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 : "strconv"
13 : "sync"
14 : "sync/atomic"
15 : "time"
16 :
17 : "github.com/cockroachdb/errors"
18 : "github.com/cockroachdb/pebble/internal/arenaskl"
19 : "github.com/cockroachdb/pebble/internal/base"
20 : "github.com/cockroachdb/pebble/internal/invalidating"
21 : "github.com/cockroachdb/pebble/internal/invariants"
22 : "github.com/cockroachdb/pebble/internal/keyspan"
23 : "github.com/cockroachdb/pebble/internal/keyspan/keyspanimpl"
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/pebble/wal"
34 : "github.com/cockroachdb/tokenbucket"
35 : "github.com/prometheus/client_golang/prometheus"
36 : )
37 :
38 : const (
39 : // minTableCacheSize is the minimum size of the table cache, for a single db.
40 : minTableCacheSize = 64
41 :
42 : // numNonTableCacheFiles is an approximation for the number of files
43 : // that we don't use for table caches, for a given db.
44 : numNonTableCacheFiles = 10
45 : )
46 :
47 : var (
48 : // ErrNotFound is returned when a get operation does not find the requested
49 : // key.
50 : ErrNotFound = base.ErrNotFound
51 : // ErrClosed is panicked when an operation is performed on a closed snapshot or
52 : // DB. Use errors.Is(err, ErrClosed) to check for this error.
53 : ErrClosed = errors.New("pebble: closed")
54 : // ErrReadOnly is returned when a write operation is performed on a read-only
55 : // database.
56 : ErrReadOnly = errors.New("pebble: read-only")
57 : // errNoSplit indicates that the user is trying to perform a range key
58 : // operation but the configured Comparer does not provide a Split
59 : // implementation.
60 : errNoSplit = errors.New("pebble: Comparer.Split required for range key operations")
61 : )
62 :
63 : // Reader is a readable key/value store.
64 : //
65 : // It is safe to call Get and NewIter from concurrent goroutines.
66 : type Reader interface {
67 : // Get gets the value for the given key. It returns ErrNotFound if the DB
68 : // does not contain the key.
69 : //
70 : // The caller should not modify the contents of the returned slice, but it is
71 : // safe to modify the contents of the argument after Get returns. The
72 : // returned slice will remain valid until the returned Closer is closed. On
73 : // success, the caller MUST call closer.Close() or a memory leak will occur.
74 : Get(key []byte) (value []byte, closer io.Closer, err error)
75 :
76 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
77 : // return false). The iterator can be positioned via a call to SeekGE,
78 : // SeekLT, First or Last.
79 : NewIter(o *IterOptions) (*Iterator, error)
80 :
81 : // NewIterWithContext is like NewIter, and additionally accepts a context
82 : // for tracing.
83 : NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error)
84 :
85 : // Close closes the Reader. It may or may not close any underlying io.Reader
86 : // or io.Writer, depending on how the DB was created.
87 : //
88 : // It is not safe to close a DB until all outstanding iterators are closed.
89 : // It is valid to call Close multiple times. Other methods should not be
90 : // called after the DB has been closed.
91 : Close() error
92 : }
93 :
94 : // Writer is a writable key/value store.
95 : //
96 : // Goroutine safety is dependent on the specific implementation.
97 : type Writer interface {
98 : // Apply the operations contained in the batch to the DB.
99 : //
100 : // It is safe to modify the contents of the arguments after Apply returns.
101 : Apply(batch *Batch, o *WriteOptions) error
102 :
103 : // Delete deletes the value for the given key. Deletes are blind all will
104 : // succeed even if the given key does not exist.
105 : //
106 : // It is safe to modify the contents of the arguments after Delete returns.
107 : Delete(key []byte, o *WriteOptions) error
108 :
109 : // DeleteSized behaves identically to Delete, but takes an additional
110 : // argument indicating the size of the value being deleted. DeleteSized
111 : // should be preferred when the caller has the expectation that there exists
112 : // a single internal KV pair for the key (eg, the key has not been
113 : // overwritten recently), and the caller knows the size of its value.
114 : //
115 : // DeleteSized will record the value size within the tombstone and use it to
116 : // inform compaction-picking heuristics which strive to reduce space
117 : // amplification in the LSM. This "calling your shot" mechanic allows the
118 : // storage engine to more accurately estimate and reduce space
119 : // amplification.
120 : //
121 : // It is safe to modify the contents of the arguments after DeleteSized
122 : // returns.
123 : DeleteSized(key []byte, valueSize uint32, _ *WriteOptions) error
124 :
125 : // SingleDelete is similar to Delete in that it deletes the value for the given key. Like Delete,
126 : // it is a blind operation that will succeed even if the given key does not exist.
127 : //
128 : // WARNING: Undefined (non-deterministic) behavior will result if a key is overwritten and
129 : // then deleted using SingleDelete. The record may appear deleted immediately, but be
130 : // resurrected at a later time after compactions have been performed. Or the record may
131 : // be deleted permanently. A Delete operation lays down a "tombstone" which shadows all
132 : // previous versions of a key. The SingleDelete operation is akin to "anti-matter" and will
133 : // only delete the most recently written version for a key. These different semantics allow
134 : // the DB to avoid propagating a SingleDelete operation during a compaction as soon as the
135 : // corresponding Set operation is encountered. These semantics require extreme care to handle
136 : // properly. Only use if you have a workload where the performance gain is critical and you
137 : // can guarantee that a record is written once and then deleted once.
138 : //
139 : // SingleDelete is internally transformed into a Delete if the most recent record for a key is either
140 : // a Merge or Delete record.
141 : //
142 : // It is safe to modify the contents of the arguments after SingleDelete returns.
143 : SingleDelete(key []byte, o *WriteOptions) error
144 :
145 : // DeleteRange deletes all of the point keys (and values) in the range
146 : // [start,end) (inclusive on start, exclusive on end). DeleteRange does NOT
147 : // delete overlapping range keys (eg, keys set via RangeKeySet).
148 : //
149 : // It is safe to modify the contents of the arguments after DeleteRange
150 : // returns.
151 : DeleteRange(start, end []byte, o *WriteOptions) error
152 :
153 : // LogData adds the specified to the batch. The data will be written to the
154 : // WAL, but not added to memtables or sstables. Log data is never indexed,
155 : // which makes it useful for testing WAL performance.
156 : //
157 : // It is safe to modify the contents of the argument after LogData returns.
158 : LogData(data []byte, opts *WriteOptions) error
159 :
160 : // Merge merges the value for the given key. The details of the merge are
161 : // dependent upon the configured merge operation.
162 : //
163 : // It is safe to modify the contents of the arguments after Merge returns.
164 : Merge(key, value []byte, o *WriteOptions) error
165 :
166 : // Set sets the value for the given key. It overwrites any previous value
167 : // for that key; a DB is not a multi-map.
168 : //
169 : // It is safe to modify the contents of the arguments after Set returns.
170 : Set(key, value []byte, o *WriteOptions) error
171 :
172 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
173 : // timestamp suffix to value. The suffix is optional. If any portion of the key
174 : // range [start, end) is already set by a range key with the same suffix value,
175 : // RangeKeySet overrides it.
176 : //
177 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
178 : RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error
179 :
180 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
181 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
182 : // range key. RangeKeyUnset only removes portions of range keys that fall within
183 : // the [start, end) key span, and only range keys with suffixes that exactly
184 : // match the unset suffix.
185 : //
186 : // It is safe to modify the contents of the arguments after RangeKeyUnset
187 : // returns.
188 : RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error
189 :
190 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
191 : // (inclusive on start, exclusive on end). It does not delete point keys (for
192 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
193 : // bounds, including those with or without suffixes.
194 : //
195 : // It is safe to modify the contents of the arguments after RangeKeyDelete
196 : // returns.
197 : RangeKeyDelete(start, end []byte, opts *WriteOptions) error
198 : }
199 :
200 : // CPUWorkHandle represents a handle used by the CPUWorkPermissionGranter API.
201 : type CPUWorkHandle interface {
202 : // Permitted indicates whether Pebble can use additional CPU resources.
203 : Permitted() bool
204 : }
205 :
206 : // CPUWorkPermissionGranter is used to request permission to opportunistically
207 : // use additional CPUs to speed up internal background work.
208 : type CPUWorkPermissionGranter interface {
209 : // GetPermission returns a handle regardless of whether permission is granted
210 : // or not. In the latter case, the handle is only useful for recording
211 : // the CPU time actually spent on this calling goroutine.
212 : GetPermission(time.Duration) CPUWorkHandle
213 : // CPUWorkDone must be called regardless of whether CPUWorkHandle.Permitted
214 : // returns true or false.
215 : CPUWorkDone(CPUWorkHandle)
216 : }
217 :
218 : // Use a default implementation for the CPU work granter to avoid excessive nil
219 : // checks in the code.
220 : type defaultCPUWorkHandle struct{}
221 :
222 2 : func (d defaultCPUWorkHandle) Permitted() bool {
223 2 : return false
224 2 : }
225 :
226 : type defaultCPUWorkGranter struct{}
227 :
228 2 : func (d defaultCPUWorkGranter) GetPermission(_ time.Duration) CPUWorkHandle {
229 2 : return defaultCPUWorkHandle{}
230 2 : }
231 :
232 2 : func (d defaultCPUWorkGranter) CPUWorkDone(_ CPUWorkHandle) {}
233 :
234 : // DB provides a concurrent, persistent ordered key/value store.
235 : //
236 : // A DB's basic operations (Get, Set, Delete) should be self-explanatory. Get
237 : // and Delete will return ErrNotFound if the requested key is not in the store.
238 : // Callers are free to ignore this error.
239 : //
240 : // A DB also allows for iterating over the key/value pairs in key order. If d
241 : // is a DB, the code below prints all key/value pairs whose keys are 'greater
242 : // than or equal to' k:
243 : //
244 : // iter := d.NewIter(readOptions)
245 : // for iter.SeekGE(k); iter.Valid(); iter.Next() {
246 : // fmt.Printf("key=%q value=%q\n", iter.Key(), iter.Value())
247 : // }
248 : // return iter.Close()
249 : //
250 : // The Options struct holds the optional parameters for the DB, including a
251 : // Comparer to define a 'less than' relationship over keys. It is always valid
252 : // to pass a nil *Options, which means to use the default parameter values. Any
253 : // zero field of a non-nil *Options also means to use the default value for
254 : // that parameter. Thus, the code below uses a custom Comparer, but the default
255 : // values for every other parameter:
256 : //
257 : // db := pebble.Open(&Options{
258 : // Comparer: myComparer,
259 : // })
260 : type DB struct {
261 : // The count and size of referenced memtables. This includes memtables
262 : // present in DB.mu.mem.queue, as well as memtables that have been flushed
263 : // but are still referenced by an inuse readState, as well as up to one
264 : // memTable waiting to be reused and stored in d.memTableRecycle.
265 : memTableCount atomic.Int64
266 : memTableReserved atomic.Int64 // number of bytes reserved in the cache for memtables
267 : // memTableRecycle holds a pointer to an obsolete memtable. The next
268 : // memtable allocation will reuse this memtable if it has not already been
269 : // recycled.
270 : memTableRecycle atomic.Pointer[memTable]
271 :
272 : // The logical size of the current WAL.
273 : logSize atomic.Uint64
274 :
275 : // The number of bytes available on disk.
276 : diskAvailBytes atomic.Uint64
277 :
278 : cacheID uint64
279 : dirname 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 :
300 : tableCache *tableCacheContainer
301 : newIters tableNewIters
302 : tableNewRangeKeyIter keyspanimpl.TableNewSpanIter
303 :
304 : commit *commitPipeline
305 :
306 : // readState provides access to the state needed for reading without needing
307 : // to acquire DB.mu.
308 : readState struct {
309 : sync.RWMutex
310 : val *readState
311 : }
312 :
313 : closed *atomic.Value
314 : closedCh chan struct{}
315 :
316 : cleanupManager *cleanupManager
317 :
318 : // During an iterator close, we may asynchronously schedule read compactions.
319 : // We want to wait for those goroutines to finish, before closing the DB.
320 : // compactionShedulers.Wait() should not be called while the DB.mu is held.
321 : compactionSchedulers sync.WaitGroup
322 :
323 : // The main mutex protecting internal DB state. This mutex encompasses many
324 : // fields because those fields need to be accessed and updated atomically. In
325 : // particular, the current version, log.*, mem.*, and snapshot list need to
326 : // be accessed and updated atomically during compaction.
327 : //
328 : // Care is taken to avoid holding DB.mu during IO operations. Accomplishing
329 : // this sometimes requires releasing DB.mu in a method that was called with
330 : // it held. See versionSet.logAndApply() and DB.makeRoomForWrite() for
331 : // examples. This is a common pattern, so be careful about expectations that
332 : // DB.mu will be held continuously across a set of calls.
333 : mu struct {
334 : sync.Mutex
335 :
336 : formatVers struct {
337 : // vers is the database's current format major version.
338 : // Backwards-incompatible features are gated behind new
339 : // format major versions and not enabled until a database's
340 : // version is ratcheted upwards.
341 : //
342 : // Although this is under the `mu` prefix, readers may read vers
343 : // atomically without holding d.mu. Writers must only write to this
344 : // value through finalizeFormatVersUpgrade which requires d.mu is
345 : // held.
346 : vers atomic.Uint64
347 : // marker is the atomic marker for the format major version.
348 : // When a database's version is ratcheted upwards, the
349 : // marker is moved in order to atomically record the new
350 : // version.
351 : marker *atomicfs.Marker
352 : // ratcheting when set to true indicates that the database is
353 : // currently in the process of ratcheting the format major version
354 : // to vers + 1. As a part of ratcheting the format major version,
355 : // migrations may drop and re-acquire the mutex.
356 : ratcheting bool
357 : }
358 :
359 : // The ID of the next job. Job IDs are passed to event listener
360 : // notifications and act as a mechanism for tying together the events and
361 : // log messages for a single job such as a flush, compaction, or file
362 : // ingestion. Job IDs are not serialized to disk or used for correctness.
363 : nextJobID JobID
364 :
365 : // The collection of immutable versions and state about the log and visible
366 : // sequence numbers. Use the pointer here to ensure the atomic fields in
367 : // version set are aligned properly.
368 : versions *versionSet
369 :
370 : log struct {
371 : // manager is not protected by mu, but calls to Create must be
372 : // serialized, and happen after the previous writer is closed.
373 : manager wal.Manager
374 : // The number of input bytes to the log. This is the raw size of the
375 : // batches written to the WAL, without the overhead of the record
376 : // envelopes.
377 : bytesIn uint64
378 : // The Writer is protected by commitPipeline.mu. This allows log writes
379 : // to be performed without holding DB.mu, but requires both
380 : // commitPipeline.mu and DB.mu to be held when rotating the WAL/memtable
381 : // (i.e. makeRoomForWrite). Can be nil.
382 : writer wal.Writer
383 : metrics struct {
384 : // fsyncLatency has its own internal synchronization, and is not
385 : // protected by mu.
386 : fsyncLatency prometheus.Histogram
387 : // Updated whenever a wal.Writer is closed.
388 : record.LogWriterMetrics
389 : }
390 : }
391 :
392 : mem struct {
393 : // The current mutable memTable.
394 : mutable *memTable
395 : // Queue of flushables (the mutable memtable is at end). Elements are
396 : // added to the end of the slice and removed from the beginning. Once an
397 : // index is set it is never modified making a fixed slice immutable and
398 : // safe for concurrent reads.
399 : queue flushableList
400 : // nextSize is the size of the next memtable. The memtable size starts at
401 : // min(256KB,Options.MemTableSize) and doubles each time a new memtable
402 : // is allocated up to Options.MemTableSize. This reduces the memory
403 : // footprint of memtables when lots of DB instances are used concurrently
404 : // in test environments.
405 : nextSize uint64
406 : }
407 :
408 : compact struct {
409 : // Condition variable used to signal when a flush or compaction has
410 : // completed. Used by the write-stall mechanism to wait for the stall
411 : // condition to clear. See DB.makeRoomForWrite().
412 : cond sync.Cond
413 : // True when a flush is in progress.
414 : flushing bool
415 : // The number of ongoing non-download compactions.
416 : compactingCount int
417 : // The number of download compactions.
418 : downloadingCount int
419 : // The list of deletion hints, suggesting ranges for delete-only
420 : // compactions.
421 : deletionHints []deleteCompactionHint
422 : // The list of manual compactions. The next manual compaction to perform
423 : // is at the start of the list. New entries are added to the end.
424 : manual []*manualCompaction
425 : // downloads is the list of pending download tasks. The next download to
426 : // perform is at the start of the list. New entries are added to the end.
