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/fifo"
19 : "github.com/cockroachdb/pebble/internal/base"
20 : "github.com/cockroachdb/pebble/internal/cache"
21 : "github.com/cockroachdb/pebble/internal/humanize"
22 : "github.com/cockroachdb/pebble/internal/keyspan"
23 : "github.com/cockroachdb/pebble/internal/manifest"
24 : "github.com/cockroachdb/pebble/internal/testkeys"
25 : "github.com/cockroachdb/pebble/objstorage/objstorageprovider"
26 : "github.com/cockroachdb/pebble/objstorage/remote"
27 : "github.com/cockroachdb/pebble/rangekey"
28 : "github.com/cockroachdb/pebble/sstable"
29 : "github.com/cockroachdb/pebble/sstable/block"
30 : "github.com/cockroachdb/pebble/sstable/colblk"
31 : "github.com/cockroachdb/pebble/vfs"
32 : "github.com/cockroachdb/pebble/wal"
33 : )
34 :
35 : const (
36 : cacheDefaultSize = 8 << 20 // 8 MB
37 : defaultLevelMultiplier = 10
38 : )
39 :
40 : // Compression exports the base.Compression type.
41 : type Compression = block.Compression
42 :
43 : // Exported Compression constants.
44 : const (
45 : DefaultCompression = block.DefaultCompression
46 : NoCompression = block.NoCompression
47 : SnappyCompression = block.SnappyCompression
48 : ZstdCompression = block.ZstdCompression
49 : )
50 :
51 : // FilterType exports the base.FilterType type.
52 : type FilterType = base.FilterType
53 :
54 : // Exported TableFilter constants.
55 : const (
56 : TableFilter = base.TableFilter
57 : )
58 :
59 : // FilterWriter exports the base.FilterWriter type.
60 : type FilterWriter = base.FilterWriter
61 :
62 : // FilterPolicy exports the base.FilterPolicy type.
63 : type FilterPolicy = base.FilterPolicy
64 :
65 : // KeySchema exports the colblk.KeySchema type.
66 : type KeySchema = colblk.KeySchema
67 :
68 : // BlockPropertyCollector exports the sstable.BlockPropertyCollector type.
69 : type BlockPropertyCollector = sstable.BlockPropertyCollector
70 :
71 : // BlockPropertyFilter exports the sstable.BlockPropertyFilter type.
72 : type BlockPropertyFilter = base.BlockPropertyFilter
73 :
74 : // ShortAttributeExtractor exports the base.ShortAttributeExtractor type.
75 : type ShortAttributeExtractor = base.ShortAttributeExtractor
76 :
77 : // UserKeyPrefixBound exports the sstable.UserKeyPrefixBound type.
78 : type UserKeyPrefixBound = sstable.UserKeyPrefixBound
79 :
80 : // IterKeyType configures which types of keys an iterator should surface.
81 : type IterKeyType int8
82 :
83 : const (
84 : // IterKeyTypePointsOnly configures an iterator to iterate over point keys
85 : // only.
86 : IterKeyTypePointsOnly IterKeyType = iota
87 : // IterKeyTypeRangesOnly configures an iterator to iterate over range keys
88 : // only.
89 : IterKeyTypeRangesOnly
90 : // IterKeyTypePointsAndRanges configures an iterator iterate over both point
91 : // keys and range keys simultaneously.
92 : IterKeyTypePointsAndRanges
93 : )
94 :
95 : // String implements fmt.Stringer.
96 1 : func (t IterKeyType) String() string {
97 1 : switch t {
98 1 : case IterKeyTypePointsOnly:
99 1 : return "points-only"
100 1 : case IterKeyTypeRangesOnly:
101 1 : return "ranges-only"
102 1 : case IterKeyTypePointsAndRanges:
103 1 : return "points-and-ranges"
104 0 : default:
105 0 : panic(fmt.Sprintf("unknown key type %d", t))
106 : }
107 : }
108 :
109 : // IterOptions hold the optional per-query parameters for NewIter.
110 : //
111 : // Like Options, a nil *IterOptions is valid and means to use the default
112 : // values.
113 : type IterOptions struct {
114 : // LowerBound specifies the smallest key (inclusive) that the iterator will
115 : // return during iteration. If the iterator is seeked or iterated past this
116 : // boundary the iterator will return Valid()==false. Setting LowerBound
117 : // effectively truncates the key space visible to the iterator.
118 : LowerBound []byte
119 : // UpperBound specifies the largest key (exclusive) that the iterator will
120 : // return during iteration. If the iterator is seeked or iterated past this
121 : // boundary the iterator will return Valid()==false. Setting UpperBound
122 : // effectively truncates the key space visible to the iterator.
123 : UpperBound []byte
124 : // SkipPoint may be used to skip over point keys that don't match an
125 : // arbitrary predicate during iteration. If set, the Iterator invokes
126 : // SkipPoint for keys encountered. If SkipPoint returns true, the iterator
127 : // will skip the key without yielding it to the iterator operation in
128 : // progress.
129 : //
130 : // SkipPoint must be a pure function and always return the same result when
131 : // provided the same arguments. The iterator may call SkipPoint multiple
132 : // times for the same user key.
133 : SkipPoint func(userKey []byte) bool
134 : // PointKeyFilters can be used to avoid scanning tables and blocks in tables
135 : // when iterating over point keys. This slice represents an intersection
136 : // across all filters, i.e., all filters must indicate that the block is
137 : // relevant.
138 : //
139 : // Performance note: When len(PointKeyFilters) > 0, the caller should ensure
140 : // that cap(PointKeyFilters) is at least len(PointKeyFilters)+1. This helps
141 : // avoid allocations in Pebble internal code that mutates the slice.
142 : PointKeyFilters []BlockPropertyFilter
143 : // RangeKeyFilters can be usefd to avoid scanning tables and blocks in tables
144 : // when iterating over range keys. The same requirements that apply to
145 : // PointKeyFilters apply here too.
146 : RangeKeyFilters []BlockPropertyFilter
147 : // KeyTypes configures which types of keys to iterate over: point keys,
148 : // range keys, or both.
149 : KeyTypes IterKeyType
150 : // RangeKeyMasking can be used to enable automatic masking of point keys by
151 : // range keys. Range key masking is only supported during combined range key
152 : // and point key iteration mode (IterKeyTypePointsAndRanges).
153 : RangeKeyMasking RangeKeyMasking
154 :
155 : // OnlyReadGuaranteedDurable is an advanced option that is only supported by
156 : // the Reader implemented by DB. When set to true, only the guaranteed to be
157 : // durable state is visible in the iterator.
158 : // - This definition is made under the assumption that the FS implementation
159 : // is providing a durability guarantee when data is synced.
160 : // - The visible state represents a consistent point in the history of the
161 : // DB.
162 : // - The implementation is free to choose a conservative definition of what
163 : // is guaranteed durable. For simplicity, the current implementation
164 : // ignores memtables. A more sophisticated implementation could track the
165 : // highest seqnum that is synced to the WAL and published and use that as
166 : // the visible seqnum for an iterator. Note that the latter approach is
167 : // not strictly better than the former since we can have DBs that are (a)
168 : // synced more rarely than memtable flushes, (b) have no WAL. (a) is
169 : // likely to be true in a future CockroachDB context where the DB
170 : // containing the state machine may be rarely synced.
171 : // NB: this current implementation relies on the fact that memtables are
172 : // flushed in seqnum order, and any ingested sstables that happen to have a
173 : // lower seqnum than a non-flushed memtable don't have any overlapping keys.
174 : // This is the fundamental level invariant used in other code too, like when
175 : // merging iterators.
176 : //
177 : // Semantically, using this option provides the caller a "snapshot" as of
178 : // the time the most recent memtable was flushed. An alternate interface
179 : // would be to add a NewSnapshot variant. Creating a snapshot is heavier
180 : // weight than creating an iterator, so we have opted to support this
181 : // iterator option.
182 : OnlyReadGuaranteedDurable bool
183 : // UseL6Filters allows the caller to opt into reading filter blocks for L6
184 : // sstables. Helpful if a lot of SeekPrefixGEs are expected in quick
185 : // succession, that are also likely to not yield a single key. Filter blocks in
186 : // L6 can be relatively large, often larger than data blocks, so the benefit of
187 : // loading them in the cache is minimized if the probability of the key
188 : // existing is not low or if we just expect a one-time Seek (where loading the
189 : // data block directly is better).
190 : UseL6Filters bool
191 : // CategoryAndQoS is used for categorized iterator stats. This should not be
192 : // changed by calling SetOptions.
193 : sstable.CategoryAndQoS
194 :
195 : DebugRangeKeyStack bool
196 :
197 : // Internal options.
198 :
199 : logger Logger
200 : // Layer corresponding to this file. Only passed in if constructed by a
201 : // levelIter.
202 : layer manifest.Layer
203 : // disableLazyCombinedIteration is an internal testing option.
204 : disableLazyCombinedIteration bool
205 : // snapshotForHideObsoletePoints is specified for/by levelIter when opening
206 : // files and is used to decide whether to hide obsolete points. A value of 0
207 : // implies obsolete points should not be hidden.
208 : snapshotForHideObsoletePoints base.SeqNum
209 :
210 : // NB: If adding new Options, you must account for them in iterator
211 : // construction and Iterator.SetOptions.
212 : }
213 :
214 : // GetLowerBound returns the LowerBound or nil if the receiver is nil.
215 2 : func (o *IterOptions) GetLowerBound() []byte {
216 2 : if o == nil {
217 2 : return nil
218 2 : }
219 2 : return o.LowerBound
220 : }
221 :
222 : // GetUpperBound returns the UpperBound or nil if the receiver is nil.
223 2 : func (o *IterOptions) GetUpperBound() []byte {
224 2 : if o == nil {
225 2 : return nil
226 2 : }
227 2 : return o.UpperBound
228 : }
229 :
230 2 : func (o *IterOptions) pointKeys() bool {
231 2 : if o == nil {
232 0 : return true
233 0 : }
234 2 : return o.KeyTypes == IterKeyTypePointsOnly || o.KeyTypes == IterKeyTypePointsAndRanges
235 : }
236 :
237 2 : func (o *IterOptions) rangeKeys() bool {
238 2 : if o == nil {
239 0 : return false
240 0 : }
241 2 : return o.KeyTypes == IterKeyTypeRangesOnly || o.KeyTypes == IterKeyTypePointsAndRanges
242 : }
243 :
244 2 : func (o *IterOptions) getLogger() Logger {
245 2 : if o == nil || o.logger == nil {
246 2 : return DefaultLogger
247 2 : }
248 2 : return o.logger
249 : }
250 :
251 : // SpanIterOptions creates a SpanIterOptions from this IterOptions.
252 2 : func (o *IterOptions) SpanIterOptions() keyspan.SpanIterOptions {
253 2 : if o == nil {
254 2 : return keyspan.SpanIterOptions{}
255 2 : }
256 2 : return keyspan.SpanIterOptions{
257 2 : RangeKeyFilters: o.RangeKeyFilters,
258 2 : }
259 : }
260 :
261 : // scanInternalOptions is similar to IterOptions, meant for use with
262 : // scanInternalIterator.
263 : type scanInternalOptions struct {
264 : sstable.CategoryAndQoS
265 : IterOptions
266 :
267 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error
268 : visitRangeDel func(start, end []byte, seqNum SeqNum) error
269 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error
270 : visitSharedFile func(sst *SharedSSTMeta) error
271 : visitExternalFile func(sst *ExternalFile) error
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 2 : func (o *WriteOptions) GetSync() bool {
382 2 : return o == nil || o.Sync
383 2 : }
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 func() 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 2 : func (o *LevelOptions) EnsureDefaults() *LevelOptions {
445 2 : if o == nil {
446 0 : o = &LevelOptions{}
447 0 : }
448 2 : if o.BlockRestartInterval <= 0 {
449 1 : o.BlockRestartInterval = base.DefaultBlockRestartInterval
450 1 : }
451 2 : if o.BlockSize <= 0 {
452 1 : o.BlockSize = base.DefaultBlockSize
453 2 : } else if o.BlockSize > sstable.MaximumBlockSize {
454 0 : panic(errors.Errorf("BlockSize %d exceeds MaximumBlockSize", o.BlockSize))
455 : }
456 2 : if o.BlockSizeThreshold <= 0 {
457 1 : o.BlockSizeThreshold = base.DefaultBlockSizeThreshold
458 1 : }
459 2 : if o.Compression == nil {
460 1 : o.Compression = func() Compression { return DefaultCompression }
461 : }
462 2 : if o.IndexBlockSize <= 0 {
463 1 : o.IndexBlockSize = o.BlockSize
464 1 : }
465 2 : if o.TargetFileSize <= 0 {
466 1 : o.TargetFileSize = 2 << 20 // 2 MB
467 1 : }
468 2 : 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 : // LoadBlockSema, if set, is used to limit the number of blocks that can be
490 : // loaded (i.e. read from the filesystem) in parallel. Each load acquires one
491 : // unit from the semaphore for the duration of the read.
