Line data Source code
1 : // Copyright 2011 The LevelDB-Go and Pebble Authors. All rights reserved. Use
2 : // of this source code is governed by a BSD-style license that can be found in
3 : // the LICENSE file.
4 :
5 : package pebble
6 :
7 : import (
8 : "bytes"
9 : "fmt"
10 : "io"
11 : "regexp"
12 : "runtime"
13 : "slices"
14 : "sort"
15 : "strconv"
16 : "strings"
17 : "time"
18 : "unicode"
19 :
20 : "github.com/cockroachdb/crlib/fifo"
21 : "github.com/cockroachdb/errors"
22 : "github.com/cockroachdb/pebble/internal/base"
23 : "github.com/cockroachdb/pebble/internal/cache"
24 : "github.com/cockroachdb/pebble/internal/humanize"
25 : "github.com/cockroachdb/pebble/internal/keyspan"
26 : "github.com/cockroachdb/pebble/internal/manifest"
27 : "github.com/cockroachdb/pebble/internal/testkeys"
28 : "github.com/cockroachdb/pebble/objstorage/objstorageprovider"
29 : "github.com/cockroachdb/pebble/objstorage/remote"
30 : "github.com/cockroachdb/pebble/rangekey"
31 : "github.com/cockroachdb/pebble/sstable"
32 : "github.com/cockroachdb/pebble/sstable/blob"
33 : "github.com/cockroachdb/pebble/sstable/block"
34 : "github.com/cockroachdb/pebble/sstable/colblk"
35 : "github.com/cockroachdb/pebble/vfs"
36 : "github.com/cockroachdb/pebble/wal"
37 : "github.com/cockroachdb/redact"
38 : )
39 :
40 : const (
41 : cacheDefaultSize = 8 << 20 // 8 MB
42 : defaultLevelMultiplier = 10
43 : )
44 :
45 : // FilterType exports the base.FilterType type.
46 : type FilterType = base.FilterType
47 :
48 : // Exported TableFilter constants.
49 : const (
50 : TableFilter = base.TableFilter
51 : )
52 :
53 : // FilterWriter exports the base.FilterWriter type.
54 : type FilterWriter = base.FilterWriter
55 :
56 : // FilterPolicy exports the base.FilterPolicy type.
57 : type FilterPolicy = base.FilterPolicy
58 :
59 : var NoFilterPolicy = base.NoFilterPolicy
60 :
61 : // KeySchema exports the colblk.KeySchema type.
62 : type KeySchema = colblk.KeySchema
63 :
64 : // BlockPropertyCollector exports the sstable.BlockPropertyCollector type.
65 : type BlockPropertyCollector = sstable.BlockPropertyCollector
66 :
67 : // BlockPropertyFilter exports the sstable.BlockPropertyFilter type.
68 : type BlockPropertyFilter = base.BlockPropertyFilter
69 :
70 : // ShortAttributeExtractor exports the base.ShortAttributeExtractor type.
71 : type ShortAttributeExtractor = base.ShortAttributeExtractor
72 :
73 : // UserKeyPrefixBound exports the sstable.UserKeyPrefixBound type.
74 : type UserKeyPrefixBound = sstable.UserKeyPrefixBound
75 :
76 : // IterKeyType configures which types of keys an iterator should surface.
77 : type IterKeyType int8
78 :
79 : const (
80 : // IterKeyTypePointsOnly configures an iterator to iterate over point keys
81 : // only.
82 : IterKeyTypePointsOnly IterKeyType = iota
83 : // IterKeyTypeRangesOnly configures an iterator to iterate over range keys
84 : // only.
85 : IterKeyTypeRangesOnly
86 : // IterKeyTypePointsAndRanges configures an iterator iterate over both point
87 : // keys and range keys simultaneously.
88 : IterKeyTypePointsAndRanges
89 : )
90 :
91 : // String implements fmt.Stringer.
92 1 : func (t IterKeyType) String() string {
93 1 : switch t {
94 1 : case IterKeyTypePointsOnly:
95 1 : return "points-only"
96 1 : case IterKeyTypeRangesOnly:
97 1 : return "ranges-only"
98 1 : case IterKeyTypePointsAndRanges:
99 1 : return "points-and-ranges"
100 0 : default:
101 0 : panic(fmt.Sprintf("unknown key type %d", t))
102 : }
103 : }
104 :
105 : // IterOptions hold the optional per-query parameters for NewIter.
106 : //
107 : // Like Options, a nil *IterOptions is valid and means to use the default
108 : // values.
109 : type IterOptions struct {
110 : // LowerBound specifies the smallest key (inclusive) that the iterator will
111 : // return during iteration. If the iterator is seeked or iterated past this
112 : // boundary the iterator will return Valid()==false. Setting LowerBound
113 : // effectively truncates the key space visible to the iterator.
114 : LowerBound []byte
115 : // UpperBound specifies the largest key (exclusive) that the iterator will
116 : // return during iteration. If the iterator is seeked or iterated past this
117 : // boundary the iterator will return Valid()==false. Setting UpperBound
118 : // effectively truncates the key space visible to the iterator.
119 : UpperBound []byte
120 : // SkipPoint may be used to skip over point keys that don't match an
121 : // arbitrary predicate during iteration. If set, the Iterator invokes
122 : // SkipPoint for keys encountered. If SkipPoint returns true, the iterator
123 : // will skip the key without yielding it to the iterator operation in
124 : // progress.
125 : //
126 : // SkipPoint must be a pure function and always return the same result when
127 : // provided the same arguments. The iterator may call SkipPoint multiple
128 : // times for the same user key.
129 : SkipPoint func(userKey []byte) bool
130 : // PointKeyFilters can be used to avoid scanning tables and blocks in tables
131 : // when iterating over point keys. This slice represents an intersection
132 : // across all filters, i.e., all filters must indicate that the block is
133 : // relevant.
134 : //
135 : // Performance note: When len(PointKeyFilters) > 0, the caller should ensure
136 : // that cap(PointKeyFilters) is at least len(PointKeyFilters)+1. This helps
137 : // avoid allocations in Pebble internal code that mutates the slice.
138 : PointKeyFilters []BlockPropertyFilter
139 : // RangeKeyFilters can be usefd to avoid scanning tables and blocks in tables
140 : // when iterating over range keys. The same requirements that apply to
141 : // PointKeyFilters apply here too.
142 : RangeKeyFilters []BlockPropertyFilter
143 : // KeyTypes configures which types of keys to iterate over: point keys,
144 : // range keys, or both.
145 : KeyTypes IterKeyType
146 : // RangeKeyMasking can be used to enable automatic masking of point keys by
147 : // range keys. Range key masking is only supported during combined range key
148 : // and point key iteration mode (IterKeyTypePointsAndRanges).
149 : RangeKeyMasking RangeKeyMasking
150 :
151 : // OnlyReadGuaranteedDurable is an advanced option that is only supported by
152 : // the Reader implemented by DB. When set to true, only the guaranteed to be
153 : // durable state is visible in the iterator.
154 : // - This definition is made under the assumption that the FS implementation
155 : // is providing a durability guarantee when data is synced.
156 : // - The visible state represents a consistent point in the history of the
157 : // DB.
158 : // - The implementation is free to choose a conservative definition of what
159 : // is guaranteed durable. For simplicity, the current implementation
160 : // ignores memtables. A more sophisticated implementation could track the
161 : // highest seqnum that is synced to the WAL and published and use that as
162 : // the visible seqnum for an iterator. Note that the latter approach is
163 : // not strictly better than the former since we can have DBs that are (a)
164 : // synced more rarely than memtable flushes, (b) have no WAL. (a) is
165 : // likely to be true in a future CockroachDB context where the DB
166 : // containing the state machine may be rarely synced.
167 : // NB: this current implementation relies on the fact that memtables are
168 : // flushed in seqnum order, and any ingested sstables that happen to have a
169 : // lower seqnum than a non-flushed memtable don't have any overlapping keys.
170 : // This is the fundamental level invariant used in other code too, like when
171 : // merging iterators.
172 : //
173 : // Semantically, using this option provides the caller a "snapshot" as of
174 : // the time the most recent memtable was flushed. An alternate interface
175 : // would be to add a NewSnapshot variant. Creating a snapshot is heavier
176 : // weight than creating an iterator, so we have opted to support this
177 : // iterator option.
178 : OnlyReadGuaranteedDurable bool
179 : // UseL6Filters allows the caller to opt into reading filter blocks for L6
180 : // sstables. Helpful if a lot of SeekPrefixGEs are expected in quick
181 : // succession, that are also likely to not yield a single key. Filter blocks in
182 : // L6 can be relatively large, often larger than data blocks, so the benefit of
183 : // loading them in the cache is minimized if the probability of the key
184 : // existing is not low or if we just expect a one-time Seek (where loading the
185 : // data block directly is better).
186 : UseL6Filters bool
187 : // Category is used for categorized iterator stats. This should not be
188 : // changed by calling SetOptions.
189 : Category block.Category
190 :
191 : DebugRangeKeyStack bool
192 :
193 : // Internal options.
194 :
195 : logger Logger
196 : // Layer corresponding to this file. Only passed in if constructed by a
197 : // levelIter.
198 : layer manifest.Layer
199 : // disableLazyCombinedIteration is an internal testing option.
200 : disableLazyCombinedIteration bool
201 : // snapshotForHideObsoletePoints is specified for/by levelIter when opening
202 : // files and is used to decide whether to hide obsolete points. A value of 0
203 : // implies obsolete points should not be hidden.
204 : snapshotForHideObsoletePoints base.SeqNum
205 :
206 : // NB: If adding new Options, you must account for them in iterator
207 : // construction and Iterator.SetOptions.
208 : }
209 :
210 : // GetLowerBound returns the LowerBound or nil if the receiver is nil.
211 1 : func (o *IterOptions) GetLowerBound() []byte {
212 1 : if o == nil {
213 1 : return nil
214 1 : }
215 1 : return o.LowerBound
216 : }
217 :
218 : // GetUpperBound returns the UpperBound or nil if the receiver is nil.
219 1 : func (o *IterOptions) GetUpperBound() []byte {
220 1 : if o == nil {
221 1 : return nil
222 1 : }
223 1 : return o.UpperBound
224 : }
225 :
226 1 : func (o *IterOptions) pointKeys() bool {
227 1 : if o == nil {
228 0 : return true
229 0 : }
230 1 : return o.KeyTypes == IterKeyTypePointsOnly || o.KeyTypes == IterKeyTypePointsAndRanges
231 : }
232 :
233 1 : func (o *IterOptions) rangeKeys() bool {
234 1 : if o == nil {
235 0 : return false
236 0 : }
237 1 : return o.KeyTypes == IterKeyTypeRangesOnly || o.KeyTypes == IterKeyTypePointsAndRanges
238 : }
239 :
240 1 : func (o *IterOptions) getLogger() Logger {
241 1 : if o == nil || o.logger == nil {
242 1 : return DefaultLogger
243 1 : }
244 1 : return o.logger
245 : }
246 :
247 : // SpanIterOptions creates a SpanIterOptions from this IterOptions.
248 1 : func (o *IterOptions) SpanIterOptions() keyspan.SpanIterOptions {
249 1 : if o == nil {
250 1 : return keyspan.SpanIterOptions{}
251 1 : }
252 1 : return keyspan.SpanIterOptions{
253 1 : RangeKeyFilters: o.RangeKeyFilters,
254 1 : }
255 : }
256 :
257 : // scanInternalOptions is similar to IterOptions, meant for use with
258 : // scanInternalIterator.
259 : type scanInternalOptions struct {
260 : IterOptions
261 :
262 : category block.Category
263 :
264 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error
265 : visitRangeDel func(start, end []byte, seqNum SeqNum) error
266 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error
267 : visitSharedFile func(sst *SharedSSTMeta) error
268 : visitExternalFile func(sst *ExternalFile) error
269 :
270 : // includeObsoleteKeys specifies whether keys shadowed by newer internal keys
271 : // are exposed. If false, only one internal key per user key is exposed.
272 : includeObsoleteKeys bool
273 :
274 : // rateLimitFunc is used to limit the amount of bytes read per second.
275 : rateLimitFunc func(key *InternalKey, value LazyValue) error
276 : }
277 :
278 : // RangeKeyMasking configures automatic hiding of point keys by range keys. A
279 : // non-nil Suffix enables range-key masking. When enabled, range keys with
280 : // suffixes ≥ Suffix behave as masks. All point keys that are contained within a
281 : // masking range key's bounds and have suffixes greater than the range key's
282 : // suffix are automatically skipped.
283 : //
284 : // Specifically, when configured with a RangeKeyMasking.Suffix _s_, and there
285 : // exists a range key with suffix _r_ covering a point key with suffix _p_, and
286 : //
287 : // _s_ ≤ _r_ < _p_
288 : //
289 : // then the point key is elided.
290 : //
291 : // Range-key masking may only be used when iterating over both point keys and
292 : // range keys with IterKeyTypePointsAndRanges.
293 : type RangeKeyMasking struct {
294 : // Suffix configures which range keys may mask point keys. Only range keys
295 : // that are defined at suffixes greater than or equal to Suffix will mask
296 : // point keys.
297 : Suffix []byte
298 : // Filter is an optional field that may be used to improve performance of
299 : // range-key masking through a block-property filter defined over key
300 : // suffixes. If non-nil, Filter is called by Pebble to construct a
301 : // block-property filter mask at iterator creation. The filter is used to
302 : // skip whole point-key blocks containing point keys with suffixes greater
303 : // than a covering range-key's suffix.
304 : //
305 : // To use this functionality, the caller must create and configure (through
306 : // Options.BlockPropertyCollectors) a block-property collector that records
307 : // the maxmimum suffix contained within a block. The caller then must write
308 : // and provide a BlockPropertyFilterMask implementation on that same
309 : // property. See the BlockPropertyFilterMask type for more information.
310 : Filter func() BlockPropertyFilterMask
311 : }
312 :
313 : // BlockPropertyFilterMask extends the BlockPropertyFilter interface for use
314 : // with range-key masking. Unlike an ordinary block property filter, a
315 : // BlockPropertyFilterMask's filtering criteria is allowed to change when Pebble
316 : // invokes its SetSuffix method.
317 : //
318 : // When a Pebble iterator steps into a range key's bounds and the range key has
319 : // a suffix greater than or equal to RangeKeyMasking.Suffix, the range key acts
320 : // as a mask. The masking range key hides all point keys that fall within the
321 : // range key's bounds and have suffixes > the range key's suffix. Without a
322 : // filter mask configured, Pebble performs this hiding by stepping through point
323 : // keys and comparing suffixes. If large numbers of point keys are masked, this
324 : // requires Pebble to load, iterate through and discard a large number of
325 : // sstable blocks containing masked point keys.
326 : //
327 : // If a block-property collector and a filter mask are configured, Pebble may
328 : // skip loading some point-key blocks altogether. If a block's keys are known to
329 : // all fall within the bounds of the masking range key and the block was
330 : // annotated by a block-property collector with the maximal suffix, Pebble can
331 : // ask the filter mask to compare the property to the current masking range
332 : // key's suffix. If the mask reports no intersection, the block may be skipped.
333 : //
334 : // If unsuffixed and suffixed keys are written to the database, care must be
335 : // taken to avoid unintentionally masking un-suffixed keys located in the same
336 : // block as suffixed keys. One solution is to interpret unsuffixed keys as
337 : // containing the maximal suffix value, ensuring that blocks containing
338 : // unsuffixed keys are always loaded.
339 : type BlockPropertyFilterMask interface {
340 : BlockPropertyFilter
341 :
342 : // SetSuffix configures the mask with the suffix of a range key. The filter
343 : // should return false from Intersects whenever it's provided with a
344 : // property encoding a block's minimum suffix that's greater (according to
345 : // Compare) than the provided suffix.
346 : SetSuffix(suffix []byte) error
347 : }
348 :
349 : // WriteOptions hold the optional per-query parameters for Set and Delete
350 : // operations.
351 : //
352 : // Like Options, a nil *WriteOptions is valid and means to use the default
353 : // values.
354 : type WriteOptions struct {
355 : // Sync is whether to sync writes through the OS buffer cache and down onto
356 : // the actual disk, if applicable. Setting Sync is required for durability of
357 : // individual write operations but can result in slower writes.
358 : //
359 : // If false, and the process or machine crashes, then a recent write may be
360 : // lost. This is due to the recently written data being buffered inside the
361 : // process running Pebble. This differs from the semantics of a write system
362 : // call in which the data is buffered in the OS buffer cache and would thus
363 : // survive a process crash.
364 : //
365 : // The default value is true.
366 : Sync bool
367 : }
368 :
369 : // Sync specifies the default write options for writes which synchronize to
370 : // disk.
371 : var Sync = &WriteOptions{Sync: true}
372 :
373 : // NoSync specifies the default write options for writes which do not
374 : // synchronize to disk.
