forked from cockroachdb/pebble
-
Notifications
You must be signed in to change notification settings - Fork 0
/
compaction_iter.go
1331 lines (1259 loc) · 50.4 KB
/
compaction_iter.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// Copyright 2018 The LevelDB-Go and Pebble Authors. All rights reserved. Use
// of this source code is governed by a BSD-style license that can be found in
// the LICENSE file.
package pebble
import (
"encoding/binary"
"io"
"sort"
"strconv"
"github.com/cockroachdb/errors"
"github.com/cockroachdb/pebble/internal/base"
"github.com/cockroachdb/pebble/internal/bytealloc"
"github.com/cockroachdb/pebble/internal/compact"
"github.com/cockroachdb/pebble/internal/keyspan"
"github.com/cockroachdb/pebble/internal/rangekey"
"github.com/cockroachdb/redact"
)
// compactionIter provides a forward-only iterator that encapsulates the logic
// for collapsing entries during compaction. It wraps an internal iterator and
// collapses entries that are no longer necessary because they are shadowed by
// newer entries. The simplest example of this is when the internal iterator
// contains two keys: a.PUT.2 and a.PUT.1. Instead of returning both entries,
// compactionIter collapses the second entry because it is no longer
// necessary. The high-level structure for compactionIter is to iterate over
// its internal iterator and output 1 entry for every user-key. There are four
// complications to this story.
//
// 1. Eliding Deletion Tombstones
//
// Consider the entries a.DEL.2 and a.PUT.1. These entries collapse to
// a.DEL.2. Do we have to output the entry a.DEL.2? Only if a.DEL.2 possibly
// shadows an entry at a lower level. If we're compacting to the base-level in
// the LSM tree then a.DEL.2 is definitely not shadowing an entry at a lower
// level and can be elided.
//
// We can do slightly better than only eliding deletion tombstones at the base
// level by observing that we can elide a deletion tombstone if there are no
// sstables that contain the entry's key. This check is performed by
// elideTombstone.
//
// 2. Merges
//
// The MERGE operation merges the value for an entry with the existing value
// for an entry. The logical value of an entry can be composed of a series of
// merge operations. When compactionIter sees a MERGE, it scans forward in its
// internal iterator collapsing MERGE operations for the same key until it
// encounters a SET or DELETE operation. For example, the keys a.MERGE.4,
// a.MERGE.3, a.MERGE.2 will be collapsed to a.MERGE.4 and the values will be
// merged using the specified Merger.
//
// An interesting case here occurs when MERGE is combined with SET. Consider
// the entries a.MERGE.3 and a.SET.2. The collapsed key will be a.SET.3. The
// reason that the kind is changed to SET is because the SET operation acts as
// a barrier preventing further merging. This can be seen better in the
// scenario a.MERGE.3, a.SET.2, a.MERGE.1. The entry a.MERGE.1 may be at lower
// (older) level and not involved in the compaction. If the compaction of
// a.MERGE.3 and a.SET.2 produced a.MERGE.3, a subsequent compaction with
// a.MERGE.1 would merge the values together incorrectly.
//
// 3. Snapshots
//
// Snapshots are lightweight point-in-time views of the DB state. At its core,
// a snapshot is a sequence number along with a guarantee from Pebble that it
// will maintain the view of the database at that sequence number. Part of this
// guarantee is relatively straightforward to achieve. When reading from the
// database Pebble will ignore sequence numbers that are larger than the
// snapshot sequence number. The primary complexity with snapshots occurs
// during compaction: the collapsing of entries that are shadowed by newer
// entries is at odds with the guarantee that Pebble will maintain the view of
// the database at the snapshot sequence number. Rather than collapsing entries
// up to the next user key, compactionIter can only collapse entries up to the
// next snapshot boundary. That is, every snapshot boundary potentially causes
// another entry for the same user-key to be emitted. Another way to view this
// is that snapshots define stripes and entries are collapsed within stripes,
// but not across stripes. Consider the following scenario:
//
// a.PUT.9
// a.DEL.8
// a.PUT.7
// a.DEL.6
// a.PUT.5
//
// In the absence of snapshots these entries would be collapsed to
// a.PUT.9. What if there is a snapshot at sequence number 7? The entries can
// be divided into two stripes and collapsed within the stripes:
//
// a.PUT.9 a.PUT.9
// a.DEL.8 --->
// a.PUT.7
// -- --
// a.DEL.6 ---> a.DEL.6
// a.PUT.5
//
// All of the rules described earlier still apply, but they are confined to
// operate within a snapshot stripe. Snapshots only affect compaction when the
// snapshot sequence number lies within the range of sequence numbers being
// compacted. In the above example, a snapshot at sequence number 10 or at
// sequence number 5 would not have any effect.
