Line data Source code
1 : // Copyright 2021 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 : "strconv"
10 :
11 : "github.com/cockroachdb/errors"
12 : "github.com/cockroachdb/pebble/internal/manifest"
13 : "github.com/cockroachdb/pebble/sstable"
14 : "github.com/cockroachdb/pebble/vfs"
15 : "github.com/cockroachdb/pebble/vfs/atomicfs"
16 : )
17 :
18 : // FormatMajorVersion is a constant controlling the format of persisted
19 : // data. Backwards incompatible changes to durable formats are gated
20 : // behind new format major versions.
21 : //
22 : // At any point, a database's format major version may be bumped.
23 : // However, once a database's format major version is increased,
24 : // previous versions of Pebble will refuse to open the database.
25 : //
26 : // The zero value format is the FormatDefault constant. The exact
27 : // FormatVersion that the default corresponds to may change with time.
28 : type FormatMajorVersion uint64
29 :
30 : // SafeValue implements redact.SafeValue.
31 0 : func (v FormatMajorVersion) SafeValue() {}
32 :
33 : // String implements fmt.Stringer.
34 2 : func (v FormatMajorVersion) String() string {
35 2 : // NB: This must not change. It's used as the value for the on-disk
36 2 : // version marker file.
37 2 : //
38 2 : // Specifically, this value must always parse as a base 10 integer
39 2 : // that fits in a uint64. We format it as zero-padded, 3-digit
40 2 : // number today, but the padding may change.
41 2 : return fmt.Sprintf("%03d", v)
42 2 : }
43 :
44 : const (
45 : // FormatDefault leaves the format version unspecified. When used to create a
46 : // new store, Pebble will choose the earliest format version it supports.
47 : FormatDefault FormatMajorVersion = iota
48 :
49 : // 21.2 versions.
50 :
51 : // FormatMostCompatible maintains the most backwards compatibility,
52 : // maintaining bi-directional compatibility with RocksDB 6.2.1 in
53 : // the particular configuration described in the Pebble README.
54 : // Deprecated.
55 : _ // FormatMostCompatible
56 :
57 : // formatVersionedManifestMarker is the first
58 : // backwards-incompatible change made to Pebble, introducing the
59 : // format-version marker file for handling backwards-incompatible
60 : // changes more broadly, and replacing the `CURRENT` file with a
61 : // marker file.
62 : //
63 : // This format version is intended as an intermediary version state.
64 : // It is deliberately unexported to discourage direct use of this
65 : // format major version. Clients should use FormatVersioned which
66 : // also ensures earlier versions of Pebble fail to open a database
67 : // written in a future format major version.
68 : // Deprecated.
69 : _ // formatVersionedManifestMarker
70 :
71 : // FormatVersioned is a new format major version that replaces the
72 : // old `CURRENT` file with a new 'marker' file scheme. Previous
73 : // Pebble versions will be unable to open the database unless
74 : // they're aware of format versions.
75 : // Deprecated.
76 : _ // FormatVersioned
77 :
78 : // FormatSetWithDelete is a format major version that introduces a new key
79 : // kind, base.InternalKeyKindSetWithDelete. Previous Pebble versions will be
80 : // unable to open this database.
81 : // Deprecated.
82 : _ // FormatSetWithDelete
83 :
84 : // 22.1 versions.
85 :
86 : // FormatBlockPropertyCollector is a format major version that introduces
87 : // BlockPropertyCollectors.
88 : // Deprecated.
89 : _ // FormatBlockPropertyCollector
90 :
91 : // FormatSplitUserKeysMarked is a format major version that guarantees that
92 : // all files that share user keys with neighbors are marked for compaction
93 : // in the manifest. Ratcheting to FormatSplitUserKeysMarked will block
94 : // (without holding mutexes) until the scan of the LSM is complete and the
95 : // manifest has been rotated.
96 : // Deprecated.
97 : _ // FormatSplitUserKeysMarked
98 :
99 : // 22.2 versions.
