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/objstorage/remote"
14 : "github.com/cockroachdb/pebble/sstable"
15 : "github.com/cockroachdb/pebble/sstable/blob"
16 : "github.com/cockroachdb/pebble/vfs"
17 : "github.com/cockroachdb/pebble/vfs/atomicfs"
18 : )
19 :
20 : // FormatMajorVersion is a constant controlling the format of persisted
21 : // data. Backwards incompatible changes to durable formats are gated
22 : // behind new format major versions.
23 : //
24 : // At any point, a database's format major version may be bumped.
25 : // However, once a database's format major version is increased,
26 : // previous versions of Pebble will refuse to open the database.
27 : //
28 : // The zero value format is the FormatDefault constant. The exact
29 : // FormatVersion that the default corresponds to may change with time.
30 : type FormatMajorVersion uint64
31 :
32 : // SafeValue implements redact.SafeValue.
33 0 : func (v FormatMajorVersion) SafeValue() {}
34 :
35 : // String implements fmt.Stringer.
36 1 : func (v FormatMajorVersion) String() string {
37 1 : // NB: This must not change. It's used as the value for the on-disk
38 1 : // version marker file.
39 1 : //
40 1 : // Specifically, this value must always parse as a base 10 integer
41 1 : // that fits in a uint64. We format it as zero-padded, 3-digit
42 1 : // number today, but the padding may change.
43 1 : return fmt.Sprintf("%03d", v)
44 1 : }
45 :
46 : const (
47 : // FormatDefault leaves the format version unspecified. When used to create a
48 : // new store, Pebble will choose the earliest format version it supports.
49 : FormatDefault FormatMajorVersion = iota
50 :
51 : // 21.2 versions.
52 :
53 : // FormatMostCompatible maintains the most backwards compatibility,
54 : // maintaining bi-directional compatibility with RocksDB 6.2.1 in
55 : // the particular configuration described in the Pebble README.
56 : // Deprecated.
57 : _ // FormatMostCompatible
58 :
59 : // formatVersionedManifestMarker is the first
60 : // backwards-incompatible change made to Pebble, introducing the
61 : // format-version marker file for handling backwards-incompatible
62 : // changes more broadly, and replacing the `CURRENT` file with a
63 : // marker file.
64 : //
65 : // This format version is intended as an intermediary version state.
66 : // It is deliberately unexported to discourage direct use of this
67 : // format major version. Clients should use FormatVersioned which
68 : // also ensures earlier versions of Pebble fail to open a database
69 : // written in a future format major version.
70 : // Deprecated.
71 : _ // formatVersionedManifestMarker
72 :
73 : // FormatVersioned is a new format major version that replaces the
74 : // old `CURRENT` file with a new 'marker' file scheme. Previous
75 : // Pebble versions will be unable to open the database unless
76 : // they're aware of format versions.
77 : // Deprecated.
78 : _ // FormatVersioned
79 :
80 : // FormatSetWithDelete is a format major version that introduces a new key
81 : // kind, base.InternalKeyKindSetWithDelete. Previous Pebble versions will be
82 : // unable to open this database.
83 : // Deprecated.
84 : _ // FormatSetWithDelete
85 :
86 : // 22.1 versions.
87 :
88 : // FormatBlockPropertyCollector is a format major version that introduces
89 : // BlockPropertyCollectors.
90 : // Deprecated.
91 : _ // FormatBlockPropertyCollector
92 :
93 : // FormatSplitUserKeysMarked is a format major version that guarantees that
94 : // all files that share user keys with neighbors are marked for compaction
95 : // in the manifest. Ratcheting to FormatSplitUserKeysMarked will block
96 : // (without holding mutexes) until the scan of the LSM is complete and the
97 : // manifest has been rotated.
98 : // Deprecated.
99 : _ // FormatSplitUserKeysMarked
100 :
101 : // 22.2 versions.
