Line data Source code
1 : // Copyright 2012 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 provides an ordered key/value store.
6 : package pebble // import "github.com/cockroachdb/pebble"
7 :
8 : import (
9 : "context"
10 : "fmt"
11 : "io"
12 : "os"
13 : "strconv"
14 : "sync"
15 : "sync/atomic"
16 : "time"
17 :
18 : "github.com/cockroachdb/errors"
19 : "github.com/cockroachdb/pebble/internal/arenaskl"
20 : "github.com/cockroachdb/pebble/internal/base"
21 : "github.com/cockroachdb/pebble/internal/invalidating"
22 : "github.com/cockroachdb/pebble/internal/invariants"
23 : "github.com/cockroachdb/pebble/internal/keyspan"
24 : "github.com/cockroachdb/pebble/internal/manifest"
25 : "github.com/cockroachdb/pebble/internal/manual"
26 : "github.com/cockroachdb/pebble/objstorage"
27 : "github.com/cockroachdb/pebble/objstorage/remote"
28 : "github.com/cockroachdb/pebble/rangekey"
29 : "github.com/cockroachdb/pebble/record"
30 : "github.com/cockroachdb/pebble/sstable"
31 : "github.com/cockroachdb/pebble/vfs"
32 : "github.com/cockroachdb/pebble/vfs/atomicfs"
33 : "github.com/cockroachdb/tokenbucket"
34 : "github.com/prometheus/client_golang/prometheus"
35 : )
36 :
37 : const (
38 : // minTableCacheSize is the minimum size of the table cache, for a single db.
39 : minTableCacheSize = 64
40 :
41 : // numNonTableCacheFiles is an approximation for the number of files
42 : // that we don't use for table caches, for a given db.
43 : numNonTableCacheFiles = 10
44 : )
45 :
46 : var (
47 : // ErrNotFound is returned when a get operation does not find the requested
48 : // key.
49 : ErrNotFound = base.ErrNotFound
50 : // ErrClosed is panicked when an operation is performed on a closed snapshot or
51 : // DB. Use errors.Is(err, ErrClosed) to check for this error.
52 : ErrClosed = errors.New("pebble: closed")
53 : // ErrReadOnly is returned when a write operation is performed on a read-only
54 : // database.
55 : ErrReadOnly = errors.New("pebble: read-only")
56 : // errNoSplit indicates that the user is trying to perform a range key
57 : // operation but the configured Comparer does not provide a Split
58 : // implementation.
59 : errNoSplit = errors.New("pebble: Comparer.Split required for range key operations")
60 : )
61 :
62 : // Reader is a readable key/value store.
63 : //
64 : // It is safe to call Get and NewIter from concurrent goroutines.
65 : type Reader interface {
66 : // Get gets the value for the given key. It returns ErrNotFound if the DB
67 : // does not contain the key.
68 : //
69 : // The caller should not modify the contents of the returned slice, but it is
70 : // safe to modify the contents of the argument after Get returns. The
71 : // returned slice will remain valid until the returned Closer is closed. On
72 : // success, the caller MUST call closer.Close() or a memory leak will occur.
73 : Get(key []byte) (value []byte, closer io.Closer, err error)
74 :
75 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
76 : // return false). The iterator can be positioned via a call to SeekGE,
77 : // SeekLT, First or Last.
78 : NewIter(o *IterOptions) (*Iterator, error)
79 :
80 : // NewIterWithContext is like NewIter, and additionally accepts a context
81 : // for tracing.
82 : NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error)
83 :
84 : // Close closes the Reader. It may or may not close any underlying io.Reader
85 : // or io.Writer, depending on how the DB was created.
86 : //
87 : // It is not safe to close a DB until all outstanding iterators are closed.
88 : // It is valid to call Close multiple times. Other methods should not be
89 : // called after the DB has been closed.
90 : Close() error
91 : }
92 :
93 : // Writer is a writable key/value store.
94 : //
95 : // Goroutine safety is dependent on the specific implementation.
96 : type Writer interface {
97 : // Apply the operations contained in the batch to the DB.
98 : //
99 : // It is safe to modify the contents of the arguments after Apply returns.
100 : Apply(batch *Batch, o *WriteOptions) error
101 :
102 : // Delete deletes the value for the given key. Deletes are blind all will
103 : // succeed even if the given key does not exist.
104 : //
105 : // It is safe to modify the contents of the arguments after Delete returns.
106 : Delete(key []byte, o *WriteOptions) error
107 :
108 : // DeleteSized behaves identically to Delete, but takes an additional
109 : // argument indicating the size of the value being deleted. DeleteSized
110 : // should be preferred when the caller has the expectation that there exists
111 : // a single internal KV pair for the key (eg, the key has not been
112 : // overwritten recently), and the caller knows the size of its value.
113 : //
114 : // DeleteSized will record the value size within the tombstone and use it to
115 : // inform compaction-picking heuristics which strive to reduce space
116 : // amplification in the LSM. This "calling your shot" mechanic allows the
117 : // storage engine to more accurately estimate and reduce space
118 : // amplification.
119 : //
120 : // It is safe to modify the contents of the arguments after DeleteSized
121 : // returns.
122 : DeleteSized(key []byte, valueSize uint32, _ *WriteOptions) error
123 :
124 : // SingleDelete is similar to Delete in that it deletes the value for the given key. Like Delete,
125 : // it is a blind operation that will succeed even if the given key does not exist.
126 : //
127 : // WARNING: Undefined (non-deterministic) behavior will result if a key is overwritten and
128 : // then deleted using SingleDelete. The record may appear deleted immediately, but be
129 : // resurrected at a later time after compactions have been performed. Or the record may
130 : // be deleted permanently. A Delete operation lays down a "tombstone" which shadows all
131 : // previous versions of a key. The SingleDelete operation is akin to "anti-matter" and will
132 : // only delete the most recently written version for a key. These different semantics allow
133 : // the DB to avoid propagating a SingleDelete operation during a compaction as soon as the
134 : // corresponding Set operation is encountered. These semantics require extreme care to handle
135 : // properly. Only use if you have a workload where the performance gain is critical and you
136 : // can guarantee that a record is written once and then deleted once.
137 : //
138 : // SingleDelete is internally transformed into a Delete if the most recent record for a key is either
139 : // a Merge or Delete record.
140 : //
141 : // It is safe to modify the contents of the arguments after SingleDelete returns.
142 : SingleDelete(key []byte, o *WriteOptions) error
143 :
144 : // DeleteRange deletes all of the point keys (and values) in the range
145 : // [start,end) (inclusive on start, exclusive on end). DeleteRange does NOT
146 : // delete overlapping range keys (eg, keys set via RangeKeySet).
147 : //
148 : // It is safe to modify the contents of the arguments after DeleteRange
149 : // returns.
150 : DeleteRange(start, end []byte, o *WriteOptions) error
151 :
152 : // LogData adds the specified to the batch. The data will be written to the
153 : // WAL, but not added to memtables or sstables. Log data is never indexed,
154 : // which makes it useful for testing WAL performance.
155 : //
156 : // It is safe to modify the contents of the argument after LogData returns.
157 : LogData(data []byte, opts *WriteOptions) error
158 :
159 : // Merge merges the value for the given key. The details of the merge are
160 : // dependent upon the configured merge operation.
161 : //
162 : // It is safe to modify the contents of the arguments after Merge returns.
163 : Merge(key, value []byte, o *WriteOptions) error
164 :
165 : // Set sets the value for the given key. It overwrites any previous value
166 : // for that key; a DB is not a multi-map.
167 : //
168 : // It is safe to modify the contents of the arguments after Set returns.
169 : Set(key, value []byte, o *WriteOptions) error
170 :
171 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
172 : // timestamp suffix to value. The suffix is optional. If any portion of the key
173 : // range [start, end) is already set by a range key with the same suffix value,
174 : // RangeKeySet overrides it.
175 : //
176 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
177 : RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error
178 :
179 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
180 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
181 : // range key. RangeKeyUnset only removes portions of range keys that fall within
182 : // the [start, end) key span, and only range keys with suffixes that exactly
183 : // match the unset suffix.
184 : //
185 : // It is safe to modify the contents of the arguments after RangeKeyUnset
186 : // returns.
187 : RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error
188 :
189 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
190 : // (inclusive on start, exclusive on end). It does not delete point keys (for
191 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
192 : // bounds, including those with or without suffixes.
193 : //
194 : // It is safe to modify the contents of the arguments after RangeKeyDelete
195 : // returns.
196 : RangeKeyDelete(start, end []byte, opts *WriteOptions) error
197 : }
198 :
199 : // CPUWorkHandle represents a handle used by the CPUWorkPermissionGranter API.
200 : type CPUWorkHandle interface {
201 : // Permitted indicates whether Pebble can use additional CPU resources.
202 : Permitted() bool
203 : }
204 :
205 : // CPUWorkPermissionGranter is used to request permission to opportunistically
206 : // use additional CPUs to speed up internal background work.
207 : type CPUWorkPermissionGranter interface {
208 : // GetPermission returns a handle regardless of whether permission is granted
209 : // or not. In the latter case, the handle is only useful for recording
210 : // the CPU time actually spent on this calling goroutine.
211 : GetPermission(time.Duration) CPUWorkHandle
212 : // CPUWorkDone must be called regardless of whether CPUWorkHandle.Permitted
213 : // returns true or false.
214 : CPUWorkDone(CPUWorkHandle)
215 : }
216 :
217 : // Use a default implementation for the CPU work granter to avoid excessive nil
218 : // checks in the code.
219 : type defaultCPUWorkHandle struct{}
220 :
221 1 : func (d defaultCPUWorkHandle) Permitted() bool {
222 1 : return false
223 1 : }
224 :
225 : type defaultCPUWorkGranter struct{}
226 :
227 2 : func (d defaultCPUWorkGranter) GetPermission(_ time.Duration) CPUWorkHandle {
228 2 : return defaultCPUWorkHandle{}
229 2 : }
230 :
231 2 : func (d defaultCPUWorkGranter) CPUWorkDone(_ CPUWorkHandle) {}
232 :
233 : // DB provides a concurrent, persistent ordered key/value store.
234 : //
235 : // A DB's basic operations (Get, Set, Delete) should be self-explanatory. Get
236 : // and Delete will return ErrNotFound if the requested key is not in the store.
237 : // Callers are free to ignore this error.
238 : //
239 : // A DB also allows for iterating over the key/value pairs in key order. If d
240 : // is a DB, the code below prints all key/value pairs whose keys are 'greater
241 : // than or equal to' k:
242 : //
243 : // iter := d.NewIter(readOptions)
244 : // for iter.SeekGE(k); iter.Valid(); iter.Next() {
245 : // fmt.Printf("key=%q value=%q\n", iter.Key(), iter.Value())
246 : // }
247 : // return iter.Close()
248 : //
249 : // The Options struct holds the optional parameters for the DB, including a
250 : // Comparer to define a 'less than' relationship over keys. It is always valid
251 : // to pass a nil *Options, which means to use the default parameter values. Any
252 : // zero field of a non-nil *Options also means to use the default value for
253 : // that parameter. Thus, the code below uses a custom Comparer, but the default
254 : // values for every other parameter:
255 : //
256 : // db := pebble.Open(&Options{
257 : // Comparer: myComparer,
258 : // })
259 : type DB struct {
260 : // The count and size of referenced memtables. This includes memtables
261 : // present in DB.mu.mem.queue, as well as memtables that have been flushed
262 : // but are still referenced by an inuse readState, as well as up to one
263 : // memTable waiting to be reused and stored in d.memTableRecycle.
264 : memTableCount atomic.Int64
265 : memTableReserved atomic.Int64 // number of bytes reserved in the cache for memtables
266 : // memTableRecycle holds a pointer to an obsolete memtable. The next
267 : // memtable allocation will reuse this memtable if it has not already been
268 : // recycled.
269 : memTableRecycle atomic.Pointer[memTable]
270 :
271 : // The size of the current log file (i.e. db.mu.log.queue[len(queue)-1].
272 : logSize atomic.Uint64
273 :
274 : // The number of bytes available on disk.
275 : diskAvailBytes atomic.Uint64
276 :
277 : cacheID uint64
278 : dirname string
279 : walDirname string
280 : opts *Options
281 : cmp Compare
282 : equal Equal
283 : merge Merge
284 : split Split
285 : abbreviatedKey AbbreviatedKey
286 : // The threshold for determining when a batch is "large" and will skip being
287 : // inserted into a memtable.
288 : largeBatchThreshold uint64
289 : // The current OPTIONS file number.
290 : optionsFileNum base.DiskFileNum
291 : // The on-disk size of the current OPTIONS file.
292 : optionsFileSize uint64
293 :
294 : // objProvider is used to access and manage SSTs.
295 : objProvider objstorage.Provider
296 :
297 : fileLock *Lock
298 : dataDir vfs.File
299 : walDir vfs.File
300 :
301 : tableCache *tableCacheContainer
302 : newIters tableNewIters
303 : tableNewRangeKeyIter keyspan.TableNewSpanIter
304 :
305 : commit *commitPipeline
306 :
307 : // readState provides access to the state needed for reading without needing
308 : // to acquire DB.mu.
309 : readState struct {
310 : sync.RWMutex
311 : val *readState
312 : }
313 : // logRecycler holds a set of log file numbers that are available for
314 : // reuse. Writing to a recycled log file is faster than to a new log file on
315 : // some common filesystems (xfs, and ext3/4) due to avoiding metadata
316 : // updates.
317 : logRecycler logRecycler
318 :
319 : closed *atomic.Value
320 : closedCh chan struct{}
321 :
322 : cleanupManager *cleanupManager
323 :
324 : // During an iterator close, we may asynchronously schedule read compactions.
325 : // We want to wait for those goroutines to finish, before closing the DB.
326 : // compactionShedulers.Wait() should not be called while the DB.mu is held.
327 : compactionSchedulers sync.WaitGroup
328 :
329 : // The main mutex protecting internal DB state. This mutex encompasses many
330 : // fields because those fields need to be accessed and updated atomically. In
331 : // particular, the current version, log.*, mem.*, and snapshot list need to
332 : // be accessed and updated atomically during compaction.
333 : //
334 : // Care is taken to avoid holding DB.mu during IO operations. Accomplishing
335 : // this sometimes requires releasing DB.mu in a method that was called with
336 : // it held. See versionSet.logAndApply() and DB.makeRoomForWrite() for
337 : // examples. This is a common pattern, so be careful about expectations that
338 : // DB.mu will be held continuously across a set of calls.
339 : mu struct {
340 : sync.Mutex
341 :
342 : formatVers struct {
343 : // vers is the database's current format major version.
344 : // Backwards-incompatible features are gated behind new
345 : // format major versions and not enabled until a database's
346 : // version is ratcheted upwards.
347 : //
348 : // Although this is under the `mu` prefix, readers may read vers
349 : // atomically without holding d.mu. Writers must only write to this
350 : // value through finalizeFormatVersUpgrade which requires d.mu is
351 : // held.
352 : vers atomic.Uint64
353 : // marker is the atomic marker for the format major version.
354 : // When a database's version is ratcheted upwards, the
355 : // marker is moved in order to atomically record the new
356 : // version.
357 : marker *atomicfs.Marker
358 : // ratcheting when set to true indicates that the database is
359 : // currently in the process of ratcheting the format major version
360 : // to vers + 1. As a part of ratcheting the format major version,
361 : // migrations may drop and re-acquire the mutex.
362 : ratcheting bool
363 : }
364 :
365 : // The ID of the next job. Job IDs are passed to event listener
366 : // notifications and act as a mechanism for tying together the events and
367 : // log messages for a single job such as a flush, compaction, or file
368 : // ingestion. Job IDs are not serialized to disk or used for correctness.
369 : nextJobID int
370 :
371 : // The collection of immutable versions and state about the log and visible
372 : // sequence numbers. Use the pointer here to ensure the atomic fields in
373 : // version set are aligned properly.
374 : versions *versionSet
375 :
376 : log struct {
377 : // The queue of logs, containing both flushed and unflushed logs. The
378 : // flushed logs will be a prefix, the unflushed logs a suffix. The
379 : // delimeter between flushed and unflushed logs is
380 : // versionSet.minUnflushedLogNum.
381 : queue []fileInfo
382 : // The number of input bytes to the log. This is the raw size of the
383 : // batches written to the WAL, without the overhead of the record
384 : // envelopes.
385 : bytesIn uint64
386 : // The LogWriter is protected by commitPipeline.mu. This allows log
387 : // writes to be performed without holding DB.mu, but requires both
388 : // commitPipeline.mu and DB.mu to be held when rotating the WAL/memtable
389 : // (i.e. makeRoomForWrite).
390 : *record.LogWriter
391 : // Can be nil.
392 : metrics struct {
393 : fsyncLatency prometheus.Histogram
394 : record.LogWriterMetrics
395 : }
396 : registerLogWriterForTesting func(w *record.LogWriter)
397 : }
398 :
399 : mem struct {
400 : // The current mutable memTable.
401 : mutable *memTable
402 : // Queue of flushables (the mutable memtable is at end). Elements are
403 : // added to the end of the slice and removed from the beginning. Once an
404 : // index is set it is never modified making a fixed slice immutable and
405 : // safe for concurrent reads.
406 : queue flushableList
407 : // nextSize is the size of the next memtable. The memtable size starts at
408 : // min(256KB,Options.MemTableSize) and doubles each time a new memtable
409 : // is allocated up to Options.MemTableSize. This reduces the memory
410 : // footprint of memtables when lots of DB instances are used concurrently
411 : // in test environments.
