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 : "reflect"
13 : "slices"
14 : "sync"
15 : "sync/atomic"
16 : "time"
17 : "unsafe"
18 :
19 : "github.com/cockroachdb/crlib/crtime"
20 : "github.com/cockroachdb/errors"
21 : "github.com/cockroachdb/pebble/internal/arenaskl"
22 : "github.com/cockroachdb/pebble/internal/base"
23 : "github.com/cockroachdb/pebble/internal/cache"
24 : "github.com/cockroachdb/pebble/internal/invalidating"
25 : "github.com/cockroachdb/pebble/internal/invariants"
26 : "github.com/cockroachdb/pebble/internal/keyspan"
27 : "github.com/cockroachdb/pebble/internal/keyspan/keyspanimpl"
28 : "github.com/cockroachdb/pebble/internal/manifest"
29 : "github.com/cockroachdb/pebble/internal/manual"
30 : "github.com/cockroachdb/pebble/internal/problemspans"
31 : "github.com/cockroachdb/pebble/objstorage"
32 : "github.com/cockroachdb/pebble/objstorage/remote"
33 : "github.com/cockroachdb/pebble/rangekey"
34 : "github.com/cockroachdb/pebble/record"
35 : "github.com/cockroachdb/pebble/sstable"
36 : "github.com/cockroachdb/pebble/sstable/blob"
37 : "github.com/cockroachdb/pebble/sstable/block"
38 : "github.com/cockroachdb/pebble/vfs"
39 : "github.com/cockroachdb/pebble/vfs/atomicfs"
40 : "github.com/cockroachdb/pebble/wal"
41 : "github.com/cockroachdb/tokenbucket"
42 : "github.com/prometheus/client_golang/prometheus"
43 : )
44 :
45 : const (
46 : // minFileCacheSize is the minimum size of the file cache, for a single db.
47 : minFileCacheSize = 64
48 :
49 : // numNonFileCacheFiles is an approximation for the number of files
50 : // that we don't account for in the file cache, for a given db.
51 : numNonFileCacheFiles = 10
52 : )
53 :
54 : var (
55 : // ErrNotFound is returned when a get operation does not find the requested
56 : // key.
57 : ErrNotFound = base.ErrNotFound
58 : // ErrClosed is panicked when an operation is performed on a closed snapshot or
59 : // DB. Use errors.Is(err, ErrClosed) to check for this error.
60 : ErrClosed = errors.New("pebble: closed")
61 : // ErrReadOnly is returned when a write operation is performed on a read-only
62 : // database.
63 : ErrReadOnly = errors.New("pebble: read-only")
64 : // errNoSplit indicates that the user is trying to perform a range key
65 : // operation but the configured Comparer does not provide a Split
66 : // implementation.
67 : errNoSplit = errors.New("pebble: Comparer.Split required for range key operations")
68 : )
69 :
70 : // Reader is a readable key/value store.
71 : //
72 : // It is safe to call Get and NewIter from concurrent goroutines.
73 : type Reader interface {
74 : // Get gets the value for the given key. It returns ErrNotFound if the DB
75 : // does not contain the key.
76 : //
77 : // The caller should not modify the contents of the returned slice, but it is
78 : // safe to modify the contents of the argument after Get returns. The
79 : // returned slice will remain valid until the returned Closer is closed. On
80 : // success, the caller MUST call closer.Close() or a memory leak will occur.
81 : Get(key []byte) (value []byte, closer io.Closer, err error)
82 :
83 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
84 : // return false). The iterator can be positioned via a call to SeekGE,
85 : // SeekLT, First or Last.
86 : NewIter(o *IterOptions) (*Iterator, error)
87 :
88 : // NewIterWithContext is like NewIter, and additionally accepts a context
89 : // for tracing.
90 : NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error)
91 :
92 : // Close closes the Reader. It may or may not close any underlying io.Reader
93 : // or io.Writer, depending on how the DB was created.
94 : //
95 : // It is not safe to close a DB until all outstanding iterators are closed.
96 : // It is valid to call Close multiple times. Other methods should not be
97 : // called after the DB has been closed.
98 : Close() error
99 : }
100 :
101 : // Writer is a writable key/value store.
102 : //
103 : // Goroutine safety is dependent on the specific implementation.
104 : type Writer interface {
105 : // Apply the operations contained in the batch to the DB.
106 : //
107 : // It is safe to modify the contents of the arguments after Apply returns.
108 : Apply(batch *Batch, o *WriteOptions) error
109 :
110 : // Delete deletes the value for the given key. Deletes are blind all will
111 : // succeed even if the given key does not exist.
112 : //
113 : // It is safe to modify the contents of the arguments after Delete returns.
114 : Delete(key []byte, o *WriteOptions) error
115 :
116 : // DeleteSized behaves identically to Delete, but takes an additional
117 : // argument indicating the size of the value being deleted. DeleteSized
118 : // should be preferred when the caller has the expectation that there exists
119 : // a single internal KV pair for the key (eg, the key has not been
120 : // overwritten recently), and the caller knows the size of its value.
121 : //
122 : // DeleteSized will record the value size within the tombstone and use it to
123 : // inform compaction-picking heuristics which strive to reduce space
124 : // amplification in the LSM. This "calling your shot" mechanic allows the
125 : // storage engine to more accurately estimate and reduce space
126 : // amplification.
127 : //
128 : // It is safe to modify the contents of the arguments after DeleteSized
129 : // returns.
130 : DeleteSized(key []byte, valueSize uint32, _ *WriteOptions) error
131 :
132 : // SingleDelete is similar to Delete in that it deletes the value for the given key. Like Delete,
133 : // it is a blind operation that will succeed even if the given key does not exist.
134 : //
135 : // WARNING: Undefined (non-deterministic) behavior will result if a key is overwritten and
136 : // then deleted using SingleDelete. The record may appear deleted immediately, but be
137 : // resurrected at a later time after compactions have been performed. Or the record may
138 : // be deleted permanently. A Delete operation lays down a "tombstone" which shadows all
139 : // previous versions of a key. The SingleDelete operation is akin to "anti-matter" and will
140 : // only delete the most recently written version for a key. These different semantics allow
141 : // the DB to avoid propagating a SingleDelete operation during a compaction as soon as the
142 : // corresponding Set operation is encountered. These semantics require extreme care to handle
143 : // properly. Only use if you have a workload where the performance gain is critical and you
144 : // can guarantee that a record is written once and then deleted once.
145 : //
146 : // Note that SINGLEDEL, SET, SINGLEDEL, SET, DEL/RANGEDEL, ... from most
147 : // recent to older will work as intended since there is a single SET
148 : // sandwiched between SINGLEDEL/DEL/RANGEDEL.
149 : //
150 : // IMPLEMENTATION WARNING: By offering SingleDelete, Pebble must guarantee
151 : // that there is no duplication of writes inside Pebble. That is, idempotent
152 : // application of writes is insufficient. For example, if a SET operation
153 : // gets duplicated inside Pebble, resulting in say SET#20 and SET#17, the
154 : // caller may issue a SINGLEDEL#25 and it will not have the desired effect.
155 : // A duplication where a SET#20 is duplicated across two sstables will have
156 : // the same correctness problem, since the SINGLEDEL may meet one of the
157 : // SETs. This guarantee is partially achieved by ensuring that a WAL and a
158 : // flushable are usually in one-to-one correspondence, and atomically
159 : // updating the MANIFEST when the flushable is flushed (which ensures the
160 : // WAL will never be replayed). There is one exception: a flushableBatch (a
161 : // batch too large to fit in a memtable) is written to the end of the WAL
162 : // that it shares with the preceding memtable. This is safe because the
163 : // memtable and the flushableBatch are part of the same flush (see DB.flush1
164 : // where this invariant is maintained). If the memtable were to be flushed
165 : // without the flushableBatch, the WAL cannot yet be deleted and if a crash
166 : // happened, the WAL would be replayed despite the memtable already being
167 : // flushed.
168 : //
169 : // It is safe to modify the contents of the arguments after SingleDelete returns.
170 : SingleDelete(key []byte, o *WriteOptions) error
171 :
172 : // DeleteRange deletes all of the point keys (and values) in the range
173 : // [start,end) (inclusive on start, exclusive on end). DeleteRange does NOT
174 : // delete overlapping range keys (eg, keys set via RangeKeySet).
175 : //
176 : // It is safe to modify the contents of the arguments after DeleteRange
177 : // returns.
178 : DeleteRange(start, end []byte, o *WriteOptions) error
179 :
180 : // LogData adds the specified to the batch. The data will be written to the
181 : // WAL, but not added to memtables or sstables. Log data is never indexed,
182 : // which makes it useful for testing WAL performance.
183 : //
184 : // It is safe to modify the contents of the argument after LogData returns.
185 : LogData(data []byte, opts *WriteOptions) error
186 :
187 : // Merge merges the value for the given key. The details of the merge are
188 : // dependent upon the configured merge operation.
189 : //
190 : // It is safe to modify the contents of the arguments after Merge returns.
191 : Merge(key, value []byte, o *WriteOptions) error
192 :
193 : // Set sets the value for the given key. It overwrites any previous value
194 : // for that key; a DB is not a multi-map.
195 : //
196 : // It is safe to modify the contents of the arguments after Set returns.
197 : Set(key, value []byte, o *WriteOptions) error
198 :
199 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
200 : // timestamp suffix to value. The suffix is optional. If any portion of the key
201 : // range [start, end) is already set by a range key with the same suffix value,
202 : // RangeKeySet overrides it.
203 : //
204 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
205 : RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error
206 :
207 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
208 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
209 : // range key. RangeKeyUnset only removes portions of range keys that fall within
210 : // the [start, end) key span, and only range keys with suffixes that exactly
211 : // match the unset suffix.
212 : //
213 : // It is safe to modify the contents of the arguments after RangeKeyUnset
214 : // returns.
215 : RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error
216 :
217 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
218 : // (inclusive on start, exclusive on end). It does not delete point keys (for
219 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
220 : // bounds, including those with or without suffixes.
221 : //
222 : // It is safe to modify the contents of the arguments after RangeKeyDelete
223 : // returns.
224 : RangeKeyDelete(start, end []byte, opts *WriteOptions) error
225 : }
226 :
227 : // DB provides a concurrent, persistent ordered key/value store.
228 : //
229 : // A DB's basic operations (Get, Set, Delete) should be self-explanatory. Get
230 : // and Delete will return ErrNotFound if the requested key is not in the store.
231 : // Callers are free to ignore this error.
232 : //
233 : // A DB also allows for iterating over the key/value pairs in key order. If d
234 : // is a DB, the code below prints all key/value pairs whose keys are 'greater
235 : // than or equal to' k:
236 : //
237 : // iter := d.NewIter(readOptions)
238 : // for iter.SeekGE(k); iter.Valid(); iter.Next() {
239 : // fmt.Printf("key=%q value=%q\n", iter.Key(), iter.Value())
240 : // }
241 : // return iter.Close()
242 : //
243 : // The Options struct holds the optional parameters for the DB, including a
244 : // Comparer to define a 'less than' relationship over keys. It is always valid
245 : // to pass a nil *Options, which means to use the default parameter values. Any
246 : // zero field of a non-nil *Options also means to use the default value for
247 : // that parameter. Thus, the code below uses a custom Comparer, but the default
248 : // values for every other parameter:
249 : //
250 : // db := pebble.Open(&Options{
251 : // Comparer: myComparer,
252 : // })
253 : type DB struct {
254 : // The count and size of referenced memtables. This includes memtables
255 : // present in DB.mu.mem.queue, as well as memtables that have been flushed
256 : // but are still referenced by an inuse readState, as well as up to one
257 : // memTable waiting to be reused and stored in d.memTableRecycle.
258 : memTableCount atomic.Int64
259 : memTableReserved atomic.Int64 // number of bytes reserved in the cache for memtables
260 : // memTableRecycle holds a pointer to an obsolete memtable. The next
261 : // memtable allocation will reuse this memtable if it has not already been
262 : // recycled.
263 : memTableRecycle atomic.Pointer[memTable]
264 :
265 : // The logical size of the current WAL.
266 : logSize atomic.Uint64
267 : // The number of input bytes to the log. This is the raw size of the
268 : // batches written to the WAL, without the overhead of the record
269 : // envelopes.
270 : logBytesIn atomic.Uint64
271 :
272 : // The number of bytes available on disk.
273 : diskAvailBytes atomic.Uint64
274 : lowDiskSpaceReporter lowDiskSpaceReporter
275 :
276 : cacheHandle *cache.Handle
277 : dirname string
278 : opts *Options
279 : cmp Compare
280 : equal Equal
281 : merge Merge
282 : split Split
283 : abbreviatedKey AbbreviatedKey
284 : // The threshold for determining when a batch is "large" and will skip being
285 : // inserted into a memtable.
286 : largeBatchThreshold uint64
287 : // The current OPTIONS file number.
288 : optionsFileNum base.DiskFileNum
289 : // The on-disk size of the current OPTIONS file.
290 : optionsFileSize uint64
291 :
292 : // objProvider is used to access and manage SSTs.
293 : objProvider objstorage.Provider
294 :
295 : dataDirLock *base.DirLock
296 : dataDir vfs.File
297 :
298 : fileCache *fileCacheHandle
299 : newIters tableNewIters
300 : tableNewRangeKeyIter keyspanimpl.TableNewSpanIter
301 :
302 : commit *commitPipeline
303 :
304 : // readState provides access to the state needed for reading without needing
305 : // to acquire DB.mu.
306 : readState struct {
307 : sync.RWMutex
308 : val *readState
309 : }
310 :
311 : closed *atomic.Value
312 : closedCh chan struct{}
313 :
314 : cleanupManager *cleanupManager
315 :
316 : compactionScheduler CompactionScheduler
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.UpdateVersionLocked() 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 Writer is protected by commitPipeline.mu. This allows log writes
375 : // to be performed without holding DB.mu, but requires both
376 : // commitPipeline.mu and DB.mu to be held when rotating the WAL/memtable
377 : // (i.e. makeRoomForWrite). Can be nil.
378 : writer wal.Writer
379 : metrics struct {
380 : // fsyncLatency has its own internal synchronization, and is not
381 : // protected by mu.
382 : fsyncLatency prometheus.Histogram
383 : // Updated whenever a wal.Writer is closed.
384 : record.LogWriterMetrics
385 : }
386 : }
387 :
388 : mem struct {
389 : // The current mutable memTable. Readers of the pointer may hold
390 : // either DB.mu or commitPipeline.mu.
391 : //
392 : // Its internal fields are protected by commitPipeline.mu. This
393 : // allows batch commits to be performed without DB.mu as long as no
394 : // memtable rotation is required.
395 : //
396 : // Both commitPipeline.mu and DB.mu must be held when rotating the
397 : // memtable.
398 : mutable *memTable
399 : // Queue of flushables (the mutable memtable is at end). Elements are
400 : // added to the end of the slice and removed from the beginning. Once an
401 : // index is set it is never modified making a fixed slice immutable and
402 : // safe for concurrent reads.
403 : queue flushableList
404 : // nextSize is the size of the next memtable. The memtable size starts at
405 : // min(256KB,Options.MemTableSize) and doubles each time a new memtable
406 : // is allocated up to Options.MemTableSize. This reduces the memory
407 : // footprint of memtables when lots of DB instances are used concurrently
408 : // in test environments.
409 : nextSize uint64
410 : }
411 :
412 : compact struct {
413 : // Condition variable used to signal when a flush or compaction has
414 : // completed. Used by the write-stall mechanism to wait for the stall
415 : // condition to clear. See DB.makeRoomForWrite().
416 : cond sync.Cond
417 : // True when a flush is in progress.
418 : flushing bool
419 : // The number of ongoing non-download compactions.
420 : compactingCount int
421 : // The number of download compactions.
422 : downloadingCount int
423 : // The list of deletion hints, suggesting ranges for delete-only
424 : // compactions.
425 : deletionHints []deleteCompactionHint
426 : // The list of manual compactions. The next manual compaction to perform
427 : // is at the start of the list. New entries are added to the end.
428 : manual []*manualCompaction
429 : manualLen atomic.Int32
430 : // manualID is used to identify manualCompactions in the manual slice.
431 : manualID uint64
432 : // downloads is the list of pending download tasks. The next download to
433 : // perform is at the start of the list. New entries are added to the end.