427 : downloads []*downloadSpanTask
428 : // inProgress is the set of in-progress flushes and compactions.
429 : // It's used in the calculation of some metrics and to initialize L0
430 : // sublevels' state. Some of the compactions contained within this
431 : // map may have already committed an edit to the version but are
432 : // lingering performing cleanup, like deleting obsolete files.
433 : inProgress map[*compaction]struct{}
434 :
435 : // rescheduleReadCompaction indicates to an iterator that a read compaction
436 : // should be scheduled.
437 : rescheduleReadCompaction bool
438 :
439 : // readCompactions is a readCompactionQueue which keeps track of the
440 : // compactions which we might have to perform.
441 : readCompactions readCompactionQueue
442 :
443 : // The cumulative duration of all completed compactions since Open.
444 : // Does not include flushes.
445 : duration time.Duration
446 : // Flush throughput metric.
447 : flushWriteThroughput ThroughputMetric
448 : // The idle start time for the flush "loop", i.e., when the flushing
449 : // bool above transitions to false.
450 : noOngoingFlushStartTime time.Time
451 : }
452 :
453 : // Non-zero when file cleaning is disabled. The disabled count acts as a
454 : // reference count to prohibit file cleaning. See
455 : // DB.{disable,Enable}FileDeletions().
456 : disableFileDeletions int
457 :
458 : snapshots struct {
459 : // The list of active snapshots.
460 : snapshotList
461 :
462 : // The cumulative count and size of snapshot-pinned keys written to
463 : // sstables.
464 : cumulativePinnedCount uint64
465 : cumulativePinnedSize uint64
466 : }
467 :
468 : tableStats struct {
469 : // Condition variable used to signal the completion of a
470 : // job to collect table stats.
471 : cond sync.Cond
472 : // True when a stat collection operation is in progress.
473 : loading bool
474 : // True if stat collection has loaded statistics for all tables
475 : // other than those listed explicitly in pending. This flag starts
476 : // as false when a database is opened and flips to true once stat
477 : // collection has caught up.
478 : loadedInitial bool
479 : // A slice of files for which stats have not been computed.
480 : // Compactions, ingests, flushes append files to be processed. An
481 : // active stat collection goroutine clears the list and processes
482 : // them.
483 : pending []manifest.NewFileEntry
484 : }
485 :
486 : tableValidation struct {
487 : // cond is a condition variable used to signal the completion of a
488 : // job to validate one or more sstables.
489 : cond sync.Cond
490 : // pending is a slice of metadata for sstables waiting to be
491 : // validated. Only physical sstables should be added to the pending
492 : // queue.
493 : pending []newFileEntry
494 : // validating is set to true when validation is running.
495 : validating bool
496 : }
497 : }
498 :
499 : // Normally equal to time.Now() but may be overridden in tests.
500 : timeNow func() time.Time
501 : // the time at database Open; may be used to compute metrics like effective
502 : // compaction concurrency
503 : openedAt time.Time
504 : }
505 :
506 : var _ Reader = (*DB)(nil)
507 : var _ Writer = (*DB)(nil)
508 :
509 : // TestOnlyWaitForCleaning MUST only be used in tests.
510 1 : func (d *DB) TestOnlyWaitForCleaning() {
511 1 : d.cleanupManager.Wait()
512 1 : }
513 :
514 : // Get gets the value for the given key. It returns ErrNotFound if the DB does
515 : // not contain the key.
516 : //
517 : // The caller should not modify the contents of the returned slice, but it is
518 : // safe to modify the contents of the argument after Get returns. The returned
519 : // slice will remain valid until the returned Closer is closed. On success, the
520 : // caller MUST call closer.Close() or a memory leak will occur.
521 2 : func (d *DB) Get(key []byte) ([]byte, io.Closer, error) {
522 2 : return d.getInternal(key, nil /* batch */, nil /* snapshot */)
523 2 : }
524 :
525 : type getIterAlloc struct {
526 : dbi Iterator
527 : keyBuf []byte
528 : get getIter
529 : }
530 :
531 : var getIterAllocPool = sync.Pool{
532 2 : New: func() interface{} {
533 2 : return &getIterAlloc{}
534 2 : },
535 : }
536 :
537 2 : func (d *DB) getInternal(key []byte, b *Batch, s *Snapshot) ([]byte, io.Closer, error) {
538 2 : if err := d.closed.Load(); err != nil {
539 1 : panic(err)
540 : }
541 :
542 : // Grab and reference the current readState. This prevents the underlying
543 : // files in the associated version from being deleted if there is a current
544 : // compaction. The readState is unref'd by Iterator.Close().
545 2 : readState := d.loadReadState()
546 2 :
547 2 : // Determine the seqnum to read at after grabbing the read state (current and
548 2 : // memtables) above.
549 2 : var seqNum uint64
550 2 : if s != nil {
551 2 : seqNum = s.seqNum
552 2 : } else {
553 2 : seqNum = d.mu.versions.visibleSeqNum.Load()
554 2 : }
555 :
556 2 : buf := getIterAllocPool.Get().(*getIterAlloc)
557 2 :
558 2 : get := &buf.get
559 2 : *get = getIter{
560 2 : logger: d.opts.Logger,
561 2 : comparer: d.opts.Comparer,
562 2 : newIters: d.newIters,
563 2 : snapshot: seqNum,
564 2 : key: key,
565 2 : batch: b,
566 2 : mem: readState.memtables,
567 2 : l0: readState.current.L0SublevelFiles,
568 2 : version: readState.current,
569 2 : }
570 2 :
571 2 : // Strip off memtables which cannot possibly contain the seqNum being read
572 2 : // at.
573 2 : for len(get.mem) > 0 {
574 2 : n := len(get.mem)
575 2 : if logSeqNum := get.mem[n-1].logSeqNum; logSeqNum < seqNum {
576 2 : break
577 : }
578 2 : get.mem = get.mem[:n-1]
579 : }
580 :
581 2 : i := &buf.dbi
582 2 : pointIter := get
583 2 : *i = Iterator{
584 2 : ctx: context.Background(),
585 2 : getIterAlloc: buf,
586 2 : iter: pointIter,
587 2 : pointIter: pointIter,
588 2 : merge: d.merge,
589 2 : comparer: *d.opts.Comparer,
590 2 : readState: readState,
591 2 : keyBuf: buf.keyBuf,
592 2 : }
593 2 :
594 2 : if !i.First() {
595 2 : err := i.Close()
596 2 : if err != nil {
597 1 : return nil, nil, err
598 1 : }
599 2 : return nil, nil, ErrNotFound
600 : }
601 2 : return i.Value(), i, nil
602 : }
603 :
604 : // Set sets the value for the given key. It overwrites any previous value
605 : // for that key; a DB is not a multi-map.
606 : //
607 : // It is safe to modify the contents of the arguments after Set returns.
608 2 : func (d *DB) Set(key, value []byte, opts *WriteOptions) error {
609 2 : b := newBatch(d)
610 2 : _ = b.Set(key, value, opts)
611 2 : if err := d.Apply(b, opts); err != nil {
612 1 : return err
613 1 : }
614 : // Only release the batch on success.
615 2 : return b.Close()
616 : }
617 :
618 : // Delete deletes the value for the given key. Deletes are blind all will
619 : // succeed even if the given key does not exist.
620 : //
621 : // It is safe to modify the contents of the arguments after Delete returns.
622 2 : func (d *DB) Delete(key []byte, opts *WriteOptions) error {
623 2 : b := newBatch(d)
624 2 : _ = b.Delete(key, opts)
625 2 : if err := d.Apply(b, opts); err != nil {
626 1 : return err
627 1 : }
628 : // Only release the batch on success.
629 2 : return b.Close()
630 : }
631 :
632 : // DeleteSized behaves identically to Delete, but takes an additional
633 : // argument indicating the size of the value being deleted. DeleteSized
634 : // should be preferred when the caller has the expectation that there exists
635 : // a single internal KV pair for the key (eg, the key has not been
636 : // overwritten recently), and the caller knows the size of its value.
637 : //
638 : // DeleteSized will record the value size within the tombstone and use it to
639 : // inform compaction-picking heuristics which strive to reduce space
640 : // amplification in the LSM. This "calling your shot" mechanic allows the
641 : // storage engine to more accurately estimate and reduce space amplification.
642 : //
643 : // It is safe to modify the contents of the arguments after DeleteSized
644 : // returns.
645 2 : func (d *DB) DeleteSized(key []byte, valueSize uint32, opts *WriteOptions) error {
646 2 : b := newBatch(d)
647 2 : _ = b.DeleteSized(key, valueSize, opts)
648 2 : if err := d.Apply(b, opts); err != nil {
649 0 : return err
650 0 : }
651 : // Only release the batch on success.
652 2 : return b.Close()
653 : }
654 :
655 : // SingleDelete adds an action to the batch that single deletes the entry for key.
656 : // See Writer.SingleDelete for more details on the semantics of SingleDelete.
657 : //
658 : // It is safe to modify the contents of the arguments after SingleDelete returns.
659 2 : func (d *DB) SingleDelete(key []byte, opts *WriteOptions) error {
660 2 : b := newBatch(d)
661 2 : _ = b.SingleDelete(key, opts)
662 2 : if err := d.Apply(b, opts); err != nil {
663 0 : return err
664 0 : }
665 : // Only release the batch on success.
666 2 : return b.Close()
667 : }
668 :
669 : // DeleteRange deletes all of the keys (and values) in the range [start,end)
670 : // (inclusive on start, exclusive on end).
671 : //
672 : // It is safe to modify the contents of the arguments after DeleteRange
673 : // returns.
674 2 : func (d *DB) DeleteRange(start, end []byte, opts *WriteOptions) error {
675 2 : b := newBatch(d)
676 2 : _ = b.DeleteRange(start, end, opts)
677 2 : if err := d.Apply(b, opts); err != nil {
678 1 : return err
679 1 : }
680 : // Only release the batch on success.
681 2 : return b.Close()
682 : }
683 :
684 : // Merge adds an action to the DB that merges the value at key with the new
685 : // value. The details of the merge are dependent upon the configured merge
686 : // operator.
687 : //
688 : // It is safe to modify the contents of the arguments after Merge returns.
689 2 : func (d *DB) Merge(key, value []byte, opts *WriteOptions) error {
690 2 : b := newBatch(d)
691 2 : _ = b.Merge(key, value, opts)
692 2 : if err := d.Apply(b, opts); err != nil {
693 1 : return err
694 1 : }
695 : // Only release the batch on success.
696 2 : return b.Close()
697 : }
698 :
699 : // LogData adds the specified to the batch. The data will be written to the
700 : // WAL, but not added to memtables or sstables. Log data is never indexed,
701 : // which makes it useful for testing WAL performance.
702 : //
703 : // It is safe to modify the contents of the argument after LogData returns.
704 2 : func (d *DB) LogData(data []byte, opts *WriteOptions) error {
705 2 : b := newBatch(d)
706 2 : _ = b.LogData(data, opts)
707 2 : if err := d.Apply(b, opts); err != nil {
708 1 : return err
709 1 : }
710 : // Only release the batch on success.
711 2 : return b.Close()
712 : }
713 :
714 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
715 : // timestamp suffix to value. The suffix is optional. If any portion of the key
716 : // range [start, end) is already set by a range key with the same suffix value,
717 : // RangeKeySet overrides it.
718 : //
719 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
720 2 : func (d *DB) RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error {
721 2 : b := newBatch(d)
722 2 : _ = b.RangeKeySet(start, end, suffix, value, opts)
723 2 : if err := d.Apply(b, opts); err != nil {
724 0 : return err
725 0 : }
726 : // Only release the batch on success.
727 2 : return b.Close()
728 : }
729 :
730 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
731 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
732 : // range key. RangeKeyUnset only removes portions of range keys that fall within
733 : // the [start, end) key span, and only range keys with suffixes that exactly
734 : // match the unset suffix.
735 : //
736 : // It is safe to modify the contents of the arguments after RangeKeyUnset
737 : // returns.
738 2 : func (d *DB) RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error {
739 2 : b := newBatch(d)
740 2 : _ = b.RangeKeyUnset(start, end, suffix, opts)
741 2 : if err := d.Apply(b, opts); err != nil {
742 0 : return err
743 0 : }
744 : // Only release the batch on success.
745 2 : return b.Close()
746 : }
747 :
748 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
749 : // (inclusive on start, exclusive on end). It does not delete point keys (for
750 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
751 : // bounds, including those with or without suffixes.
752 : //
753 : // It is safe to modify the contents of the arguments after RangeKeyDelete
754 : // returns.
755 2 : func (d *DB) RangeKeyDelete(start, end []byte, opts *WriteOptions) error {
756 2 : b := newBatch(d)
757 2 : _ = b.RangeKeyDelete(start, end, opts)
758 2 : if err := d.Apply(b, opts); err != nil {
759 0 : return err
760 0 : }
761 : // Only release the batch on success.
762 2 : return b.Close()
763 : }
764 :
765 : // Apply the operations contained in the batch to the DB. If the batch is large
766 : // the contents of the batch may be retained by the database. If that occurs
767 : // the batch contents will be cleared preventing the caller from attempting to
768 : // reuse them.
769 : //
770 : // It is safe to modify the contents of the arguments after Apply returns.
771 : //
772 : // Apply returns ErrInvalidBatch if the provided batch is invalid in any way.
773 2 : func (d *DB) Apply(batch *Batch, opts *WriteOptions) error {
774 2 : return d.applyInternal(batch, opts, false)
775 2 : }
776 :
777 : // ApplyNoSyncWait must only be used when opts.Sync is true and the caller
778 : // does not want to wait for the WAL fsync to happen. The method will return
779 : // once the mutation is applied to the memtable and is visible (note that a
780 : // mutation is visible before the WAL sync even in the wait case, so we have
781 : // not weakened the durability semantics). The caller must call Batch.SyncWait
782 : // to wait for the WAL fsync. The caller must not Close the batch without
783 : // first calling Batch.SyncWait.
784 : //
785 : // RECOMMENDATION: Prefer using Apply unless you really understand why you
786 : // need ApplyNoSyncWait.
787 : // EXPERIMENTAL: API/feature subject to change. Do not yet use outside
788 : // CockroachDB.
789 2 : func (d *DB) ApplyNoSyncWait(batch *Batch, opts *WriteOptions) error {
790 2 : if !opts.Sync {
791 0 : return errors.Errorf("cannot request asynchonous apply when WriteOptions.Sync is false")
792 0 : }
793 2 : return d.applyInternal(batch, opts, true)
794 : }
795 :
796 : // REQUIRES: noSyncWait => opts.Sync
797 2 : func (d *DB) applyInternal(batch *Batch, opts *WriteOptions, noSyncWait bool) error {
798 2 : if err := d.closed.Load(); err != nil {
799 1 : panic(err)
800 : }
801 2 : if batch.committing {
802 0 : panic("pebble: batch already committing")
803 : }
804 2 : if batch.applied.Load() {
805 0 : panic("pebble: batch already applied")
806 : }
807 2 : if d.opts.ReadOnly {
808 1 : return ErrReadOnly
809 1 : }
810 2 : if batch.db != nil && batch.db != d {
811 1 : panic(fmt.Sprintf("pebble: batch db mismatch: %p != %p", batch.db, d))
812 : }
813 :
814 2 : sync := opts.GetSync()
815 2 : if sync && d.opts.DisableWAL {
816 0 : return errors.New("pebble: WAL disabled")
817 0 : }
818 :
819 2 : if fmv := d.FormatMajorVersion(); fmv < batch.minimumFormatMajorVersion {
820 0 : panic(fmt.Sprintf(
821 0 : "pebble: batch requires at least format major version %d (current: %d)",
822 0 : batch.minimumFormatMajorVersion, fmv,
823 0 : ))
824 : }
825 :
826 2 : if batch.countRangeKeys > 0 {
827 2 : if d.split == nil {
828 0 : return errNoSplit
829 0 : }
830 : }
831 2 : batch.committing = true
832 2 :
833 2 : if batch.db == nil {
834 1 : if err := batch.refreshMemTableSize(); err != nil {
835 0 : return err
836 0 : }
837 : }
838 2 : if batch.memTableSize >= d.largeBatchThreshold {
839 2 : var err error
840 2 : batch.flushable, err = newFlushableBatch(batch, d.opts.Comparer)
841 2 : if err != nil {
842 0 : return err
843 0 : }
844 : }
845 2 : if err := d.commit.Commit(batch, sync, noSyncWait); err != nil {
846 0 : // There isn't much we can do on an error here. The commit pipeline will be
847 0 : // horked at this point.
