Line data Source code
1 : // Copyright 2021 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
6 :
7 : import (
8 : "context"
9 :
10 : "github.com/cockroachdb/errors"
11 : "github.com/cockroachdb/pebble/internal/base"
12 : "github.com/cockroachdb/pebble/internal/invariants"
13 : "github.com/cockroachdb/pebble/internal/keyspan"
14 : "github.com/cockroachdb/pebble/internal/manifest"
15 : "github.com/cockroachdb/pebble/sstable"
16 : )
17 :
18 : // constructRangeKeyIter constructs the range-key iterator stack, populating
19 : // i.rangeKey.rangeKeyIter with the resulting iterator.
20 2 : func (i *Iterator) constructRangeKeyIter() {
21 2 : i.rangeKey.rangeKeyIter = i.rangeKey.iterConfig.Init(
22 2 : &i.comparer, i.seqNum, i.opts.LowerBound, i.opts.UpperBound,
23 2 : &i.hasPrefix, &i.prefixOrFullSeekKey, false /* internalKeys */, &i.rangeKey.rangeKeyBuffers.internal)
24 2 :
25 2 : // If there's an indexed batch with range keys, include it.
26 2 : if i.batch != nil {
27 2 : if i.batch.index == nil {
28 0 : // This isn't an indexed batch. We shouldn't have gotten this far.
29 0 : panic(errors.AssertionFailedf("creating an iterator over an unindexed batch"))
30 2 : } else {
31 2 : // Only include the batch's range key iterator if it has any keys.
32 2 : // NB: This can force reconstruction of the rangekey iterator stack
33 2 : // in SetOptions if subsequently range keys are added. See
34 2 : // SetOptions.
35 2 : if i.batch.countRangeKeys > 0 {
36 2 : i.batch.initRangeKeyIter(&i.opts, &i.batchRangeKeyIter, i.batchSeqNum)
37 2 : i.rangeKey.iterConfig.AddLevel(&i.batchRangeKeyIter)
38 2 : }
39 : }
40 : }
41 :
42 2 : if !i.batchOnlyIter {
43 2 : // Next are the flushables: memtables and large batches.
44 2 : if i.readState != nil {
45 2 : for j := len(i.readState.memtables) - 1; j >= 0; j-- {
46 2 : mem := i.readState.memtables[j]
47 2 : // We only need to read from memtables which contain sequence numbers older
48 2 : // than seqNum.
49 2 : if logSeqNum := mem.logSeqNum; logSeqNum >= i.seqNum {
50 2 : continue
51 : }
52 2 : if rki := mem.newRangeKeyIter(&i.opts); rki != nil {
53 2 : i.rangeKey.iterConfig.AddLevel(rki)
54 2 : }
55 : }
56 : }
57 :
58 2 : current := i.version
59 2 : if current == nil {
60 2 : current = i.readState.current
61 2 : }
62 : // Next are the file levels: L0 sub-levels followed by lower levels.
63 :
64 : // Add file-specific iterators for L0 files containing range keys. We
65 : // maintain a separate manifest.LevelMetadata for each level containing only
66 : // files that contain range keys, however we don't compute a separate
67 : // L0Sublevels data structure too.
68 : //
69 : // We first use L0's LevelMetadata to peek and see whether L0 contains any
70 : // range keys at all. If it does, we create a range key level iterator per
71 : // level that contains range keys using the information from L0Sublevels.
72 : // Some sublevels may not contain any range keys, and we need to iterate
73 : // through the fileMetadata to determine that. Since L0's file count should
74 : // not significantly exceed ~1000 files (see L0CompactionFileThreshold),
75 : // this should be okay.
76 2 : if !current.RangeKeyLevels[0].Empty() {
77 2 : // L0 contains at least 1 file containing range keys.
78 2 : // Add level iterators for the L0 sublevels, iterating from newest to
79 2 : // oldest.
80 2 : for j := len(current.L0SublevelFiles) - 1; j >= 0; j-- {
81 2 : iter := current.L0SublevelFiles[j].Iter()
82 2 : if !containsAnyRangeKeys(iter) {
83 2 : continue
84 : }
85 :
86 2 : li := i.rangeKey.iterConfig.NewLevelIter()
87 2 : li.Init(
88 2 : i.opts.SpanIterOptions(),
89 2 : i.cmp,
90 2 : i.newIterRangeKey,
91 2 : iter.Filter(manifest.KeyTypeRange),
92 2 : manifest.L0Sublevel(j),
93 2 : manifest.KeyTypeRange,
94 2 : )
95 2 : i.rangeKey.iterConfig.AddLevel(li)
96 : }
97 : }
98 :
99 : // Add level iterators for the non-empty non-L0 levels.