434 : downloads []*downloadSpanTask
435 : // inProgress is the set of in-progress flushes and compactions.
436 : // It's used in the calculation of some metrics and to initialize L0
437 : // sublevels' state. Some of the compactions contained within this
438 : // map may have already committed an edit to the version but are
439 : // lingering performing cleanup, like deleting obsolete files.
440 : inProgress map[compaction]struct{}
441 : // burstConcurrency is the number of additional compaction
442 : // concurrency slots that have been added in order to allow
443 : // high-priority compactions to run.
444 : //
445 : // Today, it's used when scheduling 'high priority' blob file
446 : // rewrite compactions to reclaim disk space. Normally compactions
447 : // that reclaim disk space and don't reshape the LSM are considered
448 : // low priority. If there's sufficient space amplification,
449 : // heuristics may decide to let these space-reclamation compactions
450 : // run anyways.
451 : //
452 : // In this case we don't want to steal concurrency from the default
453 : // compactions, so we temporarily inflate the allowed compaction
454 : // concurrency for the duration of the compaction. This field is the
455 : // count of compactions within inProgress for which
456 : // UsesBurstConcurrency is true. It's incremented when one begins
457 : // and decremented when it completes.
458 : burstConcurrency atomic.Int32
459 :
460 : // rescheduleReadCompaction indicates to an iterator that a read compaction
461 : // should be scheduled.
462 : rescheduleReadCompaction bool
463 :
464 : // readCompactions is a readCompactionQueue which keeps track of the
465 : // compactions which we might have to perform.
466 : readCompactions readCompactionQueue
467 :
468 : // The cumulative duration of all completed compactions since Open.
469 : // Does not include flushes.
470 : duration time.Duration
471 : // Flush throughput metric.
472 : flushWriteThroughput ThroughputMetric
473 : // The idle start time for the flush "loop", i.e., when the flushing
474 : // bool above transitions to false.
475 : noOngoingFlushStartTime crtime.Mono
476 : }
477 :
478 : fileDeletions struct {
479 : // Non-zero when file cleaning is disableCount. The disableCount
480 : // count acts as a reference count to prohibit file cleaning. See
481 : // DB.{disable,enable}FileDeletions().
482 : disableCount int
483 : // queuedStats holds cumulative stats for files that have been
484 : // queued for deletion by the cleanup manager. These stats are
485 : // monotonically increasing for the *DB's lifetime.
486 : queuedStats obsoleteObjectStats
487 : }
488 :
489 : snapshots struct {
490 : // The list of active snapshots.
491 : snapshotList
492 :
493 : // The cumulative count and size of snapshot-pinned keys written to
494 : // sstables.
495 : cumulativePinnedCount uint64
496 : cumulativePinnedSize uint64
497 : }
498 :
499 : tableStats struct {
500 : // Condition variable used to signal the completion of a
501 : // job to collect table stats.
502 : cond sync.Cond
503 : // True when a stat collection operation is in progress.
504 : loading bool
505 : // True if stat collection has loaded statistics for all tables
506 : // other than those listed explicitly in pending. This flag starts
507 : // as false when a database is opened and flips to true once stat
508 : // collection has caught up.
509 : loadedInitial bool
510 : // A slice of files for which stats have not been computed.
511 : // Compactions, ingests, flushes append files to be processed. An
512 : // active stat collection goroutine clears the list and processes
513 : // them.
514 : pending []manifest.NewTableEntry
515 : }
516 :
517 : tableValidation struct {
518 : // cond is a condition variable used to signal the completion of a
519 : // job to validate one or more sstables.
520 : cond sync.Cond
521 : // pending is a slice of metadata for sstables waiting to be
522 : // validated. Only physical sstables should be added to the pending
523 : // queue.
524 : pending []manifest.NewTableEntry
525 : // validating is set to true when validation is running.
526 : validating bool
527 : }
528 : }
529 :
530 : fileSizeAnnotator manifest.TableAnnotator[fileSizeByBacking]
531 :
532 : // problemSpans keeps track of spans of keys within LSM levels where
533 : // compactions have failed; used to avoid retrying these compactions too
534 : // quickly.
535 : problemSpans problemspans.ByLevel
536 :
537 : // Normally equal to time.Now() but may be overridden in tests.
538 : timeNow func() time.Time
539 : // the time at database Open; may be used to compute metrics like effective
540 : // compaction concurrency
541 : openedAt time.Time
542 : }
543 :
544 : var _ Reader = (*DB)(nil)
545 : var _ Writer = (*DB)(nil)
546 :
547 : // TestOnlyWaitForCleaning MUST only be used in tests.
548 1 : func (d *DB) TestOnlyWaitForCleaning() {
549 1 : d.cleanupManager.Wait()
550 1 : }
551 :
552 : // Set sets the value for the given key. It overwrites any previous value
553 : // for that key; a DB is not a multi-map.
554 : //
555 : // It is safe to modify the contents of the arguments after Set returns.
556 1 : func (d *DB) Set(key, value []byte, opts *WriteOptions) error {
557 1 : b := newBatch(d)
558 1 : _ = b.Set(key, value, opts)
559 1 : if err := d.Apply(b, opts); err != nil {
560 1 : return err
561 1 : }
562 : // Only release the batch on success.
563 1 : return b.Close()
564 : }
565 :
566 : // Delete deletes the value for the given key. Deletes are blind all will
567 : // succeed even if the given key does not exist.
568 : //
569 : // It is safe to modify the contents of the arguments after Delete returns.
570 1 : func (d *DB) Delete(key []byte, opts *WriteOptions) error {
571 1 : b := newBatch(d)
572 1 : _ = b.Delete(key, opts)
573 1 : if err := d.Apply(b, opts); err != nil {
574 1 : return err
575 1 : }
576 : // Only release the batch on success.
577 1 : return b.Close()
578 : }
579 :
580 : // DeleteSized behaves identically to Delete, but takes an additional
581 : // argument indicating the size of the value being deleted. DeleteSized
582 : // should be preferred when the caller has the expectation that there exists
583 : // a single internal KV pair for the key (eg, the key has not been
584 : // overwritten recently), and the caller knows the size of its value.
585 : //
586 : // DeleteSized will record the value size within the tombstone and use it to
587 : // inform compaction-picking heuristics which strive to reduce space
588 : // amplification in the LSM. This "calling your shot" mechanic allows the
589 : // storage engine to more accurately estimate and reduce space amplification.
590 : //
591 : // It is safe to modify the contents of the arguments after DeleteSized
592 : // returns.
593 1 : func (d *DB) DeleteSized(key []byte, valueSize uint32, opts *WriteOptions) error {
594 1 : b := newBatch(d)
595 1 : _ = b.DeleteSized(key, valueSize, opts)
596 1 : if err := d.Apply(b, opts); err != nil {
597 0 : return err
598 0 : }
599 : // Only release the batch on success.
600 1 : return b.Close()
601 : }
602 :
603 : // SingleDelete adds an action to the batch that single deletes the entry for key.
604 : // See Writer.SingleDelete for more details on the semantics of SingleDelete.
605 : //
606 : // WARNING: See the detailed warning in Writer.SingleDelete before using this.
607 : //
608 : // It is safe to modify the contents of the arguments after SingleDelete returns.
609 1 : func (d *DB) SingleDelete(key []byte, opts *WriteOptions) error {
610 1 : b := newBatch(d)
611 1 : _ = b.SingleDelete(key, opts)
612 1 : if err := d.Apply(b, opts); err != nil {
613 0 : return err
614 0 : }
615 : // Only release the batch on success.
616 1 : return b.Close()
617 : }
618 :
619 : // DeleteRange deletes all of the keys (and values) in the range [start,end)
620 : // (inclusive on start, exclusive on end).
621 : //
622 : // It is safe to modify the contents of the arguments after DeleteRange
623 : // returns.
624 1 : func (d *DB) DeleteRange(start, end []byte, opts *WriteOptions) error {
625 1 : b := newBatch(d)
626 1 : _ = b.DeleteRange(start, end, opts)
627 1 : if err := d.Apply(b, opts); err != nil {
628 1 : return err
629 1 : }
630 : // Only release the batch on success.
631 1 : return b.Close()
632 : }
633 :
634 : // Merge adds an action to the DB that merges the value at key with the new
635 : // value. The details of the merge are dependent upon the configured merge
636 : // operator.
637 : //
638 : // It is safe to modify the contents of the arguments after Merge returns.
639 1 : func (d *DB) Merge(key, value []byte, opts *WriteOptions) error {
640 1 : b := newBatch(d)
641 1 : _ = b.Merge(key, value, opts)
642 1 : if err := d.Apply(b, opts); err != nil {
643 1 : return err
644 1 : }
645 : // Only release the batch on success.
646 1 : return b.Close()
647 : }
648 :
649 : // LogData adds the specified to the batch. The data will be written to the
650 : // WAL, but not added to memtables or sstables. Log data is never indexed,
651 : // which makes it useful for testing WAL performance.
652 : //
653 : // It is safe to modify the contents of the argument after LogData returns.
654 1 : func (d *DB) LogData(data []byte, opts *WriteOptions) error {
655 1 : b := newBatch(d)
656 1 : _ = b.LogData(data, opts)
657 1 : if err := d.Apply(b, opts); err != nil {
658 1 : return err
659 1 : }
660 : // Only release the batch on success.
661 1 : return b.Close()
662 : }
663 :
664 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
665 : // timestamp suffix to value. The suffix is optional. If any portion of the key
666 : // range [start, end) is already set by a range key with the same suffix value,
667 : // RangeKeySet overrides it.
668 : //
669 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
670 1 : func (d *DB) RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error {
671 1 : b := newBatch(d)
672 1 : _ = b.RangeKeySet(start, end, suffix, value, opts)
673 1 : if err := d.Apply(b, opts); err != nil {
674 0 : return err
675 0 : }
676 : // Only release the batch on success.
677 1 : return b.Close()
678 : }
679 :
680 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
681 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
682 : // range key. RangeKeyUnset only removes portions of range keys that fall within
683 : // the [start, end) key span, and only range keys with suffixes that exactly
684 : // match the unset suffix.
685 : //
686 : // It is safe to modify the contents of the arguments after RangeKeyUnset
687 : // returns.
688 1 : func (d *DB) RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error {
689 1 : b := newBatch(d)
690 1 : _ = b.RangeKeyUnset(start, end, suffix, opts)
691 1 : if err := d.Apply(b, opts); err != nil {
692 0 : return err
693 0 : }
694 : // Only release the batch on success.
695 1 : return b.Close()
696 : }
697 :
698 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
699 : // (inclusive on start, exclusive on end). It does not delete point keys (for
700 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
701 : // bounds, including those with or without suffixes.
702 : //
703 : // It is safe to modify the contents of the arguments after RangeKeyDelete
704 : // returns.
705 1 : func (d *DB) RangeKeyDelete(start, end []byte, opts *WriteOptions) error {
706 1 : b := newBatch(d)
707 1 : _ = b.RangeKeyDelete(start, end, opts)
708 1 : if err := d.Apply(b, opts); err != nil {
709 0 : return err
710 0 : }
711 : // Only release the batch on success.
712 1 : return b.Close()
713 : }
714 :
715 : // Apply the operations contained in the batch to the DB. If the batch is large
716 : // the contents of the batch may be retained by the database. If that occurs
717 : // the batch contents will be cleared preventing the caller from attempting to
718 : // reuse them.
719 : //
720 : // It is safe to modify the contents of the arguments after Apply returns.
721 : //
722 : // Apply returns ErrInvalidBatch if the provided batch is invalid in any way.
723 1 : func (d *DB) Apply(batch *Batch, opts *WriteOptions) error {
724 1 : return d.applyInternal(batch, opts, false)
725 1 : }
726 :
727 : // ApplyNoSyncWait must only be used when opts.Sync is true and the caller
728 : // does not want to wait for the WAL fsync to happen. The method will return
729 : // once the mutation is applied to the memtable and is visible (note that a
730 : // mutation is visible before the WAL sync even in the wait case, so we have
731 : // not weakened the durability semantics). The caller must call Batch.SyncWait
732 : // to wait for the WAL fsync. The caller must not Close the batch without
733 : // first calling Batch.SyncWait.
734 : //
735 : // RECOMMENDATION: Prefer using Apply unless you really understand why you
736 : // need ApplyNoSyncWait.
737 : // EXPERIMENTAL: API/feature subject to change. Do not yet use outside
738 : // CockroachDB.
739 1 : func (d *DB) ApplyNoSyncWait(batch *Batch, opts *WriteOptions) error {
740 1 : if !opts.Sync {
741 0 : return errors.Errorf("cannot request asynchonous apply when WriteOptions.Sync is false")
742 0 : }
743 1 : return d.applyInternal(batch, opts, true)
744 : }
745 :
746 : // REQUIRES: noSyncWait => opts.Sync
747 1 : func (d *DB) applyInternal(batch *Batch, opts *WriteOptions, noSyncWait bool) error {
748 1 : if err := d.closed.Load(); err != nil {
749 1 : panic(err)
750 : }
751 1 : if batch.committing {
752 0 : panic("pebble: batch already committing")
753 : }
754 1 : if batch.applied.Load() {
755 0 : panic("pebble: batch already applied")
756 : }
757 1 : if d.opts.ReadOnly {
758 1 : return ErrReadOnly
759 1 : }
760 1 : if batch.db != nil && batch.db != d {
761 1 : panic(fmt.Sprintf("pebble: batch db mismatch: %p != %p", batch.db, d))
762 : }
763 :
764 1 : sync := opts.GetSync()
765 1 : if sync && d.opts.DisableWAL {
766 0 : return errors.New("pebble: WAL disabled")
767 0 : }
768 :
769 1 : if fmv := d.FormatMajorVersion(); fmv < batch.minimumFormatMajorVersion {
770 0 : panic(fmt.Sprintf(
771 0 : "pebble: batch requires at least format major version %d (current: %d)",
772 0 : batch.minimumFormatMajorVersion, fmv,
773 0 : ))
774 : }
775 :
776 1 : if batch.countRangeKeys > 0 {
777 1 : if d.split == nil {
778 0 : return errNoSplit
779 0 : }
780 : }
781 1 : batch.committing = true
782 1 :
783 1 : if batch.db == nil {
784 1 : if err := batch.refreshMemTableSize(); err != nil {
785 0 : return err
786 0 : }
787 : }
788 1 : if batch.memTableSize >= d.largeBatchThreshold {
789 1 : var err error
790 1 : batch.flushable, err = newFlushableBatch(batch, d.opts.Comparer)
791 1 : if err != nil {
792 0 : return err
793 0 : }
794 : }
795 1 : if err := d.commit.Commit(batch, sync, noSyncWait); err != nil {
796 0 : // There isn't much we can do on an error here. The commit pipeline will be
797 0 : // horked at this point.
798 0 : d.opts.Logger.Fatalf("pebble: fatal commit error: %v", err)
799 0 : }
800 : // If this is a large batch, we need to clear the batch contents as the
801 : // flushable batch may still be present in the flushables queue.
802 : //
803 : // TODO(peter): Currently large batches are written to the WAL. We could
804 : // skip the WAL write and instead wait for the large batch to be flushed to
805 : // an sstable. For a 100 MB batch, this might actually be faster. For a 1
806 : // GB batch this is almost certainly faster.
807 1 : if batch.flushable != nil {
808 1 : batch.data = nil
809 1 : }
810 1 : return nil
811 : }
812 :
813 1 : func (d *DB) commitApply(b *Batch, mem *memTable) error {
814 1 : if b.flushable != nil {
815 1 : // This is a large batch which was already added to the immutable queue.
816 1 : return nil
817 1 : }
818 1 : err := mem.apply(b, b.SeqNum())
819 1 : if err != nil {
820 0 : return err
821 0 : }
822 :
823 : // If the batch contains range tombstones and the database is configured
824 : // to flush range deletions, schedule a delayed flush so that disk space
825 : // may be reclaimed without additional writes or an explicit flush.
826 1 : if b.countRangeDels > 0 && d.opts.FlushDelayDeleteRange > 0 {
827 1 : d.mu.Lock()
828 1 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayDeleteRange)
829 1 : d.mu.Unlock()
830 1 : }
831 :
832 : // If the batch contains range keys and the database is configured to flush
833 : // range keys, schedule a delayed flush so that the range keys are cleared
834 : // from the memtable.