//
// 4. Range Deletions
//
// Range deletions provide the ability to delete all of the keys (and values)
// in a contiguous range. Range deletions are stored indexed by their start
// key. The end key of the range is stored in the value. In order to support
// lookup of the range deletions which overlap with a particular key, the range
// deletion tombstones need to be fragmented whenever they overlap. This
// fragmentation is performed by keyspan.Fragmenter. The fragments are then
// subject to the rules for snapshots. For example, consider the two range
// tombstones [a,e)#1 and [c,g)#2:
//
// 2: c-------g
// 1: a-------e
//
// These tombstones will be fragmented into:
//
// 2: c---e---g
// 1: a---c---e
//
// Do we output the fragment [c,e)#1? Since it is covered by [c-e]#2 the answer
// depends on whether it is in a new snapshot stripe.
//
// In addition to the fragmentation of range tombstones, compaction also needs
// to take the range tombstones into consideration when outputting normal
// keys. Just as with point deletions, a range deletion covering an entry can
// cause the entry to be elided.
//
// A note on the stability of keys and values.
//
// The stability guarantees of keys and values returned by the iterator tree
// that backs a compactionIter is nuanced and care must be taken when
// referencing any returned items.
//
// Keys and values returned by exported functions (i.e. First, Next, etc.) have
// lifetimes that fall into two categories:
//
// Lifetime valid for duration of compaction. Range deletion keys and values are
// stable for the duration of the compaction, due to way in which a
// compactionIter is typically constructed (i.e. via (*compaction).newInputIter,
// which wraps the iterator over the range deletion block in a noCloseIter,
// preventing the release of the backing memory until the compaction is
// finished).
//
// Lifetime limited to duration of sstable block liveness. Point keys (SET, DEL,
// etc.) and values must be cloned / copied following the return from the
// exported function, and before a subsequent call to Next advances the iterator
// and mutates the contents of the returned key and value.
type compactionIter struct {
equal Equal
merge Merge
iter internalIterator
err error
// `key.UserKey` is set to `keyBuf` caused by saving `i.iterKey.UserKey`
// and `key.Trailer` is set to `i.iterKey.Trailer`. This is the
// case on return from all public methods -- these methods return `key`.
// Additionally, it is the internal state when the code is moving to the
// next key so it can determine whether the user key has changed from
// the previous key.
key InternalKey
// keyTrailer is updated when `i.key` is updated and holds the key's
// original trailer (eg, before any sequence-number zeroing or changes to
// key kind).
keyTrailer uint64
value []byte
valueCloser io.Closer
// Temporary buffer used for storing the previous user key in order to
// determine when iteration has advanced to a new user key and thus a new
// snapshot stripe.
keyBuf []byte
// Temporary buffer used for storing the previous value, which may be an
// unsafe, i.iter-owned slice that could be altered when the iterator is
// advanced.
valueBuf []byte
// Is the current entry valid?
valid bool
iterKey *InternalKey
iterValue []byte
iterStripeChange stripeChangeType
// `skip` indicates whether the remaining entries in the current snapshot
// stripe should be skipped or processed. `skip` has no effect when `pos ==
// iterPosNext`.
skip bool
// `pos` indicates the iterator position at the top of `Next()`. Its type's
// (`iterPos`) values take on the following meanings in the context of
// `compactionIter`.
//
// - `iterPosCur`: the iterator is at the last key returned.
// - `iterPosNext`: the iterator has already been advanced to the next
// candidate key. For example, this happens when processing merge operands,
// where we advance the iterator all the way into the next stripe or next
// user key to ensure we've seen all mergeable operands.
// - `iterPosPrev`: this is invalid as compactionIter is forward-only.
pos iterPos
// `snapshotPinned` indicates whether the last point key returned by the
// compaction iterator was only returned because an open snapshot prevents
// its elision. This field only applies to point keys, and not to range
// deletions or range keys.
snapshotPinned bool
// forceObsoleteDueToRangeDel is set to true in a subset of the cases that
// snapshotPinned is true. This value is true when the point is obsolete due
// to a RANGEDEL but could not be deleted due to a snapshot.
//
// NB: it may seem that the additional cases that snapshotPinned captures
// are harmless in that they can also be used to mark a point as obsolete
// (it is merely a duplication of some logic that happens in
// Writer.AddWithForceObsolete), but that is not quite accurate as of this
// writing -- snapshotPinned originated in stats collection and for a
// sequence MERGE, SET, where the MERGE cannot merge with the (older) SET
// due to a snapshot, the snapshotPinned value for the SET is true.
