactionHints(
2070 : cmp Compare, v *version, hints []deleteCompactionHint, snapshots compact.Snapshots,
2071 1 : ) ([]compactionLevel, []deleteCompactionHint) {
2072 1 : var files map[*fileMetadata]bool
2073 1 : var byLevel [numLevels][]*fileMetadata
2074 1 :
2075 1 : unresolvedHints := hints[:0]
2076 1 : for _, h := range hints {
2077 1 : // Check each compaction hint to see if it's resolvable. Resolvable
2078 1 : // hints are removed and trigger a delete-only compaction if any files
2079 1 : // in the current LSM still meet their criteria. Unresolvable hints
2080 1 : // are saved and don't trigger a delete-only compaction.
2081 1 : //
2082 1 : // When a compaction hint is created, the sequence numbers of the
2083 1 : // range tombstones and the covered file with the oldest key are
2084 1 : // recorded. The largest tombstone sequence number and the smallest
2085 1 : // file sequence number must be in the same snapshot stripe for the
2086 1 : // hint to be resolved. The below graphic models a compaction hint
2087 1 : // covering the keyspace [b, r). The hint completely contains two
2088 1 : // files, 000002 and 000003. The file 000003 contains the lowest
2089 1 : // covered sequence number at #90. The tombstone b.RANGEDEL.230:h has
2090 1 : // the highest tombstone sequence number incorporated into the hint.
2091 1 : // The hint may be resolved only once the snapshots at #100, #180 and
2092 1 : // #210 are all closed. File 000001 is not included within the hint
2093 1 : // because it extends beyond the range tombstones in user key space.
2094 1 : //
2095 1 : // 250
2096 1 : //
2097 1 : // |-b...230:h-|
2098 1 : // _____________________________________________________ snapshot #210
2099 1 : // 200 |--h.RANGEDEL.200:r--|
2100 1 : //
2101 1 : // _____________________________________________________ snapshot #180
2102 1 : //
2103 1 : // 150 +--------+
2104 1 : // +---------+ | 000003 |
2105 1 : // | 000002 | | |
2106 1 : // +_________+ | |
2107 1 : // 100_____________________|________|___________________ snapshot #100
2108 1 : // +--------+
2109 1 : // _____________________________________________________ snapshot #70
2110 1 : // +---------------+
2111 1 : // 50 | 000001 |
2112 1 : // | |
2113 1 : // +---------------+
2114 1 : // ______________________________________________________________
2115 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
2116 1 :
2117 1 : if snapshots.Index(h.tombstoneLargestSeqNum) != snapshots.Index(h.fileSmallestSeqNum) {
2118 1 : // Cannot resolve yet.
2119 1 : unresolvedHints = append(unresolvedHints, h)
2120 1 : continue
2121 : }
2122 :
2123 : // The hint h will be resolved and dropped, regardless of whether
2124 : // there are any tables that can be deleted.
2125 1 : for l := h.tombstoneLevel + 1; l < numLevels; l++ {
2126 1 : overlaps := v.Overlaps(l, base.UserKeyBoundsEndExclusive(h.start, h.end))
2127 1 : iter := overlaps.Iter()
2128 1 : for m := iter.First(); m != nil; m = iter.Next() {
2129 1 : if m.IsCompacting() || !h.canDelete(cmp, m, snapshots) || files[m] {
2130 1 : continue
2131 : }
2132 1 : if files == nil {
2133 1 : // Construct files lazily, assuming most calls will not
2134 1 : // produce delete-only compactions.
2135 1 : files = make(map[*fileMetadata]bool)
2136 1 : }
2137 1 : files[m] = true
2138 1 : byLevel[l] = append(byLevel[l], m)
2139 : }
2140 : }
2141 : }
2142 :
2143 1 : var compactLevels []compactionLevel
2144 1 : for l, files := range byLevel {
2145 1 : if len(files) == 0 {
2146 1 : continue
2147 : }
2148 1 : compactLevels = append(compactLevels, compactionLevel{
2149 1 : level: l,
2150 1 : files: manifest.NewLevelSliceKeySorted(cmp, files),
2151 1 : })
2152 : }
2153 1 : return compactLevels, unresolvedHints
2154 : }
2155 :
2156 : // compact runs one compaction and maybe schedules another call to compact.
2157 1 : func (d *DB) compact(c *compaction, errChannel chan error) {
2158 1 : pprof.Do(context.Background(), compactLabels, func(context.Context) {
2159 1 : d.mu.Lock()
2160 1 : defer d.mu.Unlock()
2161 1 : if err := d.compact1(c, errChannel); err != nil {
2162 1 : // TODO(peter): count consecutive compaction errors and backoff.
2163 1 : d.opts.EventListener.BackgroundError(err)
2164 1 : }
2165 1 : if c.isDownload {
2166 1 : d.mu.compact.downloadingCount--
2167 1 : } else {
2168 1 : d.mu.compact.compactingCount--
2169 1 : }
2170 1 : delete(d.mu.compact.inProgress, c)
2171 1 : // Add this compaction's duration to the cumulative duration. NB: This
2172 1 : // must be atomic with the above removal of c from
2173 1 : // d.mu.compact.InProgress to ensure Metrics.Compact.Duration does not
2174 1 : // miss or double count a completing compaction's duration.
2175 1 : d.mu.compact.duration += d.timeNow().Sub(c.beganAt)
2176 1 :
2177 1 : // The previous compaction may have produced too many files in a
2178 1 : // level, so reschedule another compaction if needed.
2179 1 : d.maybeScheduleCompaction()
2180 1 : d.mu.compact.cond.Broadcast()
2181 : })
2182 : }
2183 :
2184 : // compact1 runs one compaction.
2185 : //
2186 : // d.mu must be held when calling this, but the mutex may be dropped and
2187 : // re-acquired during the course of this method.
2188 1 : func (d *DB) compact1(c *compaction, errChannel chan error) (err error) {
2189 1 : if errChannel != nil {
2190 1 : defer func() {
2191 1 : errChannel <- err
2192 1 : }()
2193 : }
2194 :
2195 1 : jobID := d.newJobIDLocked()
2196 1 : info := c.makeInfo(jobID)
2197 1 : d.opts.EventListener.CompactionBegin(info)
2198 1 : startTime := d.timeNow()
2199 1 :
2200 1 : ve, pendingOutputs, stats, err := d.runCompaction(jobID, c)
2201 1 :
2202 1 : info.Duration = d.timeNow().Sub(startTime)
2203 1 : if err == nil {
2204 1 : validateVersionEdit(ve, d.opts.Experimental.KeyValidationFunc, d.opts.Comparer.FormatKey, d.opts.Logger)
2205 1 : err = func() error {
2206 1 : var err error
2207 1 : d.mu.versions.logLock()
2208 1 : // Check if this compaction had a conflicting operation (eg. a d.excise())
2209 1 : // that necessitates it restarting from scratch. Note that since we hold
2210 1 : // the manifest lock, we don't expect this bool to change its value
2211 1 : // as only the holder of the manifest lock will ever write to it.
