Line data Source code
1 : // Copyright 2018 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 : "bytes"
9 : "fmt"
10 : "runtime/debug"
11 : "unsafe"
12 :
13 : "github.com/cockroachdb/errors"
14 : "github.com/cockroachdb/pebble/internal/base"
15 : "github.com/cockroachdb/pebble/internal/invariants"
16 : "github.com/cockroachdb/pebble/internal/keyspan"
17 : )
18 :
19 : type mergingIterLevel struct {
20 : index int
21 : iter internalIterator
22 : // rangeDelIter is set to the range-deletion iterator for the level. When
23 : // configured with a levelIter, this pointer changes as sstable boundaries
24 : // are crossed. See levelIter.initRangeDel and the Range Deletions comment
25 : // below.
26 : rangeDelIter keyspan.FragmentIterator
27 : // iterKey and iterValue cache the current key and value iter are pointed at.
28 : iterKey *InternalKey
29 : iterValue base.LazyValue
30 :
31 : // levelIterBoundaryContext's fields are set when using levelIter, in order
32 : // to surface sstable boundary keys and file-level context. See levelIter
33 : // comment and the Range Deletions comment below.
34 : levelIterBoundaryContext
35 :
36 : // tombstone caches the tombstone rangeDelIter is currently pointed at. If
37 : // tombstone is nil, there are no further tombstones within the
38 : // current sstable in the current iterator direction. The cached tombstone is
39 : // only valid for the levels in the range [0,heap[0].index]. This avoids
40 : // positioning tombstones at lower levels which cannot possibly shadow the
41 : // current key.
42 : tombstone *keyspan.Span
43 : }
44 :
45 : type levelIterBoundaryContext struct {
46 : // smallestUserKey and largestUserKey are populated with the smallest and
47 : // largest boundaries of the current file.
48 : smallestUserKey, largestUserKey []byte
49 : // isLargestUserKeyExclusive is set to true when a file's largest boundary
50 : // is an exclusive key, (eg, a range deletion sentinel). If true, the file
51 : // does not contain any keys with the provided user key, and the
52 : // largestUserKey bound is exclusive.
53 : isLargestUserKeyExclusive bool
54 : // isSyntheticIterBoundsKey is set to true iff the key returned by the level
55 : // iterator is a synthetic key derived from the iterator bounds. This is used
56 : // to prevent the mergingIter from being stuck at such a synthetic key if it
57 : // becomes the top element of the heap. When used with a user-facing Iterator,
58 : // the only range deletions exposed by this mergingIter should be those with
59 : // `isSyntheticIterBoundsKey || isIgnorableBoundaryKey`.
60 : isSyntheticIterBoundsKey bool
61 : // isIgnorableBoundaryKey is set to true iff the key returned by the level
62 : // iterator is a file boundary key that should be ignored when returning to
63 : // the parent iterator. File boundary keys are used by the level iter to
64 : // keep a levelIter file's range deletion iterator open as long as other
65 : // levels within the merging iterator require it. When used with a user-facing
66 : // Iterator, the only range deletions exposed by this mergingIter should be
67 : // those with `isSyntheticIterBoundsKey || isIgnorableBoundaryKey`.
68 : isIgnorableBoundaryKey bool
69 : }
70 :
71 : // mergingIter provides a merged view of multiple iterators from different
72 : // levels of the LSM.
73 : //
74 : // The core of a mergingIter is a heap of internalIterators (see
75 : // mergingIterHeap). The heap can operate as either a min-heap, used during
76 : // forward iteration (First, SeekGE, Next) or a max-heap, used during reverse
77 : // iteration (Last, SeekLT, Prev). The heap is initialized in calls to First,
78 : // Last, SeekGE, and SeekLT. A call to Next or Prev takes the current top
79 : // element on the heap, advances its iterator, and then "fixes" the heap
80 : // property. When one of the child iterators is exhausted during Next/Prev
81 : // iteration, it is removed from the heap.
82 : //
83 : // # Range Deletions
84 : //
85 : // A mergingIter can optionally be configured with a slice of range deletion
86 : // iterators. The range deletion iterator slice must exactly parallel the point
87 : // iterators and the range deletion iterator must correspond to the same level
88 : // in the LSM as the point iterator. Note that each memtable and each table in
89 : // L0 is a different "level" from the mergingIter perspective. So level 0 below
90 : // does not correspond to L0 in the LSM.
91 : //
92 : // A range deletion iterator iterates over fragmented range tombstones. Range
93 : // tombstones are fragmented by splitting them at any overlapping points. This
94 : // fragmentation guarantees that within an sstable tombstones will either be
95 : // distinct or will have identical start and end user keys. While range
96 : // tombstones are fragmented within an sstable, the start and end keys are not truncated
97 : // to sstable boundaries. This is necessary because the tombstone end key is
98 : // exclusive and does not have a sequence number. Consider an sstable
99 : // containing the range tombstone [a,c)#9 and the key "b#8". The tombstone must
100 : // delete "b#8", yet older versions of "b" might spill over to the next
101 : // sstable. So the boundary key for this sstable must be "b#8". Adjusting the
102 : // end key of tombstones to be optionally inclusive or contain a sequence
103 : // number would be possible solutions (such solutions have potentially serious
104 : // issues: tombstones have exclusive end keys since an inclusive deletion end can
105 : // be converted to an exclusive one while the reverse transformation is not possible;
106 : // the semantics of a sequence number for the end key of a range tombstone are murky).
107 : //
108 : // The approach taken here performs an
109 : // implicit truncation of the tombstone to the sstable boundaries.
110 : //
111 : // During initialization of a mergingIter, the range deletion iterators for
112 : // batches, memtables, and L0 tables are populated up front. Note that Batches
113 : // and memtables index unfragmented tombstones. Batch.newRangeDelIter() and
114 : // memTable.newRangeDelIter() fragment and cache the tombstones on demand. The
115 : // L1-L6 range deletion iterators are populated by levelIter. When configured
116 : // to load range deletion iterators, whenever a levelIter loads a table it
117 : // loads both the point iterator and the range deletion
118 : // iterator. levelIter.rangeDelIter is configured to point to the right entry
119 : // in mergingIter.levels. The effect of this setup is that
120 : // mergingIter.levels[i].rangeDelIter always contains the fragmented range
121 : // tombstone for the current table in level i that the levelIter has open.
122 : //
123 : // Another crucial mechanism of levelIter is that it materializes fake point
124 : // entries for the table boundaries if the boundary is range deletion
125 : // key. Consider a table that contains only a range tombstone [a-e)#10. The
126 : // sstable boundaries for this table will be a#10,15 and
127 : // e#72057594037927935,15. During forward iteration levelIter will return
128 : // e#72057594037927935,15 as a key. During reverse iteration levelIter will
129 : // return a#10,15 as a key. These sentinel keys act as bookends to point
130 : // iteration and allow mergingIter to keep a table and its associated range
131 : // tombstones loaded as long as there are keys at lower levels that are within
132 : // the bounds of the table.
133 : //
134 : // The final piece to the range deletion puzzle is the LSM invariant that for a
135 : // given key K newer versions of K can only exist earlier in the level, or at
136 : // higher levels of the tree. For example, if K#4 exists in L3, k#5 can only
137 : // exist earlier in the L3 or in L0, L1, L2 or a memtable. Get very explicitly
138 : // uses this invariant to find the value for a key by walking the LSM level by
139 : // level. For range deletions, this invariant means that a range deletion at
140 : // level N will necessarily shadow any keys within its bounds in level Y where
141 : // Y > N. One wrinkle to this statement is that it only applies to keys that
142 : // lie within the sstable bounds as well, but we get that guarantee due to the
143 : // way the range deletion iterator and point iterator are bound together by a
144 : // levelIter.
145 : //
146 : // Tying the above all together, we get a picture where each level (index in
147 : // mergingIter.levels) is composed of both point operations (pX) and range
148 : // deletions (rX). The range deletions for level X shadow both the point
149 : // operations and range deletions for level Y where Y > X allowing mergingIter
150 : // to skip processing entries in that shadow. For example, consider the
151 : // scenario:
152 : //
153 : // r0: a---e
154 : // r1: d---h
155 : // r2: g---k
156 : // r3: j---n
157 : // r4: m---q
158 : //
159 : // This is showing 5 levels of range deletions. Consider what happens upon
160 : // SeekGE("b"). We first seek the point iterator for level 0 (the point values
161 : // are not shown above) and we then seek the range deletion iterator. That
162 : // returns the tombstone [a,e). This tombstone tells us that all keys in the
163 : // range [a,e) in lower levels are deleted so we can skip them. So we can
164 : // adjust the seek key to "e", the tombstone end key. For level 1 we seek to
165 : // "e" and find the range tombstone [d,h) and similar logic holds. By the time
166 : // we get to level 4 we're seeking to "n".