375 : var NoSync = &WriteOptions{Sync: false}
376 :
377 : // GetSync returns the Sync value or true if the receiver is nil.
378 1 : func (o *WriteOptions) GetSync() bool {
379 1 : return o == nil || o.Sync
380 1 : }
381 :
382 : // LevelOptions holds the optional per-level parameters.
383 : type LevelOptions struct {
384 : // BlockRestartInterval is the number of keys between restart points
385 : // for delta encoding of keys.
386 : //
387 : // The default value is 16 for L0, and the value from the previous level for
388 : // all other levels.
389 : BlockRestartInterval int
390 :
391 : // BlockSize is the target uncompressed size in bytes of each table block.
392 : //
393 : // The default value is 4096 for L0, and the value from the previous level for
394 : // all other levels.
395 : BlockSize int
396 :
397 : // BlockSizeThreshold finishes a block if the block size is larger than the
398 : // specified percentage of the target block size and adding the next entry
399 : // would cause the block to be larger than the target block size.
400 : //
401 : // The default value is 90 for L0, and the value from the previous level for
402 : // all other levels.
403 : BlockSizeThreshold int
404 :
405 : // Compression defines the per-block compression to use.
406 : //
407 : // The default value is Snappy for L0, or the function from the previous level
408 : // for all other levels.
409 : //
410 : // ApplyCompressionSettings can be used to initialize this field for all levels.
411 : Compression func() *sstable.CompressionProfile
412 :
413 : // FilterPolicy defines a filter algorithm (such as a Bloom filter) that can
414 : // reduce disk reads for Get calls.
415 : //
416 : // One such implementation is bloom.FilterPolicy(10) from the pebble/bloom
417 : // package.
418 : //
419 : // The default value for L0 is NoFilterPolicy (no filter), and the value from
420 : // the previous level for all other levels.
421 : FilterPolicy FilterPolicy
422 :
423 : // FilterType is a legacy field. The default and only possible value is
424 : // TableFilter.
425 : FilterType FilterType
426 :
427 : // IndexBlockSize is the target uncompressed size in bytes of each index
428 : // block. When the index block size is larger than this target, two-level
429 : // indexes are automatically enabled. Setting this option to a large value
430 : // (such as math.MaxInt32) disables the automatic creation of two-level
431 : // indexes.
432 : //
433 : // The default value is the value of BlockSize for L0, or the value from the
434 : // previous level for all other levels.
435 : IndexBlockSize int
436 : }
437 :
438 : // EnsureL0Defaults ensures that the L0 default values for the options have been
439 : // initialized.
440 1 : func (o *LevelOptions) EnsureL0Defaults() {
441 1 : if o.BlockRestartInterval <= 0 {
442 1 : o.BlockRestartInterval = base.DefaultBlockRestartInterval
443 1 : }
444 1 : if o.BlockSize <= 0 {
445 1 : o.BlockSize = base.DefaultBlockSize
446 1 : } else if o.BlockSize > sstable.MaximumRestartOffset {
447 0 : panic(errors.Errorf("BlockSize %d exceeds MaximumRestartOffset", o.BlockSize))
448 : }
449 1 : if o.BlockSizeThreshold <= 0 {
450 1 : o.BlockSizeThreshold = base.DefaultBlockSizeThreshold
451 1 : }
452 1 : if o.Compression == nil {
453 1 : o.Compression = func() *sstable.CompressionProfile { return sstable.SnappyCompression }
454 : }
455 1 : if o.FilterPolicy == nil {
456 1 : o.FilterPolicy = NoFilterPolicy
457 1 : }
458 1 : if o.IndexBlockSize <= 0 {
459 1 : o.IndexBlockSize = o.BlockSize
460 1 : }
461 : }
462 :
463 : // EnsureL1PlusDefaults ensures that the L1+ default values for the options have
464 : // been initialized. Requires the fully initialized options for the level above.
465 1 : func (o *LevelOptions) EnsureL1PlusDefaults(previousLevel *LevelOptions) {
466 1 : if o.BlockRestartInterval <= 0 {
467 1 : o.BlockRestartInterval = previousLevel.BlockRestartInterval
468 1 : }
469 1 : if o.BlockSize <= 0 {
470 1 : o.BlockSize = previousLevel.BlockSize
471 1 : } else if o.BlockSize > sstable.MaximumRestartOffset {
472 0 : panic(errors.Errorf("BlockSize %d exceeds MaximumRestartOffset", o.BlockSize))
473 : }
474 1 : if o.BlockSizeThreshold <= 0 {
475 1 : o.BlockSizeThreshold = previousLevel.BlockSizeThreshold
476 1 : }
477 1 : if o.Compression == nil {
478 1 : o.Compression = previousLevel.Compression
479 1 : }
480 1 : if o.FilterPolicy == nil {
481 1 : o.FilterPolicy = previousLevel.FilterPolicy
482 1 : }
483 1 : if o.IndexBlockSize <= 0 {
484 1 : o.IndexBlockSize = previousLevel.IndexBlockSize
485 1 : }
486 : }
487 :
488 : // Options holds the optional parameters for configuring pebble. These options
489 : // apply to the DB at large; per-query options are defined by the IterOptions
490 : // and WriteOptions types.
491 : type Options struct {
492 : // Sync sstables periodically in order to smooth out writes to disk. This
493 : // option does not provide any persistency guarantee, but is used to avoid
494 : // latency spikes if the OS automatically decides to write out a large chunk
495 : // of dirty filesystem buffers. This option only controls SSTable syncs; WAL
496 : // syncs are controlled by WALBytesPerSync.
497 : //
498 : // The default value is 512KB.
499 : BytesPerSync int
500 :
501 : // Cache is used to cache uncompressed blocks from sstables. If it is nil,
502 : // a block cache of CacheSize will be created for each DB.
503 : Cache *cache.Cache
504 : // CacheSize is used when Cache is not set. The default value is 8 MB.
505 : CacheSize int64
506 :
507 : // LoadBlockSema, if set, is used to limit the number of blocks that can be
508 : // loaded (i.e. read from the filesystem) in parallel. Each load acquires one
509 : // unit from the semaphore for the duration of the read.
510 : LoadBlockSema *fifo.Semaphore
511 :
512 : // Cleaner cleans obsolete files.
513 : //
514 : // The default cleaner uses the DeleteCleaner.
515 : Cleaner Cleaner
516 :
517 : // Local contains option that pertain to files stored on the local filesystem.
518 : Local struct {
519 : // ReadaheadConfig is used to retrieve the current readahead mode; it is
520 : // consulted whenever a read handle is initialized.
521 : ReadaheadConfig *ReadaheadConfig
522 :
523 : // TODO(radu): move BytesPerSync, LoadBlockSema, Cleaner here.
524 : }
525 :
526 : // Comparer defines a total ordering over the space of []byte keys: a 'less
527 : // than' relationship. The same comparison algorithm must be used for reads
528 : // and writes over the lifetime of the DB.
529 : //
530 : // The default value uses the same ordering as bytes.Compare.
531 : Comparer *Comparer
532 :
533 : // DebugCheck is invoked, if non-nil, whenever a new version is being
534 : // installed. Typically, this is set to pebble.DebugCheckLevels in tests
535 : // or tools only, to check invariants over all the data in the database.
536 : DebugCheck func(*DB) error
537 :
538 : // Disable the write-ahead log (WAL). Disabling the write-ahead log prohibits
539 : // crash recovery, but can improve performance if crash recovery is not
540 : // needed (e.g. when only temporary state is being stored in the database).
541 : //
542 : // TODO(peter): untested
543 : DisableWAL bool
544 :
545 : // ErrorIfExists causes an error on Open if the database already exists.
546 : // The error can be checked with errors.Is(err, ErrDBAlreadyExists).
547 : //
548 : // The default value is false.
549 : ErrorIfExists bool
550 :
551 : // ErrorIfNotExists causes an error on Open if the database does not already
552 : // exist. The error can be checked with errors.Is(err, ErrDBDoesNotExist).
553 : //
554 : // The default value is false which will cause a database to be created if it
555 : // does not already exist.
556 : ErrorIfNotExists bool
557 :
558 : // ErrorIfNotPristine causes an error on Open if the database already exists
559 : // and any operations have been performed on the database. The error can be
560 : // checked with errors.Is(err, ErrDBNotPristine).
561 : //
562 : // Note that a database that contained keys that were all subsequently deleted
563 : // may or may not trigger the error. Currently, we check if there are any live
564 : // SSTs or log records to replay.
565 : ErrorIfNotPristine bool
566 :
567 : // EventListener provides hooks to listening to significant DB events such as
568 : // flushes, compactions, and table deletion.
569 : EventListener *EventListener
570 :
571 : // Experimental contains experimental options which are off by default.
572 : // These options are temporary and will eventually either be deleted, moved
573 : // out of the experimental group, or made the non-adjustable default. These
574 : // options may change at any time, so do not rely on them.
575 : Experimental struct {
576 : // The threshold of L0 read-amplification at which compaction concurrency
577 : // is enabled (if CompactionDebtConcurrency was not already exceeded).
578 : // Every multiple of this value enables another concurrent
579 : // compaction up to CompactionConcurrencyRange.
580 : L0CompactionConcurrency int
581 :
582 : // CompactionDebtConcurrency controls the threshold of compaction debt
583 : // at which additional compaction concurrency slots are added. For every
584 : // multiple of this value in compaction debt bytes, an additional
585 : // concurrent compaction is added. This works "on top" of
586 : // L0CompactionConcurrency, so the higher of the count of compaction
587 : // concurrency slots as determined by the two options is chosen.
588 : CompactionDebtConcurrency uint64
589 :
590 : // CompactionGarbageFractionForMaxConcurrency is the fraction of garbage
591 : // due to DELs and RANGEDELs that causes MaxConcurrentCompactions to be
592 : // allowed. Concurrent compactions are allowed in a linear manner upto
593 : // this limit being reached. A value <= 0.0 disables adding concurrency
594 : // due to garbage.
595 : CompactionGarbageFractionForMaxConcurrency func() float64
596 :
597 : // IngestSplit, if it returns true, allows for ingest-time splitting of
598 : // existing sstables into two virtual sstables to allow ingestion sstables to
599 : // slot into a lower level than they otherwise would have.
600 : IngestSplit func() bool
601 :
602 : // ReadCompactionRate controls the frequency of read triggered
603 : // compactions by adjusting `AllowedSeeks` in manifest.TableMetadata:
604 : //
605 : // AllowedSeeks = FileSize / ReadCompactionRate
606 : //
607 : // From LevelDB:
608 : // ```
609 : // We arrange to automatically compact this file after
610 : // a certain number of seeks. Let's assume:
611 : // (1) One seek costs 10ms
612 : // (2) Writing or reading 1MB costs 10ms (100MB/s)
613 : // (3) A compaction of 1MB does 25MB of IO:
614 : // 1MB read from this level
615 : // 10-12MB read from next level (boundaries may be misaligned)
616 : // 10-12MB written to next level
617 : // This implies that 25 seeks cost the same as the compaction
618 : // of 1MB of data. I.e., one seek costs approximately the
619 : // same as the compaction of 40KB of data. We are a little
620 : // conservative and allow approximately one seek for every 16KB
621 : // of data before triggering a compaction.
622 : // ```
623 : ReadCompactionRate int64
624 :
625 : // ReadSamplingMultiplier is a multiplier for the readSamplingPeriod in
626 : // iterator.maybeSampleRead() to control the frequency of read sampling
627 : // to trigger a read triggered compaction. A value of -1 prevents sampling
628 : // and disables read triggered compactions. The default is 1 << 4. which
629 : // gets multiplied with a constant of 1 << 16 to yield 1 << 20 (1MB).
630 : ReadSamplingMultiplier int64
631 :
632 : // NumDeletionsThreshold defines the minimum number of point tombstones
633 : // that must be present in a single data block for that block to be
634 : // considered tombstone-dense for the purposes of triggering a
635 : // tombstone density compaction. Data blocks may also be considered
636 : // tombstone-dense if they meet the criteria defined by
637 : // DeletionSizeRatioThreshold below. Tombstone-dense blocks are identified
638 : // when sstables are written, and so this is effectively an option for
639 : // sstable writers. The default value is 100.
640 : NumDeletionsThreshold int
641 :
642 : // DeletionSizeRatioThreshold defines the minimum ratio of the size of
643 : // point tombstones to the size of a data block that must be reached
644 : // for that block to be considered tombstone-dense for the purposes of
645 : // triggering a tombstone density compaction. Data blocks may also be
646 : // considered tombstone-dense if they meet the criteria defined by
647 : // NumDeletionsThreshold above. Tombstone-dense blocks are identified
648 : // when sstables are written, and so this is effectively an option for
649 : // sstable writers. The default value is 0.5.
650 : DeletionSizeRatioThreshold float32
651 :
652 : // TombstoneDenseCompactionThreshold is the minimum percent of data
653 : // blocks in a table that must be tombstone-dense for that table to be
654 : // eligible for a tombstone density compaction. It should be defined as a
655 : // ratio out of 1. The default value is 0.10.
656 : //
657 : // If multiple tables are eligible for a tombstone density compaction, then
658 : // tables with a higher percent of tombstone-dense blocks are still
659 : // prioritized for compaction.
660 : //
661 : // A zero or negative value disables tombstone density compactions.
662 : TombstoneDenseCompactionThreshold float64
663 :
664 : // FileCacheShards is the number of shards per file cache.
665 : // Reducing the value can reduce the number of idle goroutines per DB
666 : // instance which can be useful in scenarios with a lot of DB instances
667 : // and a large number of CPUs, but doing so can lead to higher contention
668 : // in the file cache and reduced performance.
669 : //
670 : // The default value is the number of logical CPUs, which can be
671 : // limited by runtime.GOMAXPROCS.
672 : FileCacheShards int
673 :
674 : // ValidateOnIngest schedules validation of sstables after they have
675 : // been ingested.
676 : //
677 : // By default, this value is false.
678 : ValidateOnIngest bool
679 :
680 : // LevelMultiplier configures the size multiplier used to determine the
681 : // desired size of each level of the LSM. Defaults to 10.
682 : LevelMultiplier int
683 :
684 : // MultiLevelCompactionHeuristic determines whether to add an additional
685 : // level to a conventional two level compaction.
686 : MultiLevelCompactionHeuristic func() MultiLevelHeuristic
687 :
688 : // EnableColumnarBlocks is used to decide whether to enable writing
689 : // TableFormatPebblev5 sstables. This setting is only respected by
690 : // FormatColumnarBlocks. In lower format major versions, the
691 : // TableFormatPebblev5 format is prohibited. If EnableColumnarBlocks is
692 : // nil and the DB is at FormatColumnarBlocks, the DB defaults to not
693 : // writing columnar blocks.
694 : EnableColumnarBlocks func() bool
695 :
696 : // EnableValueBlocks is used to decide whether to enable writing
697 : // TableFormatPebblev3 sstables. This setting is only respected by a
698 : // specific subset of format major versions: FormatSSTableValueBlocks,
699 : // FormatFlushableIngest and FormatPrePebblev1MarkedCompacted. In lower
700 : // format major versions, value blocks are never enabled. In higher
701 : // format major versions, value blocks are always enabled.
702 : EnableValueBlocks func() bool
703 :
704 : // ShortAttributeExtractor is used iff EnableValueBlocks() returns true
705 : // (else ignored). If non-nil, a ShortAttribute can be extracted from the
706 : // value and stored with the key, when the value is stored elsewhere.
707 : ShortAttributeExtractor ShortAttributeExtractor
708 :
709 : // DisableIngestAsFlushable disables lazy ingestion of sstables through
710 : // a WAL write and memtable rotation. Only effectual if the format
711 : // major version is at least `FormatFlushableIngest`.
712 : DisableIngestAsFlushable func() bool
713 :
714 : // RemoteStorage enables use of remote storage (e.g. S3) for storing
715 : // sstables. Setting this option enables use of CreateOnShared option and
716 : // allows ingestion of external files.
717 : RemoteStorage remote.StorageFactory
718 :
719 : // If CreateOnShared is non-zero, new sstables are created on remote storage
720 : // (using CreateOnSharedLocator and with the appropriate
721 : // CreateOnSharedStrategy). These sstables can be shared between different
722 : // Pebble instances; the lifecycle of such objects is managed by the
723 : // remote.Storage constructed by options.RemoteStorage.
724 : //
725 : // Can only be used when RemoteStorage is set (and recognizes
726 : // CreateOnSharedLocator).
727 : CreateOnShared remote.CreateOnSharedStrategy
728 : CreateOnSharedLocator remote.Locator
729 :
730 : // CacheSizeBytesBytes is the size of the on-disk block cache for objects
731 : // on shared storage in bytes. If it is 0, no cache is used.
