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 : "runtime"
12 : "strconv"
13 : "strings"
14 : "time"
15 : "unicode"
16 :
17 : "github.com/cockroachdb/errors"
18 : "github.com/cockroachdb/pebble/internal/base"
19 : "github.com/cockroachdb/pebble/internal/cache"
20 : "github.com/cockroachdb/pebble/internal/humanize"
21 : "github.com/cockroachdb/pebble/internal/keyspan"
22 : "github.com/cockroachdb/pebble/internal/manifest"
23 : "github.com/cockroachdb/pebble/objstorage/remote"
24 : "github.com/cockroachdb/pebble/rangekey"
25 : "github.com/cockroachdb/pebble/sstable"
26 : "github.com/cockroachdb/pebble/vfs"
27 : )
28 :
29 : const (
30 : cacheDefaultSize = 8 << 20 // 8 MB
31 : defaultLevelMultiplier = 10
32 : )
33 :
34 : // Compression exports the base.Compression type.
35 : type Compression = sstable.Compression
36 :
37 : // Exported Compression constants.
38 : const (
39 : DefaultCompression = sstable.DefaultCompression
40 : NoCompression = sstable.NoCompression
41 : SnappyCompression = sstable.SnappyCompression
42 : ZstdCompression = sstable.ZstdCompression
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 : // TablePropertyCollector exports the sstable.TablePropertyCollector type.
60 : type TablePropertyCollector = sstable.TablePropertyCollector
61 :
62 : // BlockPropertyCollector exports the sstable.BlockPropertyCollector type.
63 : type BlockPropertyCollector = sstable.BlockPropertyCollector
64 :
65 : // BlockPropertyFilter exports the sstable.BlockPropertyFilter type.
66 : type BlockPropertyFilter = base.BlockPropertyFilter
67 :
68 : // ShortAttributeExtractor exports the base.ShortAttributeExtractor type.
69 : type ShortAttributeExtractor = base.ShortAttributeExtractor
70 :
71 : // UserKeyPrefixBound exports the sstable.UserKeyPrefixBound type.
72 : type UserKeyPrefixBound = sstable.UserKeyPrefixBound
73 :
74 : // IterKeyType configures which types of keys an iterator should surface.
75 : type IterKeyType int8
76 :
77 : const (
78 : // IterKeyTypePointsOnly configures an iterator to iterate over point keys
79 : // only.
80 : IterKeyTypePointsOnly IterKeyType = iota
81 : // IterKeyTypeRangesOnly configures an iterator to iterate over range keys
82 : // only.
83 : IterKeyTypeRangesOnly
84 : // IterKeyTypePointsAndRanges configures an iterator iterate over both point
85 : // keys and range keys simultaneously.
86 : IterKeyTypePointsAndRanges
87 : )
88 :
89 : // String implements fmt.Stringer.
90 0 : func (t IterKeyType) String() string {
91 0 : switch t {
92 0 : case IterKeyTypePointsOnly:
93 0 : return "points-only"
94 0 : case IterKeyTypeRangesOnly:
95 0 : return "ranges-only"
96 0 : case IterKeyTypePointsAndRanges:
97 0 : return "points-and-ranges"
98 0 : default:
99 0 : panic(fmt.Sprintf("unknown key type %d", t))
100 : }
101 : }
102 :
103 : // IterOptions hold the optional per-query parameters for NewIter.
104 : //
105 : // Like Options, a nil *IterOptions is valid and means to use the default
106 : // values.
107 : type IterOptions struct {
108 : // LowerBound specifies the smallest key (inclusive) that the iterator will
109 : // return during iteration. If the iterator is seeked or iterated past this
110 : // boundary the iterator will return Valid()==false. Setting LowerBound
111 : // effectively truncates the key space visible to the iterator.
112 : LowerBound []byte
113 : // UpperBound specifies the largest key (exclusive) that the iterator will
114 : // return during iteration. If the iterator is seeked or iterated past this
115 : // boundary the iterator will return Valid()==false. Setting UpperBound
116 : // effectively truncates the key space visible to the iterator.
117 : UpperBound []byte
118 : // TableFilter can be used to filter the tables that are scanned during
119 : // iteration based on the user properties. Return true to scan the table and
120 : // false to skip scanning. This function must be thread-safe since the same
121 : // function can be used by multiple iterators, if the iterator is cloned.
122 : TableFilter func(userProps map[string]string) bool
123 : // SkipPoint may be used to skip over point keys that don't match an
124 : // arbitrary predicate during iteration. If set, the Iterator invokes
125 : // SkipPoint for keys encountered. If SkipPoint returns true, the iterator
126 : // will skip the key without yielding it to the iterator operation in
127 : // progress.
128 : //
129 : // SkipPoint must be a pure function and always return the same result when
130 : // provided the same arguments. The iterator may call SkipPoint multiple
131 : // times for the same user key.
132 : SkipPoint func(userKey []byte) bool
133 : // PointKeyFilters can be used to avoid scanning tables and blocks in tables
134 : // when iterating over point keys. This slice represents an intersection
135 : // across all filters, i.e., all filters must indicate that the block is
136 : // relevant.
137 : //
138 : // Performance note: When len(PointKeyFilters) > 0, the caller should ensure
139 : // that cap(PointKeyFilters) is at least len(PointKeyFilters)+1. This helps
140 : // avoid allocations in Pebble internal code that mutates the slice.
141 : PointKeyFilters []BlockPropertyFilter
142 : // RangeKeyFilters can be usefd to avoid scanning tables and blocks in tables
143 : // when iterating over range keys. The same requirements that apply to
144 : // PointKeyFilters apply here too.
145 : RangeKeyFilters []BlockPropertyFilter
146 : // KeyTypes configures which types of keys to iterate over: point keys,
147 : // range keys, or both.
148 : KeyTypes IterKeyType
149 : // RangeKeyMasking can be used to enable automatic masking of point keys by
150 : // range keys. Range key masking is only supported during combined range key
151 : // and point key iteration mode (IterKeyTypePointsAndRanges).
152 : RangeKeyMasking RangeKeyMasking
153 :
154 : // OnlyReadGuaranteedDurable is an advanced option that is only supported by
155 : // the Reader implemented by DB. When set to true, only the guaranteed to be
156 : // durable state is visible in the iterator.
157 : // - This definition is made under the assumption that the FS implementation
158 : // is providing a durability guarantee when data is synced.
159 : // - The visible state represents a consistent point in the history of the
160 : // DB.
161 : // - The implementation is free to choose a conservative definition of what
162 : // is guaranteed durable. For simplicity, the current implementation
163 : // ignores memtables. A more sophisticated implementation could track the
164 : // highest seqnum that is synced to the WAL and published and use that as
165 : // the visible seqnum for an iterator. Note that the latter approach is
166 : // not strictly better than the former since we can have DBs that are (a)
167 : // synced more rarely than memtable flushes, (b) have no WAL. (a) is
168 : // likely to be true in a future CockroachDB context where the DB
169 : // containing the state machine may be rarely synced.
170 : // NB: this current implementation relies on the fact that memtables are
171 : // flushed in seqnum order, and any ingested sstables that happen to have a
172 : // lower seqnum than a non-flushed memtable don't have any overlapping keys.
173 : // This is the fundamental level invariant used in other code too, like when
174 : // merging iterators.
175 : //
176 : // Semantically, using this option provides the caller a "snapshot" as of
177 : // the time the most recent memtable was flushed. An alternate interface
178 : // would be to add a NewSnapshot variant. Creating a snapshot is heavier
179 : // weight than creating an iterator, so we have opted to support this
180 : // iterator option.
181 : OnlyReadGuaranteedDurable bool
182 : // UseL6Filters allows the caller to opt into reading filter blocks for L6
183 : // sstables. Helpful if a lot of SeekPrefixGEs are expected in quick
184 : // succession, that are also likely to not yield a single key. Filter blocks in
185 : // L6 can be relatively large, often larger than data blocks, so the benefit of
186 : // loading them in the cache is minimized if the probability of the key
187 : // existing is not low or if we just expect a one-time Seek (where loading the
188 : // data block directly is better).
189 : UseL6Filters bool
190 : // CategoryAndQoS is used for categorized iterator stats. This should not be
191 : // changed by calling SetOptions.
192 : sstable.CategoryAndQoS
193 :
194 : // Internal options.
195 :
196 : logger Logger
197 : // Level corresponding to this file. Only passed in if constructed by a
198 : // levelIter.
199 : level manifest.Level
200 : // disableLazyCombinedIteration is an internal testing option.
201 : disableLazyCombinedIteration bool
202 : // snapshotForHideObsoletePoints is specified for/by levelIter when opening
203 : // files and is used to decide whether to hide obsolete points. A value of 0
204 : // implies obsolete points should not be hidden.
205 : snapshotForHideObsoletePoints uint64
206 :
207 : // NB: If adding new Options, you must account for them in iterator
208 : // construction and Iterator.SetOptions.
209 : }
210 :
211 : // GetLowerBound returns the LowerBound or nil if the receiver is nil.
212 1 : func (o *IterOptions) GetLowerBound() []byte {
213 1 : if o == nil {
214 1 : return nil
215 1 : }
216 1 : return o.LowerBound
217 : }
218 :
219 : // GetUpperBound returns the UpperBound or nil if the receiver is nil.
220 1 : func (o *IterOptions) GetUpperBound() []byte {
221 1 : if o == nil {
222 1 : return nil
223 1 : }
224 1 : return o.UpperBound
225 : }
226 :
227 1 : func (o *IterOptions) pointKeys() bool {
228 1 : if o == nil {
229 0 : return true
230 0 : }
231 1 : return o.KeyTypes == IterKeyTypePointsOnly || o.KeyTypes == IterKeyTypePointsAndRanges
232 : }
233 :
234 1 : func (o *IterOptions) rangeKeys() bool {
235 1 : if o == nil {
236 0 : return false
237 0 : }
238 1 : return o.KeyTypes == IterKeyTypeRangesOnly || o.KeyTypes == IterKeyTypePointsAndRanges
239 : }
240 :
241 1 : func (o *IterOptions) getLogger() Logger {
242 1 : if o == nil || o.logger == nil {
243 1 : return DefaultLogger
244 1 : }
245 1 : return o.logger
246 : }
247 :
248 : // SpanIterOptions creates a SpanIterOptions from this IterOptions.
249 1 : func (o *IterOptions) SpanIterOptions() keyspan.SpanIterOptions {
250 1 : if o == nil {
251 0 : return keyspan.SpanIterOptions{}
252 0 : }
253 1 : return keyspan.SpanIterOptions{
254 1 : RangeKeyFilters: o.RangeKeyFilters,
255 1 : }
256 : }
257 :
258 : // scanInternalOptions is similar to IterOptions, meant for use with
259 : // scanInternalIterator.
260 : type scanInternalOptions struct {
261 : sstable.CategoryAndQoS
262 : IterOptions
263 :
264 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error
265 : visitRangeDel func(start, end []byte, seqNum uint64) error
266 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error
267 : visitSharedFile func(sst *SharedSSTMeta) error
268 :
269 : // skipSharedLevels skips levels that are shareable (level >=
270 : // sharedLevelStart).
271 : skipSharedLevels bool
272 :
273 : // includeObsoleteKeys specifies whether keys shadowed by newer internal keys
274 : // are exposed. If false, only one internal key per user key is exposed.