2212 1 : if c.cancel.Load() {
2213 1 : err = firstError(err, ErrCancelledCompaction)
2214 1 : }
2215 1 : if err != nil {
2216 1 : // logAndApply calls logUnlock. If we didn't call it, we need to call
2217 1 : // logUnlock ourselves.
2218 1 : d.mu.versions.logUnlock()
2219 1 : return err
2220 1 : }
2221 1 : return d.mu.versions.logAndApply(jobID, ve, c.metrics, false /* forceRotation */, func() []compactionInfo {
2222 1 : return d.getInProgressCompactionInfoLocked(c)
2223 1 : })
2224 : }()
2225 1 : if err != nil {
2226 1 : // TODO(peter): untested.
2227 1 : for _, f := range pendingOutputs {
2228 1 : // Note that the FileBacking for the file metadata might not have
2229 1 : // been set yet. So, we directly use the FileNum. Since these
2230 1 : // files were generated as compaction outputs, these must be
2231 1 : // physical files on disk. This property might not hold once
2232 1 : // https://github.com/cockroachdb/pebble/issues/389 is
2233 1 : // implemented if #389 creates virtual sstables as output files.
2234 1 : d.mu.versions.obsoleteTables = append(d.mu.versions.obsoleteTables, tableInfo{
2235 1 : fileInfo: fileInfo{
2236 1 : FileNum: base.PhysicalTableDiskFileNum(f.meta.FileNum),
2237 1 : FileSize: f.meta.Size,
2238 1 : },
2239 1 : isLocal: f.isLocal,
2240 1 : })
2241 1 : }
2242 1 : d.mu.versions.updateObsoleteTableMetricsLocked()
2243 : }
2244 : }
2245 :
2246 1 : info.Done = true
2247 1 : info.Err = err
2248 1 : if err == nil {
2249 1 : for i := range ve.NewFiles {
2250 1 : e := &ve.NewFiles[i]
2251 1 : info.Output.Tables = append(info.Output.Tables, e.Meta.TableInfo())
2252 1 : }
2253 1 : d.mu.snapshots.cumulativePinnedCount += stats.cumulativePinnedKeys
2254 1 : d.mu.snapshots.cumulativePinnedSize += stats.cumulativePinnedSize
2255 1 : d.mu.versions.metrics.Keys.MissizedTombstonesCount += stats.countMissizedDels
2256 1 : d.maybeUpdateDeleteCompactionHints(c)
2257 : }
2258 :
2259 : // NB: clearing compacting state must occur before updating the read state;
2260 : // L0Sublevels initialization depends on it.
2261 1 : d.clearCompactingState(c, err != nil)
2262 1 : d.mu.versions.incrementCompactions(c.kind, c.extraLevels, c.pickerMetrics)
2263 1 : d.mu.versions.incrementCompactionBytes(-c.bytesWritten)
2264 1 :
2265 1 : info.TotalDuration = d.timeNow().Sub(c.beganAt)
2266 1 : d.opts.EventListener.CompactionEnd(info)
2267 1 :
2268 1 : // Update the read state before deleting obsolete files because the
2269 1 : // read-state update will cause the previous version to be unref'd and if
2270 1 : // there are no references obsolete tables will be added to the obsolete
2271 1 : // table list.
2272 1 : if err == nil {
2273 1 : d.updateReadStateLocked(d.opts.DebugCheck)
2274 1 : d.updateTableStatsLocked(ve.NewFiles)
2275 1 : }
2276 1 : d.deleteObsoleteFiles(jobID)
2277 1 :
2278 1 : return err
2279 : }
2280 :
2281 : type compactStats struct {
2282 : cumulativePinnedKeys uint64
2283 : cumulativePinnedSize uint64
2284 : countMissizedDels uint64
2285 : }
2286 :
2287 : // runCopyCompaction runs a copy compaction where a new FileNum is created that
2288 : // is a byte-for-byte copy of the input file or span thereof in some cases. This
2289 : // is used in lieu of a move compaction when a file is being moved across the
2290 : // local/remote storage boundary. It could also be used in lieu of a rewrite
2291 : // compaction as part of a Download() call, which allows copying only a span of
2292 : // the external file, provided the file does not contain range keys or value
2293 : // blocks (see sstable.CopySpan).
2294 : //
2295 : // d.mu must be held when calling this method. The mutex will be released when
2296 : // doing IO.
2297 : func (d *DB) runCopyCompaction(
2298 : jobID JobID,
2299 : c *compaction,
2300 : inputMeta *fileMetadata,
2301 : objMeta objstorage.ObjectMetadata,
2302 : ve *versionEdit,
2303 1 : ) (pendingOutputs []compactionOutput, retErr error) {
2304 1 : ctx := context.TODO()
2305 1 :
2306 1 : if !objMeta.IsExternal() {
2307 1 : if objMeta.IsRemote() || !remote.ShouldCreateShared(d.opts.Experimental.CreateOnShared, c.outputLevel.level) {
2308 0 : panic("pebble: scheduled a copy compaction that is not actually moving files to shared storage")
2309 : }
2310 : // Note that based on logic in the compaction picker, we're guaranteed
2311 : // inputMeta.Virtual is false.
2312 1 : if inputMeta.Virtual {
2313 0 : panic(errors.AssertionFailedf("cannot do a copy compaction of a virtual sstable across local/remote storage"))
2314 : }
2315 : }
2316 :
2317 : // We are in the relatively more complex case where we need to copy this
2318 : // file to remote storage. Drop the db mutex while we do the copy
2319 : //
2320 : // To ease up cleanup of the local file and tracking of refs, we create
2321 : // a new FileNum. This has the potential of making the block cache less
2322 : // effective, however.
2323 1 : newMeta := &fileMetadata{
2324 1 : Size: inputMeta.Size,
2325 1 : CreationTime: inputMeta.CreationTime,
2326 1 : SmallestSeqNum: inputMeta.SmallestSeqNum,
2327 1 : LargestSeqNum: inputMeta.LargestSeqNum,
2328 1 : Stats: inputMeta.Stats,
2329 1 : Virtual: inputMeta.Virtual,
2330 1 : SyntheticPrefix: inputMeta.SyntheticPrefix,
2331 1 : SyntheticSuffix: inputMeta.SyntheticSuffix,
2332 1 : }
2333 1 : if inputMeta.HasPointKeys {
2334 1 : newMeta.ExtendPointKeyBounds(c.cmp, inputMeta.SmallestPointKey, inputMeta.LargestPointKey)
2335 1 : }
2336 1 : if inputMeta.HasRangeKeys {
2337 1 : newMeta.ExtendRangeKeyBounds(c.cmp, inputMeta.SmallestRangeKey, inputMeta.LargestRangeKey)
2338 1 : }
2339 1 : newMeta.FileNum = d.mu.versions.getNextFileNum()
2340 1 : if objMeta.IsExternal() {
2341 1 : // external -> local/shared copy. File must be virtual.
2342 1 : // We will update this size later after we produce the new backing file.
2343 1 : newMeta.InitProviderBacking(base.DiskFileNum(newMeta.FileNum), inputMeta.FileBacking.Size)
2344 1 : } else {
2345 1 : // local -> shared copy. New file is guaranteed to not be virtual.