732 : SecondaryCacheSizeBytes int64
733 :
734 : // EnableDeleteOnlyCompactionExcises enables delete-only compactions to also
735 : // apply delete-only compaction hints on sstables that partially overlap
736 : // with it. This application happens through an excise, similar to
737 : // the excise phase of IngestAndExcise.
738 : EnableDeleteOnlyCompactionExcises func() bool
739 :
740 : // CompactionScheduler, if set, is used to limit concurrent compactions as
741 : // well as to pace compactions already chosen. If nil, a default scheduler
742 : // is created and used.
743 : CompactionScheduler CompactionScheduler
744 :
745 : UserKeyCategories UserKeyCategories
746 :
747 : // ValueSeparationPolicy controls the policy for separating values into
748 : // external blob files. If nil, value separation defaults to disabled.
749 : // The value separation policy is ignored if EnableColumnarBlocks() is
750 : // false.
751 : ValueSeparationPolicy func() ValueSeparationPolicy
752 :
753 : // SpanPolicyFunc is used to determine the SpanPolicy for a key region.
754 : SpanPolicyFunc SpanPolicyFunc
755 : }
756 :
757 : // Filters is a map from filter policy name to filter policy. It is used for
758 : // debugging tools which may be used on multiple databases configured with
759 : // different filter policies. It is not necessary to populate this filters
760 : // map during normal usage of a DB (it will be done automatically by
761 : // EnsureDefaults).
762 : Filters map[string]FilterPolicy
763 :
764 : // FlushDelayDeleteRange configures how long the database should wait before
765 : // forcing a flush of a memtable that contains a range deletion. Disk space
766 : // cannot be reclaimed until the range deletion is flushed. No automatic
767 : // flush occurs if zero.
768 : FlushDelayDeleteRange time.Duration
769 :
770 : // FlushDelayRangeKey configures how long the database should wait before
771 : // forcing a flush of a memtable that contains a range key. Range keys in
772 : // the memtable prevent lazy combined iteration, so it's desirable to flush
773 : // range keys promptly. No automatic flush occurs if zero.
774 : FlushDelayRangeKey time.Duration
775 :
776 : // FlushSplitBytes denotes the target number of bytes per sublevel in
777 : // each flush split interval (i.e. range between two flush split keys)
778 : // in L0 sstables. When set to zero, only a single sstable is generated
779 : // by each flush. When set to a non-zero value, flushes are split at
780 : // points to meet L0's TargetFileSize, any grandparent-related overlap
781 : // options, and at boundary keys of L0 flush split intervals (which are
782 : // targeted to contain around FlushSplitBytes bytes in each sublevel
783 : // between pairs of boundary keys). Splitting sstables during flush
784 : // allows increased compaction flexibility and concurrency when those
785 : // tables are compacted to lower levels.
786 : FlushSplitBytes int64
787 :
788 : // FormatMajorVersion sets the format of on-disk files. It is
789 : // recommended to set the format major version to an explicit
790 : // version, as the default may change over time.
791 : //
792 : // At Open if the existing database is formatted using a later
793 : // format major version that is known to this version of Pebble,
794 : // Pebble will continue to use the later format major version. If
795 : // the existing database's version is unknown, the caller may use
796 : // FormatMostCompatible and will be able to open the database
797 : // regardless of its actual version.
798 : //
799 : // If the existing database is formatted using a format major
800 : // version earlier than the one specified, Open will automatically
801 : // ratchet the database to the specified format major version.
802 : FormatMajorVersion FormatMajorVersion
803 :
804 : // FS provides the interface for persistent file storage.
805 : //
806 : // The default value uses the underlying operating system's file system.
807 : FS vfs.FS
808 :
809 : // KeySchema is the name of the key schema that should be used when writing
810 : // new sstables. There must be a key schema with this name defined in
811 : // KeySchemas. If not set, colblk.DefaultKeySchema is used to construct a
812 : // default key schema.
813 : KeySchema string
814 :
815 : // KeySchemas defines the set of known schemas of user keys. When columnar
816 : // blocks are in use (see FormatColumnarBlocks), the user may specify how a
817 : // key should be decomposed into columns. Each KeySchema must have a unique
818 : // name. The schema named by Options.KeySchema is used while writing
819 : // sstables during flushes and compactions.
820 : //
821 : // Multiple KeySchemas may be used over the lifetime of a database. Once a
822 : // KeySchema is used, it must be provided in KeySchemas in subsequent calls
823 : // to Open for perpetuity.
824 : KeySchemas sstable.KeySchemas
825 :
826 : // Lock, if set, must be a database lock acquired through LockDirectory for
827 : // the same directory passed to Open. If provided, Open will skip locking
828 : // the directory. Closing the database will not release the lock, and it's
829 : // the responsibility of the caller to release the lock after closing the
830 : // database.
831 : //
832 : // Open will enforce that the Lock passed locks the same directory passed to
833 : // Open. Concurrent calls to Open using the same Lock are detected and
834 : // prohibited.
835 : Lock *Lock
836 :
837 : // The count of L0 files necessary to trigger an L0 compaction.
838 : L0CompactionFileThreshold int
839 :
840 : // The amount of L0 read-amplification necessary to trigger an L0 compaction.
841 : L0CompactionThreshold int
842 :
843 : // Hard limit on L0 read-amplification, computed as the number of L0
844 : // sublevels. Writes are stopped when this threshold is reached.
845 : L0StopWritesThreshold int
846 :
847 : // The maximum number of bytes for LBase. The base level is the level which
848 : // L0 is compacted into. The base level is determined dynamically based on
849 : // the existing data in the LSM. The maximum number of bytes for other levels
850 : // is computed dynamically based on the base level's maximum size. When the
851 : // maximum number of bytes for a level is exceeded, compaction is requested.
852 : LBaseMaxBytes int64
853 :
854 : // TargetFileSizes contains the target file size for each level, ignoring
855 : // unpopulated levels. Specifically:
856 : // - TargetFileSizes[0] is the target file size for L0;
857 : // - TargetFileSizes[1] is the target file size for Lbase;
858 : // - TargetFileSizes[2] is the target file size for Lbase+1;
859 : // and so on.
860 : //
861 : // The default value for TargetFileSizes[0] is 2MB.
862 : // The default value for TargetFileSizes[i] is TargetFileSizes[i-1] * 2.
863 : TargetFileSizes [manifest.NumLevels]int64
864 :
865 : // Per-level options. Levels[i] contains the options for Li (regardless of
866 : // what Lbase is).
867 : Levels [manifest.NumLevels]LevelOptions
868 :
869 : // LoggerAndTracer will be used, if non-nil, else Logger will be used and
870 : // tracing will be a noop.
871 :
872 : // Logger used to write log messages.
873 : //
874 : // The default logger uses the Go standard library log package.
875 : Logger Logger
876 : // LoggerAndTracer is used for writing log messages and traces.
877 : LoggerAndTracer LoggerAndTracer
878 :
879 : // MaxManifestFileSize is the maximum size the MANIFEST file is allowed to
880 : // become. When the MANIFEST exceeds this size it is rolled over and a new
881 : // MANIFEST is created.
882 : MaxManifestFileSize int64
883 :
884 : // MaxOpenFiles is a soft limit on the number of open files that can be
885 : // used by the DB.
886 : //
887 : // The default value is 1000.
888 : MaxOpenFiles int
889 :
890 : // The size of a MemTable in steady state. The actual MemTable size starts at
891 : // min(256KB, MemTableSize) and doubles for each subsequent MemTable up to
892 : // MemTableSize. This reduces the memory pressure caused by MemTables for
893 : // short lived (test) DB instances. Note that more than one MemTable can be
894 : // in existence since flushing a MemTable involves creating a new one and
895 : // writing the contents of the old one in the
896 : // background. MemTableStopWritesThreshold places a hard limit on the size of
897 : // the queued MemTables.
898 : //
899 : // The default value is 4MB.
900 : MemTableSize uint64
901 :
902 : // Hard limit on the number of queued of MemTables. Writes are stopped when
903 : // the sum of the queued memtable sizes exceeds:
904 : // MemTableStopWritesThreshold * MemTableSize.
905 : //
906 : // This value should be at least 2 or writes will stop whenever a MemTable is
907 : // being flushed.
908 : //
909 : // The default value is 2.
910 : MemTableStopWritesThreshold int
911 :
912 : // Merger defines the associative merge operation to use for merging values
913 : // written with {Batch,DB}.Merge.
914 : //
915 : // The default merger concatenates values.
916 : Merger *Merger
917 :
918 : // CompactionConcurrencyRange returns a [lower, upper] range for the number of
919 : // compactions Pebble runs in parallel (with the caveats below), not including
920 : // download compactions (which have a separate limit specified by
921 : // MaxConcurrentDownloads).
922 : //
923 : // The lower value is the concurrency allowed under normal circumstances.
924 : // Pebble can dynamically increase the concurrency based on heuristics (like
925 : // high read amplification or compaction debt) up to the maximum.
926 : //
927 : // The upper value is a rough upper bound since delete-only compactions (a) do
928 : // not use the CompactionScheduler, and (b) the CompactionScheduler may use
929 : // other criteria to decide on how many compactions to permit.
930 : //
931 : // Elaborating on (b), when the ConcurrencyLimitScheduler is being used, the
932 : // value returned by DB.GetAllowedWithoutPermission fully controls how many
933 : // compactions get to run. Other CompactionSchedulers may use additional
934 : // criteria, like resource availability.
935 : //
936 : // Elaborating on (a), we don't use the CompactionScheduler to schedule
937 : // delete-only compactions since they are expected to be almost free from a
938 : // CPU and disk usage perspective. Since the CompactionScheduler does not
939 : // know about their existence, the total running count can exceed this
940 : // value. For example, consider CompactionConcurrencyRange returns 3, and the
941 : // current value returned from DB.GetAllowedWithoutPermission is also 3. Say
942 : // 3 delete-only compactions are also running. Then the
943 : // ConcurrencyLimitScheduler can also start 3 other compactions, for a total
944 : // of 6.
945 : //
946 : // DB.GetAllowedWithoutPermission returns a value in the interval
947 : // [lower, upper]. A value > lower is returned:
948 : // - when L0 read-amplification passes the L0CompactionConcurrency threshold;
949 : // - when compaction debt passes the CompactionDebtConcurrency threshold;
950 : // - when there are multiple manual compactions waiting to run.
951 : //
952 : // lower and upper must be greater than 0. If lower > upper, then upper is
953 : // used for both.
954 : //
955 : // The default values are 1, 1.
956 : CompactionConcurrencyRange func() (lower, upper int)
957 :
958 : // MaxConcurrentDownloads specifies the maximum number of download
959 : // compactions. These are compactions that copy an external file to the local
960 : // store.
961 : //
962 : // This limit is independent of CompactionConcurrencyRange; at any point in
963 : // time, we may be running CompactionConcurrencyRange non-download compactions
964 : // and MaxConcurrentDownloads download compactions.
965 : //
966 : // MaxConcurrentDownloads() must be greater than 0.
967 : //
968 : // The default value is 1.
969 : MaxConcurrentDownloads func() int
970 :
971 : // DisableAutomaticCompactions dictates whether automatic compactions are
972 : // scheduled or not. The default is false (enabled). This option is only used
973 : // externally when running a manual compaction, and internally for tests.
974 : DisableAutomaticCompactions bool
975 :
976 : // DisableConsistencyCheck disables the consistency check that is performed on
977 : // open. Should only be used when a database cannot be opened normally (e.g.
978 : // some of the tables don't exist / aren't accessible).
979 : DisableConsistencyCheck bool
980 :
981 : // DisableTableStats dictates whether tables should be loaded asynchronously
982 : // to compute statistics that inform compaction heuristics. The collection
983 : // of table stats improves compaction of tombstones, reclaiming disk space
984 : // more quickly and in some cases reducing write amplification in the
985 : // presence of tombstones. Disabling table stats may be useful in tests
986 : // that require determinism as the asynchronicity of table stats collection
987 : // introduces significant nondeterminism.
988 : DisableTableStats bool
989 :
990 : // NoSyncOnClose decides whether the Pebble instance will enforce a
991 : // close-time synchronization (e.g., fdatasync() or sync_file_range())
992 : // on files it writes to. Setting this to true removes the guarantee for a
993 : // sync on close. Some implementations can still issue a non-blocking sync.
994 : NoSyncOnClose bool
995 :
996 : // NumPrevManifest is the number of non-current or older manifests which
997 : // we want to keep around for debugging purposes. By default, we're going
998 : // to keep one older manifest.
999 : NumPrevManifest int
1000 :
1001 : // ReadOnly indicates that the DB should be opened in read-only mode. Writes
1002 : // to the DB will return an error, background compactions are disabled, and
1003 : // the flush that normally occurs after replaying the WAL at startup is
1004 : // disabled.
1005 : ReadOnly bool
1006 :
1007 : // FileCache is an initialized FileCache which should be set as an
1008 : // option if the DB needs to be initialized with a pre-existing file cache.
1009 : // If FileCache is nil, then a file cache which is unique to the DB instance
1010 : // is created. FileCache can be shared between db instances by setting it here.
1011 : // The FileCache set here must use the same underlying cache as Options.Cache
1012 : // and pebble will panic otherwise.
1013 : FileCache *FileCache
1014 :
1015 : // BlockPropertyCollectors is a list of BlockPropertyCollector creation
1016 : // functions. A new BlockPropertyCollector is created for each sstable
1017 : // built and lives for the lifetime of writing that table.
1018 : BlockPropertyCollectors []func() BlockPropertyCollector
1019 :
1020 : // WALBytesPerSync sets the number of bytes to write to a WAL before calling
1021 : // Sync on it in the background. Just like with BytesPerSync above, this
1022 : // helps smooth out disk write latencies, and avoids cases where the OS
1023 : // writes a lot of buffered data to disk at once. However, this is less
1024 : // necessary with WALs, as many write operations already pass in
1025 : // Sync = true.
1026 : //
1027 : // The default value is 0, i.e. no background syncing. This matches the
1028 : // default behaviour in RocksDB.
1029 : WALBytesPerSync int
1030 :
1031 : // WALDir specifies the directory to store write-ahead logs (WALs) in. If
1032 : // empty (the default), WALs will be stored in the same directory as sstables
1033 : // (i.e. the directory passed to pebble.Open).
1034 : WALDir string
1035 :
1036 : // WALFailover may be set to configure Pebble to monitor writes to its
1037 : // write-ahead log and failover to writing write-ahead log entries to a
1038 : // secondary location (eg, a separate physical disk). WALFailover may be
1039 : // used to improve write availability in the presence of transient disk
1040 : // unavailability.
1041 : WALFailover *WALFailoverOptions
1042 :
1043 : // WALRecoveryDirs is a list of additional directories that should be
1044 : // scanned for the existence of additional write-ahead logs. WALRecoveryDirs
1045 : // is expected to be used when starting Pebble with a new WALDir or a new
1046 : // WALFailover configuration. The directories associated with the previous
1047 : // configuration may still contain WALs that are required for recovery of
1048 : // the current database state.
1049 : //
1050 : // If a previous WAL configuration may have stored WALs elsewhere but there
1051 : // is not a corresponding entry in WALRecoveryDirs, Open will error.
1052 : WALRecoveryDirs []wal.Dir
1053 :
1054 : // WALMinSyncInterval is the minimum duration between syncs of the WAL. If
1055 : // WAL syncs are requested faster than this interval, they will be
1056 : // artificially delayed. Introducing a small artificial delay (500us) between
1057 : // WAL syncs can allow more operations to arrive and reduce IO operations
1058 : // while having a minimal impact on throughput. This option is supplied as a
1059 : // closure in order to allow the value to be changed dynamically. The default
1060 : // value is 0.
1061 : //
1062 : // TODO(peter): rather than a closure, should there be another mechanism for
1063 : // changing options dynamically?
1064 : WALMinSyncInterval func() time.Duration
1065 :
1066 : // The controls below manage deletion pacing, which slows down
1067 : // deletions when compactions finish or when readers close and
1068 : // obsolete files must be cleaned up. Rapid deletion of many
1069 : // files simultaneously can increase disk latency on certain
1070 : // SSDs, and this functionality helps protect against that.
1071 :
1072 : // TargetByteDeletionRate is the rate (in bytes per second) at which sstable file
1073 : // deletions are limited to (under normal circumstances).
1074 : //
1075 : // This value is only a best-effort target; the effective rate can be
1076 : // higher if deletions are falling behind or disk space is running low.
1077 : //
1078 : // Setting this to 0 disables deletion pacing, which is also the default.
1079 : TargetByteDeletionRate int
1080 :
1081 : // FreeSpaceThresholdBytes specifies the minimum amount of free disk space that Pebble
1082 : // attempts to maintain. If free disk space drops below this threshold, deletions
1083 : // are accelerated above TargetByteDeletionRate until the threshold is restored.
1084 : // Default is 16GB.
