prometheus/promql/value.go

495 lines
13 KiB
Go
Raw Normal View History

// Copyright 2017 The Prometheus Authors
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package promql
import (
"encoding/json"
"errors"
"fmt"
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
"math"
"strconv"
"strings"
Style cleanup of all the changes in sparsehistogram so far A lot of this code was hacked together, literally during a hackathon. This commit intends not to change the code substantially, but just make the code obey the usual style practices. A (possibly incomplete) list of areas: * Generally address linter warnings. * The `pgk` directory is deprecated as per dev-summit. No new packages should be added to it. I moved the new `pkg/histogram` package to `model` anticipating what's proposed in #9478. * Make the naming of the Sparse Histogram more consistent. Including abbreviations, there were just too many names for it: SparseHistogram, Histogram, Histo, hist, his, shs, h. The idea is to call it "Histogram" in general. Only add "Sparse" if it is needed to avoid confusion with conventional Histograms (which is rare because the TSDB really has no notion of conventional Histograms). Use abbreviations only in local scope, and then really abbreviate (not just removing three out of seven letters like in "Histo"). This is in the spirit of https://github.com/golang/go/wiki/CodeReviewComments#variable-names * Several other minor name changes. * A lot of formatting of doc comments. For one, following https://github.com/golang/go/wiki/CodeReviewComments#comment-sentences , but also layout question, anticipating how things will look like when rendered by `godoc` (even where `godoc` doesn't render them right now because they are for unexported types or not a doc comment at all but just a normal code comment - consistency is queen!). * Re-enabled `TestQueryLog` and `TestEndopints` (they pass now, leaving them disabled was presumably an oversight). * Bucket iterator for histogram.Histogram is now created with a method. * HistogramChunk.iterator now allows iterator recycling. (I think @dieterbe only commented it out because he was confused by the question in the comment.) * HistogramAppender.Append panics now because we decided to treat staleness marker differently. Signed-off-by: beorn7 <beorn@grafana.com>
2021-10-09 13:57:07 +00:00
"github.com/prometheus/prometheus/model/histogram"
"github.com/prometheus/prometheus/model/labels"
"github.com/prometheus/prometheus/promql/parser"
"github.com/prometheus/prometheus/storage"
"github.com/prometheus/prometheus/tsdb/chunkenc"
)
func (Matrix) Type() parser.ValueType { return parser.ValueTypeMatrix }
func (Vector) Type() parser.ValueType { return parser.ValueTypeVector }
func (Scalar) Type() parser.ValueType { return parser.ValueTypeScalar }
func (String) Type() parser.ValueType { return parser.ValueTypeString }
// String represents a string value.
type String struct {
T int64
V string
}
func (s String) String() string {
return s.V
}
func (s String) MarshalJSON() ([]byte, error) {
return json.Marshal([...]interface{}{float64(s.T) / 1000, s.V})
}
// Scalar is a data point that's explicitly not associated with a metric.
type Scalar struct {
T int64
V float64
}
func (s Scalar) String() string {
v := strconv.FormatFloat(s.V, 'f', -1, 64)
return fmt.Sprintf("scalar: %v @[%v]", v, s.T)
}
func (s Scalar) MarshalJSON() ([]byte, error) {
v := strconv.FormatFloat(s.V, 'f', -1, 64)
return json.Marshal([...]interface{}{float64(s.T) / 1000, v})
}
// Series is a stream of data points belonging to a metric.
type Series struct {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
Metric labels.Labels `json:"metric"`
Floats []FPoint `json:"values,omitempty"`
Histograms []HPoint `json:"histograms,omitempty"`
}
func (s Series) String() string {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// TODO(beorn7): This currently renders floats first and then
// histograms, each sorted by timestamp. Maybe, in mixed series, that's
// fine. Maybe, however, primary sorting by timestamp is preferred, in
// which case this has to be changed.
vals := make([]string, 0, len(s.Floats)+len(s.Histograms))
for _, f := range s.Floats {
vals = append(vals, f.String())
}
for _, h := range s.Histograms {
vals = append(vals, h.String())
}
return fmt.Sprintf("%s =>\n%s", s.Metric, strings.Join(vals, "\n"))
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// FPoint represents a single float data point for a given timestamp.
type FPoint struct {
T int64
F float64
}
func (p FPoint) String() string {
s := strconv.FormatFloat(p.F, 'f', -1, 64)
return fmt.Sprintf("%s @[%v]", s, p.T)
}
// MarshalJSON implements json.Marshaler.