100 2 : for level := 1; level < len(current.RangeKeyLevels); level++ {
101 2 : if current.RangeKeyLevels[level].Empty() {
102 2 : continue
103 : }
104 2 : li := i.rangeKey.iterConfig.NewLevelIter()
105 2 : spanIterOpts := i.opts.SpanIterOptions()
106 2 : li.Init(spanIterOpts, i.cmp, i.newIterRangeKey, current.RangeKeyLevels[level].Iter(),
107 2 : manifest.Level(level), manifest.KeyTypeRange)
108 2 : i.rangeKey.iterConfig.AddLevel(li)
109 : }
110 : }
111 : }
112 :
113 2 : func containsAnyRangeKeys(iter manifest.LevelIterator) bool {
114 2 : for f := iter.First(); f != nil; f = iter.Next() {
115 2 : if f.HasRangeKeys {
116 2 : return true
117 2 : }
118 : }
119 2 : return false
120 : }
121 :
122 : // Range key masking
123 : //
124 : // Pebble iterators may be configured such that range keys with suffixes mask
125 : // point keys with lower suffixes. The intended use is implementing a MVCC
126 : // delete range operation using range keys, when suffixes are MVCC timestamps.
127 : //
128 : // To enable masking, the user populates the IterOptions's RangeKeyMasking
129 : // field. The Suffix field configures which range keys act as masks. The
130 : // intended use is to hold a MVCC read timestamp. When implementing a MVCC
131 : // delete range operation, only range keys that are visible at the read
132 : // timestamp should be visible. If a range key has a suffix ≤
133 : // RangeKeyMasking.Suffix, it acts as a mask.
134 : //
135 : // Range key masking is facilitated by the keyspan.InterleavingIter. The
136 : // interleaving iterator interleaves range keys and point keys during combined
137 : // iteration. During user iteration, the interleaving iterator is configured
138 : // with a keyspan.SpanMask, implemented by the rangeKeyMasking struct below.
139 : // The SpanMask interface defines two methods: SpanChanged and SkipPoint.
140 : //
141 : // SpanChanged is used to keep the current mask up-to-date. Whenever the point
142 : // iterator has stepped into or out of the bounds of a range key, the
143 : // interleaving iterator invokes SpanChanged passing the current covering range
144 : // key. The below rangeKeyMasking implementation scans the range keys looking
145 : // for the range key with the largest suffix that's still ≤ the suffix supplied
146 : // to IterOptions.RangeKeyMasking.Suffix (the "read timestamp"). If it finds a
147 : // range key that meets the condition, the range key should act as a mask. The
148 : // span and the relevant range key's suffix are saved.
149 : //
150 : // The above ensures that `rangeKeyMasking.maskActiveSuffix` always contains the
151 : // current masking suffix such that any point keys with lower suffixes should be
152 : // skipped.
153 : //
154 : // There are two ways in which masked point keys are skipped.
155 : //
156 : // 1. Interleaving iterator SkipPoint
157 : //
158 : // Whenever the interleaving iterator encounters a point key that falls within
159 : // the bounds of a range key, it invokes SkipPoint. The interleaving iterator
160 : // guarantees that the SpanChanged method described above has already been
161 : // invoked with the covering range key. The below rangeKeyMasking implementation
162 : // of SkipPoint splits the key into prefix and suffix, compares the suffix to
163 : // the `maskActiveSuffix` updated by SpanChanged and returns true if
164 : // suffix(point) < maskActiveSuffix.
165 : //
166 : // The SkipPoint logic is sufficient to ensure that the Pebble iterator filters
167 : // out all masked point keys. However, it requires the iterator read each masked
168 : // point key. For broad range keys that mask many points, this may be expensive.
169 : //
170 : // 2. Block property filter
171 : //
172 : // For more efficient handling of braad range keys that mask many points, the
173 : // IterOptions.RangeKeyMasking field has an optional Filter option. This Filter
174 : // field takes a superset of the block-property filter interface, adding a
175 : // method to dynamically configure the filter's filtering criteria.
176 : //
177 : // To make use of the Filter option, the user is required to define and
178 : // configure a block-property collector that collects a property containing at
179 : // least the maximum suffix of a key within a block.