//
// TODO(sumeer,jackson): improve the logic of snapshotPinned and reconsider
// whether we need forceObsoleteDueToRangeDel.
forceObsoleteDueToRangeDel bool
// The index of the snapshot for the current key within the snapshots slice.
curSnapshotIdx int
curSnapshotSeqNum uint64
// The snapshot sequence numbers that need to be maintained. These sequence
// numbers define the snapshot stripes (see the Snapshots description
// above). The sequence numbers are in ascending order.
snapshots []uint64
// frontiers holds a heap of user keys that affect compaction behavior when
// they're exceeded. Before a new key is returned, the compaction iterator
// advances the frontier, notifying any code that subscribed to be notified
// when a key was reached. The primary use today is within the
// implementation of compactionOutputSplitters in compaction.go. Many of
// these splitters wait for the compaction iterator to call Advance(k) when
// it's returning a new key. If the key that they're waiting for is
// surpassed, these splitters update internal state recording that they
// should request a compaction split next time they're asked in
// [shouldSplitBefore].
frontiers compact.Frontiers
// Reference to the range deletion tombstone fragmenter (e.g.,
// `compaction.rangeDelFrag`).
// TODO(jackson): We can eliminate range{Del,Key}Frag now that fragmentation
// is performed upfront by keyspanimpl.MergingIters.
rangeDelFrag *keyspan.Fragmenter
rangeKeyFrag *keyspan.Fragmenter
// The fragmented tombstones.
tombstones []keyspan.Span
// The fragmented range keys.
rangeKeys []keyspan.Span
// Byte allocator for the tombstone keys.
alloc bytealloc.A
allowZeroSeqNum bool
elideTombstone func(key []byte) bool
elideRangeTombstone func(start, end []byte) bool
ineffectualSingleDeleteCallback func(userKey []byte)
singleDeleteInvariantViolationCallback func(userKey []byte)
// The on-disk format major version. This informs the types of keys that
// may be written to disk during a compaction.
formatVersion FormatMajorVersion
stats struct {
// count of DELSIZED keys that were missized.
countMissizedDels uint64
}
}
func newCompactionIter(
cmp Compare,
equal Equal,
formatKey base.FormatKey,
merge Merge,
iter internalIterator,
snapshots []uint64,
rangeDelFrag *keyspan.Fragmenter,
rangeKeyFrag *keyspan.Fragmenter,
allowZeroSeqNum bool,
elideTombstone func(key []byte) bool,
elideRangeTombstone func(start, end []byte) bool,
ineffectualSingleDeleteCallback func(userKey []byte),
singleDeleteInvariantViolationCallback func(userKey []byte),
formatVersion FormatMajorVersion,
) *compactionIter {
i := &compactionIter{
equal: equal,
merge: merge,
iter: iter,
snapshots: snapshots,
rangeDelFrag: rangeDelFrag,
rangeKeyFrag: rangeKeyFrag,
allowZeroSeqNum: allowZeroSeqNum,
elideTombstone: elideTombstone,
elideRangeTombstone: elideRangeTombstone,
ineffectualSingleDeleteCallback: ineffectualSingleDeleteCallback,
singleDeleteInvariantViolationCallback: singleDeleteInvariantViolationCallback,
formatVersion: formatVersion,
}
i.frontiers.Init(cmp)
i.rangeDelFrag.Cmp = cmp
i.rangeDelFrag.Format = formatKey
i.rangeDelFrag.Emit = i.emitRangeDelChunk
i.rangeKeyFrag.Cmp = cmp
i.rangeKeyFrag.Format = formatKey
i.rangeKeyFrag.Emit = i.emitRangeKeyChunk
return i
}
func (i *compactionIter) First() (*InternalKey, []byte) {
if i.err != nil {
return nil, nil
}
var iterValue LazyValue
i.iterKey, iterValue = i.iter.First()
i.iterValue, _, i.err = iterValue.Value(nil)
if i.err != nil {
return nil, nil
}
if i.iterKey != nil {
i.curSnapshotIdx, i.curSnapshotSeqNum = snapshotIndex(i.iterKey.SeqNum(), i.snapshots)
}
i.pos = iterPosNext
i.iterStripeChange = newStripeNewKey
return i.Next()
}
func (i *compactionIter) Next() (*InternalKey, []byte) {
if i.err != nil {
return nil, nil
}
// Close the closer for the current value if one was open.
if i.closeValueCloser() != nil {
return nil, nil
}
// Prior to this call to `Next()` we are in one of three situations with
// respect to `iterKey` and related state:
//
// - `!skip && pos == iterPosNext`: `iterKey` is already at the next key.
// - `!skip && pos == iterPosCurForward`: We are at the key that has been returned.
// To move forward we advance by one key, even if that lands us in the same
// snapshot stripe.
// - `skip && pos == iterPosCurForward`: We are at the key that has been returned.