835 1 : if b.countRangeKeys > 0 && d.opts.FlushDelayRangeKey > 0 {
836 1 : d.mu.Lock()
837 1 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayRangeKey)
838 1 : d.mu.Unlock()
839 1 : }
840 :
841 1 : if mem.writerUnref() {
842 1 : d.mu.Lock()
843 1 : d.maybeScheduleFlush()
844 1 : d.mu.Unlock()
845 1 : }
846 1 : return nil
847 : }
848 :
849 1 : func (d *DB) commitWrite(b *Batch, syncWG *sync.WaitGroup, syncErr *error) (*memTable, error) {
850 1 : var size int64
851 1 : repr := b.Repr()
852 1 :
853 1 : if b.flushable != nil {
854 1 : // We have a large batch. Such batches are special in that they don't get
855 1 : // added to the memtable, and are instead inserted into the queue of
856 1 : // memtables. The call to makeRoomForWrite with this batch will force the
857 1 : // current memtable to be flushed. We want the large batch to be part of
858 1 : // the same log, so we add it to the WAL here, rather than after the call
859 1 : // to makeRoomForWrite().
860 1 : //
861 1 : // Set the sequence number since it was not set to the correct value earlier
862 1 : // (see comment in newFlushableBatch()).
863 1 : b.flushable.setSeqNum(b.SeqNum())
864 1 : if !d.opts.DisableWAL {
865 1 : var err error
866 1 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b)
867 1 : if err != nil {
868 0 : panic(err)
869 : }
870 : }
871 : }
872 :
873 1 : var err error
874 1 : // Grab a reference to the memtable. We don't hold DB.mu, but we do hold
875 1 : // d.commit.mu. It's okay for readers of d.mu.mem.mutable to only hold one of
876 1 : // d.commit.mu or d.mu, because memtable rotations require holding both.
877 1 : mem := d.mu.mem.mutable
878 1 : // Batches which contain keys of kind InternalKeyKindIngestSST will
879 1 : // never be applied to the memtable, so we don't need to make room for
880 1 : // write.
881 1 : if !b.ingestedSSTBatch {
882 1 : // Flushable batches will require a rotation of the memtable regardless,
883 1 : // so only attempt an optimistic reservation of space in the current
884 1 : // memtable if this batch is not a large flushable batch.
885 1 : if b.flushable == nil {
886 1 : err = d.mu.mem.mutable.prepare(b)
887 1 : }
888 1 : if b.flushable != nil || err == arenaskl.ErrArenaFull {
889 1 : // Slow path.
890 1 : // We need to acquire DB.mu and rotate the memtable.
891 1 : func() {
892 1 : d.mu.Lock()
893 1 : defer d.mu.Unlock()
894 1 : err = d.makeRoomForWrite(b)
895 1 : mem = d.mu.mem.mutable
896 1 : }()
897 : }
898 : }
899 1 : if err != nil {
900 0 : return nil, err
901 0 : }
902 1 : if d.opts.DisableWAL {
903 1 : return mem, nil
904 1 : }
905 1 : d.logBytesIn.Add(uint64(len(repr)))
906 1 :
907 1 : if b.flushable == nil {
908 1 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b)
909 1 : if err != nil {
910 0 : panic(err)
911 : }
912 : }
913 :
914 1 : d.logSize.Store(uint64(size))
915 1 : return mem, err
916 : }
917 :
918 : type iterAlloc struct {
919 : keyBuf []byte `invariants:"reused"`
920 : boundsBuf [2][]byte `invariants:"reused"`
921 : prefixOrFullSeekKey []byte `invariants:"reused"`
922 : batchState iteratorBatchState
923 : dbi Iterator
924 : merging mergingIter
925 : mlevels [3 + numLevels]mergingIterLevel
926 : levels [3 + numLevels]levelIter
927 : levelsPositioned [3 + numLevels]bool
928 : }
929 :
930 : // maybeAssertZeroed asserts that i is a "zeroed" value. See assertZeroed for
931 : // the definition of "zeroed". It's used to ensure we're properly zeroing out
932 : // memory before returning the iterAlloc to the shared pool.
933 1 : func (i *iterAlloc) maybeAssertZeroed() {
934 1 : if invariants.Enabled {
935 1 : v := reflect.ValueOf(i).Elem()
936 1 : if err := assertZeroed(v); err != nil {
937 0 : panic(err)
938 : }
939 : }
940 : }
941 :
942 : // assertZeroed asserts that v is a "zeroed" value. A "zeroed" value is a value
943 : // that is:
944 : // - the Go zero value for its type, or
945 : // - a pointer to a "zeroed" value, or
946 : // - a slice of len()=0, with any values in the backing array being "zeroed
947 : // values", or
948 : // - a struct with all fields being "zeroed" values,
949 : // - a struct field explicitly marked with a "invariants:reused" tag.
950 1 : func assertZeroed(v reflect.Value) error {
951 1 : if v.IsZero() {
952 1 : return nil
953 1 : }
954 1 : typ := v.Type()
955 1 : switch typ.Kind() {
956 0 : case reflect.Pointer:
957 0 : return assertZeroed(v.Elem())
958 1 : case reflect.Slice:
959 1 : if v.Len() > 0 {
960 0 : return errors.AssertionFailedf("%s is not zeroed (%d len): %#v", typ.Name(), v.Len(), v)
961 0 : }
962 1 : resliced := v.Slice(0, v.Cap())
963 1 : for i := 0; i < resliced.Len(); i++ {
964 1 : if err := assertZeroed(resliced.Index(i)); err != nil {
965 0 : return errors.Wrapf(err, "[%d]", i)
966 0 : }
967 : }
968 1 : return nil
969 1 : case reflect.Struct:
970 1 : for i := 0; i < typ.NumField(); i++ {
971 1 : if typ.Field(i).Tag.Get("invariants") == "reused" {
972 1 : continue
973 : }
974 1 : if err := assertZeroed(v.Field(i)); err != nil {
975 0 : return errors.Wrapf(err, "%q", typ.Field(i).Name)
976 0 : }
977 : }
978 1 : return nil
979 : }
980 0 : return errors.AssertionFailedf("%s (%s) is not zeroed: %#v", typ.Name(), typ.Kind(), v)
981 : }
982 :
983 1 : func newIterAlloc() *iterAlloc {
984 1 : buf := iterAllocPool.Get().(*iterAlloc)
985 1 : buf.maybeAssertZeroed()
986 1 : return buf
987 1 : }
988 :
989 : var iterAllocPool = sync.Pool{
990 1 : New: func() interface{} {
991 1 : return &iterAlloc{}
992 1 : },
993 : }
994 :
995 : // snapshotIterOpts denotes snapshot-related iterator options when calling
996 : // newIter. These are the possible cases for a snapshotIterOpts:
997 : // - No snapshot: All fields are zero values.
998 : // - Classic snapshot: Only `seqNum` is set. The latest readState will be used
999 : // and the specified seqNum will be used as the snapshot seqNum.
1000 : // - EventuallyFileOnlySnapshot (EFOS) behaving as a classic snapshot. Only
1001 : // the `seqNum` is set. The latest readState will be used
1002 : // and the specified seqNum will be used as the snapshot seqNum.
1003 : // - EFOS in file-only state: Only `seqNum` and `vers` are set. All the
1004 : // relevant SSTs are referenced by the *version.
1005 : // - EFOS that has been excised but is in alwaysCreateIters mode (tests only).
1006 : // Only `seqNum` and `readState` are set.
1007 : type snapshotIterOpts struct {
1008 : seqNum base.SeqNum
1009 : vers *manifest.Version
1010 : readState *readState
1011 : }
1012 :
1013 : type batchIterOpts struct {
1014 : batchOnly bool
1015 : }
1016 : type newIterOpts struct {
1017 : snapshot snapshotIterOpts
1018 : batch batchIterOpts
1019 : }
1020 :
1021 : // newIter constructs a new iterator, merging in batch iterators as an extra
1022 : // level.
1023 : func (d *DB) newIter(
1024 : ctx context.Context, batch *Batch, newIterOpts newIterOpts, o *IterOptions,
1025 1 : ) *Iterator {
1026 1 : if newIterOpts.batch.batchOnly {
1027 1 : if batch == nil {
1028 0 : panic("batchOnly is true, but batch is nil")
1029 : }
1030 1 : if newIterOpts.snapshot.vers != nil {
1031 0 : panic("batchOnly is true, but snapshotIterOpts is initialized")
1032 : }
1033 : }
1034 1 : if err := d.closed.Load(); err != nil {
1035 1 : panic(err)
1036 : }
1037 1 : seqNum := newIterOpts.snapshot.seqNum
1038 1 : if o != nil && o.RangeKeyMasking.Suffix != nil && o.KeyTypes != IterKeyTypePointsAndRanges {
1039 0 : panic("pebble: range key masking requires IterKeyTypePointsAndRanges")
1040 : }
1041 1 : if (batch != nil || seqNum != 0) && (o != nil && o.OnlyReadGuaranteedDurable) {
1042 1 : // We could add support for OnlyReadGuaranteedDurable on snapshots if
1043 1 : // there was a need: this would require checking that the sequence number
1044 1 : // of the snapshot has been flushed, by comparing with
1045 1 : // DB.mem.queue[0].logSeqNum.
1046 1 : panic("OnlyReadGuaranteedDurable is not supported for batches or snapshots")
1047 : }
1048 1 : var readState *readState
1049 1 : var newIters tableNewIters
1050 1 : var newIterRangeKey keyspanimpl.TableNewSpanIter
1051 1 : if !newIterOpts.batch.batchOnly {
1052 1 : // Grab and reference the current readState. This prevents the underlying
1053 1 : // files in the associated version from being deleted if there is a current
1054 1 : // compaction. The readState is unref'd by Iterator.Close().
1055 1 : if newIterOpts.snapshot.vers == nil {
1056 1 : if newIterOpts.snapshot.readState != nil {
1057 0 : readState = newIterOpts.snapshot.readState
1058 0 : readState.ref()
1059 1 : } else {
1060 1 : // NB: loadReadState() calls readState.ref().
1061 1 : readState = d.loadReadState()
1062 1 : }
1063 1 : } else {
1064 1 : // vers != nil
1065 1 : newIterOpts.snapshot.vers.Ref()
1066 1 : }
1067 :
1068 : // Determine the seqnum to read at after grabbing the read state (current and
1069 : // memtables) above.
1070 1 : if seqNum == 0 {
1071 1 : seqNum = d.mu.versions.visibleSeqNum.Load()
1072 1 : }
1073 1 : newIters = d.newIters
1074 1 : newIterRangeKey = d.tableNewRangeKeyIter
1075 : }
1076 :
1077 : // Bundle various structures under a single umbrella in order to allocate
1078 : // them together.
1079 1 : buf := newIterAlloc()
1080 1 : dbi := &buf.dbi
1081 1 : dbi.ctx = ctx
1082 1 : dbi.alloc = buf
1083 1 : dbi.merge = d.merge
1084 1 : dbi.comparer = d.opts.Comparer
1085 1 : dbi.readState = readState
1086 1 : dbi.version = newIterOpts.snapshot.vers
1087 1 : dbi.keyBuf = buf.keyBuf
1088 1 : dbi.prefixOrFullSeekKey = buf.prefixOrFullSeekKey
1089 1 : dbi.boundsBuf = buf.boundsBuf
1090 1 : dbi.fc = d.fileCache
1091 1 : dbi.newIters = newIters
1092 1 : dbi.newIterRangeKey = newIterRangeKey
1093 1 : dbi.seqNum = seqNum
1094 1 : dbi.batchOnlyIter = newIterOpts.batch.batchOnly
1095 1 : if o != nil {
1096 1 : dbi.opts = *o
1097 1 : dbi.processBounds(o.LowerBound, o.UpperBound)
1098 1 : }
1099 1 : dbi.opts.logger = d.opts.Logger
1100 1 : if d.opts.private.disableLazyCombinedIteration {
1101 0 : dbi.opts.disableLazyCombinedIteration = true
1102 0 : }
1103 1 : if batch != nil {
1104 1 : dbi.batch = &buf.batchState
1105 1 : dbi.batch.batch = batch
1106 1 : dbi.batch.batchSeqNum = batch.nextSeqNum()
1107 1 : }
1108 1 : return finishInitializingIter(ctx, buf)
1109 : }
1110 :
1111 : // finishInitializingIter is a helper for doing the non-trivial initialization
1112 : // of an Iterator. It's invoked to perform the initial initialization of an
1113 : // Iterator during NewIter or Clone, and to perform reinitialization due to a
1114 : // change in IterOptions by a call to Iterator.SetOptions.
1115 1 : func finishInitializingIter(ctx context.Context, buf *iterAlloc) *Iterator {
1116 1 : // Short-hand.
1117 1 : dbi := &buf.dbi
1118 1 : var memtables flushableList
1119 1 : if dbi.readState != nil {
1120 1 : memtables = dbi.readState.memtables
1121 1 : }
1122 1 : if dbi.opts.OnlyReadGuaranteedDurable {
1123 1 : memtables = nil
1124 1 : } else {
1125 1 : // We only need to read from memtables which contain sequence numbers older
1126 1 : // than seqNum. Trim off newer memtables.
1127 1 : for i := len(memtables) - 1; i >= 0; i-- {
1128 1 : if logSeqNum := memtables[i].logSeqNum; logSeqNum < dbi.seqNum {
1129 1 : break
1130 : }
1131 1 : memtables = memtables[:i]
1132 : }
1133 : }
1134 :
1135 1 : if dbi.opts.pointKeys() {
1136 1 : // Construct the point iterator, initializing dbi.pointIter to point to
1137 1 : // dbi.merging. If this is called during a SetOptions call and this
1138 1 : // Iterator has already initialized dbi.merging, constructPointIter is a
1139 1 : // noop and an initialized pointIter already exists in dbi.pointIter.
1140 1 : dbi.constructPointIter(ctx, memtables, buf)
1141 1 : dbi.iter = dbi.pointIter
1142 1 : } else {
1143 1 : dbi.iter = emptyIter
1144 1 : }
1145 :
1146 1 : if dbi.opts.rangeKeys() {
1147 1 : dbi.rangeKeyMasking.init(dbi, dbi.comparer)
1148 1 :
1149 1 : // When iterating over both point and range keys, don't create the
1150 1 : // range-key iterator stack immediately if we can avoid it. This
1151 1 : // optimization takes advantage of the expected sparseness of range
1152 1 : // keys, and configures the point-key iterator to dynamically switch to
1153 1 : // combined iteration when it observes a file containing range keys.
1154 1 : //
1155 1 : // Lazy combined iteration is not possible if a batch or a memtable
1156 1 : // contains any range keys.
1157 1 : useLazyCombinedIteration := dbi.rangeKey == nil &&
1158 1 : dbi.opts.KeyTypes == IterKeyTypePointsAndRanges &&
1159 1 : (dbi.batch == nil || dbi.batch.batch.countRangeKeys == 0) &&
1160 1 : !dbi.opts.disableLazyCombinedIteration
1161 1 : if useLazyCombinedIteration {
1162 1 : // The user requested combined iteration, and there's no indexed
1163 1 : // batch currently containing range keys that would prevent lazy
1164 1 : // combined iteration. Check the memtables to see if they contain
1165 1 : // any range keys.
1166 1 : for i := range memtables {
1167 1 : if memtables[i].containsRangeKeys() {
1168 1 : useLazyCombinedIteration = false
1169 1 : break
1170 : }
1171 : }
1172 : }
1173 :
1174 1 : if useLazyCombinedIteration {
1175 1 : dbi.lazyCombinedIter = lazyCombinedIter{
1176 1 : parent: dbi,
1177 1 : pointIter: dbi.pointIter,
1178 1 : combinedIterState: combinedIterState{
1179 1 : initialized: false,
1180 1 : },
1181 1 : }
1182 1 : dbi.iter = &dbi.lazyCombinedIter
1183 1 : dbi.iter = invalidating.MaybeWrapIfInvariants(dbi.iter)
1184 1 : } else {
1185 1 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{
1186 1 : initialized: true,
1187 1 : }
1188 1 : if dbi.rangeKey == nil {
1189 1 : dbi.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1190 1 : dbi.constructRangeKeyIter()
1191 1 : } else {
1192 1 : dbi.rangeKey.iterConfig.SetBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
1193 1 : }
1194 :
1195 : // Wrap the point iterator (currently dbi.iter) with an interleaving
1196 : // iterator that interleaves range keys pulled from
1197 : // dbi.rangeKey.rangeKeyIter.