848 0 : d.opts.Logger.Fatalf("pebble: fatal commit error: %v", err)
849 0 : }
850 : // If this is a large batch, we need to clear the batch contents as the
851 : // flushable batch may still be present in the flushables queue.
852 : //
853 : // TODO(peter): Currently large batches are written to the WAL. We could
854 : // skip the WAL write and instead wait for the large batch to be flushed to
855 : // an sstable. For a 100 MB batch, this might actually be faster. For a 1
856 : // GB batch this is almost certainly faster.
857 2 : if batch.flushable != nil {
858 2 : batch.data = nil
859 2 : }
860 2 : return nil
861 : }
862 :
863 2 : func (d *DB) commitApply(b *Batch, mem *memTable) error {
864 2 : if b.flushable != nil {
865 2 : // This is a large batch which was already added to the immutable queue.
866 2 : return nil
867 2 : }
868 2 : err := mem.apply(b, b.SeqNum())
869 2 : if err != nil {
870 0 : return err
871 0 : }
872 :
873 : // If the batch contains range tombstones and the database is configured
874 : // to flush range deletions, schedule a delayed flush so that disk space
875 : // may be reclaimed without additional writes or an explicit flush.
876 2 : if b.countRangeDels > 0 && d.opts.FlushDelayDeleteRange > 0 {
877 2 : d.mu.Lock()
878 2 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayDeleteRange)
879 2 : d.mu.Unlock()
880 2 : }
881 :
882 : // If the batch contains range keys and the database is configured to flush
883 : // range keys, schedule a delayed flush so that the range keys are cleared
884 : // from the memtable.
885 2 : if b.countRangeKeys > 0 && d.opts.FlushDelayRangeKey > 0 {
886 2 : d.mu.Lock()
887 2 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayRangeKey)
888 2 : d.mu.Unlock()
889 2 : }
890 :
891 2 : if mem.writerUnref() {
892 2 : d.mu.Lock()
893 2 : d.maybeScheduleFlush()
894 2 : d.mu.Unlock()
895 2 : }
896 2 : return nil
897 : }
898 :
899 2 : func (d *DB) commitWrite(b *Batch, syncWG *sync.WaitGroup, syncErr *error) (*memTable, error) {
900 2 : var size int64
901 2 : repr := b.Repr()
902 2 :
903 2 : if b.flushable != nil {
904 2 : // We have a large batch. Such batches are special in that they don't get
905 2 : // added to the memtable, and are instead inserted into the queue of
906 2 : // memtables. The call to makeRoomForWrite with this batch will force the
907 2 : // current memtable to be flushed. We want the large batch to be part of
908 2 : // the same log, so we add it to the WAL here, rather than after the call
909 2 : // to makeRoomForWrite().
910 2 : //
911 2 : // Set the sequence number since it was not set to the correct value earlier
912 2 : // (see comment in newFlushableBatch()).
913 2 : b.flushable.setSeqNum(b.SeqNum())
914 2 : if !d.opts.DisableWAL {
915 2 : var err error
916 2 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b.refData)
917 2 : if err != nil {
918 0 : panic(err)
919 : }
920 : }
921 : }
922 :
923 2 : d.mu.Lock()
924 2 :
925 2 : var err error
926 2 : if !b.ingestedSSTBatch {
927 2 : // Batches which contain keys of kind InternalKeyKindIngestSST will
928 2 : // never be applied to the memtable, so we don't need to make room for
929 2 : // write. For the other cases, switch out the memtable if there was not
930 2 : // enough room to store the batch.
931 2 : err = d.makeRoomForWrite(b)
932 2 : }
933 :
934 2 : if err == nil && !d.opts.DisableWAL {
935 2 : d.mu.log.bytesIn += uint64(len(repr))
936 2 : }
937 :
938 : // Grab a reference to the memtable while holding DB.mu. Note that for
939 : // non-flushable batches (b.flushable == nil) makeRoomForWrite() added a
940 : // reference to the memtable which will prevent it from being flushed until
941 : // we unreference it. This reference is dropped in DB.commitApply().
942 2 : mem := d.mu.mem.mutable
943 2 :
944 2 : d.mu.Unlock()
945 2 : if err != nil {
946 0 : return nil, err
947 0 : }
948 :
949 2 : if d.opts.DisableWAL {
950 2 : return mem, nil
951 2 : }
952 :
953 2 : if b.flushable == nil {
954 2 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b.refData)
955 2 : if err != nil {
956 0 : panic(err)
957 : }
958 : }
959 :
960 2 : d.logSize.Store(uint64(size))
961 2 : return mem, err
962 : }
963 :
964 : type iterAlloc struct {
965 : dbi Iterator
966 : keyBuf []byte
967 : boundsBuf [2][]byte
968 : prefixOrFullSeekKey []byte
969 : merging mergingIter
970 : mlevels [3 + numLevels]mergingIterLevel
971 : levels [3 + numLevels]levelIter
972 : levelsPositioned [3 + numLevels]bool
973 : }
974 :
975 : var iterAllocPool = sync.Pool{
976 2 : New: func() interface{} {
977 2 : return &iterAlloc{}
978 2 : },
979 : }
980 :
981 : // snapshotIterOpts denotes snapshot-related iterator options when calling
982 : // newIter. These are the possible cases for a snapshotIterOpts:
983 : // - No snapshot: All fields are zero values.
984 : // - Classic snapshot: Only `seqNum` is set. The latest readState will be used
985 : // and the specified seqNum will be used as the snapshot seqNum.
986 : // - EventuallyFileOnlySnapshot (EFOS) behaving as a classic snapshot. Only
987 : // the `seqNum` is set. The latest readState will be used
988 : // and the specified seqNum will be used as the snapshot seqNum.
989 : // - EFOS in file-only state: Only `seqNum` and `vers` are set. All the
990 : // relevant SSTs are referenced by the *version.
991 : // - EFOS that has been excised but is in alwaysCreateIters mode (tests only).
992 : // Only `seqNum` and `readState` are set.
993 : type snapshotIterOpts struct {
994 : seqNum uint64
995 : vers *version
996 : readState *readState
997 : }
998 :
999 : type batchIterOpts struct {
1000 : batchOnly bool
1001 : }
1002 : type newIterOpts struct {
1003 : snapshot snapshotIterOpts
1004 : batch batchIterOpts
1005 : }
1006 :
1007 : // newIter constructs a new iterator, merging in batch iterators as an extra
1008 : // level.
1009 : func (d *DB) newIter(
1010 : ctx context.Context, batch *Batch, internalOpts newIterOpts, o *IterOptions,
1011 2 : ) *Iterator {
1012 2 : if internalOpts.batch.batchOnly {
1013 1 : if batch == nil {
1014 0 : panic("batchOnly is true, but batch is nil")
1015 : }
1016 1 : if internalOpts.snapshot.vers != nil {
1017 0 : panic("batchOnly is true, but snapshotIterOpts is initialized")
1018 : }
1019 : }
1020 2 : if err := d.closed.Load(); err != nil {
1021 1 : panic(err)
1022 : }
1023 2 : seqNum := internalOpts.snapshot.seqNum
1024 2 : if o != nil && o.RangeKeyMasking.Suffix != nil && o.KeyTypes != IterKeyTypePointsAndRanges {
1025 0 : panic("pebble: range key masking requires IterKeyTypePointsAndRanges")
1026 : }
1027 2 : if (batch != nil || seqNum != 0) && (o != nil && o.OnlyReadGuaranteedDurable) {
1028 1 : // We could add support for OnlyReadGuaranteedDurable on snapshots if
1029 1 : // there was a need: this would require checking that the sequence number
1030 1 : // of the snapshot has been flushed, by comparing with
1031 1 : // DB.mem.queue[0].logSeqNum.
1032 1 : panic("OnlyReadGuaranteedDurable is not supported for batches or snapshots")
1033 : }
1034 2 : var readState *readState
1035 2 : var newIters tableNewIters
1036 2 : var newIterRangeKey keyspanimpl.TableNewSpanIter
1037 2 : if !internalOpts.batch.batchOnly {
1038 2 : // Grab and reference the current readState. This prevents the underlying
1039 2 : // files in the associated version from being deleted if there is a current
1040 2 : // compaction. The readState is unref'd by Iterator.Close().
1041 2 : if internalOpts.snapshot.vers == nil {
1042 2 : if internalOpts.snapshot.readState != nil {
1043 0 : readState = internalOpts.snapshot.readState
1044 0 : readState.ref()
1045 2 : } else {
1046 2 : // NB: loadReadState() calls readState.ref().
1047 2 : readState = d.loadReadState()
1048 2 : }
1049 2 : } else {
1050 2 : // vers != nil
1051 2 : internalOpts.snapshot.vers.Ref()
1052 2 : }
1053 :
1054 : // Determine the seqnum to read at after grabbing the read state (current and
1055 : // memtables) above.
1056 2 : if seqNum == 0 {
1057 2 : seqNum = d.mu.versions.visibleSeqNum.Load()
1058 2 : }
1059 2 : newIters = d.newIters
1060 2 : newIterRangeKey = d.tableNewRangeKeyIter
1061 : }
1062 :
1063 : // Bundle various structures under a single umbrella in order to allocate
1064 : // them together.
1065 2 : buf := iterAllocPool.Get().(*iterAlloc)
1066 2 : dbi := &buf.dbi
1067 2 : *dbi = Iterator{
1068 2 : ctx: ctx,
1069 2 : alloc: buf,
1070 2 : merge: d.merge,
1071 2 : comparer: *d.opts.Comparer,
1072 2 : readState: readState,
1073 2 : version: internalOpts.snapshot.vers,
1074 2 : keyBuf: buf.keyBuf,
1075 2 : prefixOrFullSeekKey: buf.prefixOrFullSeekKey,
1076 2 : boundsBuf: buf.boundsBuf,
1077 2 : batch: batch,
1078 2 : newIters: newIters,
1079 2 : newIterRangeKey: newIterRangeKey,
1080 2 : seqNum: seqNum,
1081 2 : batchOnlyIter: internalOpts.batch.batchOnly,
1082 2 : }
1083 2 : if o != nil {
1084 2 : dbi.opts = *o
1085 2 : dbi.processBounds(o.LowerBound, o.UpperBound)
1086 2 : }
1087 2 : dbi.opts.logger = d.opts.Logger
1088 2 : if d.opts.private.disableLazyCombinedIteration {
1089 1 : dbi.opts.disableLazyCombinedIteration = true
1090 1 : }
1091 2 : if batch != nil {
1092 2 : dbi.batchSeqNum = dbi.batch.nextSeqNum()
1093 2 : }
1094 2 : return finishInitializingIter(ctx, buf)
1095 : }
1096 :
1097 : // finishInitializingIter is a helper for doing the non-trivial initialization
1098 : // of an Iterator. It's invoked to perform the initial initialization of an
1099 : // Iterator during NewIter or Clone, and to perform reinitialization due to a
1100 : // change in IterOptions by a call to Iterator.SetOptions.
1101 2 : func finishInitializingIter(ctx context.Context, buf *iterAlloc) *Iterator {
1102 2 : // Short-hand.
1103 2 : dbi := &buf.dbi
1104 2 : var memtables flushableList
1105 2 : if dbi.readState != nil {
1106 2 : memtables = dbi.readState.memtables
1107 2 : }
1108 2 : if dbi.opts.OnlyReadGuaranteedDurable {
1109 1 : memtables = nil
1110 2 : } else {
1111 2 : // We only need to read from memtables which contain sequence numbers older
1112 2 : // than seqNum. Trim off newer memtables.
1113 2 : for i := len(memtables) - 1; i >= 0; i-- {
1114 2 : if logSeqNum := memtables[i].logSeqNum; logSeqNum < dbi.seqNum {
1115 2 : break
1116 : }
1117 2 : memtables = memtables[:i]
1118 : }
1119 : }
1120 :
1121 2 : if dbi.opts.pointKeys() {
1122 2 : // Construct the point iterator, initializing dbi.pointIter to point to
1123 2 : // dbi.merging. If this is called during a SetOptions call and this
1124 2 : // Iterator has already initialized dbi.merging, constructPointIter is a
1125 2 : // noop and an initialized pointIter already exists in dbi.pointIter.
1126 2 : dbi.constructPointIter(ctx, memtables, buf)
1127 2 : dbi.iter = dbi.pointIter
1128 2 : } else {
1129 2 : dbi.iter = emptyIter
1130 2 : }
1131 :
1132 2 : if dbi.opts.rangeKeys() {
1133 2 : dbi.rangeKeyMasking.init(dbi, dbi.comparer.Compare, dbi.comparer.Split)
1134 2 :
1135 2 : // When iterating over both point and range keys, don't create the
1136 2 : // range-key iterator stack immediately if we can avoid it. This
1137 2 : // optimization takes advantage of the expected sparseness of range
1138 2 : // keys, and configures the point-key iterator to dynamically switch to
1139 2 : // combined iteration when it observes a file containing range keys.
1140 2 : //
1141 2 : // Lazy combined iteration is not possible if a batch or a memtable
1142 2 : // contains any range keys.
1143 2 : useLazyCombinedIteration := dbi.rangeKey == nil &&
1144 2 : dbi.opts.KeyTypes == IterKeyTypePointsAndRanges &&
1145 2 : (dbi.batch == nil || dbi.batch.countRangeKeys == 0) &&
1146 2 : !dbi.opts.disableLazyCombinedIteration
1147 2 : if useLazyCombinedIteration {
1148 2 : // The user requested combined iteration, and there's no indexed
1149 2 : // batch currently containing range keys that would prevent lazy
1150 2 : // combined iteration. Check the memtables to see if they contain
1151 2 : // any range keys.
1152 2 : for i := range memtables {
1153 2 : if memtables[i].containsRangeKeys() {
1154 2 : useLazyCombinedIteration = false
1155 2 : break
1156 : }
1157 : }
1158 : }
1159 :
1160 2 : if useLazyCombinedIteration {
1161 2 : dbi.lazyCombinedIter = lazyCombinedIter{
1162 2 : parent: dbi,
1163 2 : pointIter: dbi.pointIter,
1164 2 : combinedIterState: combinedIterState{
1165 2 : initialized: false,
1166 2 : },
1167 2 : }
1168 2 : dbi.iter = &dbi.lazyCombinedIter
1169 2 : dbi.iter = invalidating.MaybeWrapIfInvariants(dbi.iter)
1170 2 : } else {
1171 2 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{
1172 2 : initialized: true,
1173 2 : }
1174 2 : if dbi.rangeKey == nil {
1175 2 : dbi.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1176 2 : dbi.rangeKey.init(dbi.comparer.Compare, dbi.comparer.Split, &dbi.opts)
1177 2 : dbi.constructRangeKeyIter()
1178 2 : } else {
1179 2 : dbi.rangeKey.iterConfig.SetBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
1180 2 : }
1181 :
1182 : // Wrap the point iterator (currently dbi.iter) with an interleaving
1183 : // iterator that interleaves range keys pulled from
1184 : // dbi.rangeKey.rangeKeyIter.
1185 : //
1186 : // NB: The interleaving iterator is always reinitialized, even if
1187 : // dbi already had an initialized range key iterator, in case the point
1188 : // iterator changed or the range key masking suffix changed.
1189 2 : dbi.rangeKey.iiter.Init(&dbi.comparer, dbi.iter, dbi.rangeKey.rangeKeyIter,
1190 2 : keyspan.InterleavingIterOpts{
1191 2 : Mask: &dbi.rangeKeyMasking,
1192 2 : LowerBound: dbi.opts.LowerBound,
1193 2 : UpperBound: dbi.opts.UpperBound,
1194 2 : })
1195 2 : dbi.iter = &dbi.rangeKey.iiter
1196 : }
1197 2 : } else {
1198 2 : // !dbi.opts.rangeKeys()
1199 2 : //
1200 2 : // Reset the combined iterator state. The initialized=true ensures the
1201 2 : // iterator doesn't unnecessarily try to switch to combined iteration.
1202 2 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{initialized: true}
1203 2 : }
1204 2 : return dbi
1205 : }
1206 :
1207 : // ScanInternal scans all internal keys within the specified bounds, truncating
1208 : // any rangedels and rangekeys to those bounds if they span past them. For use
1209 : // when an external user needs to be aware of all internal keys that make up a
1210 : // key range.
1211 : //
1212 : // Keys deleted by range deletions must not be returned or exposed by this
1213 : // method, while the range deletion deleting that key must be exposed using
1214 : // visitRangeDel. Keys that would be masked by range key masking (if an
1215 : // appropriate prefix were set) should be exposed, alongside the range key
1216 : // that would have masked it. This method also collapses all point keys into
1217 : // one InternalKey; so only one internal key at most per user key is returned
1218 : // to visitPointKey.