//
// JSON marshaling is only needed for the HTTP API. Since FPoint is such a
// frequently marshaled type, it gets an optimized treatment directly in
// web/api/v1/api.go. Therefore, this method is unused within Prometheus. It is
// still provided here as convenience for debugging and for other users of this
// code. Also note that the different marshaling implementations might lead to
// slightly different results in terms of formatting and rounding of the
// timestamp.
func (p FPoint) MarshalJSON() ([]byte, error) {
v := strconv.FormatFloat(p.F, 'f', -1, 64)
return json.Marshal([...]interface{}{float64(p.T) / 1000, v})
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// HPoint represents a single histogram data point for a given timestamp.
// H must never be nil.
type HPoint struct {
T int64
H *histogram.FloatHistogram
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
func (p HPoint) String() string {
return fmt.Sprintf("%s @[%v]", p.H.String(), p.T)
}
// MarshalJSON implements json.Marshaler.
//
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// JSON marshaling is only needed for the HTTP API. Since HPoint is such a
// frequently marshaled type, it gets an optimized treatment directly in
// web/api/v1/api.go. Therefore, this method is unused within Prometheus. It is
// still provided here as convenience for debugging and for other users of this
// code. Also note that the different marshaling implementations might lead to
// slightly different results in terms of formatting and rounding of the
// timestamp.
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
func (p HPoint) MarshalJSON() ([]byte, error) {
h := struct {
Count string `json:"count"`
Sum string `json:"sum"`
Buckets [][]interface{} `json:"buckets,omitempty"`
}{
Count: strconv.FormatFloat(p.H.Count, 'f', -1, 64),
Sum: strconv.FormatFloat(p.H.Sum, 'f', -1, 64),
}
it := p.H.AllBucketIterator()
for it.Next() {
bucket := it.At()
if bucket.Count == 0 {
continue // No need to expose empty buckets in JSON.
}
boundaries := 2 // Exclusive on both sides AKA open interval.
if bucket.LowerInclusive {
if bucket.UpperInclusive {
boundaries = 3 // Inclusive on both sides AKA closed interval.
} else {
boundaries = 1 // Inclusive only on lower end AKA right open.
}
} else {
if bucket.UpperInclusive {
boundaries = 0 // Inclusive only on upper end AKA left open.
}
}
bucketToMarshal := []interface{}{
boundaries,
strconv.FormatFloat(bucket.Lower, 'f', -1, 64),
strconv.FormatFloat(bucket.Upper, 'f', -1, 64),
strconv.FormatFloat(bucket.Count, 'f', -1, 64),
}
h.Buckets = append(h.Buckets, bucketToMarshal)
}
return json.Marshal([...]interface{}{float64(p.T) / 1000, h})
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// Sample is a single sample belonging to a metric. It represents either a float
// sample or a histogram sample. If H is nil, it is a float sample. Otherwise,
// it is a histogram sample.
type Sample struct {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
T int64
F float64
H *histogram.FloatHistogram
Metric labels.Labels
}
func (s Sample) String() string {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
var str string
if s.H == nil {
p := FPoint{T: s.T, F: s.F}
str = p.String()
} else {
p := HPoint{T: s.T, H: s.H}
str = p.String()
}
return fmt.Sprintf("%s => %s", s.Metric, str)
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// MarshalJSON is mirrored in web/api/v1/api.go with jsoniter because FPoint and
// HPoint wouldn't be marshaled with jsoniter otherwise.
func (s Sample) MarshalJSON() ([]byte, error) {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
if s.H == nil {
f := struct {
M labels.Labels `json:"metric"`
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
F FPoint `json:"value"`
}{
M: s.Metric,
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
F: FPoint{T: s.T, F: s.F},
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
return json.Marshal(f)
}
h := struct {
M labels.Labels `json:"metric"`
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
H HPoint `json:"histogram"`
}{
M: s.Metric,
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
H: HPoint{T: s.T, H: s.H},
}
return json.Marshal(h)
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// Vector is basically only an an alias for []Sample, but the contract is that
// in a Vector, all Samples have the same timestamp.
type Vector []Sample
func (vec Vector) String() string {
entries := make([]string, len(vec))
for i, s := range vec {
entries[i] = s.String()
}
return strings.Join(entries, "\n")
}
// ContainsSameLabelset checks if a vector has samples with the same labelset
// Such a behavior is semantically undefined
// https://github.com/prometheus/prometheus/issues/4562
func (vec Vector) ContainsSameLabelset() bool {
switch len(vec) {
case 0, 1:
return false
case 2:
return vec[0].Metric.Hash() == vec[1].Metric.Hash()
default:
l := make(map[uint64]struct{}, len(vec))
for _, ss := range vec {
hash := ss.Metric.Hash()
if _, ok := l[hash]; ok {
return true
}
l[hash] = struct{}{}
}
return false
}
}
// Matrix is a slice of Series that implements sort.Interface and
// has a String method.
