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 calls to compact that have not yet finished. This is different
422 : // from compactingCount, as calls to compact will attempt to schedule and run
423 : // follow-up compactions after the current compaction finishes, dropping db.mu
424 : // in between updating compactingCount and the scheduling operation. This value is
425 : // used in tests in order to track when all compaction activity has finished.
426 : compactProcesses int
427 : // The number of download compactions.
428 : downloadingCount int
429 : // The list of deletion hints, suggesting ranges for delete-only
430 : // compactions.
431 : deletionHints []deleteCompactionHint
432 : // The list of manual compactions. The next manual compaction to perform
433 : // is at the start of the list. New entries are added to the end.
434 : manual []*manualCompaction
435 : manualLen atomic.Int32
436 : // manualID is used to identify manualCompactions in the manual slice.
437 : manualID uint64
438 : // downloads is the list of pending download tasks. The next download to
439 : // perform is at the start of the list. New entries are added to the end.
440 : downloads []*downloadSpanTask
441 : // inProgress is the set of in-progress flushes and compactions.
442 : // It's used in the calculation of some metrics and to initialize L0
443 : // sublevels' state. Some of the compactions contained within this
444 : // map may have already committed an edit to the version but are
445 : // lingering performing cleanup, like deleting obsolete files.
446 : inProgress map[compaction]struct{}
447 : // burstConcurrency is the number of additional compaction
448 : // concurrency slots that have been added in order to allow
449 : // high-priority compactions to run.
450 : //
451 : // Today, it's used when scheduling 'high priority' blob file
452 : // rewrite compactions to reclaim disk space. Normally compactions
453 : // that reclaim disk space and don't reshape the LSM are considered
454 : // low priority. If there's sufficient space amplification,
455 : // heuristics may decide to let these space-reclamation compactions
456 : // run anyways.
457 : //
458 : // In this case we don't want to steal concurrency from the default
459 : // compactions, so we temporarily inflate the allowed compaction
460 : // concurrency for the duration of the compaction. This field is the
461 : // count of compactions within inProgress for which
462 : // UsesBurstConcurrency is true. It's incremented when one begins
463 : // and decremented when it completes.
464 : burstConcurrency atomic.Int32
465 :
466 : // rescheduleReadCompaction indicates to an iterator that a read compaction
467 : // should be scheduled.
468 : rescheduleReadCompaction bool
469 :
470 : // readCompactions is a readCompactionQueue which keeps track of the
471 : // compactions which we might have to perform.
472 : readCompactions readCompactionQueue
473 :
474 : // The cumulative duration of all completed compactions since Open.
475 : // Does not include flushes.
476 : duration time.Duration
477 : // Flush throughput metric.
478 : flushWriteThroughput ThroughputMetric
479 : // The idle start time for the flush "loop", i.e., when the flushing
480 : // bool above transitions to false.
481 : noOngoingFlushStartTime crtime.Mono
482 : }
483 :
484 : fileDeletions struct {
485 : // Non-zero when file cleaning is disableCount. The disableCount
486 : // count acts as a reference count to prohibit file cleaning. See
487 : // DB.{disable,enable}FileDeletions().
488 : disableCount int
489 : // queuedStats holds cumulative stats for files that have been
490 : // queued for deletion by the cleanup manager. These stats are
491 : // monotonically increasing for the *DB's lifetime.
492 : queuedStats obsoleteObjectStats
493 : }
494 :
495 : snapshots struct {
496 : // The list of active snapshots.
497 : snapshotList
498 :
499 : // The cumulative count and size of snapshot-pinned keys written to
500 : // sstables.
501 : cumulativePinnedCount uint64
502 : cumulativePinnedSize uint64
503 : }
504 :
505 : tableStats struct {
506 : // Condition variable used to signal the completion of a
507 : // job to collect table stats.
508 : cond sync.Cond
509 : // True when a stat collection operation is in progress.
510 : loading bool
511 : // True if stat collection has loaded statistics for all tables
512 : // other than those listed explicitly in pending. This flag starts
513 : // as false when a database is opened and flips to true once stat
514 : // collection has caught up.
515 : loadedInitial bool
516 : // A slice of files for which stats have not been computed.
517 : // Compactions, ingests, flushes append files to be processed. An
518 : // active stat collection goroutine clears the list and processes
519 : // them.
520 : pending []manifest.NewTableEntry
521 : }
522 :
523 : tableValidation struct {
524 : // cond is a condition variable used to signal the completion of a
525 : // job to validate one or more sstables.
526 : cond sync.Cond
527 : // pending is a slice of metadata for sstables waiting to be
528 : // validated. Only physical sstables should be added to the pending
529 : // queue.
530 : pending []manifest.NewTableEntry
531 : // validating is set to true when validation is running.
532 : validating bool
533 : }
534 : }
535 :
536 : fileSizeAnnotator manifest.TableAnnotator[fileSizeByBacking]
537 :
538 : // problemSpans keeps track of spans of keys within LSM levels where
539 : // compactions have failed; used to avoid retrying these compactions too
540 : // quickly.
541 : problemSpans problemspans.ByLevel
542 :
543 : // Normally equal to time.Now() but may be overridden in tests.
544 : timeNow func() time.Time
545 : // the time at database Open; may be used to compute metrics like effective
546 : // compaction concurrency
547 : openedAt time.Time
548 : }
549 :
550 : var _ Reader = (*DB)(nil)
551 : var _ Writer = (*DB)(nil)
552 :
553 : // TestOnlyWaitForCleaning MUST only be used in tests.
554 0 : func (d *DB) TestOnlyWaitForCleaning() {
555 0 : d.cleanupManager.Wait()
556 0 : }
557 :
558 : // Set sets the value for the given key. It overwrites any previous value
559 : // for that key; a DB is not a multi-map.
560 : //
561 : // It is safe to modify the contents of the arguments after Set returns.
562 1 : func (d *DB) Set(key, value []byte, opts *WriteOptions) error {
563 1 : b := newBatch(d)
564 1 : _ = b.Set(key, value, opts)
565 1 : if err := d.Apply(b, opts); err != nil {
566 0 : return err
567 0 : }
568 : // Only release the batch on success.
569 1 : return b.Close()
570 : }
571 :
572 : // Delete deletes the value for the given key. Deletes are blind all will
573 : // succeed even if the given key does not exist.
574 : //
575 : // It is safe to modify the contents of the arguments after Delete returns.
576 1 : func (d *DB) Delete(key []byte, opts *WriteOptions) error {
577 1 : b := newBatch(d)
578 1 : _ = b.Delete(key, opts)
579 1 : if err := d.Apply(b, opts); err != nil {
580 0 : return err
581 0 : }
582 : // Only release the batch on success.
583 1 : return b.Close()
584 : }
585 :
586 : // DeleteSized behaves identically to Delete, but takes an additional
587 : // argument indicating the size of the value being deleted. DeleteSized
588 : // should be preferred when the caller has the expectation that there exists
589 : // a single internal KV pair for the key (eg, the key has not been
590 : // overwritten recently), and the caller knows the size of its value.
591 : //
592 : // DeleteSized will record the value size within the tombstone and use it to
593 : // inform compaction-picking heuristics which strive to reduce space
594 : // amplification in the LSM. This "calling your shot" mechanic allows the
595 : // storage engine to more accurately estimate and reduce space amplification.
596 : //
597 : // It is safe to modify the contents of the arguments after DeleteSized
598 : // returns.
599 1 : func (d *DB) DeleteSized(key []byte, valueSize uint32, opts *WriteOptions) error {
600 1 : b := newBatch(d)
601 1 : _ = b.DeleteSized(key, valueSize, opts)
602 1 : if err := d.Apply(b, opts); err != nil {
603 0 : return err
604 0 : }
605 : // Only release the batch on success.
606 1 : return b.Close()
607 : }
608 :
609 : // SingleDelete adds an action to the batch that single deletes the entry for key.
610 : // See Writer.SingleDelete for more details on the semantics of SingleDelete.
611 : //
612 : // WARNING: See the detailed warning in Writer.SingleDelete before using this.
613 : //
614 : // It is safe to modify the contents of the arguments after SingleDelete returns.
615 1 : func (d *DB) SingleDelete(key []byte, opts *WriteOptions) error {
616 1 : b := newBatch(d)
617 1 : _ = b.SingleDelete(key, opts)
618 1 : if err := d.Apply(b, opts); err != nil {
619 0 : return err
620 0 : }
621 : // Only release the batch on success.
622 1 : return b.Close()
623 : }
624 :
625 : // DeleteRange deletes all of the keys (and values) in the range [start,end)
626 : // (inclusive on start, exclusive on end).
627 : //
628 : // It is safe to modify the contents of the arguments after DeleteRange
629 : // returns.
630 1 : func (d *DB) DeleteRange(start, end []byte, opts *WriteOptions) error {
631 1 : b := newBatch(d)
632 1 : _ = b.DeleteRange(start, end, opts)
633 1 : if err := d.Apply(b, opts); err != nil {
634 0 : return err
635 0 : }
636 : // Only release the batch on success.
637 1 : return b.Close()
638 : }
639 :
640 : // Merge adds an action to the DB that merges the value at key with the new
641 : // value. The details of the merge are dependent upon the configured merge
642 : // operator.
643 : //
644 : // It is safe to modify the contents of the arguments after Merge returns.
645 1 : func (d *DB) Merge(key, value []byte, opts *WriteOptions) error {
646 1 : b := newBatch(d)
647 1 : _ = b.Merge(key, value, opts)
648 1 : if err := d.Apply(b, opts); err != nil {
649 0 : return err
650 0 : }
651 : // Only release the batch on success.
652 1 : return b.Close()
653 : }
654 :
655 : // LogData adds the specified to the batch. The data will be written to the
656 : // WAL, but not added to memtables or sstables. Log data is never indexed,
657 : // which makes it useful for testing WAL performance.
658 : //
659 : // It is safe to modify the contents of the argument after LogData returns.
660 1 : func (d *DB) LogData(data []byte, opts *WriteOptions) error {
661 1 : b := newBatch(d)
662 1 : _ = b.LogData(data, opts)
663 1 : if err := d.Apply(b, opts); err != nil {
664 0 : return err
665 0 : }
666 : // Only release the batch on success.
667 1 : return b.Close()
668 : }
669 :
670 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
671 : // timestamp suffix to value. The suffix is optional. If any portion of the key
672 : // range [start, end) is already set by a range key with the same suffix value,
673 : // RangeKeySet overrides it.
674 : //
675 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
676 1 : func (d *DB) RangeKeySet(start, end, suffix, value []byte, opts *WriteOptions) error {
677 1 : b := newBatch(d)
678 1 : _ = b.RangeKeySet(start, end, suffix, value, opts)
679 1 : if err := d.Apply(b, opts); err != nil {
680 0 : return err
681 0 : }
682 : // Only release the batch on success.
683 1 : return b.Close()
684 : }
685 :
686 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
687 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
688 : // range key. RangeKeyUnset only removes portions of range keys that fall within
689 : // the [start, end) key span, and only range keys with suffixes that exactly
690 : // match the unset suffix.
691 : //
692 : // It is safe to modify the contents of the arguments after RangeKeyUnset
693 : // returns.
694 1 : func (d *DB) RangeKeyUnset(start, end, suffix []byte, opts *WriteOptions) error {
695 1 : b := newBatch(d)
696 1 : _ = b.RangeKeyUnset(start, end, suffix, opts)
697 1 : if err := d.Apply(b, opts); err != nil {
698 0 : return err
699 0 : }
700 : // Only release the batch on success.
701 1 : return b.Close()
702 : }
703 :
704 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
705 : // (inclusive on start, exclusive on end). It does not delete point keys (for
706 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
707 : // bounds, including those with or without suffixes.
708 : //
709 : // It is safe to modify the contents of the arguments after RangeKeyDelete
710 : // returns.
711 1 : func (d *DB) RangeKeyDelete(start, end []byte, opts *WriteOptions) error {
712 1 : b := newBatch(d)
713 1 : _ = b.RangeKeyDelete(start, end, opts)
714 1 : if err := d.Apply(b, opts); err != nil {
715 0 : return err
716 0 : }
717 : // Only release the batch on success.
718 1 : return b.Close()
719 : }
720 :
721 : // Apply the operations contained in the batch to the DB. If the batch is large
722 : // the contents of the batch may be retained by the database. If that occurs
723 : // the batch contents will be cleared preventing the caller from attempting to
724 : // reuse them.
725 : //
726 : // It is safe to modify the contents of the arguments after Apply returns.
727 : //
728 : // Apply returns ErrInvalidBatch if the provided batch is invalid in any way.
729 1 : func (d *DB) Apply(batch *Batch, opts *WriteOptions) error {
730 1 : return d.applyInternal(batch, opts, false)
731 1 : }
732 :
733 : // ApplyNoSyncWait must only be used when opts.Sync is true and the caller
734 : // does not want to wait for the WAL fsync to happen. The method will return
735 : // once the mutation is applied to the memtable and is visible (note that a
736 : // mutation is visible before the WAL sync even in the wait case, so we have
737 : // not weakened the durability semantics). The caller must call Batch.SyncWait
738 : // to wait for the WAL fsync. The caller must not Close the batch without
739 : // first calling Batch.SyncWait.
740 : //
741 : // RECOMMENDATION: Prefer using Apply unless you really understand why you
742 : // need ApplyNoSyncWait.
743 : // EXPERIMENTAL: API/feature subject to change. Do not yet use outside
744 : // CockroachDB.