2346 1 : newMeta.InitPhysicalBacking()
2347 1 : }
2348 :
2349 1 : c.metrics = map[int]*LevelMetrics{
2350 1 : c.outputLevel.level: {
2351 1 : BytesIn: inputMeta.Size,
2352 1 : BytesCompacted: inputMeta.Size,
2353 1 : TablesCompacted: 1,
2354 1 : },
2355 1 : }
2356 1 :
2357 1 : // Before dropping the db mutex, grab a ref to the current version. This
2358 1 : // prevents any concurrent excises from deleting files that this compaction
2359 1 : // needs to read/maintain a reference to.
2360 1 : vers := d.mu.versions.currentVersion()
2361 1 : vers.Ref()
2362 1 : defer vers.UnrefLocked()
2363 1 :
2364 1 : // NB: The order here is reversed, lock after unlock. This is similar to
2365 1 : // runCompaction.
2366 1 : d.mu.Unlock()
2367 1 : defer d.mu.Lock()
2368 1 :
2369 1 : // If the src obj is external, we're doing an external to local/shared copy.
2370 1 : if objMeta.IsExternal() {
2371 1 : src, err := d.objProvider.OpenForReading(
2372 1 : ctx, fileTypeTable, inputMeta.FileBacking.DiskFileNum, objstorage.OpenOptions{},
2373 1 : )
2374 1 : if err != nil {
2375 0 : return pendingOutputs, err
2376 0 : }
2377 1 : defer func() {
2378 1 : if src != nil {
2379 0 : src.Close()
2380 0 : }
2381 : }()
2382 :
2383 1 : w, outObjMeta, err := d.objProvider.Create(
2384 1 : ctx, fileTypeTable, base.PhysicalTableDiskFileNum(newMeta.FileNum),
2385 1 : objstorage.CreateOptions{
2386 1 : PreferSharedStorage: remote.ShouldCreateShared(d.opts.Experimental.CreateOnShared, c.outputLevel.level),
2387 1 : },
2388 1 : )
2389 1 : if err != nil {
2390 0 : return pendingOutputs, err
2391 0 : }
2392 1 : pendingOutputs = append(pendingOutputs, compactionOutput{
2393 1 : meta: newMeta,
2394 1 : isLocal: !outObjMeta.IsRemote(),
2395 1 : })
2396 1 :
2397 1 : start, end := newMeta.Smallest, newMeta.Largest
2398 1 : if newMeta.SyntheticPrefix.IsSet() {
2399 1 : start.UserKey = newMeta.SyntheticPrefix.Invert(start.UserKey)
2400 1 : end.UserKey = newMeta.SyntheticPrefix.Invert(end.UserKey)
2401 1 : }
2402 1 : if newMeta.SyntheticSuffix.IsSet() {
2403 0 : // Extend the bounds as necessary so that the keys don't include suffixes.
2404 0 : start.UserKey = start.UserKey[:c.comparer.Split(start.UserKey)]
2405 0 : if n := c.comparer.Split(end.UserKey); n < len(end.UserKey) {
2406 0 : end = base.MakeRangeDeleteSentinelKey(c.comparer.ImmediateSuccessor(nil, end.UserKey[:n]))
2407 0 : }
2408 : }
2409 :
2410 1 : wrote, err := sstable.CopySpan(ctx,
2411 1 : src, d.opts.MakeReaderOptions(),
2412 1 : w, d.opts.MakeWriterOptions(c.outputLevel.level, d.FormatMajorVersion().MaxTableFormat()),
2413 1 : start, end,
2414 1 : )
2415 1 : src = nil // We passed src to CopySpan; it's responsible for closing it.
2416 1 : if err != nil {
2417 0 : return pendingOutputs, err
2418 0 : }
2419 1 : newMeta.FileBacking.Size = wrote
2420 1 : newMeta.Size = wrote
2421 1 : } else {
2422 1 : pendingOutputs = append(pendingOutputs, compactionOutput{
2423 1 : meta: newMeta.PhysicalMeta().FileMetadata,
2424 1 : isLocal: true,
2425 1 : })
2426 1 : _, err := d.objProvider.LinkOrCopyFromLocal(context.TODO(), d.opts.FS,
2427 1 : d.objProvider.Path(objMeta), fileTypeTable, newMeta.FileBacking.DiskFileNum,
2428 1 : objstorage.CreateOptions{PreferSharedStorage: true})
2429 1 :
2430 1 : if err != nil {
2431 0 : return pendingOutputs, err
2432 0 : }
2433 : }
2434 1 : ve.NewFiles[0].Meta = newMeta
2435 1 : if newMeta.Virtual {
2436 1 : ve.CreatedBackingTables = []*fileBacking{newMeta.FileBacking}
2437 1 : }
2438 :
2439 1 : if err := d.objProvider.Sync(); err != nil {
2440 0 : return pendingOutputs, err
2441 0 : }
2442 1 : return pendingOutputs, nil
2443 : }
2444 :
2445 : type compactionOutput struct {
2446 : meta *fileMetadata
2447 : isLocal bool
2448 : }
2449 :
2450 : func (d *DB) runDeleteOnlyCompaction(
2451 : jobID JobID, c *compaction,
2452 1 : ) (ve *versionEdit, pendingOutputs []compactionOutput, stats compactStats, retErr error) {
2453 1 : c.metrics = make(map[int]*LevelMetrics, len(c.inputs))
2454 1 : ve = &versionEdit{
2455 1 : DeletedFiles: map[deletedFileEntry]*fileMetadata{},
2456 1 : }
2457 1 : for _, cl := range c.inputs {
2458 1 : levelMetrics := &LevelMetrics{}
2459 1 : iter := cl.files.Iter()
2460 1 : for f := iter.First(); f != nil; f = iter.Next() {
2461 1 : ve.DeletedFiles[deletedFileEntry{
2462 1 : Level: cl.level,
2463 1 : FileNum: f.FileNum,
2464 1 : }] = f
2465 1 : }
2466 1 : c.metrics[cl.level] = levelMetrics
2467 : }
2468 1 : return ve, nil, stats, nil
2469 : }
2470 :
2471 : func (d *DB) runMoveOrCopyCompaction(
2472 : jobID JobID, c *compaction,
2473 1 : ) (ve *versionEdit, pendingOutputs []compactionOutput, stats compactStats, _ error) {
2474 1 : iter := c.startLevel.files.Iter()
2475 1 : meta := iter.First()
2476 1 : if invariants.Enabled {
2477 1 : if iter.Next() != nil {
2478 0 : panic("got more than one file for a move or copy compaction")
2479 : }
2480 : }
2481 1 : if c.cancel.Load() {
2482 0 : return ve, nil, stats, ErrCancelledCompaction
2483 0 : }
2484 1 : objMeta, err := d.objProvider.Lookup(fileTypeTable, meta.FileBacking.DiskFileNum)
2485 1 : if err != nil {
2486 0 : return ve, pendingOutputs, stats, err
2487 0 : }
2488 1 : c.metrics = map[int]*LevelMetrics{
2489 1 : c.outputLevel.level: {
2490 1 : BytesMoved: meta.Size,
2491 1 : TablesMoved: 1,
2492 1 : },
2493 1 : }
2494 1 : ve = &versionEdit{
2495 1 : DeletedFiles: map[deletedFileEntry]*fileMetadata{
2496 1 : {Level: c.startLevel.level, FileNum: meta.FileNum}: meta,
2497 1 : },
2498 1 : NewFiles: []newFileEntry{
2499 1 : {Level: c.outputLevel.level, Meta: meta},
2500 1 : },
2501 1 : }
2502 1 : if c.kind == compactionKindMove {
2503 1 : return ve, nil, stats, nil
2504 1 : }
2505 :
2506 1 : pendingOutputs, err = d.runCopyCompaction(jobID, c, meta, objMeta, ve)
2507 1 : return ve, pendingOutputs, stats, err
2508 : }
2509 :
2510 : // runCompaction runs a compaction that produces new on-disk tables from
2511 : // memtables or old on-disk tables.