412 : nextSize uint64
413 : }
414 :
415 : compact struct {
416 : // Condition variable used to signal when a flush or compaction has
417 : // completed. Used by the write-stall mechanism to wait for the stall
418 : // condition to clear. See DB.makeRoomForWrite().
419 : cond sync.Cond
420 : // True when a flush is in progress.
421 : flushing bool
422 : // The number of ongoing compactions.
423 : compactingCount int
424 : // The list of deletion hints, suggesting ranges for delete-only
425 : // compactions.
426 : deletionHints []deleteCompactionHint
427 : // The list of manual compactions. The next manual compaction to perform
428 : // is at the start of the list. New entries are added to the end.
429 : manual []*manualCompaction
430 : // inProgress is the set of in-progress flushes and compactions.
431 : // It's used in the calculation of some metrics and to initialize L0
432 : // sublevels' state. Some of the compactions contained within this
433 : // map may have already committed an edit to the version but are
434 : // lingering performing cleanup, like deleting obsolete files.
435 : inProgress map[*compaction]struct{}
436 :
437 : // rescheduleReadCompaction indicates to an iterator that a read compaction
438 : // should be scheduled.
439 : rescheduleReadCompaction bool
440 :
441 : // readCompactions is a readCompactionQueue which keeps track of the
442 : // compactions which we might have to perform.
443 : readCompactions readCompactionQueue
444 :
445 : // The cumulative duration of all completed compactions since Open.
446 : // Does not include flushes.
447 : duration time.Duration
448 : // Flush throughput metric.
449 : flushWriteThroughput ThroughputMetric
450 : // The idle start time for the flush "loop", i.e., when the flushing
451 : // bool above transitions to false.
452 : noOngoingFlushStartTime time.Time
453 : }
454 :
455 : // Non-zero when file cleaning is disabled. The disabled count acts as a
456 : // reference count to prohibit file cleaning. See
457 : // DB.{disable,Enable}FileDeletions().
458 : disableFileDeletions int
459 :
460 : snapshots struct {
461 : // The list of active snapshots.
462 : snapshotList
463 :
464 : // The cumulative count and size of snapshot-pinned keys written to
465 : // sstables.
466 : cumulativePinnedCount uint64
467 : cumulativePinnedSize uint64
468 : }
469 :
470 : tableStats struct {
471 : // Condition variable used to signal the completion of a
472 : // job to collect table stats.
473 : cond sync.Cond
474 : // True when a stat collection operation is in progress.
475 : loading bool
476 : // True if stat collection has loaded statistics for all tables
477 : // other than those listed explicitly in pending. This flag starts
478 : // as false when a database is opened and flips to true once stat
479 : // collection has caught up.
480 : loadedInitial bool
481 : // A slice of files for which stats have not been computed.
482 : // Compactions, ingests, flushes append files to be processed. An
483 : // active stat collection goroutine clears the list and processes
484 : // them.
485 : pending []manifest.NewFileEntry
486 : }
487 :
488 : tableValidation struct {
489 : // cond is a condition variable used to signal the completion of a
490 : // job to validate one or more sstables.
491 : cond sync.Cond
492 : // pending is a slice of metadata for sstables waiting to be
493 : // validated. Only physical sstables should be added to the pending
494 : // queue.
495 : pending []newFileEntry
496 : // validating is set to true when validation is running.
497 : validating bool
498 : }
499 : }
500 :
501 : // Normally equal to time.Now() but may be overridden in tests.
502 : timeNow func() time.Time
503 : // the time at database Open; may be used to compute metrics like effective
504 : // compaction concurrency
505 : openedAt time.Time
506 : }
507 :
508 : var _ Reader = (*DB)(nil)
509 : var _ Writer = (*DB)(nil)
510 :
511 : // TestOnlyWaitForCleaning MUST only be used in tests.
512 1 : func (d *DB) TestOnlyWaitForCleaning() {
513 1 : d.cleanupManager.Wait()
514 1 : }
515 :
516 : // Get gets the value for the given key. It returns ErrNotFound if the DB does
517 : // not contain the key.
518 : //
519 : // The caller should not modify the contents of the returned slice, but it is
520 : // safe to modify the contents of the argument after Get returns. The returned
521 : // slice will remain valid until the returned Closer is closed. On success, the
522 : // caller MUST call closer.Close() or a memory leak will occur.
523 2 : func (d *DB) Get(key []byte) ([]byte, io.Closer, error) {
524 2 : return d.getInternal(key, nil /* batch */, nil /* snapshot */)
525 2 : }
526 :
527 : type getIterAlloc struct {
528 : dbi Iterator
529 : keyBuf []byte
530 : get getIter
531 : }
532 :
533 : var getIterAllocPool = sync.Pool{
534 2 : New: func() interface{} {
535 2 : return &getIterAlloc{}
536 2 : },
537 : }
538 :
539 2 : func (d *DB) getInternal(key []byte, b *Batch, s *Snapshot) ([]byte, io.Closer, error) {
540 2 : if err := d.closed.Load(); err != nil {
541 1 : panic(err)
542 : }
543 :
544 : // Grab and reference the current readState. This prevents the underlying
545 : // files in the associated version from being deleted if there is a current
546 : // compaction. The readState is unref'd by Iterator.Close().
547 2 : readState := d.loadReadState()
548 2 :
549 2 : // Determine the seqnum to read at after grabbing the read state (current and
550 2 : // memtables) above.
551 2 : var seqNum uint64
552 2 : if s != nil {
553 2 : seqNum = s.seqNum
554 2 : } else {
555 2 : seqNum = d.mu.versions.visibleSeqNum.Load()
556 2 : }
557 :
558 2 : buf := getIterAllocPool.Get().(*getIterAlloc)
559 2 :
560 2 : get := &buf.get
561 2 : *get = getIter{
562 2 : logger: d.opts.Logger,
563 2 : comparer: d.opts.Comparer,
564 2 : newIters: d.newIters,
565 2 : snapshot: seqNum,
566 2 : key: key,
567 2 : batch: b,
568 2 : mem: readState.memtables,
569 2 : l0: readState.current.L0SublevelFiles,
570 2 : version: readState.current,
571 2 : }
572 2 :
573 2 : // Strip off memtables which cannot possibly contain the seqNum being read
574 2 : // at.
575 2 : for len(get.mem) > 0 {
576 2 : n := len(get.mem)
577 2 : if logSeqNum := get.mem[n-1].logSeqNum; logSeqNum < seqNum {
578 2 : break
579 : }
580 2 : get.mem = get.mem[:n-1]
581 : }
582 :
583 2 : i := &buf.dbi
584 2 : pointIter := get
585 2 : *i = Iterator{
586 2 : ctx: context.Background(),
587 2 : getIterAlloc: buf,
588 2 : iter: pointIter,
589 2 : pointIter: pointIter,
590 2 : merge: d.merge,
591 2 : comparer: *d.opts.Comparer,
592 2 : readState: readState,
593 2 : keyBuf: buf.keyBuf,
594 2 : }
595 2 :
596 2 : if !i.First() {
597 2 : err := i.Close()
598 2 : if err != nil {
599 1 : return nil, nil, err
600 1 : }
601 2 : return nil, nil, ErrNotFound
602 : }
603 2 : return i.Value(), i, nil
604 : }
605 :
606 : // Set sets the value for the given key. It overwrites any previous value
607 : // for that key; a DB is not a multi-map.
608 : //
609 : // It is safe to modify the contents of the arguments after Set returns.
610 2 : func (d *DB) Set(key, value []byte, opts *WriteOptions) error {
611 2 : b := newBatch(d)
612 2 : _ = b.Set(key, value, opts)
613 2 : if err := d.Apply(b, opts); err != nil {
614 1 : return err
615 1 : }
616 : // Only release the batch on success.
617 2 : b.release()
618 2 : return nil
619 : }
620 :
621 : // Delete deletes the value for the given key. Deletes are blind all will
622 : // succeed even if the given key does not exist.
623 : //
624 : // It is safe to modify the contents of the arguments after Delete returns.
625 2 : func (d *DB) Delete(key []byte, opts *WriteOptions) error {
626 2 : b := newBatch(d)
627 2 : _ = b.Delete(key, opts)
628 2 : if err := d.Apply(b, opts); err != nil {
629 1 : return err
630 1 : }
631 : // Only release the batch on success.
632 2 : b.release()
633 2 : return nil
634 : }
635 :
636 : // DeleteSized behaves identically to Delete, but takes an additional
637 : // argument indicating the size of the value being deleted. DeleteSized
638 : // should be preferred when the caller has the expectation that there exists
639 : // a single internal KV pair for the key (eg, the key has not been
640 : // overwritten recently), and the caller knows the size of its value.
641 : //
642 : // DeleteSized will record the value size within the tombstone and use it to
643 : // inform compaction-picking heuristics which strive to reduce space
644 : // amplification in the LSM. This "calling your shot" mechanic allows the
645 : // storage engine to more accurately estimate and reduce space amplification.
646 : //
647 : // It is safe to modify the contents of the arguments after DeleteSized
648 : // returns.
649 1 : func (d *DB) DeleteSized(key []byte, valueSize uint32, opts *WriteOptions) error {
650 1 : b := newBatch(d)
651 1 : _ = b.DeleteSized(key, valueSize, opts)
652 1 : if err := d.Apply(b, opts); err != nil {
653 0 : return err
654 0 : }
655 : // Only release the batch on success.
656 1 : b.release()
657 1 : return nil
658 : }
659 :
660 : // SingleDelete adds an action to the batch that single deletes the entry for key.
661 : // See Writer.SingleDelete for more details on the semantics of SingleDelete.
662 : //
663 : // It is safe to modify the contents of the arguments after SingleDelete returns.
664 2 : func (d *DB) SingleDelete(key []byte, opts *WriteOptions) error {
665 2 : b := newBatch(d)
666 2 : _ = b.SingleDelete(key, opts)
667 2 : if err := d.Apply(b, opts); err != nil {
668 0 : return err
669 0 : }
670 : // Only release the batch on success.
671 2 : b.release()
672 2 : return nil
673 : }
674 :
675 : // DeleteRange deletes all of the keys (and values) in the range [start,end)
676 : // (inclusive on start, exclusive on end).
677 : //
678 : // It is safe to modify the contents of the arguments after DeleteRange
679 : // returns.
680 2 : func (d *DB) DeleteRange(start, end []byte, opts *WriteOptions) error {
681 2 : b := newBatch(d)
682 2 : _ = b.DeleteRange(start, end, opts)
683 2 : if err := d.Apply(b, opts); err != nil {
684 1 : return err
685 1 : }
686 : // Only release the batch on success.
687 2 : b.release()
688 2 : return nil
689 : }
690 :
691 : // Merge adds an action to the DB that merges the value at key with the new
692 : // value. The details of the merge are dependent upon the configured merge
693 : // operator.
694 : //
695 : // It is safe to modify the contents of the arguments after Merge returns.
696 2 : func (d *DB) Merge(key, value []byte, opts *WriteOptions) error {
697 2 : b := newBatch(d)
698 2 : _ = b.Merge(key, value, opts)
699 2 : if err := d.Apply(b, opts); err != nil {
700 1 : return err
701 1 : }
702 : // Only release the batch on success.
703 2 : b.release()
704 2 : return nil
705 : }
706 :
707 : // LogData adds the specified to the batch. The data will be written to the
708 : // WAL, but not added to memtables or sstables. Log data is never indexed,
709 : // which makes it useful for testing WAL performance.
710 : //
711 : // It is safe to modify the contents of the argument after LogData returns.
712 1 : func (d *DB) LogData(data []byte, opts *WriteOptions) error {
713 1 : b := newBatch(d)
714 1 : _ = b.LogData(data, opts)
715 1 : if err := d.Apply(b, opts); err != nil {
716 1 : return err
717 1 : }
718 : // Only release the batch on success.
719 1 : b.release()
720 1 : return nil
721 : }
722 :
723 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
724 : // timestamp suffix to value. The suffix is optional. If any portion of the key
725 : // range [start, end) is already set by a range key with the same suffix value,
726 : // RangeKeySet overrides it.
727 : //
728 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
729 2 : func (d *DB) RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error {
730 2 : b := newBatch(d)
731 2 : _ = b.RangeKeySet(start, end, suffix, value, opts)
732 2 : if err := d.Apply(b, opts); err != nil {
733 0 : return err
734 0 : }
735 : // Only release the batch on success.
736 2 : b.release()
737 2 : return nil
738 : }
739 :
740 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
741 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
742 : // range key. RangeKeyUnset only removes portions of range keys that fall within
743 : // the [start, end) key span, and only range keys with suffixes that exactly
744 : // match the unset suffix.
745 : //
746 : // It is safe to modify the contents of the arguments after RangeKeyUnset
747 : // returns.
748 2 : func (d *DB) RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error {
749 2 : b := newBatch(d)
750 2 : _ = b.RangeKeyUnset(start, end, suffix, opts)
751 2 : if err := d.Apply(b, opts); err != nil {
752 0 : return err
753 0 : }
754 : // Only release the batch on success.
755 2 : b.release()
756 2 : return nil
757 : }
758 :
759 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
760 : // (inclusive on start, exclusive on end). It does not delete point keys (for
761 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
762 : // bounds, including those with or without suffixes.
763 : //
764 : // It is safe to modify the contents of the arguments after RangeKeyDelete
765 : // returns.
766 2 : func (d *DB) RangeKeyDelete(start, end []byte, opts *WriteOptions) error {
767 2 : b := newBatch(d)
768 2 : _ = b.RangeKeyDelete(start, end, opts)
769 2 : if err := d.Apply(b, opts); err != nil {
770 0 : return err
771 0 : }
772 : // Only release the batch on success.
773 2 : b.release()
774 2 : return nil
775 : }
776 :
777 : // Apply the operations contained in the batch to the DB. If the batch is large
778 : // the contents of the batch may be retained by the database. If that occurs
779 : // the batch contents will be cleared preventing the caller from attempting to
780 : // reuse them.
781 : //
782 : // It is safe to modify the contents of the arguments after Apply returns.
783 2 : func (d *DB) Apply(batch *Batch, opts *WriteOptions) error {
784 2 : return d.applyInternal(batch, opts, false)
785 2 : }
786 :
787 : // ApplyNoSyncWait must only be used when opts.Sync is true and the caller
788 : // does not want to wait for the WAL fsync to happen. The method will return
789 : // once the mutation is applied to the memtable and is visible (note that a
790 : // mutation is visible before the WAL sync even in the wait case, so we have
791 : // not weakened the durability semantics). The caller must call Batch.SyncWait
792 : // to wait for the WAL fsync. The caller must not Close the batch without
793 : // first calling Batch.SyncWait.
794 : //
795 : // RECOMMENDATION: Prefer using Apply unless you really understand why you
796 : // need ApplyNoSyncWait.
797 : // EXPERIMENTAL: API/feature subject to change. Do not yet use outside
798 : // CockroachDB.
799 2 : func (d *DB) ApplyNoSyncWait(batch *Batch, opts *WriteOptions) error {
800 2 : if !opts.Sync {
801 0 : return errors.Errorf("cannot request asynchonous apply when WriteOptions.Sync is false")
802 0 : }
803 2 : return d.applyInternal(batch, opts, true)
804 : }
805 :
806 : // REQUIRES: noSyncWait => opts.Sync
807 2 : func (d *DB) applyInternal(batch *Batch, opts *WriteOptions, noSyncWait bool) error {
808 2 : if err := d.closed.Load(); err != nil {
809 1 : panic(err)
810 : }
811 2 : if batch.committing {
812 0 : panic("pebble: batch already committing")
813 : }
814 2 : if batch.applied.Load() {
815 0 : panic("pebble: batch already applied")
816 : }
817 2 : if d.opts.ReadOnly {
818 1 : return ErrReadOnly
819 1 : }
820 2 : if batch.db != nil && batch.db != d {
821 1 : panic(fmt.Sprintf("pebble: batch db mismatch: %p != %p", batch.db, d))
822 : }
823 :
824 2 : sync := opts.GetSync()
825 2 : if sync && d.opts.DisableWAL {
826 0 : return errors.New("pebble: WAL disabled")
827 0 : }
828 :
829 2 : if batch.minimumFormatMajorVersion != FormatMostCompatible {
830 2 : if fmv := d.FormatMajorVersion(); fmv < batch.minimumFormatMajorVersion {
831 1 : panic(fmt.Sprintf(
832 1 : "pebble: batch requires at least format major version %d (current: %d)",
833 1 : batch.minimumFormatMajorVersion, fmv,
834 1 : ))
835 : }
836 : }
837 :
838 2 : if batch.countRangeKeys > 0 {
839 2 : if d.split == nil {
840 0 : return errNoSplit
841 0 : }
842 : // TODO(jackson): Assert that all range key operands are suffixless.
843 : }
844 2 : batch.committing = true
845 2 :
846 2 : if batch.db == nil {
847 1 : batch.refreshMemTableSize()
848 1 : }
849 2 : if batch.memTableSize >= d.largeBatchThreshold {
850 2 : batch.flushable = newFlushableBatch(batch, d.opts.Comparer)
851 2 : }
852 2 : if err := d.commit.Commit(batch, sync, noSyncWait); err != nil {
853 0 : // There isn't much we can do on an error here. The commit pipeline will be
854 0 : // horked at this point.
855 0 : d.opts.Logger.Fatalf("pebble: fatal commit error: %v", err)
856 0 : }
857 : // If this is a large batch, we need to clear the batch contents as the
858 : // flushable batch may still be present in the flushables queue.