// To move forward we skip skippable entries in the stripe.
if i.pos == iterPosCurForward {
if i.skip {
i.skipInStripe()
} else {
i.nextInStripe()
}
} else if i.skip {
panic(errors.AssertionFailedf("compaction iterator has skip=true, but iterator is at iterPosNext"))
}
i.pos = iterPosCurForward
i.valid = false
for i.iterKey != nil {
// If we entered a new snapshot stripe with the same key, any key we
// return on this iteration is only returned because the open snapshot
// prevented it from being elided or merged with the key returned for
// the previous stripe. Mark it as pinned so that the compaction loop
// can correctly populate output tables' pinned statistics. We might
// also set snapshotPinned=true down below if we observe that the key is
// deleted by a range deletion in a higher stripe or that this key is a
// tombstone that could be elided if only it were in the last snapshot
// stripe.
i.snapshotPinned = i.iterStripeChange == newStripeSameKey
if i.iterKey.Kind() == InternalKeyKindRangeDelete || rangekey.IsRangeKey(i.iterKey.Kind()) {
// Return the span so the compaction can use it for file truncation and add
// it to the relevant fragmenter. In the case of range deletions, we do not
// set `skip` to true before returning as there may be any number of point
// keys with the same user key and sequence numbers ≥ the range deletion's
// sequence number. Such point keys must be visible (i.e., not skipped
// over) since we promise point keys are not deleted by range tombstones at
// the same sequence number (or higher).
//
// Note that `skip` must already be false here, because range keys and range
// deletions are interleaved at the maximal sequence numbers and neither will
// set `skip`=true.
if i.skip {
panic(errors.AssertionFailedf("pebble: compaction iterator: skip unexpectedly true"))
}
// NOTE: there is a subtle invariant violation here in that calling
// saveKey and returning a reference to the temporary slice violates
// the stability guarantee for range deletion keys. A potential
// mediation could return the original iterKey and iterValue
// directly, as the backing memory is guaranteed to be stable until
// the compaction completes. The violation here is only minor in
// that the caller immediately clones the range deletion InternalKey
// when passing the key to the deletion fragmenter (see the
// call-site in compaction.go).
// TODO(travers): address this violation by removing the call to
// saveKey and instead return the original iterKey and iterValue.
// This goes against the comment on i.key in the struct, and
// therefore warrants some investigation.
i.saveKey()
// TODO(jackson): Handle tracking pinned statistics for range keys
// and range deletions. This would require updating
// emitRangeDelChunk and rangeKeyCompactionTransform to update
// statistics when they apply their own snapshot striping logic.
i.snapshotPinned = false
i.value = i.iterValue
i.valid = true
return &i.key, i.value
}
// TODO(sumeer): we could avoid calling Covers if i.iterStripeChange ==
// sameStripeSameKey since that check has already been done in
// nextInStripeHelper. However, we also need to handle the case of
// CoversInvisibly below.
if cover := i.rangeDelFrag.Covers(*i.iterKey, i.curSnapshotSeqNum); cover == keyspan.CoversVisibly {
// A pending range deletion deletes this key. Skip it.
i.saveKey()
i.skipInStripe()
continue
} else if cover == keyspan.CoversInvisibly {
// i.iterKey would be deleted by a range deletion if there weren't
// any open snapshots. Mark it as pinned.
//
// NB: there are multiple places in this file where we call
// i.rangeDelFrag.Covers and this is the only one where we are writing
// to i.snapshotPinned. Those other cases occur in mergeNext where the
// caller is deciding whether the value should be merged or not, and the
// key is in the same snapshot stripe. Hence, snapshotPinned is by
// definition false in those cases.
i.snapshotPinned = true
i.forceObsoleteDueToRangeDel = true
} else {
i.forceObsoleteDueToRangeDel = false
}
switch i.iterKey.Kind() {
case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
if i.elideTombstone(i.iterKey.UserKey) {
if i.curSnapshotIdx == 0 {
// If we're at the last snapshot stripe and the tombstone
// can be elided skip skippable keys in the same stripe.
i.saveKey()
if i.key.Kind() == InternalKeyKindSingleDelete {
i.skipDueToSingleDeleteElision()
} else {
i.skipInStripe()
if !i.skip && i.iterStripeChange != newStripeNewKey {
panic(errors.AssertionFailedf("pebble: skipInStripe in last stripe disabled skip without advancing to new key"))
}
}
if i.iterStripeChange == newStripeSameKey {
panic(errors.AssertionFailedf("pebble: skipInStripe in last stripe found a new stripe within the same key"))
}
continue
} else {
// We're not at the last snapshot stripe, so the tombstone
// can NOT yet be elided. Mark it as pinned, so that it's
// included in table statistics appropriately.
i.snapshotPinned = true
}
}
switch i.iterKey.Kind() {
case InternalKeyKindDelete:
i.saveKey()
i.value = i.iterValue
i.valid = true
i.skip = true
return &i.key, i.value
case InternalKeyKindDeleteSized:
// We may skip subsequent keys because of this tombstone. Scan
// ahead to see just how much data this tombstone drops and if
// the tombstone's value should be updated accordingly.