2512 : //
2513 : // runCompaction cannot be used for compactionKindIngestedFlushable.
2514 : //
2515 : // d.mu must be held when calling this, but the mutex may be dropped and
2516 : // re-acquired during the course of this method.
2517 : func (d *DB) runCompaction(
2518 : jobID JobID, c *compaction,
2519 1 : ) (ve *versionEdit, pendingOutputs []compactionOutput, stats compactStats, retErr error) {
2520 1 : switch c.kind {
2521 1 : case compactionKindDeleteOnly:
2522 1 : return d.runDeleteOnlyCompaction(jobID, c)
2523 1 : case compactionKindMove, compactionKindCopy:
2524 1 : return d.runMoveOrCopyCompaction(jobID, c)
2525 0 : case compactionKindIngestedFlushable:
2526 0 : panic("pebble: runCompaction cannot handle compactionKindIngestedFlushable.")
2527 : }
2528 :
2529 1 : defer func() {
2530 1 : if retErr != nil {
2531 1 : pendingOutputs = nil
2532 1 : }
2533 : }()
2534 :
2535 1 : snapshots := d.mu.snapshots.toSlice()
2536 1 : formatVers := d.FormatMajorVersion()
2537 1 :
2538 1 : if c.flushing == nil {
2539 1 : // Before dropping the db mutex, grab a ref to the current version. This
2540 1 : // prevents any concurrent excises from deleting files that this compaction
2541 1 : // needs to read/maintain a reference to.
2542 1 : //
2543 1 : // Note that unlike user iterators, compactionIter does not maintain a ref
2544 1 : // of the version or read state.
2545 1 : vers := d.mu.versions.currentVersion()
2546 1 : vers.Ref()
2547 1 : defer vers.UnrefLocked()
2548 1 : }
2549 :
2550 1 : if c.cancel.Load() {
2551 0 : return ve, nil, stats, ErrCancelledCompaction
2552 0 : }
2553 :
2554 : // Release the d.mu lock while doing I/O.
2555 : // Note the unusual order: Unlock and then Lock.
2556 1 : d.mu.Unlock()
2557 1 : defer d.mu.Lock()
2558 1 :
2559 1 : // Compactions use a pool of buffers to read blocks, avoiding polluting the
2560 1 : // block cache with blocks that will not be read again. We initialize the
2561 1 : // buffer pool with a size 12. This initial size does not need to be
2562 1 : // accurate, because the pool will grow to accommodate the maximum number of
2563 1 : // blocks allocated at a given time over the course of the compaction. But
2564 1 : // choosing a size larger than that working set avoids any additional
2565 1 : // allocations to grow the size of the pool over the course of iteration.
2566 1 : //
2567 1 : // Justification for initial size 12: In a two-level compaction, at any
2568 1 : // given moment we'll have 2 index blocks in-use and 2 data blocks in-use.
2569 1 : // Additionally, when decoding a compressed block, we'll temporarily
2570 1 : // allocate 1 additional block to hold the compressed buffer. In the worst
2571 1 : // case that all input sstables have two-level index blocks (+2), value
2572 1 : // blocks (+2), range deletion blocks (+n) and range key blocks (+n), we'll
2573 1 : // additionally require 2n+4 blocks where n is the number of input sstables.
2574 1 : // Range deletion and range key blocks are relatively rare, and the cost of
2575 1 : // an additional allocation or two over the course of the compaction is
2576 1 : // considered to be okay. A larger initial size would cause the pool to hold
2577 1 : // on to more memory, even when it's not in-use because the pool will
2578 1 : // recycle buffers up to the current capacity of the pool. The memory use of
2579 1 : // a 12-buffer pool is expected to be within reason, even if all the buffers
2580 1 : // grow to the typical size of an index block (256 KiB) which would
2581 1 : // translate to 3 MiB per compaction.
2582 1 : c.bufferPool.Init(12)
2583 1 : defer c.bufferPool.Release()
2584 1 :
2585 1 : iiter, err := c.newInputIter(d.newIters, d.tableNewRangeKeyIter)
2586 1 : if err != nil {
2587 0 : return nil, pendingOutputs, stats, err
2588 0 : }
2589 1 : c.allowedZeroSeqNum = c.allowZeroSeqNum()
2590 1 : iiter = invalidating.MaybeWrapIfInvariants(iiter)
2591 1 : cfg := compact.IterConfig{
2592 1 : Cmp: c.cmp,
2593 1 : Equal: c.equal,
2594 1 : Merge: d.merge,
2595 1 : TombstoneElision: c.delElision,
2596 1 : RangeKeyElision: c.rangeKeyElision,
2597 1 : Snapshots: snapshots,
2598 1 : AllowZeroSeqNum: c.allowedZeroSeqNum,
2599 1 : IneffectualSingleDeleteCallback: d.opts.Experimental.IneffectualSingleDeleteCallback,
2600 1 : SingleDeleteInvariantViolationCallback: d.opts.Experimental.SingleDeleteInvariantViolationCallback,
2601 1 : }
2602 1 : iter := compact.NewIter(cfg, iiter)
2603 1 :
2604 1 : var (
2605 1 : createdFiles []base.DiskFileNum
2606 1 : tw *sstable.Writer
2607 1 : pinnedKeySize uint64
2608 1 : pinnedValueSize uint64
2609 1 : pinnedCount uint64
2610 1 : )
2611 1 : defer func() {
2612 1 : if iter != nil {
2613 1 : retErr = firstError(retErr, iter.Close())
2614 1 : }
2615 1 : if tw != nil {
2616 0 : retErr = firstError(retErr, tw.Close())
2617 0 : }
2618 1 : if retErr != nil {
2619 1 : for _, fileNum := range createdFiles {
2620 1 : _ = d.objProvider.Remove(fileTypeTable, fileNum)
2621 1 : }
2622 : }
2623 1 : for _, closer := range c.closers {
2624 1 : retErr = firstError(retErr, closer.Close())
2625 1 : }
2626 : }()
2627 :
2628 1 : ve = &versionEdit{
2629 1 : DeletedFiles: map[deletedFileEntry]*fileMetadata{},
2630 1 : }
2631 1 :
2632 1 : startLevelBytes := c.startLevel.files.SizeSum()
2633 1 : outputMetrics := &LevelMetrics{
2634 1 : BytesIn: startLevelBytes,
2635 1 : BytesRead: c.outputLevel.files.SizeSum(),
2636 1 : }
2637 1 : if len(c.extraLevels) > 0 {
2638 1 : outputMetrics.BytesIn += c.extraLevels[0].files.SizeSum()
2639 1 : }
2640 1 : outputMetrics.BytesRead += outputMetrics.BytesIn
2641 1 :
2642 1 : c.metrics = map[int]*LevelMetrics{
2643 1 : c.outputLevel.level: outputMetrics,
2644 1 : }
2645 1 : if len(c.flushing) == 0 && c.metrics[c.startLevel.level] == nil {
2646 1 : c.metrics[c.startLevel.level] = &LevelMetrics{}
2647 1 : }
2648 1 : if len(c.extraLevels) > 0 {
2649 1 : c.metrics[c.extraLevels[0].level] = &LevelMetrics{}
2650 1 : outputMetrics.MultiLevel.BytesInTop = startLevelBytes
2651 1 : outputMetrics.MultiLevel.BytesIn = outputMetrics.BytesIn
2652 1 : outputMetrics.MultiLevel.BytesRead = outputMetrics.BytesRead
2653 1 : }
2654 :
2655 : // The table is typically written at the maximum allowable format implied by
2656 : // the current format major version of the DB.