859 : //
860 : // TODO(peter): Currently large batches are written to the WAL. We could
861 : // skip the WAL write and instead wait for the large batch to be flushed to
862 : // an sstable. For a 100 MB batch, this might actually be faster. For a 1
863 : // GB batch this is almost certainly faster.
864 2 : if batch.flushable != nil {
865 2 : batch.data = nil
866 2 : }
867 2 : return nil
868 : }
869 :
870 2 : func (d *DB) commitApply(b *Batch, mem *memTable) error {
871 2 : if b.flushable != nil {
872 2 : // This is a large batch which was already added to the immutable queue.
873 2 : return nil
874 2 : }
875 2 : err := mem.apply(b, b.SeqNum())
876 2 : if err != nil {
877 0 : return err
878 0 : }
879 :
880 : // If the batch contains range tombstones and the database is configured
881 : // to flush range deletions, schedule a delayed flush so that disk space
882 : // may be reclaimed without additional writes or an explicit flush.
883 2 : if b.countRangeDels > 0 && d.opts.FlushDelayDeleteRange > 0 {
884 2 : d.mu.Lock()
885 2 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayDeleteRange)
886 2 : d.mu.Unlock()
887 2 : }
888 :
889 : // If the batch contains range keys and the database is configured to flush
890 : // range keys, schedule a delayed flush so that the range keys are cleared
891 : // from the memtable.
892 2 : if b.countRangeKeys > 0 && d.opts.FlushDelayRangeKey > 0 {
893 2 : d.mu.Lock()
894 2 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayRangeKey)
895 2 : d.mu.Unlock()
896 2 : }
897 :
898 2 : if mem.writerUnref() {
899 2 : d.mu.Lock()
900 2 : d.maybeScheduleFlush()
901 2 : d.mu.Unlock()
902 2 : }
903 2 : return nil
904 : }
905 :
906 2 : func (d *DB) commitWrite(b *Batch, syncWG *sync.WaitGroup, syncErr *error) (*memTable, error) {
907 2 : var size int64
908 2 : repr := b.Repr()
909 2 :
910 2 : if b.flushable != nil {
911 2 : // We have a large batch. Such batches are special in that they don't get
912 2 : // added to the memtable, and are instead inserted into the queue of
913 2 : // memtables. The call to makeRoomForWrite with this batch will force the
914 2 : // current memtable to be flushed. We want the large batch to be part of
915 2 : // the same log, so we add it to the WAL here, rather than after the call
916 2 : // to makeRoomForWrite().
917 2 : //
918 2 : // Set the sequence number since it was not set to the correct value earlier
919 2 : // (see comment in newFlushableBatch()).
920 2 : b.flushable.setSeqNum(b.SeqNum())
921 2 : if !d.opts.DisableWAL {
922 2 : var err error
923 2 : size, err = d.mu.log.SyncRecord(repr, syncWG, syncErr)
924 2 : if err != nil {
925 0 : panic(err)
926 : }
927 : }
928 : }
929 :
930 2 : d.mu.Lock()
931 2 :
932 2 : var err error
933 2 : if !b.ingestedSSTBatch {
934 2 : // Batches which contain keys of kind InternalKeyKindIngestSST will
935 2 : // never be applied to the memtable, so we don't need to make room for
936 2 : // write. For the other cases, switch out the memtable if there was not
937 2 : // enough room to store the batch.
938 2 : err = d.makeRoomForWrite(b)
939 2 : }
940 :
941 2 : if err == nil && !d.opts.DisableWAL {
942 2 : d.mu.log.bytesIn += uint64(len(repr))
943 2 : }
944 :
945 : // Grab a reference to the memtable while holding DB.mu. Note that for
946 : // non-flushable batches (b.flushable == nil) makeRoomForWrite() added a
947 : // reference to the memtable which will prevent it from being flushed until
948 : // we unreference it. This reference is dropped in DB.commitApply().
949 2 : mem := d.mu.mem.mutable
950 2 :
951 2 : d.mu.Unlock()
952 2 : if err != nil {
953 0 : return nil, err
954 0 : }
955 :
956 2 : if d.opts.DisableWAL {
957 2 : return mem, nil
958 2 : }
959 :
960 2 : if b.flushable == nil {
961 2 : size, err = d.mu.log.SyncRecord(repr, syncWG, syncErr)
962 2 : if err != nil {
963 0 : panic(err)
964 : }
965 : }
966 :
967 2 : d.logSize.Store(uint64(size))
968 2 : return mem, err
969 : }
970 :
971 : type iterAlloc struct {
972 : dbi Iterator
973 : keyBuf []byte
974 : boundsBuf [2][]byte
975 : prefixOrFullSeekKey []byte
976 : merging mergingIter
977 : mlevels [3 + numLevels]mergingIterLevel
978 : levels [3 + numLevels]levelIter
979 : levelsPositioned [3 + numLevels]bool
980 : }
981 :
982 : var iterAllocPool = sync.Pool{
983 2 : New: func() interface{} {
984 2 : return &iterAlloc{}
985 2 : },
986 : }
987 :
988 : // snapshotIterOpts denotes snapshot-related iterator options when calling
989 : // newIter. These are the possible cases for a snapshotIterOpts:
990 : // - No snapshot: All fields are zero values.
991 : // - Classic snapshot: Only `seqNum` is set. The latest readState will be used
992 : // and the specified seqNum will be used as the snapshot seqNum.
993 : // - EventuallyFileOnlySnapshot (EFOS) behaving as a classic snapshot. Only
994 : // the `seqNum` is set. The latest readState will be used
995 : // and the specified seqNum will be used as the snapshot seqNum.
996 : // - EFOS in file-only state: Only `seqNum` and `vers` are set. All the
997 : // relevant SSTs are referenced by the *version.
998 : type snapshotIterOpts struct {
999 : seqNum uint64
1000 : vers *version
1001 : }
1002 :
1003 : type batchIterOpts struct {
1004 : batchOnly bool
1005 : }
1006 : type newIterOpts struct {
1007 : snapshot snapshotIterOpts
1008 : batch batchIterOpts
1009 : }
1010 :
1011 : // newIter constructs a new iterator, merging in batch iterators as an extra
1012 : // level.
1013 : func (d *DB) newIter(
1014 : ctx context.Context, batch *Batch, internalOpts newIterOpts, o *IterOptions,
1015 2 : ) *Iterator {
1016 2 : if internalOpts.batch.batchOnly {
1017 1 : if batch == nil {
1018 0 : panic("batchOnly is true, but batch is nil")
1019 : }
1020 1 : if internalOpts.snapshot.vers != nil {
1021 0 : panic("batchOnly is true, but snapshotIterOpts is initialized")
1022 : }
1023 : }
1024 2 : if err := d.closed.Load(); err != nil {
1025 1 : panic(err)
1026 : }
1027 2 : seqNum := internalOpts.snapshot.seqNum
1028 2 : if o.rangeKeys() {
1029 2 : if d.FormatMajorVersion() < FormatRangeKeys {
1030 1 : panic(fmt.Sprintf(
1031 1 : "pebble: range keys require at least format major version %d (current: %d)",
1032 1 : FormatRangeKeys, d.FormatMajorVersion(),
1033 1 : ))
1034 : }
1035 : }
1036 2 : if o != nil && o.RangeKeyMasking.Suffix != nil && o.KeyTypes != IterKeyTypePointsAndRanges {
1037 0 : panic("pebble: range key masking requires IterKeyTypePointsAndRanges")
1038 : }
1039 2 : if (batch != nil || seqNum != 0) && (o != nil && o.OnlyReadGuaranteedDurable) {
1040 1 : // We could add support for OnlyReadGuaranteedDurable on snapshots if
1041 1 : // there was a need: this would require checking that the sequence number
1042 1 : // of the snapshot has been flushed, by comparing with
1043 1 : // DB.mem.queue[0].logSeqNum.
1044 1 : panic("OnlyReadGuaranteedDurable is not supported for batches or snapshots")
1045 : }
1046 2 : var readState *readState
1047 2 : var newIters tableNewIters
1048 2 : var newIterRangeKey keyspan.TableNewSpanIter
1049 2 : if !internalOpts.batch.batchOnly {
1050 2 : // Grab and reference the current readState. This prevents the underlying
1051 2 : // files in the associated version from being deleted if there is a current
1052 2 : // compaction. The readState is unref'd by Iterator.Close().
1053 2 : if internalOpts.snapshot.vers == nil {
1054 2 : // NB: loadReadState() calls readState.ref().
1055 2 : readState = d.loadReadState()
1056 2 : } else {
1057 2 : // vers != nil
1058 2 : internalOpts.snapshot.vers.Ref()
1059 2 : }
1060 :
1061 : // Determine the seqnum to read at after grabbing the read state (current and
1062 : // memtables) above.
1063 2 : if seqNum == 0 {
1064 2 : seqNum = d.mu.versions.visibleSeqNum.Load()
1065 2 : }
1066 2 : newIters = d.newIters
1067 2 : newIterRangeKey = d.tableNewRangeKeyIter
1068 : }
1069 :
1070 : // Bundle various structures under a single umbrella in order to allocate
1071 : // them together.
1072 2 : buf := iterAllocPool.Get().(*iterAlloc)
1073 2 : dbi := &buf.dbi
1074 2 : *dbi = Iterator{
1075 2 : ctx: ctx,
1076 2 : alloc: buf,
1077 2 : merge: d.merge,
1078 2 : comparer: *d.opts.Comparer,
1079 2 : readState: readState,
1080 2 : version: internalOpts.snapshot.vers,
1081 2 : keyBuf: buf.keyBuf,
1082 2 : prefixOrFullSeekKey: buf.prefixOrFullSeekKey,
1083 2 : boundsBuf: buf.boundsBuf,
1084 2 : batch: batch,
1085 2 : newIters: newIters,
1086 2 : newIterRangeKey: newIterRangeKey,
1087 2 : seqNum: seqNum,
1088 2 : batchOnlyIter: internalOpts.batch.batchOnly,
1089 2 : }
1090 2 : if o != nil {
1091 2 : dbi.opts = *o
1092 2 : dbi.processBounds(o.LowerBound, o.UpperBound)
1093 2 : }
1094 2 : dbi.opts.logger = d.opts.Logger
1095 2 : if d.opts.private.disableLazyCombinedIteration {
1096 1 : dbi.opts.disableLazyCombinedIteration = true
1097 1 : }
1098 2 : if batch != nil {
1099 2 : dbi.batchSeqNum = dbi.batch.nextSeqNum()
1100 2 : }
1101 2 : return finishInitializingIter(ctx, buf)
1102 : }
1103 :
1104 : // finishInitializingIter is a helper for doing the non-trivial initialization
1105 : // of an Iterator. It's invoked to perform the initial initialization of an
1106 : // Iterator during NewIter or Clone, and to perform reinitialization due to a
1107 : // change in IterOptions by a call to Iterator.SetOptions.
1108 2 : func finishInitializingIter(ctx context.Context, buf *iterAlloc) *Iterator {
1109 2 : // Short-hand.
1110 2 : dbi := &buf.dbi
1111 2 : var memtables flushableList
1112 2 : if dbi.readState != nil {
1113 2 : memtables = dbi.readState.memtables
1114 2 : }
1115 2 : if dbi.opts.OnlyReadGuaranteedDurable {
1116 1 : memtables = nil
1117 2 : } else {
1118 2 : // We only need to read from memtables which contain sequence numbers older
1119 2 : // than seqNum. Trim off newer memtables.
1120 2 : for i := len(memtables) - 1; i >= 0; i-- {
1121 2 : if logSeqNum := memtables[i].logSeqNum; logSeqNum < dbi.seqNum {
1122 2 : break
1123 : }
1124 2 : memtables = memtables[:i]
1125 : }
1126 : }
1127 :
1128 2 : if dbi.opts.pointKeys() {
1129 2 : // Construct the point iterator, initializing dbi.pointIter to point to
1130 2 : // dbi.merging. If this is called during a SetOptions call and this
1131 2 : // Iterator has already initialized dbi.merging, constructPointIter is a
1132 2 : // noop and an initialized pointIter already exists in dbi.pointIter.
1133 2 : dbi.constructPointIter(ctx, memtables, buf)
1134 2 : dbi.iter = dbi.pointIter
1135 2 : } else {
1136 2 : dbi.iter = emptyIter
1137 2 : }
1138 :
1139 2 : if dbi.opts.rangeKeys() {
1140 2 : dbi.rangeKeyMasking.init(dbi, dbi.comparer.Compare, dbi.comparer.Split)
1141 2 :
1142 2 : // When iterating over both point and range keys, don't create the
1143 2 : // range-key iterator stack immediately if we can avoid it. This
1144 2 : // optimization takes advantage of the expected sparseness of range
1145 2 : // keys, and configures the point-key iterator to dynamically switch to
1146 2 : // combined iteration when it observes a file containing range keys.
1147 2 : //
1148 2 : // Lazy combined iteration is not possible if a batch or a memtable
1149 2 : // contains any range keys.
1150 2 : useLazyCombinedIteration := dbi.rangeKey == nil &&
1151 2 : dbi.opts.KeyTypes == IterKeyTypePointsAndRanges &&
1152 2 : (dbi.batch == nil || dbi.batch.countRangeKeys == 0) &&
1153 2 : !dbi.opts.disableLazyCombinedIteration
1154 2 : if useLazyCombinedIteration {
1155 2 : // The user requested combined iteration, and there's no indexed
1156 2 : // batch currently containing range keys that would prevent lazy
1157 2 : // combined iteration. Check the memtables to see if they contain
1158 2 : // any range keys.
1159 2 : for i := range memtables {
1160 2 : if memtables[i].containsRangeKeys() {
1161 2 : useLazyCombinedIteration = false
1162 2 : break
1163 : }
1164 : }
1165 : }
1166 :
1167 2 : if useLazyCombinedIteration {
1168 2 : dbi.lazyCombinedIter = lazyCombinedIter{
1169 2 : parent: dbi,
1170 2 : pointIter: dbi.pointIter,
1171 2 : combinedIterState: combinedIterState{
1172 2 : initialized: false,
1173 2 : },
1174 2 : }
1175 2 : dbi.iter = &dbi.lazyCombinedIter
1176 2 : dbi.iter = invalidating.MaybeWrapIfInvariants(dbi.iter)
1177 2 : } else {
1178 2 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{
1179 2 : initialized: true,
1180 2 : }
1181 2 : if dbi.rangeKey == nil {
1182 2 : dbi.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1183 2 : dbi.rangeKey.init(dbi.comparer.Compare, dbi.comparer.Split, &dbi.opts)
1184 2 : dbi.constructRangeKeyIter()
1185 2 : } else {
1186 2 : dbi.rangeKey.iterConfig.SetBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
1187 2 : }
1188 :
1189 : // Wrap the point iterator (currently dbi.iter) with an interleaving
1190 : // iterator that interleaves range keys pulled from
1191 : // dbi.rangeKey.rangeKeyIter.
1192 : //
1193 : // NB: The interleaving iterator is always reinitialized, even if
1194 : // dbi already had an initialized range key iterator, in case the point
1195 : // iterator changed or the range key masking suffix changed.
1196 2 : dbi.rangeKey.iiter.Init(&dbi.comparer, dbi.iter, dbi.rangeKey.rangeKeyIter,
1197 2 : keyspan.InterleavingIterOpts{
1198 2 : Mask: &dbi.rangeKeyMasking,
1199 2 : LowerBound: dbi.opts.LowerBound,
1200 2 : UpperBound: dbi.opts.UpperBound,
1201 2 : })
1202 2 : dbi.iter = &dbi.rangeKey.iiter
1203 : }
1204 2 : } else {
1205 2 : // !dbi.opts.rangeKeys()
1206 2 : //
1207 2 : // Reset the combined iterator state. The initialized=true ensures the
1208 2 : // iterator doesn't unnecessarily try to switch to combined iteration.
1209 2 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{initialized: true}
1210 2 : }
1211 2 : return dbi
1212 : }
1213 :
1214 : // ScanInternal scans all internal keys within the specified bounds, truncating
1215 : // any rangedels and rangekeys to those bounds if they span past them. For use
1216 : // when an external user needs to be aware of all internal keys that make up a
1217 : // key range.
1218 : //
1219 : // Keys deleted by range deletions must not be returned or exposed by this
1220 : // method, while the range deletion deleting that key must be exposed using
1221 : // visitRangeDel. Keys that would be masked by range key masking (if an
1222 : // appropriate prefix were set) should be exposed, alongside the range key
1223 : // that would have masked it. This method also collapses all point keys into
1224 : // one InternalKey; so only one internal key at most per user key is returned
1225 : // to visitPointKey.
1226 : //
1227 : // If visitSharedFile is not nil, ScanInternal iterates in skip-shared iteration
1228 : // mode. In this iteration mode, sstables in levels L5 and L6 are skipped, and
1229 : // their metadatas truncated to [lower, upper) and passed into visitSharedFile.
1230 : // ErrInvalidSkipSharedIteration is returned if visitSharedFile is not nil and an
1231 : // sstable in L5 or L6 is found that is not in shared storage according to
1232 : // provider.IsShared, or an sstable in those levels contains a newer key than the
1233 : // snapshot sequence number (only applicable for snapshot.ScanInternal). Examples
1234 : // of when this could happen could be if Pebble started writing sstables before a
1235 : // creator ID was set (as creator IDs are necessary to enable shared storage)
1236 : // resulting in some lower level SSTs being on non-shared storage. Skip-shared
1237 : // iteration is invalid in those cases.