745 1 : func (d *DB) ApplyNoSyncWait(batch *Batch, opts *WriteOptions) error {
746 1 : if !opts.Sync {
747 0 : return errors.Errorf("cannot request asynchonous apply when WriteOptions.Sync is false")
748 0 : }
749 1 : return d.applyInternal(batch, opts, true)
750 : }
751 :
752 : // REQUIRES: noSyncWait => opts.Sync
753 1 : func (d *DB) applyInternal(batch *Batch, opts *WriteOptions, noSyncWait bool) error {
754 1 : if err := d.closed.Load(); err != nil {
755 0 : panic(err)
756 : }
757 1 : if batch.committing {
758 0 : panic("pebble: batch already committing")
759 : }
760 1 : if batch.applied.Load() {
761 0 : panic("pebble: batch already applied")
762 : }
763 1 : if d.opts.ReadOnly {
764 0 : return ErrReadOnly
765 0 : }
766 1 : if batch.db != nil && batch.db != d {
767 0 : panic(fmt.Sprintf("pebble: batch db mismatch: %p != %p", batch.db, d))
768 : }
769 :
770 1 : sync := opts.GetSync()
771 1 : if sync && d.opts.DisableWAL {
772 0 : return errors.New("pebble: WAL disabled")
773 0 : }
774 :
775 1 : if fmv := d.FormatMajorVersion(); fmv < batch.minimumFormatMajorVersion {
776 0 : panic(fmt.Sprintf(
777 0 : "pebble: batch requires at least format major version %d (current: %d)",
778 0 : batch.minimumFormatMajorVersion, fmv,
779 0 : ))
780 : }
781 :
782 1 : if batch.countRangeKeys > 0 {
783 1 : if d.split == nil {
784 0 : return errNoSplit
785 0 : }
786 : }
787 1 : batch.committing = true
788 1 :
789 1 : if batch.db == nil {
790 0 : if err := batch.refreshMemTableSize(); err != nil {
791 0 : return err
792 0 : }
793 : }
794 1 : if batch.memTableSize >= d.largeBatchThreshold {
795 1 : var err error
796 1 : batch.flushable, err = newFlushableBatch(batch, d.opts.Comparer)
797 1 : if err != nil {
798 0 : return err
799 0 : }
800 : }
801 1 : if err := d.commit.Commit(batch, sync, noSyncWait); err != nil {
802 0 : // There isn't much we can do on an error here. The commit pipeline will be
803 0 : // horked at this point.
804 0 : d.opts.Logger.Fatalf("pebble: fatal commit error: %v", err)
805 0 : }
806 : // If this is a large batch, we need to clear the batch contents as the
807 : // flushable batch may still be present in the flushables queue.
808 : //
809 : // TODO(peter): Currently large batches are written to the WAL. We could
810 : // skip the WAL write and instead wait for the large batch to be flushed to
811 : // an sstable. For a 100 MB batch, this might actually be faster. For a 1
812 : // GB batch this is almost certainly faster.
813 1 : if batch.flushable != nil {
814 1 : batch.data = nil
815 1 : }
816 1 : return nil
817 : }
818 :
819 1 : func (d *DB) commitApply(b *Batch, mem *memTable) error {
820 1 : if b.flushable != nil {
821 1 : // This is a large batch which was already added to the immutable queue.
822 1 : return nil
823 1 : }
824 1 : err := mem.apply(b, b.SeqNum())
825 1 : if err != nil {
826 0 : return err
827 0 : }
828 :
829 : // If the batch contains range tombstones and the database is configured
830 : // to flush range deletions, schedule a delayed flush so that disk space
831 : // may be reclaimed without additional writes or an explicit flush.
832 1 : if b.countRangeDels > 0 && d.opts.FlushDelayDeleteRange > 0 {
833 1 : d.mu.Lock()
834 1 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayDeleteRange)
835 1 : d.mu.Unlock()
836 1 : }
837 :
838 : // If the batch contains range keys and the database is configured to flush
839 : // range keys, schedule a delayed flush so that the range keys are cleared
840 : // from the memtable.
841 1 : if b.countRangeKeys > 0 && d.opts.FlushDelayRangeKey > 0 {
842 1 : d.mu.Lock()
843 1 : d.maybeScheduleDelayedFlush(mem, d.opts.FlushDelayRangeKey)
844 1 : d.mu.Unlock()
845 1 : }
846 :
847 1 : if mem.writerUnref() {
848 1 : d.mu.Lock()
849 1 : d.maybeScheduleFlush()
850 1 : d.mu.Unlock()
851 1 : }
852 1 : return nil
853 : }
854 :
855 1 : func (d *DB) commitWrite(b *Batch, syncWG *sync.WaitGroup, syncErr *error) (*memTable, error) {
856 1 : var size int64
857 1 : repr := b.Repr()
858 1 :
859 1 : if b.flushable != nil {
860 1 : // We have a large batch. Such batches are special in that they don't get
861 1 : // added to the memtable, and are instead inserted into the queue of
862 1 : // memtables. The call to makeRoomForWrite with this batch will force the
863 1 : // current memtable to be flushed. We want the large batch to be part of
864 1 : // the same log, so we add it to the WAL here, rather than after the call
865 1 : // to makeRoomForWrite().
866 1 : //
867 1 : // Set the sequence number since it was not set to the correct value earlier
868 1 : // (see comment in newFlushableBatch()).
869 1 : b.flushable.setSeqNum(b.SeqNum())
870 1 : if !d.opts.DisableWAL {
871 1 : var err error
872 1 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b)
873 1 : if err != nil {
874 0 : panic(err)
875 : }
876 : }
877 : }
878 :
879 1 : var err error
880 1 : // Grab a reference to the memtable. We don't hold DB.mu, but we do hold
881 1 : // d.commit.mu. It's okay for readers of d.mu.mem.mutable to only hold one of
882 1 : // d.commit.mu or d.mu, because memtable rotations require holding both.
883 1 : mem := d.mu.mem.mutable
884 1 : // Batches which contain keys of kind InternalKeyKindIngestSST will
885 1 : // never be applied to the memtable, so we don't need to make room for
886 1 : // write.
887 1 : if !b.ingestedSSTBatch {
888 1 : // Flushable batches will require a rotation of the memtable regardless,
889 1 : // so only attempt an optimistic reservation of space in the current
890 1 : // memtable if this batch is not a large flushable batch.
891 1 : if b.flushable == nil {
892 1 : err = d.mu.mem.mutable.prepare(b)
893 1 : }
894 1 : if b.flushable != nil || err == arenaskl.ErrArenaFull {
895 1 : // Slow path.
896 1 : // We need to acquire DB.mu and rotate the memtable.
897 1 : func() {
898 1 : d.mu.Lock()
899 1 : defer d.mu.Unlock()
900 1 : err = d.makeRoomForWrite(b)
901 1 : mem = d.mu.mem.mutable
902 1 : }()
903 : }
904 : }
905 1 : if err != nil {
906 0 : return nil, err
907 0 : }
908 1 : if d.opts.DisableWAL {
909 1 : return mem, nil
910 1 : }
911 1 : d.logBytesIn.Add(uint64(len(repr)))
912 1 :
913 1 : if b.flushable == nil {
914 1 : size, err = d.mu.log.writer.WriteRecord(repr, wal.SyncOptions{Done: syncWG, Err: syncErr}, b)
915 1 : if err != nil {
916 0 : panic(err)
917 : }
918 : }
919 :
920 1 : d.logSize.Store(uint64(size))
921 1 : return mem, err
922 : }
923 :
924 : type iterAlloc struct {
925 : keyBuf []byte `invariants:"reused"`
926 : boundsBuf [2][]byte `invariants:"reused"`
927 : prefixOrFullSeekKey []byte `invariants:"reused"`
928 : batchState iteratorBatchState
929 : dbi Iterator
930 : merging mergingIter
931 : mlevels [3 + numLevels]mergingIterLevel
932 : levels [3 + numLevels]levelIter
933 : levelsPositioned [3 + numLevels]bool
934 : }
935 :
936 : // maybeAssertZeroed asserts that i is a "zeroed" value. See assertZeroed for
937 : // the definition of "zeroed". It's used to ensure we're properly zeroing out
938 : // memory before returning the iterAlloc to the shared pool.
939 1 : func (i *iterAlloc) maybeAssertZeroed() {
940 1 : if invariants.Enabled {
941 1 : v := reflect.ValueOf(i).Elem()
942 1 : if err := assertZeroed(v); err != nil {
943 0 : panic(err)
944 : }
945 : }
946 : }
947 :
948 : // assertZeroed asserts that v is a "zeroed" value. A "zeroed" value is a value
949 : // that is:
950 : // - the Go zero value for its type, or
951 : // - a pointer to a "zeroed" value, or
952 : // - a slice of len()=0, with any values in the backing array being "zeroed
953 : // values", or
954 : // - a struct with all fields being "zeroed" values,
955 : // - a struct field explicitly marked with a "invariants:reused" tag.
956 1 : func assertZeroed(v reflect.Value) error {
957 1 : if v.IsZero() {
958 1 : return nil
959 1 : }
960 1 : typ := v.Type()
961 1 : switch typ.Kind() {
962 0 : case reflect.Pointer:
963 0 : return assertZeroed(v.Elem())
964 1 : case reflect.Slice:
965 1 : if v.Len() > 0 {
966 0 : return errors.AssertionFailedf("%s is not zeroed (%d len): %#v", typ.Name(), v.Len(), v)
967 0 : }
968 1 : resliced := v.Slice(0, v.Cap())
969 1 : for i := 0; i < resliced.Len(); i++ {
970 1 : if err := assertZeroed(resliced.Index(i)); err != nil {
971 0 : return errors.Wrapf(err, "[%d]", i)
972 0 : }
973 : }
974 1 : return nil
975 1 : case reflect.Struct:
976 1 : for i := 0; i < typ.NumField(); i++ {
977 1 : if typ.Field(i).Tag.Get("invariants") == "reused" {
978 1 : continue
979 : }
980 1 : if err := assertZeroed(v.Field(i)); err != nil {
981 0 : return errors.Wrapf(err, "%q", typ.Field(i).Name)
982 0 : }
983 : }
984 1 : return nil
985 : }
986 0 : return errors.AssertionFailedf("%s (%s) is not zeroed: %#v", typ.Name(), typ.Kind(), v)
987 : }
988 :
989 1 : func newIterAlloc() *iterAlloc {
990 1 : buf := iterAllocPool.Get().(*iterAlloc)
991 1 : buf.maybeAssertZeroed()
992 1 : return buf
993 1 : }
994 :
995 : var iterAllocPool = sync.Pool{
996 1 : New: func() interface{} {
997 1 : return &iterAlloc{}
998 1 : },
999 : }
1000 :
1001 : // snapshotIterOpts denotes snapshot-related iterator options when calling
1002 : // newIter. These are the possible cases for a snapshotIterOpts:
1003 : // - No snapshot: All fields are zero values.
1004 : // - Classic snapshot: Only `seqNum` is set. The latest readState will be used
1005 : // and the specified seqNum will be used as the snapshot seqNum.
1006 : // - EventuallyFileOnlySnapshot (EFOS) behaving as a classic snapshot. Only
1007 : // the `seqNum` is set. The latest readState will be used
1008 : // and the specified seqNum will be used as the snapshot seqNum.
1009 : // - EFOS in file-only state: Only `seqNum` and `vers` are set. All the
1010 : // relevant SSTs are referenced by the *version.
1011 : // - EFOS that has been excised but is in alwaysCreateIters mode (tests only).
1012 : // Only `seqNum` and `readState` are set.
1013 : type snapshotIterOpts struct {
1014 : seqNum base.SeqNum
1015 : vers *manifest.Version
1016 : readState *readState
1017 : }
1018 :
1019 : type batchIterOpts struct {
1020 : batchOnly bool
1021 : }
1022 : type newIterOpts struct {
1023 : snapshot snapshotIterOpts
1024 : batch batchIterOpts
1025 : }
1026 :
1027 : // newIter constructs a new iterator, merging in batch iterators as an extra
1028 : // level.
1029 : func (d *DB) newIter(
1030 : ctx context.Context, batch *Batch, newIterOpts newIterOpts, o *IterOptions,
1031 1 : ) *Iterator {
1032 1 : if newIterOpts.batch.batchOnly {
1033 0 : if batch == nil {
1034 0 : panic("batchOnly is true, but batch is nil")
1035 : }
1036 0 : if newIterOpts.snapshot.vers != nil {
1037 0 : panic("batchOnly is true, but snapshotIterOpts is initialized")
1038 : }
1039 : }
1040 1 : if err := d.closed.Load(); err != nil {
1041 0 : panic(err)
1042 : }
1043 1 : seqNum := newIterOpts.snapshot.seqNum
1044 1 : if o != nil && o.RangeKeyMasking.Suffix != nil && o.KeyTypes != IterKeyTypePointsAndRanges {
1045 0 : panic("pebble: range key masking requires IterKeyTypePointsAndRanges")
1046 : }
1047 1 : if (batch != nil || seqNum != 0) && (o != nil && o.OnlyReadGuaranteedDurable) {
1048 0 : // We could add support for OnlyReadGuaranteedDurable on snapshots if
1049 0 : // there was a need: this would require checking that the sequence number
1050 0 : // of the snapshot has been flushed, by comparing with
1051 0 : // DB.mem.queue[0].logSeqNum.
1052 0 : panic("OnlyReadGuaranteedDurable is not supported for batches or snapshots")
1053 : }
1054 1 : var readState *readState
1055 1 : var newIters tableNewIters
1056 1 : var newIterRangeKey keyspanimpl.TableNewSpanIter
1057 1 : if !newIterOpts.batch.batchOnly {
1058 1 : // Grab and reference the current readState. This prevents the underlying
1059 1 : // files in the associated version from being deleted if there is a current
1060 1 : // compaction. The readState is unref'd by Iterator.Close().
1061 1 : if newIterOpts.snapshot.vers == nil {
1062 1 : if newIterOpts.snapshot.readState != nil {
1063 0 : readState = newIterOpts.snapshot.readState
1064 0 : readState.ref()
1065 1 : } else {
1066 1 : // NB: loadReadState() calls readState.ref().
1067 1 : readState = d.loadReadState()
1068 1 : }
1069 1 : } else {
1070 1 : // vers != nil
1071 1 : newIterOpts.snapshot.vers.Ref()
1072 1 : }
1073 :
1074 : // Determine the seqnum to read at after grabbing the read state (current and
1075 : // memtables) above.
1076 1 : if seqNum == 0 {
1077 1 : seqNum = d.mu.versions.visibleSeqNum.Load()
1078 1 : }
1079 1 : newIters = d.newIters
1080 1 : newIterRangeKey = d.tableNewRangeKeyIter
1081 : }
1082 :
1083 : // Bundle various structures under a single umbrella in order to allocate
1084 : // them together.