492 : LoadBlockSema *fifo.Semaphore
493 :
494 : // Cleaner cleans obsolete files.
495 : //
496 : // The default cleaner uses the DeleteCleaner.
497 : Cleaner Cleaner
498 :
499 : // Local contains option that pertain to files stored on the local filesystem.
500 : Local struct {
501 : // ReadaheadConfig is used to retrieve the current readahead mode; it is
502 : // consulted whenever a read handle is initialized.
503 : ReadaheadConfig *ReadaheadConfig
504 :
505 : // TODO(radu): move BytesPerSync, LoadBlockSema, Cleaner here.
506 : }
507 :
508 : // Comparer defines a total ordering over the space of []byte keys: a 'less
509 : // than' relationship. The same comparison algorithm must be used for reads
510 : // and writes over the lifetime of the DB.
511 : //
512 : // The default value uses the same ordering as bytes.Compare.
513 : Comparer *Comparer
514 :
515 : // DebugCheck is invoked, if non-nil, whenever a new version is being
516 : // installed. Typically, this is set to pebble.DebugCheckLevels in tests
517 : // or tools only, to check invariants over all the data in the database.
518 : DebugCheck func(*DB) error
519 :
520 : // Disable the write-ahead log (WAL). Disabling the write-ahead log prohibits
521 : // crash recovery, but can improve performance if crash recovery is not
522 : // needed (e.g. when only temporary state is being stored in the database).
523 : //
524 : // TODO(peter): untested
525 : DisableWAL bool
526 :
527 : // ErrorIfExists causes an error on Open if the database already exists.
528 : // The error can be checked with errors.Is(err, ErrDBAlreadyExists).
529 : //
530 : // The default value is false.
531 : ErrorIfExists bool
532 :
533 : // ErrorIfNotExists causes an error on Open if the database does not already
534 : // exist. The error can be checked with errors.Is(err, ErrDBDoesNotExist).
535 : //
536 : // The default value is false which will cause a database to be created if it
537 : // does not already exist.
538 : ErrorIfNotExists bool
539 :
540 : // ErrorIfNotPristine causes an error on Open if the database already exists
541 : // and any operations have been performed on the database. The error can be
542 : // checked with errors.Is(err, ErrDBNotPristine).
543 : //
544 : // Note that a database that contained keys that were all subsequently deleted
545 : // may or may not trigger the error. Currently, we check if there are any live
546 : // SSTs or log records to replay.
547 : ErrorIfNotPristine bool
548 :
549 : // EventListener provides hooks to listening to significant DB events such as
550 : // flushes, compactions, and table deletion.
551 : EventListener *EventListener
552 :
553 : // Experimental contains experimental options which are off by default.
554 : // These options are temporary and will eventually either be deleted, moved
555 : // out of the experimental group, or made the non-adjustable default. These
556 : // options may change at any time, so do not rely on them.
557 : Experimental struct {
558 : // The threshold of L0 read-amplification at which compaction concurrency
559 : // is enabled (if CompactionDebtConcurrency was not already exceeded).
560 : // Every multiple of this value enables another concurrent
561 : // compaction up to MaxConcurrentCompactions.
562 : L0CompactionConcurrency int
563 :
564 : // CompactionDebtConcurrency controls the threshold of compaction debt
565 : // at which additional compaction concurrency slots are added. For every
566 : // multiple of this value in compaction debt bytes, an additional
567 : // concurrent compaction is added. This works "on top" of
568 : // L0CompactionConcurrency, so the higher of the count of compaction
569 : // concurrency slots as determined by the two options is chosen.
570 : CompactionDebtConcurrency uint64
571 :
572 : // IngestSplit, if it returns true, allows for ingest-time splitting of
573 : // existing sstables into two virtual sstables to allow ingestion sstables to
574 : // slot into a lower level than they otherwise would have.
575 : IngestSplit func() bool
576 :
577 : // ReadCompactionRate controls the frequency of read triggered
578 : // compactions by adjusting `AllowedSeeks` in manifest.FileMetadata:
579 : //
580 : // AllowedSeeks = FileSize / ReadCompactionRate
581 : //
582 : // From LevelDB:
583 : // ```
584 : // We arrange to automatically compact this file after
585 : // a certain number of seeks. Let's assume:
586 : // (1) One seek costs 10ms
587 : // (2) Writing or reading 1MB costs 10ms (100MB/s)
588 : // (3) A compaction of 1MB does 25MB of IO:
589 : // 1MB read from this level
590 : // 10-12MB read from next level (boundaries may be misaligned)
591 : // 10-12MB written to next level
592 : // This implies that 25 seeks cost the same as the compaction
593 : // of 1MB of data. I.e., one seek costs approximately the
594 : // same as the compaction of 40KB of data. We are a little
595 : // conservative and allow approximately one seek for every 16KB
596 : // of data before triggering a compaction.
597 : // ```
598 : ReadCompactionRate int64
599 :
600 : // ReadSamplingMultiplier is a multiplier for the readSamplingPeriod in
601 : // iterator.maybeSampleRead() to control the frequency of read sampling
602 : // to trigger a read triggered compaction. A value of -1 prevents sampling
603 : // and disables read triggered compactions. The default is 1 << 4. which
604 : // gets multiplied with a constant of 1 << 16 to yield 1 << 20 (1MB).
605 : ReadSamplingMultiplier int64
606 :
607 : // NumDeletionsThreshold defines the minimum number of point tombstones
608 : // that must be present in a single data block for that block to be
609 : // considered tombstone-dense for the purposes of triggering a
610 : // tombstone density compaction. Data blocks may also be considered
611 : // tombstone-dense if they meet the criteria defined by
612 : // DeletionSizeRatioThreshold below. Tombstone-dense blocks are identified
613 : // when sstables are written, and so this is effectively an option for
614 : // sstable writers. The default value is 100.
615 : NumDeletionsThreshold int
616 :
617 : // DeletionSizeRatioThreshold defines the minimum ratio of the size of
618 : // point tombstones to the size of a data block that must be reached
619 : // for that block to be considered tombstone-dense for the purposes of
620 : // triggering a tombstone density compaction. Data blocks may also be
621 : // considered tombstone-dense if they meet the criteria defined by
622 : // NumDeletionsThreshold above. Tombstone-dense blocks are identified
623 : // when sstables are written, and so this is effectively an option for
624 : // sstable writers. The default value is 0.5.
625 : DeletionSizeRatioThreshold float32
626 :
627 : // TombstoneDenseCompactionThreshold is the minimum percent of data
628 : // blocks in a table that must be tombstone-dense for that table to be
629 : // eligible for a tombstone density compaction. It should be defined as a
630 : // ratio out of 1. The default value is 0.05.
631 : //
632 : // If multiple tables are eligible for a tombstone density compaction, then
633 : // tables with a higher percent of tombstone-dense blocks are still
634 : // prioritized for compaction.
635 : //
636 : // A zero or negative value disables tombstone density compactions.
637 : TombstoneDenseCompactionThreshold float64
638 :
639 : // TableCacheShards is the number of shards per table cache.
640 : // Reducing the value can reduce the number of idle goroutines per DB
641 : // instance which can be useful in scenarios with a lot of DB instances
642 : // and a large number of CPUs, but doing so can lead to higher contention
643 : // in the table cache and reduced performance.
644 : //
645 : // The default value is the number of logical CPUs, which can be
646 : // limited by runtime.GOMAXPROCS.
647 : TableCacheShards int
648 :
649 : // KeyValidationFunc is a function to validate a user key in an SSTable.
650 : //
651 : // Currently, this function is used to validate the smallest and largest
652 : // keys in an SSTable undergoing compaction. In this case, returning an
653 : // error from the validation function will result in a panic at runtime,
654 : // given that there is rarely any way of recovering from malformed keys
655 : // present in compacted files. By default, validation is not performed.
656 : //
657 : // Additional use-cases may be added in the future.
658 : //
659 : // NOTE: callers should take care to not mutate the key being validated.
660 : KeyValidationFunc func(userKey []byte) error
661 :
662 : // ValidateOnIngest schedules validation of sstables after they have
663 : // been ingested.
664 : //
665 : // By default, this value is false.
666 : ValidateOnIngest bool
667 :
668 : // LevelMultiplier configures the size multiplier used to determine the
669 : // desired size of each level of the LSM. Defaults to 10.
670 : LevelMultiplier int
671 :
672 : // MultiLevelCompactionHeuristic determines whether to add an additional
673 : // level to a conventional two level compaction. If nil, a multilevel
674 : // compaction will never get triggered.
675 : MultiLevelCompactionHeuristic MultiLevelHeuristic
676 :
677 : // MaxWriterConcurrency is used to indicate the maximum number of
678 : // compression workers the compression queue is allowed to use. If
679 : // MaxWriterConcurrency > 0, then the Writer will use parallelism, to
680 : // compress and write blocks to disk. Otherwise, the writer will
681 : // compress and write blocks to disk synchronously.
682 : MaxWriterConcurrency int
683 :
684 : // ForceWriterParallelism is used to force parallelism in the sstable
685 : // Writer for the metamorphic tests. Even with the MaxWriterConcurrency
686 : // option set, we only enable parallelism in the sstable Writer if there
687 : // is enough CPU available, and this option bypasses that.
688 : ForceWriterParallelism bool
689 :
690 : // CPUWorkPermissionGranter should be set if Pebble should be given the
691 : // ability to optionally schedule additional CPU. See the documentation
692 : // for CPUWorkPermissionGranter for more details.
693 : CPUWorkPermissionGranter CPUWorkPermissionGranter
694 :
695 : // EnableColumnarBlocks is used to decide whether to enable writing
696 : // TableFormatPebblev5 sstables. This setting is only respected by
697 : // FormatColumnarBlocks. In lower format major versions, the
698 : // TableFormatPebblev5 format is prohibited. If EnableColumnarBlocks is
699 : // nil and the DB is at FormatColumnarBlocks, the DB defaults to not
700 : // writing columnar blocks.
701 : EnableColumnarBlocks func() bool
702 :
703 : // EnableValueBlocks is used to decide whether to enable writing
704 : // TableFormatPebblev3 sstables. This setting is only respected by a
705 : // specific subset of format major versions: FormatSSTableValueBlocks,
706 : // FormatFlushableIngest and FormatPrePebblev1MarkedCompacted. In lower
707 : // format major versions, value blocks are never enabled. In higher
708 : // format major versions, value blocks are always enabled.
709 : EnableValueBlocks func() bool
710 :
711 : // ShortAttributeExtractor is used iff EnableValueBlocks() returns true
712 : // (else ignored). If non-nil, a ShortAttribute can be extracted from the
713 : // value and stored with the key, when the value is stored elsewhere.
714 : ShortAttributeExtractor ShortAttributeExtractor
715 :
716 : // RequiredInPlaceValueBound specifies an optional span of user key
717 : // prefixes that are not-MVCC, but have a suffix. For these the values
718 : // must be stored with the key, since the concept of "older versions" is
719 : // not defined. It is also useful for statically known exclusions to value
720 : // separation. In CockroachDB, this will be used for the lock table key
721 : // space that has non-empty suffixes, but those locks don't represent
722 : // actual MVCC versions (the suffix ordering is arbitrary). We will also
723 : // need to add support for dynamically configured exclusions (we want the
724 : // default to be to allow Pebble to decide whether to separate the value
725 : // or not, hence this is structured as exclusions), for example, for users
726 : // of CockroachDB to dynamically exclude certain tables.
727 : //
728 : // Any change in exclusion behavior takes effect only on future written
729 : // sstables, and does not start rewriting existing sstables.
730 : //
731 : // Even ignoring changes in this setting, exclusions are interpreted as a
732 : // guidance by Pebble, and not necessarily honored. Specifically, user
733 : // keys with multiple Pebble-versions *may* have the older versions stored
734 : // in value blocks.