100 :
101 : // FormatSplitUserKeysMarkedCompacted is a format major version that
102 : // guarantees that all files explicitly marked for compaction in the manifest
103 : // have been compacted. Combined with the FormatSplitUserKeysMarked format
104 : // major version, this version guarantees that there are no user keys split
105 : // across multiple files within a level L1+. Ratcheting to this format version
106 : // will block (without holding mutexes) until all necessary compactions for
107 : // files marked for compaction are complete.
108 : // Deprecated.
109 : _ // FormatSplitUserKeysMarkedCompacted
110 :
111 : // FormatRangeKeys is a format major version that introduces range keys.
112 : // Deprecated.
113 : _ // FormatRangeKeys
114 :
115 : // FormatMinTableFormatPebblev1 is a format major version that guarantees that
116 : // tables created by or ingested into the DB at or above this format major
117 : // version will have a table format version of at least Pebblev1 (Block
118 : // Properties).
119 : // Deprecated.
120 : _ // FormatMinTableFormatPebblev1
121 :
122 : // FormatPrePebblev1Marked is a format major version that guarantees that all
123 : // sstables with a table format version pre-Pebblev1 (i.e. those that are
124 : // guaranteed to not contain block properties) are marked for compaction in
125 : // the manifest. Ratcheting to FormatPrePebblev1Marked will block (without
126 : // holding mutexes) until the scan of the LSM is complete and the manifest has
127 : // been rotated.
128 : // Deprecated.
129 : _ // FormatPrePebblev1Marked
130 :
131 : // 23.1 versions.
132 :
133 : // formatUnusedPrePebblev1MarkedCompacted is an unused format major version.
134 : // This format major version was originally intended to ship in the 23.1
135 : // release. It was later decided that this should be deferred until a
136 : // subsequent release. The original ordering is preserved so as not to
137 : // introduce breaking changes in Cockroach.
138 : _ // formatUnusedPrePebblev1MarkedCompacted
139 :
140 : // FormatSSTableValueBlocks is a format major version that adds support for
141 : // storing values in value blocks in the sstable. Value block support is not
142 : // necessarily enabled when writing sstables, when running with this format
143 : // major version.
144 : _ // FormatSSTableValueBlocks
145 :
146 : // FormatFlushableIngest is a format major version that enables lazy
147 : // addition of ingested sstables into the LSM structure. When an ingest
148 : // overlaps with a memtable, a record of the ingest is written to the WAL
149 : // without waiting for a flush. Subsequent reads treat the ingested files as
150 : // a level above the overlapping memtable. Once the memtable is flushed, the
151 : // ingested files are moved into the lowest possible levels.
152 : //
153 : // This feature is behind a format major version because it required
154 : // breaking changes to the WAL format.
155 : FormatFlushableIngest
156 :
157 : // 23.2 versions.
158 :
159 : // FormatPrePebblev1MarkedCompacted is a format major version that guarantees
160 : // that all sstables explicitly marked for compaction in the manifest (see
161 : // FormatPrePebblev1Marked) have been compacted. Ratcheting to this format
162 : // version will block (without holding mutexes) until all necessary
163 : // compactions for files marked for compaction are complete.
164 : FormatPrePebblev1MarkedCompacted
165 :
166 : // FormatDeleteSizedAndObsolete is a format major version that adds support
167 : // for deletion tombstones that encode the size of the value they're
168 : // expected to delete. This format major version is required before the
169 : // associated key kind may be committed through batch applications or
170 : // ingests. It also adds support for keys that are marked obsolete (see
171 : // sstable/format.go for details).
172 : FormatDeleteSizedAndObsolete
173 :
174 : // FormatVirtualSSTables is a format major version that adds support for
175 : // virtual sstables that can reference a sub-range of keys in an underlying
176 : // physical sstable. This information is persisted through new,
177 : // backward-incompatible fields in the Manifest, and therefore requires
178 : // a format major version.
179 : FormatVirtualSSTables
180 :
181 : // FormatSyntheticPrefixes is a format major version that adds support for
182 : // sstables to have their content exposed in a different prefix of keyspace
183 : // than the actual prefix persisted in the keys in such sstables. The prefix
184 : // replacement information is stored in new fields in the Manifest and thus
185 : // requires a format major version.
186 : FormatSyntheticPrefixes
187 :
188 : // -- Add new versions here --
189 :
190 : // FormatNewest is the most recent format major version.