275 : includeObsoleteKeys bool
276 :
277 : // rateLimitFunc is used to limit the amount of bytes read per second.
278 : rateLimitFunc func(key *InternalKey, value LazyValue) error
279 : }
280 :
281 : // RangeKeyMasking configures automatic hiding of point keys by range keys. A
282 : // non-nil Suffix enables range-key masking. When enabled, range keys with
283 : // suffixes ≥ Suffix behave as masks. All point keys that are contained within a
284 : // masking range key's bounds and have suffixes greater than the range key's
285 : // suffix are automatically skipped.
286 : //
287 : // Specifically, when configured with a RangeKeyMasking.Suffix _s_, and there
288 : // exists a range key with suffix _r_ covering a point key with suffix _p_, and
289 : //
290 : // _s_ ≤ _r_ < _p_
291 : //
292 : // then the point key is elided.
293 : //
294 : // Range-key masking may only be used when iterating over both point keys and
295 : // range keys with IterKeyTypePointsAndRanges.
296 : type RangeKeyMasking struct {
297 : // Suffix configures which range keys may mask point keys. Only range keys
298 : // that are defined at suffixes greater than or equal to Suffix will mask
299 : // point keys.
300 : Suffix []byte
301 : // Filter is an optional field that may be used to improve performance of
302 : // range-key masking through a block-property filter defined over key
303 : // suffixes. If non-nil, Filter is called by Pebble to construct a
304 : // block-property filter mask at iterator creation. The filter is used to
305 : // skip whole point-key blocks containing point keys with suffixes greater
306 : // than a covering range-key's suffix.
307 : //
308 : // To use this functionality, the caller must create and configure (through
309 : // Options.BlockPropertyCollectors) a block-property collector that records
310 : // the maxmimum suffix contained within a block. The caller then must write
311 : // and provide a BlockPropertyFilterMask implementation on that same
312 : // property. See the BlockPropertyFilterMask type for more information.
313 : Filter func() BlockPropertyFilterMask
314 : }
315 :
316 : // BlockPropertyFilterMask extends the BlockPropertyFilter interface for use
317 : // with range-key masking. Unlike an ordinary block property filter, a
318 : // BlockPropertyFilterMask's filtering criteria is allowed to change when Pebble
319 : // invokes its SetSuffix method.
320 : //
321 : // When a Pebble iterator steps into a range key's bounds and the range key has
322 : // a suffix greater than or equal to RangeKeyMasking.Suffix, the range key acts
323 : // as a mask. The masking range key hides all point keys that fall within the
324 : // range key's bounds and have suffixes > the range key's suffix. Without a
325 : // filter mask configured, Pebble performs this hiding by stepping through point
326 : // keys and comparing suffixes. If large numbers of point keys are masked, this
327 : // requires Pebble to load, iterate through and discard a large number of
328 : // sstable blocks containing masked point keys.
329 : //
330 : // If a block-property collector and a filter mask are configured, Pebble may
331 : // skip loading some point-key blocks altogether. If a block's keys are known to
332 : // all fall within the bounds of the masking range key and the block was
333 : // annotated by a block-property collector with the maximal suffix, Pebble can
334 : // ask the filter mask to compare the property to the current masking range
335 : // key's suffix. If the mask reports no intersection, the block may be skipped.
336 : //
337 : // If unsuffixed and suffixed keys are written to the database, care must be
338 : // taken to avoid unintentionally masking un-suffixed keys located in the same
339 : // block as suffixed keys. One solution is to interpret unsuffixed keys as
340 : // containing the maximal suffix value, ensuring that blocks containing
341 : // unsuffixed keys are always loaded.
342 : type BlockPropertyFilterMask interface {
343 : BlockPropertyFilter
344 :
345 : // SetSuffix configures the mask with the suffix of a range key. The filter
346 : // should return false from Intersects whenever it's provided with a
347 : // property encoding a block's minimum suffix that's greater (according to
348 : // Compare) than the provided suffix.
349 : SetSuffix(suffix []byte) error
350 : }
351 :
352 : // WriteOptions hold the optional per-query parameters for Set and Delete
353 : // operations.
354 : //
355 : // Like Options, a nil *WriteOptions is valid and means to use the default
356 : // values.
357 : type WriteOptions struct {
358 : // Sync is whether to sync writes through the OS buffer cache and down onto
359 : // the actual disk, if applicable. Setting Sync is required for durability of
360 : // individual write operations but can result in slower writes.
361 : //
362 : // If false, and the process or machine crashes, then a recent write may be
363 : // lost. This is due to the recently written data being buffered inside the
364 : // process running Pebble. This differs from the semantics of a write system
365 : // call in which the data is buffered in the OS buffer cache and would thus
366 : // survive a process crash.
367 : //
368 : // The default value is true.
369 : Sync bool
370 : }
371 :
372 : // Sync specifies the default write options for writes which synchronize to
373 : // disk.
374 : var Sync = &WriteOptions{Sync: true}
375 :
376 : // NoSync specifies the default write options for writes which do not
377 : // synchronize to disk.
378 : var NoSync = &WriteOptions{Sync: false}
379 :
380 : // GetSync returns the Sync value or true if the receiver is nil.
381 1 : func (o *WriteOptions) GetSync() bool {
382 1 : return o == nil || o.Sync
383 1 : }
384 :
385 : // LevelOptions holds the optional per-level parameters.
386 : type LevelOptions struct {
387 : // BlockRestartInterval is the number of keys between restart points
388 : // for delta encoding of keys.
389 : //
390 : // The default value is 16.
391 : BlockRestartInterval int
392 :
393 : // BlockSize is the target uncompressed size in bytes of each table block.
394 : //
395 : // The default value is 4096.
396 : BlockSize int
397 :
398 : // BlockSizeThreshold finishes a block if the block size is larger than the
399 : // specified percentage of the target block size and adding the next entry
400 : // would cause the block to be larger than the target block size.
401 : //
402 : // The default value is 90
403 : BlockSizeThreshold int
404 :
405 : // Compression defines the per-block compression to use.
406 : //
407 : // The default value (DefaultCompression) uses snappy compression.
408 : Compression Compression
409 :
410 : // FilterPolicy defines a filter algorithm (such as a Bloom filter) that can
411 : // reduce disk reads for Get calls.
412 : //
413 : // One such implementation is bloom.FilterPolicy(10) from the pebble/bloom
414 : // package.
415 : //
416 : // The default value means to use no filter.
417 : FilterPolicy FilterPolicy
418 :
419 : // FilterType defines whether an existing filter policy is applied at a
420 : // block-level or table-level. Block-level filters use less memory to create,
421 : // but are slower to access as a check for the key in the index must first be
422 : // performed to locate the filter block. A table-level filter will require
423 : // memory proportional to the number of keys in an sstable to create, but
424 : // avoids the index lookup when determining if a key is present. Table-level
425 : // filters should be preferred except under constrained memory situations.
426 : FilterType FilterType
427 :
428 : // IndexBlockSize is the target uncompressed size in bytes of each index
429 : // block. When the index block size is larger than this target, two-level
430 : // indexes are automatically enabled. Setting this option to a large value
431 : // (such as math.MaxInt32) disables the automatic creation of two-level
432 : // indexes.
433 : //
434 : // The default value is the value of BlockSize.
435 : IndexBlockSize int
436 :
437 : // The target file size for the level.
438 : TargetFileSize int64
439 : }
440 :
441 : // EnsureDefaults ensures that the default values for all of the options have
442 : // been initialized. It is valid to call EnsureDefaults on a nil receiver. A
443 : // non-nil result will always be returned.
444 1 : func (o *LevelOptions) EnsureDefaults() *LevelOptions {
445 1 : if o == nil {
446 0 : o = &LevelOptions{}
447 0 : }
448 1 : if o.BlockRestartInterval <= 0 {
449 0 : o.BlockRestartInterval = base.DefaultBlockRestartInterval
450 0 : }
451 1 : if o.BlockSize <= 0 {
452 0 : o.BlockSize = base.DefaultBlockSize
453 1 : } else if o.BlockSize > sstable.MaximumBlockSize {
454 0 : panic(errors.Errorf("BlockSize %d exceeds MaximumBlockSize", o.BlockSize))
455 : }
456 1 : if o.BlockSizeThreshold <= 0 {
457 0 : o.BlockSizeThreshold = base.DefaultBlockSizeThreshold
458 0 : }
459 1 : if o.Compression <= DefaultCompression || o.Compression >= sstable.NCompression {
460 0 : o.Compression = SnappyCompression
461 0 : }
462 1 : if o.IndexBlockSize <= 0 {
463 0 : o.IndexBlockSize = o.BlockSize
464 0 : }
465 1 : if o.TargetFileSize <= 0 {
466 0 : o.TargetFileSize = 2 << 20 // 2 MB
467 0 : }
468 1 : return o
469 : }
470 :
471 : // Options holds the optional parameters for configuring pebble. These options
472 : // apply to the DB at large; per-query options are defined by the IterOptions
473 : // and WriteOptions types.
474 : type Options struct {
475 : // Sync sstables periodically in order to smooth out writes to disk. This
476 : // option does not provide any persistency guarantee, but is used to avoid
477 : // latency spikes if the OS automatically decides to write out a large chunk
478 : // of dirty filesystem buffers. This option only controls SSTable syncs; WAL
479 : // syncs are controlled by WALBytesPerSync.
480 : //
481 : // The default value is 512KB.
482 : BytesPerSync int
483 :
484 : // Cache is used to cache uncompressed blocks from sstables.
485 : //
486 : // The default cache size is 8 MB.
487 : Cache *cache.Cache
488 :
489 : // Cleaner cleans obsolete files.
490 : //
491 : // The default cleaner uses the DeleteCleaner.
492 : Cleaner Cleaner
493 :
494 : // Comparer defines a total ordering over the space of []byte keys: a 'less
495 : // than' relationship. The same comparison algorithm must be used for reads
496 : // and writes over the lifetime of the DB.
497 : //
498 : // The default value uses the same ordering as bytes.Compare.
499 : Comparer *Comparer
500 :
501 : // DebugCheck is invoked, if non-nil, whenever a new version is being
502 : // installed. Typically, this is set to pebble.DebugCheckLevels in tests
503 : // or tools only, to check invariants over all the data in the database.
504 : DebugCheck func(*DB) error
505 :
506 : // Disable the write-ahead log (WAL). Disabling the write-ahead log prohibits
507 : // crash recovery, but can improve performance if crash recovery is not
508 : // needed (e.g. when only temporary state is being stored in the database).
509 : //
510 : // TODO(peter): untested
511 : DisableWAL bool
512 :
513 : // ErrorIfExists causes an error on Open if the database already exists.
514 : // The error can be checked with errors.Is(err, ErrDBAlreadyExists).
515 : //
516 : // The default value is false.
517 : ErrorIfExists bool
518 :
519 : // ErrorIfNotExists causes an error on Open if the database does not already
520 : // exist. The error can be checked with errors.Is(err, ErrDBDoesNotExist).
521 : //
522 : // The default value is false which will cause a database to be created if it
523 : // does not already exist.
524 : ErrorIfNotExists bool
525 :
526 : // ErrorIfNotPristine causes an error on Open if the database already exists
527 : // and any operations have been performed on the database. The error can be
528 : // checked with errors.Is(err, ErrDBNotPristine).
529 : //
530 : // Note that a database that contained keys that were all subsequently deleted
531 : // may or may not trigger the error. Currently, we check if there are any live
532 : // SSTs or log records to replay.