1238 : func (d *DB) ScanInternal(
1239 : ctx context.Context,
1240 : categoryAndQoS sstable.CategoryAndQoS,
1241 : lower, upper []byte,
1242 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error,
1243 : visitRangeDel func(start, end []byte, seqNum uint64) error,
1244 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error,
1245 : visitSharedFile func(sst *SharedSSTMeta) error,
1246 1 : ) error {
1247 1 : scanInternalOpts := &scanInternalOptions{
1248 1 : CategoryAndQoS: categoryAndQoS,
1249 1 : visitPointKey: visitPointKey,
1250 1 : visitRangeDel: visitRangeDel,
1251 1 : visitRangeKey: visitRangeKey,
1252 1 : visitSharedFile: visitSharedFile,
1253 1 : skipSharedLevels: visitSharedFile != nil,
1254 1 : IterOptions: IterOptions{
1255 1 : KeyTypes: IterKeyTypePointsAndRanges,
1256 1 : LowerBound: lower,
1257 1 : UpperBound: upper,
1258 1 : },
1259 1 : }
1260 1 : iter := d.newInternalIter(ctx, snapshotIterOpts{} /* snapshot */, scanInternalOpts)
1261 1 : defer iter.close()
1262 1 : return scanInternalImpl(ctx, lower, upper, iter, scanInternalOpts)
1263 1 : }
1264 :
1265 : // newInternalIter constructs and returns a new scanInternalIterator on this db.
1266 : // If o.skipSharedLevels is true, levels below sharedLevelsStart are *not* added
1267 : // to the internal iterator.
1268 : //
1269 : // TODO(bilal): This method has a lot of similarities with db.newIter as well as
1270 : // finishInitializingIter. Both pairs of methods should be refactored to reduce
1271 : // this duplication.
1272 : func (d *DB) newInternalIter(
1273 : ctx context.Context, sOpts snapshotIterOpts, o *scanInternalOptions,
1274 1 : ) *scanInternalIterator {
1275 1 : if err := d.closed.Load(); err != nil {
1276 0 : panic(err)
1277 : }
1278 : // Grab and reference the current readState. This prevents the underlying
1279 : // files in the associated version from being deleted if there is a current
1280 : // compaction. The readState is unref'd by Iterator.Close().
1281 1 : var readState *readState
1282 1 : if sOpts.vers == nil {
1283 1 : readState = d.loadReadState()
1284 1 : }
1285 1 : if sOpts.vers != nil {
1286 1 : sOpts.vers.Ref()
1287 1 : }
1288 :
1289 : // Determine the seqnum to read at after grabbing the read state (current and
1290 : // memtables) above.
1291 1 : seqNum := sOpts.seqNum
1292 1 : if seqNum == 0 {
1293 1 : seqNum = d.mu.versions.visibleSeqNum.Load()
1294 1 : }
1295 :
1296 : // Bundle various structures under a single umbrella in order to allocate
1297 : // them together.
1298 1 : buf := iterAllocPool.Get().(*iterAlloc)
1299 1 : dbi := &scanInternalIterator{
1300 1 : ctx: ctx,
1301 1 : db: d,
1302 1 : comparer: d.opts.Comparer,
1303 1 : merge: d.opts.Merger.Merge,
1304 1 : readState: readState,
1305 1 : version: sOpts.vers,
1306 1 : alloc: buf,
1307 1 : newIters: d.newIters,
1308 1 : newIterRangeKey: d.tableNewRangeKeyIter,
1309 1 : seqNum: seqNum,
1310 1 : mergingIter: &buf.merging,
1311 1 : }
1312 1 : dbi.opts = *o
1313 1 : dbi.opts.logger = d.opts.Logger
1314 1 : if d.opts.private.disableLazyCombinedIteration {
1315 0 : dbi.opts.disableLazyCombinedIteration = true
1316 0 : }
1317 1 : return finishInitializingInternalIter(buf, dbi)
1318 : }
1319 :
1320 1 : func finishInitializingInternalIter(buf *iterAlloc, i *scanInternalIterator) *scanInternalIterator {
1321 1 : // Short-hand.
1322 1 : var memtables flushableList
1323 1 : if i.readState != nil {
1324 1 : memtables = i.readState.memtables
1325 1 : }
1326 : // We only need to read from memtables which contain sequence numbers older
1327 : // than seqNum. Trim off newer memtables.
1328 1 : for j := len(memtables) - 1; j >= 0; j-- {
1329 1 : if logSeqNum := memtables[j].logSeqNum; logSeqNum < i.seqNum {
1330 1 : break
1331 : }
1332 1 : memtables = memtables[:j]
1333 : }
1334 1 : i.initializeBoundBufs(i.opts.LowerBound, i.opts.UpperBound)
1335 1 :
1336 1 : i.constructPointIter(i.opts.CategoryAndQoS, memtables, buf)
1337 1 :
1338 1 : // For internal iterators, we skip the lazy combined iteration optimization
1339 1 : // entirely, and create the range key iterator stack directly.
1340 1 : i.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1341 1 : i.rangeKey.init(i.comparer.Compare, i.comparer.Split, &i.opts.IterOptions)
1342 1 : i.constructRangeKeyIter()
1343 1 :
1344 1 : // Wrap the point iterator (currently i.iter) with an interleaving
1345 1 : // iterator that interleaves range keys pulled from
1346 1 : // i.rangeKey.rangeKeyIter.
1347 1 : i.rangeKey.iiter.Init(i.comparer, i.iter, i.rangeKey.rangeKeyIter,
1348 1 : keyspan.InterleavingIterOpts{
1349 1 : LowerBound: i.opts.LowerBound,
1350 1 : UpperBound: i.opts.UpperBound,
1351 1 : })
1352 1 : i.iter = &i.rangeKey.iiter
1353 1 :
1354 1 : return i
1355 : }
1356 :
1357 : func (i *Iterator) constructPointIter(
1358 : ctx context.Context, memtables flushableList, buf *iterAlloc,
1359 2 : ) {
1360 2 : if i.pointIter != nil {
1361 2 : // Already have one.
1362 2 : return
1363 2 : }
1364 2 : internalOpts := internalIterOpts{stats: &i.stats.InternalStats}
1365 2 : if i.opts.RangeKeyMasking.Filter != nil {
1366 2 : internalOpts.boundLimitedFilter = &i.rangeKeyMasking
1367 2 : }
1368 :
1369 : // Merging levels and levels from iterAlloc.
1370 2 : mlevels := buf.mlevels[:0]
1371 2 : levels := buf.levels[:0]
1372 2 :
1373 2 : // We compute the number of levels needed ahead of time and reallocate a slice if
1374 2 : // the array from the iterAlloc isn't large enough. Doing this allocation once
1375 2 : // should improve the performance.
1376 2 : numMergingLevels := 0
1377 2 : numLevelIters := 0
1378 2 : if i.batch != nil {
1379 2 : numMergingLevels++
1380 2 : }
1381 :
1382 2 : var current *version
1383 2 : if !i.batchOnlyIter {
1384 2 : numMergingLevels += len(memtables)
1385 2 :
1386 2 : current = i.version
1387 2 : if current == nil {
1388 2 : current = i.readState.current
1389 2 : }
1390 2 : numMergingLevels += len(current.L0SublevelFiles)
1391 2 : numLevelIters += len(current.L0SublevelFiles)
1392 2 : for level := 1; level < len(current.Levels); level++ {
1393 2 : if current.Levels[level].Empty() {
1394 2 : continue
1395 : }
1396 2 : numMergingLevels++
1397 2 : numLevelIters++
1398 : }
1399 : }
1400 :
1401 2 : if numMergingLevels > cap(mlevels) {
1402 2 : mlevels = make([]mergingIterLevel, 0, numMergingLevels)
1403 2 : }
1404 2 : if numLevelIters > cap(levels) {
1405 2 : levels = make([]levelIter, 0, numLevelIters)
1406 2 : }
1407 :
1408 : // Top-level is the batch, if any.
1409 2 : if i.batch != nil {
1410 2 : if i.batch.index == nil {
1411 0 : // This isn't an indexed batch. Include an error iterator so that
1412 0 : // the resulting iterator correctly surfaces ErrIndexed.
1413 0 : mlevels = append(mlevels, mergingIterLevel{
1414 0 : iter: newErrorIter(ErrNotIndexed),
1415 0 : rangeDelIter: newErrorKeyspanIter(ErrNotIndexed),
1416 0 : })
1417 2 : } else {
1418 2 : i.batch.initInternalIter(&i.opts, &i.batchPointIter)
1419 2 : i.batch.initRangeDelIter(&i.opts, &i.batchRangeDelIter, i.batchSeqNum)
1420 2 : // Only include the batch's rangedel iterator if it's non-empty.
1421 2 : // This requires some subtle logic in the case a rangedel is later
1422 2 : // written to the batch and the view of the batch is refreshed
1423 2 : // during a call to SetOptions—in this case, we need to reconstruct
1424 2 : // the point iterator to add the batch rangedel iterator.
1425 2 : var rangeDelIter keyspan.FragmentIterator
1426 2 : if i.batchRangeDelIter.Count() > 0 {
1427 2 : rangeDelIter = &i.batchRangeDelIter
1428 2 : }
1429 2 : mlevels = append(mlevels, mergingIterLevel{
1430 2 : iter: &i.batchPointIter,
1431 2 : rangeDelIter: rangeDelIter,
1432 2 : })
1433 : }
1434 : }
1435 :
1436 2 : if !i.batchOnlyIter {
1437 2 : // Next are the memtables.
1438 2 : for j := len(memtables) - 1; j >= 0; j-- {
1439 2 : mem := memtables[j]
1440 2 : mlevels = append(mlevels, mergingIterLevel{
1441 2 : iter: mem.newIter(&i.opts),
1442 2 : rangeDelIter: mem.newRangeDelIter(&i.opts),
1443 2 : })
1444 2 : }
1445 :
1446 : // Next are the file levels: L0 sub-levels followed by lower levels.
1447 2 : mlevelsIndex := len(mlevels)
1448 2 : levelsIndex := len(levels)
1449 2 : mlevels = mlevels[:numMergingLevels]
1450 2 : levels = levels[:numLevelIters]
1451 2 : i.opts.snapshotForHideObsoletePoints = buf.dbi.seqNum
1452 2 : addLevelIterForFiles := func(files manifest.LevelIterator, level manifest.Level) {
1453 2 : li := &levels[levelsIndex]
1454 2 :
1455 2 : li.init(ctx, i.opts, &i.comparer, i.newIters, files, level, internalOpts)
1456 2 : li.initRangeDel(&mlevels[mlevelsIndex].rangeDelIter)
1457 2 : li.initBoundaryContext(&mlevels[mlevelsIndex].levelIterBoundaryContext)
1458 2 : li.initCombinedIterState(&i.lazyCombinedIter.combinedIterState)
1459 2 : mlevels[mlevelsIndex].levelIter = li
1460 2 : mlevels[mlevelsIndex].iter = invalidating.MaybeWrapIfInvariants(li)
1461 2 :
1462 2 : levelsIndex++
1463 2 : mlevelsIndex++
1464 2 : }
1465 :
1466 : // Add level iterators for the L0 sublevels, iterating from newest to
1467 : // oldest.
1468 2 : for i := len(current.L0SublevelFiles) - 1; i >= 0; i-- {
1469 2 : addLevelIterForFiles(current.L0SublevelFiles[i].Iter(), manifest.L0Sublevel(i))
1470 2 : }
1471 :
1472 : // Add level iterators for the non-empty non-L0 levels.
1473 2 : for level := 1; level < len(current.Levels); level++ {
1474 2 : if current.Levels[level].Empty() {
1475 2 : continue
1476 : }
1477 2 : addLevelIterForFiles(current.Levels[level].Iter(), manifest.Level(level))
1478 : }
1479 : }
1480 2 : buf.merging.init(&i.opts, &i.stats.InternalStats, i.comparer.Compare, i.comparer.Split, mlevels...)
1481 2 : if len(mlevels) <= cap(buf.levelsPositioned) {
1482 2 : buf.merging.levelsPositioned = buf.levelsPositioned[:len(mlevels)]
1483 2 : }
1484 2 : buf.merging.snapshot = i.seqNum
1485 2 : buf.merging.batchSnapshot = i.batchSeqNum
1486 2 : buf.merging.combinedIterState = &i.lazyCombinedIter.combinedIterState
1487 2 : i.pointIter = invalidating.MaybeWrapIfInvariants(&buf.merging)
1488 2 : i.merging = &buf.merging
1489 : }
1490 :
1491 : // NewBatch returns a new empty write-only batch. Any reads on the batch will
1492 : // return an error. If the batch is committed it will be applied to the DB.
1493 2 : func (d *DB) NewBatch() *Batch {
1494 2 : return newBatch(d)
1495 2 : }
1496 :
1497 : // NewBatchWithSize is mostly identical to NewBatch, but it will allocate the
1498 : // the specified memory space for the internal slice in advance.
1499 0 : func (d *DB) NewBatchWithSize(size int) *Batch {
1500 0 : return newBatchWithSize(d, size)
1501 0 : }
1502 :
1503 : // NewIndexedBatch returns a new empty read-write batch. Any reads on the batch
1504 : // will read from both the batch and the DB. If the batch is committed it will
1505 : // be applied to the DB. An indexed batch is slower that a non-indexed batch
1506 : // for insert operations. If you do not need to perform reads on the batch, use
1507 : // NewBatch instead.
1508 2 : func (d *DB) NewIndexedBatch() *Batch {
1509 2 : return newIndexedBatch(d, d.opts.Comparer)
1510 2 : }
1511 :
1512 : // NewIndexedBatchWithSize is mostly identical to NewIndexedBatch, but it will
1513 : // allocate the the specified memory space for the internal slice in advance.
1514 0 : func (d *DB) NewIndexedBatchWithSize(size int) *Batch {
1515 0 : return newIndexedBatchWithSize(d, d.opts.Comparer, size)
1516 0 : }
1517 :
1518 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
1519 : // return false). The iterator can be positioned via a call to SeekGE, SeekLT,
1520 : // First or Last. The iterator provides a point-in-time view of the current DB
1521 : // state. This view is maintained by preventing file deletions and preventing
1522 : // memtables referenced by the iterator from being deleted. Using an iterator
1523 : // to maintain a long-lived point-in-time view of the DB state can lead to an
1524 : // apparent memory and disk usage leak. Use snapshots (see NewSnapshot) for
1525 : // point-in-time snapshots which avoids these problems.
1526 2 : func (d *DB) NewIter(o *IterOptions) (*Iterator, error) {
1527 2 : return d.NewIterWithContext(context.Background(), o)
1528 2 : }
1529 :
1530 : // NewIterWithContext is like NewIter, and additionally accepts a context for
1531 : // tracing.
1532 2 : func (d *DB) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
1533 2 : return d.newIter(ctx, nil /* batch */, newIterOpts{}, o), nil
1534 2 : }
1535 :
1536 : // NewSnapshot returns a point-in-time view of the current DB state. Iterators
1537 : // created with this handle will all observe a stable snapshot of the current
1538 : // DB state. The caller must call Snapshot.Close() when the snapshot is no
1539 : // longer needed. Snapshots are not persisted across DB restarts (close ->
1540 : // open). Unlike the implicit snapshot maintained by an iterator, a snapshot
1541 : // will not prevent memtables from being released or sstables from being
1542 : // deleted. Instead, a snapshot prevents deletion of sequence numbers
1543 : // referenced by the snapshot.
1544 2 : func (d *DB) NewSnapshot() *Snapshot {
1545 2 : if err := d.closed.Load(); err != nil {
1546 1 : panic(err)
1547 : }
1548 :
1549 2 : d.mu.Lock()
1550 2 : s := &Snapshot{
1551 2 : db: d,
1552 2 : seqNum: d.mu.versions.visibleSeqNum.Load(),
1553 2 : }
1554 2 : d.mu.snapshots.pushBack(s)
1555 2 : d.mu.Unlock()
1556 2 : return s
1557 : }
1558 :
1559 : // NewEventuallyFileOnlySnapshot returns a point-in-time view of the current DB
1560 : // state, similar to NewSnapshot, but with consistency constrained to the
1561 : // provided set of key ranges. See the comment at EventuallyFileOnlySnapshot for
1562 : // its semantics.
1563 2 : func (d *DB) NewEventuallyFileOnlySnapshot(keyRanges []KeyRange) *EventuallyFileOnlySnapshot {
1564 2 : if err := d.closed.Load(); err != nil {
1565 0 : panic(err)
1566 : }
1567 :
1568 2 : internalKeyRanges := make([]internalKeyRange, len(keyRanges))
1569 2 : for i := range keyRanges {
1570 2 : if i > 0 && d.cmp(keyRanges[i-1].End, keyRanges[i].Start) > 0 {
1571 0 : panic("pebble: key ranges for eventually-file-only-snapshot not in order")
1572 : }
1573 2 : internalKeyRanges[i] = internalKeyRange{
1574 2 : smallest: base.MakeInternalKey(keyRanges[i].Start, InternalKeySeqNumMax, InternalKeyKindMax),
1575 2 : largest: base.MakeExclusiveSentinelKey(InternalKeyKindRangeDelete, keyRanges[i].End),
1576 2 : }
1577 : }
1578 :
1579 2 : return d.makeEventuallyFileOnlySnapshot(keyRanges, internalKeyRanges)
1580 : }
1581 :
1582 : // Close closes the DB.
1583 : //
1584 : // It is not safe to close a DB until all outstanding iterators are closed
1585 : // or to call Close concurrently with any other DB method. It is not valid
1586 : // to call any of a DB's methods after the DB has been closed.
