-
Notifications
You must be signed in to change notification settings - Fork 3.4k
/
accumulator.go
436 lines (368 loc) · 12.5 KB
/
accumulator.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
package logql
import (
"container/heap"
"context"
"fmt"
"sort"
"time"
"golang.org/x/exp/maps"
"github.com/grafana/loki/v3/pkg/logproto"
"github.com/grafana/loki/v3/pkg/logqlmodel"
"github.com/grafana/loki/v3/pkg/logqlmodel/metadata"
"github.com/grafana/loki/v3/pkg/logqlmodel/stats"
"github.com/grafana/loki/v3/pkg/querier/queryrange/queryrangebase/definitions"
"github.com/grafana/loki/v3/pkg/util/math"
)
// NewBufferedAccumulator returns an accumulator which aggregates all query
// results in a slice. This is useful for metric queries, which are generally
// small payloads and the memory overhead for buffering is negligible.
func NewBufferedAccumulator(n int) *BufferedAccumulator {
return &BufferedAccumulator{
results: make([]logqlmodel.Result, n),
}
}
type BufferedAccumulator struct {
results []logqlmodel.Result
}
func (a *BufferedAccumulator) Accumulate(_ context.Context, acc logqlmodel.Result, i int) error {
a.results[i] = acc
return nil
}
func (a *BufferedAccumulator) Result() []logqlmodel.Result {
return a.results
}
type QuantileSketchAccumulator struct {
matrix ProbabilisticQuantileMatrix
stats stats.Result // for accumulating statistics from downstream requests
headers map[string][]string // for accumulating headers from downstream requests
warnings map[string]struct{} // for accumulating warnings from downstream requests}
}
// newQuantileSketchAccumulator returns an accumulator for sharded
// probabilistic quantile queries that merges results as they come in.
func newQuantileSketchAccumulator() *QuantileSketchAccumulator {
return &QuantileSketchAccumulator{
headers: make(map[string][]string),
warnings: make(map[string]struct{}),
}
}
func (a *QuantileSketchAccumulator) Accumulate(_ context.Context, res logqlmodel.Result, _ int) error {
if res.Data.Type() != QuantileSketchMatrixType {
return fmt.Errorf("unexpected matrix data type: got (%s), want (%s)", res.Data.Type(), QuantileSketchMatrixType)
}
data, ok := res.Data.(ProbabilisticQuantileMatrix)
if !ok {
return fmt.Errorf("unexpected matrix type: got (%T), want (ProbabilisticQuantileMatrix)", res.Data)
}
// TODO(owen-d/ewelch): Shard counts should be set by the querier
// so we don't have to do it in tricky ways in multiple places.
// See pkg/logql/downstream.go:DownstreamEvaluator.Downstream
// for another example.
if res.Statistics.Summary.Shards == 0 {
res.Statistics.Summary.Shards = 1
}
a.stats.Merge(res.Statistics)
metadata.ExtendHeaders(a.headers, res.Headers)
for _, w := range res.Warnings {
a.warnings[w] = struct{}{}
}
if a.matrix == nil {
a.matrix = data
return nil
}
var err error
a.matrix, err = a.matrix.Merge(data)
a.stats.Merge(res.Statistics)
return err
}
func (a *QuantileSketchAccumulator) Result() []logqlmodel.Result {
headers := make([]*definitions.PrometheusResponseHeader, 0, len(a.headers))
for name, vals := range a.headers {
headers = append(
headers,
&definitions.PrometheusResponseHeader{
Name: name,
Values: vals,
},
)
}
warnings := maps.Keys(a.warnings)
sort.Strings(warnings)
return []logqlmodel.Result{
{
Data: a.matrix,
Headers: headers,
Warnings: warnings,
Statistics: a.stats,
},
}
}
// heap impl for keeping only the top n results across m streams
// importantly, AccumulatedStreams is _bounded_, so it will only
// store the top `limit` results across all streams.
// To implement this, we use a min-heap when looking
// for the max values (logproto.FORWARD)
// and vice versa for logproto.BACKWARD.
// This allows us to easily find the 'worst' value
// and replace it with a better one.
// Once we've fully processed all log lines,
// we return the heap in opposite order and then reverse it
// to get the correct order.
// Heap implements container/heap.Interface
// solely to use heap.Interface as a library.
// It is not intended for the heap pkg functions
// to otherwise call this type.
type AccumulatedStreams struct {
count, limit int
labelmap map[string]int
streams []*logproto.Stream
order logproto.Direction
stats stats.Result // for accumulating statistics from downstream requests
headers map[string][]string // for accumulating headers from downstream requests
warnings map[string]struct{} // for accumulating warnings from downstream requests
}
// NewStreamAccumulator returns an accumulator for limited log queries.