102 :
103 : // FormatSplitUserKeysMarkedCompacted is a format major version that
104 : // guarantees that all files explicitly marked for compaction in the manifest
105 : // have been compacted. Combined with the FormatSplitUserKeysMarked format
106 : // major version, this version guarantees that there are no user keys split
107 : // across multiple files within a level L1+. Ratcheting to this format version
108 : // will block (without holding mutexes) until all necessary compactions for
109 : // files marked for compaction are complete.
110 : // Deprecated.
111 : _ // FormatSplitUserKeysMarkedCompacted
112 :
113 : // FormatRangeKeys is a format major version that introduces range keys.
114 : // Deprecated.
115 : _ // FormatRangeKeys
116 :
117 : // FormatMinTableFormatPebblev1 is a format major version that guarantees that
118 : // tables created by or ingested into the DB at or above this format major
119 : // version will have a table format version of at least Pebblev1 (Block
120 : // FileProperties).
121 : // Deprecated.
122 : _ // FormatMinTableFormatPebblev1
123 :
124 : // FormatPrePebblev1Marked is a format major version that guarantees that all
125 : // sstables with a table format version pre-Pebblev1 (i.e. those that are
126 : // guaranteed to not contain block properties) are marked for compaction in
127 : // the manifest. Ratcheting to FormatPrePebblev1Marked will block (without
128 : // holding mutexes) until the scan of the LSM is complete and the manifest has
129 : // been rotated.
130 : // Deprecated.
131 : _ // FormatPrePebblev1Marked
132 :
133 : // 23.1 versions.
134 :
135 : // formatUnusedPrePebblev1MarkedCompacted is an unused format major version.
136 : // This format major version was originally intended to ship in the 23.1
137 : // release. It was later decided that this should be deferred until a
138 : // subsequent release. The original ordering is preserved so as not to
139 : // introduce breaking changes in Cockroach.
140 : _ // formatUnusedPrePebblev1MarkedCompacted
141 :
142 : // FormatSSTableValueBlocks is a format major version that adds support for
143 : // storing values in value blocks in the sstable. Value block support is not
144 : // necessarily enabled when writing sstables, when running with this format
145 : // major version.
146 : _ // FormatSSTableValueBlocks
147 :
148 : // FormatFlushableIngest is a format major version that enables lazy
149 : // addition of ingested sstables into the LSM structure. When an ingest
150 : // overlaps with a memtable, a record of the ingest is written to the WAL
151 : // without waiting for a flush. Subsequent reads treat the ingested files as
152 : // a level above the overlapping memtable. Once the memtable is flushed, the
153 : // ingested files are moved into the lowest possible levels.
154 : //
155 : // This feature is behind a format major version because it required
156 : // breaking changes to the WAL format.
157 : FormatFlushableIngest
158 :
159 : // 23.2 versions.
160 :
161 : // FormatPrePebblev1MarkedCompacted is a format major version that guarantees
162 : // that all sstables explicitly marked for compaction in the manifest (see
163 : // FormatPrePebblev1Marked) have been compacted. Ratcheting to this format
164 : // version will block (without holding mutexes) until all necessary
165 : // compactions for files marked for compaction are complete.
166 : FormatPrePebblev1MarkedCompacted
167 :
168 : // FormatDeleteSizedAndObsolete is a format major version that adds support
169 : // for deletion tombstones that encode the size of the value they're
170 : // expected to delete. This format major version is required before the
171 : // associated key kind may be committed through batch applications or
172 : // ingests. It also adds support for keys that are marked obsolete (see
173 : // sstable/format.go for details).
174 : FormatDeleteSizedAndObsolete
175 :
176 : // FormatVirtualSSTables is a format major version that adds support for
177 : // virtual sstables that can reference a sub-range of keys in an underlying
178 : // physical sstable. This information is persisted through new,
179 : // backward-incompatible fields in the Manifest, and therefore requires
180 : // a format major version.
181 : FormatVirtualSSTables
182 :
183 : // FormatSyntheticPrefixSuffix is a format major version that adds support for
184 : // sstables to have their content exposed in a different prefix or suffix of
185 : // keyspace than the actual prefix/suffix persisted in the keys in such
186 : // sstables. The prefix and suffix replacement information is stored in new
187 : // fields in the Manifest and thus requires a format major version.
