| 1 |
// Copyright The OpenTelemetry Authors |
| 2 |
// SPDX-License-Identifier: Apache-2.0 |
| 3 |
|
| 4 |
package aggregate // import "go.opentelemetry.io/otel/sdk/metric/internal/aggregate" |
| 5 |
|
| 6 |
import ( |
| 7 |
"context" |
| 8 |
"slices" |
| 9 |
"sort" |
| 10 |
"sync/atomic" |
| 11 |
"time" |
| 12 |
|
| 13 |
"go.opentelemetry.io/otel/attribute" |
| 14 |
"go.opentelemetry.io/otel/sdk/internal/x" |
| 15 |
"go.opentelemetry.io/otel/sdk/metric/metricdata" |
| 16 |
) |
| 17 |
|
| 18 |
// histogramPoint is a single histogram point, used in delta aggregations. |
| 19 |
type histogramPoint[N int64 | float64] struct { |
| 20 |
attrs attribute.Set |
| 21 |
res FilteredExemplarReservoir[N] |
| 22 |
histogramPointCounters[N] |
| 23 |
} |
| 24 |
|
| 25 |
// hotColdHistogramPoint a hot and cold histogram points, used in cumulative |
| 26 |
// aggregations. |
| 27 |
type hotColdHistogramPoint[N int64 | float64] struct { |
| 28 |
hcwg hotColdWaitGroup |
| 29 |
hotColdPoint [2]histogramPointCounters[N] |
| 30 |
|
| 31 |
attrs attribute.Set |
| 32 |
res FilteredExemplarReservoir[N] |
| 33 |
startTime time.Time |
| 34 |
} |
| 35 |
|
| 36 |
// histogramPointCounters contains only the atomic counter data, and is used by |
| 37 |
// both histogramPoint and hotColdHistogramPoint. |
| 38 |
type histogramPointCounters[N int64 | float64] struct { |
| 39 |
counts []atomic.Uint64 |
| 40 |
total atomicCounter[N] |
| 41 |
minMax atomicMinMax[N] |
| 42 |
} |
| 43 |
|
| 44 |
func (b *histogramPointCounters[N]) loadCountsInto(into *[]uint64) uint64 { |
| 45 |
// TODO (#3047): Making copies for counts incurs a large |
| 46 |
// memory allocation footprint. Alternatives should be explored. |
| 47 |
counts := reset(*into, len(b.counts), len(b.counts)) |
| 48 |
count := uint64(0) |
| 49 |
for i := range b.counts { |
| 50 |
c := b.counts[i].Load() |
| 51 |
counts[i] = c |
| 52 |
count += c |
| 53 |
} |
| 54 |
*into = counts |
| 55 |
return count |
| 56 |
} |
| 57 |
|
| 58 |
// mergeIntoAndReset merges this set of histogram counter data into another, |
| 59 |
// and resets the state of this set of counters. This is used by |
| 60 |
// hotColdHistogramPoint to ensure that the cumulative counters continue to |
| 61 |
// accumulate after being read. |
| 62 |
func (b *histogramPointCounters[N]) mergeIntoAndReset( // nolint:revive // Intentional internal control flag |
| 63 |
into *histogramPointCounters[N], |
| 64 |
noMinMax, noSum bool, |
| 65 |
) { |
| 66 |
for i := range b.counts { |
| 67 |
into.counts[i].Add(b.counts[i].Load()) |
| 68 |
b.counts[i].Store(0) |
| 69 |
} |
| 70 |
|
| 71 |
if !noMinMax { |
| 72 |
// Do not reset min or max because cumulative min and max only ever grow |
| 73 |
// smaller or larger respectively. |
| 74 |
|
| 75 |
if b.minMax.set.Load() { |
| 76 |
into.minMax.Update(b.minMax.minimum.Load()) |
| 77 |
into.minMax.Update(b.minMax.maximum.Load()) |
| 78 |
} |
| 79 |
} |
| 80 |
if !noSum { |
| 81 |
into.total.add(b.total.load()) |
| 82 |
b.total.reset() |
| 83 |
} |
| 84 |
} |
| 85 |
|
| 86 |
// deltaHistogram is a histogram whose internal storage is reset when it is |
