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 keyspan
6 :
7 : import (
8 : "fmt"
9 : "sort"
10 :
11 : "github.com/cockroachdb/pebble/internal/base"
12 : "github.com/cockroachdb/pebble/internal/invariants"
13 : )
14 :
15 : type spansByStartKey struct {
16 : cmp base.Compare
17 : buf []Span
18 : }
19 :
20 1 : func (v *spansByStartKey) Len() int { return len(v.buf) }
21 1 : func (v *spansByStartKey) Less(i, j int) bool {
22 1 : return v.cmp(v.buf[i].Start, v.buf[j].Start) < 0
23 1 : }
24 1 : func (v *spansByStartKey) Swap(i, j int) {
25 1 : v.buf[i], v.buf[j] = v.buf[j], v.buf[i]
26 1 : }
27 :
28 : type spansByEndKey struct {
29 : cmp base.Compare
30 : buf []Span
31 : }
32 :
33 2 : func (v *spansByEndKey) Len() int { return len(v.buf) }
34 2 : func (v *spansByEndKey) Less(i, j int) bool {
35 2 : return v.cmp(v.buf[i].End, v.buf[j].End) < 0
36 2 : }
37 2 : func (v *spansByEndKey) Swap(i, j int) {
38 2 : v.buf[i], v.buf[j] = v.buf[j], v.buf[i]
39 2 : }
40 :
41 : // keysBySeqNumKind sorts spans by the start key's sequence number in
42 : // descending order. If two spans have equal sequence number, they're compared
43 : // by key kind in descending order. This ordering matches the ordering of
44 : // base.InternalCompare among keys with matching user keys.
45 : type keysBySeqNumKind []Key
46 :
47 2 : func (v *keysBySeqNumKind) Len() int { return len(*v) }
48 2 : func (v *keysBySeqNumKind) Less(i, j int) bool { return (*v)[i].Trailer > (*v)[j].Trailer }
49 2 : func (v *keysBySeqNumKind) Swap(i, j int) { (*v)[i], (*v)[j] = (*v)[j], (*v)[i] }
50 :
51 : // Sort the spans by start key. This is the ordering required by the
52 : // Fragmenter. Usually spans are naturally sorted by their start key,
53 : // but that isn't true for range deletion tombstones in the legacy
54 : // range-del-v1 block format.
55 1 : func Sort(cmp base.Compare, spans []Span) {
56 1 : sorter := spansByStartKey{
57 1 : cmp: cmp,
58 1 : buf: spans,
59 1 : }
60 1 : sort.Sort(&sorter)
61 1 : }
62 :
63 : // Fragmenter fragments a set of spans such that overlapping spans are
64 : // split at their overlap points. The fragmented spans are output to the
65 : // supplied Output function.
66 : type Fragmenter struct {
67 : Cmp base.Compare
68 : Format base.FormatKey
69 : // Emit is called to emit a fragmented span and its keys. Every key defined
70 : // within the emitted Span applies to the entirety of the Span's key span.
71 : // Keys are ordered in decreasing order of their sequence numbers, and if
72 : // equal, decreasing order of key kind.
73 : Emit func(Span)
74 : // pending contains the list of pending fragments that have not been
75 : // flushed to the block writer. Note that the spans have not been
76 : // fragmented on the end keys yet. That happens as the spans are
77 : // flushed. All pending spans have the same Start.
78 : pending []Span
79 : // doneBuf is used to buffer completed span fragments when flushing to a
80 : // specific key (e.g. TruncateAndFlushTo). It is cached in the Fragmenter to
81 : // allow reuse.
82 : doneBuf []Span
83 : // sortBuf is used to sort fragments by end key when flushing.
84 : sortBuf spansByEndKey
85 : // flushBuf is used to sort keys by (seqnum,kind) before emitting.
86 : flushBuf keysBySeqNumKind
87 : // flushedKey is the key that fragments have been flushed up to. Any
88 : // additional spans added to the fragmenter must have a start key >=
89 : // flushedKey. A nil value indicates flushedKey has not been set.
90 : flushedKey []byte
91 : finished bool
92 : }
93 :
94 0 : func (f *Fragmenter) checkInvariants(buf []Span) {
95 0 : for i := 1; i < len(buf); i++ {
96 0 : if f.Cmp(buf[i].Start, buf[i].End) >= 0 {
97 0 : panic(fmt.Sprintf("pebble: empty pending span invariant violated: %s", buf[i]))
98 : }
99 0 : if f.Cmp(buf[i-1].Start, buf[i].Start) != 0 {
100 0 : panic(fmt.Sprintf("pebble: pending span invariant violated: %s %s",
101 0 : f.Format(buf[i-1].Start), f.Format(buf[i].Start)))
102 : }
103 : }
104 : }
105 :
106 : // Add adds a span to the fragmenter. Spans may overlap and the
107 : // fragmenter will internally split them. The spans must be presented in
108 : // increasing start key order. That is, Add must be called with a series
109 : // of spans like:
110 : //
111 : // a---e
112 : // c---g
113 : // c-----i
114 : // j---n
115 : // j-l
116 : //
117 : // We need to fragment the spans at overlap points. In the above
118 : // example, we'd create:
119 : //
120 : // a-c-e
121 : // c-e-g
122 : // c-e-g-i
123 : // j-l-n
124 : // j-l
125 : //
126 : // The fragments need to be output sorted by start key, and for equal start
127 : // keys, sorted by descending sequence number. This last part requires a mild
128 : // bit of care as the fragments are not created in descending sequence number
129 : // order.