533 : ErrorIfNotPristine bool
534 :
535 : // EventListener provides hooks to listening to significant DB events such as
536 : // flushes, compactions, and table deletion.
537 : EventListener *EventListener
538 :
539 : // Experimental contains experimental options which are off by default.
540 : // These options are temporary and will eventually either be deleted, moved
541 : // out of the experimental group, or made the non-adjustable default. These
542 : // options may change at any time, so do not rely on them.
543 : Experimental struct {
544 : // The threshold of L0 read-amplification at which compaction concurrency
545 : // is enabled (if CompactionDebtConcurrency was not already exceeded).
546 : // Every multiple of this value enables another concurrent
547 : // compaction up to MaxConcurrentCompactions.
548 : L0CompactionConcurrency int
549 :
550 : // CompactionDebtConcurrency controls the threshold of compaction debt
551 : // at which additional compaction concurrency slots are added. For every
552 : // multiple of this value in compaction debt bytes, an additional
553 : // concurrent compaction is added. This works "on top" of
554 : // L0CompactionConcurrency, so the higher of the count of compaction
555 : // concurrency slots as determined by the two options is chosen.
556 : CompactionDebtConcurrency uint64
557 :
558 : // IngestSplit, if it returns true, allows for ingest-time splitting of
559 : // existing sstables into two virtual sstables to allow ingestion sstables to
560 : // slot into a lower level than they otherwise would have.
561 : IngestSplit func() bool
562 :
563 : // ReadCompactionRate controls the frequency of read triggered
564 : // compactions by adjusting `AllowedSeeks` in manifest.FileMetadata:
565 : //
566 : // AllowedSeeks = FileSize / ReadCompactionRate
567 : //
568 : // From LevelDB:
569 : // ```
570 : // We arrange to automatically compact this file after
571 : // a certain number of seeks. Let's assume:
572 : // (1) One seek costs 10ms
573 : // (2) Writing or reading 1MB costs 10ms (100MB/s)
574 : // (3) A compaction of 1MB does 25MB of IO:
575 : // 1MB read from this level
576 : // 10-12MB read from next level (boundaries may be misaligned)
577 : // 10-12MB written to next level
578 : // This implies that 25 seeks cost the same as the compaction
579 : // of 1MB of data. I.e., one seek costs approximately the
580 : // same as the compaction of 40KB of data. We are a little
581 : // conservative and allow approximately one seek for every 16KB
582 : // of data before triggering a compaction.
583 : // ```
584 : ReadCompactionRate int64
585 :
586 : // ReadSamplingMultiplier is a multiplier for the readSamplingPeriod in
587 : // iterator.maybeSampleRead() to control the frequency of read sampling
588 : // to trigger a read triggered compaction. A value of -1 prevents sampling
589 : // and disables read triggered compactions. The default is 1 << 4. which
590 : // gets multiplied with a constant of 1 << 16 to yield 1 << 20 (1MB).
591 : ReadSamplingMultiplier int64
592 :
593 : // TableCacheShards is the number of shards per table cache.
594 : // Reducing the value can reduce the number of idle goroutines per DB
595 : // instance which can be useful in scenarios with a lot of DB instances
596 : // and a large number of CPUs, but doing so can lead to higher contention
597 : // in the table cache and reduced performance.
598 : //
599 : // The default value is the number of logical CPUs, which can be
600 : // limited by runtime.GOMAXPROCS.
601 : TableCacheShards int
602 :
603 : // KeyValidationFunc is a function to validate a user key in an SSTable.
604 : //
605 : // Currently, this function is used to validate the smallest and largest
606 : // keys in an SSTable undergoing compaction. In this case, returning an
607 : // error from the validation function will result in a panic at runtime,
608 : // given that there is rarely any way of recovering from malformed keys
609 : // present in compacted files. By default, validation is not performed.
610 : //
611 : // Additional use-cases may be added in the future.
612 : //
613 : // NOTE: callers should take care to not mutate the key being validated.
614 : KeyValidationFunc func(userKey []byte) error
615 :
616 : // ValidateOnIngest schedules validation of sstables after they have
617 : // been ingested.
618 : //
619 : // By default, this value is false.
620 : ValidateOnIngest bool
621 :
622 : // LevelMultiplier configures the size multiplier used to determine the
623 : // desired size of each level of the LSM. Defaults to 10.
624 : LevelMultiplier int
625 :
626 : // MultiLevelCompactionHeuristic determines whether to add an additional
627 : // level to a conventional two level compaction. If nil, a multilevel
628 : // compaction will never get triggered.
629 : MultiLevelCompactionHeuristic MultiLevelHeuristic
630 :
631 : // MaxWriterConcurrency is used to indicate the maximum number of
632 : // compression workers the compression queue is allowed to use. If
633 : // MaxWriterConcurrency > 0, then the Writer will use parallelism, to
634 : // compress and write blocks to disk. Otherwise, the writer will
635 : // compress and write blocks to disk synchronously.
636 : MaxWriterConcurrency int
637 :
638 : // ForceWriterParallelism is used to force parallelism in the sstable
639 : // Writer for the metamorphic tests. Even with the MaxWriterConcurrency
640 : // option set, we only enable parallelism in the sstable Writer if there
641 : // is enough CPU available, and this option bypasses that.
642 : ForceWriterParallelism bool
643 :
644 : // CPUWorkPermissionGranter should be set if Pebble should be given the
645 : // ability to optionally schedule additional CPU. See the documentation
646 : // for CPUWorkPermissionGranter for more details.
647 : CPUWorkPermissionGranter CPUWorkPermissionGranter
648 :
649 : // EnableValueBlocks is used to decide whether to enable writing
650 : // TableFormatPebblev3 sstables. This setting is only respected by a
651 : // specific subset of format major versions: FormatSSTableValueBlocks,
652 : // FormatFlushableIngest and FormatPrePebblev1MarkedCompacted. In lower
653 : // format major versions, value blocks are never enabled. In higher
654 : // format major versions, value blocks are always enabled.
655 : EnableValueBlocks func() bool
656 :
657 : // ShortAttributeExtractor is used iff EnableValueBlocks() returns true
658 : // (else ignored). If non-nil, a ShortAttribute can be extracted from the
659 : // value and stored with the key, when the value is stored elsewhere.
660 : ShortAttributeExtractor ShortAttributeExtractor
661 :
662 : // RequiredInPlaceValueBound specifies an optional span of user key
663 : // prefixes that are not-MVCC, but have a suffix. For these the values
664 : // must be stored with the key, since the concept of "older versions" is
665 : // not defined. It is also useful for statically known exclusions to value
666 : // separation. In CockroachDB, this will be used for the lock table key
667 : // space that has non-empty suffixes, but those locks don't represent
668 : // actual MVCC versions (the suffix ordering is arbitrary). We will also
669 : // need to add support for dynamically configured exclusions (we want the
670 : // default to be to allow Pebble to decide whether to separate the value
671 : // or not, hence this is structured as exclusions), for example, for users
672 : // of CockroachDB to dynamically exclude certain tables.
673 : //
674 : // Any change in exclusion behavior takes effect only on future written
675 : // sstables, and does not start rewriting existing sstables.
676 : //
677 : // Even ignoring changes in this setting, exclusions are interpreted as a
678 : // guidance by Pebble, and not necessarily honored. Specifically, user
679 : // keys with multiple Pebble-versions *may* have the older versions stored
680 : // in value blocks.
681 : RequiredInPlaceValueBound UserKeyPrefixBound
682 :
683 : // DisableIngestAsFlushable disables lazy ingestion of sstables through
684 : // a WAL write and memtable rotation. Only effectual if the the format
685 : // major version is at least `FormatFlushableIngest`.
686 : DisableIngestAsFlushable func() bool
687 :
688 : // RemoteStorage enables use of remote storage (e.g. S3) for storing
689 : // sstables. Setting this option enables use of CreateOnShared option and
690 : // allows ingestion of external files.
691 : RemoteStorage remote.StorageFactory
692 :
693 : // If CreateOnShared is non-zero, new sstables are created on remote storage
694 : // (using CreateOnSharedLocator and with the appropriate
695 : // CreateOnSharedStrategy). These sstables can be shared between different
696 : // Pebble instances; the lifecycle of such objects is managed by the
697 : // remote.Storage constructed by options.RemoteStorage.
698 : //
699 : // Can only be used when RemoteStorage is set (and recognizes
700 : // CreateOnSharedLocator).
701 : CreateOnShared remote.CreateOnSharedStrategy
702 : CreateOnSharedLocator remote.Locator
703 :
704 : // CacheSizeBytesBytes is the size of the on-disk block cache for objects
705 : // on shared storage in bytes. If it is 0, no cache is used.
706 : SecondaryCacheSizeBytes int64
707 :
708 : // NB: DO NOT crash on SingleDeleteInvariantViolationCallback or
709 : // IneffectualSingleDeleteCallback, since these can be false positives
710 : // even if SingleDel has been used correctly.
711 : //
712 : // Pebble's delete-only compactions can cause a recent RANGEDEL to peek
713 : // below an older SINGLEDEL and delete an arbitrary subset of data below
714 : // that SINGLEDEL. When that SINGLEDEL gets compacted (without the
715 : // RANGEDEL), any of these callbacks can happen, without it being a real
716 : // correctness problem.
717 : //
718 : // Example 1:
719 : // RANGEDEL [a, c)#10 in L0
720 : // SINGLEDEL b#5 in L1
721 : // SET b#3 in L6
722 : //
723 : // If the L6 file containing the SET is narrow and the L1 file containing
724 : // the SINGLEDEL is wide, a delete-only compaction can remove the file in
725 : // L2 before the SINGLEDEL is compacted down. Then when the SINGLEDEL is
726 : // compacted down, it will not find any SET to delete, resulting in the
727 : // ineffectual callback.
728 : //
729 : // Example 2:
730 : // RANGEDEL [a, z)#60 in L0
731 : // SINGLEDEL g#50 in L1
732 : // SET g#40 in L2
733 : // RANGEDEL [g,h)#30 in L3
734 : // SET g#20 in L6
735 : //
736 : // In this example, the two SETs represent the same user write, and the
737 : // RANGEDELs are caused by the CockroachDB range being dropped. That is,
738 : // the user wrote to g once, range was dropped, then added back, which
739 : // caused the SET again, then at some point g was validly deleted using a
740 : // SINGLEDEL, and then the range was dropped again. The older RANGEDEL can
741 : // get fragmented due to compactions it has been part of. Say this L3 file
742 : // containing the RANGEDEL is very narrow, while the L1, L2, L6 files are
743 : // wider than the RANGEDEL in L0. Then the RANGEDEL in L3 can be dropped
744 : // using a delete-only compaction, resulting in an LSM with state:
745 : //
746 : // RANGEDEL [a, z)#60 in L0
747 : // SINGLEDEL g#50 in L1
748 : // SET g#40 in L2
749 : // SET g#20 in L6
750 : //
751 : // A multi-level compaction involving L1, L2, L6 will cause the invariant
752 : // violation callback. This example doesn't need multi-level compactions:
753 : // say there was a Pebble snapshot at g#21 preventing g#20 from being
754 : // dropped when it meets g#40 in a compaction. That snapshot will not save
755 : // RANGEDEL [g,h)#30, so we can have:
756 : //
757 : // SINGLEDEL g#50 in L1
758 : // SET g#40, SET g#20 in L6
759 : //
760 : // And say the snapshot is removed and then the L1 and L6 compaction
761 : // happens, resulting in the invariant violation callback.