1085 1 : buf := newIterAlloc()
1086 1 : dbi := &buf.dbi
1087 1 : dbi.ctx = ctx
1088 1 : dbi.alloc = buf
1089 1 : dbi.merge = d.merge
1090 1 : dbi.comparer = d.opts.Comparer
1091 1 : dbi.readState = readState
1092 1 : dbi.version = newIterOpts.snapshot.vers
1093 1 : dbi.keyBuf = buf.keyBuf
1094 1 : dbi.prefixOrFullSeekKey = buf.prefixOrFullSeekKey
1095 1 : dbi.boundsBuf = buf.boundsBuf
1096 1 : dbi.fc = d.fileCache
1097 1 : dbi.newIters = newIters
1098 1 : dbi.newIterRangeKey = newIterRangeKey
1099 1 : dbi.seqNum = seqNum
1100 1 : dbi.batchOnlyIter = newIterOpts.batch.batchOnly
1101 1 : if o != nil {
1102 1 : dbi.opts = *o
1103 1 : dbi.processBounds(o.LowerBound, o.UpperBound)
1104 1 : }
1105 1 : dbi.opts.logger = d.opts.Logger
1106 1 : if d.opts.private.disableLazyCombinedIteration {
1107 1 : dbi.opts.disableLazyCombinedIteration = true
1108 1 : }
1109 1 : if batch != nil {
1110 1 : dbi.batch = &buf.batchState
1111 1 : dbi.batch.batch = batch
1112 1 : dbi.batch.batchSeqNum = batch.nextSeqNum()
1113 1 : }
1114 1 : return finishInitializingIter(ctx, buf)
1115 : }
1116 :
1117 : // finishInitializingIter is a helper for doing the non-trivial initialization
1118 : // of an Iterator. It's invoked to perform the initial initialization of an
1119 : // Iterator during NewIter or Clone, and to perform reinitialization due to a
1120 : // change in IterOptions by a call to Iterator.SetOptions.
1121 1 : func finishInitializingIter(ctx context.Context, buf *iterAlloc) *Iterator {
1122 1 : // Short-hand.
1123 1 : dbi := &buf.dbi
1124 1 : var memtables flushableList
1125 1 : if dbi.readState != nil {
1126 1 : memtables = dbi.readState.memtables
1127 1 : }
1128 1 : if dbi.opts.OnlyReadGuaranteedDurable {
1129 0 : memtables = nil
1130 1 : } else {
1131 1 : // We only need to read from memtables which contain sequence numbers older
1132 1 : // than seqNum. Trim off newer memtables.
1133 1 : for i := len(memtables) - 1; i >= 0; i-- {
1134 1 : if logSeqNum := memtables[i].logSeqNum; logSeqNum < dbi.seqNum {
1135 1 : break
1136 : }
1137 1 : memtables = memtables[:i]
1138 : }
1139 : }
1140 :
1141 1 : if dbi.opts.pointKeys() {
1142 1 : // Construct the point iterator, initializing dbi.pointIter to point to
1143 1 : // dbi.merging. If this is called during a SetOptions call and this
1144 1 : // Iterator has already initialized dbi.merging, constructPointIter is a
1145 1 : // noop and an initialized pointIter already exists in dbi.pointIter.
1146 1 : dbi.constructPointIter(ctx, memtables, buf)
1147 1 : dbi.iter = dbi.pointIter
1148 1 : } else {
1149 1 : dbi.iter = emptyIter
1150 1 : }
1151 :
1152 1 : if dbi.opts.rangeKeys() {
1153 1 : dbi.rangeKeyMasking.init(dbi, dbi.comparer)
1154 1 :
1155 1 : // When iterating over both point and range keys, don't create the
1156 1 : // range-key iterator stack immediately if we can avoid it. This
1157 1 : // optimization takes advantage of the expected sparseness of range
1158 1 : // keys, and configures the point-key iterator to dynamically switch to
1159 1 : // combined iteration when it observes a file containing range keys.
1160 1 : //
1161 1 : // Lazy combined iteration is not possible if a batch or a memtable
1162 1 : // contains any range keys.
1163 1 : useLazyCombinedIteration := dbi.rangeKey == nil &&
1164 1 : dbi.opts.KeyTypes == IterKeyTypePointsAndRanges &&
1165 1 : (dbi.batch == nil || dbi.batch.batch.countRangeKeys == 0) &&
1166 1 : !dbi.opts.disableLazyCombinedIteration
1167 1 : if useLazyCombinedIteration {
1168 1 : // The user requested combined iteration, and there's no indexed
1169 1 : // batch currently containing range keys that would prevent lazy
1170 1 : // combined iteration. Check the memtables to see if they contain
1171 1 : // any range keys.
1172 1 : for i := range memtables {
1173 1 : if memtables[i].containsRangeKeys() {
1174 1 : useLazyCombinedIteration = false
1175 1 : break
1176 : }
1177 : }
1178 : }
1179 :
1180 1 : if useLazyCombinedIteration {
1181 1 : dbi.lazyCombinedIter = lazyCombinedIter{
1182 1 : parent: dbi,
1183 1 : pointIter: dbi.pointIter,
1184 1 : combinedIterState: combinedIterState{
1185 1 : initialized: false,
1186 1 : },
1187 1 : }
1188 1 : dbi.iter = &dbi.lazyCombinedIter
1189 1 : dbi.iter = invalidating.MaybeWrapIfInvariants(dbi.iter)
1190 1 : } else {
1191 1 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{
1192 1 : initialized: true,
1193 1 : }
1194 1 : if dbi.rangeKey == nil {
1195 1 : dbi.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
1196 1 : dbi.constructRangeKeyIter()
1197 1 : } else {
1198 1 : dbi.rangeKey.iterConfig.SetBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
1199 1 : }
1200 :
1201 : // Wrap the point iterator (currently dbi.iter) with an interleaving
1202 : // iterator that interleaves range keys pulled from
1203 : // dbi.rangeKey.rangeKeyIter.
1204 : //
1205 : // NB: The interleaving iterator is always reinitialized, even if
1206 : // dbi already had an initialized range key iterator, in case the point
1207 : // iterator changed or the range key masking suffix changed.
1208 1 : dbi.rangeKey.iiter.Init(dbi.comparer, dbi.iter, dbi.rangeKey.rangeKeyIter,
1209 1 : keyspan.InterleavingIterOpts{
1210 1 : Mask: &dbi.rangeKeyMasking,
1211 1 : LowerBound: dbi.opts.LowerBound,
1212 1 : UpperBound: dbi.opts.UpperBound,
1213 1 : })
1214 1 : dbi.iter = &dbi.rangeKey.iiter
1215 : }
1216 1 : } else {
1217 1 : // !dbi.opts.rangeKeys()
1218 1 : //
1219 1 : // Reset the combined iterator state. The initialized=true ensures the
1220 1 : // iterator doesn't unnecessarily try to switch to combined iteration.
1221 1 : dbi.lazyCombinedIter.combinedIterState = combinedIterState{initialized: true}
1222 1 : }
1223 1 : return dbi
1224 : }
1225 :
1226 : func (i *Iterator) constructPointIter(
1227 : ctx context.Context, memtables flushableList, buf *iterAlloc,
1228 1 : ) {
1229 1 : if i.pointIter != nil {
1230 1 : // Already have one.
1231 1 : return
1232 1 : }
1233 1 : readEnv := block.ReadEnv{
1234 1 : Stats: &i.stats.InternalStats,
1235 1 : // If the file cache has a sstable stats collector, ask it for an
1236 1 : // accumulator for this iterator's configured category and QoS. All SSTable
1237 1 : // iterators created by this Iterator will accumulate their stats to it as
1238 1 : // they Close during iteration.
1239 1 : IterStats: i.fc.SSTStatsCollector().Accumulator(
1240 1 : uint64(uintptr(unsafe.Pointer(i))),
1241 1 : i.opts.Category,
1242 1 : ),
1243 1 : }
1244 1 : if i.readState != nil {
1245 1 : i.blobValueFetcher.Init(&i.readState.current.BlobFiles, i.fc, readEnv,
1246 1 : blob.SuggestedCachedReaders(i.readState.current.MaxReadAmp()))
1247 1 : } else if i.version != nil {
1248 1 : i.blobValueFetcher.Init(&i.version.BlobFiles, i.fc, readEnv,
1249 1 : blob.SuggestedCachedReaders(i.version.MaxReadAmp()))
1250 1 : }
1251 1 : internalOpts := internalIterOpts{
1252 1 : readEnv: sstable.ReadEnv{Block: readEnv},
1253 1 : blobValueFetcher: &i.blobValueFetcher,
1254 1 : }
1255 1 : if i.opts.RangeKeyMasking.Filter != nil {
1256 1 : internalOpts.boundLimitedFilter = &i.rangeKeyMasking
1257 1 : }
1258 :
1259 : // Merging levels and levels from iterAlloc.
1260 1 : mlevels := buf.mlevels[:0]
1261 1 : levels := buf.levels[:0]
1262 1 :
1263 1 : // We compute the number of levels needed ahead of time and reallocate a slice if
1264 1 : // the array from the iterAlloc isn't large enough. Doing this allocation once
1265 1 : // should improve the performance.
1266 1 : numMergingLevels := 0
1267 1 : numLevelIters := 0
1268 1 : if i.batch != nil {
1269 1 : numMergingLevels++
1270 1 : }
1271 :
1272 1 : var current *manifest.Version
1273 1 : if !i.batchOnlyIter {
1274 1 : numMergingLevels += len(memtables)
1275 1 : current = i.version
1276 1 : if current == nil {
1277 1 : current = i.readState.current
1278 1 : }
1279 1 : maxReadAmp := current.MaxReadAmp()
1280 1 : numMergingLevels += maxReadAmp
1281 1 : numLevelIters += maxReadAmp
1282 : }
1283 :
1284 1 : if numMergingLevels > cap(mlevels) {
1285 1 : mlevels = make([]mergingIterLevel, 0, numMergingLevels)
1286 1 : }
1287 1 : if numLevelIters > cap(levels) {
1288 1 : levels = make([]levelIter, 0, numLevelIters)
1289 1 : }
1290 :
1291 : // Top-level is the batch, if any.
1292 1 : if i.batch != nil {
1293 1 : if i.batch.batch.index == nil {
1294 0 : // This isn't an indexed batch. We shouldn't have gotten this far.
1295 0 : panic(errors.AssertionFailedf("creating an iterator over an unindexed batch"))
1296 1 : } else {
1297 1 : i.batch.batch.initInternalIter(&i.opts, &i.batch.pointIter)
1298 1 : i.batch.batch.initRangeDelIter(&i.opts, &i.batch.rangeDelIter, i.batch.batchSeqNum)
1299 1 : // Only include the batch's rangedel iterator if it's non-empty.
1300 1 : // This requires some subtle logic in the case a rangedel is later
1301 1 : // written to the batch and the view of the batch is refreshed
1302 1 : // during a call to SetOptions—in this case, we need to reconstruct
1303 1 : // the point iterator to add the batch rangedel iterator.
1304 1 : var rangeDelIter keyspan.FragmentIterator
1305 1 : if i.batch.rangeDelIter.Count() > 0 {
1306 1 : rangeDelIter = &i.batch.rangeDelIter
1307 1 : }
1308 1 : mlevels = append(mlevels, mergingIterLevel{
1309 1 : iter: &i.batch.pointIter,
1310 1 : rangeDelIter: rangeDelIter,
1311 1 : })
1312 : }
1313 : }
1314 :
1315 1 : if !i.batchOnlyIter {
1316 1 : // Next are the memtables.
1317 1 : for j := len(memtables) - 1; j >= 0; j-- {
1318 1 : mem := memtables[j]
1319 1 : mlevels = append(mlevels, mergingIterLevel{
1320 1 : iter: mem.newIter(&i.opts),
1321 1 : rangeDelIter: mem.newRangeDelIter(&i.opts),
1322 1 : })
1323 1 : }
1324 :
1325 : // Next are the file levels: L0 sub-levels followed by lower levels.
1326 1 : mlevelsIndex := len(mlevels)
1327 1 : levelsIndex := len(levels)
1328 1 : mlevels = mlevels[:numMergingLevels]
1329 1 : levels = levels[:numLevelIters]
1330 1 : i.opts.snapshotForHideObsoletePoints = buf.dbi.seqNum
1331 1 : addLevelIterForFiles := func(files manifest.LevelIterator, level manifest.Layer) {
1332 1 : li := &levels[levelsIndex]
1333 1 :
1334 1 : li.init(ctx, i.opts, i.comparer, i.newIters, files, level, internalOpts)
1335 1 : li.initRangeDel(&mlevels[mlevelsIndex])
1336 1 : li.initCombinedIterState(&i.lazyCombinedIter.combinedIterState)
1337 1 : mlevels[mlevelsIndex].levelIter = li
1338 1 : mlevels[mlevelsIndex].iter = invalidating.MaybeWrapIfInvariants(li)
1339 1 :
1340 1 : levelsIndex++
1341 1 : mlevelsIndex++
1342 1 : }
1343 :
1344 : // Add level iterators for the L0 sublevels, iterating from newest to
1345 : // oldest.
1346 1 : for i := len(current.L0SublevelFiles) - 1; i >= 0; i-- {
1347 1 : addLevelIterForFiles(current.L0SublevelFiles[i].Iter(), manifest.L0Sublevel(i))
1348 1 : }
1349 :
1350 : // Add level iterators for the non-empty non-L0 levels.
1351 1 : for level := 1; level < len(current.Levels); level++ {
1352 1 : if current.Levels[level].Empty() {
1353 1 : continue
1354 : }
1355 1 : addLevelIterForFiles(current.Levels[level].Iter(), manifest.Level(level))
1356 : }
1357 : }
1358 1 : buf.merging.init(&i.opts, &i.stats.InternalStats, i.comparer.Compare, i.comparer.Split, mlevels...)
1359 1 : if len(mlevels) <= cap(buf.levelsPositioned) {
1360 1 : buf.merging.levelsPositioned = buf.levelsPositioned[:len(mlevels)]
1361 1 : }
1362 1 : buf.merging.snapshot = i.seqNum
1363 1 : if i.batch != nil {
1364 1 : buf.merging.batchSnapshot = i.batch.batchSeqNum
1365 1 : }
1366 1 : buf.merging.combinedIterState = &i.lazyCombinedIter.combinedIterState
1367 1 : i.pointIter = invalidating.MaybeWrapIfInvariants(&buf.merging).(topLevelIterator)
1368 1 : i.merging = &buf.merging
1369 : }
1370 :
1371 : // NewBatch returns a new empty write-only batch. Any reads on the batch will
1372 : // return an error. If the batch is committed it will be applied to the DB.
1373 1 : func (d *DB) NewBatch(opts ...BatchOption) *Batch {
1374 1 : return newBatch(d, opts...)
1375 1 : }
1376 :
1377 : // NewBatchWithSize is mostly identical to NewBatch, but it will allocate the
1378 : // the specified memory space for the internal slice in advance.
1379 0 : func (d *DB) NewBatchWithSize(size int, opts ...BatchOption) *Batch {
1380 0 : return newBatchWithSize(d, size, opts...)