130 : //
131 : // Once a start key has been seen, we know that we'll never see a smaller
132 : // start key and can thus flush all of the fragments that lie before that
133 : // start key.
134 : //
135 : // Walking through the example above, we start with:
136 : //
137 : // a---e
138 : //
139 : // Next we add [c,g) resulting in:
140 : //
141 : // a-c-e
142 : // c---g
143 : //
144 : // The fragment [a,c) is flushed leaving the pending spans as:
145 : //
146 : // c-e
147 : // c---g
148 : //
149 : // The next span is [c,i):
150 : //
151 : // c-e
152 : // c---g
153 : // c-----i
154 : //
155 : // No fragments are flushed. The next span is [j,n):
156 : //
157 : // c-e
158 : // c---g
159 : // c-----i
160 : // j---n
161 : //
162 : // The fragments [c,e), [c,g) and [c,i) are flushed. We sort these fragments
163 : // by their end key, then split the fragments on the end keys:
164 : //
165 : // c-e
166 : // c-e-g
167 : // c-e---i
168 : //
169 : // The [c,e) fragments all get flushed leaving:
170 : //
171 : // e-g
172 : // e---i
173 : //
174 : // This process continues until there are no more fragments to flush.
175 : //
176 : // WARNING: the slices backing Start, End, Keys, Key.Suffix and Key.Value are
177 : // all retained after this method returns and should not be modified. This is
178 : // safe for spans that are added from a memtable or batch. It is partially
179 : // unsafe for a span read from an sstable. Specifically, the Keys slice of a
180 : // Span returned during sstable iteration is only valid until the next iterator
181 : // operation. The stability of the user keys depend on whether the block is
182 : // prefix compressed, and in practice Pebble never prefix compresses range
183 : // deletion and range key blocks, so these keys are stable. Because of this key
184 : // stability, typically callers only need to perform a shallow clone of the Span
185 : // before Add-ing it to the fragmenter.
186 : //
187 : // Add requires the provided span's keys are sorted in Trailer descending order.
188 2 : func (f *Fragmenter) Add(s Span) {
189 2 : if f.finished {
190 0 : panic("pebble: span fragmenter already finished")
191 2 : } else if s.KeysOrder != ByTrailerDesc {
192 0 : panic("pebble: span keys unexpectedly not in trailer descending order")
193 : }
194 2 : if f.flushedKey != nil {
195 2 : switch c := f.Cmp(s.Start, f.flushedKey); {
196 0 : case c < 0:
197 0 : panic(fmt.Sprintf("pebble: start key (%s) < flushed key (%s)",
198 0 : f.Format(s.Start), f.Format(f.flushedKey)))
199 : }
200 : }
201 2 : if f.Cmp(s.Start, s.End) >= 0 {
202 1 : // An empty span, we can ignore it.
203 1 : return
204 1 : }
205 2 : if invariants.RaceEnabled {
206 0 : f.checkInvariants(f.pending)
207 0 : defer func() { f.checkInvariants(f.pending) }()
208 : }
209 :
210 2 : if len(f.pending) > 0 {
211 2 : // Since all of the pending spans have the same start key, we only need
212 2 : // to compare against the first one.
213 2 : switch c := f.Cmp(f.pending[0].Start, s.Start); {
214 1 : case c > 0:
215 1 : panic(fmt.Sprintf("pebble: keys must be added in order: %s > %s",
216 1 : f.Format(f.pending[0].Start), f.Format(s.Start)))
217 2 : case c == 0:
218 2 : // The new span has the same start key as the existing pending
219 2 : // spans. Add it to the pending buffer.
220 2 : f.pending = append(f.pending, s)
221 2 : return
222 : }
223 :
224 : // At this point we know that the new start key is greater than the pending
225 : // spans start keys.
226 2 : f.truncateAndFlush(s.Start)
227 : }
228 :
229 2 : f.pending = append(f.pending, s)
230 : }
231 :
232 : // Empty returns true if all fragments added so far have finished flushing.
233 0 : func (f *Fragmenter) Empty() bool {
234 0 : return f.finished || len(f.pending) == 0
235 0 : }
236 :
237 : // Start returns the start key of the first span in the pending buffer, or nil
238 : // if there are no pending spans. The start key of all pending spans is the same
239 : // as that of the first one.
240 1 : func (f *Fragmenter) Start() []byte {
241 1 : if len(f.pending) > 0 {
242 1 : return f.pending[0].Start
243 1 : }
244 1 : return nil
245 : }
246 :
247 : // Flushes all pending spans up to key (exclusive).