762 : //
763 : // TODO(sumeer): rename SingleDeleteInvariantViolationCallback to remove
764 : // the word "invariant".
765 :
766 : // IneffectualPointDeleteCallback is called in compactions/flushes if any
767 : // single delete is being elided without deleting a point set/merge.
768 : IneffectualSingleDeleteCallback func(userKey []byte)
769 :
770 : // SingleDeleteInvariantViolationCallback is called in compactions/flushes if any
771 : // single delete has consumed a Set/Merge, and there is another immediately older
772 : // Set/SetWithDelete/Merge. The user of Pebble has violated the invariant under
773 : // which SingleDelete can be used correctly.
774 : //
775 : // Consider the sequence SingleDelete#3, Set#2, Set#1. There are three
776 : // ways some of these keys can first meet in a compaction.
777 : //
778 : // - All 3 keys in the same compaction: this callback will detect the
779 : // violation.
780 : //
781 : // - SingleDelete#3, Set#2 meet in a compaction first: Both keys will
782 : // disappear. The violation will not be detected, and the DB will have
783 : // Set#1 which is likely incorrect (from the user's perspective).
784 : //
785 : // - Set#2, Set#1 meet in a compaction first: The output will be Set#2,
786 : // which will later be consumed by SingleDelete#3. The violation will
787 : // not be detected and the DB will be correct.
788 : SingleDeleteInvariantViolationCallback func(userKey []byte)
789 : }
790 :
791 : // Filters is a map from filter policy name to filter policy. It is used for
792 : // debugging tools which may be used on multiple databases configured with
793 : // different filter policies. It is not necessary to populate this filters
794 : // map during normal usage of a DB.
795 : Filters map[string]FilterPolicy
796 :
797 : // FlushDelayDeleteRange configures how long the database should wait before
798 : // forcing a flush of a memtable that contains a range deletion. Disk space
799 : // cannot be reclaimed until the range deletion is flushed. No automatic
800 : // flush occurs if zero.
801 : FlushDelayDeleteRange time.Duration
802 :
803 : // FlushDelayRangeKey configures how long the database should wait before
804 : // forcing a flush of a memtable that contains a range key. Range keys in
805 : // the memtable prevent lazy combined iteration, so it's desirable to flush
806 : // range keys promptly. No automatic flush occurs if zero.
807 : FlushDelayRangeKey time.Duration
808 :
809 : // FlushSplitBytes denotes the target number of bytes per sublevel in
810 : // each flush split interval (i.e. range between two flush split keys)
811 : // in L0 sstables. When set to zero, only a single sstable is generated
812 : // by each flush. When set to a non-zero value, flushes are split at
813 : // points to meet L0's TargetFileSize, any grandparent-related overlap
814 : // options, and at boundary keys of L0 flush split intervals (which are
815 : // targeted to contain around FlushSplitBytes bytes in each sublevel
816 : // between pairs of boundary keys). Splitting sstables during flush
817 : // allows increased compaction flexibility and concurrency when those
818 : // tables are compacted to lower levels.
819 : FlushSplitBytes int64
820 :
821 : // FormatMajorVersion sets the format of on-disk files. It is
822 : // recommended to set the format major version to an explicit
823 : // version, as the default may change over time.
824 : //
825 : // At Open if the existing database is formatted using a later
826 : // format major version that is known to this version of Pebble,
827 : // Pebble will continue to use the later format major version. If
828 : // the existing database's version is unknown, the caller may use
829 : // FormatMostCompatible and will be able to open the database
830 : // regardless of its actual version.
831 : //
832 : // If the existing database is formatted using a format major
833 : // version earlier than the one specified, Open will automatically
834 : // ratchet the database to the specified format major version.
835 : FormatMajorVersion FormatMajorVersion
836 :
837 : // FS provides the interface for persistent file storage.
838 : //
839 : // The default value uses the underlying operating system's file system.
840 : FS vfs.FS
841 :
842 : // Lock, if set, must be a database lock acquired through LockDirectory for
843 : // the same directory passed to Open. If provided, Open will skip locking
844 : // the directory. Closing the database will not release the lock, and it's
845 : // the responsibility of the caller to release the lock after closing the
846 : // database.
847 : //
848 : // Open will enforce that the Lock passed locks the same directory passed to
849 : // Open. Concurrent calls to Open using the same Lock are detected and
850 : // prohibited.
851 : Lock *Lock
852 :
853 : // The count of L0 files necessary to trigger an L0 compaction.
854 : L0CompactionFileThreshold int
855 :
856 : // The amount of L0 read-amplification necessary to trigger an L0 compaction.
857 : L0CompactionThreshold int
858 :
859 : // Hard limit on L0 read-amplification, computed as the number of L0
860 : // sublevels. Writes are stopped when this threshold is reached.
861 : L0StopWritesThreshold int
862 :
863 : // The maximum number of bytes for LBase. The base level is the level which
864 : // L0 is compacted into. The base level is determined dynamically based on
865 : // the existing data in the LSM. The maximum number of bytes for other levels
866 : // is computed dynamically based on the base level's maximum size. When the
867 : // maximum number of bytes for a level is exceeded, compaction is requested.
868 : LBaseMaxBytes int64
869 :
870 : // Per-level options. Options for at least one level must be specified. The
871 : // options for the last level are used for all subsequent levels.
872 : Levels []LevelOptions
873 :
874 : // LoggerAndTracer will be used, if non-nil, else Logger will be used and
875 : // tracing will be a noop.
876 :
877 : // Logger used to write log messages.
878 : //
879 : // The default logger uses the Go standard library log package.
880 : Logger Logger
881 : // LoggerAndTracer is used for writing log messages and traces.
882 : LoggerAndTracer LoggerAndTracer
883 :
884 : // MaxManifestFileSize is the maximum size the MANIFEST file is allowed to
885 : // become. When the MANIFEST exceeds this size it is rolled over and a new
886 : // MANIFEST is created.
887 : MaxManifestFileSize int64
888 :
889 : // MaxOpenFiles is a soft limit on the number of open files that can be
890 : // used by the DB.
891 : //
892 : // The default value is 1000.
893 : MaxOpenFiles int
894 :
895 : // The size of a MemTable in steady state. The actual MemTable size starts at
896 : // min(256KB, MemTableSize) and doubles for each subsequent MemTable up to
897 : // MemTableSize. This reduces the memory pressure caused by MemTables for
898 : // short lived (test) DB instances. Note that more than one MemTable can be
899 : // in existence since flushing a MemTable involves creating a new one and
900 : // writing the contents of the old one in the
901 : // background. MemTableStopWritesThreshold places a hard limit on the size of
902 : // the queued MemTables.
903 : //
904 : // The default value is 4MB.
905 : MemTableSize uint64
906 :
907 : // Hard limit on the number of queued of MemTables. Writes are stopped when
908 : // the sum of the queued memtable sizes exceeds:
909 : // MemTableStopWritesThreshold * MemTableSize.
910 : //
911 : // This value should be at least 2 or writes will stop whenever a MemTable is
912 : // being flushed.
913 : //
914 : // The default value is 2.
915 : MemTableStopWritesThreshold int
916 :
917 : // Merger defines the associative merge operation to use for merging values
918 : // written with {Batch,DB}.Merge.
919 : //
920 : // The default merger concatenates values.
921 : Merger *Merger
922 :
923 : // MaxConcurrentCompactions specifies the maximum number of concurrent
924 : // compactions. The default is 1. Concurrent compactions are performed
925 : // - when L0 read-amplification passes the L0CompactionConcurrency threshold
926 : // - for automatic background compactions
927 : // - when a manual compaction for a level is split and parallelized
928 : // MaxConcurrentCompactions must be greater than 0.
929 : MaxConcurrentCompactions func() int
930 :
931 : // DisableAutomaticCompactions dictates whether automatic compactions are
932 : // scheduled or not. The default is false (enabled). This option is only used
933 : // externally when running a manual compaction, and internally for tests.
934 : DisableAutomaticCompactions bool
935 :
936 : // DisableTableStats dictates whether tables should be loaded asynchronously
937 : // to compute statistics that inform compaction heuristics. The collection
938 : // of table stats improves compaction of tombstones, reclaiming disk space
939 : // more quickly and in some cases reducing write amplification in the
940 : // presence of tombstones. Disabling table stats may be useful in tests
941 : // that require determinism as the asynchronicity of table stats collection
942 : // introduces significant nondeterminism.
943 : DisableTableStats bool
944 :
945 : // NoSyncOnClose decides whether the Pebble instance will enforce a
946 : // close-time synchronization (e.g., fdatasync() or sync_file_range())
947 : // on files it writes to. Setting this to true removes the guarantee for a
948 : // sync on close. Some implementations can still issue a non-blocking sync.
949 : NoSyncOnClose bool
950 :
951 : // NumPrevManifest is the number of non-current or older manifests which
952 : // we want to keep around for debugging purposes. By default, we're going
953 : // to keep one older manifest.
954 : NumPrevManifest int
955 :
956 : // ReadOnly indicates that the DB should be opened in read-only mode. Writes
957 : // to the DB will return an error, background compactions are disabled, and
958 : // the flush that normally occurs after replaying the WAL at startup is
959 : // disabled.
960 : ReadOnly bool
961 :
962 : // TableCache is an initialized TableCache which should be set as an
963 : // option if the DB needs to be initialized with a pre-existing table cache.
964 : // If TableCache is nil, then a table cache which is unique to the DB instance
965 : // is created. TableCache can be shared between db instances by setting it here.
966 : // The TableCache set here must use the same underlying cache as Options.Cache
967 : // and pebble will panic otherwise.
968 : TableCache *TableCache
969 :
970 : // TablePropertyCollectors is a list of TablePropertyCollector creation
971 : // functions. A new TablePropertyCollector is created for each sstable built
972 : // and lives for the lifetime of the table.
973 : TablePropertyCollectors []func() TablePropertyCollector
974 :
975 : // BlockPropertyCollectors is a list of BlockPropertyCollector creation
976 : // functions. A new BlockPropertyCollector is created for each sstable
977 : // built and lives for the lifetime of writing that table.
978 : BlockPropertyCollectors []func() BlockPropertyCollector
979 :
980 : // WALBytesPerSync sets the number of bytes to write to a WAL before calling
981 : // Sync on it in the background. Just like with BytesPerSync above, this
982 : // helps smooth out disk write latencies, and avoids cases where the OS
983 : // writes a lot of buffered data to disk at once. However, this is less
984 : // necessary with WALs, as many write operations already pass in
985 : // Sync = true.
986 : //
987 : // The default value is 0, i.e. no background syncing. This matches the
988 : // default behaviour in RocksDB.
989 : WALBytesPerSync int
990 :
991 : // WALDir specifies the directory to store write-ahead logs (WALs) in. If
992 : // empty (the default), WALs will be stored in the same directory as sstables
993 : // (i.e. the directory passed to pebble.Open).
994 : WALDir string
995 :
996 : // WALMinSyncInterval is the minimum duration between syncs of the WAL. If
997 : // WAL syncs are requested faster than this interval, they will be
998 : // artificially delayed. Introducing a small artificial delay (500us) between
999 : // WAL syncs can allow more operations to arrive and reduce IO operations
1000 : // while having a minimal impact on throughput. This option is supplied as a
1001 : // closure in order to allow the value to be changed dynamically. The default
1002 : // value is 0.
