Line data Source code
1 : // Copyright 2012 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 : "fmt"
9 : "io"
10 : "slices"
11 : "sync"
12 : "sync/atomic"
13 : "time"
14 :
15 : "github.com/cockroachdb/crlib/crtime"
16 : "github.com/cockroachdb/errors"
17 : "github.com/cockroachdb/pebble/internal/base"
18 : "github.com/cockroachdb/pebble/internal/invariants"
19 : "github.com/cockroachdb/pebble/internal/manifest"
20 : "github.com/cockroachdb/pebble/objstorage"
21 : "github.com/cockroachdb/pebble/record"
22 : "github.com/cockroachdb/pebble/vfs"
23 : "github.com/cockroachdb/pebble/vfs/atomicfs"
24 : )
25 :
26 : const numLevels = manifest.NumLevels
27 :
28 : const manifestMarkerName = `manifest`
29 :
30 : // versionSet manages a collection of immutable versions, and manages the
31 : // creation of a new version from the most recent version. A new version is
32 : // created from an existing version by applying a version edit which is just
33 : // like it sounds: a delta from the previous version. Version edits are logged
34 : // to the MANIFEST file, which is replayed at startup.
35 : type versionSet struct {
36 : // Next seqNum to use for WAL writes.
37 : logSeqNum base.AtomicSeqNum
38 :
39 : // The upper bound on sequence numbers that have been assigned so far. A
40 : // suffix of these sequence numbers may not have been written to a WAL. Both
41 : // logSeqNum and visibleSeqNum are atomically updated by the commitPipeline.
42 : // visibleSeqNum is <= logSeqNum.
43 : visibleSeqNum base.AtomicSeqNum
44 :
45 : // Number of bytes present in sstables being written by in-progress
46 : // compactions. This value will be zero if there are no in-progress
47 : // compactions. Updated and read atomically.
48 : atomicInProgressBytes atomic.Int64
49 :
50 : // Immutable fields.
51 : dirname string
52 : provider objstorage.Provider
53 : // Set to DB.mu.
54 : mu *sync.Mutex
55 : opts *Options
56 : fs vfs.FS
57 : cmp *base.Comparer
58 : // Dynamic base level allows the dynamic base level computation to be
59 : // disabled. Used by tests which want to create specific LSM structures.
60 : dynamicBaseLevel bool
61 :
62 : // Mutable fields.
63 : versions manifest.VersionList
64 : latest latestVersionState
65 : picker compactionPicker
66 : // curCompactionConcurrency is updated whenever picker is updated.
67 : // INVARIANT: >= 1.
68 : curCompactionConcurrency atomic.Int32
69 :
70 : // Not all metrics are kept here. See DB.Metrics().
71 : metrics Metrics
72 :
73 : // A pointer to versionSet.addObsoleteLocked. Avoids allocating a new closure
74 : // on the creation of every version.
75 : obsoleteFn func(manifest.ObsoleteFiles)
76 : // obsolete{Tables,Blobs,Manifests,Options} are sorted by file number ascending.
77 : obsoleteTables []obsoleteFile
78 : obsoleteBlobs []obsoleteFile
79 : obsoleteManifests []obsoleteFile
80 : obsoleteOptions []obsoleteFile
81 :
82 : // Zombie tables which have been removed from the current version but are
83 : // still referenced by an inuse iterator.
84 : zombieTables zombieObjects
85 : // Zombie blobs which have been removed from the current version but are
86 : // still referenced by an inuse iterator.
87 : zombieBlobs zombieObjects
88 :
89 : // minUnflushedLogNum is the smallest WAL log file number corresponding to
90 : // mutations that have not been flushed to an sstable.
91 : minUnflushedLogNum base.DiskFileNum
92 :
93 : // The next file number. A single counter is used to assign file
94 : // numbers for the WAL, MANIFEST, sstable, and OPTIONS files.
95 : nextFileNum atomic.Uint64
96 :
97 : // The current manifest file number.
98 : manifestFileNum base.DiskFileNum
99 : manifestMarker *atomicfs.Marker
100 :
101 : manifestFile vfs.File
102 : manifest *record.Writer
103 : getFormatMajorVersion func() FormatMajorVersion
104 :
105 : writing bool
106 : writerCond sync.Cond
107 : // State for deciding when to write a snapshot. Protected by mu.
108 : rotationHelper record.RotationHelper
109 :
110 : pickedCompactionCache pickedCompactionCache
111 : }
112 :
113 : // latestVersionState maintains mutable state describing only the most recent
114 : // version. Unlike a *Version, the state within this struct is updated as new
115 : // version edits are applied.
116 : type latestVersionState struct {
117 : l0Organizer *manifest.L0Organizer
118 : // blobFiles is the set of blob files referenced by the current version.
119 : // blobFiles is protected by the manifest logLock (not vs.mu).
120 : blobFiles manifest.CurrentBlobFileSet
121 : // virtualBackings contains information about the FileBackings which support
122 : // virtual sstables in the latest version. It is mainly used to determine when
123 : // a backing is no longer in use by the tables in the latest version; this is
124 : // not a trivial problem because compactions and other operations can happen
125 : // in parallel (and they can finish in unpredictable order).
126 : //
127 : // This mechanism is complementary to the backing Ref/Unref mechanism, which
128 : // determines when a backing is no longer used by *any* live version and can
129 : // be removed.
130 : //
131 : // In principle this should have been a copy-on-write structure, with each
132 : // Version having its own record of its virtual backings (similar to the
133 : // B-tree that holds the tables). However, in practice we only need it for the
134 : // latest version, so we use a simpler structure and store it in the
135 : // versionSet instead.
136 : //
137 : // virtualBackings is modified under DB.mu and the log lock. If it is accessed
138 : // under DB.mu and a version update is in progress, it reflects the state of
139 : // the next version.
140 : virtualBackings manifest.VirtualBackings
141 : }
142 :
143 : func (vs *versionSet) init(
144 : dirname string,
145 : provider objstorage.Provider,
146 : opts *Options,
147 : marker *atomicfs.Marker,
148 : getFMV func() FormatMajorVersion,
149 : mu *sync.Mutex,
150 1 : ) {
151 1 : vs.dirname = dirname
152 1 : vs.provider = provider
153 1 : vs.mu = mu
154 1 : vs.writerCond.L = mu
155 1 : vs.opts = opts
156 1 : vs.fs = opts.FS
157 1 : vs.cmp = opts.Comparer
158 1 : vs.dynamicBaseLevel = true
159 1 : vs.versions.Init(mu)
160 1 : vs.latest.l0Organizer = manifest.NewL0Organizer(opts.Comparer, opts.FlushSplitBytes)
161 1 : vs.latest.virtualBackings = manifest.MakeVirtualBackings()
162 1 : vs.obsoleteFn = vs.addObsoleteLocked
163 1 : vs.zombieTables = makeZombieObjects()
164 1 : vs.zombieBlobs = makeZombieObjects()
165 1 : vs.nextFileNum.Store(1)
166 1 : vs.manifestMarker = marker
167 1 : vs.getFormatMajorVersion = getFMV
168 1 : }
169 :
170 : // create creates a version set for a fresh DB.
171 : func (vs *versionSet) create(
172 : jobID JobID,
173 : dirname string,
174 : provider objstorage.Provider,
175 : opts *Options,
176 : marker *atomicfs.Marker,
177 : getFormatMajorVersion func() FormatMajorVersion,
178 : blobRewriteHeuristic manifest.BlobRewriteHeuristic,
179 : mu *sync.Mutex,
180 1 : ) error {
181 1 : vs.init(dirname, provider, opts, marker, getFormatMajorVersion, mu)
182 1 : vs.latest.blobFiles.Init(nil, blobRewriteHeuristic)
183 1 : emptyVersion := manifest.NewInitialVersion(opts.Comparer)
184 1 : vs.append(emptyVersion)
185 1 :
186 1 : vs.setCompactionPicker(
187 1 : newCompactionPickerByScore(emptyVersion, &vs.latest, vs.opts, nil))
188 1 : // Note that a "snapshot" version edit is written to the manifest when it is
189 1 : // created.