167 : //
168 : // One consequence of not truncating tombstone end keys to sstable boundaries
169 : // is the seeking process described above cannot always seek to the tombstone
170 : // end key in the older level. For example, imagine in the above example r3 is
171 : // a partitioned level (i.e., L1+ in our LSM), and the sstable containing [j,
172 : // n) has "k" as its upper boundary. In this situation, compactions involving
173 : // keys at or after "k" can output those keys to r4+, even if they're newer
174 : // than our tombstone [j, n). So instead of seeking to "n" in r4 we can only
175 : // seek to "k". To achieve this, the instance variable `largestUserKey.`
176 : // maintains the upper bounds of the current sstables in the partitioned
177 : // levels. In this example, `levels[3].largestUserKey` holds "k", telling us to
178 : // limit the seek triggered by a tombstone in r3 to "k".
179 : //
180 : // During actual iteration levels can contain both point operations and range
181 : // deletions. Within a level, when a range deletion contains a point operation
182 : // the sequence numbers must be checked to determine if the point operation is
183 : // newer or older than the range deletion tombstone. The mergingIter maintains
184 : // the invariant that the range deletion iterators for all levels newer that
185 : // the current iteration key (L < m.heap.items[0].index) are positioned at the
186 : // next (or previous during reverse iteration) range deletion tombstone. We
187 : // know those levels don't contain a range deletion tombstone that covers the
188 : // current key because if they did the current key would be deleted. The range
189 : // deletion iterator for the current key's level is positioned at a range
190 : // tombstone covering or past the current key. The position of all of other
191 : // range deletion iterators is unspecified. Whenever a key from those levels
192 : // becomes the current key, their range deletion iterators need to be
193 : // positioned. This lazy positioning avoids seeking the range deletion
194 : // iterators for keys that are never considered. (A similar bit of lazy
195 : // evaluation can be done for the point iterators, but is still TBD).
196 : //
197 : // For a full example, consider the following setup:
198 : //
199 : // p0: o
200 : // r0: m---q
201 : //
202 : // p1: n p
203 : // r1: g---k
204 : //
205 : // p2: b d i
206 : // r2: a---e q----v
207 : //
208 : // p3: e
209 : // r3:
210 : //
211 : // If we start iterating from the beginning, the first key we encounter is "b"
212 : // in p2. When the mergingIter is pointing at a valid entry, the range deletion
213 : // iterators for all of the levels < m.heap.items[0].index are positioned at
214 : // the next range tombstone past the current key. So r0 will point at [m,q) and
215 : // r1 at [g,k). When the key "b" is encountered, we check to see if the current
216 : // tombstone for r0 or r1 contains it, and whether the tombstone for r2, [a,e),
217 : // contains and is newer than "b".
218 : //
219 : // Advancing the iterator finds the next key at "d". This is in the same level
220 : // as the previous key "b" so we don't have to reposition any of the range
221 : // deletion iterators, but merely check whether "d" is now contained by any of
222 : // the range tombstones at higher levels or has stepped past the range
223 : // tombstone in its own level or higher levels. In this case, there is nothing to be done.
224 : //
225 : // Advancing the iterator again finds "e". Since "e" comes from p3, we have to
226 : // position the r3 range deletion iterator, which is empty. "e" is past the r2
227 : // tombstone of [a,e) so we need to advance the r2 range deletion iterator to
228 : // [q,v).
229 : //
230 : // The next key is "i". Because this key is in p2, a level above "e", we don't
231 : // have to reposition any range deletion iterators and instead see that "i" is
232 : // covered by the range tombstone [g,k). The iterator is immediately advanced
233 : // to "n" which is covered by the range tombstone [m,q) causing the iterator to
234 : // advance to "o" which is visible.
235 : //
236 : // TODO(peter,rangedel): For testing, advance the iterator through various
237 : // scenarios and have each step display the current state (i.e. the current
238 : // heap and range-del iterator positioning).
239 : type mergingIter struct {
240 : logger Logger
241 : split Split
242 : dir int
243 : snapshot uint64
244 : batchSnapshot uint64
245 : levels []mergingIterLevel
246 : heap mergingIterHeap
247 : err error
248 : prefix []byte
249 : lower []byte
250 : upper []byte
251 : stats *InternalIteratorStats
252 :
253 : // levelsPositioned, if non-nil, is a slice of the same length as levels.
254 : // It's used by NextPrefix to record which levels have already been
255 : // repositioned. It's created lazily by the first call to NextPrefix.
256 : levelsPositioned []bool
257 :
258 : combinedIterState *combinedIterState
259 :
260 : // Used in some tests to disable the random disabling of seek optimizations.
261 : forceEnableSeekOpt bool
262 : }
263 :
264 : // mergingIter implements the base.InternalIterator interface.
265 : var _ base.InternalIterator = (*mergingIter)(nil)
266 :
267 : // newMergingIter returns an iterator that merges its input. Walking the
268 : // resultant iterator will return all key/value pairs of all input iterators
269 : // in strictly increasing key order, as defined by cmp. It is permissible to
270 : // pass a nil split parameter if the caller is never going to call
271 : // SeekPrefixGE.
272 : //
273 : // The input's key ranges may overlap, but there are assumed to be no duplicate
274 : // keys: if iters[i] contains a key k then iters[j] will not contain that key k.
275 : //
276 : // None of the iters may be nil.
277 : func newMergingIter(
278 : logger Logger,
279 : stats *base.InternalIteratorStats,
280 : cmp Compare,
281 : split Split,
282 : iters ...internalIterator,
283 1 : ) *mergingIter {
284 1 : m := &mergingIter{}
285 1 : levels := make([]mergingIterLevel, len(iters))
286 1 : for i := range levels {
287 1 : levels[i].iter = iters[i]
288 1 : }
289 1 : m.init(&IterOptions{logger: logger}, stats, cmp, split, levels...)
290 1 : return m
291 : }
292 :
293 : func (m *mergingIter) init(
294 : opts *IterOptions,
295 : stats *base.InternalIteratorStats,
296 : cmp Compare,
297 : split Split,
298 : levels ...mergingIterLevel,
299 1 : ) {
300 1 : m.err = nil // clear cached iteration error
301 1 : m.logger = opts.getLogger()
302 1 : if opts != nil {
303 1 : m.lower = opts.LowerBound
304 1 : m.upper = opts.UpperBound
305 1 : }
306 1 : m.snapshot = InternalKeySeqNumMax
307 1 : m.batchSnapshot = InternalKeySeqNumMax
308 1 : m.levels = levels
309 1 : m.heap.cmp = cmp
310 1 : m.split = split
311 1 : m.stats = stats
312 1 : if cap(m.heap.items) < len(levels) {
313 1 : m.heap.items = make([]*mergingIterLevel, 0, len(levels))
314 1 : } else {
315 1 : m.heap.items = m.heap.items[:0]
316 1 : }
317 1 : for l := range m.levels {
318 1 : m.levels[l].index = l
319 1 : }
320 : }
321 :
322 1 : func (m *mergingIter) initHeap() {
323 1 : m.heap.items = m.heap.items[:0]
324 1 : for i := range m.levels {
325 1 : if l := &m.levels[i]; l.iterKey != nil {
326 1 : m.heap.items = append(m.heap.items, l)
327 1 : } else {
328 1 : m.err = firstError(m.err, l.iter.Error())
329 1 : if m.err != nil {
330 0 : return
331 0 : }
332 : }
333 : }
334 1 : m.heap.init()
335 : }
336 :
337 1 : func (m *mergingIter) initMinHeap() {
338 1 : m.dir = 1
339 1 : m.heap.reverse = false
340 1 : m.initHeap()
341 1 : m.initMinRangeDelIters(-1)
342 1 : }
343 :
344 : // The level of the previous top element was oldTopLevel. Note that all range delete
345 : // iterators < oldTopLevel are positioned past the key of the previous top element and
346 : // the range delete iterator == oldTopLevel is positioned at or past the key of the
347 : // previous top element. We need to position the range delete iterators from oldTopLevel + 1
348 : // to the level of the current top element.
349 1 : func (m *mergingIter) initMinRangeDelIters(oldTopLevel int) {
350 1 : if m.heap.len() == 0 {
351 1 : return
352 1 : }
353 :
354 : // Position the range-del iterators at levels <= m.heap.items[0].index.