180 : //
181 : // When the SpanChanged method described above is invoked, rangeKeyMasking also
182 : // reconfigures the user-provided filter. It invokes a SetSuffix method,
183 : // providing the `maskActiveSuffix`, requesting that from now on the
184 : // block-property filter return Intersects()=false for any properties indicating
185 : // that a block contains exclusively keys with suffixes greater than the
186 : // provided suffix.
187 : //
188 : // Note that unlike other block-property filters, the filter used for masking
189 : // must not apply across the entire keyspace. It must only filter blocks that
190 : // lie within the bounds of the range key that set the mask suffix. To
191 : // accommodate this, rangeKeyMasking implements a special interface:
192 : // sstable.BoundLimitedBlockPropertyFilter. This interface extends the block
193 : // property filter interface with two new methods: KeyIsWithinLowerBound and
194 : // KeyIsWithinUpperBound. The rangeKeyMasking type wraps the user-provided block
195 : // property filter, implementing these two methods and overriding Intersects to
196 : // always return true if there is no active mask.
197 : //
198 : // The logic to ensure that a mask block-property filter is only applied within
199 : // the bounds of the masking range key is subtle. The interleaving iterator
200 : // guarantees that it never invokes SpanChanged until the point iterator is
201 : // positioned within the range key. During forward iteration, this guarantees
202 : // that any block that a sstable reader might attempt to load contains only keys
203 : // greater than or equal to the range key's lower bound. During backward
204 : // iteration, it provides the analagous guarantee on the range key's upper
205 : // bound.
206 : //
207 : // The above ensures that an sstable reader only needs to verify that a block
208 : // that it skips meets the opposite bound. This is where the
209 : // KeyIsWithinLowerBound and KeyIsWithinUpperBound methods are used. When an
210 : // sstable iterator is configured with a BoundLimitedBlockPropertyFilter, it
211 : // checks for intersection with the block-property filter before every block
212 : // load, like ordinary block-property filters. However, if the bound-limited
213 : // block property filter indicates that it does NOT intersect, the filter's
214 : // relevant KeyIsWithin{Lower,Upper}Bound method is queried, using a block
215 : // index separator as the bound. If the method indicates that the provided index
216 : // separator does not fall within the range key bounds, the no-intersection
217 : // result is ignored, and the block is read.
218 :
219 : type rangeKeyMasking struct {
220 : cmp base.Compare
221 : split base.Split
222 : filter BlockPropertyFilterMask
223 : // maskActiveSuffix holds the suffix of a range key currently acting as a
224 : // mask, hiding point keys with suffixes greater than it. maskActiveSuffix
225 : // is only ever non-nil if IterOptions.RangeKeyMasking.Suffix is non-nil.
226 : // maskActiveSuffix is updated whenever the iterator passes over a new range
227 : // key. The maskActiveSuffix should only be used if maskSpan is non-nil.
228 : //
229 : // See SpanChanged.
230 : maskActiveSuffix []byte
231 : // maskSpan holds the span from which the active mask suffix was extracted.
232 : // The span is used for bounds comparisons, to ensure that a range-key mask
233 : // is not applied beyond the bounds of the range key.
234 : maskSpan *keyspan.Span
235 : parent *Iterator
236 : }
237 :
238 2 : func (m *rangeKeyMasking) init(parent *Iterator, cmp base.Compare, split base.Split) {
239 2 : m.cmp = cmp
240 2 : m.split = split
241 2 : if parent.opts.RangeKeyMasking.Filter != nil {
242 2 : m.filter = parent.opts.RangeKeyMasking.Filter()
243 2 : }
244 2 : m.parent = parent
245 : }
246 :
247 : // SpanChanged implements the keyspan.SpanMask interface, used during range key
248 : // iteration.
249 2 : func (m *rangeKeyMasking) SpanChanged(s *keyspan.Span) {
250 2 : if s == nil && m.maskSpan == nil {
251 2 : return
252 2 : }
253 2 : m.maskSpan = nil
254 2 : m.maskActiveSuffix = m.maskActiveSuffix[:0]
255 2 :
256 2 : // Find the smallest suffix of a range key contained within the Span,
257 2 : // excluding suffixes less than m.opts.RangeKeyMasking.Suffix.
258 2 : if s != nil {
259 2 : m.parent.rangeKey.stale = true
260 2 : if m.parent.opts.RangeKeyMasking.Suffix != nil {
261 2 : for j := range s.Keys {
262 2 : if s.Keys[j].Suffix == nil {
263 0 : continue
264 : }
265 2 : if m.cmp(s.Keys[j].Suffix, m.parent.opts.RangeKeyMasking.Suffix) < 0 {
266 2 : continue
267 : }
268 2 : if len(m.maskActiveSuffix) == 0 || m.cmp(m.maskActiveSuffix, s.Keys[j].Suffix) > 0 {
269 2 : m.maskSpan = s
270 2 : m.maskActiveSuffix = append(m.maskActiveSuffix[:0], s.Keys[j].Suffix...)