735 : RequiredInPlaceValueBound UserKeyPrefixBound
736 :
737 : // DisableIngestAsFlushable disables lazy ingestion of sstables through
738 : // a WAL write and memtable rotation. Only effectual if the format
739 : // major version is at least `FormatFlushableIngest`.
740 : DisableIngestAsFlushable func() bool
741 :
742 : // RemoteStorage enables use of remote storage (e.g. S3) for storing
743 : // sstables. Setting this option enables use of CreateOnShared option and
744 : // allows ingestion of external files.
745 : RemoteStorage remote.StorageFactory
746 :
747 : // If CreateOnShared is non-zero, new sstables are created on remote storage
748 : // (using CreateOnSharedLocator and with the appropriate
749 : // CreateOnSharedStrategy). These sstables can be shared between different
750 : // Pebble instances; the lifecycle of such objects is managed by the
751 : // remote.Storage constructed by options.RemoteStorage.
752 : //
753 : // Can only be used when RemoteStorage is set (and recognizes
754 : // CreateOnSharedLocator).
755 : CreateOnShared remote.CreateOnSharedStrategy
756 : CreateOnSharedLocator remote.Locator
757 :
758 : // CacheSizeBytesBytes is the size of the on-disk block cache for objects
759 : // on shared storage in bytes. If it is 0, no cache is used.
760 : SecondaryCacheSizeBytes int64
761 :
762 : // NB: DO NOT crash on SingleDeleteInvariantViolationCallback or
763 : // IneffectualSingleDeleteCallback, since these can be false positives
764 : // even if SingleDel has been used correctly.
765 : //
766 : // Pebble's delete-only compactions can cause a recent RANGEDEL to peek
767 : // below an older SINGLEDEL and delete an arbitrary subset of data below
768 : // that SINGLEDEL. When that SINGLEDEL gets compacted (without the
769 : // RANGEDEL), any of these callbacks can happen, without it being a real
770 : // correctness problem.
771 : //
772 : // Example 1:
773 : // RANGEDEL [a, c)#10 in L0
774 : // SINGLEDEL b#5 in L1
775 : // SET b#3 in L6
776 : //
777 : // If the L6 file containing the SET is narrow and the L1 file containing
778 : // the SINGLEDEL is wide, a delete-only compaction can remove the file in
779 : // L2 before the SINGLEDEL is compacted down. Then when the SINGLEDEL is
780 : // compacted down, it will not find any SET to delete, resulting in the
781 : // ineffectual callback.
782 : //
783 : // Example 2:
784 : // RANGEDEL [a, z)#60 in L0
785 : // SINGLEDEL g#50 in L1
786 : // SET g#40 in L2
787 : // RANGEDEL [g,h)#30 in L3
788 : // SET g#20 in L6
789 : //
790 : // In this example, the two SETs represent the same user write, and the
791 : // RANGEDELs are caused by the CockroachDB range being dropped. That is,
792 : // the user wrote to g once, range was dropped, then added back, which
793 : // caused the SET again, then at some point g was validly deleted using a
794 : // SINGLEDEL, and then the range was dropped again. The older RANGEDEL can
795 : // get fragmented due to compactions it has been part of. Say this L3 file
796 : // containing the RANGEDEL is very narrow, while the L1, L2, L6 files are
797 : // wider than the RANGEDEL in L0. Then the RANGEDEL in L3 can be dropped
798 : // using a delete-only compaction, resulting in an LSM with state:
799 : //
800 : // RANGEDEL [a, z)#60 in L0
801 : // SINGLEDEL g#50 in L1
802 : // SET g#40 in L2
803 : // SET g#20 in L6
804 : //
805 : // A multi-level compaction involving L1, L2, L6 will cause the invariant
806 : // violation callback. This example doesn't need multi-level compactions:
807 : // say there was a Pebble snapshot at g#21 preventing g#20 from being
808 : // dropped when it meets g#40 in a compaction. That snapshot will not save
809 : // RANGEDEL [g,h)#30, so we can have:
810 : //
811 : // SINGLEDEL g#50 in L1
812 : // SET g#40, SET g#20 in L6
813 : //
814 : // And say the snapshot is removed and then the L1 and L6 compaction
815 : // happens, resulting in the invariant violation callback.
816 : //
817 : // TODO(sumeer): rename SingleDeleteInvariantViolationCallback to remove
818 : // the word "invariant".
819 :
820 : // IneffectualPointDeleteCallback is called in compactions/flushes if any
821 : // single delete is being elided without deleting a point set/merge.
822 : IneffectualSingleDeleteCallback func(userKey []byte)
823 :
824 : // SingleDeleteInvariantViolationCallback is called in compactions/flushes if any
825 : // single delete has consumed a Set/Merge, and there is another immediately older
826 : // Set/SetWithDelete/Merge. The user of Pebble has violated the invariant under
827 : // which SingleDelete can be used correctly.
828 : //
829 : // Consider the sequence SingleDelete#3, Set#2, Set#1. There are three
830 : // ways some of these keys can first meet in a compaction.
831 : //
832 : // - All 3 keys in the same compaction: this callback will detect the
833 : // violation.
834 : //
835 : // - SingleDelete#3, Set#2 meet in a compaction first: Both keys will
836 : // disappear. The violation will not be detected, and the DB will have
837 : // Set#1 which is likely incorrect (from the user's perspective).
838 : //
839 : // - Set#2, Set#1 meet in a compaction first: The output will be Set#2,
840 : // which will later be consumed by SingleDelete#3. The violation will
841 : // not be detected and the DB will be correct.
842 : SingleDeleteInvariantViolationCallback func(userKey []byte)
843 :
844 : // EnableDeleteOnlyCompactionExcises enables delete-only compactions to also
845 : // apply delete-only compaction hints on sstables that partially overlap
846 : // with it. This application happens through an excise, similar to
847 : // the excise phase of IngestAndExcise.
848 : EnableDeleteOnlyCompactionExcises func() bool
849 : }
850 :
851 : // Filters is a map from filter policy name to filter policy. It is used for
852 : // debugging tools which may be used on multiple databases configured with
853 : // different filter policies. It is not necessary to populate this filters
854 : // map during normal usage of a DB (it will be done automatically by
855 : // EnsureDefaults).
856 : Filters map[string]FilterPolicy
857 :
858 : // FlushDelayDeleteRange configures how long the database should wait before
859 : // forcing a flush of a memtable that contains a range deletion. Disk space
860 : // cannot be reclaimed until the range deletion is flushed. No automatic
861 : // flush occurs if zero.
862 : FlushDelayDeleteRange time.Duration
863 :
864 : // FlushDelayRangeKey configures how long the database should wait before
865 : // forcing a flush of a memtable that contains a range key. Range keys in
866 : // the memtable prevent lazy combined iteration, so it's desirable to flush
867 : // range keys promptly. No automatic flush occurs if zero.
868 : FlushDelayRangeKey time.Duration
869 :
870 : // FlushSplitBytes denotes the target number of bytes per sublevel in
871 : // each flush split interval (i.e. range between two flush split keys)
872 : // in L0 sstables. When set to zero, only a single sstable is generated
873 : // by each flush. When set to a non-zero value, flushes are split at
874 : // points to meet L0's TargetFileSize, any grandparent-related overlap
875 : // options, and at boundary keys of L0 flush split intervals (which are
876 : // targeted to contain around FlushSplitBytes bytes in each sublevel
877 : // between pairs of boundary keys). Splitting sstables during flush
878 : // allows increased compaction flexibility and concurrency when those
879 : // tables are compacted to lower levels.
880 : FlushSplitBytes int64
881 :
882 : // FormatMajorVersion sets the format of on-disk files. It is
883 : // recommended to set the format major version to an explicit
884 : // version, as the default may change over time.
885 : //
886 : // At Open if the existing database is formatted using a later
887 : // format major version that is known to this version of Pebble,
888 : // Pebble will continue to use the later format major version. If
889 : // the existing database's version is unknown, the caller may use
890 : // FormatMostCompatible and will be able to open the database
891 : // regardless of its actual version.
892 : //
893 : // If the existing database is formatted using a format major
894 : // version earlier than the one specified, Open will automatically
895 : // ratchet the database to the specified format major version.
896 : FormatMajorVersion FormatMajorVersion
897 :
898 : // FS provides the interface for persistent file storage.
899 : //
900 : // The default value uses the underlying operating system's file system.
901 : FS vfs.FS
902 :
903 : // KeySchema is the name of the key schema that should be used when writing
904 : // new sstables. There must be a key schema with this name defined in
905 : // KeySchemas. If not set, colblk.DefaultKeySchema is used to construct a
906 : // default key schema.
907 : KeySchema string
908 :
909 : // KeySchemas defines the set of known schemas of user keys. When columnar
910 : // blocks are in use (see FormatColumnarBlocks), the user may specify how a
911 : // key should be decomposed into columns. Each KeySchema must have a unique
912 : // name. The schema named by Options.KeySchema is used while writing
913 : // sstables during flushes and compactions.
914 : //
915 : // Multiple KeySchemas may be used over the lifetime of a database. Once a
916 : // KeySchema is used, it must be provided in KeySchemas in subsequent calls
917 : // to Open for perpetuity.
918 : KeySchemas sstable.KeySchemas
919 :
920 : // Lock, if set, must be a database lock acquired through LockDirectory for
921 : // the same directory passed to Open. If provided, Open will skip locking
922 : // the directory. Closing the database will not release the lock, and it's
923 : // the responsibility of the caller to release the lock after closing the
924 : // database.
925 : //
926 : // Open will enforce that the Lock passed locks the same directory passed to
927 : // Open. Concurrent calls to Open using the same Lock are detected and
928 : // prohibited.
929 : Lock *Lock
930 :
931 : // The count of L0 files necessary to trigger an L0 compaction.
932 : L0CompactionFileThreshold int
933 :
934 : // The amount of L0 read-amplification necessary to trigger an L0 compaction.
935 : L0CompactionThreshold int
936 :
937 : // Hard limit on L0 read-amplification, computed as the number of L0
938 : // sublevels. Writes are stopped when this threshold is reached.
939 : L0StopWritesThreshold int
940 :
941 : // The maximum number of bytes for LBase. The base level is the level which
942 : // L0 is compacted into. The base level is determined dynamically based on
943 : // the existing data in the LSM. The maximum number of bytes for other levels
944 : // is computed dynamically based on the base level's maximum size. When the
945 : // maximum number of bytes for a level is exceeded, compaction is requested.
946 : LBaseMaxBytes int64
947 :
948 : // Per-level options. Options for at least one level must be specified. The
949 : // options for the last level are used for all subsequent levels.
950 : Levels []LevelOptions
951 :
952 : // LoggerAndTracer will be used, if non-nil, else Logger will be used and
953 : // tracing will be a noop.
954 :
955 : // Logger used to write log messages.
956 : //
957 : // The default logger uses the Go standard library log package.
958 : Logger Logger
959 : // LoggerAndTracer is used for writing log messages and traces.
960 : LoggerAndTracer LoggerAndTracer
961 :
962 : // MaxManifestFileSize is the maximum size the MANIFEST file is allowed to
963 : // become. When the MANIFEST exceeds this size it is rolled over and a new
964 : // MANIFEST is created.
965 : MaxManifestFileSize int64
966 :
967 : // MaxOpenFiles is a soft limit on the number of open files that can be
968 : // used by the DB.
969 : //
970 : // The default value is 1000.
971 : MaxOpenFiles int
972 :
973 : // The size of a MemTable in steady state. The actual MemTable size starts at
974 : // min(256KB, MemTableSize) and doubles for each subsequent MemTable up to
975 : // MemTableSize. This reduces the memory pressure caused by MemTables for
976 : // short lived (test) DB instances. Note that more than one MemTable can be
977 : // in existence since flushing a MemTable involves creating a new one and
978 : // writing the contents of the old one in the
979 : // background. MemTableStopWritesThreshold places a hard limit on the size of
980 : // the queued MemTables.
981 : //
982 : // The default value is 4MB.
983 : MemTableSize uint64
984 :
985 : // Hard limit on the number of queued of MemTables. Writes are stopped when
986 : // the sum of the queued memtable sizes exceeds:
987 : // MemTableStopWritesThreshold * MemTableSize.
988 : //
989 : // This value should be at least 2 or writes will stop whenever a MemTable is
990 : // being flushed.
