forked from grafana/loki
/
recovery.go
416 lines (351 loc) · 10.5 KB
/
recovery.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
package ingester
import (
io "io"
"runtime"
"sync"
"github.com/go-kit/log/level"
"github.com/pkg/errors"
"github.com/prometheus/prometheus/tsdb/chunks"
"github.com/prometheus/prometheus/tsdb/record"
"github.com/prometheus/prometheus/tsdb/wal"
"golang.org/x/net/context"
"github.com/grafana/loki/pkg/logproto"
util_log "github.com/grafana/loki/pkg/util/log"
)
type WALReader interface {
Next() bool
Err() error
// Record should not be used across multiple calls to Next()
Record() []byte
}
type NoopWALReader struct{}
func (NoopWALReader) Next() bool { return false }
func (NoopWALReader) Err() error { return nil }
func (NoopWALReader) Record() []byte { return nil }
func (NoopWALReader) Close() error { return nil }
func newCheckpointReader(dir string) (WALReader, io.Closer, error) {
lastCheckpointDir, idx, err := lastCheckpoint(dir)
if err != nil {
return nil, nil, err
}
if idx < 0 {
level.Info(util_log.Logger).Log("msg", "no checkpoint found, treating as no-op")
var reader NoopWALReader
return reader, reader, nil
}
r, err := wal.NewSegmentsReader(lastCheckpointDir)
if err != nil {
return nil, nil, err
}
return wal.NewReader(r), r, nil
}
type Recoverer interface {
NumWorkers() int
Series(series *Series) error
SetStream(userID string, series record.RefSeries) error
Push(userID string, entries RefEntries) error
Done() <-chan struct{}
}
type ingesterRecoverer struct {
// basically map[userID]map[fingerprint]*stream
users sync.Map
ing *Ingester
done chan struct{}
}
func newIngesterRecoverer(i *Ingester) *ingesterRecoverer {
return &ingesterRecoverer{
ing: i,
done: make(chan struct{}),
}
}
// Use all available cores
func (r *ingesterRecoverer) NumWorkers() int { return runtime.GOMAXPROCS(0) }
func (r *ingesterRecoverer) Series(series *Series) error {
return r.ing.replayController.WithBackPressure(func() error {
inst, err := r.ing.GetOrCreateInstance(series.UserID)
if err != nil {
return err
}
// TODO(owen-d): create another fn to avoid unnecessary label type conversions.
stream, err := inst.getOrCreateStream(logproto.Stream{
Labels: logproto.FromLabelAdaptersToLabels(series.Labels).String(),
}, nil)
if err != nil {
return err
}
bytesAdded, entriesAdded, err := stream.setChunks(series.Chunks)
stream.lastLine.ts = series.To
stream.lastLine.content = series.LastLine
stream.entryCt = series.EntryCt
stream.highestTs = series.HighestTs
if err != nil {
return err
}
r.ing.metrics.memoryChunks.Add(float64(len(series.Chunks)))
r.ing.metrics.recoveredChunksTotal.Add(float64(len(series.Chunks)))
r.ing.metrics.recoveredEntriesTotal.Add(float64(entriesAdded))
r.ing.replayController.Add(int64(bytesAdded))
// now store the stream in the recovery map under the fingerprint originally recorded
// as it's possible the newly mapped fingerprint is different. This is because the WAL records
// will use this original reference.
got, _ := r.users.LoadOrStore(series.UserID, &sync.Map{})
streamsMap := got.(*sync.Map)
streamsMap.Store(chunks.HeadSeriesRef(series.Fingerprint), stream)
return nil
})
}
// SetStream is responsible for setting the key path for userIDs -> fingerprints -> streams.
// Internally, this uses nested sync.Maps due to their performance benefits for sets that only grow.
// Using these also allows us to bypass the ingester -> instance -> stream hierarchy internally, which
// may yield some performance gains, but is essential for the following:
// Due to the use of the instance's fingerprint mapper, stream fingerprints are NOT necessarily
// deterministic. The WAL uses the post-mapped fingerprint on the ingester that originally
// created the stream and we ensure that said fingerprint maps correctly to the newly
// created stream during WAL replay, even if the new in memory stream was assigned a different
// fingerprint from the mapper. This is paramount because subsequent WAL records will use
// the fingerprint reported in the WAL record, not the potentially differing one assigned during
// stream creation.
func (r *ingesterRecoverer) SetStream(userID string, series record.RefSeries) error {
inst, err := r.ing.GetOrCreateInstance(userID)
if err != nil {
return err
}
stream, err := inst.getOrCreateStream(
logproto.Stream{
Labels: series.Labels.String(),
},
nil,
)
if err != nil {
return err
}
// Now that we have the stream, ensure that the userID -> fingerprint -> stream
// path is set properly.
got, _ := r.users.LoadOrStore(userID, &sync.Map{})
streamsMap := got.(*sync.Map)
streamsMap.Store(series.Ref, stream)
return nil
}
func (r *ingesterRecoverer) Push(userID string, entries RefEntries) error {
return r.ing.replayController.WithBackPressure(func() error {
out, ok := r.users.Load(userID)
if !ok {
return errors.Errorf("user (%s) not set during WAL replay", userID)
}
s, ok := out.(*sync.Map).Load(entries.Ref)
if !ok {
return errors.Errorf("stream (%d) not set during WAL replay for user (%s)", entries.Ref, userID)
}
// ignore out of order errors here (it's possible for a checkpoint to already have data from the wal segments)
bytesAdded, err := s.(*stream).Push(context.Background(), entries.Entries, nil, entries.Counter, true)
r.ing.replayController.Add(int64(bytesAdded))
if err != nil && err == ErrEntriesExist {
r.ing.metrics.duplicateEntriesTotal.Add(float64(len(entries.Entries)))
}
return nil
})
}
func (r *ingesterRecoverer) Close() {
// Ensure this is only run once.