190 1 : vs.manifestFileNum = vs.getNextDiskFileNum()
191 1 : err := vs.createManifest(vs.dirname, vs.manifestFileNum, vs.minUnflushedLogNum, vs.nextFileNum.Load(), nil /* virtualBackings */)
192 1 : if err == nil {
193 1 : if err = vs.manifest.Flush(); err != nil {
194 1 : vs.opts.Logger.Fatalf("MANIFEST flush failed: %v", err)
195 1 : }
196 : }
197 1 : if err == nil {
198 1 : if err = vs.manifestFile.Sync(); err != nil {
199 1 : vs.opts.Logger.Fatalf("MANIFEST sync failed: %v", err)
200 1 : }
201 : }
202 1 : if err == nil {
203 1 : // NB: Move() is responsible for syncing the data directory.
204 1 : if err = vs.manifestMarker.Move(base.MakeFilename(base.FileTypeManifest, vs.manifestFileNum)); err != nil {
205 1 : vs.opts.Logger.Fatalf("MANIFEST set current failed: %v", err)
206 1 : }
207 : }
208 :
209 1 : vs.opts.EventListener.ManifestCreated(ManifestCreateInfo{
210 1 : JobID: int(jobID),
211 1 : Path: base.MakeFilepath(vs.fs, vs.dirname, base.FileTypeManifest, vs.manifestFileNum),
212 1 : FileNum: vs.manifestFileNum,
213 1 : Err: err,
214 1 : })
215 1 : if err != nil {
216 1 : return err
217 1 : }
218 1 : return nil
219 : }
220 :
221 : // load loads the version set from the manifest file.
222 : func (vs *versionSet) load(
223 : dirname string,
224 : provider objstorage.Provider,
225 : opts *Options,
226 : manifestFileNum base.DiskFileNum,
227 : marker *atomicfs.Marker,
228 : getFormatMajorVersion func() FormatMajorVersion,
229 : blobRewriteHeuristic manifest.BlobRewriteHeuristic,
230 : mu *sync.Mutex,
231 1 : ) error {
232 1 : vs.init(dirname, provider, opts, marker, getFormatMajorVersion, mu)
233 1 :
234 1 : vs.manifestFileNum = manifestFileNum
235 1 : manifestPath := base.MakeFilepath(opts.FS, dirname, base.FileTypeManifest, vs.manifestFileNum)
236 1 : manifestFilename := opts.FS.PathBase(manifestPath)
237 1 :
238 1 : // Read the versionEdits in the manifest file.
239 1 : var bve manifest.BulkVersionEdit
240 1 : bve.AllAddedTables = make(map[base.TableNum]*manifest.TableMetadata)
241 1 : manifestFile, err := vs.fs.Open(manifestPath)
242 1 : if err != nil {
243 0 : return errors.Wrapf(err, "pebble: could not open manifest file %q for DB %q",
244 0 : errors.Safe(manifestFilename), dirname)
245 0 : }
246 1 : defer manifestFile.Close()
247 1 : rr := record.NewReader(manifestFile, 0 /* logNum */)
248 1 : for {
249 1 : r, err := rr.Next()
250 1 : if err == io.EOF || record.IsInvalidRecord(err) {
251 1 : break
252 : }
253 1 : if err != nil {
254 0 : return errors.Wrapf(err, "pebble: error when loading manifest file %q",
255 0 : errors.Safe(manifestFilename))
256 0 : }
257 1 : var ve manifest.VersionEdit
258 1 : err = ve.Decode(r)
259 1 : if err != nil {
260 0 : // Break instead of returning an error if the record is corrupted
261 0 : // or invalid.
262 0 : if err == io.EOF || record.IsInvalidRecord(err) {
263 0 : break
264 : }
265 0 : return err
266 : }
267 1 : if ve.ComparerName != "" {
268 1 : if ve.ComparerName != vs.cmp.Name {
269 1 : return errors.Errorf("pebble: manifest file %q for DB %q: "+
270 1 : "comparer name from file %q != comparer name from Options %q",
271 1 : errors.Safe(manifestFilename), dirname, errors.Safe(ve.ComparerName), errors.Safe(vs.cmp.Name))
272 1 : }
273 : }
274 1 : if err := bve.Accumulate(&ve); err != nil {
275 0 : return err
276 0 : }
277 1 : if ve.MinUnflushedLogNum != 0 {
278 1 : vs.minUnflushedLogNum = ve.MinUnflushedLogNum
279 1 : }
280 1 : if ve.NextFileNum != 0 {
281 1 : vs.nextFileNum.Store(ve.NextFileNum)
282 1 : }
283 1 : if ve.LastSeqNum != 0 {
284 1 : // logSeqNum is the _next_ sequence number that will be assigned,
285 1 : // while LastSeqNum is the last assigned sequence number. Note that
286 1 : // this behaviour mimics that in RocksDB; the first sequence number
287 1 : // assigned is one greater than the one present in the manifest
288 1 : // (assuming no WALs contain higher sequence numbers than the
289 1 : // manifest's LastSeqNum). Increment LastSeqNum by 1 to get the
290 1 : // next sequence number that will be assigned.
291 1 : //
292 1 : // If LastSeqNum is less than SeqNumStart, increase it to at least
293 1 : // SeqNumStart to leave ample room for reserved sequence numbers.
294 1 : if ve.LastSeqNum+1 < base.SeqNumStart {
295 0 : vs.logSeqNum.Store(base.SeqNumStart)
296 1 : } else {
297 1 : vs.logSeqNum.Store(ve.LastSeqNum + 1)
298 1 : }
299 : }
300 : }
301 : // We have already set vs.nextFileNum = 2 at the beginning of the
302 : // function and could have only updated it to some other non-zero value,
303 : // so it cannot be 0 here.
304 1 : if vs.minUnflushedLogNum == 0 {
305 1 : if vs.nextFileNum.Load() >= 2 {
306 1 : // We either have a freshly created DB, or a DB created by RocksDB
307 1 : // that has not had a single flushed SSTable yet. This is because
308 1 : // RocksDB bumps up nextFileNum in this case without bumping up
309 1 : // minUnflushedLogNum, even if WALs with non-zero file numbers are
310 1 : // present in the directory.
311 1 : } else {
312 0 : return base.CorruptionErrorf("pebble: malformed manifest file %q for DB %q",
313 0 : errors.Safe(manifestFilename), dirname)
314 0 : }
315 : }
316 1 : vs.markFileNumUsed(vs.minUnflushedLogNum)
317 1 :
318 1 : // Populate the virtual backings for virtual sstables since we have finished
319 1 : // version edit accumulation.
320 1 : for _, b := range bve.AddedFileBacking {
321 1 : isLocal := objstorage.IsLocalTable(provider, b.DiskFileNum)
322 1 : vs.latest.virtualBackings.AddAndRef(b, isLocal)
323 1 : }
324 :
325 1 : for l, addedLevel := range bve.AddedTables {
326 1 : for _, m := range addedLevel {
327 1 : if m.Virtual {
328 1 : vs.latest.virtualBackings.AddTable(m, l)
329 1 : }
330 : }
331 : }
332 :
333 1 : if invariants.Enabled {
334 1 : // There should be no deleted tables or backings, since we're starting from
335 1 : // an empty state.