1587 2 : func (d *DB) Close() error {
1588 2 : // Lock the commit pipeline for the duration of Close. This prevents a race
1589 2 : // with makeRoomForWrite. Rotating the WAL in makeRoomForWrite requires
1590 2 : // dropping d.mu several times for I/O. If Close only holds d.mu, an
1591 2 : // in-progress WAL rotation may re-acquire d.mu only once the database is
1592 2 : // closed.
1593 2 : //
1594 2 : // Additionally, locking the commit pipeline makes it more likely that
1595 2 : // (illegal) concurrent writes will observe d.closed.Load() != nil, creating
1596 2 : // more understable panics if the database is improperly used concurrently
1597 2 : // during Close.
1598 2 : d.commit.mu.Lock()
1599 2 : defer d.commit.mu.Unlock()
1600 2 : d.mu.Lock()
1601 2 : defer d.mu.Unlock()
1602 2 : if err := d.closed.Load(); err != nil {
1603 1 : panic(err)
1604 : }
1605 :
1606 : // Clear the finalizer that is used to check that an unreferenced DB has been
1607 : // closed. We're closing the DB here, so the check performed by that
1608 : // finalizer isn't necessary.
1609 : //
1610 : // Note: this is a no-op if invariants are disabled or race is enabled.
1611 2 : invariants.SetFinalizer(d.closed, nil)
1612 2 :
1613 2 : d.closed.Store(errors.WithStack(ErrClosed))
1614 2 : close(d.closedCh)
1615 2 :
1616 2 : defer d.opts.Cache.Unref()
1617 2 :
1618 2 : for d.mu.compact.compactingCount > 0 || d.mu.compact.flushing {
1619 2 : d.mu.compact.cond.Wait()
1620 2 : }
1621 2 : for d.mu.tableStats.loading {
1622 2 : d.mu.tableStats.cond.Wait()
1623 2 : }
1624 2 : for d.mu.tableValidation.validating {
1625 2 : d.mu.tableValidation.cond.Wait()
1626 2 : }
1627 :
1628 2 : var err error
1629 2 : if n := len(d.mu.compact.inProgress); n > 0 {
1630 1 : err = errors.Errorf("pebble: %d unexpected in-progress compactions", errors.Safe(n))
1631 1 : }
1632 2 : err = firstError(err, d.mu.formatVers.marker.Close())
1633 2 : err = firstError(err, d.tableCache.close())
1634 2 : if !d.opts.ReadOnly {
1635 2 : err = firstError(err, d.mu.log.Close())
1636 2 : } else if d.mu.log.LogWriter != nil {
1637 0 : panic("pebble: log-writer should be nil in read-only mode")
1638 : }
1639 2 : err = firstError(err, d.fileLock.Close())
1640 2 :
1641 2 : // Note that versionSet.close() only closes the MANIFEST. The versions list
1642 2 : // is still valid for the checks below.
1643 2 : err = firstError(err, d.mu.versions.close())
1644 2 :
1645 2 : err = firstError(err, d.dataDir.Close())
1646 2 : if d.dataDir != d.walDir {
1647 2 : err = firstError(err, d.walDir.Close())
1648 2 : }
1649 :
1650 2 : d.readState.val.unrefLocked()
1651 2 :
1652 2 : current := d.mu.versions.currentVersion()
1653 2 : for v := d.mu.versions.versions.Front(); true; v = v.Next() {
1654 2 : refs := v.Refs()
1655 2 : if v == current {
1656 2 : if refs != 1 {
1657 1 : err = firstError(err, errors.Errorf("leaked iterators: current\n%s", v))
1658 1 : }
1659 2 : break
1660 : }
1661 0 : if refs != 0 {
1662 0 : err = firstError(err, errors.Errorf("leaked iterators:\n%s", v))
1663 0 : }
1664 : }
1665 :
1666 2 : for _, mem := range d.mu.mem.queue {
1667 2 : // Usually, we'd want to delete the files returned by readerUnref. But
1668 2 : // in this case, even if we're unreferencing the flushables, the
1669 2 : // flushables aren't obsolete. They will be reconstructed during WAL
1670 2 : // replay.
1671 2 : mem.readerUnrefLocked(false)
1672 2 : }
1673 : // If there's an unused, recycled memtable, we need to release its memory.
1674 2 : if obsoleteMemTable := d.memTableRecycle.Swap(nil); obsoleteMemTable != nil {
1675 2 : d.freeMemTable(obsoleteMemTable)
1676 2 : }
1677 2 : if reserved := d.memTableReserved.Load(); reserved != 0 {
1678 1 : err = firstError(err, errors.Errorf("leaked memtable reservation: %d", errors.Safe(reserved)))
1679 1 : }
1680 :
1681 : // Since we called d.readState.val.unrefLocked() above, we are expected to
1682 : // manually schedule deletion of obsolete files.
1683 2 : if len(d.mu.versions.obsoleteTables) > 0 {
1684 2 : d.deleteObsoleteFiles(d.mu.nextJobID)
1685 2 : }
1686 :
1687 2 : d.mu.Unlock()
1688 2 : d.compactionSchedulers.Wait()
1689 2 :
1690 2 : // Wait for all cleaning jobs to finish.
1691 2 : d.cleanupManager.Close()
1692 2 :
1693 2 : // Sanity check metrics.
1694 2 : if invariants.Enabled {
1695 2 : m := d.Metrics()
1696 2 : if m.Compact.NumInProgress > 0 || m.Compact.InProgressBytes > 0 {
1697 0 : d.mu.Lock()
1698 0 : panic(fmt.Sprintf("invalid metrics on close:\n%s", m))
1699 : }
1700 : }
1701 :
1702 2 : d.mu.Lock()
1703 2 :
1704 2 : // As a sanity check, ensure that there are no zombie tables. A non-zero count
1705 2 : // hints at a reference count leak.
1706 2 : if ztbls := len(d.mu.versions.zombieTables); ztbls > 0 {
1707 0 : err = firstError(err, errors.Errorf("non-zero zombie file count: %d", ztbls))
1708 0 : }
1709 :
1710 2 : err = firstError(err, d.objProvider.Close())
1711 2 :
1712 2 : // If the options include a closer to 'close' the filesystem, close it.
1713 2 : if d.opts.private.fsCloser != nil {
1714 2 : d.opts.private.fsCloser.Close()
1715 2 : }
1716 :
1717 : // Return an error if the user failed to close all open snapshots.
1718 2 : if v := d.mu.snapshots.count(); v > 0 {
1719 1 : err = firstError(err, errors.Errorf("leaked snapshots: %d open snapshots on DB %p", v, d))
1720 1 : }
1721 :
1722 2 : return err
1723 : }
1724 :
1725 : // Compact the specified range of keys in the database.
1726 2 : func (d *DB) Compact(start, end []byte, parallelize bool) error {
1727 2 : if err := d.closed.Load(); err != nil {
1728 1 : panic(err)
1729 : }
1730 2 : if d.opts.ReadOnly {
1731 1 : return ErrReadOnly
1732 1 : }
1733 2 : if d.cmp(start, end) >= 0 {
1734 2 : return errors.Errorf("Compact start %s is not less than end %s",
1735 2 : d.opts.Comparer.FormatKey(start), d.opts.Comparer.FormatKey(end))
1736 2 : }
1737 2 : iStart := base.MakeInternalKey(start, InternalKeySeqNumMax, InternalKeyKindMax)
1738 2 : iEnd := base.MakeInternalKey(end, 0, 0)
1739 2 : m := (&fileMetadata{}).ExtendPointKeyBounds(d.cmp, iStart, iEnd)
1740 2 : meta := []*fileMetadata{m}
1741 2 :
1742 2 : d.mu.Lock()
1743 2 : maxLevelWithFiles := 1
1744 2 : cur := d.mu.versions.currentVersion()
1745 2 : for level := 0; level < numLevels; level++ {
1746 2 : overlaps := cur.Overlaps(level, d.cmp, start, end, iEnd.IsExclusiveSentinel())
1747 2 : if !overlaps.Empty() {
1748 2 : maxLevelWithFiles = level + 1
1749 2 : }
1750 : }
1751 :
1752 2 : keyRanges := make([]internalKeyRange, len(meta))
1753 2 : for i := range meta {
1754 2 : keyRanges[i] = internalKeyRange{smallest: m.Smallest, largest: m.Largest}
1755 2 : }
1756 : // Determine if any memtable overlaps with the compaction range. We wait for
1757 : // any such overlap to flush (initiating a flush if necessary).
1758 2 : mem, err := func() (*flushableEntry, error) {
1759 2 : // Check to see if any files overlap with any of the memtables. The queue
1760 2 : // is ordered from oldest to newest with the mutable memtable being the
1761 2 : // last element in the slice. We want to wait for the newest table that
1762 2 : // overlaps.
1763 2 : for i := len(d.mu.mem.queue) - 1; i >= 0; i-- {
1764 2 : mem := d.mu.mem.queue[i]
1765 2 : if ingestMemtableOverlaps(d.cmp, mem, keyRanges) {
1766 2 : var err error
1767 2 : if mem.flushable == d.mu.mem.mutable {
1768 2 : // We have to hold both commitPipeline.mu and DB.mu when calling
1769 2 : // makeRoomForWrite(). Lock order requirements elsewhere force us to
1770 2 : // unlock DB.mu in order to grab commitPipeline.mu first.
1771 2 : d.mu.Unlock()
1772 2 : d.commit.mu.Lock()
1773 2 : d.mu.Lock()
1774 2 : defer d.commit.mu.Unlock()
1775 2 : if mem.flushable == d.mu.mem.mutable {
1776 2 : // Only flush if the active memtable is unchanged.
1777 2 : err = d.makeRoomForWrite(nil)
1778 2 : }
1779 : }
1780 2 : mem.flushForced = true
1781 2 : d.maybeScheduleFlush()
1782 2 : return mem, err
1783 : }
1784 : }
1785 2 : return nil, nil
1786 : }()
1787 :
1788 2 : d.mu.Unlock()
1789 2 :
1790 2 : if err != nil {
1791 0 : return err
1792 0 : }
1793 2 : if mem != nil {
1794 2 : <-mem.flushed
1795 2 : }
1796 :
1797 2 : for level := 0; level < maxLevelWithFiles; {
1798 2 : for {
1799 2 : if err := d.manualCompact(
1800 2 : iStart.UserKey, iEnd.UserKey, level, parallelize); err != nil {
1801 1 : if errors.Is(err, ErrCancelledCompaction) {
1802 1 : continue
1803 : }
1804 1 : return err
1805 : }
1806 2 : break
1807 : }
1808 2 : level++
1809 2 : if level == numLevels-1 {
1810 2 : // A manual compaction of the bottommost level occurred.
1811 2 : // There is no next level to try and compact.
1812 2 : break
1813 : }
1814 : }
1815 2 : return nil
1816 : }
1817 :
1818 2 : func (d *DB) manualCompact(start, end []byte, level int, parallelize bool) error {
1819 2 : d.mu.Lock()
1820 2 : curr := d.mu.versions.currentVersion()
1821 2 : files := curr.Overlaps(level, d.cmp, start, end, false)
1822 2 : if files.Empty() {
1823 2 : d.mu.Unlock()
1824 2 : return nil
1825 2 : }
1826 :
1827 2 : var compactions []*manualCompaction
1828 2 : if parallelize {
1829 2 : compactions = append(compactions, d.splitManualCompaction(start, end, level)...)
1830 2 : } else {
1831 2 : compactions = append(compactions, &manualCompaction{
1832 2 : level: level,
1833 2 : done: make(chan error, 1),
1834 2 : start: start,
1835 2 : end: end,
1836 2 : })
1837 2 : }
1838 2 : d.mu.compact.manual = append(d.mu.compact.manual, compactions...)
1839 2 : d.maybeScheduleCompaction()
1840 2 : d.mu.Unlock()
1841 2 :
1842 2 : // Each of the channels is guaranteed to be eventually sent to once. After a
1843 2 : // compaction is possibly picked in d.maybeScheduleCompaction(), either the
1844 2 : // compaction is dropped, executed after being scheduled, or retried later.
1845 2 : // Assuming eventual progress when a compaction is retried, all outcomes send
1846 2 : // a value to the done channel. Since the channels are buffered, it is not
1847 2 : // necessary to read from each channel, and so we can exit early in the event
1848 2 : // of an error.
1849 2 : for _, compaction := range compactions {
1850 2 : if err := <-compaction.done; err != nil {
1851 1 : return err
1852 1 : }
1853 : }
1854 2 : return nil
1855 : }
1856 :
1857 : // splitManualCompaction splits a manual compaction over [start,end] on level
1858 : // such that the resulting compactions have no key overlap.
1859 : func (d *DB) splitManualCompaction(
1860 : start, end []byte, level int,
1861 2 : ) (splitCompactions []*manualCompaction) {
1862 2 : curr := d.mu.versions.currentVersion()
1863 2 : endLevel := level + 1
1864 2 : baseLevel := d.mu.versions.picker.getBaseLevel()
1865 2 : if level == 0 {
1866 2 : endLevel = baseLevel
1867 2 : }
1868 2 : keyRanges := calculateInuseKeyRanges(curr, d.cmp, level, endLevel, start, end)
1869 2 : for _, keyRange := range keyRanges {
1870 2 : splitCompactions = append(splitCompactions, &manualCompaction{
1871 2 : level: level,
1872 2 : done: make(chan error, 1),
1873 2 : start: keyRange.Start,
1874 2 : end: keyRange.End,
1875 2 : split: true,
1876 2 : })
1877 2 : }
1878 2 : return splitCompactions
1879 : }
1880 :
1881 : // DownloadSpan is a key range passed to the Download method.
1882 : type DownloadSpan struct {
1883 : StartKey []byte
1884 : // EndKey is exclusive.
1885 : EndKey []byte
1886 : }
1887 :
1888 : // Download ensures that the LSM does not use any external sstables for the
1889 : // given key ranges. It does so by performing appropriate compactions so that
1890 : // all external data becomes available locally.
1891 : //
1892 : // Note that calling this method does not imply that all other compactions stop;
1893 : // it simply informs Pebble of a list of spans for which external data should be
1894 : // downloaded with high priority.
1895 : //
1896 : // The method returns once no external sstasbles overlap the given spans, the
1897 : // context is canceled, or an error is hit.
1898 : //
1899 : // TODO(radu): consider passing a priority/impact knob to express how important
1900 : // the download is (versus live traffic performance, LSM health).
1901 0 : func (d *DB) Download(ctx context.Context, spans []DownloadSpan) error {
1902 0 : return errors.Errorf("not implemented")
1903 0 : }
1904 :
1905 : // Flush the memtable to stable storage.
1906 2 : func (d *DB) Flush() error {
1907 2 : flushDone, err := d.AsyncFlush()
1908 2 : if err != nil {
1909 1 : return err
1910 1 : }
1911 2 : <-flushDone
1912 2 : return nil
1913 : }
1914 :
1915 : // AsyncFlush asynchronously flushes the memtable to stable storage.
1916 : //
1917 : // If no error is returned, the caller can receive from the returned channel in
1918 : // order to wait for the flush to complete.
1919 2 : func (d *DB) AsyncFlush() (<-chan struct{}, error) {
1920 2 : if err := d.closed.Load(); err != nil {
1921 1 : panic(err)
1922 : }
1923 2 : if d.opts.ReadOnly {
1924 1 : return nil, ErrReadOnly
1925 1 : }
1926 :
1927 2 : d.commit.mu.Lock()
1928 2 : defer d.commit.mu.Unlock()
1929 2 : d.mu.Lock()
1930 2 : defer d.mu.Unlock()
1931 2 : flushed := d.mu.mem.queue[len(d.mu.mem.queue)-1].flushed
1932 2 : err := d.makeRoomForWrite(nil)
1933 2 : if err != nil {
1934 0 : return nil, err
1935 0 : }
1936 2 : return flushed, nil
1937 : }
1938 :
1939 : // Metrics returns metrics about the database.
1940 2 : func (d *DB) Metrics() *Metrics {
1941 2 : metrics := &Metrics{}
1942 2 : recycledLogsCount, recycledLogSize := d.logRecycler.stats()
1943 2 :
1944 2 : d.mu.Lock()
1945 2 : vers := d.mu.versions.currentVersion()
1946 2 : *metrics = d.mu.versions.metrics
1947 2 : metrics.Compact.EstimatedDebt = d.mu.versions.picker.estimatedCompactionDebt(0)
1948 2 : metrics.Compact.InProgressBytes = d.mu.versions.atomicInProgressBytes.Load()
1949 2 : metrics.Compact.NumInProgress = int64(d.mu.compact.compactingCount)
1950 2 : metrics.Compact.MarkedFiles = vers.Stats.MarkedForCompaction
1951 2 : metrics.Compact.Duration = d.mu.compact.duration
1952 2 : for c := range d.mu.compact.inProgress {
1953 1 : if c.kind != compactionKindFlush {
1954 1 : metrics.Compact.Duration += d.timeNow().Sub(c.beganAt)
1955 1 : }
1956 : }
1957 :
1958 2 : for _, m := range d.mu.mem.queue {
1959 2 : metrics.MemTable.Size += m.totalBytes()
1960 2 : }
1961 2 : metrics.Snapshots.Count = d.mu.snapshots.count()
1962 2 : if metrics.Snapshots.Count > 0 {
1963 1 : metrics.Snapshots.EarliestSeqNum = d.mu.snapshots.earliest()
1964 1 : }
1965 2 : metrics.Snapshots.PinnedKeys = d.mu.snapshots.cumulativePinnedCount
1966 2 : metrics.Snapshots.PinnedSize = d.mu.snapshots.cumulativePinnedSize
1967 2 : metrics.MemTable.Count = int64(len(d.mu.mem.queue))
1968 2 : metrics.MemTable.ZombieCount = d.memTableCount.Load() - metrics.MemTable.Count
1969 2 : metrics.MemTable.ZombieSize = uint64(d.memTableReserved.Load()) - metrics.MemTable.Size
1970 2 : metrics.WAL.ObsoleteFiles = int64(recycledLogsCount)
1971 2 : metrics.WAL.ObsoletePhysicalSize = recycledLogSize
1972 2 : metrics.WAL.Size = d.logSize.Load()
1973 2 : // The current WAL size (d.atomic.logSize) is the current logical size,
1974 2 : // which may be less than the WAL's physical size if it was recycled.