1219 : //
1220 : // If visitSharedFile is not nil, ScanInternal iterates in skip-shared iteration
1221 : // mode. In this iteration mode, sstables in levels L5 and L6 are skipped, and
1222 : // their metadatas truncated to [lower, upper) and passed into visitSharedFile.
1223 : // ErrInvalidSkipSharedIteration is returned if visitSharedFile is not nil and an
1224 : // sstable in L5 or L6 is found that is not in shared storage according to
1225 : // provider.IsShared, or an sstable in those levels contains a newer key than the
1226 : // snapshot sequence number (only applicable for snapshot.ScanInternal). Examples
1227 : // of when this could happen could be if Pebble started writing sstables before a
1228 : // creator ID was set (as creator IDs are necessary to enable shared storage)
1229 : // resulting in some lower level SSTs being on non-shared storage. Skip-shared
1230 : // iteration is invalid in those cases.
1231 : func (d *DB) ScanInternal(
1232 : ctx context.Context,
1233 : categoryAndQoS sstable.CategoryAndQoS,
1234 : lower, upper []byte,
1235 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error,
1236 : visitRangeDel func(start, end []byte, seqNum uint64) error,
1237 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error,
1238 : visitSharedFile func(sst *SharedSSTMeta) error,
1239 : visitExternalFile func(sst *ExternalFile) error,
1240 2 : ) error {
1241 2 : scanInternalOpts := &scanInternalOptions{
1242 2 : CategoryAndQoS: categoryAndQoS,
1243 2 : visitPointKey: visitPointKey,
1244 2 : visitRangeDel: visitRangeDel,
1245 2 : visitRangeKey: visitRangeKey,
1246 2 : visitSharedFile: visitSharedFile,
1247 2 : visitExternalFile: visitExternalFile,
1248 2 : IterOptions: IterOptions{
1249 2 : KeyTypes: IterKeyTypePointsAndRanges,
1250 2 : LowerBound: lower,
1251 2 : UpperBound: upper,
1252 2 : },
1253 2 : }
1254 2 : iter, err := d.newInternalIter(ctx, snapshotIterOpts{} /* snapshot */, scanInternalOpts)
1255 2 : if err != nil {
1256 0 : return err
1257 0 : }
1258 2 : defer iter.close()
1259 2 : return scanInternalImpl(ctx, lower, upper, iter, scanInternalOpts)
1260 : }
1261 :
1262 : // newInternalIter constructs and returns a new scanInternalIterator on this db.
1263 : // If o.skipSharedLevels is true, levels below sharedLevelsStart are *not* added
1264 : // to the internal iterator.
1265 : //
1266 : // TODO(bilal): This method has a lot of similarities with db.newIter as well as
1267 : // finishInitializingIter. Both pairs of methods should be refactored to reduce
1268 : // this duplication.
1269 : func (d *DB) newInternalIter(
1270 : ctx context.Context, sOpts snapshotIterOpts, o *scanInternalOptions,
1271 2 : ) (*scanInternalIterator, error) {
1272 2 : if err := d.closed.Load(); err != nil {
1273 0 : panic(err)
1274 : }
1275 : // Grab and reference the current readState. This prevents the underlying
1276 : // files in the associated version from being deleted if there is a current
1277 : // compaction. The readState is unref'd by Iterator.Close().
1278 2 : var readState *readState
1279 2 : if sOpts.vers == nil {
1280 2 : if sOpts.readState != nil {
1281 0 : readState = sOpts.readState
1282 0 : readState.ref()
1283 2 : } else {
1284 2 : readState = d.loadReadState()
1285 2 : }
1286 : }
1287 2 : if sOpts.vers != nil {
1288 1 : sOpts.vers.Ref()
1289 1 : }
1290 :
1291 : // Determine the seqnum to read at after grabbing the read state (current and
1292 : // memtables) above.
1293 2 : seqNum := sOpts.seqNum
1294 2 : if seqNum == 0 {
1295 2 : seqNum = d.mu.versions.visibleSeqNum.Load()
1296 2 : }
1297 :
1298 : // Bundle various structures under a single umbrella in order to allocate
1299 : // them together.
1300 2 : buf := iterAllocPool.Get().(*iterAlloc)
1301 2 : dbi := &scanInternalIterator{
1302 2 : ctx: ctx,
1303 2 : db: d,
1304 2 : comparer: d.opts.Comparer,
1305 2 : merge: d.opts.Merger.Merge,
1306 2 : readState: readState,
1307 2 : version: sOpts.vers,
1308 2 : alloc: buf,
1309 2 : newIters: d.newIters,
1310 2 : newIterRangeKey: d.tableNewRangeKeyIter,
1311 2 : seqNum: seqNum,
1312 2 : mergingIter: &buf.merging,
1313 2 : }
1314 2 : dbi.opts = *o
1315 2 : dbi.opts.logger = d.opts.Logger
1316 2 : if d.opts.private.disableLazyCombinedIteration {
1317 1 : dbi.opts.disableLazyCombinedIteration = true
1318 1 : }
1319 2 : return finishInitializingInternalIter(buf, dbi)
1320 : }
1321 :
1322 : func finishInitializingInternalIter(
1323 : buf *iterAlloc, i *scanInternalIterator,
1324 2 : ) (*scanInternalIterator, error) {
1325 2 : // Short-hand.
1326 2 : var memtables flushableList
1327 2 : if i.readState != nil {
1328 2 : memtables = i.readState.memtables
1329 2 : }
1330 : // We only need to read from memtables which contain sequence numbers older
1331 : // than seqNum. Trim off newer memtables.
1332 2 : for j := len(memtables) - 1; j >= 0; j-- {
1333 2 : if logSeqNum := memtables[j].logSeqNum; logSeqNum < i.seqNum {
1334 2 : break
1335 : }
1336 2 : memtables = memtables[:j]
1337 : }
1338 2 : i.initializeBoundBufs(i.opts.LowerBound, i.opts.UpperBound)
1339 2 :
1340 2 : if err := i.constructPointIter(i.opts.CategoryAndQoS, memtables, buf); err != nil {
1341 0 : return nil, err
1342 0 : }
1343 :
1344 : // For internal iterators, we skip the lazy combined iteration optimization
1345 : // entirely, and create the range key iterator stack directly.
1346 2 : i.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1347 2 : i.rangeKey.init(i.comparer.Compare, i.comparer.Split, &i.opts.IterOptions)
1348 2 : if err := i.constructRangeKeyIter(); err != nil {
1349 0 : return nil, err
1350 0 : }
1351 :
1352 : // Wrap the point iterator (currently i.iter) with an interleaving
1353 : // iterator that interleaves range keys pulled from
1354 : // i.rangeKey.rangeKeyIter.
1355 2 : i.rangeKey.iiter.Init(i.comparer, i.iter, i.rangeKey.rangeKeyIter,
1356 2 : keyspan.InterleavingIterOpts{
1357 2 : LowerBound: i.opts.LowerBound,
1358 2 : UpperBound: i.opts.UpperBound,
1359 2 : })
1360 2 : i.iter = &i.rangeKey.iiter
1361 2 :
1362 2 : return i, nil
1363 : }
1364 :
1365 : func (i *Iterator) constructPointIter(
1366 : ctx context.Context, memtables flushableList, buf *iterAlloc,
1367 2 : ) {
1368 2 : if i.pointIter != nil {
1369 2 : // Already have one.
1370 2 : return
1371 2 : }
1372 2 : internalOpts := internalIterOpts{stats: &i.stats.InternalStats}
1373 2 : if i.opts.RangeKeyMasking.Filter != nil {
1374 2 : internalOpts.boundLimitedFilter = &i.rangeKeyMasking
1375 2 : }
1376 :
1377 : // Merging levels and levels from iterAlloc.
1378 2 : mlevels := buf.mlevels[:0]
1379 2 : levels := buf.levels[:0]
1380 2 :
1381 2 : // We compute the number of levels needed ahead of time and reallocate a slice if
1382 2 : // the array from the iterAlloc isn't large enough. Doing this allocation once
1383 2 : // should improve the performance.
1384 2 : numMergingLevels := 0
1385 2 : numLevelIters := 0
1386 2 : if i.batch != nil {
1387 2 : numMergingLevels++
1388 2 : }
1389 :
1390 2 : var current *version
1391 2 : if !i.batchOnlyIter {
1392 2 : numMergingLevels += len(memtables)
1393 2 :
1394 2 : current = i.version
1395 2 : if current == nil {
1396 2 : current = i.readState.current
1397 2 : }
1398 2 : numMergingLevels += len(current.L0SublevelFiles)
1399 2 : numLevelIters += len(current.L0SublevelFiles)
1400 2 : for level := 1; level < len(current.Levels); level++ {
1401 2 : if current.Levels[level].Empty() {
1402 2 : continue
1403 : }
1404 2 : numMergingLevels++
1405 2 : numLevelIters++
1406 : }
1407 : }
1408 :
1409 2 : if numMergingLevels > cap(mlevels) {
1410 2 : mlevels = make([]mergingIterLevel, 0, numMergingLevels)
1411 2 : }
1412 2 : if numLevelIters > cap(levels) {
1413 2 : levels = make([]levelIter, 0, numLevelIters)
1414 2 : }
1415 :
1416 : // Top-level is the batch, if any.
1417 2 : if i.batch != nil {
1418 2 : if i.batch.index == nil {
1419 0 : // This isn't an indexed batch. We shouldn't have gotten this far.
1420 0 : panic(errors.AssertionFailedf("creating an iterator over an unindexed batch"))
1421 2 : } else {
1422 2 : i.batch.initInternalIter(&i.opts, &i.batchPointIter)
1423 2 : i.batch.initRangeDelIter(&i.opts, &i.batchRangeDelIter, i.batchSeqNum)
1424 2 : // Only include the batch's rangedel iterator if it's non-empty.
1425 2 : // This requires some subtle logic in the case a rangedel is later
1426 2 : // written to the batch and the view of the batch is refreshed
1427 2 : // during a call to SetOptions—in this case, we need to reconstruct
1428 2 : // the point iterator to add the batch rangedel iterator.
1429 2 : var rangeDelIter keyspan.FragmentIterator
1430 2 : if i.batchRangeDelIter.Count() > 0 {
1431 2 : rangeDelIter = &i.batchRangeDelIter
1432 2 : }
1433 2 : mlevels = append(mlevels, mergingIterLevel{
1434 2 : iter: &i.batchPointIter,
1435 2 : rangeDelIter: rangeDelIter,
1436 2 : })
1437 : }
1438 : }
1439 :
1440 2 : if !i.batchOnlyIter {
1441 2 : // Next are the memtables.
1442 2 : for j := len(memtables) - 1; j >= 0; j-- {
1443 2 : mem := memtables[j]
1444 2 : mlevels = append(mlevels, mergingIterLevel{
1445 2 : iter: mem.newIter(&i.opts),
1446 2 : rangeDelIter: mem.newRangeDelIter(&i.opts),
1447 2 : })
1448 2 : }
1449 :
1450 : // Next are the file levels: L0 sub-levels followed by lower levels.
1451 2 : mlevelsIndex := len(mlevels)
1452 2 : levelsIndex := len(levels)
1453 2 : mlevels = mlevels[:numMergingLevels]
1454 2 : levels = levels[:numLevelIters]
1455 2 : i.opts.snapshotForHideObsoletePoints = buf.dbi.seqNum
1456 2 : addLevelIterForFiles := func(files manifest.LevelIterator, level manifest.Level) {
1457 2 : li := &levels[levelsIndex]
1458 2 :
1459 2 : li.init(ctx, i.opts, &i.comparer, i.newIters, files, level, internalOpts)
1460 2 : li.initRangeDel(&mlevels[mlevelsIndex].rangeDelIter)
1461 2 : li.initBoundaryContext(&mlevels[mlevelsIndex].levelIterBoundaryContext)
1462 2 : li.initCombinedIterState(&i.lazyCombinedIter.combinedIterState)
1463 2 : mlevels[mlevelsIndex].levelIter = li
1464 2 : mlevels[mlevelsIndex].iter = invalidating.MaybeWrapIfInvariants(li)
1465 2 :
1466 2 : levelsIndex++
1467 2 : mlevelsIndex++
1468 2 : }
1469 :
1470 : // Add level iterators for the L0 sublevels, iterating from newest to
1471 : // oldest.
1472 2 : for i := len(current.L0SublevelFiles) - 1; i >= 0; i-- {
1473 2 : addLevelIterForFiles(current.L0SublevelFiles[i].Iter(), manifest.L0Sublevel(i))
1474 2 : }
1475 :
1476 : // Add level iterators for the non-empty non-L0 levels.
1477 2 : for level := 1; level < len(current.Levels); level++ {
1478 2 : if current.Levels[level].Empty() {
1479 2 : continue
1480 : }
1481 2 : addLevelIterForFiles(current.Levels[level].Iter(), manifest.Level(level))
1482 : }
1483 : }
1484 2 : buf.merging.init(&i.opts, &i.stats.InternalStats, i.comparer.Compare, i.comparer.Split, mlevels...)
1485 2 : if len(mlevels) <= cap(buf.levelsPositioned) {
1486 2 : buf.merging.levelsPositioned = buf.levelsPositioned[:len(mlevels)]
1487 2 : }
1488 2 : buf.merging.snapshot = i.seqNum
1489 2 : buf.merging.batchSnapshot = i.batchSeqNum
1490 2 : buf.merging.combinedIterState = &i.lazyCombinedIter.combinedIterState
1491 2 : i.pointIter = invalidating.MaybeWrapIfInvariants(&buf.merging).(topLevelIterator)
1492 2 : i.merging = &buf.merging
1493 : }
1494 :
1495 : // NewBatch returns a new empty write-only batch. Any reads on the batch will
1496 : // return an error. If the batch is committed it will be applied to the DB.
1497 2 : func (d *DB) NewBatch() *Batch {
1498 2 : return newBatch(d)
1499 2 : }
1500 :
1501 : // NewBatchWithSize is mostly identical to NewBatch, but it will allocate the
1502 : // the specified memory space for the internal slice in advance.
1503 0 : func (d *DB) NewBatchWithSize(size int) *Batch {
1504 0 : return newBatchWithSize(d, size)
1505 0 : }
1506 :
1507 : // NewIndexedBatch returns a new empty read-write batch. Any reads on the batch
1508 : // will read from both the batch and the DB. If the batch is committed it will
1509 : // be applied to the DB. An indexed batch is slower that a non-indexed batch
1510 : // for insert operations. If you do not need to perform reads on the batch, use
1511 : // NewBatch instead.
1512 2 : func (d *DB) NewIndexedBatch() *Batch {
1513 2 : return newIndexedBatch(d, d.opts.Comparer)
1514 2 : }
1515 :
1516 : // NewIndexedBatchWithSize is mostly identical to NewIndexedBatch, but it will
1517 : // allocate the specified memory space for the internal slice in advance.
1518 0 : func (d *DB) NewIndexedBatchWithSize(size int) *Batch {
1519 0 : return newIndexedBatchWithSize(d, d.opts.Comparer, size)
1520 0 : }
1521 :
1522 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
1523 : // return false). The iterator can be positioned via a call to SeekGE, SeekLT,
1524 : // First or Last. The iterator provides a point-in-time view of the current DB
1525 : // state. This view is maintained by preventing file deletions and preventing
1526 : // memtables referenced by the iterator from being deleted. Using an iterator
1527 : // to maintain a long-lived point-in-time view of the DB state can lead to an
1528 : // apparent memory and disk usage leak. Use snapshots (see NewSnapshot) for
1529 : // point-in-time snapshots which avoids these problems.
1530 2 : func (d *DB) NewIter(o *IterOptions) (*Iterator, error) {
1531 2 : return d.NewIterWithContext(context.Background(), o)
1532 2 : }
1533 :
1534 : // NewIterWithContext is like NewIter, and additionally accepts a context for
1535 : // tracing.
1536 2 : func (d *DB) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
1537 2 : return d.newIter(ctx, nil /* batch */, newIterOpts{}, o), nil
1538 2 : }
1539 :
1540 : // NewSnapshot returns a point-in-time view of the current DB state. Iterators
1541 : // created with this handle will all observe a stable snapshot of the current
1542 : // DB state. The caller must call Snapshot.Close() when the snapshot is no
1543 : // longer needed. Snapshots are not persisted across DB restarts (close ->
1544 : // open). Unlike the implicit snapshot maintained by an iterator, a snapshot
1545 : // will not prevent memtables from being released or sstables from being
1546 : // deleted. Instead, a snapshot prevents deletion of sequence numbers
1547 : // referenced by the snapshot.