991 : //
992 : // The default value is 2.
993 : MemTableStopWritesThreshold int
994 :
995 : // Merger defines the associative merge operation to use for merging values
996 : // written with {Batch,DB}.Merge.
997 : //
998 : // The default merger concatenates values.
999 : Merger *Merger
1000 :
1001 : // MaxConcurrentCompactions specifies the maximum number of concurrent
1002 : // compactions (not including download compactions).
1003 : //
1004 : // Concurrent compactions are performed:
1005 : // - when L0 read-amplification passes the L0CompactionConcurrency threshold;
1006 : // - for automatic background compactions;
1007 : // - when a manual compaction for a level is split and parallelized.
1008 : //
1009 : // MaxConcurrentCompactions() must be greater than 0.
1010 : //
1011 : // The default value is 1.
1012 : MaxConcurrentCompactions func() int
1013 :
1014 : // MaxConcurrentDownloads specifies the maximum number of download
1015 : // compactions. These are compactions that copy an external file to the local
1016 : // store.
1017 : //
1018 : // This limit is independent of MaxConcurrentCompactions; at any point in
1019 : // time, we may be running MaxConcurrentCompactions non-download compactions
1020 : // and MaxConcurrentDownloads download compactions.
1021 : //
1022 : // MaxConcurrentDownloads() must be greater than 0.
1023 : //
1024 : // The default value is 1.
1025 : MaxConcurrentDownloads func() int
1026 :
1027 : // DisableAutomaticCompactions dictates whether automatic compactions are
1028 : // scheduled or not. The default is false (enabled). This option is only used
1029 : // externally when running a manual compaction, and internally for tests.
1030 : DisableAutomaticCompactions bool
1031 :
1032 : // DisableConsistencyCheck disables the consistency check that is performed on
1033 : // open. Should only be used when a database cannot be opened normally (e.g.
1034 : // some of the tables don't exist / aren't accessible).
1035 : DisableConsistencyCheck bool
1036 :
1037 : // DisableTableStats dictates whether tables should be loaded asynchronously
1038 : // to compute statistics that inform compaction heuristics. The collection
1039 : // of table stats improves compaction of tombstones, reclaiming disk space
1040 : // more quickly and in some cases reducing write amplification in the
1041 : // presence of tombstones. Disabling table stats may be useful in tests
1042 : // that require determinism as the asynchronicity of table stats collection
1043 : // introduces significant nondeterminism.
1044 : DisableTableStats bool
1045 :
1046 : // NoSyncOnClose decides whether the Pebble instance will enforce a
1047 : // close-time synchronization (e.g., fdatasync() or sync_file_range())
1048 : // on files it writes to. Setting this to true removes the guarantee for a
1049 : // sync on close. Some implementations can still issue a non-blocking sync.
1050 : NoSyncOnClose bool
1051 :
1052 : // NumPrevManifest is the number of non-current or older manifests which
1053 : // we want to keep around for debugging purposes. By default, we're going
1054 : // to keep one older manifest.
1055 : NumPrevManifest int
1056 :
1057 : // ReadOnly indicates that the DB should be opened in read-only mode. Writes
1058 : // to the DB will return an error, background compactions are disabled, and
1059 : // the flush that normally occurs after replaying the WAL at startup is
1060 : // disabled.
1061 : ReadOnly bool
1062 :
1063 : // TableCache is an initialized TableCache which should be set as an
1064 : // option if the DB needs to be initialized with a pre-existing table cache.
1065 : // If TableCache is nil, then a table cache which is unique to the DB instance
1066 : // is created. TableCache can be shared between db instances by setting it here.
1067 : // The TableCache set here must use the same underlying cache as Options.Cache
1068 : // and pebble will panic otherwise.
1069 : TableCache *TableCache
1070 :
1071 : // BlockPropertyCollectors is a list of BlockPropertyCollector creation
1072 : // functions. A new BlockPropertyCollector is created for each sstable
1073 : // built and lives for the lifetime of writing that table.
1074 : BlockPropertyCollectors []func() BlockPropertyCollector
1075 :
1076 : // WALBytesPerSync sets the number of bytes to write to a WAL before calling
1077 : // Sync on it in the background. Just like with BytesPerSync above, this
1078 : // helps smooth out disk write latencies, and avoids cases where the OS
1079 : // writes a lot of buffered data to disk at once. However, this is less
1080 : // necessary with WALs, as many write operations already pass in
1081 : // Sync = true.
1082 : //
1083 : // The default value is 0, i.e. no background syncing. This matches the
1084 : // default behaviour in RocksDB.
1085 : WALBytesPerSync int
1086 :
1087 : // WALDir specifies the directory to store write-ahead logs (WALs) in. If
1088 : // empty (the default), WALs will be stored in the same directory as sstables
1089 : // (i.e. the directory passed to pebble.Open).
1090 : WALDir string
1091 :
1092 : // WALFailover may be set to configure Pebble to monitor writes to its
1093 : // write-ahead log and failover to writing write-ahead log entries to a
1094 : // secondary location (eg, a separate physical disk). WALFailover may be
1095 : // used to improve write availability in the presence of transient disk
1096 : // unavailability.
1097 : WALFailover *WALFailoverOptions
1098 :
1099 : // WALRecoveryDirs is a list of additional directories that should be
1100 : // scanned for the existence of additional write-ahead logs. WALRecoveryDirs
1101 : // is expected to be used when starting Pebble with a new WALDir or a new
1102 : // WALFailover configuration. The directories associated with the previous
1103 : // configuration may still contain WALs that are required for recovery of
1104 : // the current database state.
1105 : //
1106 : // If a previous WAL configuration may have stored WALs elsewhere but there
1107 : // is not a corresponding entry in WALRecoveryDirs, Open will error.
1108 : WALRecoveryDirs []wal.Dir
1109 :
1110 : // WALMinSyncInterval is the minimum duration between syncs of the WAL. If
1111 : // WAL syncs are requested faster than this interval, they will be
1112 : // artificially delayed. Introducing a small artificial delay (500us) between
1113 : // WAL syncs can allow more operations to arrive and reduce IO operations
1114 : // while having a minimal impact on throughput. This option is supplied as a
1115 : // closure in order to allow the value to be changed dynamically. The default
1116 : // value is 0.
1117 : //
1118 : // TODO(peter): rather than a closure, should there be another mechanism for
1119 : // changing options dynamically?
1120 : WALMinSyncInterval func() time.Duration
1121 :
1122 : // TargetByteDeletionRate is the rate (in bytes per second) at which sstable file
1123 : // deletions are limited to (under normal circumstances).
1124 : //
1125 : // Deletion pacing is used to slow down deletions when compactions finish up
1126 : // or readers close and newly-obsolete files need cleaning up. Deleting lots
1127 : // of files at once can cause disk latency to go up on some SSDs, which this
1128 : // functionality guards against.
1129 : //
1130 : // This value is only a best-effort target; the effective rate can be
1131 : // higher if deletions are falling behind or disk space is running low.
1132 : //
1133 : // Setting this to 0 disables deletion pacing, which is also the default.
1134 : TargetByteDeletionRate int
1135 :
1136 : // EnableSQLRowSpillMetrics specifies whether the Pebble instance will only be used
1137 : // to temporarily persist data spilled to disk for row-oriented SQL query execution.
1138 : EnableSQLRowSpillMetrics bool
1139 :
1140 : // AllocatorSizeClasses provides a sorted list containing the supported size
1141 : // classes of the underlying memory allocator. This provides hints to the
1142 : // sstable block writer's flushing policy to select block sizes that
1143 : // preemptively reduce internal fragmentation when loaded into the block cache.
1144 : AllocatorSizeClasses []int
1145 :
1146 : // private options are only used by internal tests or are used internally
1147 : // for facilitating upgrade paths of unconfigurable functionality.
1148 : private struct {
1149 : // disableDeleteOnlyCompactions prevents the scheduling of delete-only
1150 : // compactions that drop sstables wholy covered by range tombstones or
1151 : // range key tombstones.
1152 : disableDeleteOnlyCompactions bool
1153 :
1154 : // disableElisionOnlyCompactions prevents the scheduling of elision-only
1155 : // compactions that rewrite sstables in place in order to elide obsolete
1156 : // keys.
1157 : disableElisionOnlyCompactions bool
1158 :
1159 : // disableLazyCombinedIteration is a private option used by the
1160 : // metamorphic tests to test equivalence between lazy-combined iteration
1161 : // and constructing the range-key iterator upfront. It's a private
1162 : // option to avoid littering the public interface with options that we
1163 : // do not want to allow users to actually configure.
1164 : disableLazyCombinedIteration bool
1165 :
1166 : // testingAlwaysWaitForCleanup is set by some tests to force waiting for
1167 : // obsolete file deletion (to make events deterministic).
1168 : testingAlwaysWaitForCleanup bool
1169 :
1170 : // fsCloser holds a closer that should be invoked after a DB using these
1171 : // Options is closed. This is used to automatically stop the
1172 : // long-running goroutine associated with the disk-health-checking FS.
1173 : // See the initialization of FS in EnsureDefaults. Note that care has
1174 : // been taken to ensure that it is still safe to continue using the FS
1175 : // after this closer has been invoked. However, if write operations
1176 : // against the FS are made after the DB is closed, the FS may leak a
1177 : // goroutine indefinitely.
1178 : fsCloser io.Closer
1179 : }
1180 : }
1181 :
1182 : // WALFailoverOptions configures the WAL failover mechanics to use during
1183 : // transient write unavailability on the primary WAL volume.
1184 : type WALFailoverOptions struct {
1185 : // Secondary indicates the secondary directory and VFS to use in the event a
1186 : // write to the primary WAL stalls.
1187 : Secondary wal.Dir
1188 : // FailoverOptions provides configuration of the thresholds and intervals
1189 : // involved in WAL failover. If any of its fields are left unspecified,
1190 : // reasonable defaults will be used.
1191 : wal.FailoverOptions
1192 : }
1193 :
1194 : // ReadaheadConfig controls the use of read-ahead.
1195 : type ReadaheadConfig = objstorageprovider.ReadaheadConfig
1196 :
1197 : // JemallocSizeClasses exports sstable.JemallocSizeClasses.
1198 : var JemallocSizeClasses = sstable.JemallocSizeClasses
1199 :
1200 : // DebugCheckLevels calls CheckLevels on the provided database.
1201 : // It may be set in the DebugCheck field of Options to check
1202 : // level invariants whenever a new version is installed.
1203 2 : func DebugCheckLevels(db *DB) error {
1204 2 : return db.CheckLevels(nil)
1205 2 : }
1206 :
1207 : // EnsureDefaults ensures that the default values for all options are set if a
1208 : // valid value was not already specified. Returns the new options.
1209 2 : func (o *Options) EnsureDefaults() *Options {
1210 2 : if o == nil {
1211 1 : o = &Options{}
1212 1 : }
1213 2 : o.Comparer = o.Comparer.EnsureDefaults()
1214 2 :
1215 2 : if o.BytesPerSync <= 0 {
1216 1 : o.BytesPerSync = 512 << 10 // 512 KB
1217 1 : }
1218 2 : if o.Cleaner == nil {
1219 1 : o.Cleaner = DeleteCleaner{}
1220 1 : }
1221 :
1222 2 : if o.Experimental.DisableIngestAsFlushable == nil {
1223 2 : o.Experimental.DisableIngestAsFlushable = func() bool { return false }
1224 : }
1225 2 : if o.Experimental.L0CompactionConcurrency <= 0 {
1226 1 : o.Experimental.L0CompactionConcurrency = 10
1227 1 : }
1228 2 : if o.Experimental.CompactionDebtConcurrency <= 0 {
1229 1 : o.Experimental.CompactionDebtConcurrency = 1 << 30 // 1 GB
1230 1 : }
1231 2 : if o.Experimental.KeyValidationFunc == nil {
1232 2 : o.Experimental.KeyValidationFunc = func([]byte) error { return nil }
1233 : }
1234 2 : if o.KeySchema == "" && len(o.KeySchemas) == 0 {
1235 1 : ks := colblk.DefaultKeySchema(o.Comparer, 16 /* bundleSize */)
1236 1 : o.KeySchema = ks.Name
1237 1 : o.KeySchemas = sstable.MakeKeySchemas(ks)
1238 1 : }
1239 2 : if o.L0CompactionThreshold <= 0 {
1240 1 : o.L0CompactionThreshold = 4
1241 1 : }
1242 2 : if o.L0CompactionFileThreshold <= 0 {
1243 1 : // Some justification for the default of 500:
1244 1 : // Why not smaller?:
1245 1 : // - The default target file size for L0 is 2MB, so 500 files is <= 1GB
1246 1 : // of data. At observed compaction speeds of > 20MB/s, L0 can be
1247 1 : // cleared of all files in < 1min, so this backlog is not huge.