355 1 : item := m.heap.items[0]
356 1 : for level := oldTopLevel + 1; level <= item.index; level++ {
357 1 : l := &m.levels[level]
358 1 : if l.rangeDelIter == nil {
359 1 : continue
360 : }
361 1 : l.tombstone = l.rangeDelIter.SeekGE(item.iterKey.UserKey)
362 : }
363 : }
364 :
365 1 : func (m *mergingIter) initMaxHeap() {
366 1 : m.dir = -1
367 1 : m.heap.reverse = true
368 1 : m.initHeap()
369 1 : m.initMaxRangeDelIters(-1)
370 1 : }
371 :
372 : // The level of the previous top element was oldTopLevel. Note that all range delete
373 : // iterators < oldTopLevel are positioned before the key of the previous top element and
374 : // the range delete iterator == oldTopLevel is positioned at or before the key of the
375 : // previous top element. We need to position the range delete iterators from oldTopLevel + 1
376 : // to the level of the current top element.
377 1 : func (m *mergingIter) initMaxRangeDelIters(oldTopLevel int) {
378 1 : if m.heap.len() == 0 {
379 1 : return
380 1 : }
381 : // Position the range-del iterators at levels <= m.heap.items[0].index.
382 1 : item := m.heap.items[0]
383 1 : for level := oldTopLevel + 1; level <= item.index; level++ {
384 1 : l := &m.levels[level]
385 1 : if l.rangeDelIter == nil {
386 1 : continue
387 : }
388 1 : l.tombstone = keyspan.SeekLE(m.heap.cmp, l.rangeDelIter, item.iterKey.UserKey)
389 : }
390 : }
391 :
392 1 : func (m *mergingIter) switchToMinHeap() {
393 1 : if m.heap.len() == 0 {
394 1 : if m.lower != nil {
395 1 : m.SeekGE(m.lower, base.SeekGEFlagsNone)
396 1 : } else {
397 1 : m.First()
398 1 : }
399 1 : return
400 : }
401 :
402 : // We're switching from using a max heap to a min heap. We need to advance
403 : // any iterator that is less than or equal to the current key. Consider the
404 : // scenario where we have 2 iterators being merged (user-key:seq-num):
405 : //
406 : // i1: *a:2 b:2
407 : // i2: a:1 b:1
408 : //
409 : // The current key is a:2 and i2 is pointed at a:1. When we switch to forward
410 : // iteration, we want to return a key that is greater than a:2.
411 :
412 1 : key := m.heap.items[0].iterKey
413 1 : cur := m.heap.items[0]
414 1 :
415 1 : for i := range m.levels {
416 1 : l := &m.levels[i]
417 1 : if l == cur {
418 1 : continue
419 : }
420 :
421 : // If the iterator is exhausted, it may be out of bounds if range
422 : // deletions modified our search key as we descended. we need to
423 : // reposition it within the search bounds. If the current key is a
424 : // range tombstone, the iterator might still be exhausted but at a
425 : // sstable boundary sentinel. It would be okay to reposition an
426 : // interator like this only through successive Next calls, except that
427 : // it would violate the levelIter's invariants by causing it to return
428 : // a key before the lower bound.
429 : //
430 : // bounds = [ f, _ )
431 : // L0: [ b ] [ f* z ]
432 : // L1: [ a |----| k y ]
433 : // L2: [ c (d) ] [ e g m ]
434 : // L3: [ x ]
435 : //
436 : // * - current key [] - table bounds () - heap item
437 : //
438 : // In the above diagram, the L2 iterator is positioned at a sstable
439 : // boundary (d) outside the lower bound (f). It arrived here from a
440 : // seek whose seek-key was modified by a range tombstone. If we called
441 : // Next on the L2 iterator, it would return e, violating its lower
442 : // bound. Instead, we seek it to >= f and Next from there.
443 :
444 1 : if l.iterKey == nil || (m.lower != nil && l.isSyntheticIterBoundsKey &&
445 1 : l.iterKey.IsExclusiveSentinel() &&
446 1 : m.heap.cmp(l.iterKey.UserKey, m.lower) <= 0) {
447 1 : if m.lower != nil {
448 1 : l.iterKey, l.iterValue = l.iter.SeekGE(m.lower, base.SeekGEFlagsNone)
449 1 : } else {
450 1 : l.iterKey, l.iterValue = l.iter.First()
451 1 : }
452 : }
453 1 : for ; l.iterKey != nil; l.iterKey, l.iterValue = l.iter.Next() {
454 1 : if base.InternalCompare(m.heap.cmp, *key, *l.iterKey) < 0 {
455 1 : // key < iter-key
456 1 : break
457 : }
458 : // key >= iter-key
459 : }
460 : }
461 :
462 : // Special handling for the current iterator because we were using its key
463 : // above. The iterator cur.iter may still be exhausted at a sstable boundary
464 : // sentinel. Similar to the logic applied to the other levels, in these
465 : // cases we seek the iterator to the first key in order to avoid violating
466 : // levelIter's invariants. See the example in the for loop above.
467 1 : if m.lower != nil && cur.isSyntheticIterBoundsKey && cur.iterKey.IsExclusiveSentinel() &&
468 1 : m.heap.cmp(cur.iterKey.UserKey, m.lower) <= 0 {
469 1 : cur.iterKey, cur.iterValue = cur.iter.SeekGE(m.lower, base.SeekGEFlagsNone)
470 1 : } else {
471 1 : cur.iterKey, cur.iterValue = cur.iter.Next()
472 1 : }
473 1 : m.initMinHeap()
474 : }
475 :
476 1 : func (m *mergingIter) switchToMaxHeap() {
477 1 : if m.heap.len() == 0 {
478 1 : if m.upper != nil {
479 1 : m.SeekLT(m.upper, base.SeekLTFlagsNone)
480 1 : } else {
481 1 : m.Last()
482 1 : }
483 1 : return
484 : }
485 :
486 : // We're switching from using a min heap to a max heap. We need to backup any
487 : // iterator that is greater than or equal to the current key. Consider the
488 : // scenario where we have 2 iterators being merged (user-key:seq-num):
489 : //
490 : // i1: a:2 *b:2
491 : // i2: a:1 b:1
492 : //
493 : // The current key is b:2 and i2 is pointing at b:1. When we switch to
494 : // reverse iteration, we want to return a key that is less than b:2.
495 1 : key := m.heap.items[0].iterKey
496 1 : cur := m.heap.items[0]
497 1 :
498 1 : for i := range m.levels {
499 1 : l := &m.levels[i]
500 1 : if l == cur {
501 1 : continue
502 : }
503 :
504 : // If the iterator is exhausted, it may be out of bounds if range
505 : // deletions modified our search key as we descended. we need to
506 : // reposition it within the search bounds. If the current key is a
507 : // range tombstone, the iterator might still be exhausted but at a
508 : // sstable boundary sentinel. It would be okay to reposition an
509 : // interator like this only through successive Prev calls, except that
510 : // it would violate the levelIter's invariants by causing it to return
511 : // a key beyond the upper bound.
512 : //
513 : // bounds = [ _, g )
514 : // L0: [ b ] [ f* z ]
515 : // L1: [ a |-------| k y ]
516 : // L2: [ c d ] h [(i) m ]
517 : // L3: [ e x ]
518 : //
519 : // * - current key [] - table bounds () - heap item
520 : //
521 : // In the above diagram, the L2 iterator is positioned at a sstable
522 : // boundary (i) outside the upper bound (g). It arrived here from a
523 : // seek whose seek-key was modified by a range tombstone. If we called
524 : // Prev on the L2 iterator, it would return h, violating its upper
525 : // bound. Instead, we seek it to < g, and Prev from there.
526 :
527 1 : if l.iterKey == nil || (m.upper != nil && l.isSyntheticIterBoundsKey &&
528 1 : l.iterKey.IsExclusiveSentinel() && m.heap.cmp(l.iterKey.UserKey, m.upper) >= 0) {
529 1 : if m.upper != nil {
530 1 : l.iterKey, l.iterValue = l.iter.SeekLT(m.upper, base.SeekLTFlagsNone)
531 1 : } else {
532 1 : l.iterKey, l.iterValue = l.iter.Last()
533 1 : }
534 : }
535 1 : for ; l.iterKey != nil; l.iterKey, l.iterValue = l.iter.Prev() {
536 1 : if base.InternalCompare(m.heap.cmp, *key, *l.iterKey) > 0 {
537 1 : // key > iter-key
538 1 : break
539 : }
540 : // key <= iter-key
541 : }
542 : }
543 :
544 : // Special handling for the current iterator because we were using its key
545 : // above. The iterator cur.iter may still be exhausted at a sstable boundary
546 : // sentinel. Similar to the logic applied to the other levels, in these
547 : // cases we seek the iterator to in order to avoid violating levelIter's
548 : // invariants by Prev-ing through files. See the example in the for loop
549 : // above.