1085 : FreeSpaceThresholdBytes uint64
1086 :
1087 : // FreeSpaceTimeframe sets the duration (in seconds) within which Pebble attempts
1088 : // to restore the free disk space back to FreeSpaceThreshold. A lower value means
1089 : // more aggressive deletions. Default is 10s.
1090 : FreeSpaceTimeframe time.Duration
1091 :
1092 : // ObsoleteBytesMaxRatio specifies the maximum allowed ratio of obsolete files to
1093 : // live files. If this ratio is exceeded, Pebble speeds up deletions above the
1094 : // TargetByteDeletionRate until the ratio is restored. Default is 0.20.
1095 : ObsoleteBytesMaxRatio float64
1096 :
1097 : // ObsoleteBytesTimeframe sets the duration (in seconds) within which Pebble aims
1098 : // to restore the obsolete-to-live bytes ratio below ObsoleteBytesMaxRatio. A lower
1099 : // value means more aggressive deletions. Default is 300s.
1100 : ObsoleteBytesTimeframe time.Duration
1101 :
1102 : // EnableSQLRowSpillMetrics specifies whether the Pebble instance will only be used
1103 : // to temporarily persist data spilled to disk for row-oriented SQL query execution.
1104 : EnableSQLRowSpillMetrics bool
1105 :
1106 : // AllocatorSizeClasses provides a sorted list containing the supported size
1107 : // classes of the underlying memory allocator. This provides hints to the
1108 : // sstable block writer's flushing policy to select block sizes that
1109 : // preemptively reduce internal fragmentation when loaded into the block cache.
1110 : AllocatorSizeClasses []int
1111 :
1112 : // private options are only used by internal tests or are used internally
1113 : // for facilitating upgrade paths of unconfigurable functionality.
1114 : private struct {
1115 : // disableDeleteOnlyCompactions prevents the scheduling of delete-only
1116 : // compactions that drop sstables wholy covered by range tombstones or
1117 : // range key tombstones.
1118 : disableDeleteOnlyCompactions bool
1119 :
1120 : // disableElisionOnlyCompactions prevents the scheduling of elision-only
1121 : // compactions that rewrite sstables in place in order to elide obsolete
1122 : // keys.
1123 : disableElisionOnlyCompactions bool
1124 :
1125 : // disableLazyCombinedIteration is a private option used by the
1126 : // metamorphic tests to test equivalence between lazy-combined iteration
1127 : // and constructing the range-key iterator upfront. It's a private
1128 : // option to avoid littering the public interface with options that we
1129 : // do not want to allow users to actually configure.
1130 : disableLazyCombinedIteration bool
1131 :
1132 : // testingAlwaysWaitForCleanup is set by some tests to force waiting for
1133 : // obsolete file deletion (to make events deterministic).
1134 : testingAlwaysWaitForCleanup bool
1135 :
1136 : // fsCloser holds a closer that should be invoked after a DB using these
1137 : // Options is closed. This is used to automatically stop the
1138 : // long-running goroutine associated with the disk-health-checking FS.
1139 : // See the initialization of FS in EnsureDefaults. Note that care has
1140 : // been taken to ensure that it is still safe to continue using the FS
1141 : // after this closer has been invoked. However, if write operations
1142 : // against the FS are made after the DB is closed, the FS may leak a
1143 : // goroutine indefinitely.
1144 : fsCloser io.Closer
1145 : }
1146 : }
1147 :
1148 : // ValueSeparationPolicy controls the policy for separating values into
1149 : // external blob files.
1150 : type ValueSeparationPolicy struct {
1151 : // Enabled controls whether value separation is enabled.
1152 : Enabled bool
1153 : // MinimumSize imposes a lower bound on the size of values that can be
1154 : // separated into a blob file. Values smaller than this are always written
1155 : // to the sstable (but may still be written to a value block within the
1156 : // sstable).
1157 : //
1158 : // MinimumSize must be > 0.
1159 : MinimumSize int
1160 : // MaxBlobReferenceDepth limits the number of potentially overlapping (in
1161 : // the keyspace) blob files that can be referenced by a single sstable. If a
1162 : // compaction may produce an output sstable referencing more than this many
1163 : // overlapping blob files, the compaction will instead rewrite referenced
1164 : // values into new blob files.
1165 : //
1166 : // MaxBlobReferenceDepth must be > 0.
1167 : MaxBlobReferenceDepth int
1168 : // RewriteMinimumAge specifies how old a blob file must be in order for it
1169 : // to be eligible for a rewrite that reclaims disk space. Lower values
1170 : // reduce space amplification at the cost of write amplification
1171 : RewriteMinimumAge time.Duration
1172 : // TargetGarbageRatio is a value in the range [0, 1.0] and configures how
1173 : // aggressively blob files should be written in order to reduce space
1174 : // amplification induced by value separation. As compactions rewrite blob
1175 : // files, data may be duplicated. Older blob files containing the
1176 : // duplicated data may need to remain because other sstables are referencing
1177 : // other values contained in the same file.
1178 : //
1179 : // The DB can rewrite these blob files in place in order to reduce this
1180 : // space amplification, but this incurs write amplification. This option
1181 : // configures how much garbage may accrue before the DB will attempt to
1182 : // rewrite blob files to reduce it. A value of 0.20 indicates that once 20%
1183 : // of values in blob files are unreferenced, the DB should attempt to
1184 : // rewrite blob files to reclaim disk space.
1185 : //
1186 : // A value of 1.0 indicates that the DB should never attempt to rewrite blob
1187 : // files.
1188 : TargetGarbageRatio float64
1189 : }
1190 :
1191 : // SpanPolicy contains policies that can vary by key range. The zero value is
1192 : // the default value.
1193 : type SpanPolicy struct {
1194 : // Prefer a faster compression algorithm for the keys in this span.
1195 : //
1196 : // This is useful for keys that are frequently read or written but which don't
1197 : // amount to a significant amount of space.
1198 : PreferFastCompression bool
1199 :
1200 : // DisableValueSeparationBySuffix disables discriminating KVs depending on
1201 : // suffix.
1202 : //
1203 : // Among a set of keys with the same prefix, Pebble's default heuristics
1204 : // optimize access to the KV with the smallest suffix. This is useful for MVCC
1205 : // keys (where the smallest suffix is the latest version), but should be
1206 : // disabled for keys where the suffix does not correspond to a version.
1207 : DisableValueSeparationBySuffix bool
1208 :
1209 : // ValueStoragePolicy is a hint used to determine where to store the values
1210 : // for KVs.
1211 : ValueStoragePolicy ValueStoragePolicy
1212 : }
1213 :
1214 : // String returns a string representation of the SpanPolicy.
1215 1 : func (p SpanPolicy) String() string {
1216 1 : var sb strings.Builder
1217 1 : if p.PreferFastCompression {
1218 0 : sb.WriteString("fast-compression,")
1219 0 : }
1220 1 : if p.DisableValueSeparationBySuffix {
1221 0 : sb.WriteString("disable-value-separation-by-suffix,")
1222 0 : }
1223 1 : switch p.ValueStoragePolicy {
1224 1 : case ValueStorageLowReadLatency:
1225 1 : sb.WriteString("low-read-latency,")
1226 1 : case ValueStorageLatencyTolerant:
1227 1 : sb.WriteString("latency-tolerant,")
1228 : }
1229 1 : return strings.TrimSuffix(sb.String(), ",")
1230 : }
1231 :
1232 : // ValueStoragePolicy is a hint used to determine where to store the values for
1233 : // KVs.
1234 : type ValueStoragePolicy uint8
1235 :
1236 : const (
1237 : // ValueStorageDefault is the default value; Pebble will respect global
1238 : // configuration for value blocks and value separation.
1239 : ValueStorageDefault ValueStoragePolicy = iota
1240 :
1241 : // ValueStorageLowReadLatency indicates Pebble should prefer storing values
1242 : // in-place.
1243 : ValueStorageLowReadLatency
1244 :
1245 : // ValueStorageLatencyTolerant indicates value retrieval can tolerate
1246 : // additional latency, so Pebble should aggressively prefer storing values
1247 : // separately if it can reduce write amplification.
1248 : //
1249 : // If the global Options' enable value separation, Pebble may choose to
1250 : // separate values under the LatencyTolerant policy even if they do not meet
1251 : // the minimum size threshold of the global Options' ValueSeparationPolicy.
1252 : ValueStorageLatencyTolerant
1253 : )
1254 :
1255 : // SpanPolicyFunc is used to determine the SpanPolicy for a key region.
1256 : //
1257 : // The returned policy is valid from the start key until (and not including) the
1258 : // end key.
1259 : //
1260 : // A flush or compaction will call this function once for the first key to be
1261 : // output. If the compaction reaches the end key, the current output sst is
1262 : // finished and the function is called again.
1263 : //
1264 : // The end key can be empty, in which case the policy is valid for the entire
1265 : // keyspace after startKey.
1266 : type SpanPolicyFunc func(startKey []byte) (policy SpanPolicy, endKey []byte, err error)
1267 :
1268 : // SpanAndPolicy defines a key range and the policy to apply to it.
1269 : type SpanAndPolicy struct {
1270 : KeyRange KeyRange
1271 : Policy SpanPolicy
1272 : }
1273 :
1274 : // MakeStaticSpanPolicyFunc returns a SpanPolicyFunc that applies a given policy
1275 : // to the given span (and the default policy outside the span). The supplied
1276 : // policies must be non-overlapping in key range.
1277 1 : func MakeStaticSpanPolicyFunc(cmp base.Compare, inputPolicies ...SpanAndPolicy) SpanPolicyFunc {
1278 1 : // Collect all the boundaries of the input policies, sort and deduplicate them.
1279 1 : uniqueKeys := make([][]byte, 0, 2*len(inputPolicies))
1280 1 : for i := range inputPolicies {
1281 1 : uniqueKeys = append(uniqueKeys, inputPolicies[i].KeyRange.Start)
1282 1 : uniqueKeys = append(uniqueKeys, inputPolicies[i].KeyRange.End)
1283 1 : }
1284 1 : slices.SortFunc(uniqueKeys, cmp)
1285 1 : uniqueKeys = slices.CompactFunc(uniqueKeys, func(a, b []byte) bool { return cmp(a, b) == 0 })
1286 :
1287 : // Create a list of policies.
1288 1 : policies := make([]SpanPolicy, len(uniqueKeys)-1)
1289 1 : for _, p := range inputPolicies {
1290 1 : idx, _ := slices.BinarySearchFunc(uniqueKeys, p.KeyRange.Start, cmp)
1291 1 : policies[idx] = p.Policy
1292 1 : }
1293 :
1294 1 : return func(startKey []byte) (_ SpanPolicy, endKey []byte, _ error) {
1295 1 : // Find the policy that applies to the start key.
1296 1 : idx, eq := slices.BinarySearchFunc(uniqueKeys, startKey, cmp)
1297 1 : switch idx {
1298 1 : case len(uniqueKeys):
1299 1 : // The start key is after the last policy.
1300 1 : return SpanPolicy{}, nil, nil
1301 1 : case len(uniqueKeys) - 1:
1302 1 : if eq {
1303 1 : // The start key is exactly the start of the last policy.
1304 1 : return SpanPolicy{}, nil, nil
1305 1 : }
1306 1 : case 0:
1307 1 : if !eq {
1308 1 : // The start key is before the first policy.
1309 1 : return SpanPolicy{}, uniqueKeys[0], nil
1310 1 : }
1311 : }
1312 1 : if eq {
1313 1 : // The start key is exactly the start of this policy.
1314 1 : return policies[idx], uniqueKeys[idx+1], nil
1315 1 : }
1316 : // The start key is between two policies.
1317 1 : return policies[idx-1], uniqueKeys[idx], nil
1318 : }
1319 : }
1320 :
1321 : // WALFailoverOptions configures the WAL failover mechanics to use during
1322 : // transient write unavailability on the primary WAL volume.
1323 : type WALFailoverOptions struct {
1324 : // Secondary indicates the secondary directory and VFS to use in the event a
1325 : // write to the primary WAL stalls.
1326 : Secondary wal.Dir
1327 : // FailoverOptions provides configuration of the thresholds and intervals
1328 : // involved in WAL failover. If any of its fields are left unspecified,
1329 : // reasonable defaults will be used.
1330 : wal.FailoverOptions
1331 : }
1332 :
1333 : // ReadaheadConfig controls the use of read-ahead.
1334 : type ReadaheadConfig = objstorageprovider.ReadaheadConfig
1335 :
1336 : // JemallocSizeClasses exports sstable.JemallocSizeClasses.
1337 : var JemallocSizeClasses = sstable.JemallocSizeClasses
1338 :
1339 : // DebugCheckLevels calls CheckLevels on the provided database.
1340 : // It may be set in the DebugCheck field of Options to check
1341 : // level invariants whenever a new version is installed.
1342 1 : func DebugCheckLevels(db *DB) error {
1343 1 : return db.CheckLevels(nil)
1344 1 : }
1345 :
1346 : // DBCompressionSettings contains compression settings for the entire store. It
1347 : // defines compression profiles for each LSM level.
1348 : type DBCompressionSettings struct {
1349 : Name string
1350 : Levels [manifest.NumLevels]*block.CompressionProfile
1351 : }
1352 :
1353 : // Predefined compression settings.
1354 : var (
1355 : DBCompressionNone = UniformDBCompressionSettings(block.NoCompression)
1356 : DBCompressionFastest = UniformDBCompressionSettings(block.FastestCompression)
1357 1 : DBCompressionBalanced = func() DBCompressionSettings {
1358 1 : cs := DBCompressionSettings{Name: "Balanced"}
1359 1 : for i := 0; i < manifest.NumLevels-2; i++ {
1360 1 : cs.Levels[i] = block.FastestCompression
1361 1 : }
1362 1 : cs.Levels[manifest.NumLevels-2] = block.FastCompression // Zstd1 for value blocks.
1363 1 : cs.Levels[manifest.NumLevels-1] = block.BalancedCompression // Zstd1 for data and value blocks.
1364 1 : return cs
1365 : }()
1366 1 : DBCompressionGood = func() DBCompressionSettings {
1367 1 : cs := DBCompressionSettings{Name: "Good"}
1368 1 : for i := 0; i < manifest.NumLevels-2; i++ {
1369 1 : cs.Levels[i] = block.FastestCompression
1370 1 : }
1371 1 : cs.Levels[manifest.NumLevels-2] = block.BalancedCompression // Zstd1 for data and value blocks.
1372 1 : cs.Levels[manifest.NumLevels-1] = block.GoodCompression // Zstd3 for data and value blocks.
1373 1 : return cs
1374 : }()
1375 : )
1376 :
1377 : // UniformDBCompressionSettings returns a DBCompressionSettings which uses the
1378 : // same compression profile on all LSM levels.
1379 1 : func UniformDBCompressionSettings(profile *block.CompressionProfile) DBCompressionSettings {
1380 1 : cs := DBCompressionSettings{Name: profile.Name}
1381 1 : for i := range cs.Levels {
1382 1 : cs.Levels[i] = profile
1383 1 : }
1384 1 : return cs
1385 : }
1386 :
1387 : // ApplyCompressionSettings sets the Compression field in each LevelOptions to
1388 : // call the given function and return the compression profile for that level.
1389 1 : func (o *Options) ApplyCompressionSettings(csFn func() DBCompressionSettings) {
1390 1 : for i := range o.Levels {
1391 1 : levelIdx := i
1392 1 : o.Levels[i].Compression = func() *block.CompressionProfile {
1393 1 : return csFn().Levels[levelIdx]
1394 1 : }
1395 : }
1396 : }
1397 :
1398 : // EnsureDefaults ensures that the default values for all options are set if a
1399 : // valid value was not already specified.
1400 1 : func (o *Options) EnsureDefaults() {
1401 1 : if o.Cache == nil && o.CacheSize == 0 {
1402 1 : o.CacheSize = cacheDefaultSize
1403 1 : }
1404 1 : o.Comparer = o.Comparer.EnsureDefaults()
1405 1 :
1406 1 : if o.BytesPerSync <= 0 {
1407 1 : o.BytesPerSync = 512 << 10 // 512 KB
1408 1 : }
1409 1 : if o.Cleaner == nil {
1410 1 : o.Cleaner = DeleteCleaner{}
1411 1 : }
1412 :
1413 1 : if o.FreeSpaceThresholdBytes == 0 {
1414 1 : o.FreeSpaceThresholdBytes = 16 << 30 // 16 GB
1415 1 : }
1416 :
1417 1 : if o.FreeSpaceTimeframe == 0 {
1418 1 : o.FreeSpaceTimeframe = 10 * time.Second
1419 1 : }
1420 :
1421 1 : if o.ObsoleteBytesMaxRatio == 0 {
1422 1 : o.ObsoleteBytesMaxRatio = 0.20
1423 1 : }
1424 :
1425 1 : if o.ObsoleteBytesTimeframe == 0 {
1426 1 : o.ObsoleteBytesTimeframe = 300 * time.Second
1427 1 : }
1428 :
1429 1 : if o.Experimental.DisableIngestAsFlushable == nil {
1430 1 : o.Experimental.DisableIngestAsFlushable = func() bool { return false }
1431 : }
1432 1 : if o.Experimental.L0CompactionConcurrency <= 0 {
1433 1 : o.Experimental.L0CompactionConcurrency = 10
1434 1 : }
1435 1 : if o.Experimental.CompactionDebtConcurrency <= 0 {
1436 1 : o.Experimental.CompactionDebtConcurrency = 1 << 30 // 1 GB
1437 1 : }
1438 1 : if o.Experimental.CompactionGarbageFractionForMaxConcurrency == nil {
1439 1 : // When 40% of the DB is garbage, the compaction concurrency is at the
1440 1 : // maximum permitted.