1548 2 : func (d *DB) NewSnapshot() *Snapshot {
1549 2 : if err := d.closed.Load(); err != nil {
1550 1 : panic(err)
1551 : }
1552 :
1553 2 : d.mu.Lock()
1554 2 : s := &Snapshot{
1555 2 : db: d,
1556 2 : seqNum: d.mu.versions.visibleSeqNum.Load(),
1557 2 : }
1558 2 : d.mu.snapshots.pushBack(s)
1559 2 : d.mu.Unlock()
1560 2 : return s
1561 : }
1562 :
1563 : // NewEventuallyFileOnlySnapshot returns a point-in-time view of the current DB
1564 : // state, similar to NewSnapshot, but with consistency constrained to the
1565 : // provided set of key ranges. See the comment at EventuallyFileOnlySnapshot for
1566 : // its semantics.
1567 2 : func (d *DB) NewEventuallyFileOnlySnapshot(keyRanges []KeyRange) *EventuallyFileOnlySnapshot {
1568 2 : if err := d.closed.Load(); err != nil {
1569 0 : panic(err)
1570 : }
1571 2 : for i := range keyRanges {
1572 2 : if i > 0 && d.cmp(keyRanges[i-1].End, keyRanges[i].Start) > 0 {
1573 0 : panic("pebble: key ranges for eventually-file-only-snapshot not in order")
1574 : }
1575 : }
1576 2 : return d.makeEventuallyFileOnlySnapshot(keyRanges)
1577 : }
1578 :
1579 : // Close closes the DB.
1580 : //
1581 : // It is not safe to close a DB until all outstanding iterators are closed
1582 : // or to call Close concurrently with any other DB method. It is not valid
1583 : // to call any of a DB's methods after the DB has been closed.
1584 2 : func (d *DB) Close() error {
1585 2 : // Lock the commit pipeline for the duration of Close. This prevents a race
1586 2 : // with makeRoomForWrite. Rotating the WAL in makeRoomForWrite requires
1587 2 : // dropping d.mu several times for I/O. If Close only holds d.mu, an
1588 2 : // in-progress WAL rotation may re-acquire d.mu only once the database is
1589 2 : // closed.
1590 2 : //
1591 2 : // Additionally, locking the commit pipeline makes it more likely that
1592 2 : // (illegal) concurrent writes will observe d.closed.Load() != nil, creating
1593 2 : // more understable panics if the database is improperly used concurrently
1594 2 : // during Close.
1595 2 : d.commit.mu.Lock()
1596 2 : defer d.commit.mu.Unlock()
1597 2 : d.mu.Lock()
1598 2 : defer d.mu.Unlock()
1599 2 : if err := d.closed.Load(); err != nil {
1600 1 : panic(err)
1601 : }
1602 :
1603 : // Clear the finalizer that is used to check that an unreferenced DB has been
1604 : // closed. We're closing the DB here, so the check performed by that
1605 : // finalizer isn't necessary.
1606 : //
1607 : // Note: this is a no-op if invariants are disabled or race is enabled.
1608 2 : invariants.SetFinalizer(d.closed, nil)
1609 2 :
1610 2 : d.closed.Store(errors.WithStack(ErrClosed))
1611 2 : close(d.closedCh)
1612 2 :
1613 2 : defer d.opts.Cache.Unref()
1614 2 :
1615 2 : for d.mu.compact.compactingCount > 0 || d.mu.compact.downloadingCount > 0 || d.mu.compact.flushing {
1616 2 : d.mu.compact.cond.Wait()
1617 2 : }
1618 2 : for d.mu.tableStats.loading {
1619 2 : d.mu.tableStats.cond.Wait()
1620 2 : }
1621 2 : for d.mu.tableValidation.validating {
1622 1 : d.mu.tableValidation.cond.Wait()
1623 1 : }
1624 :
1625 2 : var err error
1626 2 : if n := len(d.mu.compact.inProgress); n > 0 {
1627 1 : err = errors.Errorf("pebble: %d unexpected in-progress compactions", errors.Safe(n))
1628 1 : }
1629 2 : err = firstError(err, d.mu.formatVers.marker.Close())
1630 2 : err = firstError(err, d.tableCache.close())
1631 2 : if !d.opts.ReadOnly {
1632 2 : if d.mu.log.writer != nil {
1633 2 : _, err2 := d.mu.log.writer.Close()
1634 2 : err = firstError(err, err2)
1635 2 : }
1636 1 : } else if d.mu.log.writer != nil {
1637 0 : panic("pebble: log-writer should be nil in read-only mode")
1638 : }
1639 2 : err = firstError(err, d.mu.log.manager.Close())
1640 2 : err = firstError(err, d.fileLock.Close())
1641 2 :
1642 2 : // Note that versionSet.close() only closes the MANIFEST. The versions list
1643 2 : // is still valid for the checks below.
1644 2 : err = firstError(err, d.mu.versions.close())
1645 2 :
1646 2 : err = firstError(err, d.dataDir.Close())
1647 2 :
1648 2 : d.readState.val.unrefLocked()
1649 2 :
1650 2 : current := d.mu.versions.currentVersion()
1651 2 : for v := d.mu.versions.versions.Front(); true; v = v.Next() {
1652 2 : refs := v.Refs()
1653 2 : if v == current {
1654 2 : if refs != 1 {
1655 1 : err = firstError(err, errors.Errorf("leaked iterators: current\n%s", v))
1656 1 : }
1657 2 : break
1658 : }
1659 0 : if refs != 0 {
1660 0 : err = firstError(err, errors.Errorf("leaked iterators:\n%s", v))
1661 0 : }
1662 : }
1663 :
1664 2 : for _, mem := range d.mu.mem.queue {
1665 2 : // Usually, we'd want to delete the files returned by readerUnref. But
1666 2 : // in this case, even if we're unreferencing the flushables, the
1667 2 : // flushables aren't obsolete. They will be reconstructed during WAL
1668 2 : // replay.
1669 2 : mem.readerUnrefLocked(false)
1670 2 : }
1671 : // If there's an unused, recycled memtable, we need to release its memory.
1672 2 : if obsoleteMemTable := d.memTableRecycle.Swap(nil); obsoleteMemTable != nil {
1673 2 : d.freeMemTable(obsoleteMemTable)
1674 2 : }
1675 2 : if reserved := d.memTableReserved.Load(); reserved != 0 {
1676 1 : err = firstError(err, errors.Errorf("leaked memtable reservation: %d", errors.Safe(reserved)))
1677 1 : }
1678 :
1679 : // Since we called d.readState.val.unrefLocked() above, we are expected to
1680 : // manually schedule deletion of obsolete files.
1681 2 : if len(d.mu.versions.obsoleteTables) > 0 {
1682 2 : d.deleteObsoleteFiles(d.newJobIDLocked())
1683 2 : }
1684 :
1685 2 : d.mu.Unlock()
1686 2 : d.compactionSchedulers.Wait()
1687 2 :
1688 2 : // Wait for all cleaning jobs to finish.
1689 2 : d.cleanupManager.Close()
1690 2 :
1691 2 : // Sanity check metrics.
1692 2 : if invariants.Enabled {
1693 2 : m := d.Metrics()
1694 2 : if m.Compact.NumInProgress > 0 || m.Compact.InProgressBytes > 0 {
1695 0 : d.mu.Lock()
1696 0 : panic(fmt.Sprintf("invalid metrics on close:\n%s", m))
1697 : }
1698 : }
1699 :
1700 2 : d.mu.Lock()
1701 2 :
1702 2 : // As a sanity check, ensure that there are no zombie tables. A non-zero count
1703 2 : // hints at a reference count leak.
1704 2 : if ztbls := len(d.mu.versions.zombieTables); ztbls > 0 {
1705 0 : err = firstError(err, errors.Errorf("non-zero zombie file count: %d", ztbls))
1706 0 : }
1707 :
1708 2 : err = firstError(err, d.objProvider.Close())
1709 2 :
1710 2 : // If the options include a closer to 'close' the filesystem, close it.
1711 2 : if d.opts.private.fsCloser != nil {
1712 2 : d.opts.private.fsCloser.Close()
1713 2 : }
1714 :
1715 : // Return an error if the user failed to close all open snapshots.
1716 2 : if v := d.mu.snapshots.count(); v > 0 {
1717 0 : err = firstError(err, errors.Errorf("leaked snapshots: %d open snapshots on DB %p", v, d))
1718 0 : }
1719 :
1720 2 : return err
1721 : }
1722 :
1723 : // Compact the specified range of keys in the database.
1724 2 : func (d *DB) Compact(start, end []byte, parallelize bool) error {
1725 2 : if err := d.closed.Load(); err != nil {
1726 1 : panic(err)
1727 : }
1728 2 : if d.opts.ReadOnly {
1729 1 : return ErrReadOnly
1730 1 : }
1731 2 : if d.cmp(start, end) >= 0 {
1732 2 : return errors.Errorf("Compact start %s is not less than end %s",
1733 2 : d.opts.Comparer.FormatKey(start), d.opts.Comparer.FormatKey(end))
1734 2 : }
1735 :
1736 2 : d.mu.Lock()
1737 2 : maxLevelWithFiles := 1
1738 2 : cur := d.mu.versions.currentVersion()
1739 2 : for level := 0; level < numLevels; level++ {
1740 2 : overlaps := cur.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1741 2 : if !overlaps.Empty() {
1742 2 : maxLevelWithFiles = level + 1
1743 2 : }
1744 : }
1745 :
1746 : // Determine if any memtable overlaps with the compaction range. We wait for
1747 : // any such overlap to flush (initiating a flush if necessary).
1748 2 : mem, err := func() (*flushableEntry, error) {
1749 2 : // Check to see if any files overlap with any of the memtables. The queue
1750 2 : // is ordered from oldest to newest with the mutable memtable being the
1751 2 : // last element in the slice. We want to wait for the newest table that
1752 2 : // overlaps.
1753 2 : for i := len(d.mu.mem.queue) - 1; i >= 0; i-- {
1754 2 : mem := d.mu.mem.queue[i]
1755 2 : var anyOverlaps bool
1756 2 : mem.computePossibleOverlaps(func(b bounded) shouldContinue {
1757 2 : anyOverlaps = true
1758 2 : return stopIteration
1759 2 : }, KeyRange{Start: start, End: end})
1760 2 : if !anyOverlaps {
1761 2 : continue
1762 : }
1763 2 : var err error
1764 2 : if mem.flushable == d.mu.mem.mutable {
1765 2 : // We have to hold both commitPipeline.mu and DB.mu when calling
1766 2 : // makeRoomForWrite(). Lock order requirements elsewhere force us to
1767 2 : // unlock DB.mu in order to grab commitPipeline.mu first.
1768 2 : d.mu.Unlock()
1769 2 : d.commit.mu.Lock()
1770 2 : d.mu.Lock()
1771 2 : defer d.commit.mu.Unlock()
1772 2 : if mem.flushable == d.mu.mem.mutable {
1773 2 : // Only flush if the active memtable is unchanged.
1774 2 : err = d.makeRoomForWrite(nil)
1775 2 : }
1776 : }
1777 2 : mem.flushForced = true
1778 2 : d.maybeScheduleFlush()
1779 2 : return mem, err
1780 : }
1781 2 : return nil, nil
1782 : }()
1783 :
1784 2 : d.mu.Unlock()
1785 2 :
1786 2 : if err != nil {
1787 0 : return err
1788 0 : }
1789 2 : if mem != nil {
1790 2 : <-mem.flushed
1791 2 : }
1792 :
1793 2 : for level := 0; level < maxLevelWithFiles; {
1794 2 : for {
1795 2 : if err := d.manualCompact(
1796 2 : start, end, level, parallelize); err != nil {
1797 1 : if errors.Is(err, ErrCancelledCompaction) {
1798 1 : continue
1799 : }
1800 1 : return err
1801 : }
1802 2 : break
1803 : }
1804 2 : level++
1805 2 : if level == numLevels-1 {
1806 2 : // A manual compaction of the bottommost level occurred.
1807 2 : // There is no next level to try and compact.
1808 2 : break
1809 : }
1810 : }
1811 2 : return nil
1812 : }
1813 :
1814 2 : func (d *DB) manualCompact(start, end []byte, level int, parallelize bool) error {
1815 2 : d.mu.Lock()
1816 2 : curr := d.mu.versions.currentVersion()
1817 2 : files := curr.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1818 2 : if files.Empty() {
1819 2 : d.mu.Unlock()
1820 2 : return nil
1821 2 : }
1822 :
1823 2 : var compactions []*manualCompaction
1824 2 : if parallelize {
1825 2 : compactions = append(compactions, d.splitManualCompaction(start, end, level)...)
1826 2 : } else {
1827 2 : compactions = append(compactions, &manualCompaction{
1828 2 : level: level,
1829 2 : done: make(chan error, 1),
1830 2 : start: start,
1831 2 : end: end,
1832 2 : })
1833 2 : }
1834 2 : d.mu.compact.manual = append(d.mu.compact.manual, compactions...)
1835 2 : d.maybeScheduleCompaction()
1836 2 : d.mu.Unlock()
1837 2 :
1838 2 : // Each of the channels is guaranteed to be eventually sent to once. After a
1839 2 : // compaction is possibly picked in d.maybeScheduleCompaction(), either the
1840 2 : // compaction is dropped, executed after being scheduled, or retried later.
1841 2 : // Assuming eventual progress when a compaction is retried, all outcomes send
1842 2 : // a value to the done channel. Since the channels are buffered, it is not
1843 2 : // necessary to read from each channel, and so we can exit early in the event
1844 2 : // of an error.
1845 2 : for _, compaction := range compactions {
1846 2 : if err := <-compaction.done; err != nil {
1847 1 : return err
1848 1 : }
1849 : }
1850 2 : return nil
1851 : }
1852 :
1853 : // splitManualCompaction splits a manual compaction over [start,end] on level
1854 : // such that the resulting compactions have no key overlap.
1855 : func (d *DB) splitManualCompaction(
1856 : start, end []byte, level int,
1857 2 : ) (splitCompactions []*manualCompaction) {
1858 2 : curr := d.mu.versions.currentVersion()
1859 2 : endLevel := level + 1
1860 2 : baseLevel := d.mu.versions.picker.getBaseLevel()
1861 2 : if level == 0 {
1862 2 : endLevel = baseLevel
1863 2 : }
1864 2 : keyRanges := curr.CalculateInuseKeyRanges(level, endLevel, start, end)
1865 2 : for _, keyRange := range keyRanges {
1866 2 : splitCompactions = append(splitCompactions, &manualCompaction{
1867 2 : level: level,
1868 2 : done: make(chan error, 1),
1869 2 : start: keyRange.Start,
1870 2 : end: keyRange.End,
1871 2 : split: true,
1872 2 : })
1873 2 : }
1874 2 : return splitCompactions
1875 : }
1876 :
1877 : // Flush the memtable to stable storage.
1878 2 : func (d *DB) Flush() error {
1879 2 : flushDone, err := d.AsyncFlush()
1880 2 : if err != nil {
1881 1 : return err
1882 1 : }
1883 2 : <-flushDone
1884 2 : return nil
1885 : }
1886 :
1887 : // AsyncFlush asynchronously flushes the memtable to stable storage.
1888 : //
1889 : // If no error is returned, the caller can receive from the returned channel in
1890 : // order to wait for the flush to complete.
1891 2 : func (d *DB) AsyncFlush() (<-chan struct{}, error) {
1892 2 : if err := d.closed.Load(); err != nil {
1893 1 : panic(err)
1894 : }
1895 2 : if d.opts.ReadOnly {
1896 1 : return nil, ErrReadOnly
1897 1 : }
1898 :
1899 2 : d.commit.mu.Lock()
1900 2 : defer d.commit.mu.Unlock()
1901 2 : d.mu.Lock()
1902 2 : defer d.mu.Unlock()
1903 2 : flushed := d.mu.mem.queue[len(d.mu.mem.queue)-1].flushed
1904 2 : err := d.makeRoomForWrite(nil)
1905 2 : if err != nil {
1906 0 : return nil, err
1907 0 : }
1908 2 : return flushed, nil
1909 : }
1910 :
1911 : // Metrics returns metrics about the database.
1912 2 : func (d *DB) Metrics() *Metrics {
1913 2 : metrics := &Metrics{}
1914 2 : walStats := d.mu.log.manager.Stats()
1915 2 :
1916 2 : d.mu.Lock()
1917 2 : vers := d.mu.versions.currentVersion()
1918 2 : *metrics = d.mu.versions.metrics
1919 2 : metrics.Compact.EstimatedDebt = d.mu.versions.picker.estimatedCompactionDebt(0)
1920 2 : metrics.Compact.InProgressBytes = d.mu.versions.atomicInProgressBytes.Load()
1921 2 : // TODO(radu): split this to separate the download compactions.