191 : FormatNewest FormatMajorVersion = iota - 1
192 :
193 : // Experimental versions, which are excluded by FormatNewest (but can be used
194 : // in tests) can be defined here.
195 :
196 : // -- Add experimental versions here --
197 :
198 : // internalFormatNewest is the most recent, possibly experimental format major
199 : // version.
200 : internalFormatNewest FormatMajorVersion = iota - 2
201 : )
202 :
203 : // FormatMinSupported is the minimum format version that is supported by this
204 : // Pebble version.
205 : const FormatMinSupported = FormatFlushableIngest
206 :
207 : // FormatMinForSharedObjects it the minimum format version that supports shared
208 : // objects (see CreateOnShared option).
209 : const FormatMinForSharedObjects = FormatVirtualSSTables
210 :
211 : // IsSupported returns true if the version is supported by the current Pebble
212 : // version.
213 0 : func (v FormatMajorVersion) IsSupported() bool {
214 0 : return v == FormatDefault && v >= FormatMinSupported && v <= internalFormatNewest
215 0 : }
216 :
217 : // MaxTableFormat returns the maximum sstable.TableFormat that can be used at
218 : // this FormatMajorVersion.
219 2 : func (v FormatMajorVersion) MaxTableFormat() sstable.TableFormat {
220 2 : switch v {
221 2 : case FormatDefault, FormatFlushableIngest, FormatPrePebblev1MarkedCompacted:
222 2 : return sstable.TableFormatPebblev3
223 2 : case FormatDeleteSizedAndObsolete, FormatVirtualSSTables, FormatSyntheticPrefixes:
224 2 : return sstable.TableFormatPebblev4
225 1 : default:
226 1 : panic(fmt.Sprintf("pebble: unsupported format major version: %s", v))
227 : }
228 : }
229 :
230 : // MinTableFormat returns the minimum sstable.TableFormat that can be used at
231 : // this FormatMajorVersion.
232 2 : func (v FormatMajorVersion) MinTableFormat() sstable.TableFormat {
233 2 : switch v {
234 : case FormatDefault, FormatFlushableIngest, FormatPrePebblev1MarkedCompacted,
235 2 : FormatDeleteSizedAndObsolete, FormatVirtualSSTables, FormatSyntheticPrefixes:
236 2 : return sstable.TableFormatPebblev1
237 1 : default:
238 1 : panic(fmt.Sprintf("pebble: unsupported format major version: %s", v))
239 : }
240 : }
241 :
242 : // formatMajorVersionMigrations defines the migrations from one format
243 : // major version to the next. Each migration is defined as a closure
244 : // which will be invoked on the database before the new format major
245 : // version is committed. Migrations must be idempotent. Migrations are
246 : // invoked with d.mu locked.
247 : //
248 : // Each migration is responsible for invoking finalizeFormatVersUpgrade
249 : // to set the new format major version. RatchetFormatMajorVersion will
250 : // panic if a migration returns a nil error but fails to finalize the
251 : // new format major version.
252 : var formatMajorVersionMigrations = map[FormatMajorVersion]func(*DB) error{
253 0 : FormatFlushableIngest: func(d *DB) error { return nil },
254 2 : FormatPrePebblev1MarkedCompacted: func(d *DB) error {
255 2 : // Before finalizing the format major version, rewrite any sstables
256 2 : // still marked for compaction. Note all format major versions
257 2 : // migrations are invoked with DB.mu locked.
258 2 : if err := d.compactMarkedFilesLocked(); err != nil {
259 0 : return err
260 0 : }
261 2 : return d.finalizeFormatVersUpgrade(FormatPrePebblev1MarkedCompacted)
262 : },
263 2 : FormatDeleteSizedAndObsolete: func(d *DB) error {
264 2 : return d.finalizeFormatVersUpgrade(FormatDeleteSizedAndObsolete)
265 2 : },
266 2 : FormatVirtualSSTables: func(d *DB) error {
267 2 : return d.finalizeFormatVersUpgrade(FormatVirtualSSTables)
268 2 : },
269 2 : FormatSyntheticPrefixes: func(d *DB) error {
270 2 : return d.finalizeFormatVersUpgrade(FormatSyntheticPrefixes)
271 2 : },
272 : }
273 :
274 : const formatVersionMarkerName = `format-version`
275 :
276 : // lookupFormatMajorVersion retrieves the format version from the format version
277 : // marker file.