select {
case <-r.done:
return
default:
}
close(r.done)
// Enable the limiter here to accurately reflect tenant limits after recovery.
r.ing.limiter.Enable()
for _, inst := range r.ing.getInstances() {
inst.forAllStreams(context.Background(), func(s *stream) error {
s.chunkMtx.Lock()
defer s.chunkMtx.Unlock()
// reset all the incrementing stream counters after a successful WAL replay.
s.resetCounter()
if len(s.chunks) == 0 {
inst.removeStream(s)
return nil
}
// If we've replayed a WAL with unordered writes, but the new
// configuration disables them, convert all streams/head blocks
// to ensure unordered writes are disabled after the replay,
// but without dropping any previously accepted data.
isAllowed := r.ing.limiter.UnorderedWrites(s.tenant)
old := s.unorderedWrites
s.unorderedWrites = isAllowed
if !isAllowed && old {
err := s.chunks[len(s.chunks)-1].chunk.ConvertHead(headBlockType(isAllowed))
if err != nil {
level.Warn(util_log.Logger).Log(
"msg", "error converting headblock",
"err", err.Error(),
"stream", s.labels.String(),
"component", "ingesterRecoverer",
)
}
}
return nil
})
}
}
func (r *ingesterRecoverer) Done() <-chan struct{} {
return r.done
}
func RecoverWAL(reader WALReader, recoverer Recoverer) error {
dispatch := func(recoverer Recoverer, b []byte, inputs []chan recoveryInput) error {
rec := recordPool.GetRecord()
if err := decodeWALRecord(b, rec); err != nil {
return err
}
// First process all series to ensure we don't write entries to nonexistant series.
var firstErr error
for _, s := range rec.Series {
if err := recoverer.SetStream(rec.UserID, s); err != nil {
if firstErr == nil {
firstErr = err
}
}
}
for _, entries := range rec.RefEntries {
worker := int(uint64(entries.Ref) % uint64(len(inputs)))
inputs[worker] <- recoveryInput{
userID: rec.UserID,
data: entries,
}
}
return firstErr
}
process := func(recoverer Recoverer, input <-chan recoveryInput, errCh chan<- error) {
for {
select {
case <-recoverer.Done():
case next, ok := <-input:
if !ok {
return
}
entries, ok := next.data.(RefEntries)
var err error
if !ok {
err = errors.Errorf("unexpected type (%T) when recovering WAL, expecting (%T)", next.data, entries)
}
if err == nil {
err = recoverer.Push(next.userID, entries)
}
// Pass the error back, but respect the quit signal.
if err != nil {
errCh <- err
}
}
}
}
return recoverGeneric(
reader,
recoverer,
dispatch,
process,
)
}
func RecoverCheckpoint(reader WALReader, recoverer Recoverer) error {
dispatch := func(recoverer Recoverer, b []byte, inputs []chan recoveryInput) error {
s := &Series{}
if err := decodeCheckpointRecord(b, s); err != nil {
return err
}
worker := int(s.Fingerprint % uint64(len(inputs)))
inputs[worker] <- recoveryInput{
userID: s.UserID,
data: s,
}
return nil
}
process := func(recoverer Recoverer, input <-chan recoveryInput, errCh chan<- error) {
for {
select {
case <-recoverer.Done():
case next, ok := <-input:
if !ok {
return
}
series, ok := next.data.(*Series)
var err error
if !ok {
err = errors.Errorf("unexpected type (%T) when recovering WAL, expecting (%T)", next.data, series)
}
if err == nil {
err = recoverer.Series(series)
}
if err != nil {
errCh <- err
}
}
}
}
return recoverGeneric(
reader,
recoverer,
dispatch,
process,
)
}
type recoveryInput struct {
userID string
data interface{}
}
// recoverGeneric enables reusing the ability to recover from WALs of different types
// by exposing the dispatch and process functions.
// Note: it explicitly does not call the Recoverer.Close function as it's possible to layer
// multiple recoveries on top of each other, as in the case of recovering from Checkpoints
// then the WAL.
func recoverGeneric(
reader WALReader,
recoverer Recoverer,
dispatch func(Recoverer, []byte, []chan recoveryInput) error,
process func(Recoverer, <-chan recoveryInput, chan<- error),
) error {
var wg sync.WaitGroup
var firstErr error
nWorkers := recoverer.NumWorkers()
if nWorkers < 1 {
return errors.New("cannot recover with no workers")
}
errCh := make(chan error)
inputs := make([]chan recoveryInput, 0, nWorkers)
wg.Add(nWorkers)
for i := 0; i < nWorkers; i++ {
inputs = append(inputs, make(chan recoveryInput))
go func(input <-chan recoveryInput) {
defer wg.Done()
process(recoverer, input, errCh)
}(inputs[i])
}
go func() {
for reader.Next() {
b := reader.Record()
if err := reader.Err(); err != nil {
errCh <- err
continue
}
if err := dispatch(recoverer, b, inputs); err != nil {
errCh <- err
continue
}
}
for _, w := range inputs {
close(w)
}
}()
finished := make(chan struct{})
go func(finished chan<- struct{}) {
wg.Wait()
finished <- struct{}{}
}(finished)
for {
select {
case <-finished:
return firstErr
case err := <-errCh:
if firstErr == nil {
firstErr = err
}
}
}
}