1975 2 : // The file sizes in d.mu.log.queue are updated to the physical size
1976 2 : // during WAL rotation. Use the larger of the two for the current WAL. All
1977 2 : // the previous WALs's fileSizes in d.mu.log.queue are already updated.
1978 2 : metrics.WAL.PhysicalSize = metrics.WAL.Size
1979 2 : if len(d.mu.log.queue) > 0 && metrics.WAL.PhysicalSize < d.mu.log.queue[len(d.mu.log.queue)-1].fileSize {
1980 1 : metrics.WAL.PhysicalSize = d.mu.log.queue[len(d.mu.log.queue)-1].fileSize
1981 1 : }
1982 2 : for i, n := 0, len(d.mu.log.queue)-1; i < n; i++ {
1983 2 : metrics.WAL.PhysicalSize += d.mu.log.queue[i].fileSize
1984 2 : }
1985 :
1986 2 : metrics.WAL.BytesIn = d.mu.log.bytesIn // protected by d.mu
1987 2 : for i, n := 0, len(d.mu.mem.queue)-1; i < n; i++ {
1988 2 : metrics.WAL.Size += d.mu.mem.queue[i].logSize
1989 2 : }
1990 2 : metrics.WAL.BytesWritten = metrics.Levels[0].BytesIn + metrics.WAL.Size
1991 2 : if p := d.mu.versions.picker; p != nil {
1992 2 : compactions := d.getInProgressCompactionInfoLocked(nil)
1993 2 : for level, score := range p.getScores(compactions) {
1994 2 : metrics.Levels[level].Score = score
1995 2 : }
1996 : }
1997 2 : metrics.Table.ZombieCount = int64(len(d.mu.versions.zombieTables))
1998 2 : for _, size := range d.mu.versions.zombieTables {
1999 1 : metrics.Table.ZombieSize += size
2000 1 : }
2001 2 : metrics.private.optionsFileSize = d.optionsFileSize
2002 2 :
2003 2 : // TODO(jackson): Consider making these metrics optional.
2004 2 : metrics.Keys.RangeKeySetsCount = countRangeKeySetFragments(vers)
2005 2 : metrics.Keys.TombstoneCount = countTombstones(vers)
2006 2 :
2007 2 : d.mu.versions.logLock()
2008 2 : metrics.private.manifestFileSize = uint64(d.mu.versions.manifest.Size())
2009 2 : metrics.Table.BackingTableCount = uint64(len(d.mu.versions.backingState.fileBackingMap))
2010 2 : metrics.Table.BackingTableSize = d.mu.versions.backingState.fileBackingSize
2011 2 : if invariants.Enabled {
2012 2 : var totalSize uint64
2013 2 : for _, backing := range d.mu.versions.backingState.fileBackingMap {
2014 2 : totalSize += backing.Size
2015 2 : }
2016 2 : if totalSize != metrics.Table.BackingTableSize {
2017 0 : panic("pebble: invalid backing table size accounting")
2018 : }
2019 : }
2020 2 : d.mu.versions.logUnlock()
2021 2 :
2022 2 : metrics.LogWriter.FsyncLatency = d.mu.log.metrics.fsyncLatency
2023 2 : if err := metrics.LogWriter.Merge(&d.mu.log.metrics.LogWriterMetrics); err != nil {
2024 0 : d.opts.Logger.Errorf("metrics error: %s", err)
2025 0 : }
2026 2 : metrics.Flush.WriteThroughput = d.mu.compact.flushWriteThroughput
2027 2 : if d.mu.compact.flushing {
2028 1 : metrics.Flush.NumInProgress = 1
2029 1 : }
2030 2 : for i := 0; i < numLevels; i++ {
2031 2 : metrics.Levels[i].Additional.ValueBlocksSize = valueBlocksSizeForLevel(vers, i)
2032 2 : }
2033 :
2034 2 : d.mu.Unlock()
2035 2 :
2036 2 : metrics.BlockCache = d.opts.Cache.Metrics()
2037 2 : metrics.TableCache, metrics.Filter = d.tableCache.metrics()
2038 2 : metrics.TableIters = int64(d.tableCache.iterCount())
2039 2 : metrics.CategoryStats = d.tableCache.dbOpts.sstStatsCollector.GetStats()
2040 2 :
2041 2 : metrics.SecondaryCacheMetrics = d.objProvider.Metrics()
2042 2 :
2043 2 : metrics.Uptime = d.timeNow().Sub(d.openedAt)
2044 2 :
2045 2 : return metrics
2046 : }
2047 :
2048 : // sstablesOptions hold the optional parameters to retrieve TableInfo for all sstables.
2049 : type sstablesOptions struct {
2050 : // set to true will return the sstable properties in TableInfo
2051 : withProperties bool
2052 :
2053 : // if set, return sstables that overlap the key range (end-exclusive)
2054 : start []byte
2055 : end []byte
2056 :
2057 : withApproximateSpanBytes bool
2058 : }
2059 :
2060 : // SSTablesOption set optional parameter used by `DB.SSTables`.
2061 : type SSTablesOption func(*sstablesOptions)
2062 :
2063 : // WithProperties enable return sstable properties in each TableInfo.
2064 : //
2065 : // NOTE: if most of the sstable properties need to be read from disk,
2066 : // this options may make method `SSTables` quite slow.
2067 1 : func WithProperties() SSTablesOption {
2068 1 : return func(opt *sstablesOptions) {
2069 1 : opt.withProperties = true
2070 1 : }
2071 : }
2072 :
2073 : // WithKeyRangeFilter ensures returned sstables overlap start and end (end-exclusive)
2074 : // if start and end are both nil these properties have no effect.
2075 1 : func WithKeyRangeFilter(start, end []byte) SSTablesOption {
2076 1 : return func(opt *sstablesOptions) {
2077 1 : opt.end = end
2078 1 : opt.start = start
2079 1 : }
2080 : }
2081 :
2082 : // WithApproximateSpanBytes enables capturing the approximate number of bytes that
2083 : // overlap the provided key span for each sstable.
2084 : // NOTE: this option can only be used with WithKeyRangeFilter and WithProperties
2085 : // provided.
2086 1 : func WithApproximateSpanBytes() SSTablesOption {
2087 1 : return func(opt *sstablesOptions) {
2088 1 : opt.withApproximateSpanBytes = true
2089 1 : }
2090 : }
2091 :
2092 : // BackingType denotes the type of storage backing a given sstable.
2093 : type BackingType int
2094 :
2095 : const (
2096 : // BackingTypeLocal denotes an sstable stored on local disk according to the
2097 : // objprovider. This file is completely owned by us.
2098 : BackingTypeLocal BackingType = iota
2099 : // BackingTypeShared denotes an sstable stored on shared storage, created
2100 : // by this Pebble instance and possibly shared by other Pebble instances.
2101 : // These types of files have lifecycle managed by Pebble.
2102 : BackingTypeShared
2103 : // BackingTypeSharedForeign denotes an sstable stored on shared storage,
2104 : // created by a Pebble instance other than this one. These types of files have
2105 : // lifecycle managed by Pebble.
2106 : BackingTypeSharedForeign
2107 : // BackingTypeExternal denotes an sstable stored on external storage,
2108 : // not owned by any Pebble instance and with no refcounting/cleanup methods
2109 : // or lifecycle management. An example of an external file is a file restored
2110 : // from a backup.
2111 : BackingTypeExternal
2112 : )
2113 :
2114 : // SSTableInfo export manifest.TableInfo with sstable.Properties alongside
2115 : // other file backing info.
2116 : type SSTableInfo struct {
2117 : manifest.TableInfo
2118 : // Virtual indicates whether the sstable is virtual.
2119 : Virtual bool
2120 : // BackingSSTNum is the file number associated with backing sstable which
2121 : // backs the sstable associated with this SSTableInfo. If Virtual is false,
2122 : // then BackingSSTNum == FileNum.
2123 : BackingSSTNum base.FileNum
2124 : // BackingType is the type of storage backing this sstable.
2125 : BackingType BackingType
2126 : // Locator is the remote.Locator backing this sstable, if the backing type is
2127 : // not BackingTypeLocal.
2128 : Locator remote.Locator
2129 :
2130 : // Properties is the sstable properties of this table. If Virtual is true,
2131 : // then the Properties are associated with the backing sst.
2132 : Properties *sstable.Properties
2133 : }
2134 :
2135 : // SSTables retrieves the current sstables. The returned slice is indexed by
2136 : // level and each level is indexed by the position of the sstable within the
2137 : // level. Note that this information may be out of date due to concurrent
2138 : // flushes and compactions.
2139 1 : func (d *DB) SSTables(opts ...SSTablesOption) ([][]SSTableInfo, error) {
2140 1 : opt := &sstablesOptions{}
2141 1 : for _, fn := range opts {
2142 1 : fn(opt)
2143 1 : }
2144 :
2145 1 : if opt.withApproximateSpanBytes && !opt.withProperties {
2146 1 : return nil, errors.Errorf("Cannot use WithApproximateSpanBytes without WithProperties option.")
2147 1 : }
2148 1 : if opt.withApproximateSpanBytes && (opt.start == nil || opt.end == nil) {
2149 1 : return nil, errors.Errorf("Cannot use WithApproximateSpanBytes without WithKeyRangeFilter option.")
2150 1 : }
2151 :
2152 : // Grab and reference the current readState.
2153 1 : readState := d.loadReadState()
2154 1 : defer readState.unref()
2155 1 :
2156 1 : // TODO(peter): This is somewhat expensive, especially on a large
2157 1 : // database. It might be worthwhile to unify TableInfo and FileMetadata and
2158 1 : // then we could simply return current.Files. Note that RocksDB is doing
2159 1 : // something similar to the current code, so perhaps it isn't too bad.
2160 1 : srcLevels := readState.current.Levels
2161 1 : var totalTables int
2162 1 : for i := range srcLevels {
2163 1 : totalTables += srcLevels[i].Len()
2164 1 : }
2165 :
2166 1 : destTables := make([]SSTableInfo, totalTables)
2167 1 : destLevels := make([][]SSTableInfo, len(srcLevels))
2168 1 : for i := range destLevels {
2169 1 : iter := srcLevels[i].Iter()
2170 1 : j := 0
2171 1 : for m := iter.First(); m != nil; m = iter.Next() {
2172 1 : if opt.start != nil && opt.end != nil && !m.Overlaps(d.opts.Comparer.Compare, opt.start, opt.end, true /* exclusive end */) {
2173 1 : continue
2174 : }
2175 1 : destTables[j] = SSTableInfo{TableInfo: m.TableInfo()}
2176 1 : if opt.withProperties {
2177 1 : p, err := d.tableCache.getTableProperties(
2178 1 : m,
2179 1 : )
2180 1 : if err != nil {
2181 0 : return nil, err
2182 0 : }
2183 1 : destTables[j].Properties = p
2184 : }
2185 1 : destTables[j].Virtual = m.Virtual
2186 1 : destTables[j].BackingSSTNum = m.FileBacking.DiskFileNum.FileNum()
2187 1 : objMeta, err := d.objProvider.Lookup(fileTypeTable, m.FileBacking.DiskFileNum)
2188 1 : if err != nil {
2189 0 : return nil, err
2190 0 : }
2191 1 : if objMeta.IsRemote() {
2192 0 : if objMeta.IsShared() {
2193 0 : if d.objProvider.IsSharedForeign(objMeta) {
2194 0 : destTables[j].BackingType = BackingTypeSharedForeign
2195 0 : } else {
2196 0 : destTables[j].BackingType = BackingTypeShared
2197 0 : }
2198 0 : } else {
2199 0 : destTables[j].BackingType = BackingTypeExternal
2200 0 : }
2201 0 : destTables[j].Locator = objMeta.Remote.Locator
2202 1 : } else {
2203 1 : destTables[j].BackingType = BackingTypeLocal
2204 1 : }
2205 :
2206 1 : if opt.withApproximateSpanBytes {
2207 1 : var spanBytes uint64
2208 1 : if m.ContainedWithinSpan(d.opts.Comparer.Compare, opt.start, opt.end) {
2209 1 : spanBytes = m.Size
2210 1 : } else {
2211 1 : size, err := d.tableCache.estimateSize(m, opt.start, opt.end)
2212 1 : if err != nil {
2213 0 : return nil, err
2214 0 : }
2215 1 : spanBytes = size
2216 : }
2217 1 : propertiesCopy := *destTables[j].Properties
2218 1 :
2219 1 : // Deep copy user properties so approximate span bytes can be added.
2220 1 : propertiesCopy.UserProperties = make(map[string]string, len(destTables[j].Properties.UserProperties)+1)
2221 1 : for k, v := range destTables[j].Properties.UserProperties {
2222 0 : propertiesCopy.UserProperties[k] = v
2223 0 : }
2224 1 : propertiesCopy.UserProperties["approximate-span-bytes"] = strconv.FormatUint(spanBytes, 10)
2225 1 : destTables[j].Properties = &propertiesCopy
2226 : }
2227 1 : j++
2228 : }
2229 1 : destLevels[i] = destTables[:j]
2230 1 : destTables = destTables[j:]
2231 : }
2232 :
2233 1 : return destLevels, nil
2234 : }
2235 :
2236 : // EstimateDiskUsage returns the estimated filesystem space used in bytes for
2237 : // storing the range `[start, end]`. The estimation is computed as follows:
2238 : //
2239 : // - For sstables fully contained in the range the whole file size is included.
2240 : // - For sstables partially contained in the range the overlapping data block sizes
2241 : // are included. Even if a data block partially overlaps, or we cannot determine
2242 : // overlap due to abbreviated index keys, the full data block size is included in
2243 : // the estimation. Note that unlike fully contained sstables, none of the
2244 : // meta-block space is counted for partially overlapped files.
2245 : // - For virtual sstables, we use the overlap between start, end and the virtual
2246 : // sstable bounds to determine disk usage.
2247 : // - There may also exist WAL entries for unflushed keys in this range. This
2248 : // estimation currently excludes space used for the range in the WAL.
2249 1 : func (d *DB) EstimateDiskUsage(start, end []byte) (uint64, error) {
2250 1 : bytes, _, _, err := d.EstimateDiskUsageByBackingType(start, end)
2251 1 : return bytes, err
2252 1 : }
2253 :
2254 : // EstimateDiskUsageByBackingType is like EstimateDiskUsage but additionally
2255 : // returns the subsets of that size in remote ane external files.
2256 : func (d *DB) EstimateDiskUsageByBackingType(
2257 : start, end []byte,
2258 1 : ) (totalSize, remoteSize, externalSize uint64, _ error) {
2259 1 : if err := d.closed.Load(); err != nil {
2260 0 : panic(err)
2261 : }
2262 1 : if d.opts.Comparer.Compare(start, end) > 0 {
2263 0 : return 0, 0, 0, errors.New("invalid key-range specified (start > end)")
2264 0 : }
2265 :
2266 : // Grab and reference the current readState. This prevents the underlying
2267 : // files in the associated version from being deleted if there is a concurrent
2268 : // compaction.
2269 1 : readState := d.loadReadState()
2270 1 : defer readState.unref()
2271 1 :
2272 1 : for level, files := range readState.current.Levels {
2273 1 : iter := files.Iter()
2274 1 : if level > 0 {
2275 1 : // We can only use `Overlaps` to restrict `files` at L1+ since at L0 it
2276 1 : // expands the range iteratively until it has found a set of files that
2277 1 : // do not overlap any other L0 files outside that set.
2278 1 : overlaps := readState.current.Overlaps(level, d.opts.Comparer.Compare, start, end, false /* exclusiveEnd */)
2279 1 : iter = overlaps.Iter()
2280 1 : }
2281 1 : for file := iter.First(); file != nil; file = iter.Next() {
2282 1 : if d.opts.Comparer.Compare(start, file.Smallest.UserKey) <= 0 &&
2283 1 : d.opts.Comparer.Compare(file.Largest.UserKey, end) <= 0 {
2284 1 : // The range fully contains the file, so skip looking it up in
2285 1 : // table cache/looking at its indexes, and add the full file size.