550 1 : if m.upper != nil && cur.isSyntheticIterBoundsKey && cur.iterKey.IsExclusiveSentinel() &&
551 1 : m.heap.cmp(cur.iterKey.UserKey, m.upper) >= 0 {
552 1 : cur.iterKey, cur.iterValue = cur.iter.SeekLT(m.upper, base.SeekLTFlagsNone)
553 1 : } else {
554 1 : cur.iterKey, cur.iterValue = cur.iter.Prev()
555 1 : }
556 1 : m.initMaxHeap()
557 : }
558 :
559 : // maybeNextEntryWithinPrefix steps to the next entry, as long as the iteration
560 : // prefix has not already been exceeded. If it has, it exhausts the iterator by
561 : // resetting the heap to empty.
562 1 : func (m *mergingIter) maybeNextEntryWithinPrefix(l *mergingIterLevel) {
563 1 : if s := m.split(l.iterKey.UserKey); !bytes.Equal(m.prefix, l.iterKey.UserKey[:s]) {
564 1 : // The item at the root of the heap already exceeds the iteration
565 1 : // prefix. We should not advance any more. Clear the heap to reflect
566 1 : // that the iterator is now exhausted (within this prefix, at
567 1 : // least).
568 1 : m.heap.items = m.heap.items[:0]
569 1 : return
570 1 : }
571 1 : m.nextEntry(l, nil /* succKey */)
572 : }
573 :
574 : // nextEntry unconditionally steps to the next entry. item is the current top
575 : // item in the heap.
576 : //
577 : // nextEntry should be called directly when not in prefix-iteration mode, or by
578 : // Next. During prefix iteration mode, all other callers should use
579 : // maybeNextEntryWithinPrefix which will avoid advancing the iterator if the
580 : // current iteration prefix has been exhausted. See the comment within
581 : // nextEntry's body for an explanation of why other callers should call
582 : // maybeNextEntryWithinPrefix, which will ensure the documented invariant is
583 : // preserved.
584 1 : func (m *mergingIter) nextEntry(l *mergingIterLevel, succKey []byte) {
585 1 : // INVARIANT: If in prefix iteration mode, item.iterKey must have a prefix equal
586 1 : // to m.prefix. This invariant is important for ensuring TrySeekUsingNext
587 1 : // optimizations behave correctly.
588 1 : //
589 1 : // During prefix iteration, the iterator does not have a full view of the
590 1 : // LSM. Some level iterators may omit keys that are known to fall outside
591 1 : // the seek prefix (eg, due to sstable bloom filter exclusion). It's
592 1 : // important that in such cases we don't position any iterators beyond
593 1 : // m.prefix, because doing so may interfere with future seeks.
594 1 : //
595 1 : // Let prefixes P1 < P2 < P3. Imagine a SeekPrefixGE to prefix P1, followed
596 1 : // by a SeekPrefixGE to prefix P2. Imagine there exist live keys at prefix
597 1 : // P2, but they're not visible to the SeekPrefixGE(P1) (because of
598 1 : // bloom-filter exclusion or a range tombstone that deletes prefix P1 but
599 1 : // not P2). If the SeekPrefixGE(P1) is allowed to move any level iterators
600 1 : // to P3, the SeekPrefixGE(P2, TrySeekUsingNext=true) may mistakenly think
601 1 : // the level contains no point keys or range tombstones within the prefix
602 1 : // P2. Care is taken to avoid ever advancing the iterator beyond the current
603 1 : // prefix. If nextEntry is ever invoked while we're already beyond the
604 1 : // current prefix, we're violating the invariant.
605 1 : if invariants.Enabled && m.prefix != nil {
606 0 : if s := m.split(l.iterKey.UserKey); !bytes.Equal(m.prefix, l.iterKey.UserKey[:s]) {
607 0 : m.logger.Fatalf("mergingIter: prefix violation: nexting beyond prefix %q; existing heap root %q\n%s",
608 0 : m.prefix, l.iterKey, debug.Stack())
609 0 : }
610 : }
611 :
612 1 : oldTopLevel := l.index
613 1 : oldRangeDelIter := l.rangeDelIter
614 1 :
615 1 : if succKey == nil {
616 1 : l.iterKey, l.iterValue = l.iter.Next()
617 1 : } else {
618 1 : l.iterKey, l.iterValue = l.iter.NextPrefix(succKey)
619 1 : }
620 :
621 1 : if l.iterKey != nil {
622 1 : if m.heap.len() > 1 {
623 1 : m.heap.fix(0)
624 1 : }
625 1 : if l.rangeDelIter != oldRangeDelIter {
626 1 : // The rangeDelIter changed which indicates that the l.iter moved to the
627 1 : // next sstable. We have to update the tombstone for oldTopLevel as well.
628 1 : oldTopLevel--
629 1 : }
630 1 : } else {
631 1 : m.err = l.iter.Error()
632 1 : if m.err == nil {
633 1 : m.heap.pop()
634 1 : }
635 : }
636 :
637 : // The cached tombstones are only valid for the levels
638 : // [0,oldTopLevel]. Updated the cached tombstones for any levels in the range
639 : // [oldTopLevel+1,heap[0].index].
640 1 : m.initMinRangeDelIters(oldTopLevel)
641 : }
642 :
643 : // isNextEntryDeleted starts from the current entry (as the next entry) and if
644 : // it is deleted, moves the iterators forward as needed and returns true, else
645 : // it returns false. item is the top item in the heap.
646 : //
647 : // During prefix iteration mode, isNextEntryDeleted will exhaust the iterator by
648 : // clearing the heap if the deleted key(s) extend beyond the iteration prefix
649 : // during prefix-iteration mode.
650 1 : func (m *mergingIter) isNextEntryDeleted(item *mergingIterLevel) bool {
651 1 : // Look for a range deletion tombstone containing item.iterKey at higher
652 1 : // levels (level < item.index). If we find such a range tombstone we know
653 1 : // it deletes the key in the current level. Also look for a range
654 1 : // deletion at the current level (level == item.index). If we find such a
655 1 : // range deletion we need to check whether it is newer than the current
656 1 : // entry.
657 1 : for level := 0; level <= item.index; level++ {
658 1 : l := &m.levels[level]
659 1 : if l.rangeDelIter == nil || l.tombstone == nil {
660 1 : // If l.tombstone is nil, there are no further tombstones
661 1 : // in the current sstable in the current (forward) iteration
662 1 : // direction.
663 1 : continue
664 : }
665 1 : if m.heap.cmp(l.tombstone.End, item.iterKey.UserKey) <= 0 {
666 1 : // The current key is at or past the tombstone end key.
667 1 : //
668 1 : // NB: for the case that this l.rangeDelIter is provided by a levelIter we know that
669 1 : // the levelIter must be positioned at a key >= item.iterKey. So it is sufficient to seek the
670 1 : // current l.rangeDelIter (since any range del iterators that will be provided by the
671 1 : // levelIter in the future cannot contain item.iterKey). Also, it is possible that we
672 1 : // will encounter parts of the range delete that should be ignored -- we handle that
673 1 : // below.
674 1 : l.tombstone = l.rangeDelIter.SeekGE(item.iterKey.UserKey)
675 1 : }
676 1 : if l.tombstone == nil {
677 1 : continue
678 : }
679 :
680 : // Reasoning for correctness of untruncated tombstone handling when the untruncated
681 : // tombstone is at a higher level:
682 : // The iterator corresponding to this tombstone is still in the heap so it must be
683 : // positioned >= item.iterKey. Which means the Largest key bound of the sstable containing this
684 : // tombstone is >= item.iterKey. So the upper limit of this tombstone cannot be file-bounds-constrained
685 : // to < item.iterKey. But it is possible that item.key < smallestUserKey, in which
686 : // case this tombstone should be ignored.
687 : //
688 : // Example 1:
689 : // sstable bounds [c#8, g#12] containing a tombstone [b, i)#7, and key is c#6. The
690 : // smallestUserKey is c, so we know the key is within the file bounds and the tombstone
691 : // [b, i) covers it.
692 : //
693 : // Example 2:
694 : // Same sstable bounds but key is b#10. The smallestUserKey is c, so the tombstone [b, i)
695 : // does not cover this key.
696 : //
697 : // For a tombstone at the same level as the key, the file bounds are trivially satisfied.
698 1 : if (l.smallestUserKey == nil || m.heap.cmp(l.smallestUserKey, item.iterKey.UserKey) <= 0) &&
699 1 : l.tombstone.VisibleAt(m.snapshot) && l.tombstone.Contains(m.heap.cmp, item.iterKey.UserKey) {
700 1 : if level < item.index {
701 1 : // We could also do m.seekGE(..., level + 1). The levels from
702 1 : // [level + 1, item.index) are already after item.iterKey so seeking them may be
703 1 : // wasteful.
704 1 :
705 1 : // We can seek up to the min of largestUserKey and tombstone.End.
706 1 : //
707 1 : // Using example 1 above, we can seek to the smaller of g and i, which is g.