1248 1 : // - 500 files is low overhead for instantiating L0 sublevels from
1249 1 : // scratch.
1250 1 : // - Lower values were observed to cause excessive and inefficient
1251 1 : // compactions out of L0 in a TPCC import benchmark.
1252 1 : // Why not larger?:
1253 1 : // - More than 1min to compact everything out of L0.
1254 1 : // - CockroachDB's admission control system uses a threshold of 1000
1255 1 : // files to start throttling writes to Pebble. Using 500 here gives
1256 1 : // us headroom between when Pebble should start compacting L0 and
1257 1 : // when the admission control threshold is reached.
1258 1 : //
1259 1 : // We can revisit this default in the future based on better
1260 1 : // experimental understanding.
1261 1 : //
1262 1 : // TODO(jackson): Experiment with slightly lower thresholds [or higher
1263 1 : // admission control thresholds] to see whether a higher L0 score at the
1264 1 : // threshold (currently 2.0) is necessary for some workloads to avoid
1265 1 : // starving L0 in favor of lower-level compactions.
1266 1 : o.L0CompactionFileThreshold = 500
1267 1 : }
1268 2 : if o.L0StopWritesThreshold <= 0 {
1269 1 : o.L0StopWritesThreshold = 12
1270 1 : }
1271 2 : if o.LBaseMaxBytes <= 0 {
1272 1 : o.LBaseMaxBytes = 64 << 20 // 64 MB
1273 1 : }
1274 2 : if o.Levels == nil {
1275 1 : o.Levels = make([]LevelOptions, 1)
1276 1 : for i := range o.Levels {
1277 1 : if i > 0 {
1278 0 : l := &o.Levels[i]
1279 0 : if l.TargetFileSize <= 0 {
1280 0 : l.TargetFileSize = o.Levels[i-1].TargetFileSize * 2
1281 0 : }
1282 : }
1283 1 : o.Levels[i].EnsureDefaults()
1284 : }
1285 2 : } else {
1286 2 : for i := range o.Levels {
1287 2 : o.Levels[i].EnsureDefaults()
1288 2 : }
1289 : }
1290 2 : if o.Logger == nil {
1291 2 : o.Logger = DefaultLogger
1292 2 : }
1293 2 : if o.EventListener == nil {
1294 2 : o.EventListener = &EventListener{}
1295 2 : }
1296 2 : o.EventListener.EnsureDefaults(o.Logger)
1297 2 : if o.MaxManifestFileSize == 0 {
1298 1 : o.MaxManifestFileSize = 128 << 20 // 128 MB
1299 1 : }
1300 2 : if o.MaxOpenFiles == 0 {
1301 1 : o.MaxOpenFiles = 1000
1302 1 : }
1303 2 : if o.MemTableSize <= 0 {
1304 1 : o.MemTableSize = 4 << 20 // 4 MB
1305 1 : }
1306 2 : if o.MemTableStopWritesThreshold <= 0 {
1307 1 : o.MemTableStopWritesThreshold = 2
1308 1 : }
1309 2 : if o.Merger == nil {
1310 1 : o.Merger = DefaultMerger
1311 1 : }
1312 2 : if o.MaxConcurrentCompactions == nil {
1313 1 : o.MaxConcurrentCompactions = func() int { return 1 }
1314 : }
1315 2 : if o.MaxConcurrentDownloads == nil {
1316 1 : o.MaxConcurrentDownloads = func() int { return 1 }
1317 : }
1318 2 : if o.NumPrevManifest <= 0 {
1319 2 : o.NumPrevManifest = 1
1320 2 : }
1321 :
1322 2 : if o.FormatMajorVersion == FormatDefault {
1323 1 : o.FormatMajorVersion = FormatMinSupported
1324 1 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone {
1325 1 : o.FormatMajorVersion = FormatMinForSharedObjects
1326 1 : }
1327 : }
1328 :
1329 2 : if o.FS == nil {
1330 1 : o.WithFSDefaults()
1331 1 : }
1332 2 : if o.FlushSplitBytes <= 0 {
1333 1 : o.FlushSplitBytes = 2 * o.Levels[0].TargetFileSize
1334 1 : }
1335 2 : if o.WALFailover != nil {
1336 2 : o.WALFailover.FailoverOptions.EnsureDefaults()
1337 2 : }
1338 2 : if o.Experimental.LevelMultiplier <= 0 {
1339 2 : o.Experimental.LevelMultiplier = defaultLevelMultiplier
1340 2 : }
1341 2 : if o.Experimental.ReadCompactionRate == 0 {
1342 1 : o.Experimental.ReadCompactionRate = 16000
1343 1 : }
1344 2 : if o.Experimental.ReadSamplingMultiplier == 0 {
1345 1 : o.Experimental.ReadSamplingMultiplier = 1 << 4
1346 1 : }
1347 2 : if o.Experimental.NumDeletionsThreshold == 0 {
1348 1 : o.Experimental.NumDeletionsThreshold = sstable.DefaultNumDeletionsThreshold
1349 1 : }
1350 2 : if o.Experimental.DeletionSizeRatioThreshold == 0 {
1351 1 : o.Experimental.DeletionSizeRatioThreshold = sstable.DefaultDeletionSizeRatioThreshold
1352 1 : }
1353 2 : if o.Experimental.TombstoneDenseCompactionThreshold == 0 {
1354 1 : o.Experimental.TombstoneDenseCompactionThreshold = 0.05
1355 1 : }
1356 2 : if o.Experimental.TableCacheShards <= 0 {
1357 1 : o.Experimental.TableCacheShards = runtime.GOMAXPROCS(0)
1358 1 : }
1359 2 : if o.Experimental.CPUWorkPermissionGranter == nil {
1360 2 : o.Experimental.CPUWorkPermissionGranter = defaultCPUWorkGranter{}
1361 2 : }
1362 2 : if o.Experimental.MultiLevelCompactionHeuristic == nil {
1363 1 : o.Experimental.MultiLevelCompactionHeuristic = WriteAmpHeuristic{}
1364 1 : }
1365 :
1366 2 : o.initMaps()
1367 2 : return o
1368 : }
1369 :
1370 : // WithFSDefaults configures the Options to wrap the configured filesystem with
1371 : // the default virtual file system middleware, like disk-health checking.
1372 2 : func (o *Options) WithFSDefaults() *Options {
1373 2 : if o.FS == nil {
1374 1 : o.FS = vfs.Default
1375 1 : }
1376 2 : o.FS, o.private.fsCloser = vfs.WithDiskHealthChecks(o.FS, 5*time.Second, nil,
1377 2 : func(info vfs.DiskSlowInfo) {
1378 0 : o.EventListener.DiskSlow(info)
1379 0 : })
1380 2 : return o
1381 : }
1382 :
1383 : // AddEventListener adds the provided event listener to the Options, in addition
1384 : // to any existing event listener.
1385 1 : func (o *Options) AddEventListener(l EventListener) {
1386 1 : if o.EventListener != nil {
1387 1 : l = TeeEventListener(l, *o.EventListener)
1388 1 : }
1389 1 : o.EventListener = &l
1390 : }
1391 :
1392 : // initMaps initializes the Comparers, Filters, and Mergers maps.
1393 2 : func (o *Options) initMaps() {
1394 2 : for i := range o.Levels {
1395 2 : l := &o.Levels[i]
1396 2 : if l.FilterPolicy != nil {
1397 2 : if o.Filters == nil {
1398 2 : o.Filters = make(map[string]FilterPolicy)
1399 2 : }
1400 2 : name := l.FilterPolicy.Name()
1401 2 : if _, ok := o.Filters[name]; !ok {
1402 2 : o.Filters[name] = l.FilterPolicy
1403 2 : }
1404 : }
1405 : }
1406 : }
1407 :
1408 : // Level returns the LevelOptions for the specified level.
1409 2 : func (o *Options) Level(level int) LevelOptions {
1410 2 : if level < len(o.Levels) {
1411 2 : return o.Levels[level]
1412 2 : }
1413 2 : n := len(o.Levels) - 1
1414 2 : l := o.Levels[n]
1415 2 : for i := n; i < level; i++ {
1416 2 : l.TargetFileSize *= 2
1417 2 : }
1418 2 : return l
1419 : }
1420 :
1421 : // Clone creates a shallow-copy of the supplied options.
1422 2 : func (o *Options) Clone() *Options {
1423 2 : n := &Options{}
1424 2 : if o != nil {
1425 2 : *n = *o
1426 2 : }
1427 2 : return n
1428 : }
1429 :
1430 2 : func filterPolicyName(p FilterPolicy) string {
1431 2 : if p == nil {
1432 2 : return "none"
1433 2 : }
1434 2 : return p.Name()
1435 : }
1436 :
1437 2 : func (o *Options) String() string {
1438 2 : var buf bytes.Buffer
1439 2 :
1440 2 : cacheSize := int64(cacheDefaultSize)
1441 2 : if o.Cache != nil {
1442 2 : cacheSize = o.Cache.MaxSize()
1443 2 : }
1444 :
1445 2 : fmt.Fprintf(&buf, "[Version]\n")
1446 2 : fmt.Fprintf(&buf, " pebble_version=0.1\n")
1447 2 : fmt.Fprintf(&buf, "\n")
1448 2 : fmt.Fprintf(&buf, "[Options]\n")
1449 2 : fmt.Fprintf(&buf, " bytes_per_sync=%d\n", o.BytesPerSync)
1450 2 : fmt.Fprintf(&buf, " cache_size=%d\n", cacheSize)
1451 2 : fmt.Fprintf(&buf, " cleaner=%s\n", o.Cleaner)
1452 2 : fmt.Fprintf(&buf, " compaction_debt_concurrency=%d\n", o.Experimental.CompactionDebtConcurrency)
1453 2 : fmt.Fprintf(&buf, " comparer=%s\n", o.Comparer.Name)
1454 2 : fmt.Fprintf(&buf, " disable_wal=%t\n", o.DisableWAL)
1455 2 : if o.Experimental.DisableIngestAsFlushable != nil && o.Experimental.DisableIngestAsFlushable() {
1456 2 : fmt.Fprintf(&buf, " disable_ingest_as_flushable=%t\n", true)
1457 2 : }
1458 2 : if o.Experimental.EnableColumnarBlocks != nil && o.Experimental.EnableColumnarBlocks() {
1459 2 : fmt.Fprintf(&buf, " enable_columnar_blocks=%t\n", true)
1460 2 : }
1461 2 : fmt.Fprintf(&buf, " flush_delay_delete_range=%s\n", o.FlushDelayDeleteRange)
1462 2 : fmt.Fprintf(&buf, " flush_delay_range_key=%s\n", o.FlushDelayRangeKey)
1463 2 : fmt.Fprintf(&buf, " flush_split_bytes=%d\n", o.FlushSplitBytes)
1464 2 : fmt.Fprintf(&buf, " format_major_version=%d\n", o.FormatMajorVersion)
1465 2 : fmt.Fprintf(&buf, " key_schema=%s\n", o.KeySchema)
1466 2 : fmt.Fprintf(&buf, " l0_compaction_concurrency=%d\n", o.Experimental.L0CompactionConcurrency)
1467 2 : fmt.Fprintf(&buf, " l0_compaction_file_threshold=%d\n", o.L0CompactionFileThreshold)
1468 2 : fmt.Fprintf(&buf, " l0_compaction_threshold=%d\n", o.L0CompactionThreshold)
1469 2 : fmt.Fprintf(&buf, " l0_stop_writes_threshold=%d\n", o.L0StopWritesThreshold)
1470 2 : fmt.Fprintf(&buf, " lbase_max_bytes=%d\n", o.LBaseMaxBytes)
1471 2 : if o.Experimental.LevelMultiplier != defaultLevelMultiplier {
1472 2 : fmt.Fprintf(&buf, " level_multiplier=%d\n", o.Experimental.LevelMultiplier)
1473 2 : }
1474 2 : fmt.Fprintf(&buf, " max_concurrent_compactions=%d\n", o.MaxConcurrentCompactions())
1475 2 : fmt.Fprintf(&buf, " max_concurrent_downloads=%d\n", o.MaxConcurrentDownloads())
1476 2 : fmt.Fprintf(&buf, " max_manifest_file_size=%d\n", o.MaxManifestFileSize)
1477 2 : fmt.Fprintf(&buf, " max_open_files=%d\n", o.MaxOpenFiles)
1478 2 : fmt.Fprintf(&buf, " mem_table_size=%d\n", o.MemTableSize)
1479 2 : fmt.Fprintf(&buf, " mem_table_stop_writes_threshold=%d\n", o.MemTableStopWritesThreshold)
1480 2 : fmt.Fprintf(&buf, " min_deletion_rate=%d\n", o.TargetByteDeletionRate)
1481 2 : fmt.Fprintf(&buf, " merger=%s\n", o.Merger.Name)
1482 2 : if o.Experimental.MultiLevelCompactionHeuristic != nil {
1483 2 : fmt.Fprintf(&buf, " multilevel_compaction_heuristic=%s\n", o.Experimental.MultiLevelCompactionHeuristic.String())
1484 2 : }
1485 2 : fmt.Fprintf(&buf, " read_compaction_rate=%d\n", o.Experimental.ReadCompactionRate)
1486 2 : fmt.Fprintf(&buf, " read_sampling_multiplier=%d\n", o.Experimental.ReadSamplingMultiplier)
1487 2 : fmt.Fprintf(&buf, " num_deletions_threshold=%d\n", o.Experimental.NumDeletionsThreshold)
1488 2 : fmt.Fprintf(&buf, " deletion_size_ratio_threshold=%f\n", o.Experimental.DeletionSizeRatioThreshold)
1489 2 : fmt.Fprintf(&buf, " tombstone_dense_compaction_threshold=%f\n", o.Experimental.TombstoneDenseCompactionThreshold)
1490 2 : // We no longer care about strict_wal_tail, but set it to true in case an
1491 2 : // older version reads the options.