336 1 : for _, deletedLevel := range bve.DeletedTables {
337 1 : if len(deletedLevel) != 0 {
338 0 : panic("deleted files after manifest replay")
339 : }
340 : }
341 1 : if len(bve.RemovedFileBacking) > 0 {
342 0 : panic("deleted backings after manifest replay")
343 : }
344 : }
345 :
346 1 : emptyVersion := manifest.NewInitialVersion(opts.Comparer)
347 1 : newVersion, err := bve.Apply(emptyVersion, opts.Experimental.ReadCompactionRate)
348 1 : if err != nil {
349 0 : return err
350 0 : }
351 1 : vs.latest.l0Organizer.PerformUpdate(vs.latest.l0Organizer.PrepareUpdate(&bve, newVersion), newVersion)
352 1 : vs.latest.l0Organizer.InitCompactingFileInfo(nil /* in-progress compactions */)
353 1 : vs.latest.blobFiles.Init(&bve, blobRewriteHeuristic)
354 1 : vs.append(newVersion)
355 1 :
356 1 : for i := range vs.metrics.Levels {
357 1 : l := &vs.metrics.Levels[i]
358 1 : l.TablesCount = int64(newVersion.Levels[i].Len())
359 1 : files := newVersion.Levels[i].Slice()
360 1 : l.TablesSize = int64(files.TableSizeSum())
361 1 : }
362 1 : for _, l := range newVersion.Levels {
363 1 : for f := range l.All() {
364 1 : if !f.Virtual {
365 1 : isLocal, localSize := sizeIfLocal(f.TableBacking, vs.provider)
366 1 : vs.metrics.Table.Local.LiveSize = uint64(int64(vs.metrics.Table.Local.LiveSize) + localSize)
367 1 : if isLocal {
368 1 : vs.metrics.Table.Local.LiveCount++
369 1 : }
370 : }
371 : }
372 : }
373 1 : for backing := range vs.latest.virtualBackings.All() {
374 1 : isLocal, localSize := sizeIfLocal(backing, vs.provider)
375 1 : vs.metrics.Table.Local.LiveSize = uint64(int64(vs.metrics.Table.Local.LiveSize) + localSize)
376 1 : if isLocal {
377 1 : vs.metrics.Table.Local.LiveCount++
378 1 : }
379 : }
380 :
381 1 : vs.setCompactionPicker(newCompactionPickerByScore(newVersion, &vs.latest, vs.opts, nil))
382 1 : return nil
383 : }
384 :
385 1 : func (vs *versionSet) close() error {
386 1 : if vs.manifestFile != nil {
387 1 : if err := vs.manifestFile.Close(); err != nil {
388 0 : return err
389 0 : }
390 : }
391 1 : if vs.manifestMarker != nil {
392 1 : if err := vs.manifestMarker.Close(); err != nil {
393 0 : return err
394 0 : }
395 : }
396 1 : return nil
397 : }
398 :
399 : // logLock locks the manifest for writing. The lock must be released by
400 : // a call to logUnlock.
401 : //
402 : // DB.mu must be held when calling this method, but the mutex may be dropped and
403 : // re-acquired during the course of this method.
404 1 : func (vs *versionSet) logLock() {
405 1 : // Wait for any existing writing to the manifest to complete, then mark the
406 1 : // manifest as busy.
407 1 : for vs.writing {
408 1 : // Note: writerCond.L is DB.mu, so we unlock it while we wait.
409 1 : vs.writerCond.Wait()
410 1 : }
411 1 : vs.writing = true
412 : }
413 :
414 : // logUnlock releases the lock for manifest writing.
415 : //
416 : // DB.mu must be held when calling this method.
417 1 : func (vs *versionSet) logUnlock() {
418 1 : if !vs.writing {
419 0 : vs.opts.Logger.Fatalf("MANIFEST not locked for writing")
420 0 : }
421 1 : vs.writing = false
422 1 : vs.writerCond.Signal()
423 : }
424 :
425 1 : func (vs *versionSet) logUnlockAndInvalidatePickedCompactionCache() {
426 1 : vs.pickedCompactionCache.invalidate()
427 1 : vs.logUnlock()
428 1 : }
429 :
430 : // versionUpdate is returned by the function passed to UpdateVersionLocked.
431 : //
432 : // If VE is nil, there is no update to apply (but it is not an error).
433 : type versionUpdate struct {
434 : VE *manifest.VersionEdit
435 : JobID JobID
436 : Metrics levelMetricsDelta
437 : // InProgressCompactionFn is called while DB.mu is held after the I/O part of
438 : // the update was performed. It should return any compactions that are
439 : // in-progress (excluding than the one that is being applied).
440 : InProgressCompactionsFn func() []compactionInfo
441 : ForceManifestRotation bool
442 : }
443 :
444 : // UpdateVersionLocked is used to update the current version, returning the
445 : // duration of the manifest update. If the manifest update is unsuccessful, the
446 : // duration will be unset (0).
447 : //
448 : // DB.mu must be held. UpdateVersionLocked first waits for any other version
449 : // update to complete, releasing and reacquiring DB.mu.
450 : //
451 : // UpdateVersionLocked then calls updateFn which builds a versionUpdate, while
452 : // holding DB.mu. The updateFn can release and reacquire DB.mu (it should
453 : // attempt to do as much work as possible outside of the lock).
454 : //
455 : // UpdateVersionLocked fills in the following fields of the VersionEdit:
456 : // NextFileNum, LastSeqNum, RemovedBackingTables. The removed backing tables are
457 : // those backings that are no longer used (in the new version) after applying
458 : // the edit (as per vs.virtualBackings). Other than these fields, the
459 : // VersionEdit must be complete.
460 : //
461 : // New table backing references (TableBacking.Ref) are taken as part of applying
462 : // the version edit. The state of the virtual backings (vs.virtualBackings) is
463 : // updated before logging to the manifest and installing the latest version;
464 : // this is ok because any failure in those steps is fatal.
465 : //
466 : // If updateFn returns an error, no update is applied and that same error is returned.
467 : // If versionUpdate.VE is nil, the no update is applied (and no error is returned).
468 : func (vs *versionSet) UpdateVersionLocked(
469 : updateFn func() (versionUpdate, error),
470 1 : ) (time.Duration, error) {
471 1 : vs.logLock()
472 1 : defer vs.logUnlockAndInvalidatePickedCompactionCache()
473 1 :
474 1 : vu, err := updateFn()
475 1 : if err != nil || vu.VE == nil {
476 1 : return 0, err
477 1 : }
478 :
479 1 : if !vs.writing {
480 0 : vs.opts.Logger.Fatalf("MANIFEST not locked for writing")
481 0 : }
482 :
483 1 : updateManifestStart := crtime.NowMono()
484 1 : ve := vu.VE
485 1 : if ve.MinUnflushedLogNum != 0 {
486 1 : if ve.MinUnflushedLogNum < vs.minUnflushedLogNum ||
487 1 : vs.nextFileNum.Load() <= uint64(ve.MinUnflushedLogNum) {
488 0 : panic(fmt.Sprintf("pebble: inconsistent versionEdit minUnflushedLogNum %d",
489 0 : ve.MinUnflushedLogNum))
490 : }
491 : }
492 :
493 : // This is the next manifest filenum, but if the current file is too big we
494 : // will write this ve to the next file which means what ve encodes is the
495 : // current filenum and not the next one.
496 : //
497 : // TODO(sbhola): figure out why this is correct and update comment.
498 1 : ve.NextFileNum = vs.nextFileNum.Load()
499 1 :
500 1 : // LastSeqNum is set to the current upper bound on the assigned sequence
501 1 : // numbers. Note that this is exactly the behavior of RocksDB. LastSeqNum is
502 1 : // used to initialize versionSet.logSeqNum and versionSet.visibleSeqNum on
503 1 : // replay. It must be higher than or equal to any than any sequence number
504 1 : // written to an sstable, including sequence numbers in ingested files.
505 1 : // Note that LastSeqNum is not (and cannot be) the minimum unflushed sequence
506 1 : // number. This is fallout from ingestion which allows a sequence number X to
507 1 : // be assigned to an ingested sstable even though sequence number X-1 resides
508 1 : // in an unflushed memtable. logSeqNum is the _next_ sequence number that
509 1 : // will be assigned, so subtract that by 1 to get the upper bound on the
510 1 : // last assigned sequence number.