1922 2 : metrics.Compact.NumInProgress = int64(d.mu.compact.compactingCount + d.mu.compact.downloadingCount)
1923 2 : metrics.Compact.MarkedFiles = vers.Stats.MarkedForCompaction
1924 2 : metrics.Compact.Duration = d.mu.compact.duration
1925 2 : for c := range d.mu.compact.inProgress {
1926 1 : if c.kind != compactionKindFlush {
1927 1 : metrics.Compact.Duration += d.timeNow().Sub(c.beganAt)
1928 1 : }
1929 : }
1930 :
1931 2 : for _, m := range d.mu.mem.queue {
1932 2 : metrics.MemTable.Size += m.totalBytes()
1933 2 : }
1934 2 : metrics.Snapshots.Count = d.mu.snapshots.count()
1935 2 : if metrics.Snapshots.Count > 0 {
1936 0 : metrics.Snapshots.EarliestSeqNum = d.mu.snapshots.earliest()
1937 0 : }
1938 2 : metrics.Snapshots.PinnedKeys = d.mu.snapshots.cumulativePinnedCount
1939 2 : metrics.Snapshots.PinnedSize = d.mu.snapshots.cumulativePinnedSize
1940 2 : metrics.MemTable.Count = int64(len(d.mu.mem.queue))
1941 2 : metrics.MemTable.ZombieCount = d.memTableCount.Load() - metrics.MemTable.Count
1942 2 : metrics.MemTable.ZombieSize = uint64(d.memTableReserved.Load()) - metrics.MemTable.Size
1943 2 : metrics.WAL.ObsoleteFiles = int64(walStats.ObsoleteFileCount)
1944 2 : metrics.WAL.ObsoletePhysicalSize = walStats.ObsoleteFileSize
1945 2 : metrics.WAL.Files = int64(walStats.LiveFileCount)
1946 2 : // The current WAL's size (d.logSize) is the logical size, which may be less
1947 2 : // than the WAL's physical size if it was recycled. walStats.LiveFileSize
1948 2 : // includes the physical size of all live WALs, but for the current WAL it
1949 2 : // reflects the physical size when it was opened. So it is possible that
1950 2 : // d.atomic.logSize has exceeded that physical size. We allow for this
1951 2 : // anomaly.
1952 2 : metrics.WAL.PhysicalSize = walStats.LiveFileSize
1953 2 : metrics.WAL.BytesIn = d.mu.log.bytesIn // protected by d.mu
1954 2 : metrics.WAL.Size = d.logSize.Load()
1955 2 : for i, n := 0, len(d.mu.mem.queue)-1; i < n; i++ {
1956 2 : metrics.WAL.Size += d.mu.mem.queue[i].logSize
1957 2 : }
1958 2 : metrics.WAL.BytesWritten = metrics.Levels[0].BytesIn + metrics.WAL.Size
1959 2 : metrics.WAL.Failover = walStats.Failover
1960 2 :
1961 2 : if p := d.mu.versions.picker; p != nil {
1962 2 : compactions := d.getInProgressCompactionInfoLocked(nil)
1963 2 : for level, score := range p.getScores(compactions) {
1964 2 : metrics.Levels[level].Score = score
1965 2 : }
1966 : }
1967 2 : metrics.Table.ZombieCount = int64(len(d.mu.versions.zombieTables))
1968 2 : for _, info := range d.mu.versions.zombieTables {
1969 1 : metrics.Table.ZombieSize += info.FileSize
1970 1 : if info.isLocal {
1971 1 : metrics.Table.Local.ZombieSize += info.FileSize
1972 1 : }
1973 : }
1974 2 : metrics.private.optionsFileSize = d.optionsFileSize
1975 2 :
1976 2 : // TODO(jackson): Consider making these metrics optional.
1977 2 : metrics.Keys.RangeKeySetsCount = countRangeKeySetFragments(vers)
1978 2 : metrics.Keys.TombstoneCount = countTombstones(vers)
1979 2 :
1980 2 : d.mu.versions.logLock()
1981 2 : metrics.private.manifestFileSize = uint64(d.mu.versions.manifest.Size())
1982 2 : backingCount, backingTotalSize := d.mu.versions.virtualBackings.Stats()
1983 2 : metrics.Table.BackingTableCount = uint64(backingCount)
1984 2 : metrics.Table.BackingTableSize = backingTotalSize
1985 2 : d.mu.versions.logUnlock()
1986 2 :
1987 2 : metrics.LogWriter.FsyncLatency = d.mu.log.metrics.fsyncLatency
1988 2 : if err := metrics.LogWriter.Merge(&d.mu.log.metrics.LogWriterMetrics); err != nil {
1989 0 : d.opts.Logger.Errorf("metrics error: %s", err)
1990 0 : }
1991 2 : metrics.Flush.WriteThroughput = d.mu.compact.flushWriteThroughput
1992 2 : if d.mu.compact.flushing {
1993 1 : metrics.Flush.NumInProgress = 1
1994 1 : }
1995 2 : for i := 0; i < numLevels; i++ {
1996 2 : metrics.Levels[i].Additional.ValueBlocksSize = valueBlocksSizeForLevel(vers, i)
1997 2 : }
1998 :
1999 2 : d.mu.Unlock()
2000 2 :
2001 2 : metrics.BlockCache = d.opts.Cache.Metrics()
2002 2 : metrics.TableCache, metrics.Filter = d.tableCache.metrics()
2003 2 : metrics.TableIters = int64(d.tableCache.iterCount())
2004 2 : metrics.CategoryStats = d.tableCache.dbOpts.sstStatsCollector.GetStats()
2005 2 :
2006 2 : metrics.SecondaryCacheMetrics = d.objProvider.Metrics()
2007 2 :
2008 2 : metrics.Uptime = d.timeNow().Sub(d.openedAt)
2009 2 :
2010 2 : return metrics
2011 : }
2012 :
2013 : // sstablesOptions hold the optional parameters to retrieve TableInfo for all sstables.
2014 : type sstablesOptions struct {
2015 : // set to true will return the sstable properties in TableInfo
2016 : withProperties bool
2017 :
2018 : // if set, return sstables that overlap the key range (end-exclusive)
2019 : start []byte
2020 : end []byte
2021 :
2022 : withApproximateSpanBytes bool
2023 : }
2024 :
2025 : // SSTablesOption set optional parameter used by `DB.SSTables`.
2026 : type SSTablesOption func(*sstablesOptions)
2027 :
2028 : // WithProperties enable return sstable properties in each TableInfo.
2029 : //
2030 : // NOTE: if most of the sstable properties need to be read from disk,
2031 : // this options may make method `SSTables` quite slow.
2032 1 : func WithProperties() SSTablesOption {
2033 1 : return func(opt *sstablesOptions) {
2034 1 : opt.withProperties = true
2035 1 : }
2036 : }
2037 :
2038 : // WithKeyRangeFilter ensures returned sstables overlap start and end (end-exclusive)
2039 : // if start and end are both nil these properties have no effect.
2040 1 : func WithKeyRangeFilter(start, end []byte) SSTablesOption {
2041 1 : return func(opt *sstablesOptions) {
2042 1 : opt.end = end
2043 1 : opt.start = start
2044 1 : }
2045 : }
2046 :
2047 : // WithApproximateSpanBytes enables capturing the approximate number of bytes that
2048 : // overlap the provided key span for each sstable.
2049 : // NOTE: this option can only be used with WithKeyRangeFilter and WithProperties
2050 : // provided.
2051 1 : func WithApproximateSpanBytes() SSTablesOption {
2052 1 : return func(opt *sstablesOptions) {
2053 1 : opt.withApproximateSpanBytes = true
2054 1 : }
2055 : }
2056 :
2057 : // BackingType denotes the type of storage backing a given sstable.
2058 : type BackingType int
2059 :
2060 : const (
2061 : // BackingTypeLocal denotes an sstable stored on local disk according to the
2062 : // objprovider. This file is completely owned by us.
2063 : BackingTypeLocal BackingType = iota
2064 : // BackingTypeShared denotes an sstable stored on shared storage, created
2065 : // by this Pebble instance and possibly shared by other Pebble instances.
2066 : // These types of files have lifecycle managed by Pebble.
2067 : BackingTypeShared
2068 : // BackingTypeSharedForeign denotes an sstable stored on shared storage,
2069 : // created by a Pebble instance other than this one. These types of files have
2070 : // lifecycle managed by Pebble.
2071 : BackingTypeSharedForeign
2072 : // BackingTypeExternal denotes an sstable stored on external storage,
2073 : // not owned by any Pebble instance and with no refcounting/cleanup methods
2074 : // or lifecycle management. An example of an external file is a file restored
2075 : // from a backup.
2076 : BackingTypeExternal
2077 : )
2078 :
2079 : // SSTableInfo export manifest.TableInfo with sstable.Properties alongside
2080 : // other file backing info.
2081 : type SSTableInfo struct {
2082 : manifest.TableInfo
2083 : // Virtual indicates whether the sstable is virtual.
2084 : Virtual bool
2085 : // BackingSSTNum is the disk file number associated with the backing sstable.
2086 : // If Virtual is false, BackingSSTNum == PhysicalTableDiskFileNum(FileNum).
2087 : BackingSSTNum base.DiskFileNum
2088 : // BackingType is the type of storage backing this sstable.
2089 : BackingType BackingType
2090 : // Locator is the remote.Locator backing this sstable, if the backing type is
2091 : // not BackingTypeLocal.
2092 : Locator remote.Locator
2093 :
2094 : // Properties is the sstable properties of this table. If Virtual is true,
2095 : // then the Properties are associated with the backing sst.
2096 : Properties *sstable.Properties
2097 : }
2098 :
2099 : // SSTables retrieves the current sstables. The returned slice is indexed by
2100 : // level and each level is indexed by the position of the sstable within the
2101 : // level. Note that this information may be out of date due to concurrent
2102 : // flushes and compactions.
2103 1 : func (d *DB) SSTables(opts ...SSTablesOption) ([][]SSTableInfo, error) {
2104 1 : opt := &sstablesOptions{}
2105 1 : for _, fn := range opts {
2106 1 : fn(opt)
2107 1 : }
2108 :
2109 1 : if opt.withApproximateSpanBytes && !opt.withProperties {
2110 1 : return nil, errors.Errorf("Cannot use WithApproximateSpanBytes without WithProperties option.")
2111 1 : }
2112 1 : if opt.withApproximateSpanBytes && (opt.start == nil || opt.end == nil) {
2113 1 : return nil, errors.Errorf("Cannot use WithApproximateSpanBytes without WithKeyRangeFilter option.")
2114 1 : }
2115 :
2116 : // Grab and reference the current readState.
2117 1 : readState := d.loadReadState()
2118 1 : defer readState.unref()
2119 1 :
2120 1 : // TODO(peter): This is somewhat expensive, especially on a large
2121 1 : // database. It might be worthwhile to unify TableInfo and FileMetadata and
2122 1 : // then we could simply return current.Files. Note that RocksDB is doing
2123 1 : // something similar to the current code, so perhaps it isn't too bad.
2124 1 : srcLevels := readState.current.Levels
2125 1 : var totalTables int
2126 1 : for i := range srcLevels {
2127 1 : totalTables += srcLevels[i].Len()
2128 1 : }
2129 :
2130 1 : destTables := make([]SSTableInfo, totalTables)
2131 1 : destLevels := make([][]SSTableInfo, len(srcLevels))
2132 1 : for i := range destLevels {
2133 1 : iter := srcLevels[i].Iter()
2134 1 : j := 0
2135 1 : for m := iter.First(); m != nil; m = iter.Next() {
2136 1 : if opt.start != nil && opt.end != nil {
2137 1 : b := base.UserKeyBoundsEndExclusive(opt.start, opt.end)
2138 1 : if !m.Overlaps(d.opts.Comparer.Compare, &b) {
2139 1 : continue
2140 : }
2141 : }
2142 1 : destTables[j] = SSTableInfo{TableInfo: m.TableInfo()}
2143 1 : if opt.withProperties {
2144 1 : p, err := d.tableCache.getTableProperties(
2145 1 : m,
2146 1 : )
2147 1 : if err != nil {
2148 0 : return nil, err
2149 0 : }
2150 1 : destTables[j].Properties = p
2151 : }
2152 1 : destTables[j].Virtual = m.Virtual
2153 1 : destTables[j].BackingSSTNum = m.FileBacking.DiskFileNum
2154 1 : objMeta, err := d.objProvider.Lookup(fileTypeTable, m.FileBacking.DiskFileNum)
2155 1 : if err != nil {
2156 0 : return nil, err
2157 0 : }
2158 1 : if objMeta.IsRemote() {
2159 0 : if objMeta.IsShared() {
2160 0 : if d.objProvider.IsSharedForeign(objMeta) {
2161 0 : destTables[j].BackingType = BackingTypeSharedForeign
2162 0 : } else {
2163 0 : destTables[j].BackingType = BackingTypeShared
2164 0 : }
2165 0 : } else {
2166 0 : destTables[j].BackingType = BackingTypeExternal
2167 0 : }
2168 0 : destTables[j].Locator = objMeta.Remote.Locator
2169 1 : } else {
2170 1 : destTables[j].BackingType = BackingTypeLocal
2171 1 : }
2172 :
2173 1 : if opt.withApproximateSpanBytes {
2174 1 : var spanBytes uint64
2175 1 : if m.ContainedWithinSpan(d.opts.Comparer.Compare, opt.start, opt.end) {
2176 0 : spanBytes = m.Size
2177 1 : } else {
2178 1 : size, err := d.tableCache.estimateSize(m, opt.start, opt.end)
2179 1 : if err != nil {
2180 0 : return nil, err
2181 0 : }
2182 1 : spanBytes = size
2183 : }
2184 1 : propertiesCopy := *destTables[j].Properties
2185 1 :
2186 1 : // Deep copy user properties so approximate span bytes can be added.
2187 1 : propertiesCopy.UserProperties = make(map[string]string, len(destTables[j].Properties.UserProperties)+1)
2188 1 : for k, v := range destTables[j].Properties.UserProperties {
2189 0 : propertiesCopy.UserProperties[k] = v
2190 0 : }
2191 1 : propertiesCopy.UserProperties["approximate-span-bytes"] = strconv.FormatUint(spanBytes, 10)
2192 1 : destTables[j].Properties = &propertiesCopy
2193 : }
2194 1 : j++
2195 : }
2196 1 : destLevels[i] = destTables[:j]
2197 1 : destTables = destTables[j:]
2198 : }
2199 :
2200 1 : return destLevels, nil
2201 : }
2202 :
2203 : // EstimateDiskUsage returns the estimated filesystem space used in bytes for
2204 : // storing the range `[start, end]`. The estimation is computed as follows:
2205 : //
2206 : // - For sstables fully contained in the range the whole file size is included.
2207 : // - For sstables partially contained in the range the overlapping data block sizes
2208 : // are included. Even if a data block partially overlaps, or we cannot determine
2209 : // overlap due to abbreviated index keys, the full data block size is included in
2210 : // the estimation. Note that unlike fully contained sstables, none of the
2211 : // meta-block space is counted for partially overlapped files.
2212 : // - For virtual sstables, we use the overlap between start, end and the virtual
2213 : // sstable bounds to determine disk usage.
2214 : // - There may also exist WAL entries for unflushed keys in this range. This
2215 : // estimation currently excludes space used for the range in the WAL.
2216 1 : func (d *DB) EstimateDiskUsage(start, end []byte) (uint64, error) {
2217 1 : bytes, _, _, err := d.EstimateDiskUsageByBackingType(start, end)
2218 1 : return bytes, err
2219 1 : }
2220 :
2221 : // EstimateDiskUsageByBackingType is like EstimateDiskUsage but additionally
2222 : // returns the subsets of that size in remote ane external files.
2223 : func (d *DB) EstimateDiskUsageByBackingType(
2224 : start, end []byte,
2225 1 : ) (totalSize, remoteSize, externalSize uint64, _ error) {
2226 1 : if err := d.closed.Load(); err != nil {
2227 0 : panic(err)
2228 : }
2229 1 : if d.opts.Comparer.Compare(start, end) > 0 {
2230 0 : return 0, 0, 0, errors.New("invalid key-range specified (start > end)")
2231 0 : }
2232 :
2233 : // Grab and reference the current readState. This prevents the underlying
2234 : // files in the associated version from being deleted if there is a concurrent
2235 : // compaction.
2236 1 : readState := d.loadReadState()
2237 1 : defer readState.unref()
2238 1 :
2239 1 : for level, files := range readState.current.Levels {
2240 1 : iter := files.Iter()
2241 1 : if level > 0 {
2242 1 : // We can only use `Overlaps` to restrict `files` at L1+ since at L0 it
2243 1 : // expands the range iteratively until it has found a set of files that
2244 1 : // do not overlap any other L0 files outside that set.