return i.deleteSizedNext()
case InternalKeyKindSingleDelete:
if i.singleDeleteNext() {
return &i.key, i.value
} else if i.err != nil {
return nil, nil
}
continue
default:
panic(errors.AssertionFailedf(
"unexpected kind %s", redact.SafeString(i.iterKey.Kind().String())))
}
case InternalKeyKindSet, InternalKeyKindSetWithDelete:
// The key we emit for this entry is a function of the current key
// kind, and whether this entry is followed by a DEL/SINGLEDEL
// entry. setNext() does the work to move the iterator forward,
// preserving the original value, and potentially mutating the key
// kind.
i.setNext()
if i.err != nil {
return nil, nil
}
return &i.key, i.value
case InternalKeyKindMerge:
// Record the snapshot index before mergeNext as merging
// advances the iterator, adjusting curSnapshotIdx.
origSnapshotIdx := i.curSnapshotIdx
var valueMerger ValueMerger
valueMerger, i.err = i.merge(i.iterKey.UserKey, i.iterValue)
if i.err == nil {
i.mergeNext(valueMerger)
}
var needDelete bool
if i.err == nil {
// includesBase is true whenever we've transformed the MERGE record
// into a SET.
var includesBase bool
switch i.key.Kind() {
case InternalKeyKindSet, InternalKeyKindSetWithDelete:
includesBase = true
case InternalKeyKindMerge:
default:
panic(errors.AssertionFailedf(
"unexpected kind %s", redact.SafeString(i.key.Kind().String())))
}
i.value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, includesBase)
}
if i.err == nil {
if needDelete {
i.valid = false
if i.closeValueCloser() != nil {
return nil, nil
}
continue
}
i.maybeZeroSeqnum(origSnapshotIdx)
return &i.key, i.value
}
if i.err != nil {
i.valid = false
// TODO(sumeer): why is MarkCorruptionError only being called for
// MERGE?
i.err = base.MarkCorruptionError(i.err)
}
return nil, nil
default:
i.err = base.CorruptionErrorf("invalid internal key kind: %d", errors.Safe(i.iterKey.Kind()))
i.valid = false
return nil, nil
}
}
return nil, nil
}
func (i *compactionIter) closeValueCloser() error {
if i.valueCloser == nil {
return nil
}
i.err = i.valueCloser.Close()
i.valueCloser = nil
if i.err != nil {
i.valid = false
}
return i.err
}
// snapshotIndex returns the index of the first sequence number in snapshots
// which is greater than or equal to seq.
func snapshotIndex(seq uint64, snapshots []uint64) (int, uint64) {
index := sort.Search(len(snapshots), func(i int) bool {
return snapshots[i] > seq
})
if index >= len(snapshots) {
return index, InternalKeySeqNumMax
}
return index, snapshots[index]
}
// skipInStripe skips over skippable keys in the same stripe and user key. It
// may set i.err, in which case i.iterKey will be nil.
func (i *compactionIter) skipInStripe() {
i.skip = true
// TODO(sumeer): we can avoid the overhead of calling i.rangeDelFrag.Covers,
// in this case of nextInStripe, since we are skipping all of them anyway.
for i.nextInStripe() == sameStripe {
if i.err != nil {
panic(i.err)
}
}
// We landed outside the original stripe, so reset skip.
i.skip = false
}
func (i *compactionIter) iterNext() bool {
var iterValue LazyValue
i.iterKey, iterValue = i.iter.Next()
i.iterValue, _, i.err = iterValue.Value(nil)
if i.err != nil {
i.iterKey = nil
}
return i.iterKey != nil
}
// stripeChangeType indicates how the snapshot stripe changed relative to the
// previous key. If the snapshot stripe changed, it also indicates whether the
// new stripe was entered because the iterator progressed onto an entirely new
// key or entered a new stripe within the same key.
type stripeChangeType int
const (
newStripeNewKey stripeChangeType = iota
newStripeSameKey
sameStripe
)
// nextInStripe advances the iterator and returns one of the above const ints
// indicating how its state changed.
//
// All sameStripe keys that are covered by a RANGEDEL will be skipped and not
// returned.
//
// Calls to nextInStripe must be preceded by a call to saveKey to retain a
// temporary reference to the original key, so that forward iteration can
// proceed with a reference to the original key. Care should be taken to avoid
// overwriting or mutating the saved key or value before they have been returned
// to the caller of the exported function (i.e. the caller of Next, First, etc.)