511 1 : logSeqNum := vs.logSeqNum.Load()
512 1 : ve.LastSeqNum = logSeqNum - 1
513 1 : if logSeqNum == 0 {
514 0 : // logSeqNum is initialized to 1 in Open() if there are no previous WAL
515 0 : // or manifest records, so this case should never happen.
516 0 : vs.opts.Logger.Fatalf("logSeqNum must be a positive integer: %d", logSeqNum)
517 0 : }
518 :
519 1 : currentVersion := vs.currentVersion()
520 1 : var newVersion *manifest.Version
521 1 :
522 1 : // Generate a new manifest if we don't currently have one, or forceRotation
523 1 : // is true, or the current one is too large.
524 1 : //
525 1 : // For largeness, we do not exclusively use MaxManifestFileSize size
526 1 : // threshold since we have had incidents where due to either large keys or
527 1 : // large numbers of files, each edit results in a snapshot + write of the
528 1 : // edit. This slows the system down since each flush or compaction is
529 1 : // writing a new manifest snapshot. The primary goal of the size-based
530 1 : // rollover logic is to ensure that when reopening a DB, the number of edits
531 1 : // that need to be replayed on top of the snapshot is "sane". Rolling over
532 1 : // to a new manifest after each edit is not relevant to that goal.
533 1 : //
534 1 : // Consider the following cases:
535 1 : // - The number of live files F in the DB is roughly stable: after writing
536 1 : // the snapshot (with F files), say we require that there be enough edits
537 1 : // such that the cumulative number of files in those edits, E, be greater
538 1 : // than F. This will ensure that the total amount of time in
539 1 : // UpdateVersionLocked that is spent in snapshot writing is ~50%.
540 1 : //
541 1 : // - The number of live files F in the DB is shrinking drastically, say from
542 1 : // F to F/10: This can happen for various reasons, like wide range
543 1 : // tombstones, or large numbers of smaller than usual files that are being
544 1 : // merged together into larger files. And say the new files generated
545 1 : // during this shrinkage is insignificant compared to F/10, and so for
546 1 : // this example we will assume it is effectively 0. After this shrinking,
547 1 : // E = 0.9F, and so if we used the previous snapshot file count, F, as the
548 1 : // threshold that needs to be exceeded, we will further delay the snapshot
549 1 : // writing. Which means on DB reopen we will need to replay 0.9F edits to
550 1 : // get to a version with 0.1F files. It would be better to create a new
551 1 : // snapshot when E exceeds the number of files in the current version.
552 1 : //
553 1 : // - The number of live files F in the DB is growing via perfect ingests
554 1 : // into L6: Say we wrote the snapshot when there were F files and now we
555 1 : // have 10F files, so E = 9F. We will further delay writing a new
556 1 : // snapshot. This case can be critiqued as contrived, but we consider it
557 1 : // nonetheless.
558 1 : //
559 1 : // The logic below uses the min of the last snapshot file count and the file
560 1 : // count in the current version.
561 1 : vs.rotationHelper.AddRecord(int64(len(ve.DeletedTables) + len(ve.NewTables)))
562 1 : sizeExceeded := vs.manifest.Size() >= vs.opts.MaxManifestFileSize
563 1 : requireRotation := vu.ForceManifestRotation || vs.manifest == nil
564 1 :
565 1 : var nextSnapshotFilecount int64
566 1 : for i := range vs.metrics.Levels {
567 1 : nextSnapshotFilecount += vs.metrics.Levels[i].TablesCount
568 1 : }
569 1 : if sizeExceeded && !requireRotation {
570 1 : requireRotation = vs.rotationHelper.ShouldRotate(nextSnapshotFilecount)
571 1 : }
572 1 : var newManifestFileNum base.DiskFileNum
573 1 : var prevManifestFileSize uint64
574 1 : var newManifestVirtualBackings []*manifest.TableBacking
575 1 : if requireRotation {
576 1 : newManifestFileNum = vs.getNextDiskFileNum()
577 1 : prevManifestFileSize = uint64(vs.manifest.Size())
578 1 :
579 1 : // We want the virtual backings *before* applying the version edit, because
580 1 : // the new manifest will contain the pre-apply version plus the last version
581 1 : // edit.
582 1 : newManifestVirtualBackings = slices.Collect(vs.latest.virtualBackings.All())
583 1 : }
584 :
585 : // Grab certain values before releasing vs.mu, in case createManifest() needs
586 : // to be called.
587 1 : minUnflushedLogNum := vs.minUnflushedLogNum
588 1 : nextFileNum := vs.nextFileNum.Load()
589 1 :
590 1 : // Note: this call populates ve.RemovedBackingTables.
591 1 : zombieBackings, removedVirtualBackings, localTablesLiveDelta :=
592 1 : getZombieTablesAndUpdateVirtualBackings(ve, &vs.latest.virtualBackings, vs.provider)
593 1 :
594 1 : var l0Update manifest.L0PreparedUpdate
595 1 : if err := func() error {
596 1 : vs.mu.Unlock()
597 1 : defer vs.mu.Lock()
598 1 :
599 1 : if vs.getFormatMajorVersion() < FormatVirtualSSTables && len(ve.CreatedBackingTables) > 0 {
600 0 : return base.AssertionFailedf("MANIFEST cannot contain virtual sstable records due to format major version")
601 0 : }
602 :
603 : // Rotate the manifest if necessary. Rotating the manifest involves
604 : // creating a new file and writing an initial version edit containing a
605 : // snapshot of the current version. This initial version edit will
606 : // reflect the Version prior to the pending version edit (`ve`). Once
607 : // we've created the new manifest with the previous version state, we'll
608 : // append the version edit `ve` to the tail of the new manifest.
609 1 : if newManifestFileNum != 0 {
610 1 : if err := vs.createManifest(vs.dirname, newManifestFileNum, minUnflushedLogNum, nextFileNum, newManifestVirtualBackings); err != nil {
611 1 : vs.opts.EventListener.ManifestCreated(ManifestCreateInfo{
612 1 : JobID: int(vu.JobID),
613 1 : Path: base.MakeFilepath(vs.fs, vs.dirname, base.FileTypeManifest, newManifestFileNum),
614 1 : FileNum: newManifestFileNum,
615 1 : Err: err,
616 1 : })
617 1 : return errors.Wrap(err, "MANIFEST create failed")
618 1 : }
619 : }
620 :
621 : // Call ApplyAndUpdateVersionEdit before accumulating the version edit.
622 : // If any blob files are no longer referenced, the version edit will be
623 : // updated to explicitly record the deletion of the blob files. This can
624 : // happen here because vs.blobFiles is protected by the manifest logLock
625 : // (NOT vs.mu). We only read or write vs.blobFiles while holding the
626 : // manifest lock.
627 1 : if err := vs.latest.blobFiles.ApplyAndUpdateVersionEdit(ve); err != nil {
628 0 : return errors.Wrap(err, "MANIFEST blob files apply and update failed")
629 0 : }
630 :
631 1 : var bulkEdit manifest.BulkVersionEdit
632 1 : err := bulkEdit.Accumulate(ve)
633 1 : if err != nil {
634 0 : return errors.Wrap(err, "MANIFEST accumulate failed")
635 0 : }
636 1 : newVersion, err = bulkEdit.Apply(currentVersion, vs.opts.Experimental.ReadCompactionRate)
637 1 : if err != nil {
638 0 : return errors.Wrap(err, "MANIFEST apply failed")
639 0 : }
640 1 : l0Update = vs.latest.l0Organizer.PrepareUpdate(&bulkEdit, newVersion)
641 1 :
642 1 : w, err := vs.manifest.Next()
643 1 : if err != nil {
644 0 : return errors.Wrap(err, "MANIFEST next record write failed")
645 0 : }
646 :
647 : // NB: Any error from this point on is considered fatal as we don't know if
648 : // the MANIFEST write occurred or not. Trying to determine that is
649 : // fraught. Instead we rely on the standard recovery mechanism run when a
650 : // database is open. In particular, that mechanism generates a new MANIFEST
651 : // and ensures it is synced.