708 1 : //
709 1 : // Another example, where the sstable bounds are [c#8, i#InternalRangeDelSentinel],
710 1 : // and the tombstone is [b, i)#8. Seeking to i is correct since it is seeking up to
711 1 : // the exclusive bound of the tombstone. We do not need to look at
712 1 : // isLargestKeyRangeDelSentinel.
713 1 : //
714 1 : // Progress argument: Since this file is at a higher level than item.iterKey we know
715 1 : // that the iterator in this file must be positioned within its bounds and at a key
716 1 : // X > item.iterKey (otherwise it would be the min of the heap). It is not
717 1 : // possible for X.UserKey == item.iterKey.UserKey, since it is incompatible with
718 1 : // X > item.iterKey (a lower version cannot be in a higher sstable), so it must be that
719 1 : // X.UserKey > item.iterKey.UserKey. Which means l.largestUserKey > item.key.UserKey.
720 1 : // We also know that l.tombstone.End > item.iterKey.UserKey. So the min of these,
721 1 : // seekKey, computed below, is > item.iterKey.UserKey, so the call to seekGE() will
722 1 : // make forward progress.
723 1 : seekKey := l.tombstone.End
724 1 : if l.largestUserKey != nil && m.heap.cmp(l.largestUserKey, seekKey) < 0 {
725 0 : seekKey = l.largestUserKey
726 0 : }
727 : // This seek is not directly due to a SeekGE call, so we don't know
728 : // enough about the underlying iterator positions, and so we keep the
729 : // try-seek-using-next optimization disabled. Additionally, if we're in
730 : // prefix-seek mode and a re-seek would have moved us past the original
731 : // prefix, we can remove all merging iter levels below the rangedel
732 : // tombstone's level and return immediately instead of re-seeking. This
733 : // is correct since those levels cannot provide a key that matches the
734 : // prefix, and is also visible. Additionally, this is important to make
735 : // subsequent `TrySeekUsingNext` work correctly, as a re-seek on a
736 : // different prefix could have resulted in this iterator skipping visible
737 : // keys at prefixes in between m.prefix and seekKey, that are currently
738 : // not in the heap due to a bloom filter mismatch.
739 : //
740 : // Additionally, we set the relative-seek flag. This is
741 : // important when iterating with lazy combined iteration. If
742 : // there's a range key between this level's current file and the
743 : // file the seek will land on, we need to detect it in order to
744 : // trigger construction of the combined iterator.
745 1 : if m.prefix != nil {
746 1 : if n := m.split(seekKey); !bytes.Equal(m.prefix, seekKey[:n]) {
747 1 : for i := item.index; i < len(m.levels); i++ {
748 1 : // Remove this level from the heap. Setting iterKey and iterValue
749 1 : // to their zero values should be sufficient for initMinHeap to not
750 1 : // re-initialize the heap with them in it. Other fields in
751 1 : // mergingIterLevel can remain as-is; the iter/rangeDelIter needs
752 1 : // to stay intact for future trySeekUsingNexts to work, the level
753 1 : // iter boundary context is owned by the levelIter which is not
754 1 : // being repositioned, and any tombstones in these levels will be
755 1 : // irrelevant for us anyway.
756 1 : m.levels[i].iterKey = nil
757 1 : m.levels[i].iterValue = base.LazyValue{}
758 1 : }
759 : // TODO(bilal): Consider a more efficient way of removing levels from
760 : // the heap without reinitializing all of it. This would likely
761 : // necessitate tracking the heap positions of each mergingIterHeap
762 : // item in the mergingIterLevel, and then swapping that item in the
763 : // heap with the last-positioned heap item, and shrinking the heap by
764 : // one.
765 1 : m.initMinHeap()
766 1 : return true
767 : }
768 : }
769 1 : m.seekGE(seekKey, item.index, base.SeekGEFlagsNone.EnableRelativeSeek())
770 1 : return true
771 : }
772 1 : if l.tombstone.CoversAt(m.snapshot, item.iterKey.SeqNum()) {
773 1 : if m.prefix == nil {
774 1 : m.nextEntry(item, nil /* succKey */)
775 1 : } else {
776 1 : m.maybeNextEntryWithinPrefix(item)
777 1 : }
778 1 : return true
779 : }
780 : }
781 : }
782 1 : return false
783 : }
784 :
785 : // Starting from the current entry, finds the first (next) entry that can be returned.
786 1 : func (m *mergingIter) findNextEntry() (*InternalKey, base.LazyValue) {
787 1 : for m.heap.len() > 0 && m.err == nil {
788 1 : item := m.heap.items[0]
789 1 : if m.levels[item.index].isSyntheticIterBoundsKey {
790 1 : break
791 : }
792 :
793 1 : m.addItemStats(item)
794 1 :
795 1 : // Skip ignorable boundary keys. These are not real keys and exist to
796 1 : // keep sstables open until we've surpassed their end boundaries so that
797 1 : // their range deletions are visible.
798 1 : if m.levels[item.index].isIgnorableBoundaryKey {
799 1 : if m.prefix == nil {
800 1 : m.nextEntry(item, nil /* succKey */)
801 1 : } else {
802 1 : m.maybeNextEntryWithinPrefix(item)
803 1 : }
804 1 : continue
805 : }
806 :
807 : // Check if the heap root key is deleted by a range tombstone in a
808 : // higher level. If it is, isNextEntryDeleted will advance the iterator
809 : // to a later key (through seeking or nexting).
810 1 : if m.isNextEntryDeleted(item) {
811 1 : m.stats.PointsCoveredByRangeTombstones++
812 1 : continue
813 : }
814 :
815 : // Check if the key is visible at the iterator sequence numbers.
816 1 : if !item.iterKey.Visible(m.snapshot, m.batchSnapshot) {
817 1 : if m.prefix == nil {
818 1 : m.nextEntry(item, nil /* succKey */)
819 1 : } else {
820 1 : m.maybeNextEntryWithinPrefix(item)
821 1 : }
822 1 : continue
823 : }
824 :
825 : // The heap root is visible and not deleted by any range tombstones.
826 : // Return it.
827 1 : return item.iterKey, item.iterValue
828 : }
829 1 : return nil, base.LazyValue{}
830 : }
831 :
832 : // Steps to the prev entry. item is the current top item in the heap.
833 1 : func (m *mergingIter) prevEntry(l *mergingIterLevel) {
834 1 : oldTopLevel := l.index
835 1 : oldRangeDelIter := l.rangeDelIter
836 1 : if l.iterKey, l.iterValue = l.iter.Prev(); l.iterKey != nil {
837 1 : if m.heap.len() > 1 {
838 1 : m.heap.fix(0)
839 1 : }
840 1 : if l.rangeDelIter != oldRangeDelIter && l.rangeDelIter != nil {
841 1 : // The rangeDelIter changed which indicates that the l.iter moved to the
842 1 : // previous sstable. We have to update the tombstone for oldTopLevel as
843 1 : // well.
844 1 : oldTopLevel--
845 1 : }
846 1 : } else {
847 1 : m.err = l.iter.Error()
848 1 : if m.err == nil {
849 1 : m.heap.pop()
850 1 : }
851 : }
852 :
853 : // The cached tombstones are only valid for the levels
854 : // [0,oldTopLevel]. Updated the cached tombstones for any levels in the range
855 : // [oldTopLevel+1,heap[0].index].
856 1 : m.initMaxRangeDelIters(oldTopLevel)
857 : }
858 :
859 : // isPrevEntryDeleted() starts from the current entry (as the prev entry) and if it is deleted,
860 : // moves the iterators backward as needed and returns true, else it returns false. item is the top
861 : // item in the heap.
862 1 : func (m *mergingIter) isPrevEntryDeleted(item *mergingIterLevel) bool {
863 1 : // Look for a range deletion tombstone containing item.iterKey at higher
864 1 : // levels (level < item.index). If we find such a range tombstone we know
865 1 : // it deletes the key in the current level. Also look for a range
866 1 : // deletion at the current level (level == item.index). If we find such a
867 1 : // range deletion we need to check whether it is newer than the current
868 1 : // entry.
869 1 : for level := 0; level <= item.index; level++ {
870 1 : l := &m.levels[level]
871 1 : if l.rangeDelIter == nil || l.tombstone == nil {
872 1 : // If l.tombstone is nil, there are no further tombstones
873 1 : // in the current sstable in the current (reverse) iteration
874 1 : // direction.
875 1 : continue
876 : }
877 1 : if m.heap.cmp(item.iterKey.UserKey, l.tombstone.Start) < 0 {
878 1 : // The current key is before the tombstone start key.