188 : FormatSyntheticPrefixSuffix
189 :
190 : // FormatFlushableIngestExcises is a format major version that adds support for
191 : // having excises unconditionally being written as flushable ingestions. This
192 : // is implemented through adding a new key kind that can go in the same batches
193 : // as flushable ingested sstables.
194 : FormatFlushableIngestExcises
195 :
196 : // FormatColumnarBlocks is a format major version enabling use of the
197 : // TableFormatPebblev5 table format, that encodes sstable data blocks, index
198 : // blocks and keyspan blocks by organizing the KVs into columns within the
199 : // block.
200 : FormatColumnarBlocks
201 :
202 : // FormatWALSyncChunks is a format major version enabling the writing of
203 : // WAL sync chunks. These new chunks are used to disambiguate between corruption
204 : // and logical EOF during WAL replay. This is implemented by adding a new
205 : // chunk wire format that encodes an additional "Synced Offset" field which acts
206 : // as a commitment that the WAL should have been synced up until the offset.
207 : FormatWALSyncChunks
208 :
209 : // FormatTableFormatV6 is a format major version enabling the sstable table
210 : // format TableFormatPebblev6.
211 : //
212 : // The TableFormatPebblev6 sstable format introduces a checksum within the
213 : // sstable footer, allows inclusion of blob handle references within the
214 : // value column of a sstable block, and supports columnar meta index +
215 : // properties blocks.
216 : //
217 : // This format major version does not yet enable use of value separation.
218 : FormatTableFormatV6
219 :
220 : // formatDeprecatedExperimentalValueSeparation was used to enable an
221 : // experimental version of value separation, separating values into external
222 : // blob files that do not participate in every compaction.
223 : //
224 : // Value separation now depends on TableFormatPebblev7 which this format
225 : // major version precedes. This format major version is deprecated and
226 : // unexported, and value separation now requires FormatValueSeparation.
227 : formatDeprecatedExperimentalValueSeparation
228 :
229 : // formatFooterAttributes is a format major version that adds support for
230 : // writing sstable.Attributes in the footer of sstables.
231 : formatFooterAttributes
232 :
233 : // FormatValueSeparation is a format major version that adds support for
234 : // value separation, separating values into external blob files that do not
235 : // participate in every compaction.
236 : FormatValueSeparation
237 :
238 : // FormatExciseBoundsRecord is a format major version that adds support for
239 : // persisting excise bounds records in the manifest (VersionEdit).
240 : FormatExciseBoundsRecord
241 :
242 : // FormatV2BlobFiles is a format major version that adds support for V2 blob
243 : // file format (which adds compression statistics).
244 : FormatV2BlobFiles
245 :
246 : // FormatBackingValueSize is a format major version that adds support for
247 : // persisting the value size of the backing sst for virtual sstables in the
248 : // manifest (VersionEdit).
249 : FormatBackingValueSize
250 :
251 : // FormatMarkForCompactionInVersionEdit is a format major version that adds
252 : // marking tables for compaction (via tagTableMarkedForCompaction).
253 : // Previously, marking for compaction required a manifest rotation.
254 : FormatMarkForCompactionInVersionEdit
255 :
256 : // -- Add new versions here --
257 :
258 : // FormatNewest is the most recent format major version.
259 : FormatNewest FormatMajorVersion = iota - 1
260 :
261 : // Experimental versions, which are excluded by FormatNewest (but can be used
262 : // in tests) can be defined here.
263 :
264 : // -- Add experimental versions here --
265 :
266 : // internalFormatNewest is the most recent, possibly experimental format major
267 : // version.
268 : internalFormatNewest FormatMajorVersion = iota - 2
269 : )
270 :
271 : // FormatMinSupported is the minimum format version that is supported by this
272 : // Pebble version.
273 : const FormatMinSupported = FormatFlushableIngest
274 :
275 : // FormatMinForSharedObjects it the minimum format version that supports shared
276 : // objects (see CreateOnShared option).