652 1 : if err := ve.Encode(w); err != nil {
653 0 : return errors.Wrap(err, "MANIFEST write failed")
654 0 : }
655 1 : if err := vs.manifest.Flush(); err != nil {
656 1 : return errors.Wrap(err, "MANIFEST flush failed")
657 1 : }
658 1 : if err := vs.manifestFile.Sync(); err != nil {
659 1 : return errors.Wrap(err, "MANIFEST sync failed")
660 1 : }
661 1 : if newManifestFileNum != 0 {
662 1 : // NB: Move() is responsible for syncing the data directory.
663 1 : if err := vs.manifestMarker.Move(base.MakeFilename(base.FileTypeManifest, newManifestFileNum)); err != nil {
664 0 : return errors.Wrap(err, "MANIFEST set current failed")
665 0 : }
666 1 : vs.opts.EventListener.ManifestCreated(ManifestCreateInfo{
667 1 : JobID: int(vu.JobID),
668 1 : Path: base.MakeFilepath(vs.fs, vs.dirname, base.FileTypeManifest, newManifestFileNum),
669 1 : FileNum: newManifestFileNum,
670 1 : })
671 : }
672 1 : return nil
673 1 : }(); err != nil {
674 1 : // Any error encountered during any of the operations in the previous
675 1 : // closure are considered fatal. Treating such errors as fatal is preferred
676 1 : // to attempting to unwind various file and b-tree reference counts, and
677 1 : // re-generating L0 sublevel metadata. This may change in the future, if
678 1 : // certain manifest / WAL operations become retryable. For more context, see
679 1 : // #1159 and #1792.
680 1 : vs.opts.Logger.Fatalf("%s", err)
681 1 : return 0, err
682 1 : }
683 :
684 1 : if requireRotation {
685 1 : // Successfully rotated.
686 1 : vs.rotationHelper.Rotate(nextSnapshotFilecount)
687 1 : }
688 : // Now that DB.mu is held again, initialize compacting file info in
689 : // L0Sublevels.
690 1 : inProgress := vu.InProgressCompactionsFn()
691 1 :
692 1 : zombieBlobs, localBlobLiveDelta := getZombieBlobFilesAndComputeLocalMetrics(ve, vs.provider)
693 1 : vs.latest.l0Organizer.PerformUpdate(l0Update, newVersion)
694 1 : vs.latest.l0Organizer.InitCompactingFileInfo(inProgressL0Compactions(inProgress))
695 1 :
696 1 : // Update the zombie objects sets first, as installation of the new version
697 1 : // will unref the previous version which could result in addObsoleteLocked
698 1 : // being called.
699 1 : for _, b := range zombieBackings {
700 1 : vs.zombieTables.Add(objectInfo{
701 1 : fileInfo: fileInfo{
702 1 : FileNum: b.backing.DiskFileNum,
703 1 : FileSize: b.backing.Size,
704 1 : },
705 1 : isLocal: b.isLocal,
706 1 : })
707 1 : }
708 1 : for _, zb := range zombieBlobs {
709 1 : vs.zombieBlobs.Add(zb)
710 1 : }
711 : // Unref the removed backings and report those that already became obsolete.
712 : // Note that the only case where we report obsolete tables here is when
713 : // VirtualBackings.Protect/Unprotect was used to keep a backing alive without
714 : // it being used in the current version.
715 1 : var obsoleteVirtualBackings manifest.ObsoleteFiles
716 1 : for _, b := range removedVirtualBackings {
717 1 : if b.backing.Unref() == 0 {
718 1 : obsoleteVirtualBackings.TableBackings = append(obsoleteVirtualBackings.TableBackings, b.backing)
719 1 : }
720 : }
721 1 : vs.addObsoleteLocked(obsoleteVirtualBackings)
722 1 :
723 1 : // Install the new version.
724 1 : vs.append(newVersion)
725 1 :
726 1 : if ve.MinUnflushedLogNum != 0 {
727 1 : vs.minUnflushedLogNum = ve.MinUnflushedLogNum
728 1 : }
729 1 : if newManifestFileNum != 0 {
730 1 : if vs.manifestFileNum != 0 {
731 1 : vs.obsoleteManifests = append(vs.obsoleteManifests, obsoleteFile{
732 1 : fileType: base.FileTypeManifest,
733 1 : fs: vs.fs,
734 1 : path: base.MakeFilepath(vs.fs, vs.dirname, base.FileTypeManifest, vs.manifestFileNum),
735 1 : fileNum: vs.manifestFileNum,
736 1 : fileSize: prevManifestFileSize,
737 1 : isLocal: true,
738 1 : })
739 1 : }
740 1 : vs.manifestFileNum = newManifestFileNum
741 : }
742 :
743 1 : vs.metrics.updateLevelMetrics(vu.Metrics)
744 1 : for i := range vs.metrics.Levels {
745 1 : l := &vs.metrics.Levels[i]
746 1 : l.TablesCount = int64(newVersion.Levels[i].Len())
747 1 : l.VirtualTablesCount = newVersion.Levels[i].NumVirtual
748 1 : l.VirtualTablesSize = newVersion.Levels[i].VirtualTableSize
749 1 : l.TablesSize = int64(newVersion.Levels[i].TableSize())
750 1 : l.EstimatedReferencesSize = newVersion.Levels[i].EstimatedReferenceSize()
751 1 : l.Sublevels = 0
752 1 : if l.TablesCount > 0 {
753 1 : l.Sublevels = 1
754 1 : }
755 1 : if invariants.Enabled {
756 1 : levelFiles := newVersion.Levels[i].Slice()
757 1 : if size := int64(levelFiles.TableSizeSum()); l.TablesSize != size {
758 0 : vs.opts.Logger.Fatalf("versionSet metrics L%d Size = %d, actual size = %d", i, l.TablesSize, size)
759 0 : }
760 1 : refSize := uint64(0)
761 1 : for f := range levelFiles.All() {
762 1 : refSize += f.EstimatedReferenceSize()
763 1 : }
764 1 : if refSize != l.EstimatedReferencesSize {
765 0 : vs.opts.Logger.Fatalf("versionSet metrics L%d EstimatedReferencesSize = %d, recomputed size = %d", i, l.EstimatedReferencesSize, refSize)
766 0 : }
767 :
768 1 : if nVirtual := levelFiles.NumVirtual(); nVirtual != l.VirtualTablesCount {
769 0 : vs.opts.Logger.Fatalf(
770 0 : "versionSet metrics L%d NumVirtual = %d, actual NumVirtual = %d",
771 0 : i, l.VirtualTablesCount, nVirtual,
772 0 : )
773 0 : }
774 1 : if vSize := levelFiles.VirtualTableSizeSum(); vSize != l.VirtualTablesSize {
775 0 : vs.opts.Logger.Fatalf(
776 0 : "versionSet metrics L%d Virtual size = %d, actual size = %d",
777 0 : i, l.VirtualTablesSize, vSize,
778 0 : )
779 0 : }
780 : }
781 : }
782 1 : vs.metrics.Levels[0].Sublevels = int32(len(newVersion.L0SublevelFiles))
783 1 : vs.metrics.Table.Local.LiveSize = uint64(int64(vs.metrics.Table.Local.LiveSize) + localTablesLiveDelta.size)
784 1 : vs.metrics.Table.Local.LiveCount = uint64(int64(vs.metrics.Table.Local.LiveCount) + localTablesLiveDelta.count)
785 1 : vs.metrics.BlobFiles.Local.LiveSize = uint64(int64(vs.metrics.BlobFiles.Local.LiveSize) + localBlobLiveDelta.size)
786 1 : vs.metrics.BlobFiles.Local.LiveCount = uint64(int64(vs.metrics.BlobFiles.Local.LiveCount) + localBlobLiveDelta.count)
787 1 :
788 1 : vs.setCompactionPicker(newCompactionPickerByScore(newVersion, &vs.latest, vs.opts, inProgress))
789 1 : if !vs.dynamicBaseLevel {
790 1 : vs.picker.forceBaseLevel1()
791 1 : }
792 :
793 1 : return updateManifestStart.Elapsed(), nil
794 : }
795 :
796 1 : func (vs *versionSet) setCompactionPicker(picker *compactionPickerByScore) {
797 1 : vs.picker = picker
798 1 : vs.curCompactionConcurrency.Store(int32(picker.getCompactionConcurrency()))
799 1 : }
800 :
801 : type tableBackingInfo struct {
802 : backing *manifest.TableBacking
803 : isLocal bool
804 : }
805 :
806 : type fileMetricDelta struct {
807 : count int64
808 : size int64
809 : }
810 :
811 : // getZombieTablesAndUpdateVirtualBackings updates the virtual backings with the
812 : // changes in the versionEdit and populates ve.RemovedBackingTables.