2657 1 : tableFormat := formatVers.MaxTableFormat()
2658 1 :
2659 1 : // In format major versions with maximum table formats of Pebblev3, value
2660 1 : // blocks were conditional on an experimental setting. In format major
2661 1 : // versions with maximum table formats of Pebblev4 and higher, value blocks
2662 1 : // are always enabled.
2663 1 : if tableFormat == sstable.TableFormatPebblev3 &&
2664 1 : (d.opts.Experimental.EnableValueBlocks == nil || !d.opts.Experimental.EnableValueBlocks()) {
2665 1 : tableFormat = sstable.TableFormatPebblev2
2666 1 : }
2667 :
2668 1 : writerOpts := d.opts.MakeWriterOptions(c.outputLevel.level, tableFormat)
2669 1 :
2670 1 : // prevPointKey is a sstable.WriterOption that provides access to
2671 1 : // the last point key written to a writer's sstable. When a new
2672 1 : // output begins in newOutput, prevPointKey is updated to point to
2673 1 : // the new output's sstable.Writer. This allows the compaction loop
2674 1 : // to access the last written point key without requiring the
2675 1 : // compaction loop to make a copy of each key ahead of time. Users
2676 1 : // must be careful, because the byte slice returned by UnsafeKey
2677 1 : // points directly into the Writer's block buffer.
2678 1 : var cpuWorkHandle CPUWorkHandle
2679 1 : defer func() {
2680 1 : if cpuWorkHandle != nil {
2681 1 : d.opts.Experimental.CPUWorkPermissionGranter.CPUWorkDone(cpuWorkHandle)
2682 1 : }
2683 : }()
2684 :
2685 1 : newOutput := func() error {
2686 1 : // Check if we've been cancelled by a concurrent operation.
2687 1 : if c.cancel.Load() {
2688 0 : return ErrCancelledCompaction
2689 0 : }
2690 1 : d.mu.Lock()
2691 1 : fileNum := d.mu.versions.getNextFileNum()
2692 1 : d.mu.Unlock()
2693 1 :
2694 1 : ctx := context.TODO()
2695 1 : if objiotracing.Enabled {
2696 0 : ctx = objiotracing.WithLevel(ctx, c.outputLevel.level)
2697 0 : switch c.kind {
2698 0 : case compactionKindFlush:
2699 0 : ctx = objiotracing.WithReason(ctx, objiotracing.ForFlush)
2700 0 : default:
2701 0 : ctx = objiotracing.WithReason(ctx, objiotracing.ForCompaction)
2702 : }
2703 : }
2704 1 : var writeCategory vfs.DiskWriteCategory
2705 1 : switch c.kind {
2706 1 : case compactionKindFlush:
2707 1 : if d.opts.EnableSQLRowSpillMetrics {
2708 0 : // In the scenario that the Pebble engine is used for SQL row spills the data written to
2709 0 : // the memtable will correspond to spills to disk and should be categorized as such.
2710 0 : writeCategory = "sql-row-spill"
2711 1 : } else {
2712 1 : writeCategory = "pebble-memtable-flush"
2713 1 : }
2714 1 : default:
2715 1 : writeCategory = "pebble-compaction"
2716 : }
2717 : // Prefer shared storage if present.
2718 1 : createOpts := objstorage.CreateOptions{
2719 1 : PreferSharedStorage: remote.ShouldCreateShared(d.opts.Experimental.CreateOnShared, c.outputLevel.level),
2720 1 : WriteCategory: writeCategory,
2721 1 : }
2722 1 : diskFileNum := base.PhysicalTableDiskFileNum(fileNum)
2723 1 : writable, objMeta, err := d.objProvider.Create(ctx, fileTypeTable, diskFileNum, createOpts)
2724 1 : if err != nil {
2725 1 : return err
2726 1 : }
2727 1 : fileMeta := &fileMetadata{}
2728 1 : fileMeta.FileNum = fileNum
2729 1 : pendingOutputs = append(pendingOutputs, compactionOutput{
2730 1 : meta: fileMeta.PhysicalMeta().FileMetadata,
2731 1 : isLocal: !objMeta.IsRemote(),
2732 1 : })
2733 1 :
2734 1 : reason := "flushing"
2735 1 : if c.flushing == nil {
2736 1 : reason = "compacting"
2737 1 : }
2738 1 : d.opts.EventListener.TableCreated(TableCreateInfo{
2739 1 : JobID: int(jobID),
2740 1 : Reason: reason,
2741 1 : Path: d.objProvider.Path(objMeta),
2742 1 : FileNum: diskFileNum,
2743 1 : })
2744 1 : if c.kind != compactionKindFlush {
2745 1 : writable = &compactionWritable{
2746 1 : Writable: writable,
2747 1 : versions: d.mu.versions,
2748 1 : written: &c.bytesWritten,
2749 1 : }
2750 1 : }
2751 1 : createdFiles = append(createdFiles, diskFileNum)
2752 1 : cacheOpts := private.SSTableCacheOpts(d.cacheID, diskFileNum).(sstable.WriterOption)
2753 1 :
2754 1 : const MaxFileWriteAdditionalCPUTime = time.Millisecond * 100
2755 1 : cpuWorkHandle = d.opts.Experimental.CPUWorkPermissionGranter.GetPermission(
2756 1 : MaxFileWriteAdditionalCPUTime,
2757 1 : )
2758 1 : writerOpts.Parallelism =
2759 1 : d.opts.Experimental.MaxWriterConcurrency > 0 &&
2760 1 : (cpuWorkHandle.Permitted() || d.opts.Experimental.ForceWriterParallelism)
2761 1 :
2762 1 : tw = sstable.NewWriter(writable, writerOpts, cacheOpts)
2763 1 :
2764 1 : fileMeta.CreationTime = time.Now().Unix()
2765 1 : ve.NewFiles = append(ve.NewFiles, newFileEntry{
2766 1 : Level: c.outputLevel.level,
2767 1 : Meta: fileMeta,
2768 1 : })
2769 1 : return nil
2770 : }
2771 :
2772 : // splitL0Outputs is true during flushes and intra-L0 compactions with flush
2773 : // splits enabled.