1003 : //
1004 : // TODO(peter): rather than a closure, should there be another mechanism for
1005 : // changing options dynamically?
1006 : WALMinSyncInterval func() time.Duration
1007 :
1008 : // TargetByteDeletionRate is the rate (in bytes per second) at which sstable file
1009 : // deletions are limited to (under normal circumstances).
1010 : //
1011 : // Deletion pacing is used to slow down deletions when compactions finish up
1012 : // or readers close and newly-obsolete files need cleaning up. Deleting lots
1013 : // of files at once can cause disk latency to go up on some SSDs, which this
1014 : // functionality guards against.
1015 : //
1016 : // This value is only a best-effort target; the effective rate can be
1017 : // higher if deletions are falling behind or disk space is running low.
1018 : //
1019 : // Setting this to 0 disables deletion pacing, which is also the default.
1020 : TargetByteDeletionRate int
1021 :
1022 : // private options are only used by internal tests or are used internally
1023 : // for facilitating upgrade paths of unconfigurable functionality.
1024 : private struct {
1025 : // strictWALTail configures whether or not a database's WALs created
1026 : // prior to the most recent one should be interpreted strictly,
1027 : // requiring a clean EOF. RocksDB 6.2.1 and the version of Pebble
1028 : // included in CockroachDB 20.1 do not guarantee that closed WALs end
1029 : // cleanly. If this option is set within an OPTIONS file, Pebble
1030 : // interprets previous WALs strictly, requiring a clean EOF.
1031 : // Otherwise, it interprets them permissively in the same manner as
1032 : // RocksDB 6.2.1.
1033 : strictWALTail bool
1034 :
1035 : // disableDeleteOnlyCompactions prevents the scheduling of delete-only
1036 : // compactions that drop sstables wholy covered by range tombstones or
1037 : // range key tombstones.
1038 : disableDeleteOnlyCompactions bool
1039 :
1040 : // disableElisionOnlyCompactions prevents the scheduling of elision-only
1041 : // compactions that rewrite sstables in place in order to elide obsolete
1042 : // keys.
1043 : disableElisionOnlyCompactions bool
1044 :
1045 : // disableLazyCombinedIteration is a private option used by the
1046 : // metamorphic tests to test equivalence between lazy-combined iteration
1047 : // and constructing the range-key iterator upfront. It's a private
1048 : // option to avoid littering the public interface with options that we
1049 : // do not want to allow users to actually configure.
1050 : disableLazyCombinedIteration bool
1051 :
1052 : // testingAlwaysWaitForCleanup is set by some tests to force waiting for
1053 : // obsolete file deletion (to make events deterministic).
1054 : testingAlwaysWaitForCleanup bool
1055 :
1056 : // fsCloser holds a closer that should be invoked after a DB using these
1057 : // Options is closed. This is used to automatically stop the
1058 : // long-running goroutine associated with the disk-health-checking FS.
1059 : // See the initialization of FS in EnsureDefaults. Note that care has
1060 : // been taken to ensure that it is still safe to continue using the FS
1061 : // after this closer has been invoked. However, if write operations
1062 : // against the FS are made after the DB is closed, the FS may leak a
1063 : // goroutine indefinitely.
1064 : fsCloser io.Closer
1065 :
1066 : // efosAlwaysCreatesIterators is set by some tests to force
1067 : // EventuallyFileOnlySnapshots to always create iterators, even after a
1068 : // conflicting excise.
1069 : efosAlwaysCreatesIterators bool
1070 : }
1071 : }
1072 :
1073 : // DebugCheckLevels calls CheckLevels on the provided database.
1074 : // It may be set in the DebugCheck field of Options to check
1075 : // level invariants whenever a new version is installed.
1076 1 : func DebugCheckLevels(db *DB) error {
1077 1 : return db.CheckLevels(nil)
1078 1 : }
1079 :
1080 : // EnsureDefaults ensures that the default values for all options are set if a
1081 : // valid value was not already specified. Returns the new options.
1082 1 : func (o *Options) EnsureDefaults() *Options {
1083 1 : if o == nil {
1084 0 : o = &Options{}
1085 0 : }
1086 1 : if o.BytesPerSync <= 0 {
1087 0 : o.BytesPerSync = 512 << 10 // 512 KB
1088 0 : }
1089 1 : if o.Cleaner == nil {
1090 0 : o.Cleaner = DeleteCleaner{}
1091 0 : }
1092 1 : if o.Comparer == nil {
1093 0 : o.Comparer = DefaultComparer
1094 0 : }
1095 1 : if o.Experimental.DisableIngestAsFlushable == nil {
1096 1 : o.Experimental.DisableIngestAsFlushable = func() bool { return false }
1097 : }
1098 1 : if o.Experimental.L0CompactionConcurrency <= 0 {
1099 0 : o.Experimental.L0CompactionConcurrency = 10
1100 0 : }
1101 1 : if o.Experimental.CompactionDebtConcurrency <= 0 {
1102 0 : o.Experimental.CompactionDebtConcurrency = 1 << 30 // 1 GB
1103 0 : }
1104 1 : if o.Experimental.KeyValidationFunc == nil {
1105 1 : o.Experimental.KeyValidationFunc = func([]byte) error { return nil }
1106 : }
1107 1 : if o.L0CompactionThreshold <= 0 {
1108 0 : o.L0CompactionThreshold = 4
1109 0 : }
1110 1 : if o.L0CompactionFileThreshold <= 0 {
1111 0 : // Some justification for the default of 500:
1112 0 : // Why not smaller?:
1113 0 : // - The default target file size for L0 is 2MB, so 500 files is <= 1GB
1114 0 : // of data. At observed compaction speeds of > 20MB/s, L0 can be
1115 0 : // cleared of all files in < 1min, so this backlog is not huge.
1116 0 : // - 500 files is low overhead for instantiating L0 sublevels from
1117 0 : // scratch.
1118 0 : // - Lower values were observed to cause excessive and inefficient
1119 0 : // compactions out of L0 in a TPCC import benchmark.
1120 0 : // Why not larger?:
1121 0 : // - More than 1min to compact everything out of L0.
1122 0 : // - CockroachDB's admission control system uses a threshold of 1000
1123 0 : // files to start throttling writes to Pebble. Using 500 here gives
1124 0 : // us headroom between when Pebble should start compacting L0 and
1125 0 : // when the admission control threshold is reached.
1126 0 : //
1127 0 : // We can revisit this default in the future based on better
1128 0 : // experimental understanding.
1129 0 : //
1130 0 : // TODO(jackson): Experiment with slightly lower thresholds [or higher
1131 0 : // admission control thresholds] to see whether a higher L0 score at the
1132 0 : // threshold (currently 2.0) is necessary for some workloads to avoid
1133 0 : // starving L0 in favor of lower-level compactions.
1134 0 : o.L0CompactionFileThreshold = 500
1135 0 : }
1136 1 : if o.L0StopWritesThreshold <= 0 {
1137 0 : o.L0StopWritesThreshold = 12
1138 0 : }
1139 1 : if o.LBaseMaxBytes <= 0 {
1140 0 : o.LBaseMaxBytes = 64 << 20 // 64 MB
1141 0 : }
1142 1 : if o.Levels == nil {
1143 0 : o.Levels = make([]LevelOptions, 1)
1144 0 : for i := range o.Levels {
1145 0 : if i > 0 {
1146 0 : l := &o.Levels[i]
1147 0 : if l.TargetFileSize <= 0 {
1148 0 : l.TargetFileSize = o.Levels[i-1].TargetFileSize * 2
1149 0 : }
1150 : }
1151 0 : o.Levels[i].EnsureDefaults()
1152 : }
1153 1 : } else {
1154 1 : for i := range o.Levels {
1155 1 : o.Levels[i].EnsureDefaults()
1156 1 : }
1157 : }
1158 1 : if o.Logger == nil {
1159 1 : o.Logger = DefaultLogger
1160 1 : }
1161 1 : if o.EventListener == nil {
1162 1 : o.EventListener = &EventListener{}
1163 1 : }
1164 1 : o.EventListener.EnsureDefaults(o.Logger)
1165 1 : if o.MaxManifestFileSize == 0 {
1166 0 : o.MaxManifestFileSize = 128 << 20 // 128 MB
1167 0 : }
1168 1 : if o.MaxOpenFiles == 0 {
1169 0 : o.MaxOpenFiles = 1000
1170 0 : }
1171 1 : if o.MemTableSize <= 0 {
1172 0 : o.MemTableSize = 4 << 20 // 4 MB
1173 0 : }
1174 1 : if o.MemTableStopWritesThreshold <= 0 {
1175 0 : o.MemTableStopWritesThreshold = 2
1176 0 : }
1177 1 : if o.Merger == nil {
1178 0 : o.Merger = DefaultMerger
1179 0 : }
1180 1 : o.private.strictWALTail = true
1181 1 : if o.MaxConcurrentCompactions == nil {
1182 0 : o.MaxConcurrentCompactions = func() int { return 1 }
1183 : }
1184 1 : if o.NumPrevManifest <= 0 {
1185 1 : o.NumPrevManifest = 1
1186 1 : }
1187 :
1188 1 : if o.FormatMajorVersion == FormatDefault {
1189 0 : o.FormatMajorVersion = FormatMinSupported
1190 0 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone {
1191 0 : o.FormatMajorVersion = FormatMinForSharedObjects
1192 0 : }
1193 : }
1194 :
1195 1 : if o.FS == nil {
1196 0 : o.WithFSDefaults()
1197 0 : }
1198 1 : if o.FlushSplitBytes <= 0 {
1199 0 : o.FlushSplitBytes = 2 * o.Levels[0].TargetFileSize
1200 0 : }
1201 1 : if o.Experimental.LevelMultiplier <= 0 {
1202 1 : o.Experimental.LevelMultiplier = defaultLevelMultiplier
1203 1 : }
1204 1 : if o.Experimental.ReadCompactionRate == 0 {
1205 0 : o.Experimental.ReadCompactionRate = 16000
1206 0 : }
1207 1 : if o.Experimental.ReadSamplingMultiplier == 0 {
1208 0 : o.Experimental.ReadSamplingMultiplier = 1 << 4
1209 0 : }
1210 1 : if o.Experimental.TableCacheShards <= 0 {
1211 0 : o.Experimental.TableCacheShards = runtime.GOMAXPROCS(0)
1212 0 : }
1213 1 : if o.Experimental.CPUWorkPermissionGranter == nil {
1214 1 : o.Experimental.CPUWorkPermissionGranter = defaultCPUWorkGranter{}
1215 1 : }
1216 1 : if o.Experimental.MultiLevelCompactionHeuristic == nil {
1217 0 : o.Experimental.MultiLevelCompactionHeuristic = WriteAmpHeuristic{}
1218 0 : }
1219 :
1220 1 : o.initMaps()
1221 1 : return o
1222 : }
1223 :
1224 : // WithFSDefaults configures the Options to wrap the configured filesystem with
1225 : // the default virtual file system middleware, like disk-health checking.