813 : // Returns:
814 : // - zombieBackings: all backings (physical and virtual) that will no longer be
815 : // needed when we apply ve.
816 : // - removedVirtualBackings: the virtual backings that will be removed by the
817 : // VersionEdit and which must be Unref()ed by the caller. These backings
818 : // match ve.RemovedBackingTables.
819 : // - localLiveSizeDelta: the delta in local live bytes.
820 : func getZombieTablesAndUpdateVirtualBackings(
821 : ve *manifest.VersionEdit, virtualBackings *manifest.VirtualBackings, provider objstorage.Provider,
822 1 : ) (zombieBackings, removedVirtualBackings []tableBackingInfo, localLiveDelta fileMetricDelta) {
823 1 : // First, deal with the physical tables.
824 1 : //
825 1 : // A physical backing has become unused if it is in DeletedFiles but not in
826 1 : // NewFiles or CreatedBackingTables.
827 1 : //
828 1 : // Note that for the common case where there are very few elements, the map
829 1 : // will stay on the stack.
830 1 : stillUsed := make(map[base.DiskFileNum]struct{})
831 1 : for _, nf := range ve.NewTables {
832 1 : if !nf.Meta.Virtual {
833 1 : stillUsed[nf.Meta.TableBacking.DiskFileNum] = struct{}{}
834 1 : isLocal, localFileDelta := sizeIfLocal(nf.Meta.TableBacking, provider)
835 1 : localLiveDelta.size += localFileDelta
836 1 : if isLocal {
837 1 : localLiveDelta.count++
838 1 : }
839 : }
840 : }
841 1 : for _, b := range ve.CreatedBackingTables {
842 1 : stillUsed[b.DiskFileNum] = struct{}{}
843 1 : }
844 1 : for _, m := range ve.DeletedTables {
845 1 : if !m.Virtual {
846 1 : // NB: this deleted file may also be in NewFiles or
847 1 : // CreatedBackingTables, due to a file moving between levels, or
848 1 : // becoming virtualized. In which case there is no change due to this
849 1 : // file in the localLiveSizeDelta -- the subtraction below compensates
850 1 : // for the addition.
851 1 : isLocal, localFileDelta := sizeIfLocal(m.TableBacking, provider)
852 1 : localLiveDelta.size -= localFileDelta
853 1 : if isLocal {
854 1 : localLiveDelta.count--
855 1 : }
856 1 : if _, ok := stillUsed[m.TableBacking.DiskFileNum]; !ok {
857 1 : zombieBackings = append(zombieBackings, tableBackingInfo{
858 1 : backing: m.TableBacking,
859 1 : isLocal: isLocal,
860 1 : })
861 1 : }
862 : }
863 : }
864 :
865 : // Now deal with virtual tables.
866 : //
867 : // When a virtual table moves between levels we RemoveTable() then AddTable(),
868 : // which works out.
869 1 : for _, b := range ve.CreatedBackingTables {
870 1 : isLocal, localFileDelta := sizeIfLocal(b, provider)
871 1 : virtualBackings.AddAndRef(b, isLocal)
872 1 : localLiveDelta.size += localFileDelta
873 1 : if isLocal {
874 1 : localLiveDelta.count++
875 1 : }
876 : }
877 1 : for _, m := range ve.DeletedTables {
878 1 : if m.Virtual {
879 1 : virtualBackings.RemoveTable(m.TableBacking.DiskFileNum, m.TableNum)
880 1 : }
881 : }
882 1 : for _, nf := range ve.NewTables {
883 1 : if nf.Meta.Virtual {
884 1 : virtualBackings.AddTable(nf.Meta, nf.Level)
885 1 : }
886 : }
887 :
888 1 : if unused := virtualBackings.Unused(); len(unused) > 0 {
889 1 : // Virtual backings that are no longer used are zombies and are also added
890 1 : // to RemovedBackingTables (before the version edit is written to disk).
891 1 : ve.RemovedBackingTables = make([]base.DiskFileNum, len(unused))
892 1 : for i, b := range unused {
893 1 : isLocal, localFileDelta := sizeIfLocal(b, provider)
894 1 : localLiveDelta.size -= localFileDelta
895 1 : if isLocal {
896 1 : localLiveDelta.count--
897 1 : }
898 1 : ve.RemovedBackingTables[i] = b.DiskFileNum
899 1 : zombieBackings = append(zombieBackings, tableBackingInfo{
900 1 : backing: b,
901 1 : isLocal: isLocal,
902 1 : })
903 1 : virtualBackings.Remove(b.DiskFileNum)
904 : }
905 1 : removedVirtualBackings = zombieBackings[len(zombieBackings)-len(unused):]
906 : }
907 1 : return zombieBackings, removedVirtualBackings, localLiveDelta
908 : }
909 :
910 : // getZombieBlobFilesAndComputeLocalMetrics constructs objectInfos for all
911 : // zombie blob files, and computes the metric deltas for live files overall and
912 : // locally.
913 : func getZombieBlobFilesAndComputeLocalMetrics(
914 : ve *manifest.VersionEdit, provider objstorage.Provider,
915 1 : ) (zombieBlobFiles []objectInfo, localLiveDelta fileMetricDelta) {
916 1 : for _, b := range ve.NewBlobFiles {
917 1 : if objstorage.IsLocalBlobFile(provider, b.Physical.FileNum) {
918 1 : localLiveDelta.count++
919 1 : localLiveDelta.size += int64(b.Physical.Size)
920 1 : }
921 : }
922 1 : zombieBlobFiles = make([]objectInfo, 0, len(ve.DeletedBlobFiles))
923 1 : for _, physical := range ve.DeletedBlobFiles {
924 1 : isLocal := objstorage.IsLocalBlobFile(provider, physical.FileNum)
925 1 : if isLocal {
926 1 : localLiveDelta.count--
927 1 : localLiveDelta.size -= int64(physical.Size)
928 1 : }
929 1 : zombieBlobFiles = append(zombieBlobFiles, objectInfo{
930 1 : fileInfo: fileInfo{
931 1 : FileNum: physical.FileNum,
932 1 : FileSize: physical.Size,
933 1 : },
934 1 : isLocal: isLocal,
935 1 : })
936 : }
937 1 : return zombieBlobFiles, localLiveDelta
938 : }
939 :
940 : // sizeIfLocal returns backing.Size if the backing is a local file, else 0.