277 : const FormatMinForSharedObjects = FormatVirtualSSTables
278 :
279 : // resolveDefault asserts that the given version is supported, and returns the
280 : // given version, replacing FormatDefault with FormatMinSupported.
281 1 : func (v FormatMajorVersion) resolveDefault() FormatMajorVersion {
282 1 : if v == FormatDefault {
283 1 : return FormatMinSupported
284 1 : }
285 1 : if v < FormatMinSupported || v > internalFormatNewest {
286 1 : panic(fmt.Sprintf("pebble: unsupported format major version: %s", v))
287 : }
288 1 : return v
289 : }
290 :
291 : // MaxTableFormat returns the maximum sstable.TableFormat that can be used at
292 : // this FormatMajorVersion.
293 1 : func (v FormatMajorVersion) MaxTableFormat() sstable.TableFormat {
294 1 : v = v.resolveDefault()
295 1 : switch {
296 1 : case v >= formatFooterAttributes:
297 1 : return sstable.TableFormatPebblev7
298 1 : case v >= FormatTableFormatV6:
299 1 : return sstable.TableFormatPebblev6
300 1 : case v >= FormatColumnarBlocks:
301 1 : return sstable.TableFormatPebblev5
302 1 : case v >= FormatDeleteSizedAndObsolete:
303 1 : return sstable.TableFormatPebblev4
304 1 : default:
305 1 : return sstable.TableFormatPebblev3
306 : }
307 : }
308 :
309 : // MinTableFormat returns the minimum sstable.TableFormat that can be used at
310 : // this FormatMajorVersion.
311 1 : func (v FormatMajorVersion) MinTableFormat() sstable.TableFormat {
312 1 : _ = v.resolveDefault()
313 1 : return sstable.TableFormatPebblev1
314 1 : }
315 :
316 : // MaxBlobFileFormat returns the maximum blob.FileFormat that can be used at
317 : // this FormatMajorVersion. It can only be used on versions that support value
318 : // separation.
319 1 : func (v FormatMajorVersion) MaxBlobFileFormat() blob.FileFormat {
320 1 : v = v.resolveDefault()
321 1 : switch {
322 1 : case v >= FormatV2BlobFiles:
323 1 : return blob.FileFormatV2
324 1 : case v >= FormatValueSeparation:
325 1 : return blob.FileFormatV1
326 1 : default:
327 1 : panic(fmt.Sprintf("pebble: format major version %s does not support blob files", v))
328 : }
329 : }
330 :
331 : // formatMajorVersionMigrations defines the migrations from one format
332 : // major version to the next. Each migration is defined as a closure
333 : // which will be invoked on the database before the new format major
334 : // version is committed. Migrations must be idempotent. Migrations are
335 : // invoked with d.mu locked.
336 : //
337 : // Each migration is responsible for invoking finalizeFormatVersUpgrade
338 : // to set the new format major version. RatchetFormatMajorVersion will
339 : // panic if a migration returns a nil error but fails to finalize the
340 : // new format major version.
341 : var formatMajorVersionMigrations = map[FormatMajorVersion]func(*DB) error{
342 0 : FormatFlushableIngest: func(d *DB) error { return nil },
343 1 : FormatPrePebblev1MarkedCompacted: func(d *DB) error {
344 1 : // Before finalizing the format major version, rewrite any sstables
345 1 : // still marked for compaction. Note all format major versions
346 1 : // migrations are invoked with DB.mu locked.