1226 1 : func (o *Options) WithFSDefaults() *Options {
1227 1 : if o.FS == nil {
1228 0 : o.FS = vfs.Default
1229 0 : }
1230 1 : o.FS, o.private.fsCloser = vfs.WithDiskHealthChecks(o.FS, 5*time.Second,
1231 1 : func(info vfs.DiskSlowInfo) {
1232 0 : o.EventListener.DiskSlow(info)
1233 0 : })
1234 1 : return o
1235 : }
1236 :
1237 : // AddEventListener adds the provided event listener to the Options, in addition
1238 : // to any existing event listener.
1239 0 : func (o *Options) AddEventListener(l EventListener) {
1240 0 : if o.EventListener != nil {
1241 0 : l = TeeEventListener(l, *o.EventListener)
1242 0 : }
1243 0 : o.EventListener = &l
1244 : }
1245 :
1246 : // TestingAlwaysCreateEFOSIterators is used to toggle a private option for
1247 : // having EventuallyFileOnlySnapshots always create iterators. Meant to only
1248 : // be used in tests.
1249 1 : func (o *Options) TestingAlwaysCreateEFOSIterators(value bool) {
1250 1 : o.private.efosAlwaysCreatesIterators = value
1251 1 : }
1252 :
1253 1 : func (o *Options) equal() Equal {
1254 1 : if o.Comparer.Equal == nil {
1255 0 : return bytes.Equal
1256 0 : }
1257 1 : return o.Comparer.Equal
1258 : }
1259 :
1260 : // initMaps initializes the Comparers, Filters, and Mergers maps.
1261 1 : func (o *Options) initMaps() {
1262 1 : for i := range o.Levels {
1263 1 : l := &o.Levels[i]
1264 1 : if l.FilterPolicy != nil {
1265 1 : if o.Filters == nil {
1266 1 : o.Filters = make(map[string]FilterPolicy)
1267 1 : }
1268 1 : name := l.FilterPolicy.Name()
1269 1 : if _, ok := o.Filters[name]; !ok {
1270 1 : o.Filters[name] = l.FilterPolicy
1271 1 : }
1272 : }
1273 : }
1274 : }
1275 :
1276 : // Level returns the LevelOptions for the specified level.
1277 1 : func (o *Options) Level(level int) LevelOptions {
1278 1 : if level < len(o.Levels) {
1279 1 : return o.Levels[level]
1280 1 : }
1281 1 : n := len(o.Levels) - 1
1282 1 : l := o.Levels[n]
1283 1 : for i := n; i < level; i++ {
1284 1 : l.TargetFileSize *= 2
1285 1 : }
1286 1 : return l
1287 : }
1288 :
1289 : // Clone creates a shallow-copy of the supplied options.
1290 1 : func (o *Options) Clone() *Options {
1291 1 : n := &Options{}
1292 1 : if o != nil {
1293 1 : *n = *o
1294 1 : }
1295 1 : return n
1296 : }
1297 :
1298 1 : func filterPolicyName(p FilterPolicy) string {
1299 1 : if p == nil {
1300 1 : return "none"
1301 1 : }
1302 1 : return p.Name()
1303 : }
1304 :
1305 1 : func (o *Options) String() string {
1306 1 : var buf bytes.Buffer
1307 1 :
1308 1 : cacheSize := int64(cacheDefaultSize)
1309 1 : if o.Cache != nil {
1310 1 : cacheSize = o.Cache.MaxSize()
1311 1 : }
1312 :
1313 1 : fmt.Fprintf(&buf, "[Version]\n")
1314 1 : fmt.Fprintf(&buf, " pebble_version=0.1\n")
1315 1 : fmt.Fprintf(&buf, "\n")
1316 1 : fmt.Fprintf(&buf, "[Options]\n")
1317 1 : fmt.Fprintf(&buf, " bytes_per_sync=%d\n", o.BytesPerSync)
1318 1 : fmt.Fprintf(&buf, " cache_size=%d\n", cacheSize)
1319 1 : fmt.Fprintf(&buf, " cleaner=%s\n", o.Cleaner)
1320 1 : fmt.Fprintf(&buf, " compaction_debt_concurrency=%d\n", o.Experimental.CompactionDebtConcurrency)
1321 1 : fmt.Fprintf(&buf, " comparer=%s\n", o.Comparer.Name)
1322 1 : fmt.Fprintf(&buf, " disable_wal=%t\n", o.DisableWAL)
1323 1 : if o.Experimental.DisableIngestAsFlushable != nil && o.Experimental.DisableIngestAsFlushable() {
1324 1 : fmt.Fprintf(&buf, " disable_ingest_as_flushable=%t\n", true)
1325 1 : }
1326 1 : fmt.Fprintf(&buf, " flush_delay_delete_range=%s\n", o.FlushDelayDeleteRange)
1327 1 : fmt.Fprintf(&buf, " flush_delay_range_key=%s\n", o.FlushDelayRangeKey)
1328 1 : fmt.Fprintf(&buf, " flush_split_bytes=%d\n", o.FlushSplitBytes)
1329 1 : fmt.Fprintf(&buf, " format_major_version=%d\n", o.FormatMajorVersion)
1330 1 : fmt.Fprintf(&buf, " l0_compaction_concurrency=%d\n", o.Experimental.L0CompactionConcurrency)
1331 1 : fmt.Fprintf(&buf, " l0_compaction_file_threshold=%d\n", o.L0CompactionFileThreshold)
1332 1 : fmt.Fprintf(&buf, " l0_compaction_threshold=%d\n", o.L0CompactionThreshold)
1333 1 : fmt.Fprintf(&buf, " l0_stop_writes_threshold=%d\n", o.L0StopWritesThreshold)
1334 1 : fmt.Fprintf(&buf, " lbase_max_bytes=%d\n", o.LBaseMaxBytes)
1335 1 : if o.Experimental.LevelMultiplier != defaultLevelMultiplier {
1336 1 : fmt.Fprintf(&buf, " level_multiplier=%d\n", o.Experimental.LevelMultiplier)
1337 1 : }
1338 1 : fmt.Fprintf(&buf, " max_concurrent_compactions=%d\n", o.MaxConcurrentCompactions())
1339 1 : fmt.Fprintf(&buf, " max_manifest_file_size=%d\n", o.MaxManifestFileSize)
1340 1 : fmt.Fprintf(&buf, " max_open_files=%d\n", o.MaxOpenFiles)
1341 1 : fmt.Fprintf(&buf, " mem_table_size=%d\n", o.MemTableSize)
1342 1 : fmt.Fprintf(&buf, " mem_table_stop_writes_threshold=%d\n", o.MemTableStopWritesThreshold)
1343 1 : fmt.Fprintf(&buf, " min_deletion_rate=%d\n", o.TargetByteDeletionRate)
1344 1 : fmt.Fprintf(&buf, " merger=%s\n", o.Merger.Name)
1345 1 : if o.Experimental.MultiLevelCompactionHeuristic != nil {
1346 1 : fmt.Fprintf(&buf, " multilevel_compaction_heuristic=%s\n", o.Experimental.MultiLevelCompactionHeuristic.String())
1347 1 : }
1348 1 : fmt.Fprintf(&buf, " read_compaction_rate=%d\n", o.Experimental.ReadCompactionRate)
1349 1 : fmt.Fprintf(&buf, " read_sampling_multiplier=%d\n", o.Experimental.ReadSamplingMultiplier)
1350 1 : fmt.Fprintf(&buf, " strict_wal_tail=%t\n", o.private.strictWALTail)
1351 1 : fmt.Fprintf(&buf, " table_cache_shards=%d\n", o.Experimental.TableCacheShards)
1352 1 : fmt.Fprintf(&buf, " table_property_collectors=[")
1353 1 : for i := range o.TablePropertyCollectors {
1354 0 : if i > 0 {
1355 0 : fmt.Fprintf(&buf, ",")
1356 0 : }
1357 : // NB: This creates a new TablePropertyCollector, but Options.String() is
1358 : // called rarely so the overhead of doing so is not consequential.
1359 0 : fmt.Fprintf(&buf, "%s", o.TablePropertyCollectors[i]().Name())
1360 : }
1361 1 : fmt.Fprintf(&buf, "]\n")
1362 1 : fmt.Fprintf(&buf, " validate_on_ingest=%t\n", o.Experimental.ValidateOnIngest)
1363 1 : fmt.Fprintf(&buf, " wal_dir=%s\n", o.WALDir)
1364 1 : fmt.Fprintf(&buf, " wal_bytes_per_sync=%d\n", o.WALBytesPerSync)
1365 1 : fmt.Fprintf(&buf, " max_writer_concurrency=%d\n", o.Experimental.MaxWriterConcurrency)
1366 1 : fmt.Fprintf(&buf, " force_writer_parallelism=%t\n", o.Experimental.ForceWriterParallelism)
1367 1 : fmt.Fprintf(&buf, " secondary_cache_size_bytes=%d\n", o.Experimental.SecondaryCacheSizeBytes)
1368 1 : fmt.Fprintf(&buf, " create_on_shared=%d\n", o.Experimental.CreateOnShared)
1369 1 :
1370 1 : // Private options.
1371 1 : //
1372 1 : // These options are only encoded if true, because we do not want them to
1373 1 : // appear in production serialized Options files, since they're testing-only
1374 1 : // options. They're only serialized when true, which still ensures that the
1375 1 : // metamorphic tests may propagate them to subprocesses.
1376 1 : if o.private.disableDeleteOnlyCompactions {
1377 1 : fmt.Fprintln(&buf, " disable_delete_only_compactions=true")
1378 1 : }
1379 1 : if o.private.disableElisionOnlyCompactions {
1380 1 : fmt.Fprintln(&buf, " disable_elision_only_compactions=true")
1381 1 : }
1382 1 : if o.private.disableLazyCombinedIteration {
1383 1 : fmt.Fprintln(&buf, " disable_lazy_combined_iteration=true")
1384 1 : }
1385 :
1386 1 : for i := range o.Levels {
1387 1 : l := &o.Levels[i]
1388 1 : fmt.Fprintf(&buf, "\n")
1389 1 : fmt.Fprintf(&buf, "[Level \"%d\"]\n", i)
1390 1 : fmt.Fprintf(&buf, " block_restart_interval=%d\n", l.BlockRestartInterval)
1391 1 : fmt.Fprintf(&buf, " block_size=%d\n", l.BlockSize)
1392 1 : fmt.Fprintf(&buf, " block_size_threshold=%d\n", l.BlockSizeThreshold)
1393 1 : fmt.Fprintf(&buf, " compression=%s\n", l.Compression)
1394 1 : fmt.Fprintf(&buf, " filter_policy=%s\n", filterPolicyName(l.FilterPolicy))
1395 1 : fmt.Fprintf(&buf, " filter_type=%s\n", l.FilterType)
1396 1 : fmt.Fprintf(&buf, " index_block_size=%d\n", l.IndexBlockSize)
1397 1 : fmt.Fprintf(&buf, " target_file_size=%d\n", l.TargetFileSize)
1398 1 : }
1399 :
1400 1 : return buf.String()
1401 : }
1402 :
1403 1 : func parseOptions(s string, fn func(section, key, value string) error) error {
1404 1 : var section string
1405 1 : for _, line := range strings.Split(s, "\n") {
1406 1 : line = strings.TrimSpace(line)
1407 1 : if len(line) == 0 {
1408 1 : // Skip blank lines.
1409 1 : continue
1410 : }
1411 1 : if line[0] == ';' || line[0] == '#' {
1412 0 : // Skip comments.
1413 0 : continue
1414 : }
1415 1 : n := len(line)
1416 1 : if line[0] == '[' && line[n-1] == ']' {
1417 1 : // Parse section.