2774 1 : splitL0Outputs := c.outputLevel.level == 0 && d.opts.FlushSplitBytes > 0
2775 1 :
2776 1 : // finishOutput is called with the a user key up to which all tombstones
2777 1 : // should be flushed. Typically, this is the first key of the next
2778 1 : // sstable or an empty key if this output is the final sstable.
2779 1 : finishOutput := func(splitKey []byte) error {
2780 1 : // If we haven't output any point records to the sstable (tw == nil) then the
2781 1 : // sstable will only contain range tombstones and/or range keys. The smallest
2782 1 : // key in the sstable will be the start key of the first range tombstone or
2783 1 : // range key added. We need to ensure that this start key is distinct from
2784 1 : // the splitKey passed to finishOutput (if set), otherwise we would generate
2785 1 : // an sstable where the largest key is smaller than the smallest key due to
2786 1 : // how the largest key boundary is set below. NB: It is permissible for the
2787 1 : // range tombstone / range key start key to be nil.
2788 1 : //
2789 1 : // TODO: It is unfortunate that we have to do this check here rather than
2790 1 : // when we decide to finish the sstable in the runCompaction loop. A better
2791 1 : // structure currently eludes us.
2792 1 : if tw == nil {
2793 1 : startKey := iter.FirstTombstoneStart()
2794 1 : if startKey == nil {
2795 1 : startKey = iter.FirstRangeKeyStart()
2796 1 : }
2797 1 : if splitKey != nil && d.cmp(startKey, splitKey) == 0 {
2798 0 : return nil
2799 0 : }
2800 : }
2801 :
2802 1 : for _, v := range iter.TombstonesUpTo(splitKey) {
2803 1 : if tw == nil {
2804 1 : if err := newOutput(); err != nil {
2805 0 : return err
2806 0 : }
2807 : }
2808 : // The tombstone being added could be completely outside the
2809 : // eventual bounds of the sstable. Consider this example (bounds
2810 : // in square brackets next to table filename):
2811 : //
2812 : // ./000240.sst [tmgc#391,MERGE-tmgc#391,MERGE]
2813 : // tmgc#391,MERGE [786e627a]
2814 : // tmgc-udkatvs#331,RANGEDEL
2815 : //
2816 : // ./000241.sst [tmgc#384,MERGE-tmgc#384,MERGE]
2817 : // tmgc#384,MERGE [666c7070]
2818 : // tmgc-tvsalezade#383,RANGEDEL
2819 : // tmgc-tvsalezade#331,RANGEDEL
2820 : //
2821 : // ./000242.sst [tmgc#383,RANGEDEL-tvsalezade#72057594037927935,RANGEDEL]
2822 : // tmgc-tvsalezade#383,RANGEDEL
2823 : // tmgc#375,SET [72646c78766965616c72776865676e79]
2824 : // tmgc-tvsalezade#356,RANGEDEL
2825 : //
2826 : // Note that both of the top two SSTables have range tombstones
2827 : // that start after the file's end keys. Since the file bound
2828 : // computation happens well after all range tombstones have been
2829 : // added to the writer, eliding out-of-file range tombstones based
2830 : // on sequence number at this stage is difficult, and necessitates
2831 : // read-time logic to ignore range tombstones outside file bounds.
2832 1 : if err := rangedel.Encode(&v, tw.Add); err != nil {
2833 0 : return err
2834 0 : }
2835 : }
2836 1 : for _, v := range iter.RangeKeysUpTo(splitKey) {
2837 1 : // Same logic as for range tombstones, except added using tw.AddRangeKey.
2838 1 : if tw == nil {
2839 1 : if err := newOutput(); err != nil {
2840 0 : return err
2841 0 : }
2842 : }
2843 1 : if err := rangekey.Encode(&v, tw.AddRangeKey); err != nil {
2844 0 : return err
2845 0 : }
2846 : }
2847 :
2848 1 : if tw == nil {
2849 1 : return nil
2850 1 : }
2851 1 : {
2852 1 : // Set internal sstable properties.
2853 1 : p := getInternalWriterProperties(tw)
2854 1 : // Set the snapshot pinned totals.
2855 1 : p.SnapshotPinnedKeys = pinnedCount
2856 1 : p.SnapshotPinnedKeySize = pinnedKeySize
2857 1 : p.SnapshotPinnedValueSize = pinnedValueSize
2858 1 : stats.cumulativePinnedKeys += pinnedCount
2859 1 : stats.cumulativePinnedSize += pinnedKeySize + pinnedValueSize
2860 1 : pinnedCount = 0
2861 1 : pinnedKeySize = 0
2862 1 : pinnedValueSize = 0
2863 1 : }
2864 1 : if err := tw.Close(); err != nil {
2865 1 : tw = nil
2866 1 : return err
2867 1 : }
2868 1 : d.opts.Experimental.CPUWorkPermissionGranter.CPUWorkDone(cpuWorkHandle)
2869 1 : cpuWorkHandle = nil
2870 1 : writerMeta, err := tw.Metadata()
2871 1 : if err != nil {
2872 0 : tw = nil
2873 0 : return err
2874 0 : }
2875 1 : tw = nil
2876 1 : meta := ve.NewFiles[len(ve.NewFiles)-1].Meta
2877 1 : meta.Size = writerMeta.Size
2878 1 : meta.SmallestSeqNum = writerMeta.SmallestSeqNum
2879 1 : meta.LargestSeqNum = writerMeta.LargestSeqNum
2880 1 : meta.InitPhysicalBacking()
2881 1 :
2882 1 : // If the file didn't contain any range deletions, we can fill its
2883 1 : // table stats now, avoiding unnecessarily loading the table later.
2884 1 : maybeSetStatsFromProperties(
2885 1 : meta.PhysicalMeta(), &writerMeta.Properties,
2886 1 : )
2887 1 :
2888 1 : if c.flushing == nil {
2889 1 : outputMetrics.TablesCompacted++
2890 1 : outputMetrics.BytesCompacted += meta.Size
2891 1 : } else {
2892 1 : outputMetrics.TablesFlushed++
2893 1 : outputMetrics.BytesFlushed += meta.Size
2894 1 : }
2895 1 : outputMetrics.Size += int64(meta.Size)
2896 1 : outputMetrics.NumFiles++
2897 1 : outputMetrics.Additional.BytesWrittenDataBlocks += writerMeta.Properties.DataSize
2898 1 : outputMetrics.Additional.BytesWrittenValueBlocks += writerMeta.Properties.ValueBlocksSize
2899 1 :
2900 1 : if n := len(ve.NewFiles); n > 1 {
2901 1 : // This is not the first output file. Ensure the sstable boundaries
2902 1 : // are nonoverlapping.