278 : //
279 : // If such a file does not exist, returns FormatDefault. Note that this case is
280 : // only acceptable if we are creating a new store (we no longer support
281 : // FormatMostCompatible which is the only one with no version marker file).
282 : func lookupFormatMajorVersion(
283 : fs vfs.FS, dirname string,
284 2 : ) (FormatMajorVersion, *atomicfs.Marker, error) {
285 2 : m, versString, err := atomicfs.LocateMarker(fs, dirname, formatVersionMarkerName)
286 2 : if err != nil {
287 1 : return 0, nil, err
288 1 : }
289 2 : if versString == "" {
290 2 : return FormatDefault, m, nil
291 2 : }
292 2 : v, err := strconv.ParseUint(versString, 10, 64)
293 2 : if err != nil {
294 0 : return 0, nil, errors.Wrap(err, "parsing format major version")
295 0 : }
296 2 : vers := FormatMajorVersion(v)
297 2 : if vers == FormatDefault {
298 0 : return 0, nil, errors.Newf("pebble: default format major version should not persisted", vers)
299 0 : }
300 2 : if vers > internalFormatNewest {
301 1 : return 0, nil, errors.Newf("pebble: database %q written in unknown format major version %d", dirname, vers)
302 1 : }
303 2 : if vers < FormatMinSupported {
304 0 : return 0, nil, errors.Newf("pebble: database %q written in format major version %d which is no longer supported", dirname, vers)
305 0 : }
306 2 : return vers, m, nil
307 : }
308 :
309 : // FormatMajorVersion returns the database's active format major
310 : // version. The format major version may be higher than the one
311 : // provided in Options when the database was opened if the existing
312 : // database was written with a higher format version.
313 2 : func (d *DB) FormatMajorVersion() FormatMajorVersion {
314 2 : return FormatMajorVersion(d.mu.formatVers.vers.Load())
315 2 : }
316 :
317 : // RatchetFormatMajorVersion ratchets the opened database's format major
318 : // version to the provided version. It errors if the provided format
319 : // major version is below the database's current version. Once a
320 : // database's format major version is upgraded, previous Pebble versions
321 : // that do not know of the format version will be unable to open the
322 : // database.
323 2 : func (d *DB) RatchetFormatMajorVersion(fmv FormatMajorVersion) error {
324 2 : if err := d.closed.Load(); err != nil {
325 1 : panic(err)
326 : }
327 :
328 2 : d.mu.Lock()
329 2 : defer d.mu.Unlock()
330 2 : return d.ratchetFormatMajorVersionLocked(fmv)
331 : }
332 :
333 2 : func (d *DB) ratchetFormatMajorVersionLocked(formatVers FormatMajorVersion) error {
334 2 : if d.opts.ReadOnly {
335 0 : return ErrReadOnly
336 0 : }
337 2 : if formatVers > internalFormatNewest {
338 0 : // Guard against accidentally forgetting to update internalFormatNewest.
339 0 : return errors.Errorf("pebble: unknown format version %d", formatVers)
340 0 : }
341 2 : if currentVers := d.FormatMajorVersion(); currentVers > formatVers {
342 0 : return errors.Newf("pebble: database already at format major version %d; cannot reduce to %d",
343 0 : currentVers, formatVers)
344 0 : }
345 2 : if d.mu.formatVers.ratcheting {
346 0 : return errors.Newf("pebble: database format major version upgrade is in-progress")
347 0 : }
348 2 : d.mu.formatVers.ratcheting = true
349 2 : defer func() { d.mu.formatVers.ratcheting = false }()
350 :
351 2 : for nextVers := d.FormatMajorVersion() + 1; nextVers <= formatVers; nextVers++ {
352 2 : if err := formatMajorVersionMigrations[nextVers](d); err != nil {
353 0 : return errors.Wrapf(err, "migrating to version %d", nextVers)
354 0 : }
355 :
356 : // NB: The migration is responsible for calling
357 : // finalizeFormatVersUpgrade to finalize the upgrade. This
358 : // structure is necessary because some migrations may need to
359 : // update in-memory state (without ever dropping locks) after
360 : // the upgrade is finalized. Here we assert that the upgrade
361 : // did occur.