//
// nextInStripe may set i.err, in which case the return value will be
// newStripeNewKey, and i.iterKey will be nil.
func (i *compactionIter) nextInStripe() stripeChangeType {
i.iterStripeChange = i.nextInStripeHelper()
return i.iterStripeChange
}
// nextInStripeHelper is an internal helper for nextInStripe; callers should use
// nextInStripe and not call nextInStripeHelper.
func (i *compactionIter) nextInStripeHelper() stripeChangeType {
origSnapshotIdx := i.curSnapshotIdx
for {
if !i.iterNext() {
return newStripeNewKey
}
key := i.iterKey
// Is this a new key? There are two cases:
//
// 1. The new key has a different user key.
// 2. The previous key was an interleaved range deletion or range key
// boundary. These keys are interleaved in the same input iterator
// stream as point keys, but they do not obey the ordinary sequence
// number ordering within a user key. If the previous key was one
// of these keys, we consider the new key a `newStripeNewKey` to
// reflect that it's the beginning of a new stream of point keys.
if i.key.IsExclusiveSentinel() || !i.equal(i.key.UserKey, key.UserKey) {
i.curSnapshotIdx, i.curSnapshotSeqNum = snapshotIndex(key.SeqNum(), i.snapshots)
return newStripeNewKey
}
// If i.key and key have the same user key, then
// 1. i.key must not have had a zero sequence number (or it would've be the last
// key with its user key).
// 2. i.key must have a strictly larger sequence number
// There's an exception in that either key may be a range delete. Range
// deletes may share a sequence number with a point key if the keys were
// ingested together. Range keys may also share the sequence number if they
// were ingested, but range keys are interleaved into the compaction
// iterator's input iterator at the maximal sequence number so their
// original sequence number will not be observed here.
if prevSeqNum := base.SeqNumFromTrailer(i.keyTrailer); (prevSeqNum == 0 || prevSeqNum <= key.SeqNum()) &&
i.key.Kind() != InternalKeyKindRangeDelete && key.Kind() != InternalKeyKindRangeDelete {
prevKey := i.key
prevKey.Trailer = i.keyTrailer
panic(errors.AssertionFailedf("pebble: invariant violation: %s and %s out of order", prevKey, key))
}
i.curSnapshotIdx, i.curSnapshotSeqNum = snapshotIndex(key.SeqNum(), i.snapshots)
switch key.Kind() {
case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete,
InternalKeyKindRangeDelete:
// Range tombstones and range keys are interleaved at the max
// sequence number for a given user key, and the first key after one
// is always considered a newStripeNewKey, so we should never reach
// this.
panic("unreachable")
case InternalKeyKindDelete, InternalKeyKindSet, InternalKeyKindMerge, InternalKeyKindSingleDelete,
InternalKeyKindSetWithDelete, InternalKeyKindDeleteSized:
// Fall through
default:
kind := i.iterKey.Kind()
i.iterKey = nil
i.err = base.CorruptionErrorf("invalid internal key kind: %d", errors.Safe(kind))
i.valid = false
return newStripeNewKey
}
if i.curSnapshotIdx == origSnapshotIdx {
// Same snapshot.
if i.rangeDelFrag.Covers(*i.iterKey, i.curSnapshotSeqNum) == keyspan.CoversVisibly {
continue
}
return sameStripe
}
return newStripeSameKey
}
}
func (i *compactionIter) setNext() {
// Save the current key.
i.saveKey()
i.value = i.iterValue
i.valid = true
i.maybeZeroSeqnum(i.curSnapshotIdx)
// If this key is already a SETWITHDEL we can early return and skip the remaining
// records in the stripe:
if i.iterKey.Kind() == InternalKeyKindSetWithDelete {
i.skip = true
return
}
// We are iterating forward. Save the current value.
i.valueBuf = append(i.valueBuf[:0], i.iterValue...)
i.value = i.valueBuf
// Else, we continue to loop through entries in the stripe looking for a
// DEL. Note that we may stop *before* encountering a DEL, if one exists.
//
// NB: nextInStripe will skip sameStripe keys that are visibly covered by a
// RANGEDEL. This can include DELs -- this is fine since such DELs don't
// need to be combined with SET to make SETWITHDEL.
for {
switch i.nextInStripe() {
case newStripeNewKey, newStripeSameKey:
i.pos = iterPosNext
return
case sameStripe:
// We're still in the same stripe. If this is a
// DEL/SINGLEDEL/DELSIZED, we stop looking and emit a SETWITHDEL.