271 2 : }
272 : }
273 : }
274 : }
275 :
276 2 : if m.maskSpan != nil && m.parent.opts.RangeKeyMasking.Filter != nil {
277 2 : // Update the block-property filter to filter point keys with suffixes
278 2 : // greater than m.maskActiveSuffix.
279 2 : err := m.filter.SetSuffix(m.maskActiveSuffix)
280 2 : if err != nil {
281 0 : m.parent.err = err
282 0 : }
283 : }
284 : // If no span is active, we leave the inner block-property filter configured
285 : // with its existing suffix. That's okay, because Intersects calls are first
286 : // evaluated by iteratorRangeKeyState.Intersects, which considers all blocks
287 : // as intersecting if there's no active mask.
288 : }
289 :
290 : // SkipPoint implements the keyspan.SpanMask interface, used during range key
291 : // iteration. Whenever a point key is covered by a non-empty Span, the
292 : // interleaving iterator invokes SkipPoint. This function is responsible for
293 : // performing range key masking.
294 : //
295 : // If a non-nil IterOptions.RangeKeyMasking.Suffix is set, range key masking is
296 : // enabled. Masking hides point keys, transparently skipping over the keys.
297 : // Whether or not a point key is masked is determined by comparing the point
298 : // key's suffix, the overlapping span's keys' suffixes, and the user-configured
299 : // IterOption's RangeKeyMasking.Suffix. When configured with a masking threshold
300 : // _t_, and there exists a span with suffix _r_ covering a point key with suffix
301 : // _p_, and
302 : //
303 : // _t_ ≤ _r_ < _p_
304 : //
305 : // then the point key is elided. Consider the following rendering, where using
306 : // integer suffixes with higher integers sort before suffixes with lower
307 : // integers, (for example @7 ≤ @6 < @5):
308 : //
309 : // ^
310 : // @9 | •―――――――――――――――○ [e,m)@9
311 : // s 8 | • l@8
312 : // u 7 |------------------------------------ @7 RangeKeyMasking.Suffix
313 : // f 6 | [h,q)@6 •―――――――――――――――――○ (threshold)
314 : // f 5 | • h@5
315 : // f 4 | • n@4
316 : // i 3 | •―――――――――――○ [f,l)@3
317 : // x 2 | • b@2
318 : // 1 |
319 : // 0 |___________________________________
320 : // a b c d e f g h i j k l m n o p q
321 : //
322 : // An iterator scanning the entire keyspace with the masking threshold set to @7
323 : // will observe point keys b@2 and l@8. The span keys [h,q)@6 and [f,l)@3 serve
324 : // as masks, because cmp(@6,@7) ≥ 0 and cmp(@3,@7) ≥ 0. The span key [e,m)@9
325 : // does not serve as a mask, because cmp(@9,@7) < 0.
326 : //
327 : // Although point l@8 falls within the user key bounds of [e,m)@9, [e,m)@9 is
328 : // non-masking due to its suffix. The point key l@8 also falls within the user
329 : // key bounds of [h,q)@6, but since cmp(@6,@8) ≥ 0, l@8 is unmasked.
330 : //
331 : // Invariant: The userKey is within the user key bounds of the span most
332 : // recently provided to `SpanChanged`.
333 2 : func (m *rangeKeyMasking) SkipPoint(userKey []byte) bool {
334 2 : m.parent.stats.RangeKeyStats.ContainedPoints++
335 2 : if m.maskSpan == nil {
336 2 : // No range key is currently acting as a mask, so don't skip.
337 2 : return false
338 2 : }
339 : // Range key masking is enabled and the current span includes a range key
340 : // that is being used as a mask. (NB: SpanChanged already verified that the
341 : // range key's suffix is ≥ RangeKeyMasking.Suffix).
342 : //
343 : // This point key falls within the bounds of the range key (guaranteed by
344 : // the InterleavingIter). Skip the point key if the range key's suffix is
345 : // greater than the point key's suffix.