1198 : //
1199 : // NB: The interleaving iterator is always reinitialized, even if
1200 : // dbi already had an initialized range key iterator, in case the point
1201 : // iterator changed or the range key masking suffix changed.
1202 1 : dbi.rangeKey.iiter.Init(dbi.comparer, dbi.iter, dbi.rangeKey.rangeKeyIter,
1203 1 : keyspan.InterleavingIterOpts{
1204 1 : Mask: &dbi.rangeKeyMasking,
1205 1 : LowerBound: dbi.opts.LowerBound,
1206 1 : UpperBound: dbi.opts.UpperBound,
1207 1 : })
1208 1 : dbi.iter = &dbi.rangeKey.iiter
1209 : }
1210 1 : } else {
1211 1 : // !dbi.opts.rangeKeys()
1212 1 : //
1213 1 : // Reset the combined iterator state. The initialized=true ensures the
1214 1 : // iterator doesn't unnecessarily try to switch to combined iteration.
1215 1 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{initialized: true}
1216 1 : }
1217 1 : return dbi
1218 : }
1219 :
1220 : func (i *Iterator) constructPointIter(
1221 : ctx context.Context, memtables flushableList, buf *iterAlloc,
1222 1 : ) {
1223 1 : if i.pointIter != nil {
1224 1 : // Already have one.
1225 1 : return
1226 1 : }
1227 1 : readEnv := block.ReadEnv{
1228 1 : Stats: &i.stats.InternalStats,
1229 1 : // If the file cache has a sstable stats collector, ask it for an
1230 1 : // accumulator for this iterator's configured category and QoS. All SSTable
1231 1 : // iterators created by this Iterator will accumulate their stats to it as
1232 1 : // they Close during iteration.
1233 1 : IterStats: i.fc.SSTStatsCollector().Accumulator(
1234 1 : uint64(uintptr(unsafe.Pointer(i))),
1235 1 : i.opts.Category,
1236 1 : ),
1237 1 : }
1238 1 : if i.readState != nil {
1239 1 : i.blobValueFetcher.Init(&i.readState.current.BlobFiles, i.fc, readEnv,
1240 1 : blob.SuggestedCachedReaders(i.readState.current.MaxReadAmp()))
1241 1 : } else if i.version != nil {
1242 1 : i.blobValueFetcher.Init(&i.version.BlobFiles, i.fc, readEnv,
1243 1 : blob.SuggestedCachedReaders(i.version.MaxReadAmp()))
1244 1 : }
1245 1 : internalOpts := internalIterOpts{
1246 1 : readEnv: sstable.ReadEnv{Block: readEnv},
1247 1 : blobValueFetcher: &i.blobValueFetcher,
1248 1 : }
1249 1 : if i.opts.RangeKeyMasking.Filter != nil {
1250 1 : internalOpts.boundLimitedFilter = &i.rangeKeyMasking
1251 1 : }
1252 :
1253 : // Merging levels and levels from iterAlloc.
1254 1 : mlevels := buf.mlevels[:0]
1255 1 : levels := buf.levels[:0]
1256 1 :
1257 1 : // We compute the number of levels needed ahead of time and reallocate a slice if
1258 1 : // the array from the iterAlloc isn't large enough. Doing this allocation once
1259 1 : // should improve the performance.
1260 1 : numMergingLevels := 0
1261 1 : numLevelIters := 0
1262 1 : if i.batch != nil {
1263 1 : numMergingLevels++
1264 1 : }
1265 :
1266 1 : var current *manifest.Version
1267 1 : if !i.batchOnlyIter {
1268 1 : numMergingLevels += len(memtables)
1269 1 : current = i.version
1270 1 : if current == nil {
1271 1 : current = i.readState.current
1272 1 : }
1273 1 : maxReadAmp := current.MaxReadAmp()
1274 1 : numMergingLevels += maxReadAmp
1275 1 : numLevelIters += maxReadAmp
1276 : }
1277 :
1278 1 : if numMergingLevels > cap(mlevels) {
1279 1 : mlevels = make([]mergingIterLevel, 0, numMergingLevels)
1280 1 : }
1281 1 : if numLevelIters > cap(levels) {
1282 1 : levels = make([]levelIter, 0, numLevelIters)
1283 1 : }
1284 :
1285 : // Top-level is the batch, if any.
1286 1 : if i.batch != nil {
1287 1 : if i.batch.batch.index == nil {
1288 0 : // This isn't an indexed batch. We shouldn't have gotten this far.
1289 0 : panic(errors.AssertionFailedf("creating an iterator over an unindexed batch"))
1290 1 : } else {
1291 1 : i.batch.batch.initInternalIter(&i.opts, &i.batch.pointIter)
1292 1 : i.batch.batch.initRangeDelIter(&i.opts, &i.batch.rangeDelIter, i.batch.batchSeqNum)
1293 1 : // Only include the batch's rangedel iterator if it's non-empty.
1294 1 : // This requires some subtle logic in the case a rangedel is later
1295 1 : // written to the batch and the view of the batch is refreshed
1296 1 : // during a call to SetOptions—in this case, we need to reconstruct
1297 1 : // the point iterator to add the batch rangedel iterator.
1298 1 : var rangeDelIter keyspan.FragmentIterator
1299 1 : if i.batch.rangeDelIter.Count() > 0 {
1300 1 : rangeDelIter = &i.batch.rangeDelIter
1301 1 : }
1302 1 : mlevels = append(mlevels, mergingIterLevel{
1303 1 : iter: &i.batch.pointIter,
1304 1 : rangeDelIter: rangeDelIter,
1305 1 : })
1306 : }
1307 : }
1308 :
1309 1 : if !i.batchOnlyIter {
1310 1 : // Next are the memtables.
1311 1 : for j := len(memtables) - 1; j >= 0; j-- {
1312 1 : mem := memtables[j]
1313 1 : mlevels = append(mlevels, mergingIterLevel{
1314 1 : iter: mem.newIter(&i.opts),
1315 1 : rangeDelIter: mem.newRangeDelIter(&i.opts),
1316 1 : })
1317 1 : }
1318 :
1319 : // Next are the file levels: L0 sub-levels followed by lower levels.
1320 1 : mlevelsIndex := len(mlevels)
1321 1 : levelsIndex := len(levels)
1322 1 : mlevels = mlevels[:numMergingLevels]
1323 1 : levels = levels[:numLevelIters]
1324 1 : i.opts.snapshotForHideObsoletePoints = buf.dbi.seqNum
1325 1 : addLevelIterForFiles := func(files manifest.LevelIterator, level manifest.Layer) {
1326 1 : li := &levels[levelsIndex]
1327 1 :
1328 1 : li.init(ctx, i.opts, i.comparer, i.newIters, files, level, internalOpts)
1329 1 : li.initRangeDel(&mlevels[mlevelsIndex])
1330 1 : li.initCombinedIterState(&i.lazyCombinedIter.combinedIterState)
1331 1 : mlevels[mlevelsIndex].levelIter = li
1332 1 : mlevels[mlevelsIndex].iter = invalidating.MaybeWrapIfInvariants(li)
1333 1 :
1334 1 : levelsIndex++
1335 1 : mlevelsIndex++
1336 1 : }
1337 :
1338 : // Add level iterators for the L0 sublevels, iterating from newest to
1339 : // oldest.
1340 1 : for i := len(current.L0SublevelFiles) - 1; i >= 0; i-- {
1341 1 : addLevelIterForFiles(current.L0SublevelFiles[i].Iter(), manifest.L0Sublevel(i))
1342 1 : }
1343 :
1344 : // Add level iterators for the non-empty non-L0 levels.
1345 1 : for level := 1; level < len(current.Levels); level++ {
1346 1 : if current.Levels[level].Empty() {
1347 1 : continue
1348 : }
1349 1 : addLevelIterForFiles(current.Levels[level].Iter(), manifest.Level(level))
1350 : }
1351 : }
1352 1 : buf.merging.init(&i.opts, &i.stats.InternalStats, i.comparer.Compare, i.comparer.Split, mlevels...)
1353 1 : if len(mlevels) <= cap(buf.levelsPositioned) {
1354 1 : buf.merging.levelsPositioned = buf.levelsPositioned[:len(mlevels)]
1355 1 : }
1356 1 : buf.merging.snapshot = i.seqNum
1357 1 : if i.batch != nil {
1358 1 : buf.merging.batchSnapshot = i.batch.batchSeqNum
1359 1 : }
1360 1 : buf.merging.combinedIterState = &i.lazyCombinedIter.combinedIterState
1361 1 : i.pointIter = invalidating.MaybeWrapIfInvariants(&buf.merging).(topLevelIterator)
1362 1 : i.merging = &buf.merging
1363 : }
1364 :
1365 : // NewBatch returns a new empty write-only batch. Any reads on the batch will
1366 : // return an error. If the batch is committed it will be applied to the DB.
1367 1 : func (d *DB) NewBatch(opts ...BatchOption) *Batch {
1368 1 : return newBatch(d, opts...)
1369 1 : }
1370 :
1371 : // NewBatchWithSize is mostly identical to NewBatch, but it will allocate the
1372 : // the specified memory space for the internal slice in advance.
1373 0 : func (d *DB) NewBatchWithSize(size int, opts ...BatchOption) *Batch {
1374 0 : return newBatchWithSize(d, size, opts...)
1375 0 : }
1376 :
1377 : // NewIndexedBatch returns a new empty read-write batch. Any reads on the batch
1378 : // will read from both the batch and the DB. If the batch is committed it will
1379 : // be applied to the DB. An indexed batch is slower that a non-indexed batch
1380 : // for insert operations. If you do not need to perform reads on the batch, use
1381 : // NewBatch instead.
1382 1 : func (d *DB) NewIndexedBatch() *Batch {
1383 1 : return newIndexedBatch(d, d.opts.Comparer)
1384 1 : }
1385 :
1386 : // NewIndexedBatchWithSize is mostly identical to NewIndexedBatch, but it will
1387 : // allocate the specified memory space for the internal slice in advance.
1388 0 : func (d *DB) NewIndexedBatchWithSize(size int) *Batch {
1389 0 : return newIndexedBatchWithSize(d, d.opts.Comparer, size)
1390 0 : }
1391 :
1392 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
1393 : // return false). The iterator can be positioned via a call to SeekGE, SeekLT,
1394 : // First or Last. The iterator provides a point-in-time view of the current DB
1395 : // state. This view is maintained by preventing file deletions and preventing
1396 : // memtables referenced by the iterator from being deleted. Using an iterator
1397 : // to maintain a long-lived point-in-time view of the DB state can lead to an
1398 : // apparent memory and disk usage leak. Use snapshots (see NewSnapshot) for
1399 : // point-in-time snapshots which avoids these problems.
1400 1 : func (d *DB) NewIter(o *IterOptions) (*Iterator, error) {
1401 1 : return d.NewIterWithContext(context.Background(), o)
1402 1 : }
1403 :
1404 : // NewIterWithContext is like NewIter, and additionally accepts a context for
1405 : // tracing.
1406 1 : func (d *DB) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
1407 1 : return d.newIter(ctx, nil /* batch */, newIterOpts{}, o), nil
1408 1 : }
1409 :
1410 : // NewSnapshot returns a point-in-time view of the current DB state. Iterators
1411 : // created with this handle will all observe a stable snapshot of the current
1412 : // DB state. The caller must call Snapshot.Close() when the snapshot is no
1413 : // longer needed. Snapshots are not persisted across DB restarts (close ->
1414 : // open). Unlike the implicit snapshot maintained by an iterator, a snapshot
1415 : // will not prevent memtables from being released or sstables from being
1416 : // deleted. Instead, a snapshot prevents deletion of sequence numbers
1417 : // referenced by the snapshot.
1418 : //
1419 : // There exists one violation of a Snapshot's point-in-time guarantee: An excise
1420 : // (see DB.Excise and DB.IngestAndExcise) that occurs after the snapshot's
1421 : // creation will be observed by iterators created from the snapshot after the
1422 : // excise. See NewEventuallyFileOnlySnapshot for a variant of NewSnapshot that
1423 : // provides a full point-in-time guarantee.
1424 1 : func (d *DB) NewSnapshot() *Snapshot {
1425 1 : // TODO(jackson): Consider removal of regular, non-eventually-file-only
1426 1 : // snapshots given they no longer provide a true point-in-time snapshot of
1427 1 : // the database due to excises. If we had a mechanism to construct a maximal
1428 1 : // key range, we could implement NewSnapshot in terms of
1429 1 : // NewEventuallyFileOnlySnapshot and provide a true point-in-time guarantee.
1430 1 : if err := d.closed.Load(); err != nil {
1431 1 : panic(err)
1432 : }
1433 1 : d.mu.Lock()
1434 1 : s := &Snapshot{
1435 1 : db: d,
1436 1 : seqNum: d.mu.versions.visibleSeqNum.Load(),
1437 1 : }
1438 1 : d.mu.snapshots.pushBack(s)
1439 1 : d.mu.Unlock()
1440 1 : return s
1441 : }
1442 :
1443 : // NewEventuallyFileOnlySnapshot returns a point-in-time view of the current DB
1444 : // state, similar to NewSnapshot, but with consistency constrained to the
1445 : // provided set of key ranges. See the comment at EventuallyFileOnlySnapshot for
1446 : // its semantics.
1447 1 : func (d *DB) NewEventuallyFileOnlySnapshot(keyRanges []KeyRange) *EventuallyFileOnlySnapshot {
1448 1 : if err := d.closed.Load(); err != nil {
1449 0 : panic(err)
1450 : }
1451 1 : for i := range keyRanges {
1452 1 : if i > 0 && d.cmp(keyRanges[i-1].End, keyRanges[i].Start) > 0 {
1453 0 : panic("pebble: key ranges for eventually-file-only-snapshot not in order")
1454 : }
1455 : }
1456 1 : return d.makeEventuallyFileOnlySnapshot(keyRanges)
1457 : }
1458 :
1459 : // Close closes the DB.
1460 : //
1461 : // It is not safe to close a DB until all outstanding iterators are closed
1462 : // or to call Close concurrently with any other DB method. It is not valid
1463 : // to call any of a DB's methods after the DB has been closed.
1464 1 : func (d *DB) Close() error {
1465 1 : if err := d.closed.Load(); err != nil {
1466 1 : panic(err)
1467 : }
1468 1 : d.compactionSchedulers.Wait()
1469 1 : // Compactions can be asynchronously started by the CompactionScheduler
1470 1 : // calling d.Schedule. When this Unregister returns, we know that the
1471 1 : // CompactionScheduler will never again call a method on the DB. Note that
1472 1 : // this must be called without holding d.mu.
1473 1 : d.compactionScheduler.Unregister()
1474 1 : // Lock the commit pipeline for the duration of Close. This prevents a race
1475 1 : // with makeRoomForWrite. Rotating the WAL in makeRoomForWrite requires
1476 1 : // dropping d.mu several times for I/O. If Close only holds d.mu, an
1477 1 : // in-progress WAL rotation may re-acquire d.mu only once the database is
1478 1 : // closed.
1479 1 : //
1480 1 : // Additionally, locking the commit pipeline makes it more likely that
1481 1 : // (illegal) concurrent writes will observe d.closed.Load() != nil, creating
1482 1 : // more understable panics if the database is improperly used concurrently
1483 1 : // during Close.
1484 1 : d.commit.mu.Lock()
1485 1 : defer d.commit.mu.Unlock()
1486 1 : d.mu.Lock()
1487 1 : defer d.mu.Unlock()
1488 1 : // Check that the DB is not closed again. If there are two concurrent calls
1489 1 : // to DB.Close, the best-effort check at the top of DB.Close may not fire.
1490 1 : // But since this second check happens after mutex acquisition, the two
1491 1 : // concurrent calls will get serialized and the second one will see the
1492 1 : // effect of the d.closed.Store below.