347 1 : if err := d.compactMarkedFilesLocked(); err != nil {
348 0 : return err
349 0 : }
350 1 : return d.finalizeFormatVersUpgrade(FormatPrePebblev1MarkedCompacted)
351 : },
352 1 : FormatDeleteSizedAndObsolete: func(d *DB) error {
353 1 : return d.finalizeFormatVersUpgrade(FormatDeleteSizedAndObsolete)
354 1 : },
355 1 : FormatVirtualSSTables: func(d *DB) error {
356 1 : return d.finalizeFormatVersUpgrade(FormatVirtualSSTables)
357 1 : },
358 1 : FormatSyntheticPrefixSuffix: func(d *DB) error {
359 1 : return d.finalizeFormatVersUpgrade(FormatSyntheticPrefixSuffix)
360 1 : },
361 1 : FormatFlushableIngestExcises: func(d *DB) error {
362 1 : return d.finalizeFormatVersUpgrade(FormatFlushableIngestExcises)
363 1 : },
364 1 : FormatColumnarBlocks: func(d *DB) error {
365 1 : return d.finalizeFormatVersUpgrade(FormatColumnarBlocks)
366 1 : },
367 1 : FormatWALSyncChunks: func(d *DB) error {
368 1 : return d.finalizeFormatVersUpgrade(FormatWALSyncChunks)
369 1 : },
370 1 : FormatTableFormatV6: func(d *DB) error {
371 1 : return d.finalizeFormatVersUpgrade(FormatTableFormatV6)
372 1 : },
373 1 : formatDeprecatedExperimentalValueSeparation: func(d *DB) error {
374 1 : return d.finalizeFormatVersUpgrade(formatDeprecatedExperimentalValueSeparation)
375 1 : },
376 1 : formatFooterAttributes: func(d *DB) error {
377 1 : return d.finalizeFormatVersUpgrade(formatFooterAttributes)
378 1 : },
379 1 : FormatValueSeparation: func(d *DB) error {
380 1 : return d.finalizeFormatVersUpgrade(FormatValueSeparation)
381 1 : },
382 1 : FormatExciseBoundsRecord: func(d *DB) error {
383 1 : return d.finalizeFormatVersUpgrade(FormatExciseBoundsRecord)
384 1 : },
385 1 : FormatV2BlobFiles: func(d *DB) error {
386 1 : return d.finalizeFormatVersUpgrade(FormatV2BlobFiles)
387 1 : },
388 1 : FormatBackingValueSize: func(d *DB) error {
389 1 : return d.finalizeFormatVersUpgrade(FormatBackingValueSize)
390 1 : },
391 1 : FormatMarkForCompactionInVersionEdit: func(d *DB) error {
392 1 : return d.finalizeFormatVersUpgrade(FormatMarkForCompactionInVersionEdit)
393 1 : },
394 : }
395 :
396 : const formatVersionMarkerName = `format-version`
397 :
398 : // lookupFormatMajorVersion retrieves the format version from the format version
399 : // marker file.
400 : //
401 : // If such a file does not exist, returns FormatDefault. Note that this case is
402 : // only acceptable if we are creating a new store (we no longer support
403 : // FormatMostCompatible which is the only one with no version marker file).
404 : func lookupFormatMajorVersion(
405 : fs vfs.FS, dirname string, ls []string,
406 1 : ) (FormatMajorVersion, *atomicfs.Marker, error) {
407 1 : m, versString, err := atomicfs.LocateMarkerInListing(fs, dirname, formatVersionMarkerName, ls)
408 1 : if err != nil {
409 1 : return 0, nil, err
410 1 : }
411 1 : if versString == "" {
412 1 : return FormatDefault, m, nil
413 1 : }
414 1 : v, err := strconv.ParseUint(versString, 10, 64)
415 1 : if err != nil {
416 0 : return 0, nil, errors.Wrap(err, "parsing format major version")
417 0 : }
418 1 : vers := FormatMajorVersion(v)
419 1 : if vers == FormatDefault {
420 0 : return 0, nil, errors.Newf("pebble: default format major version should not persisted", vers)
421 0 : }
422 1 : if vers > internalFormatNewest {
423 1 : return 0, nil, errors.Newf("pebble: database %q written in unknown format major version %d", dirname, vers)
424 1 : }
425 1 : if vers < FormatMinSupported {
426 0 : return 0, nil, errors.Newf("pebble: database %q written in format major version %d which is no longer supported", dirname, vers)
427 0 : }
428 1 : return vers, m, nil
429 : }
430 :
431 : // FormatMajorVersion returns the database's active format major
432 : // version. The format major version may be higher than the one
433 : // provided in Options when the database was opened if the existing
434 : // database was written with a higher format version.