| 87 |
// collected. |
| 88 |
// |
| 89 |
// deltaHistogram's measure is implemented without locking, even when called |
| 90 |
// concurrently with collect. This is done by maintaining two separate maps: |
| 91 |
// one "hot" which is concurrently updated by measure(), and one "cold", which |
| 92 |
// is read and reset by collect(). The [hotcoldWaitGroup] allows collect() to |
| 93 |
// swap the hot and cold maps, and wait for updates to the cold map to complete |
| 94 |
// prior to reading. deltaHistogram swaps ald clears complete maps so that |
| 95 |
// unused attribute sets do not report in subsequent collect() calls. |
| 96 |
type deltaHistogram[N int64 | float64] struct { |
| 97 |
hcwg hotColdWaitGroup |
| 98 |
hotColdValMap [2]limitedSyncMap |
| 99 |
|
| 100 |
start time.Time |
| 101 |
noMinMax bool |
| 102 |
noSum bool |
| 103 |
bounds []float64 |
| 104 |
newRes func(attribute.Set) FilteredExemplarReservoir[N] |
| 105 |
} |
| 106 |
|
| 107 |
func (s *deltaHistogram[N]) measure( |
| 108 |
ctx context.Context, |
| 109 |
value N, |
| 110 |
fltrAttr attribute.Set, |
| 111 |
droppedAttr []attribute.KeyValue, |
| 112 |
) { |
| 113 |
hotIdx := s.hcwg.start() |
| 114 |
defer s.hcwg.done(hotIdx) |
| 115 |
h := s.hotColdValMap[hotIdx].LoadOrStoreAttr(fltrAttr, func(attr attribute.Set) any { |
| 116 |
hPt := &histogramPoint[N]{ |
| 117 |
res: s.newRes(attr), |
| 118 |
attrs: attr, |
| 119 |
// N+1 buckets. For example: |
| 120 |
// |
| 121 |
// bounds = [0, 5, 10] |
| 122 |
// |
| 123 |
// Then, |
| 124 |
// |
| 125 |
// counts = (-∞, 0], (0, 5.0], (5.0, 10.0], (10.0, +∞) |
| 126 |
histogramPointCounters: histogramPointCounters[N]{counts: make([]atomic.Uint64, len(s.bounds)+1)}, |
| 127 |
} |
| 128 |
return hPt |
| 129 |
}).(*histogramPoint[N]) |
| 130 |
|
| 131 |
// This search will return an index in the range [0, len(s.bounds)], where |
| 132 |
// it will return len(s.bounds) if value is greater than the last element |
| 133 |
// of s.bounds. This aligns with the histogramPoint in that the length of histogramPoint |
| 134 |
// is len(s.bounds)+1, with the last bucket representing: |
| 135 |
// (s.bounds[len(s.bounds)-1], +∞). |
| 136 |
idx := sort.SearchFloat64s(s.bounds, float64(value)) |
| 137 |
h.counts[idx].Add(1) |
| 138 |
if !s.noMinMax { |
| 139 |
h.minMax.Update(value) |
| 140 |
} |
| 141 |
if !s.noSum { |
| 142 |
h.total.add(value) |
| 143 |
} |
| 144 |
h.res.Offer(ctx, value, droppedAttr) |
| 145 |
} |
| 146 |
|
| 147 |
// newDeltaHistogram returns a histogram that is reset each time it is |
| 148 |
// collected. |
| 149 |
func newDeltaHistogram[N int64 | float64]( |
| 150 |
boundaries []float64, |
| 151 |
noMinMax, noSum bool, |
| 152 |
limit int, |
| 153 |
r func(attribute.Set) FilteredExemplarReservoir[N], |
| 154 |
) *deltaHistogram[N] { |
| 155 |
// The responsibility of keeping all histogramPoint correctly associated with the |
| 156 |
// passed boundaries is ultimately this type's responsibility. Make a copy |
| 157 |
// here so we can always guarantee this. Or, in the case of failure, have |
| 158 |