879 1 : //
880 1 : // NB: for the case that this l.rangeDelIter is provided by a levelIter we know that
881 1 : // the levelIter must be positioned at a key < item.iterKey. So it is sufficient to seek the
882 1 : // current l.rangeDelIter (since any range del iterators that will be provided by the
883 1 : // levelIter in the future cannot contain item.iterKey). Also, it is it is possible that we
884 1 : // will encounter parts of the range delete that should be ignored -- we handle that
885 1 : // below.
886 1 : l.tombstone = keyspan.SeekLE(m.heap.cmp, l.rangeDelIter, item.iterKey.UserKey)
887 1 : }
888 1 : if l.tombstone == nil {
889 1 : continue
890 : }
891 :
892 : // Reasoning for correctness of untruncated tombstone handling when the untruncated
893 : // tombstone is at a higher level:
894 : //
895 : // The iterator corresponding to this tombstone is still in the heap so it must be
896 : // positioned <= item.iterKey. Which means the Smallest key bound of the sstable containing this
897 : // tombstone is <= item.iterKey. So the lower limit of this tombstone cannot have been
898 : // file-bounds-constrained to > item.iterKey. But it is possible that item.key >= Largest
899 : // key bound of this sstable, in which case this tombstone should be ignored.
900 : //
901 : // Example 1:
902 : // sstable bounds [c#8, g#12] containing a tombstone [b, i)#7, and key is f#6. The
903 : // largestUserKey is g, so we know the key is within the file bounds and the tombstone
904 : // [b, i) covers it.
905 : //
906 : // Example 2:
907 : // Same sstable but the key is g#6. This cannot happen since the [b, i)#7 untruncated
908 : // tombstone was involved in a compaction which must have had a file to the right of this
909 : // sstable that is part of the same atomic compaction group for future compactions. That
910 : // file must have bounds that cover g#6 and this levelIter must be at that file.
911 : //
912 : // Example 3:
913 : // sstable bounds [c#8, g#RangeDelSentinel] containing [b, i)#7 and the key is g#10.
914 : // This key is not deleted by this tombstone. We need to look at
915 : // isLargestUserKeyExclusive.
916 : //
917 : // For a tombstone at the same level as the key, the file bounds are trivially satisfied.
918 :
919 : // Default to within bounds.
920 1 : withinLargestSSTableBound := true
921 1 : if l.largestUserKey != nil {
922 1 : cmpResult := m.heap.cmp(l.largestUserKey, item.iterKey.UserKey)
923 1 : withinLargestSSTableBound = cmpResult > 0 || (cmpResult == 0 && !l.isLargestUserKeyExclusive)
924 1 : }
925 1 : if withinLargestSSTableBound && l.tombstone.Contains(m.heap.cmp, item.iterKey.UserKey) && l.tombstone.VisibleAt(m.snapshot) {
926 1 : if level < item.index {
927 1 : // We could also do m.seekLT(..., level + 1). The levels from
928 1 : // [level + 1, item.index) are already before item.iterKey so seeking them may be
929 1 : // wasteful.
930 1 :
931 1 : // We can seek up to the max of smallestUserKey and tombstone.Start.UserKey.
932 1 : //
933 1 : // Using example 1 above, we can seek to the larger of c and b, which is c.
934 1 : //
935 1 : // Progress argument: We know that the iterator in this file is positioned within
936 1 : // its bounds and at a key X < item.iterKey (otherwise it would be the max of the heap).
937 1 : // So smallestUserKey <= item.iterKey.UserKey and we already know that
938 1 : // l.tombstone.Start.UserKey <= item.iterKey.UserKey. So the seekKey computed below
939 1 : // is <= item.iterKey.UserKey, and since we do a seekLT() we will make backwards
940 1 : // progress.
941 1 : seekKey := l.tombstone.Start
942 1 : if l.smallestUserKey != nil && m.heap.cmp(l.smallestUserKey, seekKey) > 0 {
943 0 : seekKey = l.smallestUserKey
944 0 : }
945 : // We set the relative-seek flag. This is important when
946 : // iterating with lazy combined iteration. If there's a range
947 : // key between this level's current file and the file the seek
948 : // will land on, we need to detect it in order to trigger
949 : // construction of the combined iterator.
950 1 : m.seekLT(seekKey, item.index, base.SeekLTFlagsNone.EnableRelativeSeek())
951 1 : return true
952 : }
953 1 : if l.tombstone.CoversAt(m.snapshot, item.iterKey.SeqNum()) {
954 1 : m.prevEntry(item)
955 1 : return true
956 1 : }
957 : }
958 : }
959 1 : return false
960 : }
961 :
962 : // Starting from the current entry, finds the first (prev) entry that can be returned.
963 1 : func (m *mergingIter) findPrevEntry() (*InternalKey, base.LazyValue) {
964 1 : for m.heap.len() > 0 && m.err == nil {
965 1 : item := m.heap.items[0]
966 1 : if m.levels[item.index].isSyntheticIterBoundsKey {
967 1 : break
968 : }
969 1 : m.addItemStats(item)
970 1 : if m.isPrevEntryDeleted(item) {
971 1 : m.stats.PointsCoveredByRangeTombstones++
972 1 : continue
973 : }
974 1 : if item.iterKey.Visible(m.snapshot, m.batchSnapshot) &&
975 1 : (!m.levels[item.index].isIgnorableBoundaryKey) {
976 1 : return item.iterKey, item.iterValue
977 1 : }
978 1 : m.prevEntry(item)
979 : }
980 1 : return nil, base.LazyValue{}
981 : }
982 :
983 : // Seeks levels >= level to >= key. Additionally uses range tombstones to extend the seeks.
984 1 : func (m *mergingIter) seekGE(key []byte, level int, flags base.SeekGEFlags) {
985 1 : // When seeking, we can use tombstones to adjust the key we seek to on each
986 1 : // level. Consider the series of range tombstones:
987 1 : //
988 1 : // 1: a---e
989 1 : // 2: d---h
990 1 : // 3: g---k
991 1 : // 4: j---n
992 1 : // 5: m---q
993 1 : //
994 1 : // If we SeekGE("b") we also find the tombstone "b" resides within in the
995 1 : // first level which is [a,e). Regardless of whether this tombstone deletes
996 1 : // "b" in that level, we know it deletes "b" in all lower levels, so we
997 1 : // adjust the search key in the next level to the tombstone end key "e". We
998 1 : // then SeekGE("e") in the second level and find the corresponding tombstone
999 1 : // [d,h). This process continues and we end up seeking for "h" in the 3rd
1000 1 : // level, "k" in the 4th level and "n" in the last level.
1001 1 : //
1002 1 : // TODO(peter,rangedel): In addition to the above we can delay seeking a
1003 1 : // level (and any lower levels) when the current iterator position is
1004 1 : // contained within a range tombstone at a higher level.
1005 1 :
1006 1 : // Deterministically disable the TrySeekUsingNext optimizations sometimes in
1007 1 : // invariant builds to encourage the metamorphic tests to surface bugs. Note
1008 1 : // that we cannot disable the optimization within individual levels. It must
1009 1 : // be disabled for all levels or none. If one lower-level iterator performs
1010 1 : // a fresh seek whereas another takes advantage of its current iterator
1011 1 : // position, the heap can become inconsistent. Consider the following
1012 1 : // example:
1013 1 : //
1014 1 : // L5: [ [b-c) ] [ d ]*
1015 1 : // L6: [ b ] [e]*
1016 1 : //
1017 1 : // Imagine a SeekGE(a). The [b-c) range tombstone deletes the L6 point key
1018 1 : // 'b', resulting in the iterator positioned at d with the heap:
1019 1 : //
1020 1 : // {L5: d, L6: e}
1021 1 : //
1022 1 : // A subsequent SeekGE(b) is seeking to a larger key, so the caller may set
1023 1 : // TrySeekUsingNext()=true. If the L5 iterator used the TrySeekUsingNext
1024 1 : // optimization but the L6 iterator did not, the iterator would have the
1025 1 : // heap:
1026 1 : //
1027 1 : // {L6: b, L5: d}
1028 1 : //
1029 1 : // Because the L5 iterator has already advanced to the next sstable, the
1030 1 : // merging iterator cannot observe the [b-c) range tombstone and will
1031 1 : // mistakenly return L6's deleted point key 'b'.
1032 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !m.forceEnableSeekOpt &&
1033 1 : disableSeekOpt(key, uintptr(unsafe.Pointer(m))) {
1034 0 : flags = flags.DisableTrySeekUsingNext()
1035 0 : }
1036 :
1037 1 : for ; level < len(m.levels); level++ {
1038 1 : if invariants.Enabled && m.lower != nil && m.heap.cmp(key, m.lower) < 0 {
1039 0 : m.logger.Fatalf("mergingIter: lower bound violation: %s < %s\n%s", key, m.lower, debug.Stack())
1040 0 : }
1041 :
1042 1 : l := &m.levels[level]
1043 1 : if m.prefix != nil {
1044 1 : l.iterKey, l.iterValue = l.iter.SeekPrefixGE(m.prefix, key, flags)
1045 1 : } else {
1046 1 : l.iterKey, l.iterValue = l.iter.SeekGE(key, flags)
1047 1 : }
1048 :
1049 : // If this level contains overlapping range tombstones, alter the seek
1050 : // key accordingly. Caveat: If we're performing lazy-combined iteration,
1051 : // we cannot alter the seek key: Range tombstones don't delete range
1052 : // keys, and there might exist live range keys within the range
1053 : // tombstone's span that need to be observed to trigger a switch to
1054 : // combined iteration.