1441 1 : o.Experimental.CompactionGarbageFractionForMaxConcurrency = func() float64 { return 0.4 }
1442 : }
1443 1 : if o.Experimental.ValueSeparationPolicy == nil {
1444 1 : o.Experimental.ValueSeparationPolicy = func() ValueSeparationPolicy {
1445 1 : return ValueSeparationPolicy{Enabled: false}
1446 1 : }
1447 : }
1448 1 : if o.KeySchema == "" && len(o.KeySchemas) == 0 {
1449 1 : ks := colblk.DefaultKeySchema(o.Comparer, 16 /* bundleSize */)
1450 1 : o.KeySchema = ks.Name
1451 1 : o.KeySchemas = sstable.MakeKeySchemas(&ks)
1452 1 : }
1453 1 : if o.L0CompactionThreshold <= 0 {
1454 1 : o.L0CompactionThreshold = 4
1455 1 : }
1456 1 : if o.L0CompactionFileThreshold <= 0 {
1457 1 : // Some justification for the default of 500:
1458 1 : // Why not smaller?:
1459 1 : // - The default target file size for L0 is 2MB, so 500 files is <= 1GB
1460 1 : // of data. At observed compaction speeds of > 20MB/s, L0 can be
1461 1 : // cleared of all files in < 1min, so this backlog is not huge.
1462 1 : // - 500 files is low overhead for instantiating L0 sublevels from
1463 1 : // scratch.
1464 1 : // - Lower values were observed to cause excessive and inefficient
1465 1 : // compactions out of L0 in a TPCC import benchmark.
1466 1 : // Why not larger?:
1467 1 : // - More than 1min to compact everything out of L0.
1468 1 : // - CockroachDB's admission control system uses a threshold of 1000
1469 1 : // files to start throttling writes to Pebble. Using 500 here gives
1470 1 : // us headroom between when Pebble should start compacting L0 and
1471 1 : // when the admission control threshold is reached.
1472 1 : //
1473 1 : // We can revisit this default in the future based on better
1474 1 : // experimental understanding.
1475 1 : //
1476 1 : // TODO(jackson): Experiment with slightly lower thresholds [or higher
1477 1 : // admission control thresholds] to see whether a higher L0 score at the
1478 1 : // threshold (currently 2.0) is necessary for some workloads to avoid
1479 1 : // starving L0 in favor of lower-level compactions.
1480 1 : o.L0CompactionFileThreshold = 500
1481 1 : }
1482 1 : if o.L0StopWritesThreshold <= 0 {
1483 1 : o.L0StopWritesThreshold = 12
1484 1 : }
1485 1 : if o.LBaseMaxBytes <= 0 {
1486 1 : o.LBaseMaxBytes = 64 << 20 // 64 MB
1487 1 : }
1488 1 : if o.TargetFileSizes[0] <= 0 {
1489 1 : o.TargetFileSizes[0] = 2 << 20 // 2 MB
1490 1 : }
1491 1 : for i := 1; i < len(o.TargetFileSizes); i++ {
1492 1 : if o.TargetFileSizes[i] <= 0 {
1493 1 : o.TargetFileSizes[i] = o.TargetFileSizes[i-1] * 2
1494 1 : }
1495 : }
1496 1 : o.Levels[0].EnsureL0Defaults()
1497 1 : for i := 1; i < len(o.Levels); i++ {
1498 1 : o.Levels[i].EnsureL1PlusDefaults(&o.Levels[i-1])
1499 1 : }
1500 1 : if o.Logger == nil {
1501 1 : o.Logger = DefaultLogger
1502 1 : }
1503 1 : if o.EventListener == nil {
1504 1 : o.EventListener = &EventListener{}
1505 1 : }
1506 1 : o.EventListener.EnsureDefaults(o.Logger)
1507 1 : if o.MaxManifestFileSize == 0 {
1508 1 : o.MaxManifestFileSize = 128 << 20 // 128 MB
1509 1 : }
1510 1 : if o.MaxOpenFiles == 0 {
1511 1 : o.MaxOpenFiles = 1000
1512 1 : }
1513 1 : if o.MemTableSize <= 0 {
1514 1 : o.MemTableSize = 4 << 20 // 4 MB
1515 1 : }
1516 1 : if o.MemTableStopWritesThreshold <= 0 {
1517 1 : o.MemTableStopWritesThreshold = 2
1518 1 : }
1519 1 : if o.Merger == nil {
1520 1 : o.Merger = DefaultMerger
1521 1 : }
1522 1 : if o.CompactionConcurrencyRange == nil {
1523 1 : o.CompactionConcurrencyRange = func() (int, int) { return 1, 1 }
1524 : }
1525 1 : if o.MaxConcurrentDownloads == nil {
1526 1 : o.MaxConcurrentDownloads = func() int { return 1 }
1527 : }
1528 1 : if o.NumPrevManifest <= 0 {
1529 1 : o.NumPrevManifest = 1
1530 1 : }
1531 :
1532 1 : if o.FormatMajorVersion == FormatDefault {
1533 1 : o.FormatMajorVersion = FormatMinSupported
1534 1 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone {
1535 1 : o.FormatMajorVersion = FormatMinForSharedObjects
1536 1 : }
1537 : }
1538 :
1539 1 : if o.FS == nil {
1540 1 : o.WithFSDefaults()
1541 1 : }
1542 1 : if o.FlushSplitBytes <= 0 {
1543 1 : o.FlushSplitBytes = 2 * o.TargetFileSizes[0]
1544 1 : }
1545 1 : if o.WALFailover != nil {
1546 1 : o.WALFailover.FailoverOptions.EnsureDefaults()
1547 1 : }
1548 1 : if o.Experimental.LevelMultiplier <= 0 {
1549 1 : o.Experimental.LevelMultiplier = defaultLevelMultiplier
1550 1 : }
1551 1 : if o.Experimental.ReadCompactionRate == 0 {
1552 1 : o.Experimental.ReadCompactionRate = 16000
1553 1 : }
1554 1 : if o.Experimental.ReadSamplingMultiplier == 0 {
1555 1 : o.Experimental.ReadSamplingMultiplier = 1 << 4
1556 1 : }
1557 1 : if o.Experimental.NumDeletionsThreshold == 0 {
1558 1 : o.Experimental.NumDeletionsThreshold = sstable.DefaultNumDeletionsThreshold
1559 1 : }
1560 1 : if o.Experimental.DeletionSizeRatioThreshold == 0 {
1561 1 : o.Experimental.DeletionSizeRatioThreshold = sstable.DefaultDeletionSizeRatioThreshold
1562 1 : }
1563 1 : if o.Experimental.EnableColumnarBlocks == nil {
1564 1 : o.Experimental.EnableColumnarBlocks = func() bool { return true }
1565 : }
1566 1 : if o.Experimental.TombstoneDenseCompactionThreshold == 0 {
1567 1 : o.Experimental.TombstoneDenseCompactionThreshold = 0.10
1568 1 : }
1569 1 : if o.Experimental.FileCacheShards <= 0 {
1570 1 : o.Experimental.FileCacheShards = runtime.GOMAXPROCS(0)
1571 1 : }
1572 1 : if o.Experimental.MultiLevelCompactionHeuristic == nil {
1573 1 : o.Experimental.MultiLevelCompactionHeuristic = OptionWriteAmpHeuristic
1574 1 : }
1575 1 : if o.Experimental.SpanPolicyFunc == nil {
1576 1 : o.Experimental.SpanPolicyFunc = func(startKey []byte) (SpanPolicy, []byte, error) { return SpanPolicy{}, nil, nil }
1577 : }
1578 : // TODO(jackson): Enable value separation by default once we have confidence
1579 : // in a default policy.
1580 :
1581 1 : o.initMaps()
1582 : }
1583 :
1584 : // TargetFileSize computes the target file size for the given output level.
1585 1 : func (o *Options) TargetFileSize(outputLevel int, baseLevel int) int64 {
1586 1 : if outputLevel == 0 {
1587 1 : return o.TargetFileSizes[0]
1588 1 : }
1589 1 : if baseLevel > outputLevel {
1590 0 : panic(fmt.Sprintf("invalid base level %d (output level %d)", baseLevel, outputLevel))
1591 : }
1592 1 : return o.TargetFileSizes[outputLevel-baseLevel+1]
1593 : }
1594 :
1595 : // DefaultOptions returns a new Options object with the default values set.
1596 1 : func DefaultOptions() *Options {
1597 1 : o := &Options{}
1598 1 : o.EnsureDefaults()
1599 1 : return o
1600 1 : }
1601 :
1602 : // WithFSDefaults configures the Options to wrap the configured filesystem with
1603 : // the default virtual file system middleware, like disk-health checking.
1604 1 : func (o *Options) WithFSDefaults() {
1605 1 : if o.FS == nil {
1606 1 : o.FS = vfs.Default
1607 1 : }
1608 1 : o.FS, o.private.fsCloser = vfs.WithDiskHealthChecks(o.FS, 5*time.Second, nil,
1609 1 : func(info vfs.DiskSlowInfo) {
1610 0 : o.EventListener.DiskSlow(info)
1611 0 : })
1612 : }
1613 :
1614 : // AddEventListener adds the provided event listener to the Options, in addition
1615 : // to any existing event listener.
1616 1 : func (o *Options) AddEventListener(l EventListener) {
1617 1 : if o.EventListener != nil {
1618 1 : l = TeeEventListener(l, *o.EventListener)
1619 1 : }
1620 1 : o.EventListener = &l
1621 : }
1622 :
1623 : // initMaps initializes the Comparers, Filters, and Mergers maps.
1624 1 : func (o *Options) initMaps() {
1625 1 : for i := range o.Levels {
1626 1 : l := &o.Levels[i]
1627 1 : if l.FilterPolicy != NoFilterPolicy {
1628 1 : if o.Filters == nil {
1629 1 : o.Filters = make(map[string]FilterPolicy)
1630 1 : }
1631 1 : name := l.FilterPolicy.Name()
1632 1 : if _, ok := o.Filters[name]; !ok {
1633 1 : o.Filters[name] = l.FilterPolicy
1634 1 : }
1635 : }
1636 : }
1637 : }
1638 :
1639 : // Clone creates a shallow-copy of the supplied options.
1640 1 : func (o *Options) Clone() *Options {
1641 1 : if o == nil {
1642 1 : return &Options{}
1643 1 : }
1644 1 : n := *o
1645 1 : if o.WALFailover != nil {
1646 1 : c := *o.WALFailover
1647 1 : n.WALFailover = &c
1648 1 : }
1649 1 : return &n
1650 : }
1651 :
1652 1 : func (o *Options) String() string {
1653 1 : var buf bytes.Buffer
1654 1 :
1655 1 : cacheSize := o.CacheSize
1656 1 : if o.Cache != nil {
1657 1 : cacheSize = o.Cache.MaxSize()
1658 1 : }
1659 :
1660 1 : fmt.Fprintf(&buf, "[Version]\n")
1661 1 : fmt.Fprintf(&buf, " pebble_version=0.1\n")
1662 1 : fmt.Fprintf(&buf, "\n")
1663 1 : fmt.Fprintf(&buf, "[Options]\n")
1664 1 : fmt.Fprintf(&buf, " bytes_per_sync=%d\n", o.BytesPerSync)
1665 1 : fmt.Fprintf(&buf, " cache_size=%d\n", cacheSize)
1666 1 : fmt.Fprintf(&buf, " cleaner=%s\n", o.Cleaner)
1667 1 : fmt.Fprintf(&buf, " compaction_debt_concurrency=%d\n", o.Experimental.CompactionDebtConcurrency)
1668 1 : fmt.Fprintf(&buf, " compaction_garbage_fraction_for_max_concurrency=%.2f\n",
1669 1 : o.Experimental.CompactionGarbageFractionForMaxConcurrency())
1670 1 : fmt.Fprintf(&buf, " comparer=%s\n", o.Comparer.Name)
1671 1 : fmt.Fprintf(&buf, " disable_wal=%t\n", o.DisableWAL)
1672 1 : if o.Experimental.DisableIngestAsFlushable != nil && o.Experimental.DisableIngestAsFlushable() {
1673 1 : fmt.Fprintf(&buf, " disable_ingest_as_flushable=%t\n", true)
1674 1 : }
1675 1 : if o.Experimental.EnableColumnarBlocks != nil && o.Experimental.EnableColumnarBlocks() {
1676 1 : fmt.Fprintf(&buf, " enable_columnar_blocks=%t\n", true)
1677 1 : }
1678 1 : fmt.Fprintf(&buf, " flush_delay_delete_range=%s\n", o.FlushDelayDeleteRange)
1679 1 : fmt.Fprintf(&buf, " flush_delay_range_key=%s\n", o.FlushDelayRangeKey)
1680 1 : fmt.Fprintf(&buf, " flush_split_bytes=%d\n", o.FlushSplitBytes)
1681 1 : fmt.Fprintf(&buf, " format_major_version=%d\n", o.FormatMajorVersion)
1682 1 : fmt.Fprintf(&buf, " key_schema=%s\n", o.KeySchema)
1683 1 : fmt.Fprintf(&buf, " l0_compaction_concurrency=%d\n", o.Experimental.L0CompactionConcurrency)
1684 1 : fmt.Fprintf(&buf, " l0_compaction_file_threshold=%d\n", o.L0CompactionFileThreshold)
1685 1 : fmt.Fprintf(&buf, " l0_compaction_threshold=%d\n", o.L0CompactionThreshold)
1686 1 : fmt.Fprintf(&buf, " l0_stop_writes_threshold=%d\n", o.L0StopWritesThreshold)
1687 1 : fmt.Fprintf(&buf, " lbase_max_bytes=%d\n", o.LBaseMaxBytes)
1688 1 : if o.Experimental.LevelMultiplier != defaultLevelMultiplier {
1689 1 : fmt.Fprintf(&buf, " level_multiplier=%d\n", o.Experimental.LevelMultiplier)
1690 1 : }
1691 1 : lower, upper := o.CompactionConcurrencyRange()
1692 1 : fmt.Fprintf(&buf, " concurrent_compactions=%d\n", lower)
1693 1 : fmt.Fprintf(&buf, " max_concurrent_compactions=%d\n", upper)
1694 1 : fmt.Fprintf(&buf, " max_concurrent_downloads=%d\n", o.MaxConcurrentDownloads())
1695 1 : fmt.Fprintf(&buf, " max_manifest_file_size=%d\n", o.MaxManifestFileSize)
1696 1 : fmt.Fprintf(&buf, " max_open_files=%d\n", o.MaxOpenFiles)
1697 1 : fmt.Fprintf(&buf, " mem_table_size=%d\n", o.MemTableSize)
1698 1 : fmt.Fprintf(&buf, " mem_table_stop_writes_threshold=%d\n", o.MemTableStopWritesThreshold)
1699 1 : fmt.Fprintf(&buf, " min_deletion_rate=%d\n", o.TargetByteDeletionRate)
1700 1 : fmt.Fprintf(&buf, " free_space_threshold_bytes=%d\n", o.FreeSpaceThresholdBytes)
1701 1 : fmt.Fprintf(&buf, " free_space_timeframe=%s\n", o.FreeSpaceTimeframe.String())
1702 1 : fmt.Fprintf(&buf, " obsolete_bytes_max_ratio=%f\n", o.ObsoleteBytesMaxRatio)
1703 1 : fmt.Fprintf(&buf, " obsolete_bytes_timeframe=%s\n", o.ObsoleteBytesTimeframe.String())
1704 1 : fmt.Fprintf(&buf, " merger=%s\n", o.Merger.Name)
1705 1 : if o.Experimental.MultiLevelCompactionHeuristic != nil {
1706 1 : fmt.Fprintf(&buf, " multilevel_compaction_heuristic=%s\n", o.Experimental.MultiLevelCompactionHeuristic().String())
1707 1 : }
1708 1 : fmt.Fprintf(&buf, " read_compaction_rate=%d\n", o.Experimental.ReadCompactionRate)
1709 1 : fmt.Fprintf(&buf, " read_sampling_multiplier=%d\n", o.Experimental.ReadSamplingMultiplier)
1710 1 : fmt.Fprintf(&buf, " num_deletions_threshold=%d\n", o.Experimental.NumDeletionsThreshold)
1711 1 : fmt.Fprintf(&buf, " deletion_size_ratio_threshold=%f\n", o.Experimental.DeletionSizeRatioThreshold)
1712 1 : fmt.Fprintf(&buf, " tombstone_dense_compaction_threshold=%f\n", o.Experimental.TombstoneDenseCompactionThreshold)
1713 1 : // We no longer care about strict_wal_tail, but set it to true in case an
1714 1 : // older version reads the options.