346 2 : pointSuffix := userKey[m.split(userKey):]
347 2 : if len(pointSuffix) > 0 && m.cmp(m.maskActiveSuffix, pointSuffix) < 0 {
348 2 : m.parent.stats.RangeKeyStats.SkippedPoints++
349 2 : return true
350 2 : }
351 2 : return false
352 : }
353 :
354 : // The iteratorRangeKeyState type implements the sstable package's
355 : // BoundLimitedBlockPropertyFilter interface in order to use block property
356 : // filters for range key masking. The iteratorRangeKeyState implementation wraps
357 : // the block-property filter provided in Options.RangeKeyMasking.Filter.
358 : //
359 : // Using a block-property filter for range-key masking requires limiting the
360 : // filter's effect to the bounds of the range key currently acting as a mask.
361 : // Consider the range key [a,m)@10, and an iterator positioned just before the
362 : // below block, bounded by index separators `c` and `z`:
363 : //
364 : // c z
365 : // x | c@9 c@5 c@1 d@7 e@4 y@4 | ...
366 : // iter pos
367 : //
368 : // The next block cannot be skipped, despite the range key suffix @10 is greater
369 : // than all the block's keys' suffixes, because it contains a key (y@4) outside
370 : // the bounds of the range key.
371 : //
372 : // This extended BoundLimitedBlockPropertyFilter interface adds two new methods,
373 : // KeyIsWithinLowerBound and KeyIsWithinUpperBound, for testing whether a
374 : // particular block is within bounds.
375 : //
376 : // The iteratorRangeKeyState implements these new methods by first checking if
377 : // the iterator is currently positioned within a range key. If not, the provided
378 : // key is considered out-of-bounds. If the iterator is positioned within a range
379 : // key, it compares the corresponding range key bound.
380 : var _ sstable.BoundLimitedBlockPropertyFilter = (*rangeKeyMasking)(nil)
381 :
382 : // Name implements the limitedBlockPropertyFilter interface defined in the
383 : // sstable package by passing through to the user-defined block property filter.
384 2 : func (m *rangeKeyMasking) Name() string {
385 2 : return m.filter.Name()
386 2 : }
387 :
388 : // Intersects implements the limitedBlockPropertyFilter interface defined in the
389 : // sstable package by passing the intersection decision to the user-provided
390 : // block property filter only if a range key is covering the current iterator
391 : // position.
392 2 : func (m *rangeKeyMasking) Intersects(prop []byte) (bool, error) {
393 2 : if m.maskSpan == nil {
394 2 : // No span is actively masking.
395 2 : return true, nil
396 2 : }
397 2 : return m.filter.Intersects(prop)
398 : }
399 :
400 : // KeyIsWithinLowerBound implements the limitedBlockPropertyFilter interface
401 : // defined in the sstable package. It's used to restrict the masking block
402 : // property filter to only applying within the bounds of the active range key.
403 2 : func (m *rangeKeyMasking) KeyIsWithinLowerBound(key []byte) bool {
404 2 : // Invariant: m.maskSpan != nil
405 2 : //
406 2 : // The provided `key` is an inclusive lower bound of the block we're
407 2 : // considering skipping.
408 2 : return m.cmp(m.maskSpan.Start, key) <= 0
409 2 : }
410 :
411 : // KeyIsWithinUpperBound implements the limitedBlockPropertyFilter interface
412 : // defined in the sstable package. It's used to restrict the masking block
413 : // property filter to only applying within the bounds of the active range key.
414 2 : func (m *rangeKeyMasking) KeyIsWithinUpperBound(key []byte) bool {
415 2 : // Invariant: m.maskSpan != nil
416 2 : //
417 2 : // The provided `key` is an *inclusive* upper bound of the block we're
418 2 : // considering skipping, so the range key's end must be strictly greater
419 2 : // than the block bound for the block to be within bounds.
420 2 : return m.cmp(m.maskSpan.End, key) > 0
421 2 : }
422 :
423 : // lazyCombinedIter implements the internalIterator interface, wrapping a
424 : // pointIter. It requires the pointIter's the levelIters be configured with
425 : // pointers to its combinedIterState. When the levelIter observes a file
426 : // containing a range key, the lazyCombinedIter constructs the combined
427 : // range+point key iterator stack and switches to it.
428 : type lazyCombinedIter struct {
429 : // parent holds a pointer to the root *pebble.Iterator containing this
430 : // iterator. It's used to mutate the internalIterator in use when switching
431 : // to combined iteration.
432 : parent *Iterator
433 : pointIter internalIterator
434 : combinedIterState combinedIterState
435 : }
436 :
437 : // combinedIterState encapsulates the current state of combined iteration.