1055 1 : if rangeDelIter := l.rangeDelIter; rangeDelIter != nil &&
1056 1 : (m.combinedIterState == nil || m.combinedIterState.initialized) {
1057 1 : // The level has a range-del iterator. Find the tombstone containing
1058 1 : // the search key.
1059 1 : //
1060 1 : // For untruncated tombstones that are possibly file-bounds-constrained, we are using a
1061 1 : // levelIter which will set smallestUserKey and largestUserKey. Since the levelIter
1062 1 : // is at this file we know that largestUserKey >= key, so we know that the
1063 1 : // tombstone we find cannot be file-bounds-constrained in its upper bound to something < key.
1064 1 : // We do need to compare with smallestUserKey to ensure that the tombstone is not
1065 1 : // file-bounds-constrained in its lower bound.
1066 1 : //
1067 1 : // See the detailed comments in isNextEntryDeleted() on why similar containment and
1068 1 : // seeking logic is correct. The subtle difference here is that key is a user key,
1069 1 : // so we can have a sstable with bounds [c#8, i#InternalRangeDelSentinel], and the
1070 1 : // tombstone is [b, k)#8 and the seek key is i: levelIter.SeekGE(i) will move past
1071 1 : // this sstable since it realizes the largest key is a InternalRangeDelSentinel.
1072 1 : l.tombstone = rangeDelIter.SeekGE(key)
1073 1 : if l.tombstone != nil && l.tombstone.VisibleAt(m.snapshot) && l.tombstone.Contains(m.heap.cmp, key) &&
1074 1 : (l.smallestUserKey == nil || m.heap.cmp(l.smallestUserKey, key) <= 0) {
1075 1 : // NB: Based on the comment above l.largestUserKey >= key, and based on the
1076 1 : // containment condition tombstone.End > key, so the assignment to key results
1077 1 : // in a monotonically non-decreasing key across iterations of this loop.
1078 1 : //
1079 1 : // The adjustment of key here can only move it to a larger key. Since
1080 1 : // the caller of seekGE guaranteed that the original key was greater
1081 1 : // than or equal to m.lower, the new key will continue to be greater
1082 1 : // than or equal to m.lower.
1083 1 : if l.largestUserKey != nil &&
1084 1 : m.heap.cmp(l.largestUserKey, l.tombstone.End) < 0 {
1085 0 : // Truncate the tombstone for seeking purposes. Note that this can over-truncate
1086 0 : // but that is harmless for this seek optimization.
1087 0 : key = l.largestUserKey
1088 1 : } else {
1089 1 : key = l.tombstone.End
1090 1 : }
1091 : }
1092 : }
1093 : }
1094 :
1095 1 : m.initMinHeap()
1096 : }
1097 :
1098 0 : func (m *mergingIter) String() string {
1099 0 : return "merging"
1100 0 : }
1101 :
1102 : // SeekGE implements base.InternalIterator.SeekGE. Note that SeekGE only checks
1103 : // the upper bound. It is up to the caller to ensure that key is greater than
1104 : // or equal to the lower bound.
1105 1 : func (m *mergingIter) SeekGE(key []byte, flags base.SeekGEFlags) (*InternalKey, base.LazyValue) {
1106 1 : m.err = nil // clear cached iteration error
1107 1 : m.prefix = nil
1108 1 : m.seekGE(key, 0 /* start level */, flags)
1109 1 : return m.findNextEntry()
1110 1 : }
1111 :
1112 : // SeekPrefixGE implements base.InternalIterator.SeekPrefixGE. Note that
1113 : // SeekPrefixGE only checks the upper bound. It is up to the caller to ensure
1114 : // that key is greater than or equal to the lower bound.
1115 : func (m *mergingIter) SeekPrefixGE(
1116 : prefix, key []byte, flags base.SeekGEFlags,
1117 1 : ) (*base.InternalKey, base.LazyValue) {
1118 1 : m.err = nil // clear cached iteration error
1119 1 : m.prefix = prefix
1120 1 : m.seekGE(key, 0 /* start level */, flags)
1121 1 : return m.findNextEntry()
1122 1 : }
1123 :
1124 : // Seeks levels >= level to < key. Additionally uses range tombstones to extend the seeks.
1125 1 : func (m *mergingIter) seekLT(key []byte, level int, flags base.SeekLTFlags) {
1126 1 : // See the comment in seekGE regarding using tombstones to adjust the seek
1127 1 : // target per level.
1128 1 : m.prefix = nil
1129 1 : for ; level < len(m.levels); level++ {
1130 1 : if invariants.Enabled && m.upper != nil && m.heap.cmp(key, m.upper) > 0 {
1131 0 : m.logger.Fatalf("mergingIter: upper bound violation: %s > %s\n%s", key, m.upper, debug.Stack())
1132 0 : }
1133 :
1134 1 : l := &m.levels[level]
1135 1 : l.iterKey, l.iterValue = l.iter.SeekLT(key, flags)
1136 1 :
1137 1 : // If this level contains overlapping range tombstones, alter the seek
1138 1 : // key accordingly. Caveat: If we're performing lazy-combined iteration,
1139 1 : // we cannot alter the seek key: Range tombstones don't delete range
1140 1 : // keys, and there might exist live range keys within the range
1141 1 : // tombstone's span that need to be observed to trigger a switch to
1142 1 : // combined iteration.
1143 1 : if rangeDelIter := l.rangeDelIter; rangeDelIter != nil &&
1144 1 : (m.combinedIterState == nil || m.combinedIterState.initialized) {
1145 1 : // The level has a range-del iterator. Find the tombstone containing
1146 1 : // the search key.
1147 1 : //
1148 1 : // For untruncated tombstones that are possibly file-bounds-constrained we are using a
1149 1 : // levelIter which will set smallestUserKey and largestUserKey. Since the levelIter
1150 1 : // is at this file we know that smallestUserKey <= key, so we know that the
1151 1 : // tombstone we find cannot be file-bounds-constrained in its lower bound to something > key.
1152 1 : // We do need to compare with largestUserKey to ensure that the tombstone is not
1153 1 : // file-bounds-constrained in its upper bound.
1154 1 : //
1155 1 : // See the detailed comments in isPrevEntryDeleted() on why similar containment and
1156 1 : // seeking logic is correct.
1157 1 :
1158 1 : // Default to within bounds.
1159 1 : withinLargestSSTableBound := true
1160 1 : if l.largestUserKey != nil {
1161 1 : cmpResult := m.heap.cmp(l.largestUserKey, key)
1162 1 : withinLargestSSTableBound = cmpResult > 0 || (cmpResult == 0 && !l.isLargestUserKeyExclusive)
1163 1 : }
1164 :
1165 1 : l.tombstone = keyspan.SeekLE(m.heap.cmp, rangeDelIter, key)
1166 1 : if l.tombstone != nil && l.tombstone.VisibleAt(m.snapshot) &&
1167 1 : l.tombstone.Contains(m.heap.cmp, key) && withinLargestSSTableBound {
1168 1 : // NB: Based on the comment above l.smallestUserKey <= key, and based
1169 1 : // on the containment condition tombstone.Start.UserKey <= key, so the
1170 1 : // assignment to key results in a monotonically non-increasing key
1171 1 : // across iterations of this loop.
1172 1 : //
1173 1 : // The adjustment of key here can only move it to a smaller key. Since
1174 1 : // the caller of seekLT guaranteed that the original key was less than
1175 1 : // or equal to m.upper, the new key will continue to be less than or
1176 1 : // equal to m.upper.
1177 1 : if l.smallestUserKey != nil &&
1178 1 : m.heap.cmp(l.smallestUserKey, l.tombstone.Start) >= 0 {
1179 1 : // Truncate the tombstone for seeking purposes. Note that this can over-truncate
1180 1 : // but that is harmless for this seek optimization.
1181 1 : key = l.smallestUserKey
1182 1 : } else {
1183 1 : key = l.tombstone.Start
1184 1 : }
1185 : }
1186 : }
1187 : }
1188 :
1189 1 : m.initMaxHeap()
1190 : }
1191 :
1192 : // SeekLT implements base.InternalIterator.SeekLT. Note that SeekLT only checks
1193 : // the lower bound. It is up to the caller to ensure that key is less than the
1194 : // upper bound.