1715 1 : fmt.Fprintf(&buf, " strict_wal_tail=%t\n", true)
1716 1 : fmt.Fprintf(&buf, " table_cache_shards=%d\n", o.Experimental.FileCacheShards)
1717 1 : fmt.Fprintf(&buf, " validate_on_ingest=%t\n", o.Experimental.ValidateOnIngest)
1718 1 : fmt.Fprintf(&buf, " wal_dir=%s\n", o.WALDir)
1719 1 : fmt.Fprintf(&buf, " wal_bytes_per_sync=%d\n", o.WALBytesPerSync)
1720 1 : fmt.Fprintf(&buf, " secondary_cache_size_bytes=%d\n", o.Experimental.SecondaryCacheSizeBytes)
1721 1 : fmt.Fprintf(&buf, " create_on_shared=%d\n", o.Experimental.CreateOnShared)
1722 1 :
1723 1 : // Private options.
1724 1 : //
1725 1 : // These options are only encoded if true, because we do not want them to
1726 1 : // appear in production serialized Options files, since they're testing-only
1727 1 : // options. They're only serialized when true, which still ensures that the
1728 1 : // metamorphic tests may propagate them to subprocesses.
1729 1 : if o.private.disableDeleteOnlyCompactions {
1730 1 : fmt.Fprintln(&buf, " disable_delete_only_compactions=true")
1731 1 : }
1732 1 : if o.private.disableElisionOnlyCompactions {
1733 1 : fmt.Fprintln(&buf, " disable_elision_only_compactions=true")
1734 1 : }
1735 1 : if o.private.disableLazyCombinedIteration {
1736 1 : fmt.Fprintln(&buf, " disable_lazy_combined_iteration=true")
1737 1 : }
1738 :
1739 1 : if o.Experimental.ValueSeparationPolicy != nil {
1740 1 : policy := o.Experimental.ValueSeparationPolicy()
1741 1 : if policy.Enabled {
1742 1 : fmt.Fprintln(&buf)
1743 1 : fmt.Fprintln(&buf, "[Value Separation]")
1744 1 : fmt.Fprintf(&buf, " enabled=%t\n", policy.Enabled)
1745 1 : fmt.Fprintf(&buf, " minimum_size=%d\n", policy.MinimumSize)
1746 1 : fmt.Fprintf(&buf, " max_blob_reference_depth=%d\n", policy.MaxBlobReferenceDepth)
1747 1 : fmt.Fprintf(&buf, " rewrite_minimum_age=%s\n", policy.RewriteMinimumAge)
1748 1 : fmt.Fprintf(&buf, " target_garbage_ratio=%.2f\n", policy.TargetGarbageRatio)
1749 1 : }
1750 : }
1751 :
1752 1 : if o.WALFailover != nil {
1753 1 : unhealthyThreshold, _ := o.WALFailover.FailoverOptions.UnhealthyOperationLatencyThreshold()
1754 1 : fmt.Fprintf(&buf, "\n")
1755 1 : fmt.Fprintf(&buf, "[WAL Failover]\n")
1756 1 : fmt.Fprintf(&buf, " secondary_dir=%s\n", o.WALFailover.Secondary.Dirname)
1757 1 : fmt.Fprintf(&buf, " primary_dir_probe_interval=%s\n", o.WALFailover.FailoverOptions.PrimaryDirProbeInterval)
1758 1 : fmt.Fprintf(&buf, " healthy_probe_latency_threshold=%s\n", o.WALFailover.FailoverOptions.HealthyProbeLatencyThreshold)
1759 1 : fmt.Fprintf(&buf, " healthy_interval=%s\n", o.WALFailover.FailoverOptions.HealthyInterval)
1760 1 : fmt.Fprintf(&buf, " unhealthy_sampling_interval=%s\n", o.WALFailover.FailoverOptions.UnhealthySamplingInterval)
1761 1 : fmt.Fprintf(&buf, " unhealthy_operation_latency_threshold=%s\n", unhealthyThreshold)
1762 1 : fmt.Fprintf(&buf, " elevated_write_stall_threshold_lag=%s\n", o.WALFailover.FailoverOptions.ElevatedWriteStallThresholdLag)
1763 1 : }
1764 :
1765 1 : for i := range o.Levels {
1766 1 : l := &o.Levels[i]
1767 1 : fmt.Fprintf(&buf, "\n")
1768 1 : fmt.Fprintf(&buf, "[Level \"%d\"]\n", i)
1769 1 : fmt.Fprintf(&buf, " block_restart_interval=%d\n", l.BlockRestartInterval)
1770 1 : fmt.Fprintf(&buf, " block_size=%d\n", l.BlockSize)
1771 1 : fmt.Fprintf(&buf, " block_size_threshold=%d\n", l.BlockSizeThreshold)
1772 1 : fmt.Fprintf(&buf, " compression=%s\n", l.Compression().Name)
1773 1 : fmt.Fprintf(&buf, " filter_policy=%s\n", l.FilterPolicy.Name())
1774 1 : fmt.Fprintf(&buf, " filter_type=%s\n", l.FilterType)
1775 1 : fmt.Fprintf(&buf, " index_block_size=%d\n", l.IndexBlockSize)
1776 1 : fmt.Fprintf(&buf, " target_file_size=%d\n", o.TargetFileSizes[i])
1777 1 : }
1778 :
1779 1 : return buf.String()
1780 : }
1781 :
1782 : type parseOptionsFuncs struct {
1783 : visitNewSection func(i, j int, section string) error
1784 : visitKeyValue func(i, j int, section, key, value string) error
1785 : visitCommentOrWhitespace func(i, j int, whitespace string) error
1786 : }
1787 :
1788 : // parseOptions takes options serialized by Options.String() and parses them
1789 : // into keys and values. It calls fns.visitNewSection for the beginning of each
1790 : // new section, fns.visitKeyValue for each key-value pair, and
1791 : // visitCommentOrWhitespace for comments and whitespace between key-value pairs.
1792 1 : func parseOptions(s string, fns parseOptionsFuncs) error {
1793 1 : var section, mappedSection string
1794 1 : i := 0
1795 1 : for i < len(s) {
1796 1 : rem := s[i:]
1797 1 : j := strings.IndexByte(rem, '\n')
1798 1 : if j < 0 {
1799 1 : j = len(rem)
1800 1 : } else {
1801 1 : j += 1 // Include the newline.
1802 1 : }
1803 1 : line := strings.TrimSpace(s[i : i+j])
1804 1 : startOff, endOff := i, i+j
1805 1 : i += j
1806 1 :
1807 1 : if len(line) == 0 || line[0] == ';' || line[0] == '#' {
1808 1 : // Skip blank lines and comments.
1809 1 : if fns.visitCommentOrWhitespace != nil {
1810 1 : if err := fns.visitCommentOrWhitespace(startOff, endOff, line); err != nil {
1811 0 : return err
1812 0 : }
1813 : }
1814 1 : continue
1815 : }
1816 1 : n := len(line)
1817 1 : if line[0] == '[' && line[n-1] == ']' {
1818 1 : // Parse section.
1819 1 : section = line[1 : n-1]
1820 1 : // RocksDB uses a similar (INI-style) syntax for the OPTIONS file, but
1821 1 : // different section names and keys. The "CFOptions ..." paths are the
1822 1 : // RocksDB versions which we map to the Pebble paths.
1823 1 : mappedSection = section
1824 1 : if section == `CFOptions "default"` {
1825 1 : mappedSection = "Options"
1826 1 : }
1827 1 : if fns.visitNewSection != nil {
1828 1 : if err := fns.visitNewSection(startOff, endOff, mappedSection); err != nil {
1829 0 : return err
1830 0 : }
1831 : }
1832 1 : continue
1833 : }
1834 :
1835 1 : pos := strings.Index(line, "=")
1836 1 : if pos < 0 {
1837 1 : const maxLen = 50
1838 1 : if len(line) > maxLen {
1839 0 : line = line[:maxLen-3] + "..."
1840 0 : }
1841 1 : return base.CorruptionErrorf("invalid key=value syntax: %q", errors.Safe(line))
1842 : }
1843 :
1844 1 : key := strings.TrimSpace(line[:pos])
1845 1 : value := strings.TrimSpace(line[pos+1:])
1846 1 :
1847 1 : if section == `CFOptions "default"` {
1848 1 : switch key {
1849 1 : case "comparator":
1850 1 : key = "comparer"
1851 1 : case "merge_operator":
1852 1 : key = "merger"
1853 : }
1854 : }
1855 1 : if fns.visitKeyValue != nil {
1856 1 : if err := fns.visitKeyValue(startOff, endOff, mappedSection, key, value); err != nil {
1857 1 : return err
1858 1 : }
1859 : }
1860 : }
1861 1 : return nil
1862 : }
1863 :
1864 : // ParseHooks contains callbacks to create options fields which can have
1865 : // user-defined implementations.
1866 : type ParseHooks struct {
1867 : NewCleaner func(name string) (Cleaner, error)
1868 : NewComparer func(name string) (*Comparer, error)
1869 : NewFilterPolicy func(name string) (FilterPolicy, error)
1870 : NewKeySchema func(name string) (KeySchema, error)
1871 : NewMerger func(name string) (*Merger, error)
1872 : SkipUnknown func(name, value string) bool
1873 : }
1874 :
1875 : // Parse parses the options from the specified string. Note that certain
1876 : // options cannot be parsed into populated fields. For example, comparer and
1877 : // merger.
1878 1 : func (o *Options) Parse(s string, hooks *ParseHooks) error {
1879 1 : var valSepPolicy ValueSeparationPolicy
1880 1 : var concurrencyLimit struct {
1881 1 : lower int
1882 1 : lowerSet bool
1883 1 : upper int
1884 1 : upperSet bool
1885 1 : }
1886 1 :
1887 1 : visitKeyValue := func(i, j int, section, key, value string) error {
1888 1 : // WARNING: DO NOT remove entries from the switches below because doing so
1889 1 : // causes a key previously written to the OPTIONS file to be considered unknown,
1890 1 : // a backwards incompatible change. Instead, leave in support for parsing the
1891 1 : // key but simply don't parse the value.
1892 1 :
1893 1 : parseComparer := func(name string) (*Comparer, error) {
1894 1 : switch name {
1895 1 : case DefaultComparer.Name:
1896 1 : return DefaultComparer, nil
1897 1 : case testkeys.Comparer.Name:
1898 1 : return testkeys.Comparer, nil
1899 1 : default:
1900 1 : if hooks != nil && hooks.NewComparer != nil {
1901 1 : return hooks.NewComparer(name)
1902 1 : }
1903 1 : return nil, nil
1904 : }
1905 : }
1906 :
1907 1 : switch {
1908 1 : case section == "Version":
1909 1 : switch key {
1910 1 : case "pebble_version":
1911 0 : default:
1912 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1913 0 : return nil
1914 0 : }
1915 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1916 0 : errors.Safe(section), errors.Safe(key))
1917 : }
1918 1 : return nil
1919 :
1920 1 : case section == "Options":
1921 1 : var err error
1922 1 : switch key {
1923 1 : case "bytes_per_sync":
1924 1 : o.BytesPerSync, err = strconv.Atoi(value)
1925 1 : case "cache_size":
1926 1 : o.CacheSize, err = strconv.ParseInt(value, 10, 64)
1927 1 : case "cleaner":
1928 1 : switch value {
1929 1 : case "archive":
1930 1 : o.Cleaner = ArchiveCleaner{}
1931 1 : case "delete":
1932 1 : o.Cleaner = DeleteCleaner{}
1933 0 : default:
1934 0 : if hooks != nil && hooks.NewCleaner != nil {
1935 0 : o.Cleaner, err = hooks.NewCleaner(value)
1936 0 : }
1937 : }
1938 1 : case "comparer":
1939 1 : var comparer *Comparer
1940 1 : comparer, err = parseComparer(value)
1941 1 : if comparer != nil {
1942 1 : o.Comparer = comparer
1943 1 : }
1944 1 : case "compaction_debt_concurrency":
1945 1 : o.Experimental.CompactionDebtConcurrency, err = strconv.ParseUint(value, 10, 64)
1946 1 : case "compaction_garbage_fraction_for_max_concurrency":
1947 1 : var frac float64
1948 1 : frac, err = strconv.ParseFloat(value, 64)
1949 1 : if err == nil {
1950 1 : o.Experimental.CompactionGarbageFractionForMaxConcurrency =
1951 1 : func() float64 { return frac }
1952 : }
1953 0 : case "delete_range_flush_delay":
1954 0 : // NB: This is a deprecated serialization of the
1955 0 : // `flush_delay_delete_range`.
1956 0 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1957 1 : case "disable_delete_only_compactions":
1958 1 : o.private.disableDeleteOnlyCompactions, err = strconv.ParseBool(value)
1959 1 : case "disable_elision_only_compactions":
1960 1 : o.private.disableElisionOnlyCompactions, err = strconv.ParseBool(value)
1961 1 : case "disable_ingest_as_flushable":
1962 1 : var v bool
1963 1 : v, err = strconv.ParseBool(value)
1964 1 : if err == nil {
1965 1 : o.Experimental.DisableIngestAsFlushable = func() bool { return v }
1966 : }
1967 1 : case "disable_lazy_combined_iteration":
1968 1 : o.private.disableLazyCombinedIteration, err = strconv.ParseBool(value)
1969 1 : case "disable_wal":
1970 1 : o.DisableWAL, err = strconv.ParseBool(value)
1971 1 : case "enable_columnar_blocks":
1972 1 : var v bool
1973 1 : if v, err = strconv.ParseBool(value); err == nil {
1974 1 : o.Experimental.EnableColumnarBlocks = func() bool { return v }
1975 : }
1976 1 : case "flush_delay_delete_range":
1977 1 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1978 1 : case "flush_delay_range_key":
1979 1 : o.FlushDelayRangeKey, err = time.ParseDuration(value)
1980 1 : case "flush_split_bytes":
1981 1 : o.FlushSplitBytes, err = strconv.ParseInt(value, 10, 64)
1982 1 : case "format_major_version":
1983 1 : // NB: The version written here may be stale. Open does
1984 1 : // not use the format major version encoded in the
1985 1 : // OPTIONS file other than to validate that the encoded
1986 1 : // version is valid right here.
1987 1 : var v uint64
1988 1 : v, err = strconv.ParseUint(value, 10, 64)
1989 1 : if vers := FormatMajorVersion(v); vers > internalFormatNewest || vers == FormatDefault {
1990 0 : err = errors.Newf("unsupported format major version %d", o.FormatMajorVersion)
1991 0 : }
1992 1 : if err == nil {
1993 1 : o.FormatMajorVersion = FormatMajorVersion(v)
1994 1 : }
1995 1 : case "key_schema":
1996 1 : o.KeySchema = value
1997 1 : if o.KeySchemas == nil {
1998 1 : o.KeySchemas = make(map[string]*KeySchema)
1999 1 : }
2000 1 : if _, ok := o.KeySchemas[o.KeySchema]; !ok {
2001 1 : if strings.HasPrefix(value, "DefaultKeySchema(") && strings.HasSuffix(value, ")") {
2002 1 : argsStr := strings.TrimSuffix(strings.TrimPrefix(value, "DefaultKeySchema("), ")")
2003 1 : args := strings.FieldsFunc(argsStr, func(r rune) bool {
2004 1 : return unicode.IsSpace(r) || r == ','
2005 1 : })
2006 1 : var comparer *base.Comparer
2007 1 : var bundleSize int
2008 1 : comparer, err = parseComparer(args[0])
2009 1 : if err == nil {
2010 1 : bundleSize, err = strconv.Atoi(args[1])
2011 1 : }
2012 1 : if err == nil {
2013 1 : schema := colblk.DefaultKeySchema(comparer, bundleSize)
2014 1 : o.KeySchema = schema.Name
2015 1 : o.KeySchemas[o.KeySchema] = &schema
2016 1 : }
2017 1 : } else if hooks != nil && hooks.NewKeySchema != nil {
2018 0 : var schema KeySchema
2019 0 : schema, err = hooks.NewKeySchema(value)
2020 0 : if err == nil {
2021 0 : o.KeySchemas[value] = &schema
2022 0 : }
2023 : }
2024 : }
2025 1 : case "l0_compaction_concurrency":
2026 1 : o.Experimental.L0CompactionConcurrency, err = strconv.Atoi(value)
2027 1 : case "l0_compaction_file_threshold":
2028 1 : o.L0CompactionFileThreshold, err = strconv.Atoi(value)
2029 1 : case "l0_compaction_threshold":
2030 1 : o.L0CompactionThreshold, err = strconv.Atoi(value)
2031 1 : case "l0_stop_writes_threshold":
2032 1 : o.L0StopWritesThreshold, err = strconv.Atoi(value)
2033 0 : case "l0_sublevel_compactions":
2034 : // Do nothing; option existed in older versions of pebble.