438 : // Various low-level iterators (mergingIter, leveliter) hold pointers to the
439 : // *pebble.Iterator's combinedIterState. This allows them to check whether or
440 : // not they must monitor for files containing range keys (!initialized), or not.
441 : //
442 : // When !initialized, low-level iterators watch for files containing range keys.
443 : // When one is discovered, they set triggered=true and key to the smallest
444 : // (forward direction) or largest (reverse direction) range key that's been
445 : // observed.
446 : type combinedIterState struct {
447 : // key holds the smallest (forward direction) or largest (backward
448 : // direction) user key from a range key bound discovered during the iterator
449 : // operation that triggered the switch to combined iteration.
450 : //
451 : // Slices stored here must be stable. This is possible because callers pass
452 : // a Smallest/Largest bound from a fileMetadata, which are immutable. A key
453 : // slice's bytes must not be overwritten.
454 : key []byte
455 : triggered bool
456 : initialized bool
457 : }
458 :
459 : // Assert that *lazyCombinedIter implements internalIterator.
460 : var _ internalIterator = (*lazyCombinedIter)(nil)
461 :
462 : // initCombinedIteration is invoked after a pointIter positioning operation
463 : // resulted in i.combinedIterState.triggered=true.
464 : //
465 : // The `dir` parameter is `+1` or `-1` indicating forward iteration or backward
466 : // iteration respectively.
467 : //
468 : // The `pointKey` and `pointValue` parameters provide the new point key-value
469 : // pair that the iterator was just positioned to. The combined iterator should
470 : // be seeded with this point key-value pair and return the smaller (forward
471 : // iteration) or largest (backward iteration) of the two.
472 : //
473 : // The `seekKey` parameter is non-nil only if the iterator operation that
474 : // triggered the switch to combined iteration was a SeekGE, SeekPrefixGE or
475 : // SeekLT. It provides the seek key supplied and is used to seek the range-key
476 : // iterator using the same key. This is necessary for SeekGE/SeekPrefixGE
477 : // operations that land in the middle of a range key and must truncate to the
478 : // user-provided seek key.
479 : func (i *lazyCombinedIter) initCombinedIteration(
480 : dir int8, pointKey *InternalKey, pointValue base.LazyValue, seekKey []byte,
481 2 : ) (*InternalKey, base.LazyValue) {
482 2 : // Invariant: i.parent.rangeKey is nil.
483 2 : // Invariant: !i.combinedIterState.initialized.
484 2 : if invariants.Enabled {
485 2 : if i.combinedIterState.initialized {
486 0 : panic("pebble: combined iterator already initialized")
487 : }
488 2 : if i.parent.rangeKey != nil {
489 0 : panic("pebble: iterator already has a range-key iterator stack")
490 : }
491 : }
492 :
493 : // We need to determine the key to seek the range key iterator to. If
494 : // seekKey is not nil, the user-initiated operation that triggered the
495 : // switch to combined iteration was itself a seek, and we can use that key.
496 : // Otherwise, a First/Last or relative positioning operation triggered the
497 : // switch to combined iteration.
498 : //
499 : // The levelIter that observed a file containing range keys populated
500 : // combinedIterState.key with the smallest (forward) or largest (backward)
501 : // range key it observed. If multiple levelIters observed files with range
502 : // keys during the same operation on the mergingIter, combinedIterState.key
503 : // is the smallest [during forward iteration; largest in reverse iteration]
504 : // such key.
505 2 : if seekKey == nil {
506 2 : // Use the levelIter-populated key.
507 2 : seekKey = i.combinedIterState.key
508 2 :
509 2 : // We may need to adjust the levelIter-populated seek key to the
510 2 : // surfaced point key. If the key observed is beyond [in the iteration
511 2 : // direction] the current point key, there may still exist a range key
512 2 : // at an earlier key. Consider the following example:
513 2 : //
514 2 : // L5: 000003:[bar.DEL.5, foo.RANGEKEYSET.9]
515 2 : // L6: 000001:[bar.SET.2] 000002:[bax.RANGEKEYSET.8]
516 2 : //
517 2 : // A call to First() seeks the levels to files L5.000003 and L6.000001.
518 2 : // The L5 levelIter observes that L5.000003 contains the range key with
519 2 : // start key `foo`, and triggers a switch to combined iteration, setting
520 2 : // `combinedIterState.key` = `foo`.