1418 1 : section = line[1 : n-1]
1419 1 : continue
1420 : }
1421 :
1422 1 : pos := strings.Index(line, "=")
1423 1 : if pos < 0 {
1424 0 : const maxLen = 50
1425 0 : if len(line) > maxLen {
1426 0 : line = line[:maxLen-3] + "..."
1427 0 : }
1428 0 : return base.CorruptionErrorf("invalid key=value syntax: %q", errors.Safe(line))
1429 : }
1430 :
1431 1 : key := strings.TrimSpace(line[:pos])
1432 1 : value := strings.TrimSpace(line[pos+1:])
1433 1 :
1434 1 : // RocksDB uses a similar (INI-style) syntax for the OPTIONS file, but
1435 1 : // different section names and keys. The "CFOptions ..." paths are the
1436 1 : // RocksDB versions which we map to the Pebble paths.
1437 1 : mappedSection := section
1438 1 : if section == `CFOptions "default"` {
1439 0 : mappedSection = "Options"
1440 0 : switch key {
1441 0 : case "comparator":
1442 0 : key = "comparer"
1443 0 : case "merge_operator":
1444 0 : key = "merger"
1445 : }
1446 : }
1447 :
1448 1 : if err := fn(mappedSection, key, value); err != nil {
1449 0 : return err
1450 0 : }
1451 : }
1452 1 : return nil
1453 : }
1454 :
1455 : // ParseHooks contains callbacks to create options fields which can have
1456 : // user-defined implementations.
1457 : type ParseHooks struct {
1458 : NewCache func(size int64) *Cache
1459 : NewCleaner func(name string) (Cleaner, error)
1460 : NewComparer func(name string) (*Comparer, error)
1461 : NewFilterPolicy func(name string) (FilterPolicy, error)
1462 : NewMerger func(name string) (*Merger, error)
1463 : SkipUnknown func(name, value string) bool
1464 : }
1465 :
1466 : // Parse parses the options from the specified string. Note that certain
1467 : // options cannot be parsed into populated fields. For example, comparer and
1468 : // merger.
1469 1 : func (o *Options) Parse(s string, hooks *ParseHooks) error {
1470 1 : return parseOptions(s, func(section, key, value string) error {
1471 1 : // WARNING: DO NOT remove entries from the switches below because doing so
1472 1 : // causes a key previously written to the OPTIONS file to be considered unknown,
1473 1 : // a backwards incompatible change. Instead, leave in support for parsing the
1474 1 : // key but simply don't parse the value.
1475 1 :
1476 1 : switch {
1477 1 : case section == "Version":
1478 1 : switch key {
1479 1 : case "pebble_version":
1480 0 : default:
1481 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1482 0 : return nil
1483 0 : }
1484 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1485 0 : errors.Safe(section), errors.Safe(key))
1486 : }
1487 1 : return nil
1488 :
1489 1 : case section == "Options":
1490 1 : var err error
1491 1 : switch key {
1492 1 : case "bytes_per_sync":
1493 1 : o.BytesPerSync, err = strconv.Atoi(value)
1494 1 : case "cache_size":
1495 1 : var n int64
1496 1 : n, err = strconv.ParseInt(value, 10, 64)
1497 1 : if err == nil && hooks != nil && hooks.NewCache != nil {
1498 1 : if o.Cache != nil {
1499 0 : o.Cache.Unref()
1500 0 : }
1501 1 : o.Cache = hooks.NewCache(n)
1502 : }
1503 : // We avoid calling cache.New in parsing because it makes it
1504 : // too easy to leak a cache.
1505 1 : case "cleaner":
1506 1 : switch value {
1507 1 : case "archive":
1508 1 : o.Cleaner = ArchiveCleaner{}
1509 0 : case "delete":
1510 0 : o.Cleaner = DeleteCleaner{}
1511 0 : default:
1512 0 : if hooks != nil && hooks.NewCleaner != nil {
1513 0 : o.Cleaner, err = hooks.NewCleaner(value)
1514 0 : }
1515 : }
1516 1 : case "comparer":
1517 1 : switch value {
1518 0 : case "leveldb.BytewiseComparator":
1519 0 : o.Comparer = DefaultComparer
1520 1 : default:
1521 1 : if hooks != nil && hooks.NewComparer != nil {
1522 0 : o.Comparer, err = hooks.NewComparer(value)
1523 0 : }
1524 : }
1525 1 : case "compaction_debt_concurrency":
1526 1 : o.Experimental.CompactionDebtConcurrency, err = strconv.ParseUint(value, 10, 64)
1527 0 : case "delete_range_flush_delay":
1528 0 : // NB: This is a deprecated serialization of the
1529 0 : // `flush_delay_delete_range`.
1530 0 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1531 1 : case "disable_delete_only_compactions":
1532 1 : o.private.disableDeleteOnlyCompactions, err = strconv.ParseBool(value)
1533 1 : case "disable_elision_only_compactions":
1534 1 : o.private.disableElisionOnlyCompactions, err = strconv.ParseBool(value)
1535 1 : case "disable_ingest_as_flushable":
1536 1 : var v bool
1537 1 : v, err = strconv.ParseBool(value)
1538 1 : if err == nil {
1539 1 : o.Experimental.DisableIngestAsFlushable = func() bool { return v }
1540 : }
1541 1 : case "disable_lazy_combined_iteration":
1542 1 : o.private.disableLazyCombinedIteration, err = strconv.ParseBool(value)
1543 1 : case "disable_wal":
1544 1 : o.DisableWAL, err = strconv.ParseBool(value)
1545 1 : case "flush_delay_delete_range":
1546 1 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1547 1 : case "flush_delay_range_key":
1548 1 : o.FlushDelayRangeKey, err = time.ParseDuration(value)
1549 1 : case "flush_split_bytes":
1550 1 : o.FlushSplitBytes, err = strconv.ParseInt(value, 10, 64)
1551 1 : case "format_major_version":
1552 1 : // NB: The version written here may be stale. Open does
1553 1 : // not use the format major version encoded in the
1554 1 : // OPTIONS file other than to validate that the encoded
1555 1 : // version is valid right here.
1556 1 : var v uint64
1557 1 : v, err = strconv.ParseUint(value, 10, 64)
1558 1 : if vers := FormatMajorVersion(v); vers > internalFormatNewest || vers == FormatDefault {
1559 0 : err = errors.Newf("unsupported format major version %d", o.FormatMajorVersion)
1560 0 : }
1561 1 : if err == nil {
1562 1 : o.FormatMajorVersion = FormatMajorVersion(v)
1563 1 : }
1564 1 : case "l0_compaction_concurrency":
1565 1 : o.Experimental.L0CompactionConcurrency, err = strconv.Atoi(value)
1566 1 : case "l0_compaction_file_threshold":
1567 1 : o.L0CompactionFileThreshold, err = strconv.Atoi(value)
1568 1 : case "l0_compaction_threshold":
1569 1 : o.L0CompactionThreshold, err = strconv.Atoi(value)
1570 1 : case "l0_stop_writes_threshold":
1571 1 : o.L0StopWritesThreshold, err = strconv.Atoi(value)
1572 0 : case "l0_sublevel_compactions":
1573 : // Do nothing; option existed in older versions of pebble.
1574 1 : case "lbase_max_bytes":
1575 1 : o.LBaseMaxBytes, err = strconv.ParseInt(value, 10, 64)
1576 1 : case "level_multiplier":
1577 1 : o.Experimental.LevelMultiplier, err = strconv.Atoi(value)
1578 1 : case "max_concurrent_compactions":
1579 1 : var concurrentCompactions int
1580 1 : concurrentCompactions, err = strconv.Atoi(value)
1581 1 : if concurrentCompactions <= 0 {
1582 0 : err = errors.New("max_concurrent_compactions cannot be <= 0")
1583 1 : } else {
1584 1 : o.MaxConcurrentCompactions = func() int { return concurrentCompactions }
1585 : }
1586 1 : case "max_manifest_file_size":
1587 1 : o.MaxManifestFileSize, err = strconv.ParseInt(value, 10, 64)
1588 1 : case "max_open_files":
1589 1 : o.MaxOpenFiles, err = strconv.Atoi(value)
1590 1 : case "mem_table_size":
1591 1 : o.MemTableSize, err = strconv.ParseUint(value, 10, 64)
1592 1 : case "mem_table_stop_writes_threshold":
1593 1 : o.MemTableStopWritesThreshold, err = strconv.Atoi(value)
1594 0 : case "min_compaction_rate":
1595 : // Do nothing; option existed in older versions of pebble, and
1596 : // may be meaningful again eventually.
1597 1 : case "min_deletion_rate":
1598 1 : o.TargetByteDeletionRate, err = strconv.Atoi(value)
1599 0 : case "min_flush_rate":
1600 : // Do nothing; option existed in older versions of pebble, and
1601 : // may be meaningful again eventually.
1602 1 : case "multilevel_compaction_heuristic":
1603 1 : switch {
1604 1 : case value == "none":
1605 1 : o.Experimental.MultiLevelCompactionHeuristic = NoMultiLevel{}
1606 1 : case strings.HasPrefix(value, "wamp"):
1607 1 : fields := strings.FieldsFunc(strings.TrimPrefix(value, "wamp"), func(r rune) bool {
1608 1 : return unicode.IsSpace(r) || r == ',' || r == '(' || r == ')'
1609 1 : })
1610 1 : if len(fields) != 2 {
1611 0 : err = errors.Newf("require 2 arguments")
1612 0 : }
1613 1 : var h WriteAmpHeuristic
1614 1 : if err == nil {
1615 1 : h.AddPropensity, err = strconv.ParseFloat(fields[0], 64)
1616 1 : }
1617 1 : if err == nil {
1618 1 : h.AllowL0, err = strconv.ParseBool(fields[1])
1619 1 : }
1620 1 : if err == nil {
1621 1 : o.Experimental.MultiLevelCompactionHeuristic = h
1622 1 : } else {
1623 0 : err = errors.Wrapf(err, "unexpected wamp heuristic arguments: %s", value)
1624 0 : }
1625 0 : default:
1626 0 : err = errors.Newf("unrecognized multilevel compaction heuristic: %s", value)
1627 : }
1628 0 : case "point_tombstone_weight":
1629 : // Do nothing; deprecated.