362 2 : if d.FormatMajorVersion() != nextVers {
363 0 : d.opts.Logger.Fatalf("pebble: successful migration to format version %d never finalized the upgrade", nextVers)
364 0 : }
365 : }
366 2 : return nil
367 : }
368 :
369 : // finalizeFormatVersUpgrade is typically only be called from within a
370 : // format major version migration.
371 : //
372 : // See formatMajorVersionMigrations.
373 2 : func (d *DB) finalizeFormatVersUpgrade(formatVers FormatMajorVersion) error {
374 2 : if err := d.writeFormatVersionMarker(formatVers); err != nil {
375 0 : return err
376 0 : }
377 2 : d.mu.formatVers.vers.Store(uint64(formatVers))
378 2 : d.opts.EventListener.FormatUpgrade(formatVers)
379 2 : return nil
380 : }
381 :
382 2 : func (d *DB) writeFormatVersionMarker(formatVers FormatMajorVersion) error {
383 2 : // We use the marker to encode the active format version in the
384 2 : // marker filename. Unlike other uses of the atomic marker, there is
385 2 : // no file with the filename `formatVers.String()` on the
386 2 : // filesystem.
387 2 : return d.mu.formatVers.marker.Move(formatVers.String())
388 2 : }
389 :
390 : // compactMarkedFilesLocked performs a migration that schedules rewrite
391 : // compactions to compact away any sstables marked for compaction.
392 : // compactMarkedFilesLocked is run while ratcheting the database's format major
393 : // version to FormatSplitUserKeysMarkedCompacted.
394 : //
395 : // Note that while this method is called with the DB.mu held, and will not
396 : // return until all marked files have been compacted, the mutex is dropped while
397 : // waiting for compactions to complete (or for slots to free up).
398 2 : func (d *DB) compactMarkedFilesLocked() error {
399 2 : curr := d.mu.versions.currentVersion()
400 2 : for curr.Stats.MarkedForCompaction > 0 {
401 0 : // Attempt to schedule a compaction to rewrite a file marked for
402 0 : // compaction.
403 0 : d.maybeScheduleCompactionPicker(func(picker compactionPicker, env compactionEnv) *pickedCompaction {
404 0 : return picker.pickRewriteCompaction(env)
405 0 : })
406 :
407 : // The above attempt might succeed and schedule a rewrite compaction. Or
408 : // there might not be available compaction concurrency to schedule the
409 : // compaction. Or compaction of the file might have already been in
410 : // progress. In any scenario, wait until there's some change in the
411 : // state of active compactions.
412 :
413 : // Before waiting, check that the database hasn't been closed. Trying to
414 : // schedule the compaction may have dropped d.mu while waiting for a
415 : // manifest write to complete. In that dropped interim, the database may
416 : // have been closed.
417 0 : if err := d.closed.Load(); err != nil {
418 0 : return err.(error)
419 0 : }
420 :
421 : // Some flush or compaction may have scheduled or completed while we waited
422 : // for the manifest lock in maybeScheduleCompactionPicker. Get the latest
423 : // Version before waiting on a compaction.
424 0 : curr = d.mu.versions.currentVersion()
425 0 :
426 0 : // Only wait on compactions if there are files still marked for compaction.
427 0 : // NB: Waiting on this condition variable drops d.mu while blocked.
428 0 : if curr.Stats.MarkedForCompaction > 0 {
429 0 : if d.mu.compact.compactingCount == 0 {
430 0 : panic("expected a compaction of marked files in progress")
431 : }
432 0 : d.mu.compact.cond.Wait()
433 0 : // Refresh the current version again.
434 0 : curr = d.mu.versions.currentVersion()
435 : }
436 : }
437 2 : return nil
438 : }
439 :
440 : // findFilesFunc scans the LSM for files, returning true if at least one
441 : // file was found. The returned array contains the matched files, if any, per
442 : // level.
443 : type findFilesFunc func(v *version) (found bool, files [numLevels][]*fileMetadata, _ error)
444 :
445 : // This method is not used currently, but it will be useful the next time we need
446 : // to mark files for compaction.