435 1 : func (d *DB) FormatMajorVersion() FormatMajorVersion {
436 1 : return FormatMajorVersion(d.mu.formatVers.vers.Load())
437 1 : }
438 :
439 : // TableFormat returns the TableFormat that the database is currently using when
440 : // writing sstables. The table format is determined by the database's format
441 : // major version, as well as experimental settings like EnableValueBlocks and
442 : // EnableColumnarBlocks.
443 1 : func (d *DB) TableFormat() sstable.TableFormat {
444 1 : // The table is typically written at the maximum allowable format implied by
445 1 : // the current format major version of the DB.
446 1 : f := d.FormatMajorVersion().MaxTableFormat()
447 1 : if f == sstable.TableFormatPebblev3 {
448 1 : // In format major versions with maximum table formats of Pebblev3,
449 1 : // value blocks were conditional on an experimental setting. In format
450 1 : // major versions with maximum table formats of Pebblev4 and higher,
451 1 : // value blocks are always enabled.
452 1 : if d.opts.Experimental.EnableValueBlocks == nil || !d.opts.Experimental.EnableValueBlocks() {
453 1 : f = sstable.TableFormatPebblev2
454 1 : }
455 : }
456 1 : return f
457 : }
458 :
459 : // BlobFileFormat returns the blob.FileFormat that the database is currently
460 : // using when writing blob files.
461 1 : func (d *DB) BlobFileFormat() blob.FileFormat {
462 1 : return d.FormatMajorVersion().MaxBlobFileFormat()
463 1 : }
464 :
465 : // shouldCreateShared returns true if the database should use shared objects
466 : // when creating new objects on the given level.
467 1 : func (d *DB) shouldCreateShared(targetLevel int) bool {
468 1 : return remote.ShouldCreateShared(d.opts.Experimental.CreateOnShared, targetLevel) &&
469 1 : d.FormatMajorVersion() >= FormatMinForSharedObjects
470 1 : }
471 :
472 : // RatchetFormatMajorVersion ratchets the opened database's format major
473 : // version to the provided version. It errors if the provided format
474 : // major version is below the database's current version. Once a
475 : // database's format major version is upgraded, previous Pebble versions
476 : // that do not know of the format version will be unable to open the
477 : // database.
478 1 : func (d *DB) RatchetFormatMajorVersion(fmv FormatMajorVersion) error {
479 1 : if err := d.closed.Load(); err != nil {
480 1 : panic(err)
481 : }
482 :
483 1 : d.mu.Lock()
484 1 : defer d.mu.Unlock()
485 1 : return d.ratchetFormatMajorVersionLocked(fmv)
486 : }
487 :
488 1 : func (d *DB) ratchetFormatMajorVersionLocked(formatVers FormatMajorVersion) error {
489 1 : if d.opts.ReadOnly {
490 0 : return ErrReadOnly
491 0 : }
492 1 : if formatVers > internalFormatNewest {
493 0 : // Guard against accidentally forgetting to update internalFormatNewest.
494 0 : return errors.Errorf("pebble: unknown format version %d", formatVers)
495 0 : }
496 1 : if currentVers := d.FormatMajorVersion(); currentVers > formatVers {
497 0 : return errors.Newf("pebble: database already at format major version %d; cannot reduce to %d",
498 0 : currentVers, formatVers)
499 0 : }
500 1 : if d.mu.formatVers.ratcheting {
501 0 : return errors.Newf("pebble: database format major version upgrade is in-progress")
502 0 : }
503 1 : d.mu.formatVers.ratcheting = true
504 1 : defer func() { d.mu.formatVers.ratcheting = false }()
505 :
506 1 : for nextVers := d.FormatMajorVersion() + 1; nextVers <= formatVers; nextVers++ {
507 1 : if err := formatMajorVersionMigrations[nextVers](d); err != nil {
508 0 : return errors.Wrapf(err, "migrating to version %d", nextVers)
509 0 : }
510 :
511 : // NB: The migration is responsible for calling
512 : // finalizeFormatVersUpgrade to finalize the upgrade. This
513 : // structure is necessary because some migrations may need to
514 : // update in-memory state (without ever dropping locks) after
515 : // the upgrade is finalized. Here we assert that the upgrade
516 : // did occur.