2903 1 : prevMeta := ve.NewFiles[n-2].Meta
2904 1 : if writerMeta.SmallestRangeDel.UserKey != nil {
2905 1 : c := d.cmp(writerMeta.SmallestRangeDel.UserKey, prevMeta.Largest.UserKey)
2906 1 : if c < 0 {
2907 0 : return errors.Errorf(
2908 0 : "pebble: smallest range tombstone start key is less than previous sstable largest key: %s < %s",
2909 0 : writerMeta.SmallestRangeDel.Pretty(d.opts.Comparer.FormatKey),
2910 0 : prevMeta.Largest.Pretty(d.opts.Comparer.FormatKey))
2911 1 : } else if c == 0 && !prevMeta.Largest.IsExclusiveSentinel() {
2912 0 : // The user key portion of the range boundary start key is
2913 0 : // equal to the previous table's largest key user key, and
2914 0 : // the previous table's largest key is not exclusive. This
2915 0 : // violates the invariant that tables are key-space
2916 0 : // partitioned.
2917 0 : return errors.Errorf(
2918 0 : "pebble: invariant violation: previous sstable largest key %s, current sstable smallest rangedel: %s",
2919 0 : prevMeta.Largest.Pretty(d.opts.Comparer.FormatKey),
2920 0 : writerMeta.SmallestRangeDel.Pretty(d.opts.Comparer.FormatKey),
2921 0 : )
2922 0 : }
2923 : }
2924 : }
2925 :
2926 : // Verify that all range deletions outputted to the sstable are
2927 : // truncated to split key.
2928 1 : if splitKey != nil && writerMeta.LargestRangeDel.UserKey != nil &&
2929 1 : d.cmp(writerMeta.LargestRangeDel.UserKey, splitKey) > 0 {
2930 0 : return errors.Errorf(
2931 0 : "pebble: invariant violation: rangedel largest key %q extends beyond split key %q",
2932 0 : writerMeta.LargestRangeDel.Pretty(d.opts.Comparer.FormatKey),
2933 0 : d.opts.Comparer.FormatKey(splitKey),
2934 0 : )
2935 0 : }
2936 :
2937 1 : if writerMeta.HasPointKeys {
2938 1 : meta.ExtendPointKeyBounds(d.cmp, writerMeta.SmallestPoint, writerMeta.LargestPoint)
2939 1 : }
2940 1 : if writerMeta.HasRangeDelKeys {
2941 1 : meta.ExtendPointKeyBounds(d.cmp, writerMeta.SmallestRangeDel, writerMeta.LargestRangeDel)
2942 1 : }
2943 1 : if writerMeta.HasRangeKeys {
2944 1 : meta.ExtendRangeKeyBounds(d.cmp, writerMeta.SmallestRangeKey, writerMeta.LargestRangeKey)
2945 1 : }
2946 :
2947 : // Verify that the sstable bounds fall within the compaction input
2948 : // bounds. This is a sanity check that we don't have a logic error
2949 : // elsewhere that causes the sstable bounds to accidentally expand past the
2950 : // compaction input bounds as doing so could lead to various badness such
2951 : // as keys being deleted by a range tombstone incorrectly.
2952 1 : if c.smallest.UserKey != nil {
2953 1 : switch v := d.cmp(meta.Smallest.UserKey, c.smallest.UserKey); {
2954 1 : case v >= 0:
2955 : // Nothing to do.
2956 0 : case v < 0:
2957 0 : return errors.Errorf("pebble: compaction output grew beyond bounds of input: %s < %s",
2958 0 : meta.Smallest.Pretty(d.opts.Comparer.FormatKey),
2959 0 : c.smallest.Pretty(d.opts.Comparer.FormatKey))
2960 : }
2961 : }
2962 1 : if c.largest.UserKey != nil {
2963 1 : switch v := d.cmp(meta.Largest.UserKey, c.largest.UserKey); {
2964 1 : case v <= 0:
2965 : // Nothing to do.
2966 0 : case v > 0:
2967 0 : return errors.Errorf("pebble: compaction output grew beyond bounds of input: %s > %s",
2968 0 : meta.Largest.Pretty(d.opts.Comparer.FormatKey),
2969 0 : c.largest.Pretty(d.opts.Comparer.FormatKey))
2970 : }
2971 : }
2972 : // Verify that we never split different revisions of the same user key
2973 : // across two different sstables.
2974 1 : if err := c.errorOnUserKeyOverlap(ve); err != nil {
2975 0 : return err
2976 0 : }
2977 1 : if err := meta.Validate(d.cmp, d.opts.Comparer.FormatKey); err != nil {
2978 0 : return err
2979 0 : }
2980 1 : return nil
2981 : }
2982 :
2983 : // Build a compactionOutputSplitter that contains all logic to determine
2984 : // whether the compaction loop should stop writing to one output sstable and
2985 : // switch to a new one. Some splitters can wrap other splitters, and the
2986 : // splitterGroup can be composed of multiple splitters. In this case, we
2987 : // start off with splitters for file sizes, grandparent limits, and (for L0
2988 : // splits) L0 limits, before wrapping them in an splitterGroup.
2989 1 : unsafePrevUserKey := func() []byte {
2990 1 : // Return the largest point key written to tw or the start of
2991 1 : // the current range deletion in the fragmenter, whichever is
2992 1 : // greater.
2993 1 : prevPoint := tw.UnsafeLastPointKey()
2994 1 : if c.cmp(prevPoint.UserKey, iter.FirstTombstoneStart()) > 0 {
2995 1 : return prevPoint.UserKey
2996 1 : }
2997 1 : return iter.FirstTombstoneStart()
2998 : }
2999 1 : outputSplitters := []compact.OutputSplitter{
3000 1 : // We do not split the same user key across different sstables within
3001 1 : // one flush or compaction. The FileSizeSplitter may request a split in
3002 1 : // the middle of a user key, so PreventSplitUserKeys ensures we are at a
3003 1 : // user key change boundary when doing a split.
3004 1 : compact.PreventSplitUserKeys(
3005 1 : c.cmp,
3006 1 : compact.FileSizeSplitter(iter.Frontiers(), c.maxOutputFileSize, c.grandparents.Iter()),
3007 1 : unsafePrevUserKey,
3008 1 : ),
3009 1 : compact.LimitFuncSplitter(iter.Frontiers(), c.findGrandparentLimit),
3010 1 : }
3011 1 : if splitL0Outputs {
3012 1 : outputSplitters = append(outputSplitters, compact.LimitFuncSplitter(iter.Frontiers(), c.findL0Limit))
3013 1 : }
3014 1 : splitter := compact.CombineSplitters(c.cmp, outputSplitters...)
3015 1 :
3016 1 : // Each outer loop iteration produces one output file. An iteration that
3017 1 : // produces a file containing point keys (and optionally range tombstones)
3018 1 : // guarantees that the input iterator advanced. An iteration that produces
3019 1 : // a file containing only range tombstones guarantees the limit passed to
3020 1 : // `finishOutput()` advanced to a strictly greater user key corresponding
3021 1 : // to a grandparent file largest key, or nil. Taken together, these
3022 1 : // progress guarantees ensure that eventually the input iterator will be
3023 1 : // exhausted and the range tombstone fragments will all be flushed.