1381 0 : }
1382 :
1383 : // NewIndexedBatch returns a new empty read-write batch. Any reads on the batch
1384 : // will read from both the batch and the DB. If the batch is committed it will
1385 : // be applied to the DB. An indexed batch is slower that a non-indexed batch
1386 : // for insert operations. If you do not need to perform reads on the batch, use
1387 : // NewBatch instead.
1388 1 : func (d *DB) NewIndexedBatch() *Batch {
1389 1 : return newIndexedBatch(d, d.opts.Comparer)
1390 1 : }
1391 :
1392 : // NewIndexedBatchWithSize is mostly identical to NewIndexedBatch, but it will
1393 : // allocate the specified memory space for the internal slice in advance.
1394 0 : func (d *DB) NewIndexedBatchWithSize(size int) *Batch {
1395 0 : return newIndexedBatchWithSize(d, d.opts.Comparer, size)
1396 0 : }
1397 :
1398 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
1399 : // return false). The iterator can be positioned via a call to SeekGE, SeekLT,
1400 : // First or Last. The iterator provides a point-in-time view of the current DB
1401 : // state. This view is maintained by preventing file deletions and preventing
1402 : // memtables referenced by the iterator from being deleted. Using an iterator
1403 : // to maintain a long-lived point-in-time view of the DB state can lead to an
1404 : // apparent memory and disk usage leak. Use snapshots (see NewSnapshot) for
1405 : // point-in-time snapshots which avoids these problems.
1406 1 : func (d *DB) NewIter(o *IterOptions) (*Iterator, error) {
1407 1 : return d.NewIterWithContext(context.Background(), o)
1408 1 : }
1409 :
1410 : // NewIterWithContext is like NewIter, and additionally accepts a context for
1411 : // tracing.
1412 1 : func (d *DB) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
1413 1 : return d.newIter(ctx, nil /* batch */, newIterOpts{}, o), nil
1414 1 : }
1415 :
1416 : // NewSnapshot returns a point-in-time view of the current DB state. Iterators
1417 : // created with this handle will all observe a stable snapshot of the current
1418 : // DB state. The caller must call Snapshot.Close() when the snapshot is no
1419 : // longer needed. Snapshots are not persisted across DB restarts (close ->
1420 : // open). Unlike the implicit snapshot maintained by an iterator, a snapshot
1421 : // will not prevent memtables from being released or sstables from being
1422 : // deleted. Instead, a snapshot prevents deletion of sequence numbers
1423 : // referenced by the snapshot.
1424 : //
1425 : // There exists one violation of a Snapshot's point-in-time guarantee: An excise
1426 : // (see DB.Excise and DB.IngestAndExcise) that occurs after the snapshot's
1427 : // creation will be observed by iterators created from the snapshot after the
1428 : // excise. See NewEventuallyFileOnlySnapshot for a variant of NewSnapshot that
1429 : // provides a full point-in-time guarantee.
1430 1 : func (d *DB) NewSnapshot() *Snapshot {
1431 1 : // TODO(jackson): Consider removal of regular, non-eventually-file-only
1432 1 : // snapshots given they no longer provide a true point-in-time snapshot of
1433 1 : // the database due to excises. If we had a mechanism to construct a maximal
1434 1 : // key range, we could implement NewSnapshot in terms of
1435 1 : // NewEventuallyFileOnlySnapshot and provide a true point-in-time guarantee.
1436 1 : if err := d.closed.Load(); err != nil {
1437 0 : panic(err)
1438 : }
1439 1 : d.mu.Lock()
1440 1 : s := &Snapshot{
1441 1 : db: d,
1442 1 : seqNum: d.mu.versions.visibleSeqNum.Load(),
1443 1 : }
1444 1 : d.mu.snapshots.pushBack(s)
1445 1 : d.mu.Unlock()
1446 1 : return s
1447 : }
1448 :
1449 : // NewEventuallyFileOnlySnapshot returns a point-in-time view of the current DB
1450 : // state, similar to NewSnapshot, but with consistency constrained to the
1451 : // provided set of key ranges. See the comment at EventuallyFileOnlySnapshot for
1452 : // its semantics.
1453 1 : func (d *DB) NewEventuallyFileOnlySnapshot(keyRanges []KeyRange) *EventuallyFileOnlySnapshot {
1454 1 : if err := d.closed.Load(); err != nil {
1455 0 : panic(err)
1456 : }
1457 1 : for i := range keyRanges {
1458 1 : if i > 0 && d.cmp(keyRanges[i-1].End, keyRanges[i].Start) > 0 {
1459 0 : panic("pebble: key ranges for eventually-file-only-snapshot not in order")
1460 : }
1461 : }
1462 1 : return d.makeEventuallyFileOnlySnapshot(keyRanges)
1463 : }
1464 :
1465 : // Close closes the DB.
1466 : //
1467 : // It is not safe to close a DB until all outstanding iterators are closed
1468 : // or to call Close concurrently with any other DB method. It is not valid
1469 : // to call any of a DB's methods after the DB has been closed.
1470 1 : func (d *DB) Close() error {
1471 1 : if err := d.closed.Load(); err != nil {
1472 0 : panic(err)
1473 : }
1474 1 : d.compactionSchedulers.Wait()
1475 1 : // Compactions can be asynchronously started by the CompactionScheduler
1476 1 : // calling d.Schedule. When this Unregister returns, we know that the
1477 1 : // CompactionScheduler will never again call a method on the DB. Note that
1478 1 : // this must be called without holding d.mu.
1479 1 : d.compactionScheduler.Unregister()
1480 1 : // Lock the commit pipeline for the duration of Close. This prevents a race
1481 1 : // with makeRoomForWrite. Rotating the WAL in makeRoomForWrite requires
1482 1 : // dropping d.mu several times for I/O. If Close only holds d.mu, an
1483 1 : // in-progress WAL rotation may re-acquire d.mu only once the database is
1484 1 : // closed.
1485 1 : //
1486 1 : // Additionally, locking the commit pipeline makes it more likely that
1487 1 : // (illegal) concurrent writes will observe d.closed.Load() != nil, creating
1488 1 : // more understable panics if the database is improperly used concurrently
1489 1 : // during Close.
1490 1 : d.commit.mu.Lock()
1491 1 : defer d.commit.mu.Unlock()
1492 1 : d.mu.Lock()
1493 1 : defer d.mu.Unlock()
1494 1 : // Check that the DB is not closed again. If there are two concurrent calls
1495 1 : // to DB.Close, the best-effort check at the top of DB.Close may not fire.
1496 1 : // But since this second check happens after mutex acquisition, the two
1497 1 : // concurrent calls will get serialized and the second one will see the
1498 1 : // effect of the d.closed.Store below.
1499 1 : if err := d.closed.Load(); err != nil {
1500 0 : panic(err)
1501 : }
1502 : // Clear the finalizer that is used to check that an unreferenced DB has been
1503 : // closed. We're closing the DB here, so the check performed by that
1504 : // finalizer isn't necessary.
1505 : //
1506 : // Note: this is a no-op if invariants are disabled or race is enabled.
1507 1 : invariants.SetFinalizer(d.closed, nil)
1508 1 :
1509 1 : d.closed.Store(errors.WithStack(ErrClosed))
1510 1 : close(d.closedCh)
1511 1 :
1512 1 : defer d.cacheHandle.Close()
1513 1 :
1514 1 : for d.mu.compact.compactingCount > 0 || d.mu.compact.downloadingCount > 0 || d.mu.compact.flushing {
1515 1 : d.mu.compact.cond.Wait()
1516 1 : }
1517 1 : for d.mu.tableStats.loading {
1518 1 : d.mu.tableStats.cond.Wait()
1519 1 : }
1520 1 : for d.mu.tableValidation.validating {
1521 1 : d.mu.tableValidation.cond.Wait()
1522 1 : }
1523 :
1524 1 : var err error
1525 1 : if n := len(d.mu.compact.inProgress); n > 0 {
1526 0 : err = errors.Errorf("pebble: %d unexpected in-progress compactions", errors.Safe(n))
1527 0 : }
1528 1 : err = firstError(err, d.mu.formatVers.marker.Close())
1529 1 : if !d.opts.ReadOnly {
1530 1 : if d.mu.log.writer != nil {
1531 1 : _, err2 := d.mu.log.writer.Close()
1532 1 : err = firstError(err, err2)
1533 1 : }
1534 0 : } else if d.mu.log.writer != nil {
1535 0 : panic("pebble: log-writer should be nil in read-only mode")
1536 : }
1537 1 : err = firstError(err, d.mu.log.manager.Close())
1538 1 : err = firstError(err, d.dataDirLock.Close())
1539 1 :
1540 1 : // Note that versionSet.close() only closes the MANIFEST. The versions list
1541 1 : // is still valid for the checks below.
1542 1 : err = firstError(err, d.mu.versions.close())
1543 1 :
1544 1 : err = firstError(err, d.dataDir.Close())
1545 1 :
1546 1 : d.readState.val.unrefLocked()
1547 1 :
1548 1 : current := d.mu.versions.currentVersion()
1549 1 : for v := d.mu.versions.versions.Front(); true; v = v.Next() {
1550 1 : refs := v.Refs()
1551 1 : if v == current {
1552 1 : if refs != 1 {
1553 0 : err = firstError(err, errors.Errorf("leaked iterators: current\n%s", v))
1554 0 : }
1555 1 : break
1556 : }
1557 0 : if refs != 0 {
1558 0 : err = firstError(err, errors.Errorf("leaked iterators:\n%s", v))
1559 0 : }
1560 : }
1561 :
1562 1 : for _, mem := range d.mu.mem.queue {
1563 1 : // Usually, we'd want to delete the files returned by readerUnref. But
1564 1 : // in this case, even if we're unreferencing the flushables, the
1565 1 : // flushables aren't obsolete. They will be reconstructed during WAL
1566 1 : // replay.
1567 1 : mem.readerUnrefLocked(false)
1568 1 : }
1569 : // If there's an unused, recycled memtable, we need to release its memory.
1570 1 : if obsoleteMemTable := d.memTableRecycle.Swap(nil); obsoleteMemTable != nil {
1571 1 : d.freeMemTable(obsoleteMemTable)
1572 1 : }
1573 1 : if reserved := d.memTableReserved.Load(); reserved != 0 {
1574 0 : err = firstError(err, errors.Errorf("leaked memtable reservation: %d", errors.Safe(reserved)))
1575 0 : }
1576 :
1577 : // Since we called d.readState.val.unrefLocked() above, we are expected to
1578 : // manually schedule deletion of obsolete files.
1579 1 : if len(d.mu.versions.obsoleteTables) > 0 || len(d.mu.versions.obsoleteBlobs) > 0 {
1580 1 : d.deleteObsoleteFiles(d.newJobIDLocked())
1581 1 : }
1582 :
1583 1 : d.mu.Unlock()
1584 1 :
1585 1 : // Wait for all cleaning jobs to finish.
1586 1 : d.cleanupManager.Close()
1587 1 :
1588 1 : d.mu.Lock()
1589 1 : // Sanity check compaction metrics.
1590 1 : if invariants.Enabled {
1591 1 : if d.mu.compact.compactingCount > 0 || d.mu.compact.downloadingCount > 0 || d.mu.versions.atomicInProgressBytes.Load() > 0 {
1592 0 : panic("compacting counts not 0 on close")
1593 : }
1594 : }
1595 :
1596 : // As a sanity check, ensure that there are no zombie tables or blob files.
1597 : // A non-zero count hints at a reference count leak.
1598 1 : if ztbls := d.mu.versions.zombieTables.Count(); ztbls > 0 {
1599 0 : err = firstError(err, errors.Errorf("non-zero zombie file count: %d", ztbls))
1600 0 : }
1601 1 : if zblobs := d.mu.versions.zombieBlobs.Count(); zblobs > 0 {
1602 0 : err = firstError(err, errors.Errorf("non-zero zombie blob count: %d", zblobs))
1603 0 : }
1604 :
1605 1 : err = firstError(err, d.fileCache.Close())
1606 1 :
1607 1 : err = firstError(err, d.objProvider.Close())
1608 1 :
1609 1 : // If the options include a closer to 'close' the filesystem, close it.
1610 1 : if d.opts.private.fsCloser != nil {
1611 1 : d.opts.private.fsCloser.Close()
1612 1 : }
1613 :
1614 : // Return an error if the user failed to close all open snapshots.
1615 1 : if v := d.mu.snapshots.count(); v > 0 {
1616 0 : err = firstError(err, errors.Errorf("leaked snapshots: %d open snapshots on DB %p", v, d))
1617 0 : }
1618 :
1619 1 : return err
1620 : }
1621 :
1622 : // Compact the specified range of keys in the database.
1623 1 : func (d *DB) Compact(ctx context.Context, start, end []byte, parallelize bool) error {
1624 1 : if err := d.closed.Load(); err != nil {
1625 0 : panic(err)
1626 : }
1627 1 : if d.opts.ReadOnly {
1628 0 : return ErrReadOnly
1629 0 : }
1630 1 : if d.cmp(start, end) >= 0 {
1631 1 : return errors.Errorf("Compact start %s is not less than end %s",
1632 1 : d.opts.Comparer.FormatKey(start), d.opts.Comparer.FormatKey(end))
1633 1 : }
1634 :
1635 1 : d.mu.Lock()
1636 1 : maxLevelWithFiles := 1
1637 1 : cur := d.mu.versions.currentVersion()
1638 1 : for level := 0; level < numLevels; level++ {
1639 1 : overlaps := cur.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1640 1 : if !overlaps.Empty() {
1641 1 : maxLevelWithFiles = level + 1
1642 1 : }
1643 : }
1644 :
1645 : // Determine if any memtable overlaps with the compaction range. We wait for
1646 : // any such overlap to flush (initiating a flush if necessary).
1647 1 : mem, err := func() (*flushableEntry, error) {
1648 1 : // Check to see if any files overlap with any of the memtables. The queue
1649 1 : // is ordered from oldest to newest with the mutable memtable being the
1650 1 : // last element in the slice. We want to wait for the newest table that
1651 1 : // overlaps.