517 1 : if d.FormatMajorVersion() != nextVers {
518 0 : d.opts.Logger.Fatalf("pebble: successful migration to format version %d never finalized the upgrade", nextVers)
519 0 : }
520 : }
521 1 : return nil
522 : }
523 :
524 : // finalizeFormatVersUpgrade is typically only be called from within a
525 : // format major version migration.
526 : //
527 : // See formatMajorVersionMigrations.
528 1 : func (d *DB) finalizeFormatVersUpgrade(formatVers FormatMajorVersion) error {
529 1 : if err := d.writeFormatVersionMarker(formatVers); err != nil {
530 0 : return err
531 0 : }
532 1 : d.mu.formatVers.vers.Store(uint64(formatVers))
533 1 : d.opts.EventListener.FormatUpgrade(formatVers)
534 1 : return nil
535 : }
536 :
537 1 : func (d *DB) writeFormatVersionMarker(formatVers FormatMajorVersion) error {
538 1 : // We use the marker to encode the active format version in the
539 1 : // marker filename. Unlike other uses of the atomic marker, there is
540 1 : // no file with the filename `formatVers.String()` on the
541 1 : // filesystem.
542 1 : return d.mu.formatVers.marker.Move(formatVers.String())
543 1 : }
544 :
545 : // compactMarkedFilesLocked performs a migration that schedules rewrite
546 : // compactions to compact away any sstables marked for compaction.
547 : // compactMarkedFilesLocked is run while ratcheting the database's format major
548 : // version to FormatSplitUserKeysMarkedCompacted.
549 : //
550 : // Note that while this method is called with the DB.mu held, and will not
551 : // return until all marked files have been compacted, the mutex is dropped while
552 : // waiting for compactions to complete (or for slots to free up).
553 1 : func (d *DB) compactMarkedFilesLocked() error {
554 1 : curr := d.mu.versions.currentVersion()
555 1 : if curr.MarkedForCompaction.Count() == 0 {
556 1 : return nil
557 1 : }
558 : // Attempt to schedule a compaction to rewrite a file marked for compaction.
559 : // We simply call maybeScheduleCompaction since it also picks rewrite
560 : // compactions. Note that we don't need to call this repeatedly in the for
561 : // loop below since the completion of a compaction either starts a new one
562 : // or ensures a compaction is queued for scheduling. By calling
563 : // maybeScheduleCompaction here we are simply kicking off this behavior.
564 0 : d.maybeScheduleCompaction()
565 0 :
566 0 : // The above attempt might succeed and schedule a rewrite compaction. Or
567 0 : // there might not be available compaction concurrency to schedule the
568 0 : // compaction. Or compaction of the file might have already been in
569 0 : // progress. In any scenario, wait until there's some change in the
570 0 : // state of active compactions.
571 0 : for curr.MarkedForCompaction.Count() > 0 {
572 0 : // Before waiting, check that the database hasn't been closed. Trying to
573 0 : // schedule the compaction may have dropped d.mu while waiting for a
574 0 : // manifest write to complete. In that dropped interim, the database may
575 0 : // have been closed.
576 0 : if err := d.closed.Load(); err != nil {
577 0 : return err.(error)
578 0 : }
579 :
580 : // Some flush or compaction may have scheduled or completed while we waited
581 : // for the manifest lock in maybeScheduleCompactionPicker. Get the latest
582 : // Version before waiting on a compaction.
583 0 : curr = d.mu.versions.currentVersion()
584 0 :
585 0 : // Only wait on compactions if there are files still marked for compaction.
586 0 : // NB: Waiting on this condition variable drops d.mu while blocked.