447 : var _ = (*DB)(nil).markFilesLocked
448 :
449 : // markFilesLocked durably marks the files that match the given findFilesFunc for
450 : // compaction.
451 0 : func (d *DB) markFilesLocked(findFn findFilesFunc) error {
452 0 : jobID := d.mu.nextJobID
453 0 : d.mu.nextJobID++
454 0 :
455 0 : // Acquire a read state to have a view of the LSM and a guarantee that none
456 0 : // of the referenced files will be deleted until we've unreferenced the read
457 0 : // state. Some findFilesFuncs may read the files, requiring they not be
458 0 : // deleted.
459 0 : rs := d.loadReadState()
460 0 : var (
461 0 : found bool
462 0 : files [numLevels][]*fileMetadata
463 0 : err error
464 0 : )
465 0 : func() {
466 0 : defer rs.unrefLocked()
467 0 : // Note the unusual locking: unlock, defer Lock(). The scan of the files in
468 0 : // the version does not need to block other operations that require the
469 0 : // DB.mu. Drop it for the scan, before re-acquiring it.
470 0 : d.mu.Unlock()
471 0 : defer d.mu.Lock()
472 0 : found, files, err = findFn(rs.current)
473 0 : }()
474 0 : if err != nil {
475 0 : return err
476 0 : }
477 :
478 : // The database lock has been acquired again by the defer within the above
479 : // anonymous function.
480 0 : if !found {
481 0 : // Nothing to do.
482 0 : return nil
483 0 : }
484 :
485 : // After scanning, if we found files to mark, we fetch the current state of
486 : // the LSM (which may have changed) and set MarkedForCompaction on the files,
487 : // and update the version's Stats.MarkedForCompaction count, which are both
488 : // protected by d.mu.
489 :
490 : // Lock the manifest for a coherent view of the LSM. The database lock has
491 : // been re-acquired by the defer within the above anonymous function.
492 0 : d.mu.versions.logLock()
493 0 : vers := d.mu.versions.currentVersion()
494 0 : for l, filesToMark := range files {
495 0 : if len(filesToMark) == 0 {
496 0 : continue
497 : }
498 0 : for _, f := range filesToMark {
499 0 : // Ignore files to be marked that have already been compacted or marked.
500 0 : if f.CompactionState == manifest.CompactionStateCompacted ||
501 0 : f.MarkedForCompaction {
502 0 : continue
503 : }
504 : // Else, mark the file for compaction in this version.
505 0 : vers.Stats.MarkedForCompaction++
506 0 : f.MarkedForCompaction = true
507 : }
508 : // The compaction picker uses the markedForCompactionAnnotator to
509 : // quickly find files marked for compaction, or to quickly determine
510 : // that there are no such files marked for compaction within a level.
511 : // A b-tree node may be annotated with an annotation recording that
512 : // there are no files marked for compaction within the node's subtree,
513 : // based on the assumption that it's static.
514 : //
515 : // Since we're marking files for compaction, these b-tree nodes'
516 : // annotations will be out of date. Clear the compaction-picking
517 : // annotation, so that it's recomputed the next time the compaction
518 : // picker looks for a file marked for compaction.
519 0 : vers.Levels[l].InvalidateAnnotation(markedForCompactionAnnotator{})
520 : }
521 :
522 : // The 'marked-for-compaction' bit is persisted in the MANIFEST file
523 : // metadata. We've already modified the in-memory file metadata, but the
524 : // manifest hasn't been updated. Force rotation to a new MANIFEST file,
525 : // which will write every file metadata to the new manifest file and ensure
526 : // that the now marked-for-compaction file metadata are persisted as marked.
527 : // NB: This call to logAndApply will unlockthe MANIFEST, which we locked up
528 : // above before obtaining `vers`.
529 0 : return d.mu.versions.logAndApply(
530 0 : jobID,
531 0 : &manifest.VersionEdit{},
532 0 : map[int]*LevelMetrics{},
533 0 : true, /* forceRotation */
534 0 : func() []compactionInfo { return d.getInProgressCompactionInfoLocked(nil) })
535 : }
|