1630 1 : case "strict_wal_tail":
1631 1 : o.private.strictWALTail, err = strconv.ParseBool(value)
1632 1 : case "merger":
1633 1 : switch value {
1634 0 : case "nullptr":
1635 0 : o.Merger = nil
1636 1 : case "pebble.concatenate":
1637 1 : o.Merger = DefaultMerger
1638 0 : default:
1639 0 : if hooks != nil && hooks.NewMerger != nil {
1640 0 : o.Merger, err = hooks.NewMerger(value)
1641 0 : }
1642 : }
1643 1 : case "read_compaction_rate":
1644 1 : o.Experimental.ReadCompactionRate, err = strconv.ParseInt(value, 10, 64)
1645 1 : case "read_sampling_multiplier":
1646 1 : o.Experimental.ReadSamplingMultiplier, err = strconv.ParseInt(value, 10, 64)
1647 1 : case "table_cache_shards":
1648 1 : o.Experimental.TableCacheShards, err = strconv.Atoi(value)
1649 0 : case "table_format":
1650 0 : switch value {
1651 0 : case "leveldb":
1652 0 : case "rocksdbv2":
1653 0 : default:
1654 0 : return errors.Errorf("pebble: unknown table format: %q", errors.Safe(value))
1655 : }
1656 1 : case "table_property_collectors":
1657 : // TODO(peter): set o.TablePropertyCollectors
1658 1 : case "validate_on_ingest":
1659 1 : o.Experimental.ValidateOnIngest, err = strconv.ParseBool(value)
1660 1 : case "wal_dir":
1661 1 : o.WALDir = value
1662 1 : case "wal_bytes_per_sync":
1663 1 : o.WALBytesPerSync, err = strconv.Atoi(value)
1664 1 : case "max_writer_concurrency":
1665 1 : o.Experimental.MaxWriterConcurrency, err = strconv.Atoi(value)
1666 1 : case "force_writer_parallelism":
1667 1 : o.Experimental.ForceWriterParallelism, err = strconv.ParseBool(value)
1668 1 : case "secondary_cache_size_bytes":
1669 1 : o.Experimental.SecondaryCacheSizeBytes, err = strconv.ParseInt(value, 10, 64)
1670 1 : case "create_on_shared":
1671 1 : var createOnSharedInt int64
1672 1 : createOnSharedInt, err = strconv.ParseInt(value, 10, 64)
1673 1 : o.Experimental.CreateOnShared = remote.CreateOnSharedStrategy(createOnSharedInt)
1674 0 : default:
1675 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1676 0 : return nil
1677 0 : }
1678 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1679 0 : errors.Safe(section), errors.Safe(key))
1680 : }
1681 1 : return err
1682 :
1683 1 : case strings.HasPrefix(section, "Level "):
1684 1 : var index int
1685 1 : if n, err := fmt.Sscanf(section, `Level "%d"`, &index); err != nil {
1686 0 : return err
1687 1 : } else if n != 1 {
1688 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section, value) {
1689 0 : return nil
1690 0 : }
1691 0 : return errors.Errorf("pebble: unknown section: %q", errors.Safe(section))
1692 : }
1693 :
1694 1 : if len(o.Levels) <= index {
1695 0 : newLevels := make([]LevelOptions, index+1)
1696 0 : copy(newLevels, o.Levels)
1697 0 : o.Levels = newLevels
1698 0 : }
1699 1 : l := &o.Levels[index]
1700 1 :
1701 1 : var err error
1702 1 : switch key {
1703 1 : case "block_restart_interval":
1704 1 : l.BlockRestartInterval, err = strconv.Atoi(value)
1705 1 : case "block_size":
1706 1 : l.BlockSize, err = strconv.Atoi(value)
1707 1 : case "block_size_threshold":
1708 1 : l.BlockSizeThreshold, err = strconv.Atoi(value)
1709 1 : case "compression":
1710 1 : switch value {
1711 0 : case "Default":
1712 0 : l.Compression = DefaultCompression
1713 1 : case "NoCompression":
1714 1 : l.Compression = NoCompression
1715 1 : case "Snappy":
1716 1 : l.Compression = SnappyCompression
1717 1 : case "ZSTD":
1718 1 : l.Compression = ZstdCompression
1719 0 : default:
1720 0 : return errors.Errorf("pebble: unknown compression: %q", errors.Safe(value))
1721 : }
1722 1 : case "filter_policy":
1723 1 : if hooks != nil && hooks.NewFilterPolicy != nil {
1724 1 : l.FilterPolicy, err = hooks.NewFilterPolicy(value)
1725 1 : }
1726 1 : case "filter_type":
1727 1 : switch value {
1728 1 : case "table":
1729 1 : l.FilterType = TableFilter
1730 0 : default:
1731 0 : return errors.Errorf("pebble: unknown filter type: %q", errors.Safe(value))
1732 : }
1733 1 : case "index_block_size":
1734 1 : l.IndexBlockSize, err = strconv.Atoi(value)
1735 1 : case "target_file_size":
1736 1 : l.TargetFileSize, err = strconv.ParseInt(value, 10, 64)
1737 0 : default:
1738 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1739 0 : return nil
1740 0 : }
1741 0 : return errors.Errorf("pebble: unknown option: %s.%s", errors.Safe(section), errors.Safe(key))
1742 : }
1743 1 : return err
1744 : }
1745 1 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1746 1 : return nil
1747 1 : }
1748 0 : return errors.Errorf("pebble: unknown section: %q", errors.Safe(section))
1749 : })
1750 : }
1751 :
1752 1 : func (o *Options) checkOptions(s string) (strictWALTail bool, err error) {
1753 1 : // TODO(jackson): Refactor to avoid awkwardness of the strictWALTail return value.
1754 1 : return strictWALTail, parseOptions(s, func(section, key, value string) error {
1755 1 : switch section + "." + key {
1756 1 : case "Options.comparer":
1757 1 : if value != o.Comparer.Name {
1758 0 : return errors.Errorf("pebble: comparer name from file %q != comparer name from options %q",
1759 0 : errors.Safe(value), errors.Safe(o.Comparer.Name))
1760 0 : }
1761 1 : case "Options.merger":
1762 1 : // RocksDB allows the merge operator to be unspecified, in which case it
1763 1 : // shows up as "nullptr".
1764 1 : if value != "nullptr" && value != o.Merger.Name {
1765 0 : return errors.Errorf("pebble: merger name from file %q != merger name from options %q",
1766 0 : errors.Safe(value), errors.Safe(o.Merger.Name))
1767 0 : }
1768 1 : case "Options.strict_wal_tail":
1769 1 : strictWALTail, err = strconv.ParseBool(value)
1770 1 : if err != nil {
1771 0 : return errors.Errorf("pebble: error parsing strict_wal_tail value %q: %w", value, err)
1772 0 : }
1773 : }
1774 1 : return nil
1775 : })
1776 : }
1777 :
1778 : // Check verifies the options are compatible with the previous options
1779 : // serialized by Options.String(). For example, the Comparer and Merger must be
1780 : // the same, or data will not be able to be properly read from the DB.
1781 0 : func (o *Options) Check(s string) error {
1782 0 : _, err := o.checkOptions(s)
1783 0 : return err
1784 0 : }
1785 :
1786 : // Validate verifies that the options are mutually consistent. For example,
1787 : // L0StopWritesThreshold must be >= L0CompactionThreshold, otherwise a write
1788 : // stall would persist indefinitely.
1789 1 : func (o *Options) Validate() error {
1790 1 : // Note that we can presume Options.EnsureDefaults has been called, so there
1791 1 : // is no need to check for zero values.
1792 1 :
1793 1 : var buf strings.Builder
1794 1 : if o.Experimental.L0CompactionConcurrency < 1 {
1795 0 : fmt.Fprintf(&buf, "L0CompactionConcurrency (%d) must be >= 1\n",
1796 0 : o.Experimental.L0CompactionConcurrency)
1797 0 : }
1798 1 : if o.L0StopWritesThreshold < o.L0CompactionThreshold {
1799 0 : fmt.Fprintf(&buf, "L0StopWritesThreshold (%d) must be >= L0CompactionThreshold (%d)\n",
1800 0 : o.L0StopWritesThreshold, o.L0CompactionThreshold)
1801 0 : }
1802 1 : if uint64(o.MemTableSize) >= maxMemTableSize {
1803 0 : fmt.Fprintf(&buf, "MemTableSize (%s) must be < %s\n",
1804 0 : humanize.Bytes.Uint64(uint64(o.MemTableSize)), humanize.Bytes.Uint64(maxMemTableSize))
1805 0 : }
1806 1 : if o.MemTableStopWritesThreshold < 2 {
1807 0 : fmt.Fprintf(&buf, "MemTableStopWritesThreshold (%d) must be >= 2\n",
1808 0 : o.MemTableStopWritesThreshold)
1809 0 : }
1810 1 : if o.FormatMajorVersion < FormatMinSupported || o.FormatMajorVersion > internalFormatNewest {
1811 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) must be between %d and %d\n",
1812 0 : o.FormatMajorVersion, FormatMinSupported, internalFormatNewest)
1813 0 : }
1814 1 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone && o.FormatMajorVersion < FormatMinForSharedObjects {
1815 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) when CreateOnShared is set must be at least %d\n",
1816 0 : o.FormatMajorVersion, FormatMinForSharedObjects)
1817 0 :
1818 0 : }
1819 1 : if o.TableCache != nil && o.Cache != o.TableCache.cache {
1820 0 : fmt.Fprintf(&buf, "underlying cache in the TableCache and the Cache dont match\n")
1821 0 : }
1822 1 : if buf.Len() == 0 {
1823 1 : return nil
1824 1 : }
1825 0 : return errors.New(buf.String())
1826 : }
1827 :
1828 : // MakeReaderOptions constructs sstable.ReaderOptions from the corresponding
1829 : // options in the receiver.
1830 1 : func (o *Options) MakeReaderOptions() sstable.ReaderOptions {
1831 1 : var readerOpts sstable.ReaderOptions
1832 1 : if o != nil {
1833 1 : readerOpts.Cache = o.Cache
1834 1 : readerOpts.Comparer = o.Comparer
1835 1 : readerOpts.Filters = o.Filters
1836 1 : if o.Merger != nil {
1837 1 : readerOpts.Merge = o.Merger.Merge
1838 1 : readerOpts.MergerName = o.Merger.Name
1839 1 : }
1840 1 : readerOpts.LoggerAndTracer = o.LoggerAndTracer
1841 : }
1842 1 : return readerOpts
1843 : }
1844 :
1845 : // MakeWriterOptions constructs sstable.WriterOptions for the specified level
1846 : // from the corresponding options in the receiver.
1847 1 : func (o *Options) MakeWriterOptions(level int, format sstable.TableFormat) sstable.WriterOptions {
1848 1 : var writerOpts sstable.WriterOptions
1849 1 : writerOpts.TableFormat = format
1850 1 : if o != nil {
1851 1 : writerOpts.Cache = o.Cache
1852 1 : writerOpts.Comparer = o.Comparer
1853 1 : if o.Merger != nil {
1854 1 : writerOpts.MergerName = o.Merger.Name
1855 1 : }
1856 1 : writerOpts.TablePropertyCollectors = o.TablePropertyCollectors
1857 1 : writerOpts.BlockPropertyCollectors = o.BlockPropertyCollectors
1858 : }
1859 1 : if format >= sstable.TableFormatPebblev3 {
1860 1 : writerOpts.ShortAttributeExtractor = o.Experimental.ShortAttributeExtractor
1861 1 : writerOpts.RequiredInPlaceValueBound = o.Experimental.RequiredInPlaceValueBound
1862 1 : if format >= sstable.TableFormatPebblev4 && level == numLevels-1 {
1863 1 : writerOpts.WritingToLowestLevel = true
1864 1 : }
1865 : }
1866 1 : levelOpts := o.Level(level)
1867 1 : writerOpts.BlockRestartInterval = levelOpts.BlockRestartInterval
1868 1 : writerOpts.BlockSize = levelOpts.BlockSize
1869 1 : writerOpts.BlockSizeThreshold = levelOpts.BlockSizeThreshold
1870 1 : writerOpts.Compression = levelOpts.Compression
1871 1 : writerOpts.FilterPolicy = levelOpts.FilterPolicy
1872 1 : writerOpts.FilterType = levelOpts.FilterType
1873 1 : writerOpts.IndexBlockSize = levelOpts.IndexBlockSize
1874 1 : return writerOpts
1875 : }
|