2245 1 : overlaps := readState.current.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
2246 1 : iter = overlaps.Iter()
2247 1 : }
2248 1 : for file := iter.First(); file != nil; file = iter.Next() {
2249 1 : if d.opts.Comparer.Compare(start, file.Smallest.UserKey) <= 0 &&
2250 1 : d.opts.Comparer.Compare(file.Largest.UserKey, end) <= 0 {
2251 1 : // The range fully contains the file, so skip looking it up in
2252 1 : // table cache/looking at its indexes, and add the full file size.
2253 1 : meta, err := d.objProvider.Lookup(fileTypeTable, file.FileBacking.DiskFileNum)
2254 1 : if err != nil {
2255 0 : return 0, 0, 0, err
2256 0 : }
2257 1 : if meta.IsRemote() {
2258 0 : remoteSize += file.Size
2259 0 : if meta.Remote.CleanupMethod == objstorage.SharedNoCleanup {
2260 0 : externalSize += file.Size
2261 0 : }
2262 : }
2263 1 : totalSize += file.Size
2264 1 : } else if d.opts.Comparer.Compare(file.Smallest.UserKey, end) <= 0 &&
2265 1 : d.opts.Comparer.Compare(start, file.Largest.UserKey) <= 0 {
2266 1 : var size uint64
2267 1 : var err error
2268 1 : if file.Virtual {
2269 0 : err = d.tableCache.withVirtualReader(
2270 0 : file.VirtualMeta(),
2271 0 : func(r sstable.VirtualReader) (err error) {
2272 0 : size, err = r.EstimateDiskUsage(start, end)
2273 0 : return err
2274 0 : },
2275 : )
2276 1 : } else {
2277 1 : err = d.tableCache.withReader(
2278 1 : file.PhysicalMeta(),
2279 1 : func(r *sstable.Reader) (err error) {
2280 1 : size, err = r.EstimateDiskUsage(start, end)
2281 1 : return err
2282 1 : },
2283 : )
2284 : }
2285 1 : if err != nil {
2286 0 : return 0, 0, 0, err
2287 0 : }
2288 1 : meta, err := d.objProvider.Lookup(fileTypeTable, file.FileBacking.DiskFileNum)
2289 1 : if err != nil {
2290 0 : return 0, 0, 0, err
2291 0 : }
2292 1 : if meta.IsRemote() {
2293 0 : remoteSize += size
2294 0 : if meta.Remote.CleanupMethod == objstorage.SharedNoCleanup {
2295 0 : externalSize += size
2296 0 : }
2297 : }
2298 1 : totalSize += size
2299 : }
2300 : }
2301 : }
2302 1 : return totalSize, remoteSize, externalSize, nil
2303 : }
2304 :
2305 2 : func (d *DB) walPreallocateSize() int {
2306 2 : // Set the WAL preallocate size to 110% of the memtable size. Note that there
2307 2 : // is a bit of apples and oranges in units here as the memtabls size
2308 2 : // corresponds to the memory usage of the memtable while the WAL size is the
2309 2 : // size of the batches (plus overhead) stored in the WAL.
2310 2 : //
2311 2 : // TODO(peter): 110% of the memtable size is quite hefty for a block
2312 2 : // size. This logic is taken from GetWalPreallocateBlockSize in
2313 2 : // RocksDB. Could a smaller preallocation block size be used?
2314 2 : size := d.opts.MemTableSize
2315 2 : size = (size / 10) + size
2316 2 : return int(size)
2317 2 : }
2318 :
2319 2 : func (d *DB) newMemTable(logNum base.DiskFileNum, logSeqNum uint64) (*memTable, *flushableEntry) {
2320 2 : size := d.mu.mem.nextSize
2321 2 : if d.mu.mem.nextSize < d.opts.MemTableSize {
2322 2 : d.mu.mem.nextSize *= 2
2323 2 : if d.mu.mem.nextSize > d.opts.MemTableSize {
2324 1 : d.mu.mem.nextSize = d.opts.MemTableSize
2325 1 : }
2326 : }
2327 :
2328 2 : memtblOpts := memTableOptions{
2329 2 : Options: d.opts,
2330 2 : logSeqNum: logSeqNum,
2331 2 : }
2332 2 :
2333 2 : // Before attempting to allocate a new memtable, check if there's one
2334 2 : // available for recycling in memTableRecycle. Large contiguous allocations
2335 2 : // can be costly as fragmentation makes it more difficult to find a large
2336 2 : // contiguous free space. We've observed 64MB allocations taking 10ms+.
2337 2 : //
2338 2 : // To reduce these costly allocations, up to 1 obsolete memtable is stashed
2339 2 : // in `d.memTableRecycle` to allow a future memtable rotation to reuse
2340 2 : // existing memory.
2341 2 : var mem *memTable
2342 2 : mem = d.memTableRecycle.Swap(nil)
2343 2 : if mem != nil && uint64(len(mem.arenaBuf)) != size {
2344 2 : d.freeMemTable(mem)
2345 2 : mem = nil
2346 2 : }
2347 2 : if mem != nil {
2348 2 : // Carry through the existing buffer and memory reservation.
2349 2 : memtblOpts.arenaBuf = mem.arenaBuf
2350 2 : memtblOpts.releaseAccountingReservation = mem.releaseAccountingReservation
2351 2 : } else {
2352 2 : mem = new(memTable)
2353 2 : memtblOpts.arenaBuf = manual.New(int(size))
2354 2 : memtblOpts.releaseAccountingReservation = d.opts.Cache.Reserve(int(size))
2355 2 : d.memTableCount.Add(1)
2356 2 : d.memTableReserved.Add(int64(size))
2357 2 :
2358 2 : // Note: this is a no-op if invariants are disabled or race is enabled.
2359 2 : invariants.SetFinalizer(mem, checkMemTable)
2360 2 : }
2361 2 : mem.init(memtblOpts)
2362 2 :
2363 2 : entry := d.newFlushableEntry(mem, logNum, logSeqNum)
2364 2 : entry.releaseMemAccounting = func() {
2365 2 : // If the user leaks iterators, we may be releasing the memtable after
2366 2 : // the DB is already closed. In this case, we want to just release the
2367 2 : // memory because DB.Close won't come along to free it for us.
2368 2 : if err := d.closed.Load(); err != nil {
2369 2 : d.freeMemTable(mem)
2370 2 : return
2371 2 : }
2372 :
2373 : // The next memtable allocation might be able to reuse this memtable.
2374 : // Stash it on d.memTableRecycle.
2375 2 : if unusedMem := d.memTableRecycle.Swap(mem); unusedMem != nil {
2376 2 : // There was already a memtable waiting to be recycled. We're now
2377 2 : // responsible for freeing it.
2378 2 : d.freeMemTable(unusedMem)
2379 2 : }
2380 : }
2381 2 : return mem, entry
2382 : }
2383 :
2384 2 : func (d *DB) freeMemTable(m *memTable) {
2385 2 : d.memTableCount.Add(-1)
2386 2 : d.memTableReserved.Add(-int64(len(m.arenaBuf)))
2387 2 : m.free()
2388 2 : }
2389 :
2390 : func (d *DB) newFlushableEntry(
2391 : f flushable, logNum base.DiskFileNum, logSeqNum uint64,
2392 2 : ) *flushableEntry {
2393 2 : fe := &flushableEntry{
2394 2 : flushable: f,
2395 2 : flushed: make(chan struct{}),
2396 2 : logNum: logNum,
2397 2 : logSeqNum: logSeqNum,
2398 2 : deleteFn: d.mu.versions.addObsolete,
2399 2 : deleteFnLocked: d.mu.versions.addObsoleteLocked,
2400 2 : }
2401 2 : fe.readerRefs.Store(1)
2402 2 : return fe
2403 2 : }
2404 :
2405 : // makeRoomForWrite ensures that the memtable has room to hold the contents of
2406 : // Batch. It reserves the space in the memtable and adds a reference to the
2407 : // memtable. The caller must later ensure that the memtable is unreferenced. If
2408 : // the memtable is full, or a nil Batch is provided, the current memtable is
2409 : // rotated (marked as immutable) and a new mutable memtable is allocated. This
2410 : // memtable rotation also causes a log rotation.
2411 : //
2412 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2413 : // may be released and reacquired.
2414 2 : func (d *DB) makeRoomForWrite(b *Batch) error {
2415 2 : if b != nil && b.ingestedSSTBatch {
2416 0 : panic("pebble: invalid function call")
2417 : }
2418 :
2419 2 : force := b == nil || b.flushable != nil
2420 2 : stalled := false
2421 2 : for {
2422 2 : if b != nil && b.flushable == nil {
2423 2 : err := d.mu.mem.mutable.prepare(b)
2424 2 : if err != arenaskl.ErrArenaFull {
2425 2 : if stalled {
2426 2 : d.opts.EventListener.WriteStallEnd()
2427 2 : }
2428 2 : return err
2429 : }
2430 2 : } else if !force {
2431 2 : if stalled {
2432 2 : d.opts.EventListener.WriteStallEnd()
2433 2 : }
2434 2 : return nil
2435 : }
2436 : // force || err == ErrArenaFull, so we need to rotate the current memtable.
2437 2 : {
2438 2 : var size uint64
2439 2 : for i := range d.mu.mem.queue {
2440 2 : size += d.mu.mem.queue[i].totalBytes()
2441 2 : }
2442 : // If ElevateWriteStallThresholdForFailover is true, we give an
2443 : // unlimited memory budget for memtables. This is simpler than trying to
2444 : // configure an explicit value, given that memory resources can vary.
2445 : // When using WAL failover in CockroachDB, an OOM risk is worth
2446 : // tolerating for workloads that have a strict latency SLO. Also, an
2447 : // unlimited budget here does not mean that the disk stall in the
2448 : // primary will go unnoticed until the OOM -- CockroachDB is monitoring
2449 : // disk stalls, and we expect it to fail the node after ~60s if the
2450 : // primary is stalled.
2451 2 : if size >= uint64(d.opts.MemTableStopWritesThreshold)*d.opts.MemTableSize &&
2452 2 : !d.mu.log.manager.ElevateWriteStallThresholdForFailover() {
2453 2 : // We have filled up the current memtable, but already queued memtables
2454 2 : // are still flushing, so we wait.
2455 2 : if !stalled {
2456 2 : stalled = true
2457 2 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2458 2 : Reason: "memtable count limit reached",
2459 2 : })
2460 2 : }
2461 2 : now := time.Now()
2462 2 : d.mu.compact.cond.Wait()
2463 2 : if b != nil {
2464 2 : b.commitStats.MemTableWriteStallDuration += time.Since(now)
2465 2 : }
2466 2 : continue
2467 : }
2468 : }
2469 2 : l0ReadAmp := d.mu.versions.currentVersion().L0Sublevels.ReadAmplification()
2470 2 : if l0ReadAmp >= d.opts.L0StopWritesThreshold {
2471 2 : // There are too many level-0 files, so we wait.
2472 2 : if !stalled {
2473 2 : stalled = true
2474 2 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2475 2 : Reason: "L0 file count limit exceeded",
2476 2 : })
2477 2 : }
2478 2 : now := time.Now()
2479 2 : d.mu.compact.cond.Wait()
2480 2 : if b != nil {
2481 2 : b.commitStats.L0ReadAmpWriteStallDuration += time.Since(now)
2482 2 : }
2483 2 : continue
2484 : }
2485 :
2486 2 : var newLogNum base.DiskFileNum
2487 2 : var prevLogSize uint64
2488 2 : if !d.opts.DisableWAL {
2489 2 : now := time.Now()
2490 2 : newLogNum, prevLogSize = d.recycleWAL()
2491 2 : if b != nil {
2492 2 : b.commitStats.WALRotationDuration += time.Since(now)
2493 2 : }
2494 : }
2495 :
2496 2 : immMem := d.mu.mem.mutable
2497 2 : imm := d.mu.mem.queue[len(d.mu.mem.queue)-1]
2498 2 : imm.logSize = prevLogSize
2499 2 : imm.flushForced = imm.flushForced || (b == nil)
2500 2 :
2501 2 : // If we are manually flushing and we used less than half of the bytes in
2502 2 : // the memtable, don't increase the size for the next memtable. This
2503 2 : // reduces memtable memory pressure when an application is frequently
2504 2 : // manually flushing.
2505 2 : if (b == nil) && uint64(immMem.availBytes()) > immMem.totalBytes()/2 {
2506 2 : d.mu.mem.nextSize = immMem.totalBytes()
2507 2 : }
2508 :
2509 2 : if b != nil && b.flushable != nil {
2510 2 : // The batch is too large to fit in the memtable so add it directly to
2511 2 : // the immutable queue. The flushable batch is associated with the same
2512 2 : // log as the immutable memtable, but logically occurs after it in
2513 2 : // seqnum space. We ensure while flushing that the flushable batch
2514 2 : // is flushed along with the previous memtable in the flushable
2515 2 : // queue. See the top level comment in DB.flush1 to learn how this
2516 2 : // is ensured.
2517 2 : //
2518 2 : // See DB.commitWrite for the special handling of log writes for large
2519 2 : // batches. In particular, the large batch has already written to
2520 2 : // imm.logNum.
2521 2 : entry := d.newFlushableEntry(b.flushable, imm.logNum, b.SeqNum())
2522 2 : // The large batch is by definition large. Reserve space from the cache
2523 2 : // for it until it is flushed.
2524 2 : entry.releaseMemAccounting = d.opts.Cache.Reserve(int(b.flushable.totalBytes()))
2525 2 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2526 2 : }
2527 :
2528 2 : var logSeqNum uint64
2529 2 : if b != nil {
2530 2 : logSeqNum = b.SeqNum()
2531 2 : if b.flushable != nil {
2532 2 : logSeqNum += uint64(b.Count())
2533 2 : }
2534 2 : } else {
2535 2 : logSeqNum = d.mu.versions.logSeqNum.Load()
2536 2 : }
2537 2 : d.rotateMemtable(newLogNum, logSeqNum, immMem)
2538 2 : force = false
2539 : }
2540 : }
2541 :
2542 : // Both DB.mu and commitPipeline.mu must be held by the caller.
2543 2 : func (d *DB) rotateMemtable(newLogNum base.DiskFileNum, logSeqNum uint64, prev *memTable) {
2544 2 : // Create a new memtable, scheduling the previous one for flushing. We do
2545 2 : // this even if the previous memtable was empty because the DB.Flush
2546 2 : // mechanism is dependent on being able to wait for the empty memtable to
2547 2 : // flush. We can't just mark the empty memtable as flushed here because we
2548 2 : // also have to wait for all previous immutable tables to
2549 2 : // flush. Additionally, the memtable is tied to particular WAL file and we
2550 2 : // want to go through the flush path in order to recycle that WAL file.
2551 2 : //
2552 2 : // NB: newLogNum corresponds to the WAL that contains mutations that are
2553 2 : // present in the new memtable. When immutable memtables are flushed to
2554 2 : // disk, a VersionEdit will be created telling the manifest the minimum
2555 2 : // unflushed log number (which will be the next one in d.mu.mem.mutable
2556 2 : // that was not flushed).
2557 2 : //
2558 2 : // NB: prev should be the current mutable memtable.
2559 2 : var entry *flushableEntry
2560 2 : d.mu.mem.mutable, entry = d.newMemTable(newLogNum, logSeqNum)
2561 2 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2562 2 : d.updateReadStateLocked(nil)
2563 2 : if prev.writerUnref() {
2564 2 : d.maybeScheduleFlush()
2565 2 : }
2566 : }
2567 :
2568 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2569 : // may be released and reacquired.
2570 2 : func (d *DB) recycleWAL() (newLogNum base.DiskFileNum, prevLogSize uint64) {
2571 2 : if d.opts.DisableWAL {
2572 0 : panic("pebble: invalid function call")
2573 : }
2574 2 : jobID := d.newJobIDLocked()
2575 2 : newLogNum = d.mu.versions.getNextDiskFileNum()
2576 2 :
2577 2 : d.mu.Unlock()
2578 2 : // Close the previous log first. This writes an EOF trailer
2579 2 : // signifying the end of the file and syncs it to disk. We must
2580 2 : // close the previous log before linking the new log file,
2581 2 : // otherwise a crash could leave both logs with unclean tails, and
2582 2 : // Open will treat the previous log as corrupt.
2583 2 : offset, err := d.mu.log.writer.Close()
2584 2 : if err != nil {
2585 0 : // What to do here? Stumbling on doesn't seem worthwhile. If we failed to
2586 0 : // close the previous log it is possible we lost a write.