// Log queries, sharded thousands of times and each returning <limit>
// results, can be _considerably_ larger. In this case, we eagerly
// accumulate the results into a logsAccumulator, discarding values
// over the limit to keep memory pressure down while other subqueries
// are executing.
func NewStreamAccumulator(params Params) *AccumulatedStreams {
// the stream accumulator stores a heap with reversed order
// from the results we expect, so we need to reverse the direction
order := logproto.FORWARD
if params.Direction() == logproto.FORWARD {
order = logproto.BACKWARD
}
return &AccumulatedStreams{
labelmap: make(map[string]int),
order: order,
limit: int(params.Limit()),
headers: make(map[string][]string),
warnings: make(map[string]struct{}),
}
}
// returns the top priority
func (acc *AccumulatedStreams) top() (time.Time, bool) {
if len(acc.streams) > 0 && len(acc.streams[0].Entries) > 0 {
return acc.streams[0].Entries[len(acc.streams[0].Entries)-1].Timestamp, true
}
return time.Time{}, false
}
func (acc *AccumulatedStreams) Find(labels string) (int, bool) {
i, ok := acc.labelmap[labels]
return i, ok
}
// number of streams
func (acc *AccumulatedStreams) Len() int { return len(acc.streams) }
func (acc *AccumulatedStreams) Swap(i, j int) {
// for i=0, j=1
// {'a': 0, 'b': 1}
// [a, b]
acc.streams[i], acc.streams[j] = acc.streams[j], acc.streams[i]
// {'a': 0, 'b': 1}
// [b, a]
acc.labelmap[acc.streams[i].Labels] = i
acc.labelmap[acc.streams[j].Labels] = j
// {'a': 1, 'b': 0}
// [b, a]
}
// first order by timestamp, then by labels
func (acc *AccumulatedStreams) Less(i, j int) bool {
// order by the 'oldest' entry in the stream
if a, b := acc.streams[i].Entries[len(acc.streams[i].Entries)-1].Timestamp, acc.streams[j].Entries[len(acc.streams[j].Entries)-1].Timestamp; !a.Equal(b) {
return acc.less(a, b)
}
return acc.streams[i].Labels <= acc.streams[j].Labels
}
func (acc *AccumulatedStreams) less(a, b time.Time) bool {
// use after for stable sort
if acc.order == logproto.FORWARD {
return !a.After(b)
}
return !b.After(a)
}
func (acc *AccumulatedStreams) Push(x any) {
s := x.(*logproto.Stream)
if len(s.Entries) == 0 {
return
}
if room := acc.limit - acc.count; room >= len(s.Entries) {
if i, ok := acc.Find(s.Labels); ok {
// stream already exists, append entries
// these are already guaranteed to be sorted
// Reasoning: we shard subrequests so each stream exists on only one
// shard. Therefore, the only time a stream should already exist
// is in successive splits, which are already guaranteed to be ordered
// and we can just append.
acc.appendTo(acc.streams[i], s)
return
}
// new stream
acc.addStream(s)
return
}
// there's not enough room for all the entries,
// so we need to
acc.push(s)
}
// there's not enough room for all the entries.
// since we store them in a reverse heap relative to what we _want_
// (i.e. the max value for FORWARD, the min value for BACKWARD),
// we test if the new entry is better than the worst entry,
// swapping them if so.
func (acc *AccumulatedStreams) push(s *logproto.Stream) {
worst, ok := acc.top()
room := math.Min(acc.limit-acc.count, len(s.Entries))
if !ok {
if room == 0 {
// special case: limit must be zero since there's no room and no worst entry
return
}
s.Entries = s.Entries[:room]
// special case: there are no entries in the heap. Push entries up to the limit
acc.addStream(s)
return
}
// since entries are sorted by timestamp from best -> worst,
// we can discard the entire stream if the incoming best entry
// is worse than the worst entry in the heap.