1493 1 : if err := d.closed.Load(); err != nil {
1494 0 : panic(err)
1495 : }
1496 : // Clear the finalizer that is used to check that an unreferenced DB has been
1497 : // closed. We're closing the DB here, so the check performed by that
1498 : // finalizer isn't necessary.
1499 : //
1500 : // Note: this is a no-op if invariants are disabled or race is enabled.
1501 1 : invariants.SetFinalizer(d.closed, nil)
1502 1 :
1503 1 : d.closed.Store(errors.WithStack(ErrClosed))
1504 1 : close(d.closedCh)
1505 1 :
1506 1 : defer d.cacheHandle.Close()
1507 1 :
1508 1 : for d.mu.compact.compactingCount > 0 || d.mu.compact.downloadingCount > 0 || d.mu.compact.flushing {
1509 1 : d.mu.compact.cond.Wait()
1510 1 : }
1511 1 : for d.mu.tableStats.loading {
1512 1 : d.mu.tableStats.cond.Wait()
1513 1 : }
1514 1 : for d.mu.tableValidation.validating {
1515 0 : d.mu.tableValidation.cond.Wait()
1516 0 : }
1517 :
1518 1 : var err error
1519 1 : if n := len(d.mu.compact.inProgress); n > 0 {
1520 1 : err = errors.Errorf("pebble: %d unexpected in-progress compactions", errors.Safe(n))
1521 1 : }
1522 1 : err = firstError(err, d.mu.formatVers.marker.Close())
1523 1 : if !d.opts.ReadOnly {
1524 1 : if d.mu.log.writer != nil {
1525 1 : _, err2 := d.mu.log.writer.Close()
1526 1 : err = firstError(err, err2)
1527 1 : }
1528 1 : } else if d.mu.log.writer != nil {
1529 0 : panic("pebble: log-writer should be nil in read-only mode")
1530 : }
1531 1 : err = firstError(err, d.mu.log.manager.Close())
1532 1 : err = firstError(err, d.dataDirLock.Close())
1533 1 :
1534 1 : // Note that versionSet.close() only closes the MANIFEST. The versions list
1535 1 : // is still valid for the checks below.
1536 1 : err = firstError(err, d.mu.versions.close())
1537 1 :
1538 1 : err = firstError(err, d.dataDir.Close())
1539 1 :
1540 1 : d.readState.val.unrefLocked()
1541 1 :
1542 1 : current := d.mu.versions.currentVersion()
1543 1 : for v := d.mu.versions.versions.Front(); true; v = v.Next() {
1544 1 : refs := v.Refs()
1545 1 : if v == current {
1546 1 : if refs != 1 {
1547 1 : err = firstError(err, errors.Errorf("leaked iterators: current\n%s", v))
1548 1 : }
1549 1 : break
1550 : }
1551 0 : if refs != 0 {
1552 0 : err = firstError(err, errors.Errorf("leaked iterators:\n%s", v))
1553 0 : }
1554 : }
1555 :
1556 1 : for _, mem := range d.mu.mem.queue {
1557 1 : // Usually, we'd want to delete the files returned by readerUnref. But
1558 1 : // in this case, even if we're unreferencing the flushables, the
1559 1 : // flushables aren't obsolete. They will be reconstructed during WAL
1560 1 : // replay.
1561 1 : mem.readerUnrefLocked(false)
1562 1 : }
1563 : // If there's an unused, recycled memtable, we need to release its memory.
1564 1 : if obsoleteMemTable := d.memTableRecycle.Swap(nil); obsoleteMemTable != nil {
1565 1 : d.freeMemTable(obsoleteMemTable)
1566 1 : }
1567 1 : if reserved := d.memTableReserved.Load(); reserved != 0 {
1568 1 : err = firstError(err, errors.Errorf("leaked memtable reservation: %d", errors.Safe(reserved)))
1569 1 : }
1570 :
1571 : // Since we called d.readState.val.unrefLocked() above, we are expected to
1572 : // manually schedule deletion of obsolete files.
1573 1 : if len(d.mu.versions.obsoleteTables) > 0 || len(d.mu.versions.obsoleteBlobs) > 0 {
1574 1 : d.deleteObsoleteFiles(d.newJobIDLocked())
1575 1 : }
1576 :
1577 1 : d.mu.Unlock()
1578 1 :
1579 1 : // Wait for all cleaning jobs to finish.
1580 1 : d.cleanupManager.Close()
1581 1 :
1582 1 : d.mu.Lock()
1583 1 : // Sanity check compaction metrics.
1584 1 : if invariants.Enabled {
1585 1 : if d.mu.compact.compactingCount > 0 || d.mu.compact.downloadingCount > 0 || d.mu.versions.atomicInProgressBytes.Load() > 0 {
1586 0 : panic("compacting counts not 0 on close")
1587 : }
1588 : }
1589 :
1590 : // As a sanity check, ensure that there are no zombie tables or blob files.
1591 : // A non-zero count hints at a reference count leak.
1592 1 : if ztbls := d.mu.versions.zombieTables.Count(); ztbls > 0 {
1593 0 : err = firstError(err, errors.Errorf("non-zero zombie file count: %d", ztbls))
1594 0 : }
1595 1 : if zblobs := d.mu.versions.zombieBlobs.Count(); zblobs > 0 {
1596 0 : err = firstError(err, errors.Errorf("non-zero zombie blob count: %d", zblobs))
1597 0 : }
1598 :
1599 1 : err = firstError(err, d.fileCache.Close())
1600 1 :
1601 1 : err = firstError(err, d.objProvider.Close())
1602 1 :
1603 1 : // If the options include a closer to 'close' the filesystem, close it.
1604 1 : if d.opts.private.fsCloser != nil {
1605 1 : d.opts.private.fsCloser.Close()
1606 1 : }
1607 :
1608 : // Return an error if the user failed to close all open snapshots.
1609 1 : if v := d.mu.snapshots.count(); v > 0 {
1610 0 : err = firstError(err, errors.Errorf("leaked snapshots: %d open snapshots on DB %p", v, d))
1611 0 : }
1612 :
1613 1 : return err
1614 : }
1615 :
1616 : // Compact the specified range of keys in the database.
1617 1 : func (d *DB) Compact(ctx context.Context, start, end []byte, parallelize bool) error {
1618 1 : if err := d.closed.Load(); err != nil {
1619 1 : panic(err)
1620 : }
1621 1 : if d.opts.ReadOnly {
1622 1 : return ErrReadOnly
1623 1 : }
1624 1 : if d.cmp(start, end) >= 0 {
1625 1 : return errors.Errorf("Compact start %s is not less than end %s",
1626 1 : d.opts.Comparer.FormatKey(start), d.opts.Comparer.FormatKey(end))
1627 1 : }
1628 :
1629 1 : d.mu.Lock()
1630 1 : maxLevelWithFiles := 1
1631 1 : cur := d.mu.versions.currentVersion()
1632 1 : for level := 0; level < numLevels; level++ {
1633 1 : overlaps := cur.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1634 1 : if !overlaps.Empty() {
1635 1 : maxLevelWithFiles = level + 1
1636 1 : }
1637 : }
1638 :
1639 : // Determine if any memtable overlaps with the compaction range. We wait for
1640 : // any such overlap to flush (initiating a flush if necessary).
1641 1 : mem, err := func() (*flushableEntry, error) {
1642 1 : // Check to see if any files overlap with any of the memtables. The queue
1643 1 : // is ordered from oldest to newest with the mutable memtable being the
1644 1 : // last element in the slice. We want to wait for the newest table that
1645 1 : // overlaps.
1646 1 : for i := len(d.mu.mem.queue) - 1; i >= 0; i-- {
1647 1 : mem := d.mu.mem.queue[i]
1648 1 : var anyOverlaps bool
1649 1 : mem.computePossibleOverlaps(func(b bounded) shouldContinue {
1650 1 : anyOverlaps = true
1651 1 : return stopIteration
1652 1 : }, KeyRange{Start: start, End: end})
1653 1 : if !anyOverlaps {
1654 1 : continue
1655 : }
1656 1 : var err error
1657 1 : if mem.flushable == d.mu.mem.mutable {
1658 1 : // We have to hold both commitPipeline.mu and DB.mu when calling
1659 1 : // makeRoomForWrite(). Lock order requirements elsewhere force us to
1660 1 : // unlock DB.mu in order to grab commitPipeline.mu first.
1661 1 : d.mu.Unlock()
1662 1 : d.commit.mu.Lock()
1663 1 : d.mu.Lock()
1664 1 : defer d.commit.mu.Unlock() //nolint:deferloop
1665 1 : if mem.flushable == d.mu.mem.mutable {
1666 1 : // Only flush if the active memtable is unchanged.
1667 1 : err = d.makeRoomForWrite(nil)
1668 1 : }
1669 : }
1670 1 : mem.flushForced = true
1671 1 : d.maybeScheduleFlush()
1672 1 : return mem, err
1673 : }
1674 1 : return nil, nil
1675 : }()
1676 :
1677 1 : d.mu.Unlock()
1678 1 :
1679 1 : if err != nil {
1680 0 : return err
1681 0 : }
1682 1 : if mem != nil {
1683 1 : select {
1684 1 : case <-mem.flushed:
1685 0 : case <-ctx.Done():
1686 0 : return ctx.Err()
1687 : }
1688 : }
1689 :
1690 1 : for level := 0; level < maxLevelWithFiles; {
1691 1 : for {
1692 1 : if err := d.manualCompact(
1693 1 : ctx, start, end, level, parallelize); err != nil {
1694 1 : if errors.Is(err, ErrCancelledCompaction) {
1695 1 : continue
1696 : }
1697 1 : return err
1698 : }
1699 1 : break
1700 : }
1701 1 : level++
1702 1 : if level == numLevels-1 {
1703 1 : // A manual compaction of the bottommost level occurred.
1704 1 : // There is no next level to try and compact.
1705 1 : break
1706 : }
1707 : }
1708 1 : return nil
1709 : }
1710 :
1711 : func (d *DB) manualCompact(
1712 : ctx context.Context, start, end []byte, level int, parallelize bool,
1713 1 : ) error {
1714 1 : d.mu.Lock()
1715 1 : curr := d.mu.versions.currentVersion()
1716 1 : files := curr.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1717 1 : if files.Empty() {
1718 1 : d.mu.Unlock()
1719 1 : return nil
1720 1 : }
1721 :
1722 1 : var compactions []*manualCompaction
1723 1 : if parallelize {
1724 1 : compactions = append(compactions, d.splitManualCompaction(start, end, level)...)
1725 1 : } else {
1726 1 : compactions = append(compactions, &manualCompaction{
1727 1 : level: level,
1728 1 : done: make(chan error, 1),
1729 1 : start: start,
1730 1 : end: end,
1731 1 : })
1732 1 : }
1733 1 : n := len(compactions)
1734 1 : if n == 0 {
1735 0 : d.mu.Unlock()
1736 0 : return nil
1737 0 : }
1738 1 : for i := range compactions {
1739 1 : d.mu.compact.manualID++
1740 1 : compactions[i].id = d.mu.compact.manualID
1741 1 : }
1742 : // [manualIDStart, manualIDEnd] are the compactions that have been added to
1743 : // d.mu.compact.manual.
1744 1 : manualIDStart := compactions[0].id
1745 1 : manualIDEnd := compactions[n-1].id
1746 1 : d.mu.compact.manual = append(d.mu.compact.manual, compactions...)
1747 1 : d.mu.compact.manualLen.Store(int32(len(d.mu.compact.manual)))
1748 1 : d.maybeScheduleCompaction()
1749 1 : d.mu.Unlock()
1750 1 :
1751 1 : // On context cancellation, we only cancel the compactions that have not yet
1752 1 : // started. The assumption is that it is relatively harmless to have the
1753 1 : // already started compactions run to completion. We don't wait for the
1754 1 : // ongoing compactions to finish, since the assumption is that the caller
1755 1 : // has already given up on the operation (and the cancellation error is
1756 1 : // going to be returned anyway).
1757 1 : //
1758 1 : // An alternative would be to store the context in each *manualCompaction,
1759 1 : // and have the goroutine that retrieves the *manualCompaction for running
1760 1 : // notice the cancellation and write the cancellation error to
1761 1 : // manualCompaction.done. That approach would require this method to wait
1762 1 : // for all the *manualCompactions it has enqueued to finish before returning
1763 1 : // (to not leak a context). Since there is no timeliness guarantee on when a
1764 1 : // *manualCompaction will be retrieved for running, the wait until a
1765 1 : // cancelled context causes this method to return is not bounded. Hence, we
1766 1 : // don't adopt that approach.
1767 1 : cancelPendingCompactions := func() {
1768 1 : d.mu.Lock()
1769 1 : for i := 0; i < len(d.mu.compact.manual); {
1770 1 : if d.mu.compact.manual[i].id >= manualIDStart && d.mu.compact.manual[i].id <= manualIDEnd {
1771 1 : d.mu.compact.manual = slices.Delete(d.mu.compact.manual, i, i+1)
1772 1 : d.mu.compact.manualLen.Store(int32(len(d.mu.compact.manual)))
1773 1 : } else {
1774 0 : i++
1775 0 : }
1776 : }
1777 1 : d.mu.Unlock()
1778 : }
1779 : // Each of the channels is guaranteed to be eventually sent to once. After a
1780 : // compaction is possibly picked in d.maybeScheduleCompaction(), either the
1781 : // compaction is dropped, executed after being scheduled, or retried later.
1782 : // Assuming eventual progress when a compaction is retried, all outcomes send
1783 : // a value to the done channel. Since the channels are buffered, it is not
1784 : // necessary to read from each channel, and so we can exit early in the event
1785 : // of an error.
1786 1 : for _, compaction := range compactions {
1787 1 : select {
1788 1 : case <-ctx.Done():
1789 1 : cancelPendingCompactions()
1790 1 : return ctx.Err()
1791 1 : case err := <-compaction.done:
1792 1 : if err != nil {
1793 1 : cancelPendingCompactions()
1794 1 : return err
1795 1 : }
1796 : }
1797 : }
1798 1 : return nil
1799 : }
1800 :
1801 : // splitManualCompaction splits a manual compaction over [start,end] on level
1802 : // such that the resulting compactions have no key overlap.
1803 : func (d *DB) splitManualCompaction(
1804 : start, end []byte, level int,
1805 1 : ) (splitCompactions []*manualCompaction) {
1806 1 : curr := d.mu.versions.currentVersion()
1807 1 : endLevel := level + 1
1808 1 : baseLevel := d.mu.versions.picker.getBaseLevel()
1809 1 : if level == 0 {
1810 1 : endLevel = baseLevel
1811 1 : }
1812 1 : keyRanges := curr.CalculateInuseKeyRanges(d.mu.versions.latest.l0Organizer, level, endLevel, start, end)
1813 1 : for _, keyRange := range keyRanges {
1814 1 : splitCompactions = append(splitCompactions, &manualCompaction{
1815 1 : level: level,
1816 1 : done: make(chan error, 1),
1817 1 : start: keyRange.Start,
1818 1 : end: keyRange.End.Key,
1819 1 : split: true,
1820 1 : })
1821 1 : }
1822 1 : return splitCompactions
1823 : }
1824 :
1825 : // Flush the memtable to stable storage.
1826 1 : func (d *DB) Flush() error {
1827 1 : flushDone, err := d.AsyncFlush()
1828 1 : if err != nil {
1829 1 : return err
1830 1 : }
1831 1 : <-flushDone
1832 1 : return nil
1833 : }
1834 :
1835 : // AsyncFlush asynchronously flushes the memtable to stable storage.
1836 : //
1837 : // If no error is returned, the caller can receive from the returned channel in
1838 : // order to wait for the flush to complete.
1839 1 : func (d *DB) AsyncFlush() (<-chan struct{}, error) {
1840 1 : if err := d.closed.Load(); err != nil {
1841 1 : panic(err)
1842 : }
1843 1 : if d.opts.ReadOnly {
1844 1 : return nil, ErrReadOnly
1845 1 : }
1846 :
1847 1 : d.commit.mu.Lock()
1848 1 : defer d.commit.mu.Unlock()
1849 1 : d.mu.Lock()
1850 1 : defer d.mu.Unlock()
1851 1 : flushed := d.mu.mem.queue[len(d.mu.mem.queue)-1].flushed
1852 1 : err := d.makeRoomForWrite(nil)
1853 1 : if err != nil {
1854 0 : return nil, err
1855 0 : }
1856 1 : return flushed, nil
1857 : }
1858 :
1859 : // Metrics returns metrics about the database.