941 : func sizeIfLocal(
942 : backing *manifest.TableBacking, provider objstorage.Provider,
943 1 : ) (isLocal bool, localSize int64) {
944 1 : isLocal = objstorage.IsLocalTable(provider, backing.DiskFileNum)
945 1 : if isLocal {
946 1 : return true, int64(backing.Size)
947 1 : }
948 1 : return false, 0
949 : }
950 :
951 : func (vs *versionSet) incrementCompactions(
952 : kind compactionKind, extraLevels []*compactionLevel, bytesWritten int64, compactionErr error,
953 1 : ) {
954 1 : if kind == compactionKindFlush || kind == compactionKindIngestedFlushable {
955 1 : vs.metrics.Flush.Count++
956 1 : } else {
957 1 : vs.metrics.Compact.Count++
958 1 : if compactionErr != nil {
959 1 : if errors.Is(compactionErr, ErrCancelledCompaction) {
960 1 : vs.metrics.Compact.CancelledCount++
961 1 : vs.metrics.Compact.CancelledBytes += bytesWritten
962 1 : } else {
963 1 : vs.metrics.Compact.FailedCount++
964 1 : }
965 : }
966 : }
967 :
968 1 : switch kind {
969 1 : case compactionKindDefault:
970 1 : vs.metrics.Compact.DefaultCount++
971 :
972 1 : case compactionKindFlush, compactionKindIngestedFlushable:
973 :
974 1 : case compactionKindMove:
975 1 : vs.metrics.Compact.MoveCount++
976 :
977 1 : case compactionKindDeleteOnly:
978 1 : vs.metrics.Compact.DeleteOnlyCount++
979 :
980 1 : case compactionKindElisionOnly:
981 1 : vs.metrics.Compact.ElisionOnlyCount++
982 :
983 1 : case compactionKindRead:
984 1 : vs.metrics.Compact.ReadCount++
985 :
986 1 : case compactionKindTombstoneDensity:
987 1 : vs.metrics.Compact.TombstoneDensityCount++
988 :
989 1 : case compactionKindRewrite:
990 1 : vs.metrics.Compact.RewriteCount++
991 :
992 1 : case compactionKindCopy:
993 1 : vs.metrics.Compact.CopyCount++
994 :
995 1 : case compactionKindBlobFileRewrite:
996 1 : vs.metrics.Compact.BlobFileRewriteCount++
997 :
998 1 : case compactionKindVirtualRewrite:
999 1 : vs.metrics.Compact.VirtualRewriteCount++
1000 :
1001 0 : default:
1002 0 : if invariants.Enabled {
1003 0 : panic("unhandled compaction kind")
1004 : }
1005 :
1006 : }
1007 1 : if len(extraLevels) > 0 {
1008 1 : vs.metrics.Compact.MultiLevelCount++
1009 1 : }
1010 : }
1011 :
1012 1 : func (vs *versionSet) incrementCompactionBytes(numBytes int64) {
1013 1 : vs.atomicInProgressBytes.Add(numBytes)
1014 1 : }
1015 :
1016 : // createManifest creates a manifest file that contains a snapshot of vs.
1017 : func (vs *versionSet) createManifest(
1018 : dirname string,
1019 : fileNum, minUnflushedLogNum base.DiskFileNum,
1020 : nextFileNum uint64,
1021 : virtualBackings []*manifest.TableBacking,
1022 1 : ) (err error) {
1023 1 : var (
1024 1 : filename = base.MakeFilepath(vs.fs, dirname, base.FileTypeManifest, fileNum)
1025 1 : manifestFile vfs.File
1026 1 : manifestWriter *record.Writer
1027 1 : )
1028 1 : defer func() {
1029 1 : if manifestWriter != nil {
1030 0 : _ = manifestWriter.Close()
1031 0 : }
1032 1 : if manifestFile != nil {
1033 0 : _ = manifestFile.Close()
1034 0 : }
1035 1 : if err != nil {
1036 1 : _ = vs.fs.Remove(filename)
1037 1 : }
1038 : }()
1039 1 : manifestFile, err = vs.fs.Create(filename, "pebble-manifest")
1040 1 : if err != nil {
1041 1 : return err
1042 1 : }
1043 1 : manifestWriter = record.NewWriter(manifestFile)
1044 1 :
1045 1 : snapshot := manifest.VersionEdit{
1046 1 : ComparerName: vs.cmp.Name,
1047 1 : // When creating a version snapshot for an existing DB, this snapshot
1048 1 : // VersionEdit will be immediately followed by another VersionEdit
1049 1 : // (being written in UpdateVersionLocked()). That VersionEdit always
1050 1 : // contains a LastSeqNum, so we don't need to include that in the
1051 1 : // snapshot. But it does not necessarily include MinUnflushedLogNum,
1052 1 : // NextFileNum, so we initialize those using the corresponding fields in
1053 1 : // the versionSet (which came from the latest preceding VersionEdit that
1054 1 : // had those fields).
1055 1 : MinUnflushedLogNum: minUnflushedLogNum,
1056 1 : NextFileNum: nextFileNum,
1057 1 : CreatedBackingTables: virtualBackings,
1058 1 : NewBlobFiles: vs.latest.blobFiles.Metadatas(),
1059 1 : }
1060 1 :
1061 1 : // Add all extant sstables in the current version.
1062 1 : for level, levelMetadata := range vs.currentVersion().Levels {
1063 1 : for meta := range levelMetadata.All() {
1064 1 : snapshot.NewTables = append(snapshot.NewTables, manifest.NewTableEntry{
1065 1 : Level: level,
1066 1 : Meta: meta,
1067 1 : })
1068 1 : }
1069 : }
1070 :
1071 : // Add all tables marked for compaction.
1072 1 : if vs.getFormatMajorVersion() >= FormatMarkForCompactionInVersionEdit {
1073 1 : for meta, level := range vs.currentVersion().MarkedForCompaction.Ascending() {
1074 1 : snapshot.TablesMarkedForCompaction = append(snapshot.TablesMarkedForCompaction, manifest.TableMarkedForCompactionEntry{
1075 1 : Level: level,
1076 1 : TableNum: meta.TableNum,
1077 1 : })
1078 1 : }
1079 : }
1080 :
1081 1 : w, err1 := manifestWriter.Next()
1082 1 : if err1 != nil {
1083 0 : return err1
1084 0 : }
1085 1 : if err := snapshot.Encode(w); err != nil {
1086 0 : return err
1087 0 : }
1088 :
1089 1 : if vs.manifest != nil {
1090 1 : if err := vs.manifest.Close(); err != nil {
1091 0 : return err
1092 0 : }
1093 1 : vs.manifest = nil
1094 : }
1095 1 : if vs.manifestFile != nil {
1096 1 : if err := vs.manifestFile.Close(); err != nil {
1097 0 : return err
1098 0 : }
1099 1 : vs.manifestFile = nil
1100 : }
1101 :
1102 1 : vs.manifest, manifestWriter = manifestWriter, nil
1103 1 : vs.manifestFile, manifestFile = manifestFile, nil
1104 1 : return nil
1105 : }
1106 :
1107 : // NB: This method is not safe for concurrent use. It is only safe
1108 : // to be called when concurrent changes to nextFileNum are not expected.
1109 1 : func (vs *versionSet) markFileNumUsed(fileNum base.DiskFileNum) {
1110 1 : if vs.nextFileNum.Load() <= uint64(fileNum) {
1111 1 : vs.nextFileNum.Store(uint64(fileNum + 1))
1112 1 : }
1113 : }
1114 :
1115 : // getNextTableNum returns a new table number.
1116 : //
1117 : // Can be called without the versionSet's mutex being held.
1118 1 : func (vs *versionSet) getNextTableNum() base.TableNum {
1119 1 : x := vs.nextFileNum.Add(1) - 1
1120 1 : return base.TableNum(x)
1121 1 : }
1122 :
1123 : // Can be called without the versionSet's mutex being held.
1124 1 : func (vs *versionSet) getNextDiskFileNum() base.DiskFileNum {
1125 1 : x := vs.nextFileNum.Add(1) - 1
1126 1 : return base.DiskFileNum(x)
1127 1 : }
1128 :
1129 1 : func (vs *versionSet) append(v *manifest.Version) {
1130 1 : if v.Refs() != 0 {
1131 0 : panic("pebble: version should be unreferenced")
1132 : }
1133 1 : if !vs.versions.Empty() {
1134 1 : vs.versions.Back().UnrefLocked()
1135 1 : }
1136 1 : v.Deleted = vs.obsoleteFn
1137 1 : v.Ref()
1138 1 : vs.versions.PushBack(v)
1139 1 : if invariants.Enabled {
1140 1 : // Verify that the virtualBackings contains all the backings referenced by
1141 1 : // the version.