2035 1 : case "lbase_max_bytes":
2036 1 : o.LBaseMaxBytes, err = strconv.ParseInt(value, 10, 64)
2037 1 : case "level_multiplier":
2038 1 : o.Experimental.LevelMultiplier, err = strconv.Atoi(value)
2039 1 : case "concurrent_compactions":
2040 1 : concurrencyLimit.lowerSet = true
2041 1 : concurrencyLimit.lower, err = strconv.Atoi(value)
2042 1 : case "max_concurrent_compactions":
2043 1 : concurrencyLimit.upperSet = true
2044 1 : concurrencyLimit.upper, err = strconv.Atoi(value)
2045 1 : case "max_concurrent_downloads":
2046 1 : var concurrentDownloads int
2047 1 : concurrentDownloads, err = strconv.Atoi(value)
2048 1 : if concurrentDownloads <= 0 {
2049 0 : err = errors.New("max_concurrent_compactions cannot be <= 0")
2050 1 : } else {
2051 1 : o.MaxConcurrentDownloads = func() int { return concurrentDownloads }
2052 : }
2053 1 : case "max_manifest_file_size":
2054 1 : o.MaxManifestFileSize, err = strconv.ParseInt(value, 10, 64)
2055 1 : case "max_open_files":
2056 1 : o.MaxOpenFiles, err = strconv.Atoi(value)
2057 1 : case "mem_table_size":
2058 1 : o.MemTableSize, err = strconv.ParseUint(value, 10, 64)
2059 1 : case "mem_table_stop_writes_threshold":
2060 1 : o.MemTableStopWritesThreshold, err = strconv.Atoi(value)
2061 0 : case "min_compaction_rate":
2062 : // Do nothing; option existed in older versions of pebble, and
2063 : // may be meaningful again eventually.
2064 1 : case "min_deletion_rate":
2065 1 : o.TargetByteDeletionRate, err = strconv.Atoi(value)
2066 1 : case "free_space_threshold_bytes":
2067 1 : o.FreeSpaceThresholdBytes, err = strconv.ParseUint(value, 10, 64)
2068 1 : case "free_space_timeframe":
2069 1 : o.FreeSpaceTimeframe, err = time.ParseDuration(value)
2070 1 : case "obsolete_bytes_max_ratio":
2071 1 : o.ObsoleteBytesMaxRatio, err = strconv.ParseFloat(value, 64)
2072 1 : case "obsolete_bytes_timeframe":
2073 1 : o.ObsoleteBytesTimeframe, err = time.ParseDuration(value)
2074 0 : case "min_flush_rate":
2075 : // Do nothing; option existed in older versions of pebble, and
2076 : // may be meaningful again eventually.
2077 1 : case "multilevel_compaction_heuristic":
2078 1 : switch {
2079 1 : case value == "none":
2080 1 : o.Experimental.MultiLevelCompactionHeuristic = OptionNoMultiLevel
2081 1 : case strings.HasPrefix(value, "wamp"):
2082 1 : o.Experimental.MultiLevelCompactionHeuristic = OptionWriteAmpHeuristic
2083 1 : fields := strings.FieldsFunc(strings.TrimPrefix(value, "wamp"), func(r rune) bool {
2084 1 : return unicode.IsSpace(r) || r == ',' || r == '(' || r == ')'
2085 1 : })
2086 1 : if len(fields) != 2 {
2087 0 : err = errors.Newf("require 2 arguments")
2088 0 : }
2089 1 : var h WriteAmpHeuristic
2090 1 : if err == nil {
2091 1 : h.AddPropensity, err = strconv.ParseFloat(fields[0], 64)
2092 1 : }
2093 1 : if err == nil {
2094 1 : h.AllowL0, err = strconv.ParseBool(fields[1])
2095 1 : }
2096 :
2097 1 : if err == nil {
2098 1 : if h.AllowL0 || h.AddPropensity != 0 {
2099 1 : o.Experimental.MultiLevelCompactionHeuristic = func() MultiLevelHeuristic {
2100 1 : return &h
2101 1 : }
2102 : }
2103 0 : } else {
2104 0 : err = errors.Wrapf(err, "unexpected wamp heuristic arguments: %s", value)
2105 0 : }
2106 0 : default:
2107 0 : err = errors.Newf("unrecognized multilevel compaction heuristic: %s", value)
2108 : }
2109 0 : case "point_tombstone_weight":
2110 : // Do nothing; deprecated.
2111 1 : case "strict_wal_tail":
2112 1 : var strictWALTail bool
2113 1 : strictWALTail, err = strconv.ParseBool(value)
2114 1 : if err == nil && !strictWALTail {
2115 0 : err = errors.Newf("reading from versions with strict_wal_tail=false no longer supported")
2116 0 : }
2117 1 : case "merger":
2118 1 : switch value {
2119 0 : case "nullptr":
2120 0 : o.Merger = nil
2121 1 : case "pebble.concatenate":
2122 1 : o.Merger = DefaultMerger
2123 1 : default:
2124 1 : if hooks != nil && hooks.NewMerger != nil {
2125 1 : o.Merger, err = hooks.NewMerger(value)
2126 1 : }
2127 : }
2128 1 : case "read_compaction_rate":
2129 1 : o.Experimental.ReadCompactionRate, err = strconv.ParseInt(value, 10, 64)
2130 1 : case "read_sampling_multiplier":
2131 1 : o.Experimental.ReadSamplingMultiplier, err = strconv.ParseInt(value, 10, 64)
2132 1 : case "num_deletions_threshold":
2133 1 : o.Experimental.NumDeletionsThreshold, err = strconv.Atoi(value)
2134 1 : case "deletion_size_ratio_threshold":
2135 1 : val, parseErr := strconv.ParseFloat(value, 32)
2136 1 : o.Experimental.DeletionSizeRatioThreshold = float32(val)
2137 1 : err = parseErr
2138 1 : case "tombstone_dense_compaction_threshold":
2139 1 : o.Experimental.TombstoneDenseCompactionThreshold, err = strconv.ParseFloat(value, 64)
2140 1 : case "table_cache_shards":
2141 1 : o.Experimental.FileCacheShards, err = strconv.Atoi(value)
2142 0 : case "table_format":
2143 0 : switch value {
2144 0 : case "leveldb":
2145 0 : case "rocksdbv2":
2146 0 : default:
2147 0 : return errors.Errorf("pebble: unknown table format: %q", errors.Safe(value))
2148 : }
2149 1 : case "table_property_collectors":
2150 : // No longer implemented; ignore.
2151 1 : case "validate_on_ingest":
2152 1 : o.Experimental.ValidateOnIngest, err = strconv.ParseBool(value)
2153 1 : case "wal_dir":
2154 1 : o.WALDir = value
2155 1 : case "wal_bytes_per_sync":
2156 1 : o.WALBytesPerSync, err = strconv.Atoi(value)
2157 1 : case "max_writer_concurrency":
2158 : // No longer implemented; ignore.
2159 1 : case "force_writer_parallelism":
2160 : // No longer implemented; ignore.
2161 1 : case "secondary_cache_size_bytes":
2162 1 : o.Experimental.SecondaryCacheSizeBytes, err = strconv.ParseInt(value, 10, 64)
2163 1 : case "create_on_shared":
2164 1 : var createOnSharedInt int64
2165 1 : createOnSharedInt, err = strconv.ParseInt(value, 10, 64)
2166 1 : o.Experimental.CreateOnShared = remote.CreateOnSharedStrategy(createOnSharedInt)
2167 0 : default:
2168 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
2169 0 : return nil
2170 0 : }
2171 0 : return errors.Errorf("pebble: unknown option: %s.%s",
2172 0 : errors.Safe(section), errors.Safe(key))
2173 : }
2174 1 : return err
2175 :
2176 1 : case section == "Value Separation":
2177 1 : var err error
2178 1 : switch key {
2179 1 : case "enabled":
2180 1 : valSepPolicy.Enabled, err = strconv.ParseBool(value)
2181 1 : case "minimum_size":
2182 1 : var minimumSize int
2183 1 : minimumSize, err = strconv.Atoi(value)
2184 1 : valSepPolicy.MinimumSize = minimumSize
2185 1 : case "max_blob_reference_depth":
2186 1 : valSepPolicy.MaxBlobReferenceDepth, err = strconv.Atoi(value)
2187 1 : case "rewrite_minimum_age":
2188 1 : valSepPolicy.RewriteMinimumAge, err = time.ParseDuration(value)
2189 1 : case "target_garbage_ratio":
2190 1 : valSepPolicy.TargetGarbageRatio, err = strconv.ParseFloat(value, 64)
2191 0 : default:
2192 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
2193 0 : return nil
2194 0 : }
2195 0 : return errors.Errorf("pebble: unknown option: %s.%s", errors.Safe(section), errors.Safe(key))
2196 : }
2197 1 : return err
2198 :
2199 1 : case section == "WAL Failover":
2200 1 : if o.WALFailover == nil {
2201 1 : o.WALFailover = new(WALFailoverOptions)
2202 1 : }
2203 1 : var err error
2204 1 : switch key {
2205 1 : case "secondary_dir":
2206 1 : o.WALFailover.Secondary = wal.Dir{Dirname: value, FS: vfs.Default}
2207 1 : case "primary_dir_probe_interval":
2208 1 : o.WALFailover.PrimaryDirProbeInterval, err = time.ParseDuration(value)
2209 1 : case "healthy_probe_latency_threshold":
2210 1 : o.WALFailover.HealthyProbeLatencyThreshold, err = time.ParseDuration(value)
2211 1 : case "healthy_interval":
2212 1 : o.WALFailover.HealthyInterval, err = time.ParseDuration(value)
2213 1 : case "unhealthy_sampling_interval":
2214 1 : o.WALFailover.UnhealthySamplingInterval, err = time.ParseDuration(value)
2215 1 : case "unhealthy_operation_latency_threshold":
2216 1 : var threshold time.Duration
2217 1 : threshold, err = time.ParseDuration(value)
2218 1 : o.WALFailover.UnhealthyOperationLatencyThreshold = func() (time.Duration, bool) {
2219 1 : return threshold, true
2220 1 : }
2221 1 : case "elevated_write_stall_threshold_lag":
2222 1 : o.WALFailover.ElevatedWriteStallThresholdLag, err = time.ParseDuration(value)
2223 0 : default:
2224 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
2225 0 : return nil
2226 0 : }
2227 0 : return errors.Errorf("pebble: unknown option: %s.%s",
2228 0 : errors.Safe(section), errors.Safe(key))
2229 : }
2230 1 : return err
2231 :
2232 1 : case strings.HasPrefix(section, "Level "):
2233 1 : m := regexp.MustCompile(`Level\s*"?(\d+)"?\s*$`).FindStringSubmatch(section)
2234 1 : if m == nil {
2235 0 : return errors.Errorf("pebble: unknown section: %q", errors.Safe(section))
2236 0 : }
2237 1 : index, _ := strconv.Atoi(m[1])
2238 1 :
2239 1 : l := &o.Levels[index]
2240 1 :
2241 1 : var err error
2242 1 : switch key {
2243 1 : case "block_restart_interval":
2244 1 : l.BlockRestartInterval, err = strconv.Atoi(value)
2245 1 : case "block_size":
2246 1 : l.BlockSize, err = strconv.Atoi(value)
2247 1 : case "block_size_threshold":
2248 1 : l.BlockSizeThreshold, err = strconv.Atoi(value)
2249 1 : case "compression":
2250 1 : profile := block.CompressionProfileByName(value)
2251 1 : if profile == nil {
2252 0 : return errors.Errorf("pebble: unknown compression: %q", errors.Safe(value))
2253 0 : }
2254 1 : l.Compression = func() *sstable.CompressionProfile { return profile }
2255 1 : case "filter_policy":
2256 1 : if hooks != nil && hooks.NewFilterPolicy != nil {
2257 1 : l.FilterPolicy, err = hooks.NewFilterPolicy(value)
2258 1 : } else {
2259 1 : l.FilterPolicy = NoFilterPolicy
2260 1 : }
2261 1 : case "filter_type":
2262 1 : switch value {
2263 1 : case "table":
2264 1 : l.FilterType = TableFilter
2265 0 : default:
2266 0 : return errors.Errorf("pebble: unknown filter type: %q", errors.Safe(value))
2267 : }
2268 1 : case "index_block_size":
2269 1 : l.IndexBlockSize, err = strconv.Atoi(value)
2270 1 : case "target_file_size":
2271 1 : o.TargetFileSizes[index], err = strconv.ParseInt(value, 10, 64)
2272 0 : default:
2273 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
2274 0 : return nil
2275 0 : }
2276 0 : return errors.Errorf("pebble: unknown option: %s.%s", errors.Safe(section), errors.Safe(key))
2277 : }
2278 1 : return err
2279 : }
2280 1 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
2281 1 : return nil
2282 1 : }
2283 0 : return errors.Errorf("pebble: unknown section %q or key %q", errors.Safe(section), errors.Safe(key))
2284 : }
2285 1 : err := parseOptions(s, parseOptionsFuncs{
2286 1 : visitKeyValue: visitKeyValue,
2287 1 : })
2288 1 : if err != nil {
2289 1 : return err
2290 1 : }
2291 1 : o.Experimental.ValueSeparationPolicy = func() ValueSeparationPolicy { return valSepPolicy }
2292 1 : if concurrencyLimit.lowerSet || concurrencyLimit.upperSet {
2293 1 : if !concurrencyLimit.lowerSet {
2294 1 : concurrencyLimit.lower = 1
2295 1 : } else if concurrencyLimit.lower < 1 {
2296 0 : return errors.New("baseline_concurrent_compactions cannot be <= 0")
2297 0 : }
2298 1 : if !concurrencyLimit.upperSet {
2299 1 : concurrencyLimit.upper = concurrencyLimit.lower
2300 1 : } else if concurrencyLimit.upper < concurrencyLimit.lower {
2301 0 : return errors.Newf("max_concurrent_compactions cannot be < %d", concurrencyLimit.lower)
2302 0 : }
2303 1 : o.CompactionConcurrencyRange = func() (int, int) {
2304 1 : return concurrencyLimit.lower, concurrencyLimit.upper
2305 1 : }
2306 : }
2307 1 : return nil
2308 : }
2309 :
2310 : // ErrMissingWALRecoveryDir is an error returned when a database is attempted to be
2311 : // opened without supplying a Options.WALRecoveryDir entry for a directory that
2312 : // may contain WALs required to recover a consistent database state.
2313 : type ErrMissingWALRecoveryDir struct {
2314 : Dir string
2315 : ExtraInfo string
2316 : }
2317 :
2318 : // Error implements error.
2319 1 : func (e ErrMissingWALRecoveryDir) Error() string {
2320 1 : return fmt.Sprintf("directory %q may contain relevant WALs but is not in WALRecoveryDirs%s", e.Dir, e.ExtraInfo)
2321 1 : }
2322 :
2323 : // CheckCompatibility verifies the options are compatible with the previous options
2324 : // serialized by Options.String(). For example, the Comparer and Merger must be
2325 : // the same, or data will not be able to be properly read from the DB.
2326 : //
2327 : // This function only looks at specific keys and does not error out if the
2328 : // options are newer and contain unknown keys.
2329 1 : func (o *Options) CheckCompatibility(storeDir string, previousOptions string) error {
2330 1 : previousWALDir := ""
2331 1 :
2332 1 : visitKeyValue := func(i, j int, section, key, value string) error {
2333 1 : switch section + "." + key {
2334 1 : case "Options.comparer":
2335 1 : if value != o.Comparer.Name {
2336 1 : return errors.Errorf("pebble: comparer name from file %q != comparer name from options %q",
2337 1 : errors.Safe(value), errors.Safe(o.Comparer.Name))
2338 1 : }
2339 1 : case "Options.merger":
2340 1 : // RocksDB allows the merge operator to be unspecified, in which case it
2341 1 : // shows up as "nullptr".