1492 2 : fmt.Fprintf(&buf, " strict_wal_tail=%t\n", true)
1493 2 : fmt.Fprintf(&buf, " table_cache_shards=%d\n", o.Experimental.TableCacheShards)
1494 2 : fmt.Fprintf(&buf, " validate_on_ingest=%t\n", o.Experimental.ValidateOnIngest)
1495 2 : fmt.Fprintf(&buf, " wal_dir=%s\n", o.WALDir)
1496 2 : fmt.Fprintf(&buf, " wal_bytes_per_sync=%d\n", o.WALBytesPerSync)
1497 2 : fmt.Fprintf(&buf, " max_writer_concurrency=%d\n", o.Experimental.MaxWriterConcurrency)
1498 2 : fmt.Fprintf(&buf, " force_writer_parallelism=%t\n", o.Experimental.ForceWriterParallelism)
1499 2 : fmt.Fprintf(&buf, " secondary_cache_size_bytes=%d\n", o.Experimental.SecondaryCacheSizeBytes)
1500 2 : fmt.Fprintf(&buf, " create_on_shared=%d\n", o.Experimental.CreateOnShared)
1501 2 :
1502 2 : // Private options.
1503 2 : //
1504 2 : // These options are only encoded if true, because we do not want them to
1505 2 : // appear in production serialized Options files, since they're testing-only
1506 2 : // options. They're only serialized when true, which still ensures that the
1507 2 : // metamorphic tests may propagate them to subprocesses.
1508 2 : if o.private.disableDeleteOnlyCompactions {
1509 2 : fmt.Fprintln(&buf, " disable_delete_only_compactions=true")
1510 2 : }
1511 2 : if o.private.disableElisionOnlyCompactions {
1512 2 : fmt.Fprintln(&buf, " disable_elision_only_compactions=true")
1513 2 : }
1514 2 : if o.private.disableLazyCombinedIteration {
1515 2 : fmt.Fprintln(&buf, " disable_lazy_combined_iteration=true")
1516 2 : }
1517 :
1518 2 : if o.WALFailover != nil {
1519 2 : unhealthyThreshold, _ := o.WALFailover.FailoverOptions.UnhealthyOperationLatencyThreshold()
1520 2 : fmt.Fprintf(&buf, "\n")
1521 2 : fmt.Fprintf(&buf, "[WAL Failover]\n")
1522 2 : fmt.Fprintf(&buf, " secondary_dir=%s\n", o.WALFailover.Secondary.Dirname)
1523 2 : fmt.Fprintf(&buf, " primary_dir_probe_interval=%s\n", o.WALFailover.FailoverOptions.PrimaryDirProbeInterval)
1524 2 : fmt.Fprintf(&buf, " healthy_probe_latency_threshold=%s\n", o.WALFailover.FailoverOptions.HealthyProbeLatencyThreshold)
1525 2 : fmt.Fprintf(&buf, " healthy_interval=%s\n", o.WALFailover.FailoverOptions.HealthyInterval)
1526 2 : fmt.Fprintf(&buf, " unhealthy_sampling_interval=%s\n", o.WALFailover.FailoverOptions.UnhealthySamplingInterval)
1527 2 : fmt.Fprintf(&buf, " unhealthy_operation_latency_threshold=%s\n", unhealthyThreshold)
1528 2 : fmt.Fprintf(&buf, " elevated_write_stall_threshold_lag=%s\n", o.WALFailover.FailoverOptions.ElevatedWriteStallThresholdLag)
1529 2 : }
1530 :
1531 2 : for i := range o.Levels {
1532 2 : l := &o.Levels[i]
1533 2 : fmt.Fprintf(&buf, "\n")
1534 2 : fmt.Fprintf(&buf, "[Level \"%d\"]\n", i)
1535 2 : fmt.Fprintf(&buf, " block_restart_interval=%d\n", l.BlockRestartInterval)
1536 2 : fmt.Fprintf(&buf, " block_size=%d\n", l.BlockSize)
1537 2 : fmt.Fprintf(&buf, " block_size_threshold=%d\n", l.BlockSizeThreshold)
1538 2 : fmt.Fprintf(&buf, " compression=%s\n", resolveDefaultCompression(l.Compression()))
1539 2 : fmt.Fprintf(&buf, " filter_policy=%s\n", filterPolicyName(l.FilterPolicy))
1540 2 : fmt.Fprintf(&buf, " filter_type=%s\n", l.FilterType)
1541 2 : fmt.Fprintf(&buf, " index_block_size=%d\n", l.IndexBlockSize)
1542 2 : fmt.Fprintf(&buf, " target_file_size=%d\n", l.TargetFileSize)
1543 2 : }
1544 :
1545 2 : return buf.String()
1546 : }
1547 :
1548 : // parseOptions takes options serialized by Options.String() and parses them into
1549 : // keys and values, calling fn for each one.
1550 2 : func parseOptions(s string, fn func(section, key, value string) error) error {
1551 2 : var section string
1552 2 : for _, line := range strings.Split(s, "\n") {
1553 2 : line = strings.TrimSpace(line)
1554 2 : if len(line) == 0 {
1555 2 : // Skip blank lines.
1556 2 : continue
1557 : }
1558 2 : if line[0] == ';' || line[0] == '#' {
1559 0 : // Skip comments.
1560 0 : continue
1561 : }
1562 2 : n := len(line)
1563 2 : if line[0] == '[' && line[n-1] == ']' {
1564 2 : // Parse section.
1565 2 : section = line[1 : n-1]
1566 2 : continue
1567 : }
1568 :
1569 2 : pos := strings.Index(line, "=")
1570 2 : if pos < 0 {
1571 1 : const maxLen = 50
1572 1 : if len(line) > maxLen {
1573 0 : line = line[:maxLen-3] + "..."
1574 0 : }
1575 1 : return base.CorruptionErrorf("invalid key=value syntax: %q", errors.Safe(line))
1576 : }
1577 :
1578 2 : key := strings.TrimSpace(line[:pos])
1579 2 : value := strings.TrimSpace(line[pos+1:])
1580 2 :
1581 2 : // RocksDB uses a similar (INI-style) syntax for the OPTIONS file, but
1582 2 : // different section names and keys. The "CFOptions ..." paths are the
1583 2 : // RocksDB versions which we map to the Pebble paths.
1584 2 : mappedSection := section
1585 2 : if section == `CFOptions "default"` {
1586 1 : mappedSection = "Options"
1587 1 : switch key {
1588 1 : case "comparator":
1589 1 : key = "comparer"
1590 1 : case "merge_operator":
1591 1 : key = "merger"
1592 : }
1593 : }
1594 :
1595 2 : if err := fn(mappedSection, key, value); err != nil {
1596 1 : return err
1597 1 : }
1598 : }
1599 2 : return nil
1600 : }
1601 :
1602 : // ParseHooks contains callbacks to create options fields which can have
1603 : // user-defined implementations.
1604 : type ParseHooks struct {
1605 : NewCache func(size int64) *Cache
1606 : NewCleaner func(name string) (Cleaner, error)
1607 : NewComparer func(name string) (*Comparer, error)
1608 : NewFilterPolicy func(name string) (FilterPolicy, error)
1609 : NewKeySchema func(name string) (KeySchema, error)
1610 : NewMerger func(name string) (*Merger, error)
1611 : SkipUnknown func(name, value string) bool
1612 : }
1613 :
1614 : // Parse parses the options from the specified string. Note that certain
1615 : // options cannot be parsed into populated fields. For example, comparer and
1616 : // merger.
1617 2 : func (o *Options) Parse(s string, hooks *ParseHooks) error {
1618 2 : return parseOptions(s, func(section, key, value string) error {
1619 2 : // WARNING: DO NOT remove entries from the switches below because doing so
1620 2 : // causes a key previously written to the OPTIONS file to be considered unknown,
1621 2 : // a backwards incompatible change. Instead, leave in support for parsing the
1622 2 : // key but simply don't parse the value.
1623 2 :
1624 2 : parseComparer := func(name string) (*Comparer, error) {
1625 2 : switch name {
1626 1 : case DefaultComparer.Name:
1627 1 : return DefaultComparer, nil
1628 2 : case testkeys.Comparer.Name:
1629 2 : return testkeys.Comparer, nil
1630 1 : default:
1631 1 : if hooks != nil && hooks.NewComparer != nil {
1632 1 : return hooks.NewComparer(name)
1633 1 : }
1634 1 : return nil, nil
1635 : }
1636 : }
1637 :
1638 2 : switch {
1639 2 : case section == "Version":
1640 2 : switch key {
1641 2 : case "pebble_version":
1642 0 : default:
1643 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1644 0 : return nil
1645 0 : }
1646 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1647 0 : errors.Safe(section), errors.Safe(key))
1648 : }
1649 2 : return nil
1650 :
1651 2 : case section == "Options":
1652 2 : var err error
1653 2 : switch key {
1654 2 : case "bytes_per_sync":
1655 2 : o.BytesPerSync, err = strconv.Atoi(value)
1656 2 : case "cache_size":
1657 2 : var n int64
1658 2 : n, err = strconv.ParseInt(value, 10, 64)
1659 2 : if err == nil && hooks != nil && hooks.NewCache != nil {
1660 2 : if o.Cache != nil {
1661 0 : o.Cache.Unref()
1662 0 : }
1663 2 : o.Cache = hooks.NewCache(n)
1664 : }
1665 : // We avoid calling cache.New in parsing because it makes it
1666 : // too easy to leak a cache.
1667 2 : case "cleaner":
1668 2 : switch value {
1669 2 : case "archive":
1670 2 : o.Cleaner = ArchiveCleaner{}
1671 1 : case "delete":
1672 1 : o.Cleaner = DeleteCleaner{}
1673 0 : default:
1674 0 : if hooks != nil && hooks.NewCleaner != nil {
1675 0 : o.Cleaner, err = hooks.NewCleaner(value)
1676 0 : }
1677 : }
1678 2 : case "comparer":
1679 2 : var comparer *Comparer
1680 2 : comparer, err = parseComparer(value)
1681 2 : if comparer != nil {
1682 2 : o.Comparer = comparer
1683 2 : }
1684 2 : case "compaction_debt_concurrency":
1685 2 : o.Experimental.CompactionDebtConcurrency, err = strconv.ParseUint(value, 10, 64)
1686 0 : case "delete_range_flush_delay":
1687 0 : // NB: This is a deprecated serialization of the
1688 0 : // `flush_delay_delete_range`.