cutoff := sort.Search(len(s.Entries), func(i int) bool {
// TODO(refactor label comparison -- should be in another fn)
if worst.Equal(s.Entries[i].Timestamp) {
return acc.streams[0].Labels < s.Labels
}
return acc.less(s.Entries[i].Timestamp, worst)
})
s.Entries = s.Entries[:cutoff]
for i := 0; i < len(s.Entries) && acc.less(worst, s.Entries[i].Timestamp); i++ {
// push one entry at a time
room = acc.limit - acc.count
// pop if there's no room to make the heap small enough for an append;
// in the short path of Push() we know that there's room for at least one entry
if room == 0 {
acc.Pop()
}
cpy := *s
cpy.Entries = []logproto.Entry{s.Entries[i]}
acc.Push(&cpy)
// update worst
worst, _ = acc.top()
}
}
func (acc *AccumulatedStreams) addStream(s *logproto.Stream) {
// ensure entries conform to order we expect
// TODO(owen-d): remove? should be unnecessary since we insert in appropriate order
// but it's nice to have the safeguard
sort.Slice(s.Entries, func(i, j int) bool {
return acc.less(s.Entries[j].Timestamp, s.Entries[i].Timestamp)
})
acc.streams = append(acc.streams, s)
i := len(acc.streams) - 1
acc.labelmap[s.Labels] = i
acc.count += len(s.Entries)
heap.Fix(acc, i)
}
// dst must already exist in acc
func (acc *AccumulatedStreams) appendTo(dst, src *logproto.Stream) {
// these are already guaranteed to be sorted
// Reasoning: we shard subrequests so each stream exists on only one
// shard. Therefore, the only time a stream should already exist
// is in successive splits, which are already guaranteed to be ordered
// and we can just append.
var needsSort bool
for _, e := range src.Entries {
// sort if order has broken
if len(dst.Entries) > 0 && acc.less(dst.Entries[len(dst.Entries)-1].Timestamp, e.Timestamp) {
needsSort = true
}
dst.Entries = append(dst.Entries, e)
}
if needsSort {
sort.Slice(dst.Entries, func(i, j int) bool {
// store in reverse order so we can more reliably insert without sorting and pop from end
return acc.less(dst.Entries[j].Timestamp, dst.Entries[i].Timestamp)
})
}
acc.count += len(src.Entries)
heap.Fix(acc, acc.labelmap[dst.Labels])
}
// Pop returns a stream with one entry. It pops the first entry of the first stream
func (acc *AccumulatedStreams) Pop() any {
n := acc.Len()
if n == 0 {
return nil
}
stream := acc.streams[0]
cpy := *stream
cpy.Entries = []logproto.Entry{cpy.Entries[len(stream.Entries)-1]}
stream.Entries = stream.Entries[:len(stream.Entries)-1]
acc.count--
if len(stream.Entries) == 0 {
// remove stream
acc.Swap(0, n-1)
acc.streams[n-1] = nil // avoid leaking reference
delete(acc.labelmap, stream.Labels)
acc.streams = acc.streams[:n-1]
}
if acc.Len() > 0 {
heap.Fix(acc, 0)
}
return &cpy
}
// Note: can only be called once as it will alter stream ordreing.
func (acc *AccumulatedStreams) Result() []logqlmodel.Result {
// sort streams by label
sort.Slice(acc.streams, func(i, j int) bool {
return acc.streams[i].Labels < acc.streams[j].Labels
})
streams := make(logqlmodel.Streams, 0, len(acc.streams))
for _, s := range acc.streams {
// sort entries by timestamp, inversely based on direction
sort.Slice(s.Entries, func(i, j int) bool {
return acc.less(s.Entries[j].Timestamp, s.Entries[i].Timestamp)
})
streams = append(streams, *s)
}
res := logqlmodel.Result{
// stats & headers are already aggregated in the context
Data: streams,
Statistics: acc.stats,
Headers: make([]*definitions.PrometheusResponseHeader, 0, len(acc.headers)),
}
for name, vals := range acc.headers {
res.Headers = append(
res.Headers,
&definitions.PrometheusResponseHeader{
Name: name,
Values: vals,
},
)
}
warnings := maps.Keys(acc.warnings)
sort.Strings(warnings)
res.Warnings = warnings
return []logqlmodel.Result{res}
}
func (acc *AccumulatedStreams) Accumulate(_ context.Context, x logqlmodel.Result, _ int) error {
// TODO(owen-d/ewelch): Shard counts should be set by the querier
// so we don't have to do it in tricky ways in multiple places.
// See pkg/logql/downstream.go:DownstreamEvaluator.Downstream
// for another example.
if x.Statistics.Summary.Shards == 0 {
x.Statistics.Summary.Shards = 1
}
acc.stats.Merge(x.Statistics)
metadata.ExtendHeaders(acc.headers, x.Headers)
for _, w := range x.Warnings {
acc.warnings[w] = struct{}{}
}
switch got := x.Data.(type) {
case logqlmodel.Streams:
for i := range got {
acc.Push(&got[i])
}
default:
return fmt.Errorf("unexpected response type during response result accumulation. Got (%T), wanted %s", got, logqlmodel.ValueTypeStreams)
}
return nil
}