1195 1 : func (m *mergingIter) SeekLT(key []byte, flags base.SeekLTFlags) (*InternalKey, base.LazyValue) {
1196 1 : m.err = nil // clear cached iteration error
1197 1 : m.prefix = nil
1198 1 : m.seekLT(key, 0 /* start level */, flags)
1199 1 : return m.findPrevEntry()
1200 1 : }
1201 :
1202 : // First implements base.InternalIterator.First. Note that First only checks
1203 : // the upper bound. It is up to the caller to ensure that key is greater than
1204 : // or equal to the lower bound (e.g. via a call to SeekGE(lower)).
1205 1 : func (m *mergingIter) First() (*InternalKey, base.LazyValue) {
1206 1 : m.err = nil // clear cached iteration error
1207 1 : m.prefix = nil
1208 1 : m.heap.items = m.heap.items[:0]
1209 1 : for i := range m.levels {
1210 1 : l := &m.levels[i]
1211 1 : l.iterKey, l.iterValue = l.iter.First()
1212 1 : }
1213 1 : m.initMinHeap()
1214 1 : return m.findNextEntry()
1215 : }
1216 :
1217 : // Last implements base.InternalIterator.Last. Note that Last only checks the
1218 : // lower bound. It is up to the caller to ensure that key is less than the
1219 : // upper bound (e.g. via a call to SeekLT(upper))
1220 1 : func (m *mergingIter) Last() (*InternalKey, base.LazyValue) {
1221 1 : m.err = nil // clear cached iteration error
1222 1 : m.prefix = nil
1223 1 : for i := range m.levels {
1224 1 : l := &m.levels[i]
1225 1 : l.iterKey, l.iterValue = l.iter.Last()
1226 1 : }
1227 1 : m.initMaxHeap()
1228 1 : return m.findPrevEntry()
1229 : }
1230 :
1231 1 : func (m *mergingIter) Next() (*InternalKey, base.LazyValue) {
1232 1 : if m.err != nil {
1233 0 : return nil, base.LazyValue{}
1234 0 : }
1235 :
1236 1 : if m.dir != 1 {
1237 1 : m.switchToMinHeap()
1238 1 : return m.findNextEntry()
1239 1 : }
1240 :
1241 1 : if m.heap.len() == 0 {
1242 0 : return nil, base.LazyValue{}
1243 0 : }
1244 :
1245 : // NB: It's okay to call nextEntry directly even during prefix iteration
1246 : // mode (as opposed to indirectly through maybeNextEntryWithinPrefix).
1247 : // During prefix iteration mode, we rely on the caller to not call Next if
1248 : // the iterator has already advanced beyond the iteration prefix. See the
1249 : // comment above the base.InternalIterator interface.
1250 1 : m.nextEntry(m.heap.items[0], nil /* succKey */)
1251 1 : return m.findNextEntry()
1252 : }
1253 :
1254 1 : func (m *mergingIter) NextPrefix(succKey []byte) (*InternalKey, LazyValue) {
1255 1 : if m.dir != 1 {
1256 0 : panic("pebble: cannot switch directions with NextPrefix")
1257 : }
1258 1 : if m.err != nil || m.heap.len() == 0 {
1259 0 : return nil, LazyValue{}
1260 0 : }
1261 1 : if m.levelsPositioned == nil {
1262 1 : m.levelsPositioned = make([]bool, len(m.levels))
1263 1 : } else {
1264 1 : for i := range m.levelsPositioned {
1265 1 : m.levelsPositioned[i] = false
1266 1 : }
1267 : }
1268 :
1269 : // The heap root necessarily must be positioned at a key < succKey, because
1270 : // NextPrefix was invoked.
1271 1 : root := &m.heap.items[0]
1272 1 : m.levelsPositioned[(*root).index] = true
1273 1 : if invariants.Enabled && m.heap.cmp((*root).iterKey.UserKey, succKey) >= 0 {
1274 0 : m.logger.Fatalf("pebble: invariant violation: NextPrefix(%q) called on merging iterator already positioned at %q",
1275 0 : succKey, (*root).iterKey)
1276 0 : }
1277 1 : m.nextEntry(*root, succKey)
1278 1 : // NB: root is a pointer to the heap root. nextEntry may have changed
1279 1 : // the heap root, so we must not expect root to still point to the same
1280 1 : // level (or to even be valid, if the heap is now exhaused).
1281 1 :
1282 1 : for m.heap.len() > 0 {
1283 1 : if m.levelsPositioned[(*root).index] {
1284 1 : // A level we've previously positioned is at the top of the heap, so
1285 1 : // there are no other levels positioned at keys < succKey. We've
1286 1 : // advanced as far as we need to.
1287 1 : break
1288 : }
1289 : // Since this level was not the original heap root when NextPrefix was
1290 : // called, we don't know whether this level's current key has the
1291 : // previous prefix or a new one.
1292 1 : if m.heap.cmp((*root).iterKey.UserKey, succKey) >= 0 {
1293 1 : break
1294 : }
1295 1 : m.levelsPositioned[(*root).index] = true
1296 1 : m.nextEntry(*root, succKey)
1297 : }
1298 1 : return m.findNextEntry()
1299 : }
1300 :
1301 1 : func (m *mergingIter) Prev() (*InternalKey, base.LazyValue) {
1302 1 : if m.err != nil {
1303 0 : return nil, base.LazyValue{}
1304 0 : }
1305 :
1306 1 : if m.dir != -1 {
1307 1 : if m.prefix != nil {
1308 0 : m.err = errors.New("pebble: unsupported reverse prefix iteration")
1309 0 : return nil, base.LazyValue{}
1310 0 : }
1311 1 : m.switchToMaxHeap()
1312 1 : return m.findPrevEntry()
1313 : }
1314 :
1315 1 : if m.heap.len() == 0 {
1316 0 : return nil, base.LazyValue{}
1317 0 : }
1318 :
1319 1 : m.prevEntry(m.heap.items[0])
1320 1 : return m.findPrevEntry()
1321 : }
1322 :
1323 1 : func (m *mergingIter) Error() error {
1324 1 : if m.heap.len() == 0 || m.err != nil {
1325 1 : return m.err
1326 1 : }
1327 1 : return m.levels[m.heap.items[0].index].iter.Error()
1328 : }
1329 :
1330 1 : func (m *mergingIter) Close() error {
1331 1 : for i := range m.levels {
1332 1 : iter := m.levels[i].iter
1333 1 : if err := iter.Close(); err != nil && m.err == nil {
1334 0 : m.err = err
1335 0 : }
1336 1 : if rangeDelIter := m.levels[i].rangeDelIter; rangeDelIter != nil {
1337 1 : if err := rangeDelIter.Close(); err != nil && m.err == nil {
1338 0 : m.err = err
1339 0 : }
1340 : }
1341 : }
1342 1 : m.levels = nil
1343 1 : m.heap.items = m.heap.items[:0]
1344 1 : return m.err
1345 : }
1346 :
1347 1 : func (m *mergingIter) SetBounds(lower, upper []byte) {
1348 1 : m.prefix = nil
1349 1 : m.lower = lower
1350 1 : m.upper = upper
1351 1 : for i := range m.levels {
1352 1 : m.levels[i].iter.SetBounds(lower, upper)
1353 1 : }
1354 1 : m.heap.clear()
1355 : }
1356 :
1357 0 : func (m *mergingIter) DebugString() string {
1358 0 : var buf bytes.Buffer
1359 0 : sep := ""
1360 0 : for m.heap.len() > 0 {
1361 0 : item := m.heap.pop()
1362 0 : fmt.Fprintf(&buf, "%s%s", sep, item.iterKey)
1363 0 : sep = " "
1364 0 : }
1365 0 : if m.dir == 1 {
1366 0 : m.initMinHeap()
1367 0 : } else {
1368 0 : m.initMaxHeap()
1369 0 : }
1370 0 : return buf.String()
1371 : }
1372 :
1373 1 : func (m *mergingIter) ForEachLevelIter(fn func(li *levelIter) bool) {
1374 1 : for _, iter := range m.levels {
1375 1 : if li, ok := iter.iter.(*levelIter); ok {
1376 1 : if done := fn(li); done {
1377 1 : break
1378 : }
1379 : }
1380 : }
1381 : }
1382 :
1383 1 : func (m *mergingIter) addItemStats(l *mergingIterLevel) {
1384 1 : m.stats.PointCount++
1385 1 : m.stats.KeyBytes += uint64(len(l.iterKey.UserKey))
1386 1 : m.stats.ValueBytes += uint64(len(l.iterValue.ValueOrHandle))
1387 1 : }
1388 :
1389 : var _ internalIterator = &mergingIter{}
|