2587 0 : panic(err)
2588 : }
2589 2 : prevLogSize = uint64(offset)
2590 2 : metrics := d.mu.log.writer.Metrics()
2591 2 :
2592 2 : d.mu.Lock()
2593 2 : if err := d.mu.log.metrics.LogWriterMetrics.Merge(&metrics); err != nil {
2594 0 : d.opts.Logger.Errorf("metrics error: %s", err)
2595 0 : }
2596 :
2597 2 : d.mu.Unlock()
2598 2 : writer, err := d.mu.log.manager.Create(wal.NumWAL(newLogNum), int(jobID))
2599 2 : if err != nil {
2600 0 : panic(err)
2601 : }
2602 :
2603 2 : d.mu.Lock()
2604 2 : d.mu.log.writer = writer
2605 2 : return newLogNum, prevLogSize
2606 : }
2607 :
2608 2 : func (d *DB) getEarliestUnflushedSeqNumLocked() uint64 {
2609 2 : seqNum := InternalKeySeqNumMax
2610 2 : for i := range d.mu.mem.queue {
2611 2 : logSeqNum := d.mu.mem.queue[i].logSeqNum
2612 2 : if seqNum > logSeqNum {
2613 2 : seqNum = logSeqNum
2614 2 : }
2615 : }
2616 2 : return seqNum
2617 : }
2618 :
2619 2 : func (d *DB) getInProgressCompactionInfoLocked(finishing *compaction) (rv []compactionInfo) {
2620 2 : for c := range d.mu.compact.inProgress {
2621 2 : if len(c.flushing) == 0 && (finishing == nil || c != finishing) {
2622 2 : info := compactionInfo{
2623 2 : versionEditApplied: c.versionEditApplied,
2624 2 : inputs: c.inputs,
2625 2 : smallest: c.smallest,
2626 2 : largest: c.largest,
2627 2 : outputLevel: -1,
2628 2 : }
2629 2 : if c.outputLevel != nil {
2630 2 : info.outputLevel = c.outputLevel.level
2631 2 : }
2632 2 : rv = append(rv, info)
2633 : }
2634 : }
2635 2 : return
2636 : }
2637 :
2638 2 : func inProgressL0Compactions(inProgress []compactionInfo) []manifest.L0Compaction {
2639 2 : var compactions []manifest.L0Compaction
2640 2 : for _, info := range inProgress {
2641 2 : // Skip in-progress compactions that have already committed; the L0
2642 2 : // sublevels initialization code requires the set of in-progress
2643 2 : // compactions to be consistent with the current version. Compactions
2644 2 : // with versionEditApplied=true are already applied to the current
2645 2 : // version and but are performing cleanup without the database mutex.
2646 2 : if info.versionEditApplied {
2647 2 : continue
2648 : }
2649 2 : l0 := false
2650 2 : for _, cl := range info.inputs {
2651 2 : l0 = l0 || cl.level == 0
2652 2 : }
2653 2 : if !l0 {
2654 2 : continue
2655 : }
2656 2 : compactions = append(compactions, manifest.L0Compaction{
2657 2 : Smallest: info.smallest,
2658 2 : Largest: info.largest,
2659 2 : IsIntraL0: info.outputLevel == 0,
2660 2 : })
2661 : }
2662 2 : return compactions
2663 : }
2664 :
2665 : // firstError returns the first non-nil error of err0 and err1, or nil if both
2666 : // are nil.
2667 2 : func firstError(err0, err1 error) error {
2668 2 : if err0 != nil {
2669 2 : return err0
2670 2 : }
2671 2 : return err1
2672 : }
2673 :
2674 : // SetCreatorID sets the CreatorID which is needed in order to use shared objects.
2675 : // Remote object usage is disabled until this method is called the first time.
2676 : // Once set, the Creator ID is persisted and cannot change.
2677 : //
2678 : // Does nothing if SharedStorage was not set in the options when the DB was
2679 : // opened or if the DB is in read-only mode.
2680 2 : func (d *DB) SetCreatorID(creatorID uint64) error {
2681 2 : if d.opts.Experimental.RemoteStorage == nil || d.opts.ReadOnly {
2682 0 : return nil
2683 0 : }
2684 2 : return d.objProvider.SetCreatorID(objstorage.CreatorID(creatorID))
2685 : }
2686 :
2687 : // KeyStatistics keeps track of the number of keys that have been pinned by a
2688 : // snapshot as well as counts of the different key kinds in the lsm.
2689 : //
2690 : // One way of using the accumulated stats, when we only have sets and dels,
2691 : // and say the counts are represented as del_count, set_count,
2692 : // del_latest_count, set_latest_count, snapshot_pinned_count.
2693 : //
2694 : // - del_latest_count + set_latest_count is the set of unique user keys
2695 : // (unique).
2696 : //
2697 : // - set_latest_count is the set of live unique user keys (live_unique).
2698 : //
2699 : // - Garbage is del_count + set_count - live_unique.
2700 : //
2701 : // - If everything were in the LSM, del_count+set_count-snapshot_pinned_count
2702 : // would also be the set of unique user keys (note that
2703 : // snapshot_pinned_count is counting something different -- see comment below).
2704 : // But snapshot_pinned_count only counts keys in the LSM so the excess here
2705 : // must be keys in memtables.
2706 : type KeyStatistics struct {
2707 : // TODO(sumeer): the SnapshotPinned* are incorrect in that these older
2708 : // versions can be in a different level. Either fix the accounting or
2709 : // rename these fields.
2710 :
2711 : // SnapshotPinnedKeys represents obsolete keys that cannot be elided during
2712 : // a compaction, because they are required by an open snapshot.
2713 : SnapshotPinnedKeys int
2714 : // SnapshotPinnedKeysBytes is the total number of bytes of all snapshot
2715 : // pinned keys.
2716 : SnapshotPinnedKeysBytes uint64
2717 : // KindsCount is the count for each kind of key. It includes point keys,
2718 : // range deletes and range keys.
2719 : KindsCount [InternalKeyKindMax + 1]int
2720 : // LatestKindsCount is the count for each kind of key when it is the latest
2721 : // kind for a user key. It is only populated for point keys.
2722 : LatestKindsCount [InternalKeyKindMax + 1]int
2723 : }
2724 :
2725 : // LSMKeyStatistics is used by DB.ScanStatistics.
2726 : type LSMKeyStatistics struct {
2727 : Accumulated KeyStatistics
2728 : // Levels contains statistics only for point keys. Range deletions and range keys will
2729 : // appear in Accumulated but not Levels.
2730 : Levels [numLevels]KeyStatistics
2731 : // BytesRead represents the logical, pre-compression size of keys and values read
2732 : BytesRead uint64
2733 : }
2734 :
2735 : // ScanStatisticsOptions is used by DB.ScanStatistics.
2736 : type ScanStatisticsOptions struct {
2737 : // LimitBytesPerSecond indicates the number of bytes that are able to be read
2738 : // per second using ScanInternal.
2739 : // A value of 0 indicates that there is no limit set.
2740 : LimitBytesPerSecond int64
2741 : }
2742 :
2743 : // ScanStatistics returns the count of different key kinds within the lsm for a
2744 : // key span [lower, upper) as well as the number of snapshot keys.
2745 : func (d *DB) ScanStatistics(
2746 : ctx context.Context, lower, upper []byte, opts ScanStatisticsOptions,
2747 1 : ) (LSMKeyStatistics, error) {
2748 1 : stats := LSMKeyStatistics{}
2749 1 : var prevKey InternalKey
2750 1 : var rateLimitFunc func(key *InternalKey, val LazyValue) error
2751 1 : tb := tokenbucket.TokenBucket{}
2752 1 :
2753 1 : if opts.LimitBytesPerSecond != 0 {
2754 0 : // Each "token" roughly corresponds to a byte that was read.
2755 0 : tb.Init(tokenbucket.TokensPerSecond(opts.LimitBytesPerSecond), tokenbucket.Tokens(1024))
2756 0 : rateLimitFunc = func(key *InternalKey, val LazyValue) error {
2757 0 : return tb.WaitCtx(ctx, tokenbucket.Tokens(key.Size()+val.Len()))
2758 0 : }
2759 : }
2760 :
2761 1 : scanInternalOpts := &scanInternalOptions{
2762 1 : visitPointKey: func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error {
2763 1 : // If the previous key is equal to the current point key, the current key was
2764 1 : // pinned by a snapshot.
2765 1 : size := uint64(key.Size())
2766 1 : kind := key.Kind()
2767 1 : sameKey := d.equal(prevKey.UserKey, key.UserKey)
2768 1 : if iterInfo.Kind == IteratorLevelLSM && sameKey {
2769 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeys++
2770 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeysBytes += size
2771 1 : stats.Accumulated.SnapshotPinnedKeys++
2772 1 : stats.Accumulated.SnapshotPinnedKeysBytes += size
2773 1 : }
2774 1 : if iterInfo.Kind == IteratorLevelLSM {
2775 1 : stats.Levels[iterInfo.Level].KindsCount[kind]++
2776 1 : }
2777 1 : if !sameKey {
2778 1 : if iterInfo.Kind == IteratorLevelLSM {
2779 1 : stats.Levels[iterInfo.Level].LatestKindsCount[kind]++
2780 1 : }
2781 1 : stats.Accumulated.LatestKindsCount[kind]++
2782 : }
2783 :
2784 1 : stats.Accumulated.KindsCount[kind]++
2785 1 : prevKey.CopyFrom(*key)
2786 1 : stats.BytesRead += uint64(key.Size() + value.Len())
2787 1 : return nil
2788 : },
2789 0 : visitRangeDel: func(start, end []byte, seqNum uint64) error {
2790 0 : stats.Accumulated.KindsCount[InternalKeyKindRangeDelete]++
2791 0 : stats.BytesRead += uint64(len(start) + len(end))
2792 0 : return nil
2793 0 : },
2794 0 : visitRangeKey: func(start, end []byte, keys []rangekey.Key) error {
2795 0 : stats.BytesRead += uint64(len(start) + len(end))
2796 0 : for _, key := range keys {
2797 0 : stats.Accumulated.KindsCount[key.Kind()]++
2798 0 : stats.BytesRead += uint64(len(key.Value) + len(key.Suffix))
2799 0 : }
2800 0 : return nil
2801 : },
2802 : includeObsoleteKeys: true,
2803 : IterOptions: IterOptions{
2804 : KeyTypes: IterKeyTypePointsAndRanges,
2805 : LowerBound: lower,
2806 : UpperBound: upper,
2807 : },
2808 : rateLimitFunc: rateLimitFunc,
2809 : }
2810 1 : iter, err := d.newInternalIter(ctx, snapshotIterOpts{}, scanInternalOpts)
2811 1 : if err != nil {
2812 0 : return LSMKeyStatistics{}, err
2813 0 : }
2814 1 : defer iter.close()
2815 1 :
2816 1 : err = scanInternalImpl(ctx, lower, upper, iter, scanInternalOpts)
2817 1 :
2818 1 : if err != nil {
2819 0 : return LSMKeyStatistics{}, err
2820 0 : }
2821 :
2822 1 : return stats, nil
2823 : }
2824 :
2825 : // ObjProvider returns the objstorage.Provider for this database. Meant to be
2826 : // used for internal purposes only.
2827 2 : func (d *DB) ObjProvider() objstorage.Provider {
2828 2 : return d.objProvider
2829 2 : }
2830 :
2831 1 : func (d *DB) checkVirtualBounds(m *fileMetadata) {
2832 1 : if !invariants.Enabled {
2833 0 : return
2834 0 : }
2835 :
2836 1 : objMeta, err := d.objProvider.Lookup(fileTypeTable, m.FileBacking.DiskFileNum)
2837 1 : if err != nil {
2838 0 : panic(err)
2839 : }
2840 1 : if objMeta.IsExternal() {
2841 0 : // Nothing to do; bounds are expected to be loose.
2842 0 : return
2843 0 : }
2844 :
2845 1 : iters, err := d.newIters(context.TODO(), m, nil, internalIterOpts{}, iterPointKeys|iterRangeDeletions|iterRangeKeys)
2846 1 : if err != nil {
2847 0 : panic(errors.Wrap(err, "pebble: error creating iterators"))
2848 : }
2849 1 : defer iters.CloseAll()
2850 1 :
2851 1 : if m.HasPointKeys {
2852 1 : pointIter := iters.Point()
2853 1 : rangeDelIter := iters.RangeDeletion()
2854 1 :
2855 1 : // Check that the lower bound is tight.
2856 1 : pointKey, _ := pointIter.First()
2857 1 : rangeDel, err := rangeDelIter.First()
2858 1 : if err != nil {
2859 0 : panic(err)
2860 : }
2861 1 : if (rangeDel == nil || d.cmp(rangeDel.SmallestKey().UserKey, m.SmallestPointKey.UserKey) != 0) &&
2862 1 : (pointKey == nil || d.cmp(pointKey.UserKey, m.SmallestPointKey.UserKey) != 0) {
2863 0 : panic(errors.Newf("pebble: virtual sstable %s lower point key bound is not tight", m.FileNum))
2864 : }
2865 :
2866 : // Check that the upper bound is tight.
2867 1 : pointKey, _ = pointIter.Last()
2868 1 : rangeDel, err = rangeDelIter.Last()
2869 1 : if err != nil {
2870 0 : panic(err)
2871 : }
2872 1 : if (rangeDel == nil || d.cmp(rangeDel.LargestKey().UserKey, m.LargestPointKey.UserKey) != 0) &&
2873 1 : (pointKey == nil || d.cmp(pointKey.UserKey, m.LargestPointKey.UserKey) != 0) {
2874 0 : panic(errors.Newf("pebble: virtual sstable %s upper point key bound is not tight", m.FileNum))
2875 : }
2876 :
2877 : // Check that iterator keys are within bounds.
2878 1 : for key, _ := pointIter.First(); key != nil; key, _ = pointIter.Next() {
2879 1 : if d.cmp(key.UserKey, m.SmallestPointKey.UserKey) < 0 || d.cmp(key.UserKey, m.LargestPointKey.UserKey) > 0 {
2880 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, key.UserKey))
2881 : }
2882 : }
2883 1 : s, err := rangeDelIter.First()
2884 1 : for ; s != nil; s, err = rangeDelIter.Next() {
2885 1 : if d.cmp(s.SmallestKey().UserKey, m.SmallestPointKey.UserKey) < 0 {
2886 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, s.SmallestKey().UserKey))
2887 : }
2888 1 : if d.cmp(s.LargestKey().UserKey, m.LargestPointKey.UserKey) > 0 {
2889 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, s.LargestKey().UserKey))
2890 : }
2891 : }
2892 1 : if err != nil {
2893 0 : panic(err)
2894 : }
2895 : }
2896 :
2897 1 : if !m.HasRangeKeys {
2898 1 : return
2899 1 : }
2900 0 : rangeKeyIter := iters.RangeKey()
2901 0 :
2902 0 : // Check that the lower bound is tight.
2903 0 : if s, err := rangeKeyIter.First(); err != nil {
2904 0 : panic(err)
2905 0 : } else if d.cmp(s.SmallestKey().UserKey, m.SmallestRangeKey.UserKey) != 0 {
2906 0 : panic(errors.Newf("pebble: virtual sstable %s lower range key bound is not tight", m.FileNum))
2907 : }
2908 :
2909 : // Check that upper bound is tight.
2910 0 : if s, err := rangeKeyIter.Last(); err != nil {
2911 0 : panic(err)
2912 0 : } else if d.cmp(s.LargestKey().UserKey, m.LargestRangeKey.UserKey) != 0 {
2913 0 : panic(errors.Newf("pebble: virtual sstable %s upper range key bound is not tight", m.FileNum))
2914 : }
2915 :
2916 0 : s, err := rangeKeyIter.First()
2917 0 : for ; s != nil; s, err = rangeKeyIter.Next() {
2918 0 : if d.cmp(s.SmallestKey().UserKey, m.SmallestRangeKey.UserKey) < 0 {
2919 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, s.SmallestKey().UserKey))
2920 : }
2921 0 : if d.cmp(s.LargestKey().UserKey, m.LargestRangeKey.UserKey) > 0 {
2922 0 : panic(errors.Newf("pebble: virtual sstable %s point key %s is not within bounds", m.FileNum, s.LargestKey().UserKey))
2923 : }
2924 : }
2925 0 : if err != nil {
2926 0 : panic(err)
2927 : }
2928 : }
2929 :
2930 : // DebugString returns a debugging string describing the LSM.
2931 0 : func (d *DB) DebugString() string {
2932 0 : d.mu.Lock()
2933 0 : defer d.mu.Unlock()
2934 0 : return d.mu.versions.currentVersion().DebugString()
2935 0 : }
|