1142 1 : for _, l := range v.Levels {
1143 1 : for f := range l.All() {
1144 1 : if f.Virtual {
1145 1 : if _, ok := vs.latest.virtualBackings.Get(f.TableBacking.DiskFileNum); !ok {
1146 0 : panic(fmt.Sprintf("%s is not in virtualBackings", f.TableBacking.DiskFileNum))
1147 : }
1148 : }
1149 : }
1150 : }
1151 : }
1152 : }
1153 :
1154 1 : func (vs *versionSet) currentVersion() *manifest.Version {
1155 1 : return vs.versions.Back()
1156 1 : }
1157 :
1158 1 : func (vs *versionSet) addLiveFileNums(m map[base.DiskFileNum]struct{}) {
1159 1 : current := vs.currentVersion()
1160 1 : for v := vs.versions.Front(); true; v = v.Next() {
1161 1 : for _, lm := range v.Levels {
1162 1 : for f := range lm.All() {
1163 1 : m[f.TableBacking.DiskFileNum] = struct{}{}
1164 1 : }
1165 : }
1166 1 : for bf := range v.BlobFiles.All() {
1167 1 : m[bf.Physical.FileNum] = struct{}{}
1168 1 : }
1169 1 : if v == current {
1170 1 : break
1171 : }
1172 : }
1173 : // virtualBackings contains backings that are referenced by some virtual
1174 : // tables in the latest version (which are handled above), and backings that
1175 : // are not but are still alive because of the protection mechanism (see
1176 : // manifset.VirtualBackings). This loop ensures the latter get added to the
1177 : // map.
1178 1 : for b := range vs.latest.virtualBackings.All() {
1179 1 : m[b.DiskFileNum] = struct{}{}
1180 1 : }
1181 : }
1182 :
1183 : // addObsoleteLocked will add the fileInfo associated with obsolete backing
1184 : // sstables to the obsolete tables list.
1185 : //
1186 : // The file backings in the obsolete list must not appear more than once.
1187 : //
1188 : // DB.mu must be held when addObsoleteLocked is called.
1189 1 : func (vs *versionSet) addObsoleteLocked(obsolete manifest.ObsoleteFiles) {
1190 1 : if obsolete.Count() == 0 {
1191 1 : return
1192 1 : }
1193 :
1194 : // Note that the zombie objects transition from zombie *to* obsolete, and
1195 : // will no longer be considered zombie.
1196 :
1197 1 : newlyObsoleteTables := make([]obsoleteFile, len(obsolete.TableBackings))
1198 1 : for i, bs := range obsolete.TableBackings {
1199 1 : newlyObsoleteTables[i] = vs.zombieTables.Extract(bs.DiskFileNum).
1200 1 : asObsoleteFile(vs.fs, base.FileTypeTable, vs.dirname)
1201 1 : }
1202 1 : vs.obsoleteTables = mergeObsoleteFiles(vs.obsoleteTables, newlyObsoleteTables)
1203 1 :
1204 1 : newlyObsoleteBlobFiles := make([]obsoleteFile, len(obsolete.BlobFiles))
1205 1 : for i, bf := range obsolete.BlobFiles {
1206 1 : newlyObsoleteBlobFiles[i] = vs.zombieBlobs.Extract(bf.FileNum).
1207 1 : asObsoleteFile(vs.fs, base.FileTypeBlob, vs.dirname)
1208 1 : }
1209 1 : vs.obsoleteBlobs = mergeObsoleteFiles(vs.obsoleteBlobs, newlyObsoleteBlobFiles)
1210 1 : vs.updateObsoleteObjectMetricsLocked()
1211 : }
1212 :
1213 : // addObsolete will acquire DB.mu, so DB.mu must not be held when this is
1214 : // called.
1215 1 : func (vs *versionSet) addObsolete(obsolete manifest.ObsoleteFiles) {
1216 1 : vs.mu.Lock()
1217 1 : defer vs.mu.Unlock()
1218 1 : vs.addObsoleteLocked(obsolete)
1219 1 : }
1220 :
1221 1 : func (vs *versionSet) updateObsoleteObjectMetricsLocked() {
1222 1 : // TODO(jackson): Ideally we would update vs.fileDeletions.queuedStats to
1223 1 : // include the files on vs.obsolete{Tables,Blobs}, but there's subtlety in
1224 1 : // deduplicating the files before computing the stats. It might also be
1225 1 : // possible to refactor to remove the vs.obsolete{Tables,Blobs} intermediary
1226 1 : // step. Revisit this.
1227 1 : vs.metrics.Table.ObsoleteCount = int64(len(vs.obsoleteTables))
1228 1 : vs.metrics.Table.ObsoleteSize = 0
1229 1 : vs.metrics.Table.Local.ObsoleteSize = 0
1230 1 : vs.metrics.Table.Local.ObsoleteCount = 0
1231 1 : for _, fi := range vs.obsoleteTables {
1232 1 : vs.metrics.Table.ObsoleteSize += fi.fileSize
1233 1 : if fi.isLocal {
1234 1 : vs.metrics.Table.Local.ObsoleteSize += fi.fileSize
1235 1 : vs.metrics.Table.Local.ObsoleteCount++
1236 1 : }
1237 : }
1238 1 : vs.metrics.BlobFiles.ObsoleteCount = uint64(len(vs.obsoleteBlobs))
1239 1 : vs.metrics.BlobFiles.ObsoleteSize = 0
1240 1 : vs.metrics.BlobFiles.Local.ObsoleteSize = 0
1241 1 : vs.metrics.BlobFiles.Local.ObsoleteCount = 0
1242 1 : for _, fi := range vs.obsoleteBlobs {
1243 1 : vs.metrics.BlobFiles.ObsoleteSize += fi.fileSize
1244 1 : if fi.isLocal {
1245 1 : vs.metrics.BlobFiles.Local.ObsoleteSize += fi.fileSize
1246 1 : vs.metrics.BlobFiles.Local.ObsoleteCount++
1247 1 : }
1248 : }
1249 : }
1250 :
1251 : func findCurrentManifest(
1252 : fs vfs.FS, dirname string, ls []string,
1253 1 : ) (marker *atomicfs.Marker, manifestNum base.DiskFileNum, exists bool, err error) {
1254 1 : // Locating a marker should succeed even if the marker has never been placed.
1255 1 : var filename string
1256 1 : marker, filename, err = atomicfs.LocateMarkerInListing(fs, dirname, manifestMarkerName, ls)
1257 1 : if err != nil {
1258 1 : return nil, 0, false, err
1259 1 : }
1260 :
1261 1 : if filename == "" {
1262 1 : // The marker hasn't been set yet. This database doesn't exist.
1263 1 : return marker, 0, false, nil
1264 1 : }
1265 :
1266 1 : var ok bool
1267 1 : _, manifestNum, ok = base.ParseFilename(fs, filename)
1268 1 : if !ok {
1269 0 : return marker, 0, false, base.CorruptionErrorf("pebble: MANIFEST name %q is malformed", errors.Safe(filename))
1270 0 : }
1271 1 : return marker, manifestNum, true, nil
1272 : }
1273 :
1274 1 : func newFileMetrics(newFiles []manifest.NewTableEntry) levelMetricsDelta {
1275 1 : var m levelMetricsDelta
1276 1 : for _, nf := range newFiles {
1277 1 : lm := m[nf.Level]
1278 1 : if lm == nil {
1279 1 : lm = &LevelMetrics{}
1280 1 : m[nf.Level] = lm
1281 1 : }
1282 1 : lm.TablesCount++
1283 1 : lm.TablesSize += int64(nf.Meta.Size)
1284 1 : lm.EstimatedReferencesSize += nf.Meta.EstimatedReferenceSize()
1285 : }
1286 1 : return m
1287 : }
|