2286 1 : meta, err := d.objProvider.Lookup(fileTypeTable, file.FileBacking.DiskFileNum)
2287 1 : if err != nil {
2288 0 : return 0, 0, 0, err
2289 0 : }
2290 1 : if meta.IsRemote() {
2291 0 : remoteSize += file.Size
2292 0 : if meta.Remote.CleanupMethod == objstorage.SharedNoCleanup {
2293 0 : externalSize += file.Size
2294 0 : }
2295 : }
2296 1 : totalSize += file.Size
2297 1 : } else if d.opts.Comparer.Compare(file.Smallest.UserKey, end) <= 0 &&
2298 1 : d.opts.Comparer.Compare(start, file.Largest.UserKey) <= 0 {
2299 1 : var size uint64
2300 1 : var err error
2301 1 : if file.Virtual {
2302 0 : err = d.tableCache.withVirtualReader(
2303 0 : file.VirtualMeta(),
2304 0 : func(r sstable.VirtualReader) (err error) {
2305 0 : size, err = r.EstimateDiskUsage(start, end)
2306 0 : return err
2307 0 : },
2308 : )
2309 1 : } else {
2310 1 : err = d.tableCache.withReader(
2311 1 : file.PhysicalMeta(),
2312 1 : func(r *sstable.Reader) (err error) {
2313 1 : size, err = r.EstimateDiskUsage(start, end)
2314 1 : return err
2315 1 : },
2316 : )
2317 : }
2318 1 : if err != nil {
2319 0 : return 0, 0, 0, err
2320 0 : }
2321 1 : meta, err := d.objProvider.Lookup(fileTypeTable, file.FileBacking.DiskFileNum)
2322 1 : if err != nil {
2323 0 : return 0, 0, 0, err
2324 0 : }
2325 1 : if meta.IsRemote() {
2326 0 : remoteSize += size
2327 0 : if meta.Remote.CleanupMethod == objstorage.SharedNoCleanup {
2328 0 : externalSize += size
2329 0 : }
2330 : }
2331 1 : totalSize += size
2332 : }
2333 : }
2334 : }
2335 1 : return totalSize, remoteSize, externalSize, nil
2336 : }
2337 :
2338 2 : func (d *DB) walPreallocateSize() int {
2339 2 : // Set the WAL preallocate size to 110% of the memtable size. Note that there
2340 2 : // is a bit of apples and oranges in units here as the memtabls size
2341 2 : // corresponds to the memory usage of the memtable while the WAL size is the
2342 2 : // size of the batches (plus overhead) stored in the WAL.
2343 2 : //
2344 2 : // TODO(peter): 110% of the memtable size is quite hefty for a block
2345 2 : // size. This logic is taken from GetWalPreallocateBlockSize in
2346 2 : // RocksDB. Could a smaller preallocation block size be used?
2347 2 : size := d.opts.MemTableSize
2348 2 : size = (size / 10) + size
2349 2 : return int(size)
2350 2 : }
2351 :
2352 2 : func (d *DB) newMemTable(logNum base.DiskFileNum, logSeqNum uint64) (*memTable, *flushableEntry) {
2353 2 : size := d.mu.mem.nextSize
2354 2 : if d.mu.mem.nextSize < d.opts.MemTableSize {
2355 2 : d.mu.mem.nextSize *= 2
2356 2 : if d.mu.mem.nextSize > d.opts.MemTableSize {
2357 1 : d.mu.mem.nextSize = d.opts.MemTableSize
2358 1 : }
2359 : }
2360 :
2361 2 : memtblOpts := memTableOptions{
2362 2 : Options: d.opts,
2363 2 : logSeqNum: logSeqNum,
2364 2 : }
2365 2 :
2366 2 : // Before attempting to allocate a new memtable, check if there's one
2367 2 : // available for recycling in memTableRecycle. Large contiguous allocations
2368 2 : // can be costly as fragmentation makes it more difficult to find a large
2369 2 : // contiguous free space. We've observed 64MB allocations taking 10ms+.
2370 2 : //
2371 2 : // To reduce these costly allocations, up to 1 obsolete memtable is stashed
2372 2 : // in `d.memTableRecycle` to allow a future memtable rotation to reuse
2373 2 : // existing memory.
2374 2 : var mem *memTable
2375 2 : mem = d.memTableRecycle.Swap(nil)
2376 2 : if mem != nil && uint64(len(mem.arenaBuf)) != size {
2377 2 : d.freeMemTable(mem)
2378 2 : mem = nil
2379 2 : }
2380 2 : if mem != nil {
2381 2 : // Carry through the existing buffer and memory reservation.
2382 2 : memtblOpts.arenaBuf = mem.arenaBuf
2383 2 : memtblOpts.releaseAccountingReservation = mem.releaseAccountingReservation
2384 2 : } else {
2385 2 : mem = new(memTable)
2386 2 : memtblOpts.arenaBuf = manual.New(int(size))
2387 2 : memtblOpts.releaseAccountingReservation = d.opts.Cache.Reserve(int(size))
2388 2 : d.memTableCount.Add(1)
2389 2 : d.memTableReserved.Add(int64(size))
2390 2 :
2391 2 : // Note: this is a no-op if invariants are disabled or race is enabled.
2392 2 : invariants.SetFinalizer(mem, checkMemTable)
2393 2 : }
2394 2 : mem.init(memtblOpts)
2395 2 :
2396 2 : entry := d.newFlushableEntry(mem, logNum, logSeqNum)
2397 2 : entry.releaseMemAccounting = func() {
2398 2 : // If the user leaks iterators, we may be releasing the memtable after
2399 2 : // the DB is already closed. In this case, we want to just release the
2400 2 : // memory because DB.Close won't come along to free it for us.
2401 2 : if err := d.closed.Load(); err != nil {
2402 2 : d.freeMemTable(mem)
2403 2 : return
2404 2 : }
2405 :
2406 : // The next memtable allocation might be able to reuse this memtable.
2407 : // Stash it on d.memTableRecycle.
2408 2 : if unusedMem := d.memTableRecycle.Swap(mem); unusedMem != nil {
2409 2 : // There was already a memtable waiting to be recycled. We're now
2410 2 : // responsible for freeing it.
2411 2 : d.freeMemTable(unusedMem)
2412 2 : }
2413 : }
2414 2 : return mem, entry
2415 : }
2416 :
2417 2 : func (d *DB) freeMemTable(m *memTable) {
2418 2 : d.memTableCount.Add(-1)
2419 2 : d.memTableReserved.Add(-int64(len(m.arenaBuf)))
2420 2 : m.free()
2421 2 : }
2422 :
2423 : func (d *DB) newFlushableEntry(
2424 : f flushable, logNum base.DiskFileNum, logSeqNum uint64,
2425 2 : ) *flushableEntry {
2426 2 : fe := &flushableEntry{
2427 2 : flushable: f,
2428 2 : flushed: make(chan struct{}),
2429 2 : logNum: logNum,
2430 2 : logSeqNum: logSeqNum,
2431 2 : deleteFn: d.mu.versions.addObsolete,
2432 2 : deleteFnLocked: d.mu.versions.addObsoleteLocked,
2433 2 : }
2434 2 : fe.readerRefs.Store(1)
2435 2 : return fe
2436 2 : }
2437 :
2438 : // makeRoomForWrite ensures that the memtable has room to hold the contents of
2439 : // Batch. It reserves the space in the memtable and adds a reference to the
2440 : // memtable. The caller must later ensure that the memtable is unreferenced. If
2441 : // the memtable is full, or a nil Batch is provided, the current memtable is
2442 : // rotated (marked as immutable) and a new mutable memtable is allocated. This
2443 : // memtable rotation also causes a log rotation.
2444 : //
2445 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2446 : // may be released and reacquired.
2447 2 : func (d *DB) makeRoomForWrite(b *Batch) error {
2448 2 : if b != nil && b.ingestedSSTBatch {
2449 0 : panic("pebble: invalid function call")
2450 : }
2451 :
2452 2 : force := b == nil || b.flushable != nil
2453 2 : stalled := false
2454 2 : for {
2455 2 : if b != nil && b.flushable == nil {
2456 2 : err := d.mu.mem.mutable.prepare(b)
2457 2 : if err != arenaskl.ErrArenaFull {
2458 2 : if stalled {
2459 2 : d.opts.EventListener.WriteStallEnd()
2460 2 : }
2461 2 : return err
2462 : }
2463 2 : } else if !force {
2464 2 : if stalled {
2465 2 : d.opts.EventListener.WriteStallEnd()
2466 2 : }
2467 2 : return nil
2468 : }
2469 : // force || err == ErrArenaFull, so we need to rotate the current memtable.
2470 2 : {
2471 2 : var size uint64
2472 2 : for i := range d.mu.mem.queue {
2473 2 : size += d.mu.mem.queue[i].totalBytes()
2474 2 : }
2475 2 : if size >= uint64(d.opts.MemTableStopWritesThreshold)*d.opts.MemTableSize {
2476 2 : // We have filled up the current memtable, but already queued memtables
2477 2 : // are still flushing, so we wait.
2478 2 : if !stalled {
2479 2 : stalled = true
2480 2 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2481 2 : Reason: "memtable count limit reached",
2482 2 : })
2483 2 : }
2484 2 : now := time.Now()
2485 2 : d.mu.compact.cond.Wait()
2486 2 : if b != nil {
2487 2 : b.commitStats.MemTableWriteStallDuration += time.Since(now)
2488 2 : }
2489 2 : continue
2490 : }
2491 : }
2492 2 : l0ReadAmp := d.mu.versions.currentVersion().L0Sublevels.ReadAmplification()
2493 2 : if l0ReadAmp >= d.opts.L0StopWritesThreshold {
2494 2 : // There are too many level-0 files, so we wait.
2495 2 : if !stalled {
2496 2 : stalled = true
2497 2 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2498 2 : Reason: "L0 file count limit exceeded",
2499 2 : })
2500 2 : }
2501 2 : now := time.Now()
2502 2 : d.mu.compact.cond.Wait()
2503 2 : if b != nil {
2504 1 : b.commitStats.L0ReadAmpWriteStallDuration += time.Since(now)
2505 1 : }
2506 2 : continue
2507 : }
2508 :
2509 2 : var newLogNum base.DiskFileNum
2510 2 : var prevLogSize uint64
2511 2 : if !d.opts.DisableWAL {
2512 2 : now := time.Now()
2513 2 : newLogNum, prevLogSize = d.recycleWAL()
2514 2 : if b != nil {
2515 2 : b.commitStats.WALRotationDuration += time.Since(now)
2516 2 : }
2517 : }
2518 :
2519 2 : immMem := d.mu.mem.mutable
2520 2 : imm := d.mu.mem.queue[len(d.mu.mem.queue)-1]
2521 2 : imm.logSize = prevLogSize
2522 2 : imm.flushForced = imm.flushForced || (b == nil)
2523 2 :
2524 2 : // If we are manually flushing and we used less than half of the bytes in
2525 2 : // the memtable, don't increase the size for the next memtable. This
2526 2 : // reduces memtable memory pressure when an application is frequently
2527 2 : // manually flushing.
2528 2 : if (b == nil) && uint64(immMem.availBytes()) > immMem.totalBytes()/2 {
2529 2 : d.mu.mem.nextSize = immMem.totalBytes()
2530 2 : }
2531 :
2532 2 : if b != nil && b.flushable != nil {
2533 2 : // The batch is too large to fit in the memtable so add it directly to
2534 2 : // the immutable queue. The flushable batch is associated with the same
2535 2 : // log as the immutable memtable, but logically occurs after it in
2536 2 : // seqnum space. We ensure while flushing that the flushable batch
2537 2 : // is flushed along with the previous memtable in the flushable
2538 2 : // queue. See the top level comment in DB.flush1 to learn how this
2539 2 : // is ensured.
2540 2 : //
2541 2 : // See DB.commitWrite for the special handling of log writes for large
2542 2 : // batches. In particular, the large batch has already written to
2543 2 : // imm.logNum.
2544 2 : entry := d.newFlushableEntry(b.flushable, imm.logNum, b.SeqNum())
2545 2 : // The large batch is by definition large. Reserve space from the cache
2546 2 : // for it until it is flushed.
2547 2 : entry.releaseMemAccounting = d.opts.Cache.Reserve(int(b.flushable.totalBytes()))
2548 2 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2549 2 : }
2550 :
2551 2 : var logSeqNum uint64
2552 2 : if b != nil {
2553 2 : logSeqNum = b.SeqNum()
2554 2 : if b.flushable != nil {
2555 2 : logSeqNum += uint64(b.Count())
2556 2 : }
2557 2 : } else {
2558 2 : logSeqNum = d.mu.versions.logSeqNum.Load()
2559 2 : }
2560 2 : d.rotateMemtable(newLogNum, logSeqNum, immMem)
2561 2 : force = false
2562 : }
2563 : }
2564 :
2565 : // Both DB.mu and commitPipeline.mu must be held by the caller.
2566 2 : func (d *DB) rotateMemtable(newLogNum base.DiskFileNum, logSeqNum uint64, prev *memTable) {
2567 2 : // Create a new memtable, scheduling the previous one for flushing. We do
2568 2 : // this even if the previous memtable was empty because the DB.Flush
2569 2 : // mechanism is dependent on being able to wait for the empty memtable to
2570 2 : // flush. We can't just mark the empty memtable as flushed here because we
2571 2 : // also have to wait for all previous immutable tables to
2572 2 : // flush. Additionally, the memtable is tied to particular WAL file and we
2573 2 : // want to go through the flush path in order to recycle that WAL file.
2574 2 : //
2575 2 : // NB: newLogNum corresponds to the WAL that contains mutations that are
2576 2 : // present in the new memtable. When immutable memtables are flushed to
2577 2 : // disk, a VersionEdit will be created telling the manifest the minimum
2578 2 : // unflushed log number (which will be the next one in d.mu.mem.mutable
2579 2 : // that was not flushed).
2580 2 : //
2581 2 : // NB: prev should be the current mutable memtable.
2582 2 : var entry *flushableEntry
2583 2 : d.mu.mem.mutable, entry = d.newMemTable(newLogNum, logSeqNum)
2584 2 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2585 2 : d.updateReadStateLocked(nil)
2586 2 : if prev.writerUnref() {
2587 2 : d.maybeScheduleFlush()
2588 2 : }
2589 : }
2590 :
2591 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2592 : // may be released and reacquired.
2593 2 : func (d *DB) recycleWAL() (newLogNum base.DiskFileNum, prevLogSize uint64) {
2594 2 : if d.opts.DisableWAL {
2595 0 : panic("pebble: invalid function call")
2596 : }
2597 :
2598 2 : jobID := d.mu.nextJobID
2599 2 : d.mu.nextJobID++
2600 2 : newLogNum = d.mu.versions.getNextDiskFileNum()
2601 2 :
2602 2 : prevLogSize = uint64(d.mu.log.Size())
2603 2 :
2604 2 : // The previous log may have grown past its original physical
2605 2 : // size. Update its file size in the queue so we have a proper
2606 2 : // accounting of its file size.
2607 2 : if d.mu.log.queue[len(d.mu.log.queue)-1].fileSize < prevLogSize {
2608 2 : d.mu.log.queue[len(d.mu.log.queue)-1].fileSize = prevLogSize
2609 2 : }
2610 2 : d.mu.Unlock()
2611 2 :
2612 2 : var err error
2613 2 : // Close the previous log first. This writes an EOF trailer
2614 2 : // signifying the end of the file and syncs it to disk. We must
2615 2 : // close the previous log before linking the new log file,
2616 2 : // otherwise a crash could leave both logs with unclean tails, and
2617 2 : // Open will treat the previous log as corrupt.
2618 2 : err = d.mu.log.LogWriter.Close()
2619 2 : metrics := d.mu.log.LogWriter.Metrics()
2620 2 : d.mu.Lock()
2621 2 : if err := d.mu.log.metrics.Merge(metrics); err != nil {
2622 0 : d.opts.Logger.Errorf("metrics error: %s", err)
2623 0 : }
2624 2 : d.mu.Unlock()
2625 2 :
2626 2 : newLogName := base.MakeFilepath(d.opts.FS, d.walDirname, fileTypeLog, newLogNum)
2627 2 :
2628 2 : // Try to use a recycled log file. Recycling log files is an important
2629 2 : // performance optimization as it is faster to sync a file that has
2630 2 : // already been written, than one which is being written for the first
2631 2 : // time. This is due to the need to sync file metadata when a file is
2632 2 : // being written for the first time. Note this is true even if file
2633 2 : // preallocation is performed (e.g. fallocate).
2634 2 : var recycleLog fileInfo
2635 2 : var recycleOK bool
2636 2 : var newLogFile vfs.File
2637 2 : if err == nil {
2638 2 : recycleLog, recycleOK = d.logRecycler.peek()
2639 2 : if recycleOK {
2640 1 : recycleLogName := base.MakeFilepath(d.opts.FS, d.walDirname, fileTypeLog, recycleLog.fileNum)
2641 1 : newLogFile, err = d.opts.FS.ReuseForWrite(recycleLogName, newLogName)
2642 1 : base.MustExist(d.opts.FS, newLogName, d.opts.Logger, err)
2643 2 : } else {
2644 2 : newLogFile, err = d.opts.FS.Create(newLogName)
2645 2 : base.MustExist(d.opts.FS, newLogName, d.opts.Logger, err)
2646 2 : }
2647 : }
2648 :
2649 2 : var newLogSize uint64
2650 2 : if err == nil && recycleOK {
2651 1 : // Figure out the recycled WAL size. This Stat is necessary
2652 1 : // because ReuseForWrite's contract allows for removing the
2653 1 : // old file and creating a new one. We don't know whether the
2654 1 : // WAL was actually recycled.
2655 1 : // TODO(jackson): Adding a boolean to the ReuseForWrite return
2656 1 : // value indicating whether or not the file was actually
2657 1 : // reused would allow us to skip the stat and use
2658 1 : // recycleLog.fileSize.
2659 1 : var finfo os.FileInfo
2660 1 : finfo, err = newLogFile.Stat()
2661 1 : if err == nil {
2662 1 : newLogSize = uint64(finfo.Size())
2663 1 : }
2664 : }
2665 :
2666 2 : if err == nil {
2667 2 : // TODO(peter): RocksDB delays sync of the parent directory until the
2668 2 : // first time the log is synced. Is that worthwhile?