521 2 : //
522 2 : // The L6 levelIter did not observe the true first range key
523 2 : // (bax.RANGEKEYSET.8), because it appears in a later sstable. When the
524 2 : // combined iterator is initialized, the range key iterator must be
525 2 : // seeked to a key that will find `bax`. To accomplish this, we seek the
526 2 : // key instead to `bar`. It is guaranteed that no range key exists
527 2 : // earlier than `bar`, otherwise a levelIter would've observed it and
528 2 : // set `combinedIterState.key` to its start key.
529 2 : if pointKey != nil {
530 2 : if dir == +1 && i.parent.cmp(i.combinedIterState.key, pointKey.UserKey) > 0 {
531 2 : seekKey = pointKey.UserKey
532 2 : } else if dir == -1 && i.parent.cmp(seekKey, pointKey.UserKey) < 0 {
533 2 : seekKey = pointKey.UserKey
534 2 : }
535 : }
536 : }
537 :
538 : // An operation on the point iterator observed a file containing range keys,
539 : // so we must switch to combined interleaving iteration. First, construct
540 : // the range key iterator stack. It must not exist, otherwise we'd already
541 : // be performing combined iteration.
542 2 : i.parent.rangeKey = iterRangeKeyStateAllocPool.Get().(*iteratorRangeKeyState)
543 2 : i.parent.rangeKey.init(i.parent.comparer.Compare, i.parent.comparer.Split, &i.parent.opts)
544 2 : i.parent.constructRangeKeyIter()
545 2 :
546 2 : // Initialize the Iterator's interleaving iterator.
547 2 : i.parent.rangeKey.iiter.Init(
548 2 : &i.parent.comparer, i.parent.pointIter, i.parent.rangeKey.rangeKeyIter,
549 2 : keyspan.InterleavingIterOpts{
550 2 : Mask: &i.parent.rangeKeyMasking,
551 2 : LowerBound: i.parent.opts.LowerBound,
552 2 : UpperBound: i.parent.opts.UpperBound,
553 2 : })
554 2 :
555 2 : // Set the parent's primary iterator to point to the combined, interleaving
556 2 : // iterator that's now initialized with our current state.
557 2 : i.parent.iter = &i.parent.rangeKey.iiter
558 2 : i.combinedIterState.initialized = true
559 2 : i.combinedIterState.key = nil
560 2 :
561 2 : // All future iterator operations will go directly through the combined
562 2 : // iterator.
563 2 : //
564 2 : // Initialize the interleaving iterator. We pass the point key-value pair so
565 2 : // that the interleaving iterator knows where the point iterator is
566 2 : // positioned. Additionally, we pass the seek key to which the range-key
567 2 : // iterator should be seeked in order to initialize its position.
568 2 : //
569 2 : // In the forward direction (invert for backwards), the seek key is a key
570 2 : // guaranteed to find the smallest range key that's greater than the last
571 2 : // key the iterator returned. The range key may be less than pointKey, in
572 2 : // which case the range key will be interleaved next instead of the point
573 2 : // key.
574 2 : if dir == +1 {
575 2 : var prefix []byte
576 2 : if i.parent.hasPrefix {
577 2 : prefix = i.parent.prefixOrFullSeekKey
578 2 : }
579 2 : return i.parent.rangeKey.iiter.InitSeekGE(prefix, seekKey, pointKey, pointValue)
580 : }
581 2 : return i.parent.rangeKey.iiter.InitSeekLT(seekKey, pointKey, pointValue)
582 : }
583 :
584 : func (i *lazyCombinedIter) SeekGE(
585 : key []byte, flags base.SeekGEFlags,
586 2 : ) (*InternalKey, base.LazyValue) {
587 2 : if i.combinedIterState.initialized {
588 0 : return i.parent.rangeKey.iiter.SeekGE(key, flags)
589 0 : }
590 2 : k, v := i.pointIter.SeekGE(key, flags)
591 2 : if i.combinedIterState.triggered {
592 2 : return i.initCombinedIteration(+1, k, v, key)
593 2 : }
594 2 : return k, v
595 : }
596 :
597 : func (i *lazyCombinedIter) SeekPrefixGE(
598 : prefix, key []byte, flags base.SeekGEFlags,
599 2 : ) (*InternalKey, base.LazyValue) {