1860 1 : func (d *DB) Metrics() *Metrics {
1861 1 : metrics := &Metrics{}
1862 1 : walStats := d.mu.log.manager.Stats()
1863 1 : completedObsoleteFileStats := d.cleanupManager.CompletedStats()
1864 1 :
1865 1 : d.mu.Lock()
1866 1 : vers := d.mu.versions.currentVersion()
1867 1 : vers.Ref()
1868 1 : defer vers.Unref()
1869 1 :
1870 1 : *metrics = d.mu.versions.metrics
1871 1 : metrics.Compact.EstimatedDebt = d.mu.versions.picker.estimatedCompactionDebt()
1872 1 : metrics.Compact.InProgressBytes = d.mu.versions.atomicInProgressBytes.Load()
1873 1 : // TODO(radu): split this to separate the download compactions.
1874 1 : metrics.Compact.NumInProgress = int64(d.mu.compact.compactingCount + d.mu.compact.downloadingCount)
1875 1 : metrics.Compact.MarkedFiles = vers.MarkedForCompaction.Count()
1876 1 : metrics.Compact.Duration = d.mu.compact.duration
1877 1 : for c := range d.mu.compact.inProgress {
1878 1 : if !c.IsFlush() {
1879 1 : metrics.Compact.Duration += d.timeNow().Sub(c.BeganAt())
1880 1 : }
1881 : }
1882 1 : metrics.Compact.NumProblemSpans = d.problemSpans.Len()
1883 1 :
1884 1 : for _, m := range d.mu.mem.queue {
1885 1 : metrics.MemTable.Size += m.totalBytes()
1886 1 : }
1887 1 : metrics.Snapshots.Count = d.mu.snapshots.count()
1888 1 : if metrics.Snapshots.Count > 0 {
1889 0 : metrics.Snapshots.EarliestSeqNum = d.mu.snapshots.earliest()
1890 0 : }
1891 1 : metrics.Snapshots.PinnedKeys = d.mu.snapshots.cumulativePinnedCount
1892 1 : metrics.Snapshots.PinnedSize = d.mu.snapshots.cumulativePinnedSize
1893 1 : metrics.MemTable.Count = int64(len(d.mu.mem.queue))
1894 1 : metrics.MemTable.ZombieCount = d.memTableCount.Load() - metrics.MemTable.Count
1895 1 : metrics.MemTable.ZombieSize = uint64(d.memTableReserved.Load()) - metrics.MemTable.Size
1896 1 : metrics.WAL.ObsoleteFiles = int64(walStats.ObsoleteFileCount)
1897 1 : metrics.WAL.ObsoletePhysicalSize = walStats.ObsoleteFileSize
1898 1 : metrics.WAL.Files = int64(walStats.LiveFileCount)
1899 1 : // The current WAL's size (d.logSize) is the logical size, which may be less
1900 1 : // than the WAL's physical size if it was recycled. walStats.LiveFileSize
1901 1 : // includes the physical size of all live WALs, but for the current WAL it
1902 1 : // reflects the physical size when it was opened. So it is possible that
1903 1 : // d.atomic.logSize has exceeded that physical size. We allow for this
1904 1 : // anomaly.
1905 1 : metrics.WAL.PhysicalSize = walStats.LiveFileSize
1906 1 : metrics.WAL.BytesIn = d.logBytesIn.Load()
1907 1 : metrics.WAL.Size = d.logSize.Load()
1908 1 : for i, n := 0, len(d.mu.mem.queue)-1; i < n; i++ {
1909 1 : metrics.WAL.Size += d.mu.mem.queue[i].logSize
1910 1 : }
1911 1 : metrics.WAL.BytesWritten = metrics.Levels[0].TableBytesIn + metrics.WAL.Size
1912 1 : metrics.WAL.Failover = walStats.Failover
1913 1 :
1914 1 : if p := d.mu.versions.picker; p != nil {
1915 1 : compactions := d.getInProgressCompactionInfoLocked(nil)
1916 1 : m := p.getMetrics(compactions)
1917 1 : for level, lm := range m.levels {
1918 1 : metrics.Levels[level].Score = lm.score
1919 1 : metrics.Levels[level].FillFactor = lm.fillFactor
1920 1 : metrics.Levels[level].CompensatedFillFactor = lm.compensatedFillFactor
1921 1 : }
1922 : }
1923 1 : metrics.Table.ZombieCount = int64(d.mu.versions.zombieTables.Count())
1924 1 : metrics.Table.ZombieSize = d.mu.versions.zombieTables.TotalSize()
1925 1 : metrics.Table.Local.ZombieCount, metrics.Table.Local.ZombieSize = d.mu.versions.zombieTables.LocalStats()
1926 1 :
1927 1 : // The obsolete blob/table metrics have a subtle calculation:
1928 1 : //
1929 1 : // (A) The vs.metrics.{Table,BlobFiles}.[Local.]{ObsoleteCount,ObsoleteSize}
1930 1 : // fields reflect the set of files currently sitting in
1931 1 : // vs.obsolete{Tables,Blobs} but not yet enqueued to the cleanup manager.
1932 1 : //
1933 1 : // (B) The d.mu.fileDeletions.queuedStats field holds the set of files that have
1934 1 : // been queued for deletion by the cleanup manager.
1935 1 : //
1936 1 : // (C) The cleanup manager also maintains cumulative stats for the set of
1937 1 : // files that have been deleted.
1938 1 : //
1939 1 : // The value of currently pending obsolete files is (A) + (B) - (C).
1940 1 : pendingObsoleteFileStats := d.mu.fileDeletions.queuedStats
1941 1 : pendingObsoleteFileStats.Sub(completedObsoleteFileStats)
1942 1 : metrics.Table.Local.ObsoleteCount += pendingObsoleteFileStats.tablesLocal.count
1943 1 : metrics.Table.Local.ObsoleteSize += pendingObsoleteFileStats.tablesLocal.size
1944 1 : metrics.Table.ObsoleteCount += int64(pendingObsoleteFileStats.tablesAll.count)
1945 1 : metrics.Table.ObsoleteSize += pendingObsoleteFileStats.tablesAll.size
1946 1 : metrics.BlobFiles.Local.ObsoleteCount += pendingObsoleteFileStats.blobFilesLocal.count
1947 1 : metrics.BlobFiles.Local.ObsoleteSize += pendingObsoleteFileStats.blobFilesLocal.size
1948 1 : metrics.BlobFiles.ObsoleteCount += pendingObsoleteFileStats.blobFilesAll.count
1949 1 : metrics.BlobFiles.ObsoleteSize += pendingObsoleteFileStats.blobFilesAll.size
1950 1 : metrics.private.optionsFileSize = d.optionsFileSize
1951 1 :
1952 1 : d.mu.versions.logLock()
1953 1 : metrics.private.manifestFileSize = uint64(d.mu.versions.manifest.Size())
1954 1 : backingCount, backingTotalSize := d.mu.versions.latest.virtualBackings.Stats()
1955 1 : metrics.Table.BackingTableCount = uint64(backingCount)
1956 1 : metrics.Table.BackingTableSize = backingTotalSize
1957 1 : blobStats, _ := d.mu.versions.latest.blobFiles.Stats()
1958 1 : d.mu.versions.logUnlock()
1959 1 : metrics.BlobFiles.LiveCount = blobStats.Count
1960 1 : metrics.BlobFiles.LiveSize = blobStats.PhysicalSize
1961 1 : metrics.BlobFiles.ValueSize = blobStats.ValueSize
1962 1 : metrics.BlobFiles.ReferencedValueSize = blobStats.ReferencedValueSize
1963 1 : metrics.BlobFiles.ReferencedBackingValueSize = blobStats.ReferencedBackingValueSize
1964 1 :
1965 1 : metrics.LogWriter.FsyncLatency = d.mu.log.metrics.fsyncLatency
1966 1 : if err := metrics.LogWriter.Merge(&d.mu.log.metrics.LogWriterMetrics); err != nil {
1967 0 : d.opts.Logger.Errorf("metrics error: %s", err)
1968 0 : }
1969 1 : metrics.Flush.WriteThroughput = d.mu.compact.flushWriteThroughput
1970 1 : if d.mu.compact.flushing {
1971 1 : metrics.Flush.NumInProgress = 1
1972 1 : }
1973 :
1974 1 : metrics.Table.PendingStatsCollectionCount = int64(len(d.mu.tableStats.pending))
1975 1 : metrics.Table.InitialStatsCollectionComplete = d.mu.tableStats.loadedInitial
1976 1 :
1977 1 : d.mu.Unlock()
1978 1 :
1979 1 : // TODO(jackson): Consider making these metrics optional.
1980 1 : aggProps := tablePropsAnnotator.MultiLevelAnnotation(vers.Levels[:])
1981 1 : metrics.Keys.RangeKeySetsCount = aggProps.NumRangeKeySets
1982 1 : metrics.Keys.TombstoneCount = aggProps.NumDeletions
1983 1 :
1984 1 : delBytes := deletionBytesAnnotator.MultiLevelAnnotation(vers.Levels[:])
1985 1 : metrics.Table.Garbage.PointDeletionsBytesEstimate = delBytes.PointDels
1986 1 : metrics.Table.Garbage.RangeDeletionsBytesEstimate = delBytes.RangeDels
1987 1 :
1988 1 : for i := 0; i < numLevels; i++ {
1989 1 : aggProps := tablePropsAnnotator.LevelAnnotation(vers.Levels[i])
1990 1 : metrics.Levels[i].Additional.ValueBlocksSize = aggProps.ValueBlocksSize
1991 1 : metrics.Table.Compression.MergeWith(&aggProps.CompressionMetrics)
1992 1 : }
1993 :
1994 1 : blobCompressionMetrics := blobCompressionStatsAnnotator.Annotation(&vers.BlobFiles)
1995 1 : metrics.BlobFiles.Compression.MergeWith(&blobCompressionMetrics)
1996 1 :
1997 1 : metrics.BlockCache = d.opts.Cache.Metrics()
1998 1 : metrics.FileCache, metrics.Filter = d.fileCache.Metrics()
1999 1 : metrics.TableIters = d.fileCache.IterCount()
2000 1 : metrics.CategoryStats = d.fileCache.SSTStatsCollector().GetStats()
2001 1 :
2002 1 : metrics.SecondaryCacheMetrics = d.objProvider.Metrics()
2003 1 :
2004 1 : metrics.Uptime = d.timeNow().Sub(d.openedAt)
2005 1 :
2006 1 : metrics.manualMemory = manual.GetMetrics()
2007 1 :
2008 1 : return metrics
2009 : }
2010 :
2011 : // sstablesOptions hold the optional parameters to retrieve TableInfo for all sstables.
2012 : type sstablesOptions struct {
2013 : // set to true will return the sstable properties in TableInfo
2014 : withProperties bool
2015 :
2016 : // if set, return sstables that overlap the key range (end-exclusive)
2017 : start []byte
2018 : end []byte
2019 :
2020 : withApproximateSpanBytes bool
2021 : }
2022 :
2023 : // SSTablesOption set optional parameter used by `DB.SSTables`.
2024 : type SSTablesOption func(*sstablesOptions)
2025 :
2026 : // WithProperties enable return sstable properties in each TableInfo.
2027 : //
2028 : // NOTE: if most of the sstable properties need to be read from disk,
2029 : // this options may make method `SSTables` quite slow.
2030 1 : func WithProperties() SSTablesOption {
2031 1 : return func(opt *sstablesOptions) {
2032 1 : opt.withProperties = true
2033 1 : }
2034 : }
2035 :
2036 : // WithKeyRangeFilter ensures returned sstables overlap start and end (end-exclusive)
2037 : // if start and end are both nil these properties have no effect.
2038 1 : func WithKeyRangeFilter(start, end []byte) SSTablesOption {
2039 1 : return func(opt *sstablesOptions) {
2040 1 : opt.end = end
2041 1 : opt.start = start
2042 1 : }
2043 : }
2044 :
2045 : // WithApproximateSpanBytes enables capturing the approximate number of bytes that
2046 : // overlap the provided key span for each sstable.
2047 : // NOTE: This option requires WithKeyRangeFilter.
2048 1 : func WithApproximateSpanBytes() SSTablesOption {
2049 1 : return func(opt *sstablesOptions) {
2050 1 : opt.withApproximateSpanBytes = true
2051 1 : }
2052 : }
2053 :
2054 : // BackingType denotes the type of storage backing a given sstable.
2055 : type BackingType int
2056 :
2057 : const (
2058 : // BackingTypeLocal denotes an sstable stored on local disk according to the
2059 : // objprovider. This file is completely owned by us.
2060 : BackingTypeLocal BackingType = iota
2061 : // BackingTypeShared denotes an sstable stored on shared storage, created
2062 : // by this Pebble instance and possibly shared by other Pebble instances.
2063 : // These types of files have lifecycle managed by Pebble.
2064 : BackingTypeShared
2065 : // BackingTypeSharedForeign denotes an sstable stored on shared storage,
2066 : // created by a Pebble instance other than this one. These types of files have
2067 : // lifecycle managed by Pebble.
2068 : BackingTypeSharedForeign
2069 : // BackingTypeExternal denotes an sstable stored on external storage,
2070 : // not owned by any Pebble instance and with no refcounting/cleanup methods
2071 : // or lifecycle management. An example of an external file is a file restored
2072 : // from a backup.
2073 : BackingTypeExternal
2074 : backingTypeCount
2075 : )
2076 :
2077 : var backingTypeToString = [backingTypeCount]string{
2078 : BackingTypeLocal: "local",
2079 : BackingTypeShared: "shared",
2080 : BackingTypeSharedForeign: "shared-foreign",
2081 : BackingTypeExternal: "external",
2082 : }
2083 :
2084 : // String implements fmt.Stringer.
2085 0 : func (b BackingType) String() string {
2086 0 : return backingTypeToString[b]
2087 0 : }
2088 :
2089 : // SSTableInfo export manifest.TableInfo with sstable.Properties alongside
2090 : // other file backing info.
2091 : type SSTableInfo struct {
2092 : manifest.TableInfo
2093 : TableStats manifest.TableStats
2094 : // Virtual indicates whether the sstable is virtual.
2095 : Virtual bool
2096 : // BackingSSTNum is the disk file number associated with the backing sstable.
2097 : // If Virtual is false, BackingSSTNum == PhysicalTableDiskFileNum(TableNum).
2098 : BackingSSTNum base.DiskFileNum
2099 : // BackingSize is the size of the backing sstable. This is the same with
2100 : // TableInfo.Size when the table is not virtual.
2101 : BackingSize uint64
2102 : // BackingType is the type of storage backing this sstable.
2103 : BackingType BackingType
2104 : // Locator is the remote.Locator backing this sstable, if the backing type is
2105 : // not BackingTypeLocal.
2106 : Locator remote.Locator
2107 : // ApproximateSpanBytes describes the approximate number of bytes within the
2108 : // sstable that fall within a particular span. It's populated only when the
2109 : // ApproximateSpanBytes option is passed into DB.SSTables.
2110 : ApproximateSpanBytes uint64 `json:"ApproximateSpanBytes,omitempty"`
2111 :
2112 : // Properties is the sstable properties of this table. If Virtual is true,
2113 : // then the Properties are associated with the backing sst.
2114 : Properties *sstable.Properties
2115 : }
2116 :
2117 : // SSTables retrieves the current sstables. The returned slice is indexed by
2118 : // level and each level is indexed by the position of the sstable within the
2119 : // level. Note that this information may be out of date due to concurrent
2120 : // flushes and compactions.
2121 1 : func (d *DB) SSTables(opts ...SSTablesOption) ([][]SSTableInfo, error) {
2122 1 : opt := &sstablesOptions{}
2123 1 : for _, fn := range opts {
2124 1 : fn(opt)
2125 1 : }
2126 :
2127 1 : if opt.withApproximateSpanBytes && (opt.start == nil || opt.end == nil) {
2128 1 : return nil, errors.Errorf("cannot use WithApproximateSpanBytes without WithKeyRangeFilter option")
2129 1 : }
2130 :
2131 : // Grab and reference the current readState.