2669 2 : err = d.walDir.Sync()
2670 2 : }
2671 :
2672 2 : if err != nil && newLogFile != nil {
2673 0 : newLogFile.Close()
2674 2 : } else if err == nil {
2675 2 : newLogFile = vfs.NewSyncingFile(newLogFile, vfs.SyncingFileOptions{
2676 2 : NoSyncOnClose: d.opts.NoSyncOnClose,
2677 2 : BytesPerSync: d.opts.WALBytesPerSync,
2678 2 : PreallocateSize: d.walPreallocateSize(),
2679 2 : })
2680 2 : }
2681 :
2682 2 : if recycleOK {
2683 1 : err = firstError(err, d.logRecycler.pop(recycleLog.fileNum.FileNum()))
2684 1 : }
2685 :
2686 2 : d.opts.EventListener.WALCreated(WALCreateInfo{
2687 2 : JobID: jobID,
2688 2 : Path: newLogName,
2689 2 : FileNum: newLogNum,
2690 2 : RecycledFileNum: recycleLog.fileNum.FileNum(),
2691 2 : Err: err,
2692 2 : })
2693 2 :
2694 2 : d.mu.Lock()
2695 2 :
2696 2 : d.mu.versions.metrics.WAL.Files++
2697 2 :
2698 2 : if err != nil {
2699 0 : // TODO(peter): avoid chewing through file numbers in a tight loop if there
2700 0 : // is an error here.
2701 0 : //
2702 0 : // What to do here? Stumbling on doesn't seem worthwhile. If we failed to
2703 0 : // close the previous log it is possible we lost a write.
2704 0 : panic(err)
2705 : }
2706 :
2707 2 : d.mu.log.queue = append(d.mu.log.queue, fileInfo{fileNum: newLogNum, fileSize: newLogSize})
2708 2 : d.mu.log.LogWriter = record.NewLogWriter(newLogFile, newLogNum, record.LogWriterConfig{
2709 2 : WALFsyncLatency: d.mu.log.metrics.fsyncLatency,
2710 2 : WALMinSyncInterval: d.opts.WALMinSyncInterval,
2711 2 : QueueSemChan: d.commit.logSyncQSem,
2712 2 : })
2713 2 : if d.mu.log.registerLogWriterForTesting != nil {
2714 0 : d.mu.log.registerLogWriterForTesting(d.mu.log.LogWriter)
2715 0 : }
2716 :
2717 2 : return
2718 : }
2719 :
2720 2 : func (d *DB) getEarliestUnflushedSeqNumLocked() uint64 {
2721 2 : seqNum := InternalKeySeqNumMax
2722 2 : for i := range d.mu.mem.queue {
2723 2 : logSeqNum := d.mu.mem.queue[i].logSeqNum
2724 2 : if seqNum > logSeqNum {
2725 2 : seqNum = logSeqNum
2726 2 : }
2727 : }
2728 2 : return seqNum
2729 : }
2730 :
2731 2 : func (d *DB) getInProgressCompactionInfoLocked(finishing *compaction) (rv []compactionInfo) {
2732 2 : for c := range d.mu.compact.inProgress {
2733 2 : if len(c.flushing) == 0 && (finishing == nil || c != finishing) {
2734 2 : info := compactionInfo{
2735 2 : versionEditApplied: c.versionEditApplied,
2736 2 : inputs: c.inputs,
2737 2 : smallest: c.smallest,
2738 2 : largest: c.largest,
2739 2 : outputLevel: -1,
2740 2 : }
2741 2 : if c.outputLevel != nil {
2742 2 : info.outputLevel = c.outputLevel.level
2743 2 : }
2744 2 : rv = append(rv, info)
2745 : }
2746 : }
2747 2 : return
2748 : }
2749 :
2750 2 : func inProgressL0Compactions(inProgress []compactionInfo) []manifest.L0Compaction {
2751 2 : var compactions []manifest.L0Compaction
2752 2 : for _, info := range inProgress {
2753 2 : // Skip in-progress compactions that have already committed; the L0
2754 2 : // sublevels initialization code requires the set of in-progress
2755 2 : // compactions to be consistent with the current version. Compactions
2756 2 : // with versionEditApplied=true are already applied to the current
2757 2 : // version and but are performing cleanup without the database mutex.
2758 2 : if info.versionEditApplied {
2759 2 : continue
2760 : }
2761 2 : l0 := false
2762 2 : for _, cl := range info.inputs {
2763 2 : l0 = l0 || cl.level == 0
2764 2 : }
2765 2 : if !l0 {
2766 2 : continue
2767 : }
2768 2 : compactions = append(compactions, manifest.L0Compaction{
2769 2 : Smallest: info.smallest,
2770 2 : Largest: info.largest,
2771 2 : IsIntraL0: info.outputLevel == 0,
2772 2 : })
2773 : }
2774 2 : return compactions
2775 : }
2776 :
2777 : // firstError returns the first non-nil error of err0 and err1, or nil if both
2778 : // are nil.
2779 2 : func firstError(err0, err1 error) error {
2780 2 : if err0 != nil {
2781 2 : return err0
2782 2 : }
2783 2 : return err1
2784 : }
2785 :
2786 : // SetCreatorID sets the CreatorID which is needed in order to use shared objects.
2787 : // Remote object usage is disabled until this method is called the first time.
2788 : // Once set, the Creator ID is persisted and cannot change.
2789 : //
2790 : // Does nothing if SharedStorage was not set in the options when the DB was
2791 : // opened or if the DB is in read-only mode.
2792 2 : func (d *DB) SetCreatorID(creatorID uint64) error {
2793 2 : if d.opts.Experimental.RemoteStorage == nil || d.opts.ReadOnly {
2794 0 : return nil
2795 0 : }
2796 2 : return d.objProvider.SetCreatorID(objstorage.CreatorID(creatorID))
2797 : }
2798 :
2799 : // KeyStatistics keeps track of the number of keys that have been pinned by a
2800 : // snapshot as well as counts of the different key kinds in the lsm.
2801 : //
2802 : // One way of using the accumulated stats, when we only have sets and dels,
2803 : // and say the counts are represented as del_count, set_count,
2804 : // del_latest_count, set_latest_count, snapshot_pinned_count.
2805 : //
2806 : // - del_latest_count + set_latest_count is the set of unique user keys
2807 : // (unique).
2808 : //
2809 : // - set_latest_count is the set of live unique user keys (live_unique).
2810 : //
2811 : // - Garbage is del_count + set_count - live_unique.
2812 : //
2813 : // - If everything were in the LSM, del_count+set_count-snapshot_pinned_count
2814 : // would also be the set of unique user keys (note that
2815 : // snapshot_pinned_count is counting something different -- see comment below).
2816 : // But snapshot_pinned_count only counts keys in the LSM so the excess here
2817 : // must be keys in memtables.
2818 : type KeyStatistics struct {
2819 : // TODO(sumeer): the SnapshotPinned* are incorrect in that these older
2820 : // versions can be in a different level. Either fix the accounting or
2821 : // rename these fields.
2822 :
2823 : // SnapshotPinnedKeys represents obsolete keys that cannot be elided during
2824 : // a compaction, because they are required by an open snapshot.
2825 : SnapshotPinnedKeys int
2826 : // SnapshotPinnedKeysBytes is the total number of bytes of all snapshot
2827 : // pinned keys.
2828 : SnapshotPinnedKeysBytes uint64
2829 : // KindsCount is the count for each kind of key. It includes point keys,
2830 : // range deletes and range keys.
2831 : KindsCount [InternalKeyKindMax + 1]int
2832 : // LatestKindsCount is the count for each kind of key when it is the latest
2833 : // kind for a user key. It is only populated for point keys.
2834 : LatestKindsCount [InternalKeyKindMax + 1]int
2835 : }
2836 :
2837 : // LSMKeyStatistics is used by DB.ScanStatistics.
2838 : type LSMKeyStatistics struct {
2839 : Accumulated KeyStatistics
2840 : // Levels contains statistics only for point keys. Range deletions and range keys will
2841 : // appear in Accumulated but not Levels.
2842 : Levels [numLevels]KeyStatistics
2843 : // BytesRead represents the logical, pre-compression size of keys and values read
2844 : BytesRead uint64
2845 : }
2846 :
2847 : // ScanStatisticsOptions is used by DB.ScanStatistics.
2848 : type ScanStatisticsOptions struct {
2849 : // LimitBytesPerSecond indicates the number of bytes that are able to be read
2850 : // per second using ScanInternal.
2851 : // A value of 0 indicates that there is no limit set.
2852 : LimitBytesPerSecond int64
2853 : }
2854 :
2855 : // ScanStatistics returns the count of different key kinds within the lsm for a
2856 : // key span [lower, upper) as well as the number of snapshot keys.
2857 : func (d *DB) ScanStatistics(
2858 : ctx context.Context, lower, upper []byte, opts ScanStatisticsOptions,
2859 1 : ) (LSMKeyStatistics, error) {
2860 1 : stats := LSMKeyStatistics{}
2861 1 : var prevKey InternalKey
2862 1 : var rateLimitFunc func(key *InternalKey, val LazyValue) error
2863 1 : tb := tokenbucket.TokenBucket{}
2864 1 :
2865 1 : if opts.LimitBytesPerSecond != 0 {
2866 0 : // Each "token" roughly corresponds to a byte that was read.
2867 0 : tb.Init(tokenbucket.TokensPerSecond(opts.LimitBytesPerSecond), tokenbucket.Tokens(1024))
2868 0 : rateLimitFunc = func(key *InternalKey, val LazyValue) error {
2869 0 : return tb.WaitCtx(ctx, tokenbucket.Tokens(key.Size()+val.Len()))
2870 0 : }
2871 : }
2872 :
2873 1 : scanInternalOpts := &scanInternalOptions{
2874 1 : visitPointKey: func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error {
2875 1 : // If the previous key is equal to the current point key, the current key was
2876 1 : // pinned by a snapshot.
2877 1 : size := uint64(key.Size())
2878 1 : kind := key.Kind()
2879 1 : sameKey := d.equal(prevKey.UserKey, key.UserKey)
2880 1 : if iterInfo.Kind == IteratorLevelLSM && sameKey {
2881 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeys++
2882 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeysBytes += size
2883 1 : stats.Accumulated.SnapshotPinnedKeys++
2884 1 : stats.Accumulated.SnapshotPinnedKeysBytes += size
2885 1 : }
2886 1 : if iterInfo.Kind == IteratorLevelLSM {
2887 1 : stats.Levels[iterInfo.Level].KindsCount[kind]++
2888 1 : }
2889 1 : if !sameKey {
2890 1 : if iterInfo.Kind == IteratorLevelLSM {
2891 1 : stats.Levels[iterInfo.Level].LatestKindsCount[kind]++
2892 1 : }
2893 1 : stats.Accumulated.LatestKindsCount[kind]++
2894 : }
2895 :
2896 1 : stats.Accumulated.KindsCount[kind]++
2897 1 : prevKey.CopyFrom(*key)
2898 1 : stats.BytesRead += uint64(key.Size() + value.Len())
2899 1 : return nil
2900 : },
2901 0 : visitRangeDel: func(start, end []byte, seqNum uint64) error {
2902 0 : stats.Accumulated.KindsCount[InternalKeyKindRangeDelete]++
2903 0 : stats.BytesRead += uint64(len(start) + len(end))
2904 0 : return nil
2905 0 : },
2906 0 : visitRangeKey: func(start, end []byte, keys []rangekey.Key) error {
2907 0 : stats.BytesRead += uint64(len(start) + len(end))
2908 0 : for _, key := range keys {
2909 0 : stats.Accumulated.KindsCount[key.Kind()]++
2910 0 : stats.BytesRead += uint64(len(key.Value) + len(key.Suffix))
2911 0 : }
2912 0 : return nil
2913 : },
2914 : includeObsoleteKeys: true,
2915 : IterOptions: IterOptions{
2916 : KeyTypes: IterKeyTypePointsAndRanges,
2917 : LowerBound: lower,
2918 : UpperBound: upper,
2919 : },
2920 : rateLimitFunc: rateLimitFunc,
2921 : }
2922 1 : iter := d.newInternalIter(ctx, snapshotIterOpts{}, scanInternalOpts)
2923 1 : defer iter.close()
2924 1 :
2925 1 : err := scanInternalImpl(ctx, lower, upper, iter, scanInternalOpts)
2926 1 :
2927 1 : if err != nil {
2928 0 : return LSMKeyStatistics{}, err
2929 0 : }
2930 :
2931 1 : return stats, nil
2932 : }
2933 :
2934 : // ObjProvider returns the objstorage.Provider for this database. Meant to be
2935 : // used for internal purposes only.
2936 2 : func (d *DB) ObjProvider() objstorage.Provider {
2937 2 : return d.objProvider
2938 2 : }
2939 :
2940 1 : func (d *DB) checkVirtualBounds(m *fileMetadata) {
2941 1 : if !invariants.Enabled {
2942 0 : return
2943 0 : }
2944 :
2945 1 : objMeta, err := d.objProvider.Lookup(fileTypeTable, m.FileBacking.DiskFileNum)
2946 1 : if err != nil {
2947 0 : panic(err)
2948 : }
2949 1 : if objMeta.IsExternal() {
2950 0 : // Nothing to do; bounds are expected to be loose.
2951 0 : return
2952 0 : }
2953 :
2954 1 : if m.HasPointKeys {
2955 1 : pointIter, rangeDelIter, err := d.newIters(context.TODO(), m, nil, internalIterOpts{})
2956 1 : if err != nil {
2957 0 : panic(errors.Wrap(err, "pebble: error creating point iterator"))
2958 : }
2959 :
2960 1 : defer pointIter.Close()
2961 1 : if rangeDelIter != nil {
2962 1 : defer rangeDelIter.Close()
2963 1 : }
2964 :
2965 1 : pointKey, _ := pointIter.First()
2966 1 : var rangeDel *keyspan.Span
2967 1 : if rangeDelIter != nil {
2968 1 : rangeDel = rangeDelIter.First()
2969 1 : }
2970 :
2971 : // Check that the lower bound is tight.
2972 1 : if (rangeDel == nil || d.cmp(rangeDel.SmallestKey().UserKey, m.SmallestPointKey.UserKey) != 0) &&
2973 1 : (pointKey == nil || d.cmp(pointKey.UserKey, m.SmallestPointKey.UserKey) != 0) {
2974 0 : panic(errors.Newf("pebble: virtual sstable %s lower point key bound is not tight", m.FileNum))
2975 : }
2976 :
2977 1 : pointKey, _ = pointIter.Last()
2978 1 : rangeDel = nil
2979 1 : if rangeDelIter != nil {
2980 1 : rangeDel = rangeDelIter.Last()
2981 1 : }
2982 :
2983 : // Check that the upper bound is tight.
2984 1 : if (rangeDel == nil || d.cmp(rangeDel.LargestKey().UserKey, m.LargestPointKey.UserKey) != 0) &&
2985 1 : (pointKey == nil || d.cmp(pointKey.UserKey, m.LargestPointKey.UserKey) != 0) {
2986 0 : panic(errors.Newf("pebble: virtual sstable %s upper point key bound is not tight", m.FileNum))
2987 : }
2988 :
2989 : // Check that iterator keys are within bounds.
2990 1 : for key, _ := pointIter.First(); key != nil; key, _ = pointIter.Next() {
2991 1 : if d.cmp(key.UserKey, m.SmallestPointKey.UserKey) < 0 || d.cmp(key.UserKey, m.LargestPointKey.UserKey) > 0 {
2992 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.UserKey))
2993 : }
2994 : }
2995 :
2996 1 : if rangeDelIter != nil {
2997 1 : for key := rangeDelIter.First(); key != nil; key = rangeDelIter.Next() {
2998 1 : if d.cmp(key.SmallestKey().UserKey, m.SmallestPointKey.UserKey) < 0 {
2999 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.SmallestKey().UserKey))
3000 : }
3001 :
3002 1 : if d.cmp(key.LargestKey().UserKey, m.LargestPointKey.UserKey) > 0 {
3003 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.LargestKey().UserKey))
3004 : }
3005 : }
3006 : }
3007 : }
3008 :
3009 1 : if !m.HasRangeKeys {
3010 1 : return
3011 1 : }
3012 :
3013 0 : rangeKeyIter, err := d.tableNewRangeKeyIter(m, keyspan.SpanIterOptions{})
3014 0 : defer rangeKeyIter.Close()
3015 0 :
3016 0 : if err != nil {
3017 0 : panic(errors.Wrap(err, "pebble: error creating range key iterator"))
3018 : }
3019 :
3020 : // Check that the lower bound is tight.
3021 0 : if d.cmp(rangeKeyIter.First().SmallestKey().UserKey, m.SmallestRangeKey.UserKey) != 0 {
3022 0 : panic(errors.Newf("pebble: virtual sstable %s lower range key bound is not tight", m.FileNum))
3023 : }
3024 :
3025 : // Check that upper bound is tight.
3026 0 : if d.cmp(rangeKeyIter.Last().LargestKey().UserKey, m.LargestRangeKey.UserKey) != 0 {
3027 0 : panic(errors.Newf("pebble: virtual sstable %s upper range key bound is not tight", m.FileNum))
3028 : }
3029 :
3030 0 : for key := rangeKeyIter.First(); key != nil; key = rangeKeyIter.Next() {
3031 0 : if d.cmp(key.SmallestKey().UserKey, m.SmallestRangeKey.UserKey) < 0 {
3032 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.SmallestKey().UserKey))
3033 : }
3034 0 : if d.cmp(key.LargestKey().UserKey, m.LargestRangeKey.UserKey) > 0 {
3035 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.LargestKey().UserKey))
3036 : }
3037 : }
3038 : }
|