// complete control over the fix. |
| 159 |
b := slices.Clone(boundaries) |
| 160 |
slices.Sort(b) |
| 161 |
return &deltaHistogram[N]{ |
| 162 |
start: now(), |
| 163 |
noMinMax: noMinMax, |
| 164 |
noSum: noSum, |
| 165 |
bounds: b, |
| 166 |
newRes: r, |
| 167 |
hotColdValMap: [2]limitedSyncMap{ |
| 168 |
{aggLimit: limit}, |
| 169 |
{aggLimit: limit}, |
| 170 |
}, |
| 171 |
} |
| 172 |
} |
| 173 |
|
| 174 |
func (s *deltaHistogram[N]) collect( |
| 175 |
dest *metricdata.Aggregation, //nolint:gocritic // The pointer is needed for the ComputeAggregation interface |
| 176 |
) int { |
| 177 |
t := now() |
| 178 |
|
| 179 |
// If *dest is not a metricdata.Histogram, memory reuse is missed. In that |
| 180 |
// case, use the zero-value h and hope for better alignment next cycle. |
| 181 |
h, _ := (*dest).(metricdata.Histogram[N]) |
| 182 |
h.Temporality = metricdata.DeltaTemporality |
| 183 |
|
| 184 |
// delta always clears values on collection |
| 185 |
readIdx := s.hcwg.swapHotAndWait() |
| 186 |
|
| 187 |
// Do not allow modification of our copy of bounds. |
| 188 |
bounds := slices.Clone(s.bounds) |
| 189 |
|
| 190 |
// The len will not change while we iterate over values, since we waited |
| 191 |
// for all writes to finish to the cold values and len. |
| 192 |
n := s.hotColdValMap[readIdx].Len() |
| 193 |
hDPts := reset(h.DataPoints, n, n) |
| 194 |
|
| 195 |
var i int |
| 196 |
s.hotColdValMap[readIdx].Range(func(_, value any) bool { |
| 197 |
val := value.(*histogramPoint[N]) |
| 198 |
|
| 199 |
count := val.loadCountsInto(&hDPts[i].BucketCounts) |
| 200 |
hDPts[i].Attributes = val.attrs |
| 201 |
hDPts[i].StartTime = s.start |
| 202 |
hDPts[i].Time = t |
| 203 |
hDPts[i].Count = count |
| 204 |
hDPts[i].Bounds = bounds |
| 205 |
|
| 206 |
if !s.noSum { |
| 207 |
hDPts[i].Sum = val.total.load() |
| 208 |
} |
| 209 |
|
| 210 |
if !s.noMinMax { |
| 211 |
if val.minMax.set.Load() { |
| 212 |
hDPts[i].Min = metricdata.NewExtrema(val.minMax.minimum.Load()) |
| 213 |
hDPts[i].Max = metricdata.NewExtrema(val.minMax.maximum.Load()) |
| 214 |
} |
| 215 |
} |
| 216 |
|
| 217 |
collectExemplars(&hDPts[i].Exemplars, val.res.Collect) |
| 218 |
|
| 219 |
i++ |
| 220 |
return true |
| 221 |
}) |
| 222 |
// Unused attribute sets do not report. |
| 223 |
s.hotColdValMap[readIdx].Clear() |
| 224 |
// The delta collection cycle resets. |
| 225 |
s.start = t |
| 226 |
|
| 227 |
h.DataPoints = hDPts |
| 228 |
*dest = h |
| 229 |
|
| 230 |
return n |
| 231 |
} |
| 232 |
|
| 233 |
// cumulativeHistogram summarizes a set of measurements as an histogram with explicitly |
| 234 |
// defined histogramPoint. |
| 235 |
// |
| 236 |
// cumulativeHistogram's measure is implemented without locking, even when |
| 237 |
// called concurrently with collect. This is done by maintaining two separate |
| 238 |
// histogramPointCounters for each attribute set: one "hot" which is |
| 239 |
// concurrently updated by measure(), and one "cold", which is read and reset |
| 240 |
// by collect(). The [hotcoldWaitGroup] allows collect() to swap the hot and |
| 241 |
// cold counters, and wait for updates to the cold counters to complete prior |
| 242 |
// to reading. Unlike deltaHistogram, this maintains a single map so that the |
| 243 |
// preserved attribute sets do not change when collect() is called. |
| 244 |
type cumulativeHistogram[N int64 | float64] struct { |
| 245 |
values limitedSyncMap |
| 246 |
|
| 247 |
start time.Time |
| 248 |
noMinMax bool |
| 249 |
noSum bool |
| 250 |
bounds []float64 |
| 251 |
newRes func(attribute.Set) FilteredExemplarReservoir[N] |
| 252 |
} |
| 253 |
|
| 254 |
// newCumulativeHistogram returns a histogram that accumulates measurements |
| 255 |
// into a histogram data structure. It is never reset. |
| 256 |
func newCumulativeHistogram[N int64 | float64]( |
| 257 |
boundaries []float64, |
| 258 |
noMinMax, noSum bool, |
| 259 |
limit int, |
| 260 |
r func(attribute.Set) FilteredExemplarReservoir[N], |
| 261 |
) *cumulativeHistogram[N] { |
| 262 |
// The responsibility of keeping all histogramPoint correctly associated with the |
| 263 |
// passed boundaries is ultimately this type's responsibility. Make a copy |
| 264 |
// here so we can always guarantee this. Or, in the case of failure, have |
| 265 |
// complete control over the fix. |
| 266 |
b := slices.Clone(boundaries) |
| 267 |
slices.Sort(b) |
| 268 |
return &cumulativeHistogram[N]{ |
| 269 |
start: now(), |
| 270 |
noMinMax: noMinMax, |
| 271 |
noSum: noSum, |
| 272 |
bounds: b, |
| 273 |
newRes: r, |
| 274 |
values: limitedSyncMap{aggLimit: limit}, |
| 275 |
} |
| 276 |
} |
| 277 |
|
| 278 |
func (s *cumulativeHistogram[N]) measure( |
| 279 |
ctx context.Context, |
| 280 |
value N, |
| 281 |
fltrAttr attribute.Set, |
| 282 |
droppedAttr []attribute.KeyValue, |
| 283 |
) { |
| 284 |
h := s.values.LoadOrStoreAttr(fltrAttr, func(attr attribute.Set) any { |
| 285 |
hPt := &hotColdHistogramPoint[N]{ |
| 286 |
res: s.newRes(attr), |
| 287 |
attrs: attr, |
| 288 |
// N+1 buckets. For example: |
| 289 |
// |
| 290 |
// bounds = [0, 5, 10] |
| 291 |
// |
| 292 |
// Then, |
| 293 |
// |
| 294 |
// count = (-∞, 0], (0, 5.0], (5.0, 10.0], (10.0, +∞) |
| 295 |
hotColdPoint: [2]histogramPointCounters[N]{ |
| 296 |
{ |
| 297 |
counts: make([]atomic.Uint64, len(s.bounds)+1), |
| 298 |
}, |
| 299 |
{ |
| 300 |
counts: make([]atomic.Uint64, len(s.bounds)+1), |
| 301 |
}, |
| 302 |
}, |
| 303 |
startTime: now(), |
| 304 |
} |
| 305 |
return hPt |
| 306 |
}).(*hotColdHistogramPoint[N]) |
| 307 |
|
| 308 |
// This search will return an index in the range [0, len(s.bounds)], where |
| 309 |
// it will return len(s.bounds) if value is greater than the last element |
| 310 |
// of s.bounds. This aligns with the histogramPoint in that the length of histogramPoint |
| 311 |
// is len(s.bounds)+1, with the last bucket representing: |
| 312 |
// (s.bounds[len(s.bounds)-1], +∞). |
| 313 |
idx := sort.SearchFloat64s(s.bounds, float64(value)) |
| 314 |
|
| 315 |
hotIdx := h.hcwg.start() |
| 316 |
defer h.hcwg.done(hotIdx) |
| 317 |
|
| 318 |
h.hotColdPoint[hotIdx].counts[idx].Add(1) |
| 319 |
if !s.noMinMax { |
| 320 |
h.hotColdPoint[hotIdx].minMax.Update(value) |
| 321 |
} |