248 : //
249 : // WARNING: The specified key is stored without making a copy, so all callers
250 : // must ensure it is safe.
251 2 : func (f *Fragmenter) truncateAndFlush(key []byte) {
252 2 : f.flushedKey = append(f.flushedKey[:0], key...)
253 2 : done := f.doneBuf[:0]
254 2 : pending := f.pending
255 2 : f.pending = f.pending[:0]
256 2 :
257 2 : // pending and f.pending share the same underlying storage. As we iterate
258 2 : // over pending we append to f.pending, but only one entry is appended in
259 2 : // each iteration, after we have read the entry being overwritten.
260 2 : for _, s := range pending {
261 2 : if f.Cmp(key, s.End) < 0 {
262 2 : // s: a--+--e
263 2 : // new: c------
264 2 : if f.Cmp(s.Start, key) < 0 {
265 2 : done = append(done, Span{
266 2 : Start: s.Start,
267 2 : End: key,
268 2 : Keys: s.Keys,
269 2 : })
270 2 : }
271 2 : f.pending = append(f.pending, Span{
272 2 : Start: key,
273 2 : End: s.End,
274 2 : Keys: s.Keys,
275 2 : })
276 2 : } else {
277 2 : // s: a-----e
278 2 : // new: e----
279 2 : done = append(done, s)
280 2 : }
281 : }
282 :
283 2 : f.doneBuf = done[:0]
284 2 : f.flush(done, nil)
285 : }
286 :
287 : // flush a group of range spans to the block. The spans are required to all have
288 : // the same start key. We flush all span fragments until startKey > lastKey. If
289 : // lastKey is nil, all span fragments are flushed. The specification of a
290 : // non-nil lastKey occurs for range deletion tombstones during compaction where
291 : // we want to flush (but not truncate) all range tombstones that start at or
292 : // before the first key in the next sstable. Consider:
293 : //
294 : // a---e#10
295 : // a------h#9
296 : //
297 : // If a compaction splits the sstables at key c we want the first sstable to
298 : // contain the tombstones [a,e)#10 and [a,e)#9. Fragmentation would naturally
299 : // produce a tombstone [e,h)#9, but we don't need to output that tombstone to
300 : // the first sstable.
301 2 : func (f *Fragmenter) flush(buf []Span, lastKey []byte) {
302 2 : if invariants.RaceEnabled {
303 0 : f.checkInvariants(buf)
304 0 : }
305 :
306 : // Sort the spans by end key. This will allow us to walk over the spans and
307 : // easily determine the next split point (the smallest end-key).
308 2 : f.sortBuf.cmp = f.Cmp
309 2 : f.sortBuf.buf = buf
310 2 : sort.Sort(&f.sortBuf)
311 2 :
312 2 : // Loop over the spans, splitting by end key.
313 2 : for len(buf) > 0 {
314 2 : // A prefix of spans will end at split. remove represents the count of
315 2 : // that prefix.
316 2 : remove := 1
317 2 : split := buf[0].End
318 2 : f.flushBuf = append(f.flushBuf[:0], buf[0].Keys...)
319 2 :
320 2 : for i := 1; i < len(buf); i++ {
321 2 : if f.Cmp(split, buf[i].End) == 0 {
322 2 : remove++
323 2 : }
324 2 : f.flushBuf = append(f.flushBuf, buf[i].Keys...)
325 : }
326 :
327 2 : sort.Sort(&f.flushBuf)
328 2 :
329 2 : f.Emit(Span{
330 2 : Start: buf[0].Start,
331 2 : End: split,
332 2 : // Copy the sorted keys to a new slice.
333 2 : //
334 2 : // This allocation is an unfortunate side effect of the Fragmenter and
335 2 : // the expectation that the spans it produces are available in-memory
336 2 : // indefinitely.
337 2 : //
338 2 : // Eventually, we should be able to replace the fragmenter with the
339 2 : // keyspanimpl.MergingIter which will perform just-in-time
340 2 : // fragmentation, and only guaranteeing the memory lifetime for the
341 2 : // current span. The MergingIter fragments while only needing to
342 2 : // access one Span per level. It only accesses the Span at the
343 2 : // current position for each level. During compactions, we can write
344 2 : // these spans to sstables without retaining previous Spans.
345 2 : Keys: append([]Key(nil), f.flushBuf...),
346 2 : })
347 2 :
348 2 : if lastKey != nil && f.Cmp(split, lastKey) > 0 {
349 0 : break
350 : }
351 :
352 : // Adjust the start key for every remaining span.
353 2 : buf = buf[remove:]
354 2 : for i := range buf {
355 2 : buf[i].Start = split
356 2 : }
357 : }
358 : }
359 :
360 : // Finish flushes any remaining fragments to the output. It is an error to call
361 : // this if any other spans will be added.
362 2 : func (f *Fragmenter) Finish() {
363 2 : if f.finished {
364 0 : panic("pebble: span fragmenter already finished")
365 : }
366 2 : f.flush(f.pending, nil)
367 2 : f.finished = true
368 : }
|