// Subsequent keys are eligible for skipping.
switch i.iterKey.Kind() {
case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
i.key.SetKind(InternalKeyKindSetWithDelete)
i.skip = true
return
case InternalKeyKindSet, InternalKeyKindMerge, InternalKeyKindSetWithDelete:
// Do nothing
default:
i.err = base.CorruptionErrorf("invalid internal key kind: %d", errors.Safe(i.iterKey.Kind()))
i.valid = false
}
default:
panic("pebble: unexpected stripeChangeType: " + strconv.Itoa(int(i.iterStripeChange)))
}
}
}
func (i *compactionIter) mergeNext(valueMerger ValueMerger) {
// Save the current key.
i.saveKey()
i.valid = true
// Loop looking for older values in the current snapshot stripe and merge
// them.
for {
if i.nextInStripe() != sameStripe {
i.pos = iterPosNext
return
}
if i.err != nil {
panic(i.err)
}
// NB: MERGE#10+RANGEDEL#9 stays a MERGE, since nextInStripe skips
// sameStripe keys that are visibly covered by a RANGEDEL. There may be
// MERGE#7 that is invisibly covered and will be preserved, but there is
// no risk that MERGE#10 and MERGE#7 will get merged in the future as
// the RANGEDEL still exists and will be used in user-facing reads that
// see MERGE#10, and will also eventually cause MERGE#7 to be deleted in
// a compaction.
key := i.iterKey
switch key.Kind() {
case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
// We've hit a deletion tombstone. Return everything up to this point and
// then skip entries until the next snapshot stripe. We change the kind
// of the result key to a Set so that it shadows keys in lower
// levels. That is, MERGE+DEL -> SETWITHDEL.
//
// We do the same for SingleDelete since SingleDelete is only
// permitted (with deterministic behavior) for keys that have been
// set once since the last SingleDelete/Delete, so everything
// older is acceptable to shadow. Note that this is slightly
// different from singleDeleteNext() which implements stricter
// semantics in terms of applying the SingleDelete to the single
// next Set. But those stricter semantics are not observable to
// the end-user since Iterator interprets SingleDelete as Delete.
// We could do something more complicated here and consume only a
// single Set, and then merge in any following Sets, but that is
// complicated wrt code and unnecessary given the narrow permitted
// use of SingleDelete.
i.key.SetKind(InternalKeyKindSetWithDelete)
i.skip = true
return
case InternalKeyKindSet, InternalKeyKindSetWithDelete:
// We've hit a Set or SetWithDel value. Merge with the existing
// value and return. We change the kind of the resulting key to a
// Set so that it shadows keys in lower levels. That is:
// MERGE + (SET*) -> SET.
i.err = valueMerger.MergeOlder(i.iterValue)
if i.err != nil {
i.valid = false
return
}
i.key.SetKind(InternalKeyKindSet)
i.skip = true
return
case InternalKeyKindMerge:
// We've hit another Merge value. Merge with the existing value and
// continue looping.
i.err = valueMerger.MergeOlder(i.iterValue)
if i.err != nil {
i.valid = false
return
}
default:
i.err = base.CorruptionErrorf("invalid internal key kind: %d", errors.Safe(i.iterKey.Kind()))
i.valid = false
return
}
}
}
// singleDeleteNext processes a SingleDelete point tombstone. A SingleDelete, or
// SINGLEDEL, is unique in that it deletes exactly 1 internal key. It's a
// performance optimization when the client knows a user key has not been
// overwritten, allowing the elision of the tombstone earlier, avoiding write
// amplification.
//
// singleDeleteNext returns a boolean indicating whether or not the caller
// should yield the SingleDelete key to the consumer of the compactionIter. If
// singleDeleteNext returns false, the caller may consume/elide the
// SingleDelete.
func (i *compactionIter) singleDeleteNext() bool {
// Save the current key.
i.saveKey()
i.value = i.iterValue
i.valid = true
// Loop until finds a key to be passed to the next level.
for {
// If we find a key that can't be skipped, return true so that the
// caller yields the SingleDelete to the caller.
if i.nextInStripe() != sameStripe {
// This defers additional error checking regarding single delete
// invariants to the compaction where the keys with the same user key as
// the single delete are in the same stripe.
i.pos = iterPosNext
return i.err == nil
}
if i.err != nil {
panic(i.err)
}
// INVARIANT: sameStripe.
key := i.iterKey
kind := key.Kind()
switch kind {
case InternalKeyKindDelete, InternalKeyKindSetWithDelete, InternalKeyKindDeleteSized:
if (kind == InternalKeyKindDelete || kind == InternalKeyKindDeleteSized) &&
i.ineffectualSingleDeleteCallback != nil {
i.ineffectualSingleDeleteCallback(i.key.UserKey)
}
// We've hit a Delete, DeleteSized, SetWithDelete, transform
// the SingleDelete into a full Delete.
i.key.SetKind(InternalKeyKindDelete)
i.skip = true
return true
case InternalKeyKindSet, InternalKeyKindMerge:
// This SingleDelete deletes the Set/Merge, and we can now elide the
// SingleDel as well. We advance past the Set and return false to
// indicate to the main compaction loop that we should NOT yield the
// current SingleDel key to the compaction loop.