1652 1 : for i := len(d.mu.mem.queue) - 1; i >= 0; i-- {
1653 1 : mem := d.mu.mem.queue[i]
1654 1 : var anyOverlaps bool
1655 1 : mem.computePossibleOverlaps(func(b bounded) shouldContinue {
1656 1 : anyOverlaps = true
1657 1 : return stopIteration
1658 1 : }, KeyRange{Start: start, End: end})
1659 1 : if !anyOverlaps {
1660 1 : continue
1661 : }
1662 1 : var err error
1663 1 : if mem.flushable == d.mu.mem.mutable {
1664 1 : // We have to hold both commitPipeline.mu and DB.mu when calling
1665 1 : // makeRoomForWrite(). Lock order requirements elsewhere force us to
1666 1 : // unlock DB.mu in order to grab commitPipeline.mu first.
1667 1 : d.mu.Unlock()
1668 1 : d.commit.mu.Lock()
1669 1 : d.mu.Lock()
1670 1 : defer d.commit.mu.Unlock() //nolint:deferloop
1671 1 : if mem.flushable == d.mu.mem.mutable {
1672 1 : // Only flush if the active memtable is unchanged.
1673 1 : err = d.makeRoomForWrite(nil)
1674 1 : }
1675 : }
1676 1 : mem.flushForced = true
1677 1 : d.maybeScheduleFlush()
1678 1 : return mem, err
1679 : }
1680 1 : return nil, nil
1681 : }()
1682 :
1683 1 : d.mu.Unlock()
1684 1 :
1685 1 : if err != nil {
1686 0 : return err
1687 0 : }
1688 1 : if mem != nil {
1689 1 : select {
1690 1 : case <-mem.flushed:
1691 0 : case <-ctx.Done():
1692 0 : return ctx.Err()
1693 : }
1694 : }
1695 :
1696 1 : for level := 0; level < maxLevelWithFiles; {
1697 1 : for {
1698 1 : if err := d.manualCompact(
1699 1 : ctx, start, end, level, parallelize); err != nil {
1700 0 : if errors.Is(err, ErrCancelledCompaction) {
1701 0 : continue
1702 : }
1703 0 : return err
1704 : }
1705 1 : break
1706 : }
1707 1 : level++
1708 1 : if level == numLevels-1 {
1709 1 : // A manual compaction of the bottommost level occurred.
1710 1 : // There is no next level to try and compact.
1711 1 : break
1712 : }
1713 : }
1714 1 : return nil
1715 : }
1716 :
1717 : func (d *DB) manualCompact(
1718 : ctx context.Context, start, end []byte, level int, parallelize bool,
1719 1 : ) error {
1720 1 : d.mu.Lock()
1721 1 : curr := d.mu.versions.currentVersion()
1722 1 : files := curr.Overlaps(level, base.UserKeyBoundsInclusive(start, end))
1723 1 : if files.Empty() {
1724 1 : d.mu.Unlock()
1725 1 : return nil
1726 1 : }
1727 :
1728 1 : var compactions []*manualCompaction
1729 1 : if parallelize {
1730 1 : compactions = append(compactions, d.splitManualCompaction(start, end, level)...)
1731 1 : } else {
1732 1 : compactions = append(compactions, &manualCompaction{
1733 1 : level: level,
1734 1 : done: make(chan error, 1),
1735 1 : start: start,
1736 1 : end: end,
1737 1 : })
1738 1 : }
1739 1 : n := len(compactions)
1740 1 : if n == 0 {
1741 0 : d.mu.Unlock()
1742 0 : return nil
1743 0 : }
1744 1 : for i := range compactions {
1745 1 : d.mu.compact.manualID++
1746 1 : compactions[i].id = d.mu.compact.manualID
1747 1 : }
1748 : // [manualIDStart, manualIDEnd] are the compactions that have been added to
1749 : // d.mu.compact.manual.
1750 1 : manualIDStart := compactions[0].id
1751 1 : manualIDEnd := compactions[n-1].id
1752 1 : d.mu.compact.manual = append(d.mu.compact.manual, compactions...)
1753 1 : d.mu.compact.manualLen.Store(int32(len(d.mu.compact.manual)))
1754 1 : d.maybeScheduleCompaction()
1755 1 : d.mu.Unlock()
1756 1 :
1757 1 : // On context cancellation, we only cancel the compactions that have not yet
1758 1 : // started. The assumption is that it is relatively harmless to have the
1759 1 : // already started compactions run to completion. We don't wait for the
1760 1 : // ongoing compactions to finish, since the assumption is that the caller
1761 1 : // has already given up on the operation (and the cancellation error is
1762 1 : // going to be returned anyway).
1763 1 : //
1764 1 : // An alternative would be to store the context in each *manualCompaction,
1765 1 : // and have the goroutine that retrieves the *manualCompaction for running
1766 1 : // notice the cancellation and write the cancellation error to
1767 1 : // manualCompaction.done. That approach would require this method to wait
1768 1 : // for all the *manualCompactions it has enqueued to finish before returning
1769 1 : // (to not leak a context). Since there is no timeliness guarantee on when a
1770 1 : // *manualCompaction will be retrieved for running, the wait until a
1771 1 : // cancelled context causes this method to return is not bounded. Hence, we
1772 1 : // don't adopt that approach.
1773 1 : cancelPendingCompactions := func() {
1774 0 : d.mu.Lock()
1775 0 : for i := 0; i < len(d.mu.compact.manual); {
1776 0 : if d.mu.compact.manual[i].id >= manualIDStart && d.mu.compact.manual[i].id <= manualIDEnd {
1777 0 : d.mu.compact.manual = slices.Delete(d.mu.compact.manual, i, i+1)
1778 0 : d.mu.compact.manualLen.Store(int32(len(d.mu.compact.manual)))
1779 0 : } else {
1780 0 : i++
1781 0 : }
1782 : }
1783 0 : d.mu.Unlock()
1784 : }
1785 : // Each of the channels is guaranteed to be eventually sent to once. After a
1786 : // compaction is possibly picked in d.maybeScheduleCompaction(), either the
1787 : // compaction is dropped, executed after being scheduled, or retried later.
1788 : // Assuming eventual progress when a compaction is retried, all outcomes send
1789 : // a value to the done channel. Since the channels are buffered, it is not
1790 : // necessary to read from each channel, and so we can exit early in the event
1791 : // of an error.
1792 1 : for _, compaction := range compactions {
1793 1 : select {
1794 0 : case <-ctx.Done():
1795 0 : cancelPendingCompactions()
1796 0 : return ctx.Err()
1797 1 : case err := <-compaction.done:
1798 1 : if err != nil {
1799 0 : cancelPendingCompactions()
1800 0 : return err
1801 0 : }
1802 : }
1803 : }
1804 1 : return nil
1805 : }
1806 :
1807 : // splitManualCompaction splits a manual compaction over [start,end] on level
1808 : // such that the resulting compactions have no key overlap.
1809 : func (d *DB) splitManualCompaction(
1810 : start, end []byte, level int,
1811 1 : ) (splitCompactions []*manualCompaction) {
1812 1 : curr := d.mu.versions.currentVersion()
1813 1 : endLevel := level + 1
1814 1 : baseLevel := d.mu.versions.picker.getBaseLevel()
1815 1 : if level == 0 {
1816 1 : endLevel = baseLevel
1817 1 : }
1818 1 : keyRanges := curr.CalculateInuseKeyRanges(d.mu.versions.latest.l0Organizer, level, endLevel, start, end)
1819 1 : for _, keyRange := range keyRanges {
1820 1 : splitCompactions = append(splitCompactions, &manualCompaction{
1821 1 : level: level,
1822 1 : done: make(chan error, 1),
1823 1 : start: keyRange.Start,
1824 1 : end: keyRange.End.Key,
1825 1 : split: true,
1826 1 : })
1827 1 : }
1828 1 : return splitCompactions
1829 : }
1830 :
1831 : // Flush the memtable to stable storage.
1832 1 : func (d *DB) Flush() error {
1833 1 : flushDone, err := d.AsyncFlush()
1834 1 : if err != nil {
1835 0 : return err
1836 0 : }
1837 1 : <-flushDone
1838 1 : return nil
1839 : }
1840 :
1841 : // AsyncFlush asynchronously flushes the memtable to stable storage.
1842 : //
1843 : // If no error is returned, the caller can receive from the returned channel in
1844 : // order to wait for the flush to complete.
1845 1 : func (d *DB) AsyncFlush() (<-chan struct{}, error) {
1846 1 : if err := d.closed.Load(); err != nil {
1847 0 : panic(err)
1848 : }
1849 1 : if d.opts.ReadOnly {
1850 0 : return nil, ErrReadOnly
1851 0 : }
1852 :
1853 1 : d.commit.mu.Lock()
1854 1 : defer d.commit.mu.Unlock()
1855 1 : d.mu.Lock()
1856 1 : defer d.mu.Unlock()
1857 1 : flushed := d.mu.mem.queue[len(d.mu.mem.queue)-1].flushed
1858 1 : err := d.makeRoomForWrite(nil)
1859 1 : if err != nil {
1860 0 : return nil, err
1861 0 : }
1862 1 : return flushed, nil
1863 : }
1864 :
1865 : // Metrics returns metrics about the database.
1866 0 : func (d *DB) Metrics() *Metrics {
1867 0 : metrics := &Metrics{}
1868 0 : walStats := d.mu.log.manager.Stats()
1869 0 : completedObsoleteFileStats := d.cleanupManager.CompletedStats()
1870 0 :
1871 0 : d.mu.Lock()
1872 0 : vers := d.mu.versions.currentVersion()
1873 0 : vers.Ref()
1874 0 : defer vers.Unref()
1875 0 :
1876 0 : *metrics = d.mu.versions.metrics
1877 0 : metrics.Compact.EstimatedDebt = d.mu.versions.picker.estimatedCompactionDebt()
1878 0 : metrics.Compact.InProgressBytes = d.mu.versions.atomicInProgressBytes.Load()
1879 0 : // TODO(radu): split this to separate the download compactions.
1880 0 : metrics.Compact.NumInProgress = int64(d.mu.compact.compactingCount + d.mu.compact.downloadingCount)
1881 0 : metrics.Compact.MarkedFiles = vers.MarkedForCompaction.Count()
1882 0 : metrics.Compact.Duration = d.mu.compact.duration
1883 0 : for c := range d.mu.compact.inProgress {
1884 0 : if !c.IsFlush() {
1885 0 : metrics.Compact.Duration += d.timeNow().Sub(c.BeganAt())
1886 0 : }
1887 : }
1888 0 : metrics.Compact.NumProblemSpans = d.problemSpans.Len()
1889 0 :
1890 0 : for _, m := range d.mu.mem.queue {
1891 0 : metrics.MemTable.Size += m.totalBytes()
1892 0 : }
1893 0 : metrics.Snapshots.Count = d.mu.snapshots.count()
1894 0 : if metrics.Snapshots.Count > 0 {
1895 0 : metrics.Snapshots.EarliestSeqNum = d.mu.snapshots.earliest()
1896 0 : }
1897 0 : metrics.Snapshots.PinnedKeys = d.mu.snapshots.cumulativePinnedCount
1898 0 : metrics.Snapshots.PinnedSize = d.mu.snapshots.cumulativePinnedSize
1899 0 : metrics.MemTable.Count = int64(len(d.mu.mem.queue))
1900 0 : metrics.MemTable.ZombieCount = d.memTableCount.Load() - metrics.MemTable.Count
1901 0 : metrics.MemTable.ZombieSize = uint64(d.memTableReserved.Load()) - metrics.MemTable.Size
1902 0 : metrics.WAL.ObsoleteFiles = int64(walStats.ObsoleteFileCount)
1903 0 : metrics.WAL.ObsoletePhysicalSize = walStats.ObsoleteFileSize
1904 0 : metrics.WAL.Files = int64(walStats.LiveFileCount)
1905 0 : // The current WAL's size (d.logSize) is the logical size, which may be less
1906 0 : // than the WAL's physical size if it was recycled. walStats.LiveFileSize
1907 0 : // includes the physical size of all live WALs, but for the current WAL it
1908 0 : // reflects the physical size when it was opened. So it is possible that
1909 0 : // d.atomic.logSize has exceeded that physical size. We allow for this
1910 0 : // anomaly.
1911 0 : metrics.WAL.PhysicalSize = walStats.LiveFileSize
1912 0 : metrics.WAL.BytesIn = d.logBytesIn.Load()
1913 0 : metrics.WAL.Size = d.logSize.Load()
1914 0 : for i, n := 0, len(d.mu.mem.queue)-1; i < n; i++ {
1915 0 : metrics.WAL.Size += d.mu.mem.queue[i].logSize
1916 0 : }
1917 0 : metrics.WAL.BytesWritten = metrics.Levels[0].TableBytesIn + metrics.WAL.Size
1918 0 : metrics.WAL.Failover = walStats.Failover
1919 0 :
1920 0 : if p := d.mu.versions.picker; p != nil {
1921 0 : compactions := d.getInProgressCompactionInfoLocked(nil)
1922 0 : m := p.getMetrics(compactions)
1923 0 : for level, lm := range m.levels {
1924 0 : metrics.Levels[level].Score = lm.score
1925 0 : metrics.Levels[level].FillFactor = lm.fillFactor
1926 0 : metrics.Levels[level].CompensatedFillFactor = lm.compensatedFillFactor
1927 0 : }
1928 : }
1929 0 : metrics.Table.ZombieCount = int64(d.mu.versions.zombieTables.Count())
1930 0 : metrics.Table.ZombieSize = d.mu.versions.zombieTables.TotalSize()
1931 0 : metrics.Table.Local.ZombieCount, metrics.Table.Local.ZombieSize = d.mu.versions.zombieTables.LocalStats()
1932 0 :
1933 0 : // The obsolete blob/table metrics have a subtle calculation:
1934 0 : //
1935 0 : // (A) The vs.metrics.{Table,BlobFiles}.[Local.]{ObsoleteCount,ObsoleteSize}
1936 0 : // fields reflect the set of files currently sitting in
1937 0 : // vs.obsolete{Tables,Blobs} but not yet enqueued to the cleanup manager.
1938 0 : //
1939 0 : // (B) The d.mu.fileDeletions.queuedStats field holds the set of files that have
1940 0 : // been queued for deletion by the cleanup manager.
1941 0 : //
1942 0 : // (C) The cleanup manager also maintains cumulative stats for the set of
1943 0 : // files that have been deleted.
1944 0 : //
1945 0 : // The value of currently pending obsolete files is (A) + (B) - (C).
1946 0 : pendingObsoleteFileStats := d.mu.fileDeletions.queuedStats
1947 0 : pendingObsoleteFileStats.Sub(completedObsoleteFileStats)
1948 0 : metrics.Table.Local.ObsoleteCount += pendingObsoleteFileStats.tablesLocal.count
1949 0 : metrics.Table.Local.ObsoleteSize += pendingObsoleteFileStats.tablesLocal.size
1950 0 : metrics.Table.ObsoleteCount += int64(pendingObsoleteFileStats.tablesAll.count)
1951 0 : metrics.Table.ObsoleteSize += pendingObsoleteFileStats.tablesAll.size
1952 0 : metrics.BlobFiles.Local.ObsoleteCount += pendingObsoleteFileStats.blobFilesLocal.count
1953 0 : metrics.BlobFiles.Local.ObsoleteSize += pendingObsoleteFileStats.blobFilesLocal.size
1954 0 : metrics.BlobFiles.ObsoleteCount += pendingObsoleteFileStats.blobFilesAll.count
1955 0 : metrics.BlobFiles.ObsoleteSize += pendingObsoleteFileStats.blobFilesAll.size
1956 0 : metrics.private.optionsFileSize = d.optionsFileSize
1957 0 :
1958 0 : d.mu.versions.logLock()
1959 0 : metrics.private.manifestFileSize = uint64(d.mu.versions.manifest.Size())
1960 0 : backingCount, backingTotalSize := d.mu.versions.latest.virtualBackings.Stats()
1961 0 : metrics.Table.BackingTableCount = uint64(backingCount)
1962 0 : metrics.Table.BackingTableSize = backingTotalSize
1963 0 : blobStats, _ := d.mu.versions.latest.blobFiles.Stats()
1964 0 : d.mu.versions.logUnlock()
1965 0 : metrics.BlobFiles.LiveCount = blobStats.Count
1966 0 : metrics.BlobFiles.LiveSize = blobStats.PhysicalSize
1967 0 : metrics.BlobFiles.ValueSize = blobStats.ValueSize
1968 0 : metrics.BlobFiles.ReferencedValueSize = blobStats.ReferencedValueSize
1969 0 : metrics.BlobFiles.ReferencedBackingValueSize = blobStats.ReferencedBackingValueSize
1970 0 :
1971 0 : metrics.LogWriter.FsyncLatency = d.mu.log.metrics.fsyncLatency
1972 0 : if err := metrics.LogWriter.Merge(&d.mu.log.metrics.LogWriterMetrics); err != nil {
1973 0 : d.opts.Logger.Errorf("metrics error: %s", err)
1974 0 : }
1975 0 : metrics.Flush.WriteThroughput = d.mu.compact.flushWriteThroughput
1976 0 : if d.mu.compact.flushing {
1977 0 : metrics.Flush.NumInProgress = 1
1978 0 : }
1979 :
1980 0 : metrics.Table.PendingStatsCollectionCount = int64(len(d.mu.tableStats.pending))
1981 0 : metrics.Table.InitialStatsCollectionComplete = d.mu.tableStats.loadedInitial
1982 0 :
1983 0 : d.mu.Unlock()
1984 0 :
1985 0 : // TODO(jackson): Consider making these metrics optional.
1986 0 : aggProps := tablePropsAnnotator.MultiLevelAnnotation(vers.Levels[:])
1987 0 : metrics.Keys.RangeKeySetsCount = aggProps.NumRangeKeySets
1988 0 : metrics.Keys.TombstoneCount = aggProps.NumDeletions
1989 0 :
1990 0 : delBytes := deletionBytesAnnotator.MultiLevelAnnotation(vers.Levels[:])
1991 0 : metrics.Table.Garbage.PointDeletionsBytesEstimate = delBytes.PointDels
1992 0 : metrics.Table.Garbage.RangeDeletionsBytesEstimate = delBytes.RangeDels
1993 0 :
1994 0 : for i := 0; i < numLevels; i++ {
1995 0 : aggProps := tablePropsAnnotator.LevelAnnotation(vers.Levels[i])
1996 0 : metrics.Levels[i].Additional.ValueBlocksSize = aggProps.ValueBlocksSize
1997 0 : metrics.Table.Compression.MergeWith(&aggProps.CompressionMetrics)
1998 0 : }
1999 :
2000 0 : blobCompressionMetrics := blobCompressionStatsAnnotator.Annotation(&vers.BlobFiles)
2001 0 : metrics.BlobFiles.Compression.MergeWith(&blobCompressionMetrics)
2002 0 :
2003 0 : metrics.BlockCache = d.opts.Cache.Metrics()
2004 0 : metrics.FileCache, metrics.Filter = d.fileCache.Metrics()
2005 0 : metrics.TableIters = d.fileCache.IterCount()
2006 0 : metrics.CategoryStats = d.fileCache.SSTStatsCollector().GetStats()
2007 0 :
2008 0 : metrics.SecondaryCacheMetrics = d.objProvider.Metrics()
2009 0 :
2010 0 : metrics.Uptime = d.timeNow().Sub(d.openedAt)
2011 0 :
2012 0 : metrics.manualMemory = manual.GetMetrics()
2013 0 :
2014 0 : return metrics
2015 : }
2016 :
2017 : // sstablesOptions hold the optional parameters to retrieve TableInfo for all sstables.
2018 : type sstablesOptions struct {
2019 : // set to true will return the sstable properties in TableInfo
2020 : withProperties bool
2021 :
2022 : // if set, return sstables that overlap the key range (end-exclusive)
2023 : start []byte
2024 : end []byte
2025 :
2026 : withApproximateSpanBytes bool
2027 : }
2028 :
2029 : // SSTablesOption set optional parameter used by `DB.SSTables`.
2030 : type SSTablesOption func(*sstablesOptions)
2031 :
2032 : // WithProperties enable return sstable properties in each TableInfo.
2033 : //
2034 : // NOTE: if most of the sstable properties need to be read from disk,
2035 : // this options may make method `SSTables` quite slow.
2036 0 : func WithProperties() SSTablesOption {
2037 0 : return func(opt *sstablesOptions) {
2038 0 : opt.withProperties = true
2039 0 : }
2040 : }
2041 :
2042 : // WithKeyRangeFilter ensures returned sstables overlap start and end (end-exclusive)
2043 : // if start and end are both nil these properties have no effect.
2044 0 : func WithKeyRangeFilter(start, end []byte) SSTablesOption {
2045 0 : return func(opt *sstablesOptions) {
2046 0 : opt.end = end
2047 0 : opt.start = start
2048 0 : }
2049 : }
2050 :
2051 : // WithApproximateSpanBytes enables capturing the approximate number of bytes that
2052 : // overlap the provided key span for each sstable.
2053 : // NOTE: This option requires WithKeyRangeFilter.
2054 0 : func WithApproximateSpanBytes() SSTablesOption {
2055 0 : return func(opt *sstablesOptions) {
2056 0 : opt.withApproximateSpanBytes = true
2057 0 : }
2058 : }
2059 :
2060 : // BackingType denotes the type of storage backing a given sstable.
2061 : type BackingType int
2062 :
2063 : const (
2064 : // BackingTypeLocal denotes an sstable stored on local disk according to the
2065 : // objprovider. This file is completely owned by us.
2066 : BackingTypeLocal BackingType = iota
2067 : // BackingTypeShared denotes an sstable stored on shared storage, created
2068 : // by this Pebble instance and possibly shared by other Pebble instances.
2069 : // These types of files have lifecycle managed by Pebble.
2070 : BackingTypeShared
2071 : // BackingTypeSharedForeign denotes an sstable stored on shared storage,
2072 : // created by a Pebble instance other than this one. These types of files have
2073 : // lifecycle managed by Pebble.
2074 : BackingTypeSharedForeign
2075 : // BackingTypeExternal denotes an sstable stored on external storage,
2076 : // not owned by any Pebble instance and with no refcounting/cleanup methods
2077 : // or lifecycle management. An example of an external file is a file restored
2078 : // from a backup.
2079 : BackingTypeExternal
2080 : backingTypeCount
2081 : )
2082 :
2083 : var backingTypeToString = [backingTypeCount]string{
2084 : BackingTypeLocal: "local",
2085 : BackingTypeShared: "shared",
2086 : BackingTypeSharedForeign: "shared-foreign",
2087 : BackingTypeExternal: "external",
2088 : }
2089 :
2090 : // String implements fmt.Stringer.
2091 0 : func (b BackingType) String() string {
2092 0 : return backingTypeToString[b]
2093 0 : }
2094 :
2095 : // SSTableInfo export manifest.TableInfo with sstable.Properties alongside
2096 : // other file backing info.
2097 : type SSTableInfo struct {
2098 : manifest.TableInfo
2099 : TableStats manifest.TableStats
2100 : // Virtual indicates whether the sstable is virtual.
2101 : Virtual bool
2102 : // BackingSSTNum is the disk file number associated with the backing sstable.
2103 : // If Virtual is false, BackingSSTNum == PhysicalTableDiskFileNum(TableNum).
2104 : BackingSSTNum base.DiskFileNum
2105 : // BackingSize is the size of the backing sstable. This is the same with
2106 : // TableInfo.Size when the table is not virtual.
2107 : BackingSize uint64
2108 : // BackingType is the type of storage backing this sstable.
2109 : BackingType BackingType
2110 : // Locator is the remote.Locator backing this sstable, if the backing type is
2111 : // not BackingTypeLocal.
2112 : Locator remote.Locator
2113 : // ApproximateSpanBytes describes the approximate number of bytes within the
2114 : // sstable that fall within a particular span. It's populated only when the
2115 : // ApproximateSpanBytes option is passed into DB.SSTables.
2116 : ApproximateSpanBytes uint64 `json:"ApproximateSpanBytes,omitempty"`
2117 :
2118 : // Properties is the sstable properties of this table. If Virtual is true,
2119 : // then the Properties are associated with the backing sst.
2120 : Properties *sstable.Properties
2121 : }
2122 :
2123 : // SSTables retrieves the current sstables. The returned slice is indexed by
2124 : // level and each level is indexed by the position of the sstable within the
2125 : // level. Note that this information may be out of date due to concurrent
2126 : // flushes and compactions.
2127 0 : func (d *DB) SSTables(opts ...SSTablesOption) ([][]SSTableInfo, error) {
2128 0 : opt := &sstablesOptions{}
2129 0 : for _, fn := range opts {
2130 0 : fn(opt)
2131 0 : }
2132 :
2133 0 : if opt.withApproximateSpanBytes && (opt.start == nil || opt.end == nil) {
2134 0 : return nil, errors.Errorf("cannot use WithApproximateSpanBytes without WithKeyRangeFilter option")
2135 0 : }
2136 :
2137 : // Grab and reference the current readState.
2138 0 : readState := d.loadReadState()
2139 0 : defer readState.unref()
2140 0 :
2141 0 : // TODO(peter): This is somewhat expensive, especially on a large
2142 0 : // database. It might be worthwhile to unify TableInfo and TableMetadata and
2143 0 : // then we could simply return current.Files. Note that RocksDB is doing
2144 0 : // something similar to the current code, so perhaps it isn't too bad.
2145 0 : srcLevels := readState.current.Levels
2146 0 : var totalTables int
2147 0 : for i := range srcLevels {
2148 0 : totalTables += srcLevels[i].Len()
2149 0 : }
2150 :
2151 0 : destTables := make([]SSTableInfo, totalTables)
2152 0 : destLevels := make([][]SSTableInfo, len(srcLevels))
2153 0 : for i := range destLevels {
2154 0 : j := 0
2155 0 : for m := range srcLevels[i].All() {
2156 0 : if opt.start != nil && opt.end != nil {
2157 0 : b := base.UserKeyBoundsEndExclusive(opt.start, opt.end)
2158 0 : if !m.Overlaps(d.opts.Comparer.Compare, &b) {
2159 0 : continue
2160 : }
2161 : }
2162 0 : destTables[j] = SSTableInfo{
2163 0 : TableInfo: m.TableInfo(),
2164 0 : }
2165 0 : if stats, ok := m.Stats(); ok {
2166 0 : destTables[j].TableStats = *stats
2167 0 : }
2168 0 : if opt.withProperties {
2169 0 : p, err := d.fileCache.getTableProperties(
2170 0 : m,
2171 0 : )
2172 0 : if err != nil {
2173 0 : return nil, err
2174 0 : }
2175 0 : destTables[j].Properties = p
2176 : }
2177 0 : destTables[j].Virtual = m.Virtual
2178 0 : destTables[j].BackingSSTNum = m.TableBacking.DiskFileNum
2179 0 : destTables[j].BackingSize = m.TableBacking.Size
2180 0 : objMeta, err := d.objProvider.Lookup(base.FileTypeTable, m.TableBacking.DiskFileNum)
2181 0 : if err != nil {
2182 0 : return nil, err
2183 0 : }
2184 0 : if objMeta.IsRemote() {
2185 0 : if objMeta.IsShared() {
2186 0 : if d.objProvider.IsSharedForeign(objMeta) {
2187 0 : destTables[j].BackingType = BackingTypeSharedForeign
2188 0 : } else {
2189 0 : destTables[j].BackingType = BackingTypeShared
2190 0 : }
2191 0 : } else {
2192 0 : destTables[j].BackingType = BackingTypeExternal
2193 0 : }
2194 0 : destTables[j].Locator = objMeta.Remote.Locator
2195 0 : } else {
2196 0 : destTables[j].BackingType = BackingTypeLocal
2197 0 : }
2198 :
2199 0 : if opt.withApproximateSpanBytes {
2200 0 : if m.ContainedWithinSpan(d.opts.Comparer.Compare, opt.start, opt.end) {
2201 0 : destTables[j].ApproximateSpanBytes = m.Size
2202 0 : } else {
2203 0 : size, err := d.fileCache.estimateSize(m, opt.start, opt.end)
2204 0 : if err != nil {
2205 0 : return nil, err
2206 0 : }
2207 0 : destTables[j].ApproximateSpanBytes = size
2208 : }
2209 : }
2210 0 : j++
2211 : }
2212 0 : destLevels[i] = destTables[:j]
2213 0 : destTables = destTables[j:]
2214 : }
2215 :
2216 0 : return destLevels, nil
2217 : }
2218 :
2219 1 : func (d *DB) walPreallocateSize() int {
2220 1 : // Set the WAL preallocate size to 110% of the memtable size. Note that there
2221 1 : // is a bit of apples and oranges in units here as the memtabls size
2222 1 : // corresponds to the memory usage of the memtable while the WAL size is the
2223 1 : // size of the batches (plus overhead) stored in the WAL.
2224 1 : //
2225 1 : // TODO(peter): 110% of the memtable size is quite hefty for a block
2226 1 : // size. This logic is taken from GetWalPreallocateBlockSize in
2227 1 : // RocksDB. Could a smaller preallocation block size be used?
2228 1 : size := d.opts.MemTableSize
2229 1 : size = (size / 10) + size
2230 1 : return int(size)
2231 1 : }
2232 :
2233 : func (d *DB) newMemTable(
2234 : logNum base.DiskFileNum, logSeqNum base.SeqNum, minSize uint64,
2235 1 : ) (*memTable, *flushableEntry) {
2236 1 : targetSize := minSize + uint64(memTableEmptySize)
2237 1 : // The targetSize should be less than MemTableSize, because any batch >=
2238 1 : // MemTableSize/2 should be treated as a large flushable batch.
2239 1 : if targetSize > d.opts.MemTableSize {
2240 0 : panic(errors.AssertionFailedf("attempting to allocate memtable larger than MemTableSize"))
2241 : }
2242 : // Double until the next memtable size is at least large enough to fit
2243 : // minSize.
2244 1 : for d.mu.mem.nextSize < targetSize {
2245 0 : d.mu.mem.nextSize = min(2*d.mu.mem.nextSize, d.opts.MemTableSize)
2246 0 : }
2247 1 : size := d.mu.mem.nextSize
2248 1 : // The next memtable should be double the size, up to Options.MemTableSize.
2249 1 : if d.mu.mem.nextSize < d.opts.MemTableSize {
2250 1 : d.mu.mem.nextSize = min(2*d.mu.mem.nextSize, d.opts.MemTableSize)
2251 1 : }
2252 :
2253 1 : memtblOpts := memTableOptions{
2254 1 : Options: d.opts,
2255 1 : logSeqNum: logSeqNum,
2256 1 : }
2257 1 :
2258 1 : // Before attempting to allocate a new memtable, check if there's one
2259 1 : // available for recycling in memTableRecycle. Large contiguous allocations
2260 1 : // can be costly as fragmentation makes it more difficult to find a large
2261 1 : // contiguous free space. We've observed 64MB allocations taking 10ms+.
2262 1 : //
2263 1 : // To reduce these costly allocations, up to 1 obsolete memtable is stashed
2264 1 : // in `d.memTableRecycle` to allow a future memtable rotation to reuse
2265 1 : // existing memory.
2266 1 : var mem *memTable
2267 1 : mem = d.memTableRecycle.Swap(nil)
2268 1 : if mem != nil && uint64(mem.arenaBuf.Len()) != size {
2269 1 : d.freeMemTable(mem)
2270 1 : mem = nil
2271 1 : }
2272 1 : if mem != nil {
2273 1 : // Carry through the existing buffer and memory reservation.
2274 1 : memtblOpts.arenaBuf = mem.arenaBuf
2275 1 : memtblOpts.releaseAccountingReservation = mem.releaseAccountingReservation
2276 1 : } else {
2277 1 : mem = new(memTable)
2278 1 : memtblOpts.arenaBuf = manual.New(manual.MemTable, uintptr(size))
2279 1 : memtblOpts.releaseAccountingReservation = d.opts.Cache.Reserve(int(size))
2280 1 : d.memTableCount.Add(1)
2281 1 : d.memTableReserved.Add(int64(size))
2282 1 :
2283 1 : // Note: this is a no-op if invariants are disabled or race is enabled.
2284 1 : invariants.SetFinalizer(mem, checkMemTable)
2285 1 : }
2286 1 : mem.init(memtblOpts)
2287 1 :
2288 1 : entry := d.newFlushableEntry(mem, logNum, logSeqNum)
2289 1 : entry.releaseMemAccounting = func() {
2290 1 : // If the user leaks iterators, we may be releasing the memtable after
2291 1 : // the DB is already closed. In this case, we want to just release the
2292 1 : // memory because DB.Close won't come along to free it for us.
2293 1 : if err := d.closed.Load(); err != nil {
2294 1 : d.freeMemTable(mem)
2295 1 : return
2296 1 : }
2297 :
2298 : // The next memtable allocation might be able to reuse this memtable.
2299 : // Stash it on d.memTableRecycle.
2300 1 : if unusedMem := d.memTableRecycle.Swap(mem); unusedMem != nil {
2301 1 : // There was already a memtable waiting to be recycled. We're now
2302 1 : // responsible for freeing it.
2303 1 : d.freeMemTable(unusedMem)
2304 1 : }
2305 : }
2306 1 : return mem, entry
2307 : }
2308 :
2309 1 : func (d *DB) freeMemTable(m *memTable) {
2310 1 : d.memTableCount.Add(-1)
2311 1 : d.memTableReserved.Add(-int64(m.arenaBuf.Len()))
2312 1 : m.free()
2313 1 : }
2314 :
2315 : func (d *DB) newFlushableEntry(
2316 : f flushable, logNum base.DiskFileNum, logSeqNum base.SeqNum,
2317 1 : ) *flushableEntry {
2318 1 : fe := &flushableEntry{
2319 1 : flushable: f,
2320 1 : flushed: make(chan struct{}),
2321 1 : logNum: logNum,
2322 1 : logSeqNum: logSeqNum,
2323 1 : deleteFn: d.mu.versions.addObsolete,
2324 1 : deleteFnLocked: d.mu.versions.addObsoleteLocked,
2325 1 : }
2326 1 : fe.readerRefs.Store(1)
2327 1 : return fe
2328 1 : }
2329 :
2330 : // maybeInduceWriteStall is called before performing a memtable rotation in
2331 : // makeRoomForWrite. In some conditions, we prefer to stall the user's write
2332 : // workload rather than continuing to accept writes that may result in resource
2333 : // exhaustion or prohibitively slow reads.
2334 : //
2335 : // There are a couple reasons we might wait to rotate the memtable and
2336 : // instead induce a write stall:
2337 : // 1. If too many memtables have queued, we wait for a flush to finish before
2338 : // creating another memtable.
2339 : // 2. If L0 read amplification has grown too high, we wait for compactions
2340 : // to reduce the read amplification before accepting more writes that will
2341 : // increase write pressure.
2342 : //
2343 : // maybeInduceWriteStall checks these stall conditions, and if present, waits
2344 : // for them to abate.
2345 1 : func (d *DB) maybeInduceWriteStall(b *Batch) {
2346 1 : stalled := false
2347 1 : // This function will call EventListener.WriteStallBegin at most once. If
2348 1 : // it does call it, it will call EventListener.WriteStallEnd once before
2349 1 : // returning.
2350 1 : var timer *time.Timer
2351 1 : for {
2352 1 : var size uint64
2353 1 : for i := range d.mu.mem.queue {
2354 1 : size += d.mu.mem.queue[i].totalBytes()
2355 1 : }
2356 1 : if size >= uint64(d.opts.MemTableStopWritesThreshold)*d.opts.MemTableSize &&
2357 1 : !d.mu.log.manager.ElevateWriteStallThresholdForFailover() {
2358 1 : // We have filled up the current memtable, but already queued memtables
2359 1 : // are still flushing, so we wait.
2360 1 : if !stalled {
2361 1 : stalled = true
2362 1 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2363 1 : Reason: "memtable count limit reached",
2364 1 : })
2365 1 : }
2366 1 : beforeWait := crtime.NowMono()
2367 1 : // NB: In a rare case, we can start a write stall, and then the system
2368 1 : // may detect WAL failover, resulting in
2369 1 : // ElevateWriteStallThresholdForFailover returning true. So we want to
2370 1 : // recheck the predicate periodically, which we do by signaling the
2371 1 : // condition variable.
2372 1 : if timer == nil {
2373 1 : timer = time.AfterFunc(time.Second, d.mu.compact.cond.Broadcast)
2374 1 : } else {
2375 1 : timer.Reset(time.Second)
2376 1 : }
2377 1 : d.mu.compact.cond.Wait()
2378 1 : timer.Stop()
2379 1 : if b != nil {
2380 1 : b.commitStats.MemTableWriteStallDuration += beforeWait.Elapsed()
2381 1 : }
2382 1 : continue
2383 : }
2384 1 : l0ReadAmp := d.mu.versions.latest.l0Organizer.ReadAmplification()
2385 1 : if l0ReadAmp >= d.opts.L0StopWritesThreshold {
2386 1 : // There are too many level-0 files, so we wait.
2387 1 : if !stalled {
2388 1 : stalled = true
2389 1 : d.opts.EventListener.WriteStallBegin(WriteStallBeginInfo{
2390 1 : Reason: "L0 file count limit exceeded",
2391 1 : })
2392 1 : }
2393 1 : beforeWait := crtime.NowMono()
2394 1 : d.mu.compact.cond.Wait()
2395 1 : if b != nil {
2396 0 : b.commitStats.L0ReadAmpWriteStallDuration += beforeWait.Elapsed()
2397 0 : }
2398 1 : continue
2399 : }
2400 : // Not stalled.
2401 1 : if stalled {
2402 1 : d.opts.EventListener.WriteStallEnd()
2403 1 : }
2404 1 : return
2405 : }
2406 : }
2407 :
2408 : // makeRoomForWrite rotates the current mutable memtable, ensuring that the
2409 : // resulting mutable memtable has room to hold the contents of the provided
2410 : // Batch. The current memtable is rotated (marked as immutable) and a new
2411 : // mutable memtable is allocated. It reserves space in the new memtable and adds
2412 : // a reference to the memtable. The caller must later ensure that the memtable
2413 : // is unreferenced. This memtable rotation also causes a log rotation.
2414 : //
2415 : // If the current memtable is not full but the caller wishes to trigger a
2416 : // rotation regardless, the caller may pass a nil Batch, and no space in the
2417 : // resulting mutable memtable will be reserved.
2418 : //
2419 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2420 : // may be released and reacquired.
2421 1 : func (d *DB) makeRoomForWrite(b *Batch) error {
2422 1 : if b != nil && b.ingestedSSTBatch {
2423 0 : panic("pebble: invalid function call")
2424 : }
2425 1 : d.maybeInduceWriteStall(b)
2426 1 :
2427 1 : var newLogNum base.DiskFileNum
2428 1 : var prevLogSize uint64
2429 1 : if !d.opts.DisableWAL {
2430 1 : beforeRotate := crtime.NowMono()
2431 1 : newLogNum, prevLogSize = d.rotateWAL()
2432 1 : if b != nil {
2433 1 : b.commitStats.WALRotationDuration += beforeRotate.Elapsed()
2434 1 : }
2435 : }
2436 1 : immMem := d.mu.mem.mutable
2437 1 : imm := d.mu.mem.queue[len(d.mu.mem.queue)-1]
2438 1 : imm.logSize = prevLogSize
2439 1 :
2440 1 : var logSeqNum base.SeqNum
2441 1 : var minSize uint64
2442 1 : if b != nil {
2443 1 : logSeqNum = b.SeqNum()
2444 1 : if b.flushable != nil {
2445 1 : logSeqNum += base.SeqNum(b.Count())
2446 1 : // The batch is too large to fit in the memtable so add it directly to
2447 1 : // the immutable queue. The flushable batch is associated with the same
2448 1 : // log as the immutable memtable, but logically occurs after it in
2449 1 : // seqnum space. We ensure while flushing that the flushable batch
2450 1 : // is flushed along with the previous memtable in the flushable
2451 1 : // queue. See the top level comment in DB.flush1 to learn how this
2452 1 : // is ensured.
2453 1 : //
2454 1 : // See DB.commitWrite for the special handling of log writes for large
2455 1 : // batches. In particular, the large batch has already written to
2456 1 : // imm.logNum.
2457 1 : entry := d.newFlushableEntry(b.flushable, imm.logNum, b.SeqNum())
2458 1 : // The large batch is by definition large. Reserve space from the cache
2459 1 : // for it until it is flushed.
2460 1 : entry.releaseMemAccounting = d.opts.Cache.Reserve(int(b.flushable.totalBytes()))
2461 1 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2462 1 : } else {
2463 1 : minSize = b.memTableSize
2464 1 : }
2465 1 : } else {
2466 1 : // b == nil
2467 1 : //
2468 1 : // This is a manual forced flush.
2469 1 : logSeqNum = base.SeqNum(d.mu.versions.logSeqNum.Load())
2470 1 : imm.flushForced = true
2471 1 : // If we are manually flushing and we used less than half of the bytes in
2472 1 : // the memtable, don't increase the size for the next memtable. This
2473 1 : // reduces memtable memory pressure when an application is frequently
2474 1 : // manually flushing.
2475 1 : if uint64(immMem.availBytes()) > immMem.totalBytes()/2 {
2476 1 : d.mu.mem.nextSize = immMem.totalBytes()
2477 1 : }
2478 : }
2479 1 : d.rotateMemtable(newLogNum, logSeqNum, immMem, minSize)
2480 1 : if b != nil && b.flushable == nil {
2481 1 : err := d.mu.mem.mutable.prepare(b)
2482 1 : // Reserving enough space for the batch after rotation must never fail.
2483 1 : // We pass in a minSize that's equal to b.memtableSize to ensure that
2484 1 : // memtable rotation allocates a memtable sufficiently large. We also
2485 1 : // held d.commit.mu for the entirety of this function, ensuring that no
2486 1 : // other committers may have reserved memory in the new memtable yet.
2487 1 : if err == arenaskl.ErrArenaFull {
2488 0 : panic(errors.AssertionFailedf("memtable still full after rotation"))
2489 : }
2490 1 : return err
2491 : }
2492 1 : return nil
2493 : }
2494 :
2495 : // Both DB.mu and commitPipeline.mu must be held by the caller.
2496 : func (d *DB) rotateMemtable(
2497 : newLogNum base.DiskFileNum, logSeqNum base.SeqNum, prev *memTable, minSize uint64,
2498 1 : ) {
2499 1 : // Create a new memtable, scheduling the previous one for flushing. We do
2500 1 : // this even if the previous memtable was empty because the DB.Flush
2501 1 : // mechanism is dependent on being able to wait for the empty memtable to
2502 1 : // flush. We can't just mark the empty memtable as flushed here because we
2503 1 : // also have to wait for all previous immutable tables to
2504 1 : // flush. Additionally, the memtable is tied to particular WAL file and we
2505 1 : // want to go through the flush path in order to recycle that WAL file.
2506 1 : //
2507 1 : // NB: newLogNum corresponds to the WAL that contains mutations that are
2508 1 : // present in the new memtable. When immutable memtables are flushed to
2509 1 : // disk, a VersionEdit will be created telling the manifest the minimum
2510 1 : // unflushed log number (which will be the next one in d.mu.mem.mutable
2511 1 : // that was not flushed).