2342 1 : if value != "nullptr" && value != o.Merger.Name {
2343 1 : return errors.Errorf("pebble: merger name from file %q != merger name from options %q",
2344 1 : errors.Safe(value), errors.Safe(o.Merger.Name))
2345 1 : }
2346 1 : case "Options.wal_dir":
2347 1 : previousWALDir = value
2348 1 : case "WAL Failover.secondary_dir":
2349 1 : previousWALSecondaryDir := value
2350 1 : if err := o.checkWALDir(storeDir, previousWALSecondaryDir, "WALFailover.Secondary changed from previous options"); err != nil {
2351 1 : return err
2352 1 : }
2353 : }
2354 1 : return nil
2355 : }
2356 1 : if err := parseOptions(previousOptions, parseOptionsFuncs{visitKeyValue: visitKeyValue}); err != nil {
2357 1 : return err
2358 1 : }
2359 1 : if err := o.checkWALDir(storeDir, previousWALDir, "WALDir changed from previous options"); err != nil {
2360 1 : return err
2361 1 : }
2362 1 : return nil
2363 : }
2364 :
2365 : // checkWALDir verifies that walDir is among o.WALDir, o.WALFailover.Secondary,
2366 : // or o.WALRecoveryDirs. An empty "walDir" maps to the storeDir.
2367 1 : func (o *Options) checkWALDir(storeDir, walDir, errContext string) error {
2368 1 : walPath := resolveStorePath(storeDir, walDir)
2369 1 : if walDir == "" {
2370 1 : walPath = storeDir
2371 1 : }
2372 :
2373 1 : if o.WALDir == "" {
2374 1 : if walPath == storeDir {
2375 1 : return nil
2376 1 : }
2377 1 : } else {
2378 1 : if walPath == resolveStorePath(storeDir, o.WALDir) {
2379 1 : return nil
2380 1 : }
2381 : }
2382 :
2383 1 : if o.WALFailover != nil && walPath == resolveStorePath(storeDir, o.WALFailover.Secondary.Dirname) {
2384 1 : return nil
2385 1 : }
2386 :
2387 1 : for _, d := range o.WALRecoveryDirs {
2388 1 : // TODO(radu): should we also check that d.FS is the same as walDir's FS?
2389 1 : if walPath == resolveStorePath(storeDir, d.Dirname) {
2390 1 : return nil
2391 1 : }
2392 : }
2393 :
2394 1 : var buf bytes.Buffer
2395 1 : fmt.Fprintf(&buf, "\n %s\n", errContext)
2396 1 : fmt.Fprintf(&buf, " o.WALDir: %q\n", o.WALDir)
2397 1 : if o.WALFailover != nil {
2398 0 : fmt.Fprintf(&buf, " o.WALFailover.Secondary.Dirname: %q\n", o.WALFailover.Secondary.Dirname)
2399 0 : }
2400 1 : fmt.Fprintf(&buf, " o.WALRecoveryDirs: %d", len(o.WALRecoveryDirs))
2401 1 : for _, d := range o.WALRecoveryDirs {
2402 0 : fmt.Fprintf(&buf, "\n %q", d.Dirname)
2403 0 : }
2404 1 : return ErrMissingWALRecoveryDir{Dir: walPath, ExtraInfo: buf.String()}
2405 : }
2406 :
2407 : // Validate verifies that the options are mutually consistent. For example,
2408 : // L0StopWritesThreshold must be >= L0CompactionThreshold, otherwise a write
2409 : // stall would persist indefinitely.
2410 1 : func (o *Options) Validate() error {
2411 1 : // Note that we can presume Options.EnsureDefaults has been called, so there
2412 1 : // is no need to check for zero values.
2413 1 :
2414 1 : var buf strings.Builder
2415 1 : if o.Experimental.L0CompactionConcurrency < 1 {
2416 1 : fmt.Fprintf(&buf, "L0CompactionConcurrency (%d) must be >= 1\n",
2417 1 : o.Experimental.L0CompactionConcurrency)
2418 1 : }
2419 1 : if o.L0StopWritesThreshold < o.L0CompactionThreshold {
2420 1 : fmt.Fprintf(&buf, "L0StopWritesThreshold (%d) must be >= L0CompactionThreshold (%d)\n",
2421 1 : o.L0StopWritesThreshold, o.L0CompactionThreshold)
2422 1 : }
2423 1 : if uint64(o.MemTableSize) >= maxMemTableSize {
2424 1 : fmt.Fprintf(&buf, "MemTableSize (%s) must be < %s\n",
2425 1 : humanize.Bytes.Uint64(uint64(o.MemTableSize)), humanize.Bytes.Uint64(maxMemTableSize))
2426 1 : }
2427 1 : if o.MemTableStopWritesThreshold < 2 {
2428 1 : fmt.Fprintf(&buf, "MemTableStopWritesThreshold (%d) must be >= 2\n",
2429 1 : o.MemTableStopWritesThreshold)
2430 1 : }
2431 1 : if o.FormatMajorVersion < FormatMinSupported || o.FormatMajorVersion > internalFormatNewest {
2432 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) must be between %d and %d\n",
2433 0 : o.FormatMajorVersion, FormatMinSupported, internalFormatNewest)
2434 0 : }
2435 1 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone && o.FormatMajorVersion < FormatMinForSharedObjects {
2436 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) when CreateOnShared is set must be at least %d\n",
2437 0 : o.FormatMajorVersion, FormatMinForSharedObjects)
2438 0 : }
2439 1 : if len(o.KeySchemas) > 0 {
2440 1 : if o.KeySchema == "" {
2441 0 : fmt.Fprintf(&buf, "KeySchemas is set but KeySchema is not\n")
2442 0 : }
2443 1 : if _, ok := o.KeySchemas[o.KeySchema]; !ok {
2444 0 : fmt.Fprintf(&buf, "KeySchema %q not found in KeySchemas\n", o.KeySchema)
2445 0 : }
2446 : }
2447 1 : if policy := o.Experimental.ValueSeparationPolicy(); policy.Enabled {
2448 1 : if policy.MinimumSize <= 0 {
2449 0 : fmt.Fprintf(&buf, "ValueSeparationPolicy.MinimumSize (%d) must be > 0\n", policy.MinimumSize)
2450 0 : }
2451 1 : if policy.MaxBlobReferenceDepth <= 0 {
2452 0 : fmt.Fprintf(&buf, "ValueSeparationPolicy.MaxBlobReferenceDepth (%d) must be > 0\n", policy.MaxBlobReferenceDepth)
2453 0 : }
2454 : }
2455 :
2456 1 : if buf.Len() == 0 {
2457 1 : return nil
2458 1 : }
2459 1 : return errors.New(buf.String())
2460 : }
2461 :
2462 : // MakeReaderOptions constructs sstable.ReaderOptions from the corresponding
2463 : // options in the receiver.
2464 1 : func (o *Options) MakeReaderOptions() sstable.ReaderOptions {
2465 1 : var readerOpts sstable.ReaderOptions
2466 1 : if o != nil {
2467 1 : readerOpts.Comparer = o.Comparer
2468 1 : readerOpts.Filters = o.Filters
2469 1 : readerOpts.KeySchemas = o.KeySchemas
2470 1 : readerOpts.LoadBlockSema = o.LoadBlockSema
2471 1 : readerOpts.LoggerAndTracer = o.LoggerAndTracer
2472 1 : readerOpts.Merger = o.Merger
2473 1 : }
2474 1 : return readerOpts
2475 : }
2476 :
2477 : // MakeWriterOptions constructs sstable.WriterOptions for the specified level
2478 : // from the corresponding options in the receiver.
2479 1 : func (o *Options) MakeWriterOptions(level int, format sstable.TableFormat) sstable.WriterOptions {
2480 1 : var writerOpts sstable.WriterOptions
2481 1 : writerOpts.TableFormat = format
2482 1 : if o != nil {
2483 1 : writerOpts.Comparer = o.Comparer
2484 1 : if o.Merger != nil {
2485 1 : writerOpts.MergerName = o.Merger.Name
2486 1 : }
2487 1 : writerOpts.BlockPropertyCollectors = o.BlockPropertyCollectors
2488 : }
2489 1 : if format >= sstable.TableFormatPebblev3 {
2490 1 : writerOpts.ShortAttributeExtractor = o.Experimental.ShortAttributeExtractor
2491 1 : if format >= sstable.TableFormatPebblev4 && level == numLevels-1 {
2492 1 : writerOpts.WritingToLowestLevel = true
2493 1 : }
2494 : }
2495 1 : levelOpts := o.Levels[level]
2496 1 : writerOpts.BlockRestartInterval = levelOpts.BlockRestartInterval
2497 1 : writerOpts.BlockSize = levelOpts.BlockSize
2498 1 : writerOpts.BlockSizeThreshold = levelOpts.BlockSizeThreshold
2499 1 : writerOpts.Compression = levelOpts.Compression()
2500 1 : writerOpts.FilterPolicy = levelOpts.FilterPolicy
2501 1 : writerOpts.FilterType = levelOpts.FilterType
2502 1 : writerOpts.IndexBlockSize = levelOpts.IndexBlockSize
2503 1 : if o.KeySchema != "" {
2504 1 : var ok bool
2505 1 : writerOpts.KeySchema, ok = o.KeySchemas[o.KeySchema]
2506 1 : if !ok {
2507 0 : panic(fmt.Sprintf("invalid schema %q", redact.Safe(o.KeySchema)))
2508 : }
2509 : }
2510 1 : writerOpts.AllocatorSizeClasses = o.AllocatorSizeClasses
2511 1 : writerOpts.NumDeletionsThreshold = o.Experimental.NumDeletionsThreshold
2512 1 : writerOpts.DeletionSizeRatioThreshold = o.Experimental.DeletionSizeRatioThreshold
2513 1 : return writerOpts
2514 : }
2515 :
2516 : // MakeBlobWriterOptions constructs blob.FileWriterOptions from the corresponding
2517 : // options in the receiver.
2518 1 : func (o *Options) MakeBlobWriterOptions(level int) blob.FileWriterOptions {
2519 1 : lo := o.Levels[level]
2520 1 : return blob.FileWriterOptions{
2521 1 : Compression: lo.Compression(),
2522 1 : ChecksumType: block.ChecksumTypeCRC32c,
2523 1 : FlushGovernor: block.MakeFlushGovernor(
2524 1 : lo.BlockSize,
2525 1 : lo.BlockSizeThreshold,
2526 1 : base.SizeClassAwareBlockSizeThreshold,
2527 1 : o.AllocatorSizeClasses,
2528 1 : ),
2529 1 : }
2530 1 : }
2531 :
2532 1 : func (o *Options) MakeObjStorageProviderSettings(dirname string) objstorageprovider.Settings {
2533 1 : s := objstorageprovider.Settings{
2534 1 : Logger: o.Logger,
2535 1 : FS: o.FS,
2536 1 : FSDirName: dirname,
2537 1 : FSCleaner: o.Cleaner,
2538 1 : NoSyncOnClose: o.NoSyncOnClose,
2539 1 : BytesPerSync: o.BytesPerSync,
2540 1 : }
2541 1 : s.Local.ReadaheadConfig = o.Local.ReadaheadConfig
2542 1 : s.Remote.StorageFactory = o.Experimental.RemoteStorage
2543 1 : s.Remote.CreateOnShared = o.Experimental.CreateOnShared
2544 1 : s.Remote.CreateOnSharedLocator = o.Experimental.CreateOnSharedLocator
2545 1 : s.Remote.CacheSizeBytes = o.Experimental.SecondaryCacheSizeBytes
2546 1 : return s
2547 1 : }
2548 :
2549 : // UserKeyCategories describes a partitioning of the user key space. Each
2550 : // partition is a category with a name. The categories are used for informative
2551 : // purposes only (like pprof labels). Pebble does not treat keys differently
2552 : // based on the UserKeyCategories.
2553 : //
2554 : // The partitions are defined by their upper bounds. The last partition is
2555 : // assumed to go until the end of keyspace; its UpperBound is ignored. The rest
2556 : // of the partitions are ordered by their UpperBound.
2557 : type UserKeyCategories struct {
2558 : categories []UserKeyCategory
2559 : cmp base.Compare
2560 : // rangeNames[i][j] contains the string referring to the categories in the
2561 : // range [i, j], with j > i.
2562 : rangeNames [][]string
2563 : }
2564 :
2565 : // UserKeyCategory describes a partition of the user key space.
2566 : //
2567 : // User keys >= the previous category's UpperBound and < this category's
2568 : // UpperBound are part of this category.
2569 : type UserKeyCategory struct {
2570 : Name string
2571 : // UpperBound is the exclusive upper bound of the category. All user keys >= the
2572 : // previous category's UpperBound and < this UpperBound are part of this
2573 : // category.
2574 : UpperBound []byte
2575 : }
2576 :
2577 : // MakeUserKeyCategories creates a UserKeyCategories object with the given
2578 : // categories. The object is immutable and can be reused across different
2579 : // stores.
2580 1 : func MakeUserKeyCategories(cmp base.Compare, categories ...UserKeyCategory) UserKeyCategories {
2581 1 : n := len(categories)
2582 1 : if n == 0 {
2583 0 : return UserKeyCategories{}
2584 0 : }
2585 1 : if categories[n-1].UpperBound != nil {
2586 0 : panic("last category UpperBound must be nil")
2587 : }
2588 : // Verify that the partitions are ordered as expected.
2589 1 : for i := 1; i < n-1; i++ {
2590 1 : if cmp(categories[i-1].UpperBound, categories[i].UpperBound) >= 0 {
2591 0 : panic("invalid UserKeyCategories: key prefixes must be sorted")
2592 : }
2593 : }
2594 :
2595 : // Precalculate a table of range names to avoid allocations in the
2596 : // categorization path.
2597 1 : rangeNamesBuf := make([]string, n*n)
2598 1 : rangeNames := make([][]string, n)
2599 1 : for i := range rangeNames {
2600 1 : rangeNames[i] = rangeNamesBuf[:n]
2601 1 : rangeNamesBuf = rangeNamesBuf[n:]
2602 1 : for j := i + 1; j < n; j++ {
2603 1 : rangeNames[i][j] = categories[i].Name + "-" + categories[j].Name
2604 1 : }
2605 : }
2606 1 : return UserKeyCategories{
2607 1 : categories: categories,
2608 1 : cmp: cmp,
2609 1 : rangeNames: rangeNames,
2610 1 : }
2611 : }
2612 :
2613 : // Len returns the number of categories defined.
2614 1 : func (kc *UserKeyCategories) Len() int {
2615 1 : return len(kc.categories)
2616 1 : }
2617 :
2618 : // CategorizeKey returns the name of the category containing the key.
2619 1 : func (kc *UserKeyCategories) CategorizeKey(userKey []byte) string {
2620 1 : idx := sort.Search(len(kc.categories)-1, func(i int) bool {
2621 1 : return kc.cmp(userKey, kc.categories[i].UpperBound) < 0
2622 1 : })
2623 1 : return kc.categories[idx].Name
2624 : }
2625 :
2626 : // CategorizeKeyRange returns the name of the category containing the key range.
2627 : // If the key range spans multiple categories, the result shows the first and
2628 : // last category separated by a dash, e.g. `cat1-cat5`.
2629 1 : func (kc *UserKeyCategories) CategorizeKeyRange(startUserKey, endUserKey []byte) string {
2630 1 : n := len(kc.categories)
2631 1 : p := sort.Search(n-1, func(i int) bool {
2632 1 : return kc.cmp(startUserKey, kc.categories[i].UpperBound) < 0
2633 1 : })
2634 1 : if p == n-1 || kc.cmp(endUserKey, kc.categories[p].UpperBound) < 0 {
2635 1 : // Fast path for a single category.
2636 1 : return kc.categories[p].Name
2637 1 : }
2638 : // Binary search among the remaining categories.
2639 1 : q := p + 1 + sort.Search(n-2-p, func(i int) bool {
2640 1 : return kc.cmp(endUserKey, kc.categories[p+1+i].UpperBound) < 0
2641 1 : })
2642 1 : return kc.rangeNames[p][q]
2643 : }
2644 :
2645 : const storePathIdentifier = "{store_path}"
2646 :
2647 : // MakeStoreRelativePath takes a path that is relative to the store directory
2648 : // and creates a path that can be used for Options.WALDir and wal.Dir.Dirname.
2649 : //
2650 : // This is used in metamorphic tests, so that the test run directory can be
2651 : // copied or moved.
2652 1 : func MakeStoreRelativePath(fs vfs.FS, relativePath string) string {
2653 1 : if relativePath == "" {
2654 0 : return storePathIdentifier
2655 0 : }
2656 1 : return fs.PathJoin(storePathIdentifier, relativePath)
2657 : }
2658 :
2659 : // resolveStorePath is the inverse of MakeStoreRelativePath(). It replaces any
2660 : // storePathIdentifier prefix with the store dir.
2661 1 : func resolveStorePath(storeDir, path string) string {
2662 1 : if remainder, ok := strings.CutPrefix(path, storePathIdentifier); ok {
2663 1 : return storeDir + remainder
2664 1 : }
2665 1 : return path
2666 : }
|