1689 0 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1690 2 : case "disable_delete_only_compactions":
1691 2 : o.private.disableDeleteOnlyCompactions, err = strconv.ParseBool(value)
1692 2 : case "disable_elision_only_compactions":
1693 2 : o.private.disableElisionOnlyCompactions, err = strconv.ParseBool(value)
1694 2 : case "disable_ingest_as_flushable":
1695 2 : var v bool
1696 2 : v, err = strconv.ParseBool(value)
1697 2 : if err == nil {
1698 2 : o.Experimental.DisableIngestAsFlushable = func() bool { return v }
1699 : }
1700 2 : case "disable_lazy_combined_iteration":
1701 2 : o.private.disableLazyCombinedIteration, err = strconv.ParseBool(value)
1702 2 : case "disable_wal":
1703 2 : o.DisableWAL, err = strconv.ParseBool(value)
1704 2 : case "enable_columnar_blocks":
1705 2 : var v bool
1706 2 : if v, err = strconv.ParseBool(value); err == nil {
1707 2 : o.Experimental.EnableColumnarBlocks = func() bool { return v }
1708 : }
1709 2 : case "flush_delay_delete_range":
1710 2 : o.FlushDelayDeleteRange, err = time.ParseDuration(value)
1711 2 : case "flush_delay_range_key":
1712 2 : o.FlushDelayRangeKey, err = time.ParseDuration(value)
1713 2 : case "flush_split_bytes":
1714 2 : o.FlushSplitBytes, err = strconv.ParseInt(value, 10, 64)
1715 2 : case "format_major_version":
1716 2 : // NB: The version written here may be stale. Open does
1717 2 : // not use the format major version encoded in the
1718 2 : // OPTIONS file other than to validate that the encoded
1719 2 : // version is valid right here.
1720 2 : var v uint64
1721 2 : v, err = strconv.ParseUint(value, 10, 64)
1722 2 : if vers := FormatMajorVersion(v); vers > internalFormatNewest || vers == FormatDefault {
1723 0 : err = errors.Newf("unsupported format major version %d", o.FormatMajorVersion)
1724 0 : }
1725 2 : if err == nil {
1726 2 : o.FormatMajorVersion = FormatMajorVersion(v)
1727 2 : }
1728 2 : case "key_schema":
1729 2 : o.KeySchema = value
1730 2 : if o.KeySchemas == nil {
1731 2 : o.KeySchemas = make(map[string]KeySchema)
1732 2 : }
1733 2 : if _, ok := o.KeySchemas[o.KeySchema]; !ok {
1734 2 : if strings.HasPrefix(value, "DefaultKeySchema(") && strings.HasSuffix(value, ")") {
1735 2 : argsStr := strings.TrimSuffix(strings.TrimPrefix(value, "DefaultKeySchema("), ")")
1736 2 : args := strings.FieldsFunc(argsStr, func(r rune) bool {
1737 2 : return unicode.IsSpace(r) || r == ','
1738 2 : })
1739 2 : var comparer *base.Comparer
1740 2 : var bundleSize int
1741 2 : comparer, err = parseComparer(args[0])
1742 2 : if err == nil {
1743 2 : bundleSize, err = strconv.Atoi(args[1])
1744 2 : }
1745 2 : if err == nil {
1746 2 : schema := colblk.DefaultKeySchema(comparer, bundleSize)
1747 2 : o.KeySchema = schema.Name
1748 2 : o.KeySchemas[o.KeySchema] = schema
1749 2 : }
1750 0 : } else if hooks != nil && hooks.NewKeySchema != nil {
1751 0 : o.KeySchemas[value], err = hooks.NewKeySchema(value)
1752 0 : }
1753 : }
1754 2 : case "l0_compaction_concurrency":
1755 2 : o.Experimental.L0CompactionConcurrency, err = strconv.Atoi(value)
1756 2 : case "l0_compaction_file_threshold":
1757 2 : o.L0CompactionFileThreshold, err = strconv.Atoi(value)
1758 2 : case "l0_compaction_threshold":
1759 2 : o.L0CompactionThreshold, err = strconv.Atoi(value)
1760 2 : case "l0_stop_writes_threshold":
1761 2 : o.L0StopWritesThreshold, err = strconv.Atoi(value)
1762 0 : case "l0_sublevel_compactions":
1763 : // Do nothing; option existed in older versions of pebble.
1764 2 : case "lbase_max_bytes":
1765 2 : o.LBaseMaxBytes, err = strconv.ParseInt(value, 10, 64)
1766 2 : case "level_multiplier":
1767 2 : o.Experimental.LevelMultiplier, err = strconv.Atoi(value)
1768 2 : case "max_concurrent_compactions":
1769 2 : var concurrentCompactions int
1770 2 : concurrentCompactions, err = strconv.Atoi(value)
1771 2 : if concurrentCompactions <= 0 {
1772 0 : err = errors.New("max_concurrent_compactions cannot be <= 0")
1773 2 : } else {
1774 2 : o.MaxConcurrentCompactions = func() int { return concurrentCompactions }
1775 : }
1776 2 : case "max_concurrent_downloads":
1777 2 : var concurrentDownloads int
1778 2 : concurrentDownloads, err = strconv.Atoi(value)
1779 2 : if concurrentDownloads <= 0 {
1780 0 : err = errors.New("max_concurrent_compactions cannot be <= 0")
1781 2 : } else {
1782 2 : o.MaxConcurrentDownloads = func() int { return concurrentDownloads }
1783 : }
1784 2 : case "max_manifest_file_size":
1785 2 : o.MaxManifestFileSize, err = strconv.ParseInt(value, 10, 64)
1786 2 : case "max_open_files":
1787 2 : o.MaxOpenFiles, err = strconv.Atoi(value)
1788 2 : case "mem_table_size":
1789 2 : o.MemTableSize, err = strconv.ParseUint(value, 10, 64)
1790 2 : case "mem_table_stop_writes_threshold":
1791 2 : o.MemTableStopWritesThreshold, err = strconv.Atoi(value)
1792 0 : case "min_compaction_rate":
1793 : // Do nothing; option existed in older versions of pebble, and
1794 : // may be meaningful again eventually.
1795 2 : case "min_deletion_rate":
1796 2 : o.TargetByteDeletionRate, err = strconv.Atoi(value)
1797 0 : case "min_flush_rate":
1798 : // Do nothing; option existed in older versions of pebble, and
1799 : // may be meaningful again eventually.
1800 2 : case "multilevel_compaction_heuristic":
1801 2 : switch {
1802 2 : case value == "none":
1803 2 : o.Experimental.MultiLevelCompactionHeuristic = NoMultiLevel{}
1804 2 : case strings.HasPrefix(value, "wamp"):
1805 2 : fields := strings.FieldsFunc(strings.TrimPrefix(value, "wamp"), func(r rune) bool {
1806 2 : return unicode.IsSpace(r) || r == ',' || r == '(' || r == ')'
1807 2 : })
1808 2 : if len(fields) != 2 {
1809 0 : err = errors.Newf("require 2 arguments")
1810 0 : }
1811 2 : var h WriteAmpHeuristic
1812 2 : if err == nil {
1813 2 : h.AddPropensity, err = strconv.ParseFloat(fields[0], 64)
1814 2 : }
1815 2 : if err == nil {
1816 2 : h.AllowL0, err = strconv.ParseBool(fields[1])
1817 2 : }
1818 2 : if err == nil {
1819 2 : o.Experimental.MultiLevelCompactionHeuristic = h
1820 2 : } else {
1821 0 : err = errors.Wrapf(err, "unexpected wamp heuristic arguments: %s", value)
1822 0 : }
1823 0 : default:
1824 0 : err = errors.Newf("unrecognized multilevel compaction heuristic: %s", value)
1825 : }
1826 0 : case "point_tombstone_weight":
1827 : // Do nothing; deprecated.
1828 2 : case "strict_wal_tail":
1829 2 : var strictWALTail bool
1830 2 : strictWALTail, err = strconv.ParseBool(value)
1831 2 : if err == nil && !strictWALTail {
1832 0 : err = errors.Newf("reading from versions with strict_wal_tail=false no longer supported")
1833 0 : }
1834 2 : case "merger":
1835 2 : switch value {
1836 0 : case "nullptr":
1837 0 : o.Merger = nil
1838 2 : case "pebble.concatenate":
1839 2 : o.Merger = DefaultMerger
1840 1 : default:
1841 1 : if hooks != nil && hooks.NewMerger != nil {
1842 1 : o.Merger, err = hooks.NewMerger(value)
1843 1 : }
1844 : }
1845 2 : case "read_compaction_rate":
1846 2 : o.Experimental.ReadCompactionRate, err = strconv.ParseInt(value, 10, 64)
1847 2 : case "read_sampling_multiplier":
1848 2 : o.Experimental.ReadSamplingMultiplier, err = strconv.ParseInt(value, 10, 64)
1849 2 : case "num_deletions_threshold":
1850 2 : o.Experimental.NumDeletionsThreshold, err = strconv.Atoi(value)
1851 2 : case "deletion_size_ratio_threshold":
1852 2 : val, parseErr := strconv.ParseFloat(value, 32)
1853 2 : o.Experimental.DeletionSizeRatioThreshold = float32(val)
1854 2 : err = parseErr
1855 2 : case "tombstone_dense_compaction_threshold":
1856 2 : o.Experimental.TombstoneDenseCompactionThreshold, err = strconv.ParseFloat(value, 64)
1857 2 : case "table_cache_shards":
1858 2 : o.Experimental.TableCacheShards, err = strconv.Atoi(value)
1859 0 : case "table_format":
1860 0 : switch value {
1861 0 : case "leveldb":
1862 0 : case "rocksdbv2":
1863 0 : default:
1864 0 : return errors.Errorf("pebble: unknown table format: %q", errors.Safe(value))
1865 : }
1866 1 : case "table_property_collectors":
1867 : // No longer implemented; ignore.