2512 1 : //
2513 1 : // NB: prev should be the current mutable memtable.
2514 1 : var entry *flushableEntry
2515 1 : d.mu.mem.mutable, entry = d.newMemTable(newLogNum, logSeqNum, minSize)
2516 1 : d.mu.mem.queue = append(d.mu.mem.queue, entry)
2517 1 : // d.logSize tracks the log size of the WAL file corresponding to the most
2518 1 : // recent flushable. The log size of the previous mutable memtable no longer
2519 1 : // applies to the current mutable memtable.
2520 1 : //
2521 1 : // It's tempting to perform this update in rotateWAL, but that would not be
2522 1 : // atomic with the enqueue of the new flushable. A call to DB.Metrics()
2523 1 : // could acquire DB.mu after the WAL has been rotated but before the new
2524 1 : // memtable has been appended; this would result in omitting the log size of
2525 1 : // the most recent flushable.
2526 1 : d.logSize.Store(0)
2527 1 : d.updateReadStateLocked(nil)
2528 1 : if prev.writerUnref() {
2529 1 : d.maybeScheduleFlush()
2530 1 : }
2531 : }
2532 :
2533 : // rotateWAL creates a new write-ahead log, possibly recycling a previous WAL's
2534 : // files. It returns the file number assigned to the new WAL, and the size of
2535 : // the previous WAL file.
2536 : //
2537 : // Both DB.mu and commitPipeline.mu must be held by the caller. Note that DB.mu
2538 : // may be released and reacquired.
2539 1 : func (d *DB) rotateWAL() (newLogNum base.DiskFileNum, prevLogSize uint64) {
2540 1 : if d.opts.DisableWAL {
2541 0 : panic("pebble: invalid function call")
2542 : }
2543 1 : jobID := d.newJobIDLocked()
2544 1 : newLogNum = d.mu.versions.getNextDiskFileNum()
2545 1 :
2546 1 : d.mu.Unlock()
2547 1 : // Close the previous log first. This writes an EOF trailer
2548 1 : // signifying the end of the file and syncs it to disk. We must
2549 1 : // close the previous log before linking the new log file,
2550 1 : // otherwise a crash could leave both logs with unclean tails, and
2551 1 : // Open will treat the previous log as corrupt.
2552 1 : offset, err := d.mu.log.writer.Close()
2553 1 : if err != nil {
2554 0 : // What to do here? Stumbling on doesn't seem worthwhile. If we failed to
2555 0 : // close the previous log it is possible we lost a write.
2556 0 : d.opts.Logger.Fatalf("pebble: error closing WAL; data loss possible if we continue: %s", err)
2557 0 : }
2558 1 : prevLogSize = uint64(offset)
2559 1 : metrics := d.mu.log.writer.Metrics()
2560 1 :
2561 1 : d.mu.Lock()
2562 1 : if err := d.mu.log.metrics.LogWriterMetrics.Merge(&metrics); err != nil {
2563 0 : d.opts.Logger.Errorf("metrics error: %s", err)
2564 0 : }
2565 :
2566 1 : d.mu.Unlock()
2567 1 : writer, err := d.mu.log.manager.Create(wal.NumWAL(newLogNum), int(jobID))
2568 1 : if err != nil {
2569 0 : panic(err)
2570 : }
2571 :
2572 1 : d.mu.Lock()
2573 1 : d.mu.log.writer = writer
2574 1 : return newLogNum, prevLogSize
2575 : }
2576 :
2577 1 : func (d *DB) getEarliestUnflushedSeqNumLocked() base.SeqNum {
2578 1 : seqNum := base.SeqNumMax
2579 1 : for i := range d.mu.mem.queue {
2580 1 : logSeqNum := d.mu.mem.queue[i].logSeqNum
2581 1 : if seqNum > logSeqNum {
2582 1 : seqNum = logSeqNum
2583 1 : }
2584 : }
2585 1 : return seqNum
2586 : }
2587 :
2588 1 : func (d *DB) getInProgressCompactionInfoLocked(finishing compaction) (rv []compactionInfo) {
2589 1 : for c := range d.mu.compact.inProgress {
2590 1 : if !c.IsFlush() && (finishing == nil || c != finishing) {
2591 1 : rv = append(rv, c.Info())
2592 1 : }
2593 : }
2594 1 : return
2595 : }
2596 :
2597 1 : func inProgressL0Compactions(inProgress []compactionInfo) []manifest.L0Compaction {
2598 1 : var compactions []manifest.L0Compaction
2599 1 : for _, info := range inProgress {
2600 1 : // Skip in-progress compactions that have already committed; the L0
2601 1 : // sublevels initialization code requires the set of in-progress
2602 1 : // compactions to be consistent with the current version. Compactions
2603 1 : // with versionEditApplied=true are already applied to the current
2604 1 : // version and but are performing cleanup without the database mutex.
2605 1 : if info.versionEditApplied {
2606 1 : continue
2607 : }
2608 1 : l0 := false
2609 1 : for _, cl := range info.inputs {
2610 1 : l0 = l0 || cl.level == 0
2611 1 : }
2612 1 : if !l0 {
2613 1 : continue
2614 : }
2615 1 : compactions = append(compactions, manifest.L0Compaction{
2616 1 : Bounds: *info.bounds,
2617 1 : IsIntraL0: info.outputLevel == 0,
2618 1 : })
2619 : }
2620 1 : return compactions
2621 : }
2622 :
2623 : // firstError returns the first non-nil error of err0 and err1, or nil if both
2624 : // are nil.
2625 1 : func firstError(err0, err1 error) error {
2626 1 : if err0 != nil {
2627 1 : return err0
2628 1 : }
2629 1 : return err1
2630 : }
2631 :
2632 : // SetCreatorID sets the CreatorID which is needed in order to use shared objects.
2633 : // Remote object usage is disabled until this method is called the first time.
2634 : // Once set, the Creator ID is persisted and cannot change.
2635 : //
2636 : // Does nothing if SharedStorage was not set in the options when the DB was
2637 : // opened or if the DB is in read-only mode.
2638 0 : func (d *DB) SetCreatorID(creatorID uint64) error {
2639 0 : if d.opts.Experimental.RemoteStorage == nil || d.opts.ReadOnly {
2640 0 : return nil
2641 0 : }
2642 0 : return d.objProvider.SetCreatorID(objstorage.CreatorID(creatorID))
2643 : }
2644 :
2645 : // KeyStatistics keeps track of the number of keys that have been pinned by a
2646 : // snapshot as well as counts of the different key kinds in the lsm.
2647 : //
2648 : // One way of using the accumulated stats, when we only have sets and dels,
2649 : // and say the counts are represented as del_count, set_count,
2650 : // del_latest_count, set_latest_count, snapshot_pinned_count.
2651 : //
2652 : // - del_latest_count + set_latest_count is the set of unique user keys
2653 : // (unique).
2654 : //
2655 : // - set_latest_count is the set of live unique user keys (live_unique).
2656 : //
2657 : // - Garbage is del_count + set_count - live_unique.
2658 : //
2659 : // - If everything were in the LSM, del_count+set_count-snapshot_pinned_count
2660 : // would also be the set of unique user keys (note that
2661 : // snapshot_pinned_count is counting something different -- see comment below).
2662 : // But snapshot_pinned_count only counts keys in the LSM so the excess here
2663 : // must be keys in memtables.
2664 : type KeyStatistics struct {
2665 : // TODO(sumeer): the SnapshotPinned* are incorrect in that these older
2666 : // versions can be in a different level. Either fix the accounting or
2667 : // rename these fields.
2668 :
2669 : // SnapshotPinnedKeys represents obsolete keys that cannot be elided during
2670 : // a compaction, because they are required by an open snapshot.
2671 : SnapshotPinnedKeys int
2672 : // SnapshotPinnedKeysBytes is the total number of bytes of all snapshot
2673 : // pinned keys.
2674 : SnapshotPinnedKeysBytes uint64
2675 : // KindsCount is the count for each kind of key. It includes point keys,
2676 : // range deletes and range keys.
2677 : KindsCount [InternalKeyKindMax + 1]int
2678 : // LatestKindsCount is the count for each kind of key when it is the latest
2679 : // kind for a user key. It is only populated for point keys.
2680 : LatestKindsCount [InternalKeyKindMax + 1]int
2681 : }
2682 :
2683 : // LSMKeyStatistics is used by DB.ScanStatistics.
2684 : type LSMKeyStatistics struct {
2685 : Accumulated KeyStatistics
2686 : // Levels contains statistics only for point keys. Range deletions and range keys will
2687 : // appear in Accumulated but not Levels.
2688 : Levels [numLevels]KeyStatistics
2689 : // BytesRead represents the logical, pre-compression size of keys and values read
2690 : BytesRead uint64
2691 : }
2692 :
2693 : // ScanStatisticsOptions is used by DB.ScanStatistics.
2694 : type ScanStatisticsOptions struct {
2695 : // LimitBytesPerSecond indicates the number of bytes that are able to be read
2696 : // per second using ScanInternal.
2697 : // A value of 0 indicates that there is no limit set.
2698 : LimitBytesPerSecond int64
2699 : }
2700 :
2701 : // ScanStatistics returns the count of different key kinds within the lsm for a
2702 : // key span [lower, upper) as well as the number of snapshot keys.
2703 : func (d *DB) ScanStatistics(
2704 : ctx context.Context, lower, upper []byte, opts ScanStatisticsOptions,
2705 0 : ) (stats LSMKeyStatistics, err error) {
2706 0 : var prevKey InternalKey
2707 0 : var rateLimitFunc func(key *InternalKey, val LazyValue) error
2708 0 : tb := tokenbucket.TokenBucket{}
2709 0 :
2710 0 : if opts.LimitBytesPerSecond != 0 {
2711 0 : const minBytesPerSec = 100 * 1024
2712 0 : if opts.LimitBytesPerSecond < minBytesPerSec {
2713 0 : return stats, errors.Newf("pebble: ScanStatistics read bandwidth limit %d is below minimum %d", opts.LimitBytesPerSecond, minBytesPerSec)
2714 0 : }
2715 : // Each "token" roughly corresponds to a byte that was read.
2716 0 : tb.Init(tokenbucket.TokensPerSecond(opts.LimitBytesPerSecond), tokenbucket.Tokens(1024))
2717 0 : rateLimitFunc = func(key *InternalKey, val LazyValue) error {
2718 0 : return tb.WaitCtx(ctx, tokenbucket.Tokens(key.Size()+val.Len()))
2719 0 : }
2720 : }
2721 :
2722 0 : scanInternalOpts := &ScanInternalOptions{
2723 0 : VisitPointKey: func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error {
2724 0 : // If the previous key is equal to the current point key, the current key was
2725 0 : // pinned by a snapshot.
2726 0 : size := uint64(key.Size())
2727 0 : kind := key.Kind()
2728 0 : sameKey := d.equal(prevKey.UserKey, key.UserKey)
2729 0 : if iterInfo.Kind == IteratorLevelLSM && sameKey {
2730 0 : stats.Levels[iterInfo.Level].SnapshotPinnedKeys++
2731 0 : stats.Levels[iterInfo.Level].SnapshotPinnedKeysBytes += size
2732 0 : stats.Accumulated.SnapshotPinnedKeys++
2733 0 : stats.Accumulated.SnapshotPinnedKeysBytes += size
2734 0 : }
2735 0 : if iterInfo.Kind == IteratorLevelLSM {
2736 0 : stats.Levels[iterInfo.Level].KindsCount[kind]++
2737 0 : }
2738 0 : if !sameKey {
2739 0 : if iterInfo.Kind == IteratorLevelLSM {
2740 0 : stats.Levels[iterInfo.Level].LatestKindsCount[kind]++
2741 0 : }
2742 0 : stats.Accumulated.LatestKindsCount[kind]++
2743 : }
2744 :
2745 0 : stats.Accumulated.KindsCount[kind]++
2746 0 : prevKey.CopyFrom(*key)
2747 0 : stats.BytesRead += uint64(key.Size() + value.Len())
2748 0 : return nil
2749 : },
2750 0 : VisitRangeDel: func(start, end []byte, seqNum base.SeqNum) error {
2751 0 : stats.Accumulated.KindsCount[InternalKeyKindRangeDelete]++
2752 0 : stats.BytesRead += uint64(len(start) + len(end))
2753 0 : return nil
2754 0 : },
2755 0 : VisitRangeKey: func(start, end []byte, keys []rangekey.Key) error {
2756 0 : stats.BytesRead += uint64(len(start) + len(end))
2757 0 : for _, key := range keys {
2758 0 : stats.Accumulated.KindsCount[key.Kind()]++
2759 0 : stats.BytesRead += uint64(len(key.Value) + len(key.Suffix))
2760 0 : }
2761 0 : return nil
2762 : },
2763 : IncludeObsoleteKeys: true,
2764 : IterOptions: IterOptions{
2765 : KeyTypes: IterKeyTypePointsAndRanges,
2766 : LowerBound: lower,
2767 : UpperBound: upper,
2768 : },
2769 : RateLimitFunc: rateLimitFunc,
2770 : }
2771 0 : iter, err := d.newInternalIter(ctx, snapshotIterOpts{}, scanInternalOpts)
2772 0 : if err != nil {
2773 0 : return LSMKeyStatistics{}, err
2774 0 : }
2775 0 : defer func() { err = errors.CombineErrors(err, iter.Close()) }()
2776 :
2777 0 : err = scanInternalImpl(ctx, iter, scanInternalOpts)
2778 0 : if err != nil {
2779 0 : return LSMKeyStatistics{}, err
2780 0 : }
2781 0 : return stats, nil
2782 : }
2783 :
2784 : // ObjProvider returns the objstorage.Provider for this database. Meant to be
2785 : // used for internal purposes only.
2786 1 : func (d *DB) ObjProvider() objstorage.Provider {
2787 1 : return d.objProvider
2788 1 : }
2789 :
2790 : // DebugString returns a debugging string describing the LSM.
2791 0 : func (d *DB) DebugString() string {
2792 0 : return d.DebugCurrentVersion().DebugString()
2793 0 : }
2794 :
2795 : // DebugCurrentVersion returns the current LSM tree metadata. Should only be
2796 : // used for testing/debugging.
2797 0 : func (d *DB) DebugCurrentVersion() *manifest.Version {
2798 0 : d.mu.Lock()
2799 0 : defer d.mu.Unlock()
2800 0 : return d.mu.versions.currentVersion()
2801 0 : }
|