3024 1 : for key, val := iter.First(); key != nil || iter.FirstTombstoneStart() != nil || iter.FirstRangeKeyStart() != nil; {
3025 1 : var firstKey []byte
3026 1 : if key != nil {
3027 1 : firstKey = key.UserKey
3028 1 : } else if startKey := iter.FirstTombstoneStart(); startKey != nil {
3029 1 : // Pass the start key of the first pending tombstone to find the
3030 1 : // next limit. All pending tombstones have the same start key. We
3031 1 : // use this as opposed to the end key of the last written sstable to
3032 1 : // effectively handle cases like these:
3033 1 : //
3034 1 : // a.SET.3
3035 1 : // (lf.limit at b)
3036 1 : // d.RANGEDEL.4:f
3037 1 : //
3038 1 : // In this case, the partition after b has only range deletions, so
3039 1 : // if we were to find the limit after the last written key at the
3040 1 : // split point (key a), we'd get the limit b again, and
3041 1 : // finishOutput() would not advance any further because the next
3042 1 : // range tombstone to write does not start until after the L0 split
3043 1 : // point.
3044 1 : firstKey = startKey
3045 1 : }
3046 1 : splitterSuggestion := splitter.OnNewOutput(firstKey)
3047 1 :
3048 1 : // Each inner loop iteration processes one key from the input iterator.
3049 1 : for ; key != nil; key, val = iter.Next() {
3050 1 : if split := splitter.ShouldSplitBefore(key, tw); split == compact.SplitNow {
3051 1 : break
3052 : }
3053 :
3054 1 : switch key.Kind() {
3055 1 : case InternalKeyKindRangeDelete:
3056 1 : // Range tombstones are handled specially. They are not written until
3057 1 : // later during `finishOutput()`. We add them to the compaction iterator
3058 1 : // so covered keys in the same snapshot stripe can be elided.
3059 1 : //
3060 1 : // Since the keys' Suffix and Value fields are not deep cloned, the
3061 1 : // underlying blockIter must be kept open for the lifetime of the
3062 1 : // compaction.
3063 1 : iter.AddTombstoneSpan(c.rangeDelInterleaving.Span())
3064 1 : continue
3065 1 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
3066 1 : // Range keys are handled in the same way as range tombstones.
3067 1 : // Since the keys' Suffix and Value fields are not deep cloned, the
3068 1 : // underlying blockIter must be kept open for the lifetime of the
3069 1 : // compaction.
3070 1 : iter.AddRangeKeySpan(c.rangeKeyInterleaving.Span())
3071 1 : continue
3072 : }
3073 1 : if tw == nil {
3074 1 : if err := newOutput(); err != nil {
3075 1 : return nil, pendingOutputs, stats, err
3076 1 : }
3077 : }
3078 1 : if err := tw.AddWithForceObsolete(*key, val, iter.ForceObsoleteDueToRangeDel()); err != nil {
3079 0 : return nil, pendingOutputs, stats, err
3080 0 : }
3081 1 : if iter.SnapshotPinned() {
3082 1 : // The kv pair we just added to the sstable was only surfaced by
3083 1 : // the compaction iterator because an open snapshot prevented
3084 1 : // its elision. Increment the stats.
3085 1 : pinnedCount++
3086 1 : pinnedKeySize += uint64(len(key.UserKey)) + base.InternalTrailerLen
3087 1 : pinnedValueSize += uint64(len(val))
3088 1 : }
3089 : }
3090 :
3091 : // A splitter requested a split, and we're ready to finish the output.
3092 : // We need to choose the key at which to split any pending range
3093 : // tombstones. There are two options:
3094 : // 1. splitterSuggestion — The key suggested by the splitter. This key
3095 : // is guaranteed to be greater than the last key written to the
3096 : // current output.
3097 : // 2. key.UserKey — the first key of the next sstable output. This user
3098 : // key is also guaranteed to be greater than the last user key
3099 : // written to the current output (see userKeyChangeSplitter).
3100 : //
3101 : // Use whichever is smaller. Using the smaller of the two limits
3102 : // overlap with grandparents. Consider the case where the
3103 : // grandparent limit is calculated to be 'b', key is 'x', and
3104 : // there exist many sstables between 'b' and 'x'. If the range
3105 : // deletion fragmenter has a pending tombstone [a,x), splitting
3106 : // at 'x' would cause the output table to overlap many
3107 : // grandparents well beyond the calculated grandparent limit
3108 : // 'b'. Splitting at the smaller `splitterSuggestion` avoids
3109 : // this unbounded overlap with grandparent tables.
3110 1 : splitKey := splitterSuggestion
3111 1 : if key != nil && (splitKey == nil || c.cmp(splitKey, key.UserKey) > 0) {
3112 1 : splitKey = key.UserKey
3113 1 : }
3114 1 : if err := finishOutput(splitKey); err != nil {
3115 1 : return nil, pendingOutputs, stats, err
3116 1 : }
3117 : }
3118 :
3119 1 : for _, cl := range c.inputs {
3120 1 : iter := cl.files.Iter()
3121 1 : for f := iter.First(); f != nil; f = iter.Next() {
3122 1 : ve.DeletedFiles[deletedFileEntry{
3123 1 : Level: cl.level,
3124 1 : FileNum: f.FileNum,
3125 1 : }] = f
3126 1 : }
3127 : }
3128 :
3129 : // The compaction iterator keeps track of a count of the number of DELSIZED
3130 : // keys that encoded an incorrect size. Propagate it up as a part of
3131 : // compactStats.
3132 1 : stats.countMissizedDels = iter.Stats().CountMissizedDels
3133 1 :
3134 1 : if err := d.objProvider.Sync(); err != nil {
3135 0 : return nil, pendingOutputs, stats, err
3136 0 : }
3137 :
3138 : // Refresh the disk available statistic whenever a compaction/flush
3139 : // completes, before re-acquiring the mutex.
3140 1 : _ = d.calculateDiskAvailableBytes()
3141 1 :
3142 1 : return ve, pendingOutputs, stats, nil
3143 : }
3144 :
3145 : // validateVersionEdit validates that start and end keys across new and deleted
3146 : // files in a versionEdit pass the given validation function.
3147 : func validateVersionEdit(
3148 : ve *versionEdit, validateFn func([]byte) error, format base.FormatKey, logger Logger,
3149 1 : ) {
3150 1 : validateKey := func(f *manifest.FileMetadata, key []byte) {
3151 1 : if err := validateFn(key); err != nil {
3152 1 : logger.Fatalf("pebble: version edit validation failed (key=%s file=%s): %v", format(key), f, err)
3153 1 : }
3154 : }
3155 :
3156 : // Validate both new and deleted files.
3157 1 : for _, f := range ve.NewFiles {
3158 1 : validateKey(f.Meta, f.Meta.Smallest.UserKey)
3159 1 : validateKey(f.Meta, f.Meta.Largest.UserKey)
3160 1 : }
3161 1 : for _, m := range ve.DeletedFiles {
3162 1 : validateKey(m, m.Smallest.UserKey)
3163 1 : validateKey(m, m.Largest.UserKey)
3164 1 : }
3165 : }
|