2132 1 : readState := d.loadReadState()
2133 1 : defer readState.unref()
2134 1 :
2135 1 : // TODO(peter): This is somewhat expensive, especially on a large
2136 1 : // database. It might be worthwhile to unify TableInfo and TableMetadata and
2137 1 : // then we could simply return current.Files. Note that RocksDB is doing
2138 1 : // something similar to the current code, so perhaps it isn't too bad.
2139 1 : srcLevels := readState.current.Levels
2140 1 : var totalTables int
2141 1 : for i := range srcLevels {
2142 1 : totalTables += srcLevels[i].Len()
2143 1 : }
2144 :
2145 1 : destTables := make([]SSTableInfo, totalTables)
2146 1 : destLevels := make([][]SSTableInfo, len(srcLevels))
2147 1 : for i := range destLevels {
2148 1 : j := 0
2149 1 : for m := range srcLevels[i].All() {
2150 1 : if opt.start != nil && opt.end != nil {
2151 1 : b := base.UserKeyBoundsEndExclusive(opt.start, opt.end)
2152 1 : if !m.Overlaps(d.opts.Comparer.Compare, &b) {
2153 1 : continue
2154 : }
2155 : }
2156 1 : destTables[j] = SSTableInfo{
2157 1 : TableInfo: m.TableInfo(),
2158 1 : }
2159 1 : if stats, ok := m.Stats(); ok {
2160 1 : destTables[j].TableStats = *stats
2161 1 : }
2162 1 : if opt.withProperties {
2163 1 : p, err := d.fileCache.getTableProperties(
2164 1 : m,
2165 1 : )
2166 1 : if err != nil {
2167 0 : return nil, err
2168 0 : }
2169 1 : destTables[j].Properties = p
2170 : }
2171 1 : destTables[j].Virtual = m.Virtual
2172 1 : destTables[j].BackingSSTNum = m.TableBacking.DiskFileNum
2173 1 : destTables[j].BackingSize = m.TableBacking.Size
2174 1 : objMeta, err := d.objProvider.Lookup(base.FileTypeTable, m.TableBacking.DiskFileNum)
2175 1 : if err != nil {
2176 0 : return nil, err
2177 0 : }
2178 1 : if objMeta.IsRemote() {
2179 0 : if objMeta.IsShared() {
2180 0 : if d.objProvider.IsSharedForeign(objMeta) {
2181 0 : destTables[j].BackingType = BackingTypeSharedForeign
2182 0 : } else {
2183 0 : destTables[j].BackingType = BackingTypeShared
2184 0 : }
2185 0 : } else {
2186 0 : destTables[j].BackingType = BackingTypeExternal
2187 0 : }
2188 0 : destTables[j].Locator = objMeta.Remote.Locator
2189 1 : } else {
2190 1 : destTables[j].BackingType = BackingTypeLocal
2191 1 : }
2192 :
2193 1 : if opt.withApproximateSpanBytes {
2194 1 : if m.ContainedWithinSpan(d.opts.Comparer.Compare, opt.start, opt.end) {
2195 0 : destTables[j].ApproximateSpanBytes = m.Size
2196 1 : } else {
2197 1 : size, err := d.fileCache.estimateSize(m, opt.start, opt.end)
2198 1 : if err != nil {
2199 0 : return nil, err
2200 0 : }
2201 1 : destTables[j].ApproximateSpanBytes = size
2202 : }
2203 : }
2204 1 : j++
2205 : }
2206 1 : destLevels[i] = destTables[:j]
2207 1 : destTables = destTables[j:]
2208 : }
2209 :
2210 1 : return destLevels, nil
2211 : }
2212 :
2213 1 : func (d *DB) walPreallocateSize() int {
2214 1 : // Set the WAL preallocate size to 110% of the memtable size. Note that there
2215 1 : // is a bit of apples and oranges in units here as the memtabls size
2216 1 : // corresponds to the memory usage of the memtable while the WAL size is the
2217 1 : // size of the batches (plus overhead) stored in the WAL.
2218 1 : //
2219 1 : // TODO(peter): 110% of the memtable size is quite hefty for a block
2220 1 : // size. This logic is taken from GetWalPreallocateBlockSize in
2221 1 : // RocksDB. Could a smaller preallocation block size be used?
2222 1 : size := d.opts.MemTableSize
2223 1 : size = (size / 10) + size
2224 1 : return int(size)
2225 1 : }
2226 :
2227 : func (d *DB) newMemTable(
2228 : logNum base.DiskFileNum, logSeqNum base.SeqNum, minSize uint64,
2229 1 : ) (*memTable, *flushableEntry) {
2230 1 : targetSize := minSize + uint64(memTableEmptySize)
2231 1 : // The targetSize should be less than MemTableSize, because any batch >=
2232 1 : // MemTableSize/2 should be treated as a large flushable batch.
2233 1 : if targetSize > d.opts.MemTableSize {
2234 0 : panic(errors.AssertionFailedf("attempting to allocate memtable larger than MemTableSize"))
2235 : }
2236 : // Double until the next memtable size is at least large enough to fit
2237 : // minSize.
2238 1 : for d.mu.mem.nextSize < targetSize {
2239 1 : d.mu.mem.nextSize = min(2*d.mu.mem.nextSize, d.opts.MemTableSize)
2240 1 : }
2241 1 : size := d.mu.mem.nextSize
2242 1 : // The next memtable should be double the size, up to Options.MemTableSize.
2243 1 : if d.mu.mem.nextSize < d.opts.MemTableSize {
2244 1 : d.mu.mem.nextSize = min(2*d.mu.mem.nextSize, d.opts.MemTableSize)
2245 1 : }
2246 :
2247 1 : memtblOpts := memTableOptions{
2248 1 : Options: d.opts,
2249 1 : logSeqNum: logSeqNum,
2250 1 : }
2251 1 :
2252 1 : // Before attempting to allocate a new memtable, check if there's one
2253 1 : // available for recycling in memTableRecycle. Large contiguous allocations
2254 1 : // can be costly as fragmentation makes it more difficult to find a large
2255 1 : // contiguous free space. We've observed 64MB allocations taking 10ms+.
2256 1 : //
2257 1 : // To reduce these costly allocations, up to 1 obsolete memtable is stashed
2258 1 : // in `d.memTableRecycle` to allow a future memtable rotation to reuse
2259 1 : // existing memory.
2260 1 : var mem *memTable
2261 1 : mem = d.memTableRecycle.Swap(nil)
2262 1 : if mem != nil && uint64(mem.arenaBuf.Len()) != size {
2263 1 : d.freeMemTable(mem)
2264 1 : mem = nil
2265 1 : }
2266 1 : if mem != nil {
2267 1 : // Carry through the existing buffer and memory reservation.
2268 1 : memtblOpts.arenaBuf = mem.arenaBuf
2269 1 : memtblOpts.releaseAccountingReservation = mem.releaseAccountingReservation
2270 1 : } else {
2271 1 : mem = new(memTable)
2272 1 : memtblOpts.arenaBuf = manual.New(manual.MemTable, uintptr(size))
2273 1 : memtblOpts.releaseAccountingReservation = d.opts.Cache.Reserve(int(size))
2274 1 : d.memTableCount.Add(1)
2275 1 : d.memTableReserved.Add(int64(size))
2276 1 :
2277 1 : // Note: this is a no-op if invariants are disabled or race is enabled.
2278 1 : invariants.SetFinalizer(mem, checkMemTable)
2279 1 : }
2280 1 : mem.init(memtblOpts)
2281 1 :
2282 1 : entry := d.newFlushableEntry(mem, logNum, logSeqNum)
2283 1 : entry.releaseMemAccounting = func() {
2284 1 : // If the user leaks iterators, we may be releasing the memtable after
2285 1 : // the DB is already closed. In this case, we want to just release the
2286 1 : // memory because DB.Close won't come along to free it for us.
2287 1 : if err := d.closed.Load(); err != nil {
2288 1 : d.freeMemTable(mem)
2289 1 : return
2290 1 : }
2291 :
2292 : // The next memtable allocation might be able to reuse this memtable.
2293 : // Stash it on d.memTableRecycle.
2294 1 : if unusedMem := d.memTableRecycle.Swap(mem); unusedMem != nil {
2295 1 : // There was already a memtable waiting to be recycled. We're now
2296 1 : // responsible for freeing it.
2297 1 : d.freeMemTable(unusedMem)
2298 1 : }
2299 : }
2300 1 : return mem, entry
2301 : }
2302 :
2303 1 : func (d *DB) freeMemTable(m *memTable) {
2304 1 : d.memTableCount.Add(-1)
2305 1 : d.memTableReserved.Add(-int64(m.arenaBuf.Len()))
2306 1 : m.free()
2307 1 : }
2308 :
2309 : func (d *DB) newFlushableEntry(
2310 : f flushable, logNum base.DiskFileNum, logSeqNum base.SeqNum,
2311 1 : ) *flushableEntry {
2312 1 : fe := &flushableEntry{
2313 1 : flushable: f,
2314 1 : flushed: make(chan struct{}),
2315 1 : logNum: logNum,
2316 1 : logSeqNum: logSeqNum,
2317 1 : deleteFn: d.mu.versions.addObsolete,
2318 1 : deleteFnLocked: d.mu.versions.addObsoleteLocked,
2319 1 : }
2320 1 : fe.readerRefs.Store(1)
2321 1 : return fe
2322 1 : }
2323 :
2324 : // maybeInduceWriteStall is called before performing a memtable rotation in
2325 : // makeRoomForWrite. In some conditions, we prefer to stall the user's write
2326 : // workload rather than continuing to accept writes that may result in resource
2327 : // exhaustion or prohibitively slow reads.
2328 : //
2329 : // There are a couple reasons we might wait to rotate the memtable and
2330 : // instead induce a write stall:
2331 : // 1. If too many memtables have queued, we wait for a flush to finish before
2332 : // creating another memtable.
2333 : // 2. If L0 read amplification has grown too high, we wait for compactions
2334 : // to reduce the read amplification before accepting more writes that will
2335 : // increase write pressure.
2336 : //
2337 : // maybeInduceWriteStall checks these stall conditions, and if present, waits
2338 : // for them to abate.
2339 1 : func (d *DB) maybeInduceWriteStall(b *Batch) {
2340 1 : stalled := false
2341 1 : // This function will call EventListener.WriteStallBegin at most once. If
2342 1 : // it does call it, it will call EventListener.WriteStallEnd once before
2343 1 : // returning.
2344 1 : var timer *time.Timer
2345 1 : for {
2346 1 : var size uint64
2347 1 : for i := range d.mu.mem.queue {
2348 1 : size += d.mu.mem.queue[i].totalBytes()
2349 1 : }
2350 1 : if size >= uint64(d.opts.MemTableStopWritesThreshold)*d.opts.MemTableSize &&
2351 1 : !d.mu.log.manager.ElevateWriteStallThresholdForFailover() {
2352 1 : // We have filled up the current memtable, but already queued memtables
2353 1 : // are still flushing, so we wait.
2354 1 : if !stalled {
2355 1 : stalled = true
2356 1 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2357 1 : Reason: "memtable count limit reached",
2358 1 : })
2359 1 : }
2360 1 : beforeWait := crtime.NowMono()
2361 1 : // NB: In a rare case, we can start a write stall, and then the system
2362 1 : // may detect WAL failover, resulting in
2363 1 : // ElevateWriteStallThresholdForFailover returning true. So we want to
2364 1 : // recheck the predicate periodically, which we do by signaling the
2365 1 : // condition variable.
2366 1 : if timer == nil {
2367 1 : timer = time.AfterFunc(time.Second, d.mu.compact.cond.Broadcast)
2368 1 : } else {
2369 1 : timer.Reset(time.Second)
2370 1 : }
2371 1 : d.mu.compact.cond.Wait()
2372 1 : timer.Stop()
2373 1 : if b != nil {
2374 1 : b.commitStats.MemTableWriteStallDuration += beforeWait.Elapsed()
2375 1 : }
2376 1 : continue
2377 : }
2378 1 : l0ReadAmp := d.mu.versions.latest.l0Organizer.ReadAmplification()
2379 1 : if l0ReadAmp >= d.opts.L0StopWritesThreshold {
2380 1 : // There are too many level-0 files, so we wait.
2381 1 : if !stalled {
2382 1 : stalled = true
2383 1 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2384 1 : Reason: "L0 file count limit exceeded",
2385 1 : })
2386 1 : }
2387 1 : beforeWait := crtime.NowMono()
2388 1 : d.mu.compact.cond.Wait()
2389 1 : if b != nil {
2390 1 : b.commitStats.L0ReadAmpWriteStallDuration += beforeWait.Elapsed()
2391 1 : }
2392 1 : continue
2393 : }
2394 : // Not stalled.
2395 1 : if stalled {
2396 1 : d.opts.EventListener.WriteStallEnd()
2397 1 : }
2398 1 : return
2399 : }
2400 : }
2401 :
2402 : // makeRoomForWrite rotates the current mutable memtable, ensuring that the
2403 : // resulting mutable memtable has room to hold the contents of the provided
2404 : // Batch. The current memtable is rotated (marked as immutable) and a new
2405 : // mutable memtable is allocated. It reserves space in the new memtable and adds
2406 : // a reference to the memtable. The caller must later ensure that the memtable
2407 : // is unreferenced. This memtable rotation also causes a log rotation.
2408 : //
2409 : // If the current memtable is not full but the caller wishes to trigger a
2410 : // rotation regardless, the caller may pass a nil Batch, and no space in the
2411 : // resulting mutable memtable will be reserved.
2412 : //
2413 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2414 : // may be released and reacquired.
2415 1 : func (d *DB) makeRoomForWrite(b *Batch) error {
2416 1 : if b != nil && b.ingestedSSTBatch {
2417 0 : panic("pebble: invalid function call")
2418 : }
2419 1 : d.maybeInduceWriteStall(b)
2420 1 :
2421 1 : var newLogNum base.DiskFileNum
2422 1 : var prevLogSize uint64
2423 1 : if !d.opts.DisableWAL {
2424 1 : beforeRotate := crtime.NowMono()
2425 1 : newLogNum, prevLogSize = d.rotateWAL()
2426 1 : if b != nil {
2427 1 : b.commitStats.WALRotationDuration += beforeRotate.Elapsed()
2428 1 : }
2429 : }
2430 1 : immMem := d.mu.mem.mutable
2431 1 : imm := d.mu.mem.queue[len(d.mu.mem.queue)-1]
2432 1 : imm.logSize = prevLogSize
2433 1 :
2434 1 : var logSeqNum base.SeqNum
2435 1 : var minSize uint64
2436 1 : if b != nil {
2437 1 : logSeqNum = b.SeqNum()
2438 1 : if b.flushable != nil {
2439 1 : logSeqNum += base.SeqNum(b.Count())
2440 1 : // The batch is too large to fit in the memtable so add it directly to
2441 1 : // the immutable queue. The flushable batch is associated with the same
2442 1 : // log as the immutable memtable, but logically occurs after it in
2443 1 : // seqnum space. We ensure while flushing that the flushable batch
2444 1 : // is flushed along with the previous memtable in the flushable
2445 1 : // queue. See the top level comment in DB.flush1 to learn how this
2446 1 : // is ensured.
2447 1 : //
2448 1 : // See DB.commitWrite for the special handling of log writes for large
2449 1 : // batches. In particular, the large batch has already written to
2450 1 : // imm.logNum.
2451 1 : entry := d.newFlushableEntry(b.flushable, imm.logNum, b.SeqNum())
2452 1 : // The large batch is by definition large. Reserve space from the cache
2453 1 : // for it until it is flushed.
2454 1 : entry.releaseMemAccounting = d.opts.Cache.Reserve(int(b.flushable.totalBytes()))
2455 1 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2456 1 : } else {
2457 1 : minSize = b.memTableSize
2458 1 : }
2459 1 : } else {
2460 1 : // b == nil
2461 1 : //
2462 1 : // This is a manual forced flush.
2463 1 : logSeqNum = base.SeqNum(d.mu.versions.logSeqNum.Load())
2464 1 : imm.flushForced = true
2465 1 : // If we are manually flushing and we used less than half of the bytes in
2466 1 : // the memtable, don't increase the size for the next memtable. This
2467 1 : // reduces memtable memory pressure when an application is frequently
2468 1 : // manually flushing.