1868 2 : case "validate_on_ingest":
1869 2 : o.Experimental.ValidateOnIngest, err = strconv.ParseBool(value)
1870 2 : case "wal_dir":
1871 2 : o.WALDir = value
1872 2 : case "wal_bytes_per_sync":
1873 2 : o.WALBytesPerSync, err = strconv.Atoi(value)
1874 2 : case "max_writer_concurrency":
1875 2 : o.Experimental.MaxWriterConcurrency, err = strconv.Atoi(value)
1876 2 : case "force_writer_parallelism":
1877 2 : o.Experimental.ForceWriterParallelism, err = strconv.ParseBool(value)
1878 2 : case "secondary_cache_size_bytes":
1879 2 : o.Experimental.SecondaryCacheSizeBytes, err = strconv.ParseInt(value, 10, 64)
1880 2 : case "create_on_shared":
1881 2 : var createOnSharedInt int64
1882 2 : createOnSharedInt, err = strconv.ParseInt(value, 10, 64)
1883 2 : o.Experimental.CreateOnShared = remote.CreateOnSharedStrategy(createOnSharedInt)
1884 0 : default:
1885 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1886 0 : return nil
1887 0 : }
1888 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1889 0 : errors.Safe(section), errors.Safe(key))
1890 : }
1891 2 : return err
1892 :
1893 2 : case section == "WAL Failover":
1894 2 : if o.WALFailover == nil {
1895 2 : o.WALFailover = new(WALFailoverOptions)
1896 2 : }
1897 2 : var err error
1898 2 : switch key {
1899 2 : case "secondary_dir":
1900 2 : o.WALFailover.Secondary = wal.Dir{Dirname: value, FS: vfs.Default}
1901 2 : case "primary_dir_probe_interval":
1902 2 : o.WALFailover.PrimaryDirProbeInterval, err = time.ParseDuration(value)
1903 2 : case "healthy_probe_latency_threshold":
1904 2 : o.WALFailover.HealthyProbeLatencyThreshold, err = time.ParseDuration(value)
1905 2 : case "healthy_interval":
1906 2 : o.WALFailover.HealthyInterval, err = time.ParseDuration(value)
1907 2 : case "unhealthy_sampling_interval":
1908 2 : o.WALFailover.UnhealthySamplingInterval, err = time.ParseDuration(value)
1909 2 : case "unhealthy_operation_latency_threshold":
1910 2 : var threshold time.Duration
1911 2 : threshold, err = time.ParseDuration(value)
1912 2 : o.WALFailover.UnhealthyOperationLatencyThreshold = func() (time.Duration, bool) {
1913 2 : return threshold, true
1914 2 : }
1915 2 : case "elevated_write_stall_threshold_lag":
1916 2 : o.WALFailover.ElevatedWriteStallThresholdLag, err = time.ParseDuration(value)
1917 0 : default:
1918 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1919 0 : return nil
1920 0 : }
1921 0 : return errors.Errorf("pebble: unknown option: %s.%s",
1922 0 : errors.Safe(section), errors.Safe(key))
1923 : }
1924 2 : return err
1925 :
1926 2 : case strings.HasPrefix(section, "Level "):
1927 2 : var index int
1928 2 : if n, err := fmt.Sscanf(section, `Level "%d"`, &index); err != nil {
1929 0 : return err
1930 2 : } else if n != 1 {
1931 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section, value) {
1932 0 : return nil
1933 0 : }
1934 0 : return errors.Errorf("pebble: unknown section: %q", errors.Safe(section))
1935 : }
1936 :
1937 2 : if len(o.Levels) <= index {
1938 1 : newLevels := make([]LevelOptions, index+1)
1939 1 : copy(newLevels, o.Levels)
1940 1 : o.Levels = newLevels
1941 1 : }
1942 2 : l := &o.Levels[index]
1943 2 :
1944 2 : var err error
1945 2 : switch key {
1946 2 : case "block_restart_interval":
1947 2 : l.BlockRestartInterval, err = strconv.Atoi(value)
1948 2 : case "block_size":
1949 2 : l.BlockSize, err = strconv.Atoi(value)
1950 2 : case "block_size_threshold":
1951 2 : l.BlockSizeThreshold, err = strconv.Atoi(value)
1952 2 : case "compression":
1953 2 : switch value {
1954 0 : case "Default":
1955 0 : l.Compression = func() Compression { return DefaultCompression }
1956 2 : case "NoCompression":
1957 2 : l.Compression = func() Compression { return NoCompression }
1958 2 : case "Snappy":
1959 2 : l.Compression = func() Compression { return SnappyCompression }
1960 2 : case "ZSTD":
1961 2 : l.Compression = func() Compression { return ZstdCompression }
1962 0 : default:
1963 0 : return errors.Errorf("pebble: unknown compression: %q", errors.Safe(value))
1964 : }
1965 2 : case "filter_policy":
1966 2 : if hooks != nil && hooks.NewFilterPolicy != nil {
1967 2 : l.FilterPolicy, err = hooks.NewFilterPolicy(value)
1968 2 : }
1969 2 : case "filter_type":
1970 2 : switch value {
1971 2 : case "table":
1972 2 : l.FilterType = TableFilter
1973 0 : default:
1974 0 : return errors.Errorf("pebble: unknown filter type: %q", errors.Safe(value))
1975 : }
1976 2 : case "index_block_size":
1977 2 : l.IndexBlockSize, err = strconv.Atoi(value)
1978 2 : case "target_file_size":
1979 2 : l.TargetFileSize, err = strconv.ParseInt(value, 10, 64)
1980 0 : default:
1981 0 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1982 0 : return nil
1983 0 : }
1984 0 : return errors.Errorf("pebble: unknown option: %s.%s", errors.Safe(section), errors.Safe(key))
1985 : }
1986 2 : return err
1987 : }
1988 2 : if hooks != nil && hooks.SkipUnknown != nil && hooks.SkipUnknown(section+"."+key, value) {
1989 2 : return nil
1990 2 : }
1991 0 : return errors.Errorf("pebble: unknown section: %q", errors.Safe(section))
1992 : })
1993 : }
1994 :
1995 : // ErrMissingWALRecoveryDir is an error returned when a database is attempted to be
1996 : // opened without supplying a Options.WALRecoveryDir entry for a directory that
1997 : // may contain WALs required to recover a consistent database state.
1998 : type ErrMissingWALRecoveryDir struct {
1999 : Dir string
2000 : }
2001 :
2002 : // Error implements error.
2003 1 : func (e ErrMissingWALRecoveryDir) Error() string {
2004 1 : return fmt.Sprintf("directory %q may contain relevant WALs", e.Dir)
2005 1 : }
2006 :
2007 : // CheckCompatibility verifies the options are compatible with the previous options
2008 : // serialized by Options.String(). For example, the Comparer and Merger must be
2009 : // the same, or data will not be able to be properly read from the DB.
2010 : //
2011 : // This function only looks at specific keys and does not error out if the
2012 : // options are newer and contain unknown keys.
2013 2 : func (o *Options) CheckCompatibility(previousOptions string) error {
2014 2 : return parseOptions(previousOptions, func(section, key, value string) error {
2015 2 : switch section + "." + key {
2016 2 : case "Options.comparer":
2017 2 : if value != o.Comparer.Name {
2018 1 : return errors.Errorf("pebble: comparer name from file %q != comparer name from options %q",
2019 1 : errors.Safe(value), errors.Safe(o.Comparer.Name))
2020 1 : }
2021 2 : case "Options.merger":
2022 2 : // RocksDB allows the merge operator to be unspecified, in which case it
2023 2 : // shows up as "nullptr".
2024 2 : if value != "nullptr" && value != o.Merger.Name {
2025 1 : return errors.Errorf("pebble: merger name from file %q != merger name from options %q",
2026 1 : errors.Safe(value), errors.Safe(o.Merger.Name))
2027 1 : }
2028 2 : case "Options.wal_dir", "WAL Failover.secondary_dir":
2029 2 : switch {
2030 2 : case o.WALDir == value:
2031 2 : return nil
2032 2 : case o.WALFailover != nil && o.WALFailover.Secondary.Dirname == value:
2033 2 : return nil
2034 1 : default:
2035 1 : for _, d := range o.WALRecoveryDirs {
2036 1 : if d.Dirname == value {
2037 1 : return nil
2038 1 : }
2039 : }
2040 1 : return ErrMissingWALRecoveryDir{Dir: value}
2041 : }
2042 : }
2043 2 : return nil
2044 : })
2045 : }
2046 :
2047 : // Validate verifies that the options are mutually consistent. For example,
2048 : // L0StopWritesThreshold must be >= L0CompactionThreshold, otherwise a write
2049 : // stall would persist indefinitely.
2050 2 : func (o *Options) Validate() error {
2051 2 : // Note that we can presume Options.EnsureDefaults has been called, so there
2052 2 : // is no need to check for zero values.
2053 2 :
2054 2 : var buf strings.Builder
2055 2 : if o.Experimental.L0CompactionConcurrency < 1 {
2056 1 : fmt.Fprintf(&buf, "L0CompactionConcurrency (%d) must be >= 1\n",
2057 1 : o.Experimental.L0CompactionConcurrency)
2058 1 : }
2059 2 : if o.L0StopWritesThreshold < o.L0CompactionThreshold {
2060 1 : fmt.Fprintf(&buf, "L0StopWritesThreshold (%d) must be >= L0CompactionThreshold (%d)\n",
2061 1 : o.L0StopWritesThreshold, o.L0CompactionThreshold)
2062 1 : }
2063 2 : if uint64(o.MemTableSize) >= maxMemTableSize {
2064 1 : fmt.Fprintf(&buf, "MemTableSize (%s) must be < %s\n",
2065 1 : humanize.Bytes.Uint64(uint64(o.MemTableSize)), humanize.Bytes.Uint64(maxMemTableSize))
2066 1 : }
2067 2 : if o.MemTableStopWritesThreshold < 2 {
2068 1 : fmt.Fprintf(&buf, "MemTableStopWritesThreshold (%d) must be >= 2\n",
2069 1 : o.MemTableStopWritesThreshold)
2070 1 : }
2071 2 : if o.FormatMajorVersion < FormatMinSupported || o.FormatMajorVersion > internalFormatNewest {
2072 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) must be between %d and %d\n",
2073 0 : o.FormatMajorVersion, FormatMinSupported, internalFormatNewest)
2074 0 : }
2075 2 : if o.Experimental.CreateOnShared != remote.CreateOnSharedNone && o.FormatMajorVersion < FormatMinForSharedObjects {
2076 0 : fmt.Fprintf(&buf, "FormatMajorVersion (%d) when CreateOnShared is set must be at least %d\n",
2077 0 : o.FormatMajorVersion, FormatMinForSharedObjects)
2078 0 : }
2079 2 : if o.TableCache != nil && o.Cache != o.TableCache.cache {
2080 1 : fmt.Fprintf(&buf, "underlying cache in the TableCache and the Cache dont match\n")
2081 1 : }
2082 2 : if len(o.KeySchemas) > 0 {
2083 2 : if o.KeySchema == "" {
2084 0 : fmt.Fprintf(&buf, "KeySchemas is set but KeySchema is not\n")
2085 0 : }
2086 2 : if _, ok := o.KeySchemas[o.KeySchema]; !ok {
2087 0 : fmt.Fprintf(&buf, "KeySchema %q not found in KeySchemas\n", o.KeySchema)
2088 0 : }
2089 : }
2090 2 : if buf.Len() == 0 {
2091 2 : return nil
2092 2 : }
2093 1 : return errors.New(buf.String())
2094 : }
2095 :
2096 : // MakeReaderOptions constructs sstable.ReaderOptions from the corresponding
2097 : // options in the receiver.
2098 2 : func (o *Options) MakeReaderOptions() sstable.ReaderOptions {
2099 2 : var readerOpts sstable.ReaderOptions
2100 2 : if o != nil {
2101 2 : readerOpts.Comparer = o.Comparer
2102 2 : readerOpts.Filters = o.Filters
2103 2 : readerOpts.KeySchemas = o.KeySchemas
2104 2 : readerOpts.LoadBlockSema = o.LoadBlockSema
2105 2 : readerOpts.LoggerAndTracer = o.LoggerAndTracer
2106 2 : readerOpts.Merger = o.Merger
2107 2 : }
2108 2 : return readerOpts
2109 : }
2110 :
2111 : // MakeWriterOptions constructs sstable.WriterOptions for the specified level
2112 : // from the corresponding options in the receiver.
2113 2 : func (o *Options) MakeWriterOptions(level int, format sstable.TableFormat) sstable.WriterOptions {
2114 2 : var writerOpts sstable.WriterOptions
2115 2 : writerOpts.TableFormat = format
2116 2 : if o != nil {
2117 2 : writerOpts.Comparer = o.Comparer
2118 2 : if o.Merger != nil {
2119 2 : writerOpts.MergerName = o.Merger.Name
2120 2 : }
2121 2 : writerOpts.BlockPropertyCollectors = o.BlockPropertyCollectors
2122 : }
2123 2 : if format >= sstable.TableFormatPebblev3 {
2124 2 : writerOpts.ShortAttributeExtractor = o.Experimental.ShortAttributeExtractor
2125 2 : writerOpts.RequiredInPlaceValueBound = o.Experimental.RequiredInPlaceValueBound
2126 2 : if format >= sstable.TableFormatPebblev4 && level == numLevels-1 {
2127 2 : writerOpts.WritingToLowestLevel = true
2128 2 : }
2129 : }
2130 2 : levelOpts := o.Level(level)
2131 2 : writerOpts.BlockRestartInterval = levelOpts.BlockRestartInterval
2132 2 : writerOpts.BlockSize = levelOpts.BlockSize
2133 2 : writerOpts.BlockSizeThreshold = levelOpts.BlockSizeThreshold
2134 2 : writerOpts.Compression = resolveDefaultCompression(levelOpts.Compression())
2135 2 : writerOpts.FilterPolicy = levelOpts.FilterPolicy
2136 2 : writerOpts.FilterType = levelOpts.FilterType
2137 2 : writerOpts.IndexBlockSize = levelOpts.IndexBlockSize
2138 2 : writerOpts.KeySchema = o.KeySchemas[o.KeySchema]
2139 2 : writerOpts.AllocatorSizeClasses = o.AllocatorSizeClasses
2140 2 : writerOpts.NumDeletionsThreshold = o.Experimental.NumDeletionsThreshold
2141 2 : writerOpts.DeletionSizeRatioThreshold = o.Experimental.DeletionSizeRatioThreshold
2142 2 : return writerOpts
2143 : }
2144 :
2145 2 : func resolveDefaultCompression(c Compression) Compression {
2146 2 : if c <= DefaultCompression || c >= block.NCompression {
2147 1 : c = SnappyCompression
2148 1 : }
2149 2 : return c
2150 : }
|