587 0 : if curr.MarkedForCompaction.Count() > 0 {
588 0 : // NB: we cannot assert that d.mu.compact.compactingCount > 0, since
589 0 : // with a CompactionScheduler a DB may not have even one ongoing
590 0 : // compaction (if other competing activities are being preferred by the
591 0 : // scheduler).
592 0 : d.mu.compact.cond.Wait()
593 0 : // Refresh the current version again.
594 0 : curr = d.mu.versions.currentVersion()
595 0 : }
596 : }
597 0 : return nil
598 : }
599 :
600 : // findFilesFunc scans the LSM for files, returning true if at least one
601 : // file was found. The returned array contains the matched files, if any, per
602 : // level.
603 : type findFilesFunc func(v *manifest.Version) (found bool, files [numLevels][]*manifest.TableMetadata, _ error)
604 :
605 : // markFilesForCompactionLocked durably marks the files that match the given findFilesFunc for
606 : // compaction.
607 1 : func (d *DB) markFilesForCompactionLocked(findFn findFilesFunc) error {
608 1 : if d.FormatMajorVersion() < FormatMarkForCompactionInVersionEdit {
609 0 : return errors.Newf("pebble: marking files for compaction requires format major version %d or higher", FormatMarkForCompactionInVersionEdit)
610 0 : }
611 1 : jobID := d.newJobIDLocked()
612 1 :
613 1 : // Acquire a read state to have a view of the LSM and a guarantee that none
614 1 : // of the referenced files will be deleted until we've unreferenced the read
615 1 : // state. Some findFilesFuncs may read the files, requiring they not be
616 1 : // deleted.
617 1 : rs := d.loadReadState()
618 1 : var (
619 1 : found bool
620 1 : files [numLevels][]*manifest.TableMetadata
621 1 : err error
622 1 : )
623 1 : func() {
624 1 : defer rs.unrefLocked()
625 1 : // Note the unusual locking: unlock, defer Lock(). The scan of the files in
626 1 : // the version does not need to block other operations that require the
627 1 : // DB.mu. Drop it for the scan, before re-acquiring it.
628 1 : d.mu.Unlock()
629 1 : defer d.mu.Lock()
630 1 : found, files, err = findFn(rs.current)
631 1 : }()
632 1 : if err != nil {
633 0 : return err
634 0 : }
635 :
636 : // The database lock has been acquired again by the defer within the above
637 : // anonymous function.
638 1 : if !found {
639 0 : // Nothing to do.
640 0 : return nil
641 0 : }
642 :
643 : // After scanning, if we found files to mark, we fetch the current state of
644 : // the LSM (which may have changed) and build the list of tables to mark for
645 : // compaction.
646 :
647 : // Lock the manifest for a coherent view of the LSM. The database lock has
648 : // been re-acquired by the defer within the above anonymous function.
649 1 : _, err = d.mu.versions.UpdateVersionLocked(func() (versionUpdate, error) {
650 1 : var ve manifest.VersionEdit
651 1 : vers := d.mu.versions.currentVersion()
652 1 : for level, filesToMark := range files {
653 1 : for _, f := range filesToMark {
654 1 : // Ignore files to be marked that have already been compacted or marked.
655 1 : if f.CompactionState == manifest.CompactionStateCompacted ||
656 1 : vers.MarkedForCompaction.Contains(f, level) {
657 0 : continue
658 : }
659 : // Else, mark the file for compaction in this version.
660 1 : ve.TablesMarkedForCompaction = append(ve.TablesMarkedForCompaction, manifest.TableMarkedForCompactionEntry{
661 1 : TableNum: f.TableNum,
662 1 : Level: level,
663 1 : Meta: f,
664 1 : })
665 : }
666 : }
667 1 : return versionUpdate{
668 1 : VE: &ve,
669 1 : JobID: jobID,
670 1 : InProgressCompactionsFn: func() []compactionInfo { return d.getInProgressCompactionInfoLocked(nil) },
671 : }, nil
672 : })
673 1 : return err
674 : }
|