600 2 : if i.combinedIterState.initialized {
601 0 : return i.parent.rangeKey.iiter.SeekPrefixGE(prefix, key, flags)
602 0 : }
603 2 : k, v := i.pointIter.SeekPrefixGE(prefix, key, flags)
604 2 : if i.combinedIterState.triggered {
605 2 : return i.initCombinedIteration(+1, k, v, key)
606 2 : }
607 2 : return k, v
608 : }
609 :
610 : func (i *lazyCombinedIter) SeekLT(
611 : key []byte, flags base.SeekLTFlags,
612 2 : ) (*InternalKey, base.LazyValue) {
613 2 : if i.combinedIterState.initialized {
614 0 : return i.parent.rangeKey.iiter.SeekLT(key, flags)
615 0 : }
616 2 : k, v := i.pointIter.SeekLT(key, flags)
617 2 : if i.combinedIterState.triggered {
618 2 : return i.initCombinedIteration(-1, k, v, key)
619 2 : }
620 2 : return k, v
621 : }
622 :
623 2 : func (i *lazyCombinedIter) First() (*InternalKey, base.LazyValue) {
624 2 : if i.combinedIterState.initialized {
625 0 : return i.parent.rangeKey.iiter.First()
626 0 : }
627 2 : k, v := i.pointIter.First()
628 2 : if i.combinedIterState.triggered {
629 2 : return i.initCombinedIteration(+1, k, v, nil)
630 2 : }
631 2 : return k, v
632 : }
633 :
634 2 : func (i *lazyCombinedIter) Last() (*InternalKey, base.LazyValue) {
635 2 : if i.combinedIterState.initialized {
636 0 : return i.parent.rangeKey.iiter.Last()
637 0 : }
638 2 : k, v := i.pointIter.Last()
639 2 : if i.combinedIterState.triggered {
640 2 : return i.initCombinedIteration(-1, k, v, nil)
641 2 : }
642 2 : return k, v
643 : }
644 :
645 2 : func (i *lazyCombinedIter) Next() (*InternalKey, base.LazyValue) {
646 2 : if i.combinedIterState.initialized {
647 0 : return i.parent.rangeKey.iiter.Next()
648 0 : }
649 2 : k, v := i.pointIter.Next()
650 2 : if i.combinedIterState.triggered {
651 2 : return i.initCombinedIteration(+1, k, v, nil)
652 2 : }
653 2 : return k, v
654 : }
655 :
656 2 : func (i *lazyCombinedIter) NextPrefix(succKey []byte) (*InternalKey, base.LazyValue) {
657 2 : if i.combinedIterState.initialized {
658 0 : return i.parent.rangeKey.iiter.NextPrefix(succKey)
659 0 : }
660 2 : k, v := i.pointIter.NextPrefix(succKey)
661 2 : if i.combinedIterState.triggered {
662 0 : return i.initCombinedIteration(+1, k, v, nil)
663 0 : }
664 2 : return k, v
665 : }
666 :
667 2 : func (i *lazyCombinedIter) Prev() (*InternalKey, base.LazyValue) {
668 2 : if i.combinedIterState.initialized {
669 0 : return i.parent.rangeKey.iiter.Prev()
670 0 : }
671 2 : k, v := i.pointIter.Prev()
672 2 : if i.combinedIterState.triggered {
673 2 : return i.initCombinedIteration(-1, k, v, nil)
674 2 : }
675 2 : return k, v
676 : }
677 :
678 2 : func (i *lazyCombinedIter) Error() error {
679 2 : if i.combinedIterState.initialized {
680 0 : return i.parent.rangeKey.iiter.Error()
681 0 : }
682 2 : return i.pointIter.Error()
683 : }
684 :
685 2 : func (i *lazyCombinedIter) Close() error {
686 2 : if i.combinedIterState.initialized {
687 0 : return i.parent.rangeKey.iiter.Close()
688 0 : }
689 2 : return i.pointIter.Close()
690 : }
691 :
692 2 : func (i *lazyCombinedIter) SetBounds(lower, upper []byte) {
693 2 : if i.combinedIterState.initialized {
694 0 : i.parent.rangeKey.iiter.SetBounds(lower, upper)
695 0 : return
696 0 : }
697 2 : i.pointIter.SetBounds(lower, upper)
698 : }
699 :
700 0 : func (i *lazyCombinedIter) SetContext(ctx context.Context) {
701 0 : if i.combinedIterState.initialized {
702 0 : i.parent.rangeKey.iiter.SetContext(ctx)
703 0 : return
704 0 : }
705 0 : i.pointIter.SetContext(ctx)
706 : }
707 :
708 0 : func (i *lazyCombinedIter) String() string {
709 0 : if i.combinedIterState.initialized {
710 0 : return i.parent.rangeKey.iiter.String()
711 0 : }
712 0 : return i.pointIter.String()
713 : }
|