2469 1 : if uint64(immMem.availBytes()) > immMem.totalBytes()/2 {
2470 1 : d.mu.mem.nextSize = immMem.totalBytes()
2471 1 : }
2472 : }
2473 1 : d.rotateMemtable(newLogNum, logSeqNum, immMem, minSize)
2474 1 : if b != nil && b.flushable == nil {
2475 1 : err := d.mu.mem.mutable.prepare(b)
2476 1 : // Reserving enough space for the batch after rotation must never fail.
2477 1 : // We pass in a minSize that's equal to b.memtableSize to ensure that
2478 1 : // memtable rotation allocates a memtable sufficiently large. We also
2479 1 : // held d.commit.mu for the entirety of this function, ensuring that no
2480 1 : // other committers may have reserved memory in the new memtable yet.
2481 1 : if err == arenaskl.ErrArenaFull {
2482 0 : panic(errors.AssertionFailedf("memtable still full after rotation"))
2483 : }
2484 1 : return err
2485 : }
2486 1 : return nil
2487 : }
2488 :
2489 : // Both DB.mu and commitPipeline.mu must be held by the caller.
2490 : func (d *DB) rotateMemtable(
2491 : newLogNum base.DiskFileNum, logSeqNum base.SeqNum, prev *memTable, minSize uint64,
2492 1 : ) {
2493 1 : // Create a new memtable, scheduling the previous one for flushing. We do
2494 1 : // this even if the previous memtable was empty because the DB.Flush
2495 1 : // mechanism is dependent on being able to wait for the empty memtable to
2496 1 : // flush. We can't just mark the empty memtable as flushed here because we
2497 1 : // also have to wait for all previous immutable tables to
2498 1 : // flush. Additionally, the memtable is tied to particular WAL file and we
2499 1 : // want to go through the flush path in order to recycle that WAL file.
2500 1 : //
2501 1 : // NB: newLogNum corresponds to the WAL that contains mutations that are
2502 1 : // present in the new memtable. When immutable memtables are flushed to
2503 1 : // disk, a VersionEdit will be created telling the manifest the minimum
2504 1 : // unflushed log number (which will be the next one in d.mu.mem.mutable
2505 1 : // that was not flushed).
2506 1 : //
2507 1 : // NB: prev should be the current mutable memtable.
2508 1 : var entry *flushableEntry
2509 1 : d.mu.mem.mutable, entry = d.newMemTable(newLogNum, logSeqNum, minSize)
2510 1 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2511 1 : // d.logSize tracks the log size of the WAL file corresponding to the most
2512 1 : // recent flushable. The log size of the previous mutable memtable no longer
2513 1 : // applies to the current mutable memtable.
2514 1 : //
2515 1 : // It's tempting to perform this update in rotateWAL, but that would not be
2516 1 : // atomic with the enqueue of the new flushable. A call to DB.Metrics()
2517 1 : // could acquire DB.mu after the WAL has been rotated but before the new
2518 1 : // memtable has been appended; this would result in omitting the log size of
2519 1 : // the most recent flushable.
2520 1 : d.logSize.Store(0)
2521 1 : d.updateReadStateLocked(nil)
2522 1 : if prev.writerUnref() {
2523 1 : d.maybeScheduleFlush()
2524 1 : }
2525 : }
2526 :
2527 : // rotateWAL creates a new write-ahead log, possibly recycling a previous WAL's
2528 : // files. It returns the file number assigned to the new WAL, and the size of
2529 : // the previous WAL file.
2530 : //
2531 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2532 : // may be released and reacquired.
2533 1 : func (d *DB) rotateWAL() (newLogNum base.DiskFileNum, prevLogSize uint64) {
2534 1 : if d.opts.DisableWAL {
2535 0 : panic("pebble: invalid function call")
2536 : }
2537 1 : jobID := d.newJobIDLocked()
2538 1 : newLogNum = d.mu.versions.getNextDiskFileNum()
2539 1 :
2540 1 : d.mu.Unlock()
2541 1 : // Close the previous log first. This writes an EOF trailer
2542 1 : // signifying the end of the file and syncs it to disk. We must
2543 1 : // close the previous log before linking the new log file,
2544 1 : // otherwise a crash could leave both logs with unclean tails, and
2545 1 : // Open will treat the previous log as corrupt.
2546 1 : offset, err := d.mu.log.writer.Close()
2547 1 : if err != nil {
2548 0 : // What to do here? Stumbling on doesn't seem worthwhile. If we failed to
2549 0 : // close the previous log it is possible we lost a write.
2550 0 : d.opts.Logger.Fatalf("pebble: error closing WAL; data loss possible if we continue: %s", err)
2551 0 : }
2552 1 : prevLogSize = uint64(offset)
2553 1 : metrics := d.mu.log.writer.Metrics()
2554 1 :
2555 1 : d.mu.Lock()
2556 1 : if err := d.mu.log.metrics.LogWriterMetrics.Merge(&metrics); err != nil {
2557 0 : d.opts.Logger.Errorf("metrics error: %s", err)
2558 0 : }
2559 :
2560 1 : d.mu.Unlock()
2561 1 : writer, err := d.mu.log.manager.Create(wal.NumWAL(newLogNum), int(jobID))
2562 1 : if err != nil {
2563 0 : panic(err)
2564 : }
2565 :
2566 1 : d.mu.Lock()
2567 1 : d.mu.log.writer = writer
2568 1 : return newLogNum, prevLogSize
2569 : }
2570 :
2571 1 : func (d *DB) getEarliestUnflushedSeqNumLocked() base.SeqNum {
2572 1 : seqNum := base.SeqNumMax
2573 1 : for i := range d.mu.mem.queue {
2574 1 : logSeqNum := d.mu.mem.queue[i].logSeqNum
2575 1 : if seqNum > logSeqNum {
2576 1 : seqNum = logSeqNum
2577 1 : }
2578 : }
2579 1 : return seqNum
2580 : }
2581 :
2582 1 : func (d *DB) getInProgressCompactionInfoLocked(finishing compaction) (rv []compactionInfo) {
2583 1 : for c := range d.mu.compact.inProgress {
2584 1 : if !c.IsFlush() && (finishing == nil || c != finishing) {
2585 1 : rv = append(rv, c.Info())
2586 1 : }
2587 : }
2588 1 : return
2589 : }
2590 :
2591 1 : func inProgressL0Compactions(inProgress []compactionInfo) []manifest.L0Compaction {
2592 1 : var compactions []manifest.L0Compaction
2593 1 : for _, info := range inProgress {
2594 1 : // Skip in-progress compactions that have already committed; the L0
2595 1 : // sublevels initialization code requires the set of in-progress
2596 1 : // compactions to be consistent with the current version. Compactions
2597 1 : // with versionEditApplied=true are already applied to the current
2598 1 : // version and but are performing cleanup without the database mutex.
2599 1 : if info.versionEditApplied {
2600 1 : continue
2601 : }
2602 1 : l0 := false
2603 1 : for _, cl := range info.inputs {
2604 1 : l0 = l0 || cl.level == 0
2605 1 : }
2606 1 : if !l0 {
2607 1 : continue
2608 : }
2609 1 : compactions = append(compactions, manifest.L0Compaction{
2610 1 : Bounds: *info.bounds,
2611 1 : IsIntraL0: info.outputLevel == 0,
2612 1 : })
2613 : }
2614 1 : return compactions
2615 : }
2616 :
2617 : // firstError returns the first non-nil error of err0 and err1, or nil if both
2618 : // are nil.
2619 1 : func firstError(err0, err1 error) error {
2620 1 : if err0 != nil {
2621 1 : return err0
2622 1 : }
2623 1 : return err1
2624 : }
2625 :
2626 : // SetCreatorID sets the CreatorID which is needed in order to use shared objects.
2627 : // Remote object usage is disabled until this method is called the first time.
2628 : // Once set, the Creator ID is persisted and cannot change.
2629 : //
2630 : // Does nothing if SharedStorage was not set in the options when the DB was
2631 : // opened or if the DB is in read-only mode.
2632 1 : func (d *DB) SetCreatorID(creatorID uint64) error {
2633 1 : if d.opts.Experimental.RemoteStorage == nil || d.opts.ReadOnly {
2634 0 : return nil
2635 0 : }
2636 1 : return d.objProvider.SetCreatorID(objstorage.CreatorID(creatorID))
2637 : }
2638 :
2639 : // KeyStatistics keeps track of the number of keys that have been pinned by a
2640 : // snapshot as well as counts of the different key kinds in the lsm.
2641 : //
2642 : // One way of using the accumulated stats, when we only have sets and dels,
2643 : // and say the counts are represented as del_count, set_count,
2644 : // del_latest_count, set_latest_count, snapshot_pinned_count.
2645 : //
2646 : // - del_latest_count + set_latest_count is the set of unique user keys
2647 : // (unique).
2648 : //
2649 : // - set_latest_count is the set of live unique user keys (live_unique).
2650 : //
2651 : // - Garbage is del_count + set_count - live_unique.
2652 : //
2653 : // - If everything were in the LSM, del_count+set_count-snapshot_pinned_count
2654 : // would also be the set of unique user keys (note that
2655 : // snapshot_pinned_count is counting something different -- see comment below).
2656 : // But snapshot_pinned_count only counts keys in the LSM so the excess here
2657 : // must be keys in memtables.
2658 : type KeyStatistics struct {
2659 : // TODO(sumeer): the SnapshotPinned* are incorrect in that these older
2660 : // versions can be in a different level. Either fix the accounting or
2661 : // rename these fields.
2662 :
2663 : // SnapshotPinnedKeys represents obsolete keys that cannot be elided during
2664 : // a compaction, because they are required by an open snapshot.
2665 : SnapshotPinnedKeys int
2666 : // SnapshotPinnedKeysBytes is the total number of bytes of all snapshot
2667 : // pinned keys.
2668 : SnapshotPinnedKeysBytes uint64
2669 : // KindsCount is the count for each kind of key. It includes point keys,
2670 : // range deletes and range keys.
2671 : KindsCount [InternalKeyKindMax + 1]int
2672 : // LatestKindsCount is the count for each kind of key when it is the latest
2673 : // kind for a user key. It is only populated for point keys.
2674 : LatestKindsCount [InternalKeyKindMax + 1]int
2675 : }
2676 :
2677 : // LSMKeyStatistics is used by DB.ScanStatistics.
2678 : type LSMKeyStatistics struct {
2679 : Accumulated KeyStatistics
2680 : // Levels contains statistics only for point keys. Range deletions and range keys will
2681 : // appear in Accumulated but not Levels.
2682 : Levels [numLevels]KeyStatistics
2683 : // BytesRead represents the logical, pre-compression size of keys and values read
2684 : BytesRead uint64
2685 : }
2686 :
2687 : // ScanStatisticsOptions is used by DB.ScanStatistics.
2688 : type ScanStatisticsOptions struct {
2689 : // LimitBytesPerSecond indicates the number of bytes that are able to be read
2690 : // per second using ScanInternal.
2691 : // A value of 0 indicates that there is no limit set.
2692 : LimitBytesPerSecond int64
2693 : }
2694 :
2695 : // ScanStatistics returns the count of different key kinds within the lsm for a
2696 : // key span [lower, upper) as well as the number of snapshot keys.
2697 : func (d *DB) ScanStatistics(
2698 : ctx context.Context, lower, upper []byte, opts ScanStatisticsOptions,
2699 1 : ) (stats LSMKeyStatistics, err error) {
2700 1 : var prevKey InternalKey
2701 1 : var rateLimitFunc func(key *InternalKey, val LazyValue) error
2702 1 : tb := tokenbucket.TokenBucket{}
2703 1 :
2704 1 : if opts.LimitBytesPerSecond != 0 {
2705 0 : const minBytesPerSec = 100 * 1024
2706 0 : if opts.LimitBytesPerSecond < minBytesPerSec {
2707 0 : return stats, errors.Newf("pebble: ScanStatistics read bandwidth limit %d is below minimum %d", opts.LimitBytesPerSecond, minBytesPerSec)
2708 0 : }
2709 : // Each "token" roughly corresponds to a byte that was read.
2710 0 : tb.Init(tokenbucket.TokensPerSecond(opts.LimitBytesPerSecond), tokenbucket.Tokens(1024))
2711 0 : rateLimitFunc = func(key *InternalKey, val LazyValue) error {
2712 0 : return tb.WaitCtx(ctx, tokenbucket.Tokens(key.Size()+val.Len()))
2713 0 : }
2714 : }
2715 :
2716 1 : scanInternalOpts := &ScanInternalOptions{
2717 1 : VisitPointKey: func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error {
2718 1 : // If the previous key is equal to the current point key, the current key was
2719 1 : // pinned by a snapshot.
2720 1 : size := uint64(key.Size())
2721 1 : kind := key.Kind()
2722 1 : sameKey := d.equal(prevKey.UserKey, key.UserKey)
2723 1 : if iterInfo.Kind == IteratorLevelLSM && sameKey {
2724 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeys++
2725 1 : stats.Levels[iterInfo.Level].SnapshotPinnedKeysBytes += size
2726 1 : stats.Accumulated.SnapshotPinnedKeys++
2727 1 : stats.Accumulated.SnapshotPinnedKeysBytes += size
2728 1 : }
2729 1 : if iterInfo.Kind == IteratorLevelLSM {
2730 1 : stats.Levels[iterInfo.Level].KindsCount[kind]++
2731 1 : }
2732 1 : if !sameKey {
2733 1 : if iterInfo.Kind == IteratorLevelLSM {
2734 1 : stats.Levels[iterInfo.Level].LatestKindsCount[kind]++
2735 1 : }
2736 1 : stats.Accumulated.LatestKindsCount[kind]++
2737 : }
2738 :
2739 1 : stats.Accumulated.KindsCount[kind]++
2740 1 : prevKey.CopyFrom(*key)
2741 1 : stats.BytesRead += uint64(key.Size() + value.Len())
2742 1 : return nil
2743 : },
2744 0 : VisitRangeDel: func(start, end []byte, seqNum base.SeqNum) error {
2745 0 : stats.Accumulated.KindsCount[InternalKeyKindRangeDelete]++
2746 0 : stats.BytesRead += uint64(len(start) + len(end))
2747 0 : return nil
2748 0 : },
2749 0 : VisitRangeKey: func(start, end []byte, keys []rangekey.Key) error {
2750 0 : stats.BytesRead += uint64(len(start) + len(end))
2751 0 : for _, key := range keys {
2752 0 : stats.Accumulated.KindsCount[key.Kind()]++
2753 0 : stats.BytesRead += uint64(len(key.Value) + len(key.Suffix))
2754 0 : }
2755 0 : return nil
2756 : },
2757 : IncludeObsoleteKeys: true,
2758 : IterOptions: IterOptions{
2759 : KeyTypes: IterKeyTypePointsAndRanges,
2760 : LowerBound: lower,
2761 : UpperBound: upper,
2762 : },
2763 : RateLimitFunc: rateLimitFunc,
2764 : }
2765 1 : iter, err := d.newInternalIter(ctx, snapshotIterOpts{}, scanInternalOpts)
2766 1 : if err != nil {
2767 0 : return LSMKeyStatistics{}, err
2768 0 : }
2769 1 : defer func() { err = errors.CombineErrors(err, iter.Close()) }()
2770 :
2771 1 : err = scanInternalImpl(ctx, iter, scanInternalOpts)
2772 1 : if err != nil {
2773 0 : return LSMKeyStatistics{}, err
2774 0 : }
2775 1 : return stats, nil
2776 : }
2777 :
2778 : // ObjProvider returns the objstorage.Provider for this database. Meant to be
2779 : // used for internal purposes only.
2780 1 : func (d *DB) ObjProvider() objstorage.Provider {
2781 1 : return d.objProvider
2782 1 : }
2783 :
2784 : // DebugString returns a debugging string describing the LSM.
2785 0 : func (d *DB) DebugString() string {
2786 0 : return d.DebugCurrentVersion().DebugString()
2787 0 : }
2788 :
2789 : // DebugCurrentVersion returns the current LSM tree metadata. Should only be
2790 : // used for testing/debugging.
2791 1 : func (d *DB) DebugCurrentVersion() *manifest.Version {
2792 1 : d.mu.Lock()
2793 1 : defer d.mu.Unlock()
2794 1 : return d.mu.versions.currentVersion()
2795 1 : }
|