//
// NB: singleDeleteNext was called with i.pos == iterPosCurForward, and
// after the call to nextInStripe, we are still at iterPosCurForward,
// since we are at the key after the Set/Merge that was single deleted.
change := i.nextInStripe()
switch change {
case sameStripe, newStripeSameKey:
// On the same user key.
nextKind := i.iterKey.Kind()
switch nextKind {
case InternalKeyKindSet, InternalKeyKindSetWithDelete, InternalKeyKindMerge:
if i.singleDeleteInvariantViolationCallback != nil {
// sameStripe keys returned by nextInStripe() are already
// known to not be covered by a RANGEDEL, so it is an invariant
// violation. The rare case is newStripeSameKey, where it is a
// violation if not covered by a RANGEDEL.
if change == sameStripe ||
i.rangeDelFrag.Covers(*i.iterKey, i.curSnapshotSeqNum) == keyspan.NoCover {
i.singleDeleteInvariantViolationCallback(i.key.UserKey)
}
}
case InternalKeyKindDelete, InternalKeyKindDeleteSized, InternalKeyKindSingleDelete:
default:
panic(errors.AssertionFailedf(
"unexpected internal key kind: %d", errors.Safe(i.iterKey.Kind())))
}
case newStripeNewKey:
default:
panic("unreachable")
}
i.valid = false
return false
case InternalKeyKindSingleDelete:
// Two single deletes met in a compaction. The first single delete is
// ineffectual.
if i.ineffectualSingleDeleteCallback != nil {
i.ineffectualSingleDeleteCallback(i.key.UserKey)
}
// Continue to apply the second single delete.
continue
default:
i.err = base.CorruptionErrorf("invalid internal key kind: %d", errors.Safe(i.iterKey.Kind()))
i.valid = false
return false
}
}
}
// skipDueToSingleDeleteElision is called when the SingleDelete is being
// elided because it is in the final snapshot stripe and there are no keys
// with the same user key in lower levels in the LSM (below the files in this
// compaction).
//
// TODO(sumeer): the only difference between singleDeleteNext and
// skipDueToSingleDeleteElision is the fact that the caller knows it will be
// eliding the single delete in the latter case. There are some similar things
// happening in both implementations. My first attempt at combining them into
// a single method was hard to comprehend. Try again.
func (i *compactionIter) skipDueToSingleDeleteElision() {
for {
stripeChange := i.nextInStripe()
if i.err != nil {
panic(i.err)
}
switch stripeChange {
case newStripeNewKey:
// The single delete is only now being elided, meaning it did not elide
// any keys earlier in its descent down the LSM. We stepped onto a new
// user key, meaning that even now at its moment of elision, it still
// hasn't elided any other keys. The single delete was ineffectual (a
// no-op).
if i.ineffectualSingleDeleteCallback != nil {
i.ineffectualSingleDeleteCallback(i.key.UserKey)
}
i.skip = false
return
case newStripeSameKey:
// This should be impossible. If we're eliding a single delete, we
// determined that the tombstone is in the final snapshot stripe, but we
// stepped into a new stripe of the same key.
panic(errors.AssertionFailedf("eliding single delete followed by same key in new stripe"))
case sameStripe:
kind := i.iterKey.Kind()
switch kind {
case InternalKeyKindDelete, InternalKeyKindDeleteSized, InternalKeyKindSingleDelete:
if i.ineffectualSingleDeleteCallback != nil {
i.ineffectualSingleDeleteCallback(i.key.UserKey)
}
switch kind {
case InternalKeyKindDelete, InternalKeyKindDeleteSized:
i.skipInStripe()
return
case InternalKeyKindSingleDelete:
// Repeat the same with this SingleDelete. We don't want to simply
// call skipInStripe(), since it increases the strength of the
// SingleDel, which hides bugs in the use of single delete.
continue
default:
panic(errors.AssertionFailedf(
"unexpected internal key kind: %d", errors.Safe(i.iterKey.Kind())))
}
case InternalKeyKindSetWithDelete:
// The SingleDelete should behave like a Delete.
i.skipInStripe()
return
case InternalKeyKindSet, InternalKeyKindMerge:
// This SingleDelete deletes the Set/Merge, and we are eliding the
// SingleDel as well. Step to the next key (this is not deleted by the
// SingleDelete).
//
// NB: skipDueToSingleDeleteElision was called with i.pos ==
// iterPosCurForward, and after the call to nextInStripe, we are still
// at iterPosCurForward, since we are at the key after the Set/Merge
// that was single deleted.
change := i.nextInStripe()
if i.err != nil {
panic(i.err)
}
switch change {
case newStripeSameKey:
panic(errors.AssertionFailedf("eliding single delete followed by same key in new stripe"))
case newStripeNewKey: