1
0
mirror of https://github.com/open-telemetry/opentelemetry-go.git synced 2025-01-26 03:52:03 +02:00
David Ashpole 0485de287e
Move time.Now call into exemplar reservoir to improve performance (#5545)
Part of addressing
https://github.com/open-telemetry/opentelemetry-go/issues/5542.

### Motivation

This removes the `time.Now()` call from filtered-out Exemplars by only
invoking `time.Now()` after the filtering decision is made. This
improvement is especially noticeable for measurements without any
attributes.

```
goos: linux
goarch: amd64
pkg: go.opentelemetry.io/otel/sdk/metric
cpu: AMD EPYC 7B12
                                                                 │   old.txt    │                new.txt                │
                                                                 │    sec/op    │   sec/op     vs base                  │
SyncMeasure/NoView/Int64Counter/Attributes/0-24                    158.20n ± 4%   99.83n ± 1%  -36.90% (p=0.000 n=10)
SyncMeasure/NoView/Int64Counter/Attributes/1-24                     333.3n ± 4%   274.8n ± 1%  -17.55% (p=0.000 n=10)
SyncMeasure/NoView/Int64Counter/Attributes/10-24                    1.640µ ± 1%   1.600µ ± 1%   -2.41% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/0-24                   159.0n ± 3%   101.3n ± 0%  -36.27% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/1-24                   340.0n ± 2%   272.0n ± 1%  -20.00% (p=0.000 n=10)
SyncMeasure/NoView/Float64Counter/Attributes/10-24                  1.661µ ± 1%   1.597µ ± 0%   -3.85% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/0-24               159.8n ± 1%   103.1n ± 0%  -35.50% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/1-24               339.5n ± 1%   273.1n ± 0%  -19.57% (p=0.000 n=10)
SyncMeasure/NoView/Int64UpDownCounter/Attributes/10-24              1.656µ ± 0%   1.589µ ± 0%   -4.05% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/0-24             159.3n ± 2%   100.8n ± 0%  -36.74% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/1-24             337.9n ± 2%   271.8n ± 1%  -19.55% (p=0.000 n=10)
SyncMeasure/NoView/Float64UpDownCounter/Attributes/10-24            1.657µ ± 0%   1.593µ ± 1%   -3.83% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/0-24                  144.65n ± 4%   89.38n ± 0%  -38.21% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/1-24                   235.7n ± 2%   183.5n ± 0%  -22.15% (p=0.000 n=10)
SyncMeasure/NoView/Int64Histogram/Attributes/10-24                  900.8n ± 1%   836.8n ± 0%   -7.10% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/0-24                145.60n ± 5%   93.48n ± 1%  -35.80% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/1-24                 240.9n ± 1%   183.0n ± 0%  -24.06% (p=0.000 n=10)
SyncMeasure/NoView/Float64Histogram/Attributes/10-24                905.6n ± 1%   826.3n ± 0%   -8.76% (p=0.000 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/0-24                   20.33n ± 0%   20.35n ± 0%        ~ (p=0.302 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/1-24                   26.46n ± 0%   26.45n ± 1%        ~ (p=0.868 n=10)
SyncMeasure/DropView/Int64Counter/Attributes/10-24                  26.50n ± 0%   26.47n ± 0%        ~ (p=0.208 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/0-24                 20.34n ± 1%   20.27n ± 0%   -0.34% (p=0.009 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/1-24                 26.55n ± 0%   26.60n ± 1%        ~ (p=0.109 n=10)
SyncMeasure/DropView/Float64Counter/Attributes/10-24                26.59n ± 1%   26.57n ± 1%        ~ (p=0.926 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/0-24             20.38n ± 1%   20.38n ± 0%        ~ (p=0.725 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/1-24             26.39n ± 0%   26.44n ± 0%        ~ (p=0.238 n=10)
SyncMeasure/DropView/Int64UpDownCounter/Attributes/10-24            26.52n ± 0%   26.42n ± 0%   -0.36% (p=0.049 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/0-24           20.30n ± 0%   20.25n ± 0%        ~ (p=0.196 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/1-24           26.57n ± 0%   26.54n ± 1%        ~ (p=0.540 n=10)
SyncMeasure/DropView/Float64UpDownCounter/Attributes/10-24          26.57n ± 0%   26.51n ± 1%        ~ (p=0.643 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/0-24                 20.37n ± 0%   20.36n ± 1%        ~ (p=1.000 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/1-24                 26.41n ± 0%   26.50n ± 0%   +0.32% (p=0.007 n=10)
SyncMeasure/DropView/Int64Histogram/Attributes/10-24                26.44n ± 0%   26.55n ± 1%   +0.42% (p=0.012 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/0-24               20.30n ± 0%   20.45n ± 0%   +0.74% (p=0.000 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/1-24               26.52n ± 0%   26.48n ± 0%        ~ (p=0.127 n=10)
SyncMeasure/DropView/Float64Histogram/Attributes/10-24              26.55n ± 0%   26.48n ± 0%   -0.26% (p=0.002 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/0-24             170.5n ± 2%   110.8n ± 0%  -35.03% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/1-24             402.5n ± 1%   331.5n ± 1%  -17.64% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Counter/Attributes/10-24            1.363µ ± 1%   1.281µ ± 1%   -6.02% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/0-24           170.6n ± 1%   111.5n ± 1%  -34.64% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/1-24           397.1n ± 1%   335.9n ± 0%  -15.41% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64Counter/Attributes/10-24          1.371µ ± 1%   1.279µ ± 1%   -6.71% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/0-24       170.1n ± 1%   112.2n ± 0%  -34.09% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/1-24       397.5n ± 1%   330.2n ± 0%  -16.93% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64UpDownCounter/Attributes/10-24      1.371µ ± 1%   1.289µ ± 1%   -5.95% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/0-24     171.4n ± 2%   112.9n ± 0%  -34.13% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/1-24     397.0n ± 3%   336.4n ± 0%  -15.24% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Float64UpDownCounter/Attributes/10-24    1.383µ ± 1%   1.305µ ± 1%   -5.61% (p=0.000 n=10)
SyncMeasure/AttrFilterView/Int64Histogram/Attributes/0-24          157.30n ± 2%   98.58n ± 1%  -37.33% (p=0.000 n=6+10)
```

### Changes

* Introduce `exemplar.Filter`, which is a filter function based on the
context. It will not be user-facing, so we can always add other
parameters later if needed.
* Introduce `exemplar.FilteredReservoir`, which is similar to a
reservoir, except it does not receive a timestamp. It gets the current
time after the filter decision has been made. It uses generics to avoid
the call to exemplar.NewValue(), since it is internal-only.
* The `exemplar.Reservoir` is left as-is, so that it can be made public
when exemplars are stable. It still includes a timestamp argument.
* Unit tests are updated to expect a much lower number of calls to
time.Now
* `exemplar.Drop` is now an `exemplar.FilteredReservoir` instead of a
`Reservoir`, since we don't need a Reservoir to store things in if the
measurement is always dropped.

Co-authored-by: Sam Xie <sam@samxie.me>
Co-authored-by: Tyler Yahn <MrAlias@users.noreply.github.com>
2024-07-01 09:36:11 -07:00

163 lines
4.2 KiB
Go

// Copyright The OpenTelemetry Authors
// SPDX-License-Identifier: Apache-2.0
package aggregate // import "go.opentelemetry.io/otel/sdk/metric/internal/aggregate"
import (
"context"
"sync"
"time"
"go.opentelemetry.io/otel/attribute"
"go.opentelemetry.io/otel/sdk/metric/internal/exemplar"
"go.opentelemetry.io/otel/sdk/metric/metricdata"
)
// datapoint is timestamped measurement data.
type datapoint[N int64 | float64] struct {
attrs attribute.Set
value N
res exemplar.FilteredReservoir[N]
}
func newLastValue[N int64 | float64](limit int, r func() exemplar.FilteredReservoir[N]) *lastValue[N] {
return &lastValue[N]{
newRes: r,
limit: newLimiter[datapoint[N]](limit),
values: make(map[attribute.Distinct]datapoint[N]),
start: now(),
}
}
// lastValue summarizes a set of measurements as the last one made.
type lastValue[N int64 | float64] struct {
sync.Mutex
newRes func() exemplar.FilteredReservoir[N]
limit limiter[datapoint[N]]
values map[attribute.Distinct]datapoint[N]
start time.Time
}
func (s *lastValue[N]) measure(ctx context.Context, value N, fltrAttr attribute.Set, droppedAttr []attribute.KeyValue) {
s.Lock()
defer s.Unlock()
attr := s.limit.Attributes(fltrAttr, s.values)
d, ok := s.values[attr.Equivalent()]
if !ok {
d.res = s.newRes()
}
d.attrs = attr
d.value = value
d.res.Offer(ctx, value, droppedAttr)
s.values[attr.Equivalent()] = d
}
func (s *lastValue[N]) delta(dest *metricdata.Aggregation) int {
t := now()
// Ignore if dest is not a metricdata.Gauge. The chance for memory reuse of
// the DataPoints is missed (better luck next time).
gData, _ := (*dest).(metricdata.Gauge[N])
s.Lock()
defer s.Unlock()
n := s.copyDpts(&gData.DataPoints, t)
// Do not report stale values.
clear(s.values)
// Update start time for delta temporality.
s.start = t
*dest = gData
return n
}
func (s *lastValue[N]) cumulative(dest *metricdata.Aggregation) int {
t := now()
// Ignore if dest is not a metricdata.Gauge. The chance for memory reuse of
// the DataPoints is missed (better luck next time).
gData, _ := (*dest).(metricdata.Gauge[N])
s.Lock()
defer s.Unlock()
n := s.copyDpts(&gData.DataPoints, t)
// TODO (#3006): This will use an unbounded amount of memory if there
// are unbounded number of attribute sets being aggregated. Attribute
// sets that become "stale" need to be forgotten so this will not
// overload the system.
*dest = gData
return n
}
// copyDpts copies the datapoints held by s into dest. The number of datapoints
// copied is returned.
func (s *lastValue[N]) copyDpts(dest *[]metricdata.DataPoint[N], t time.Time) int {
n := len(s.values)
*dest = reset(*dest, n, n)
var i int
for _, v := range s.values {
(*dest)[i].Attributes = v.attrs
(*dest)[i].StartTime = s.start
(*dest)[i].Time = t
(*dest)[i].Value = v.value
collectExemplars(&(*dest)[i].Exemplars, v.res.Collect)
i++
}
return n
}
// newPrecomputedLastValue returns an aggregator that summarizes a set of
// observations as the last one made.
func newPrecomputedLastValue[N int64 | float64](limit int, r func() exemplar.FilteredReservoir[N]) *precomputedLastValue[N] {
return &precomputedLastValue[N]{lastValue: newLastValue[N](limit, r)}
}
// precomputedLastValue summarizes a set of observations as the last one made.
type precomputedLastValue[N int64 | float64] struct {
*lastValue[N]
}
func (s *precomputedLastValue[N]) delta(dest *metricdata.Aggregation) int {
t := now()
// Ignore if dest is not a metricdata.Gauge. The chance for memory reuse of
// the DataPoints is missed (better luck next time).
gData, _ := (*dest).(metricdata.Gauge[N])
s.Lock()
defer s.Unlock()
n := s.copyDpts(&gData.DataPoints, t)
// Do not report stale values.
clear(s.values)
// Update start time for delta temporality.
s.start = t
*dest = gData
return n
}
func (s *precomputedLastValue[N]) cumulative(dest *metricdata.Aggregation) int {
t := now()
// Ignore if dest is not a metricdata.Gauge. The chance for memory reuse of
// the DataPoints is missed (better luck next time).
gData, _ := (*dest).(metricdata.Gauge[N])
s.Lock()
defer s.Unlock()
n := s.copyDpts(&gData.DataPoints, t)
// Do not report stale values.
clear(s.values)
*dest = gData
return n
}