| 322 |
if !s.noSum { |
| 323 |
h.hotColdPoint[hotIdx].total.add(value) |
| 324 |
} |
| 325 |
h.res.Offer(ctx, value, droppedAttr) |
| 326 |
} |
| 327 |
|
| 328 |
func (s *cumulativeHistogram[N]) collect( |
| 329 |
dest *metricdata.Aggregation, //nolint:gocritic // The pointer is needed for the ComputeAggregation interface |
| 330 |
) int { |
| 331 |
t := now() |
| 332 |
|
| 333 |
// If *dest is not a metricdata.Histogram, memory reuse is missed. In that |
| 334 |
// case, use the zero-value h and hope for better alignment next cycle. |
| 335 |
h, _ := (*dest).(metricdata.Histogram[N]) |
| 336 |
h.Temporality = metricdata.CumulativeTemporality |
| 337 |
|
| 338 |
// Do not allow modification of our copy of bounds. |
| 339 |
bounds := slices.Clone(s.bounds) |
| 340 |
|
| 341 |
// Values are being concurrently written while we iterate, so only use the |
| 342 |
// current length for capacity. |
| 343 |
hDPts := reset(h.DataPoints, 0, s.values.Len()) |
| 344 |
|
| 345 |
perSeriesStartTimeEnabled := x.PerSeriesStartTimestamps.Enabled() |
| 346 |
|
| 347 |
var i int |
| 348 |
s.values.Range(func(_, value any) bool { |
| 349 |
val := value.(*hotColdHistogramPoint[N]) |
| 350 |
|
| 351 |
startTime := s.start |
| 352 |
if perSeriesStartTimeEnabled { |
| 353 |
startTime = val.startTime |
| 354 |
} |
| 355 |
// swap, observe, and clear the point |
| 356 |
readIdx := val.hcwg.swapHotAndWait() |
| 357 |
var bucketCounts []uint64 |
| 358 |
count := val.hotColdPoint[readIdx].loadCountsInto(&bucketCounts) |
| 359 |
newPt := metricdata.HistogramDataPoint[N]{ |
| 360 |
Attributes: val.attrs, |
| 361 |
StartTime: startTime, |
| 362 |
Time: t, |
| 363 |
Count: count, |
| 364 |
Bounds: bounds, |
| 365 |
// The HistogramDataPoint field values returned need to be copies of |
| 366 |
// the histogramPoint value as we will keep updating them. |
| 367 |
BucketCounts: bucketCounts, |
| 368 |
} |
| 369 |
|
| 370 |
if !s.noSum { |
| 371 |
newPt.Sum = val.hotColdPoint[readIdx].total.load() |
| 372 |
} |
| 373 |
if !s.noMinMax { |
| 374 |
if val.hotColdPoint[readIdx].minMax.set.Load() { |
| 375 |
newPt.Min = metricdata.NewExtrema(val.hotColdPoint[readIdx].minMax.minimum.Load()) |
| 376 |
newPt.Max = metricdata.NewExtrema(val.hotColdPoint[readIdx].minMax.maximum.Load()) |
| 377 |
} |
| 378 |
} |
| 379 |
// Once we've read the point, merge it back into the hot histogram |
| 380 |
// point since it is cumulative. |
| 381 |
hotIdx := (readIdx + 1) % 2 |
| 382 |
val.hotColdPoint[readIdx].mergeIntoAndReset(&val.hotColdPoint[hotIdx], s.noMinMax, s.noSum) |
| 383 |
|
| 384 |
collectExemplars(&newPt.Exemplars, val.res.Collect) |
| 385 |
hDPts = append(hDPts, newPt) |
| 386 |
|
| 387 |
i++ |
| 388 |
// TODO (#3006): This will use an unbounded amount of memory if there |
| 389 |
// are unbounded number of attribute sets being aggregated. Attribute |
| 390 |
// sets that become "stale" need to be forgotten so this will not |
| 391 |
// overload the system. |
| 392 |
return true |
| 393 |
}) |
| 394 |
|
| 395 |
h.DataPoints = hDPts |
| 396 |
*dest = h |
| 397 |
|
| 398 |
return i |
| 399 |
} |