type Matrix []Series
func (m Matrix) String() string {
// TODO(fabxc): sort, or can we rely on order from the querier?
strs := make([]string, len(m))
for i, ss := range m {
strs[i] = ss.String()
}
return strings.Join(strs, "\n")
}
// TotalSamples returns the total number of samples in the series within a matrix.
func (m Matrix) TotalSamples() int {
numSamples := 0
for _, series := range m {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
numSamples += len(series.Floats) + len(series.Histograms)
}
return numSamples
}
func (m Matrix) Len() int { return len(m) }
func (m Matrix) Less(i, j int) bool { return labels.Compare(m[i].Metric, m[j].Metric) < 0 }
func (m Matrix) Swap(i, j int) { m[i], m[j] = m[j], m[i] }
// ContainsSameLabelset checks if a matrix has samples with the same labelset.
// Such a behavior is semantically undefined.
// https://github.com/prometheus/prometheus/issues/4562
func (m Matrix) ContainsSameLabelset() bool {
switch len(m) {
case 0, 1:
return false
case 2:
return m[0].Metric.Hash() == m[1].Metric.Hash()
default:
l := make(map[uint64]struct{}, len(m))
for _, ss := range m {
hash := ss.Metric.Hash()
if _, ok := l[hash]; ok {
return true
}
l[hash] = struct{}{}
}
return false
}
}
// Result holds the resulting value of an execution or an error
// if any occurred.
type Result struct {
Err error
Value parser.Value
Warnings storage.Warnings
}
// Vector returns a Vector if the result value is one. An error is returned if
// the result was an error or the result value is not a Vector.
func (r *Result) Vector() (Vector, error) {
if r.Err != nil {
return nil, r.Err
}
v, ok := r.Value.(Vector)
if !ok {
return nil, errors.New("query result is not a Vector")
}
return v, nil
}
// Matrix returns a Matrix. An error is returned if
// the result was an error or the result value is not a Matrix.
func (r *Result) Matrix() (Matrix, error) {
if r.Err != nil {
return nil, r.Err
}
v, ok := r.Value.(Matrix)
if !ok {
return nil, errors.New("query result is not a range Vector")
}
return v, nil
}
// Scalar returns a Scalar value. An error is returned if
// the result was an error or the result value is not a Scalar.
func (r *Result) Scalar() (Scalar, error) {
if r.Err != nil {
return Scalar{}, r.Err
}
v, ok := r.Value.(Scalar)
if !ok {
return Scalar{}, errors.New("query result is not a Scalar")
}
return v, nil
}
func (r *Result) String() string {
if r.Err != nil {
return r.Err.Error()
}
if r.Value == nil {
return ""
}
return r.Value.String()
}
// StorageSeries simulates promql.Series as storage.Series.
type StorageSeries struct {
series Series
}
// NewStorageSeries returns a StorageSeries from a Series.
func NewStorageSeries(series Series) *StorageSeries {
return &StorageSeries{
series: series,
}
}
func (ss *StorageSeries) Labels() labels.Labels {
return ss.series.Metric
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
// Iterator returns a new iterator of the data of the series. In case of
// multiple samples with the same timestamp, it returns the float samples first.
func (ss *StorageSeries) Iterator(it chunkenc.Iterator) chunkenc.Iterator {
if ssi, ok := it.(*storageSeriesIterator); ok {
ssi.reset(ss.series)
return ssi
}
return newStorageSeriesIterator(ss.series)
}
type storageSeriesIterator struct {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
floats []FPoint
histograms []HPoint
iFloats, iHistograms int
currT int64
currF float64
currH *histogram.FloatHistogram
}
func newStorageSeriesIterator(series Series) *storageSeriesIterator {
return &storageSeriesIterator{
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
floats: series.Floats,
histograms: series.Histograms,
iFloats: -1,
iHistograms: 0,
currT: math.MinInt64,
}
}
func (ssi *storageSeriesIterator) reset(series Series) {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
ssi.floats = series.Floats
ssi.histograms = series.Histograms
ssi.iFloats = -1
ssi.iHistograms = 0
ssi.currT = math.MinInt64
ssi.currF = 0
ssi.currH = nil
}
func (ssi *storageSeriesIterator) Seek(t int64) chunkenc.ValueType {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
if ssi.iFloats >= len(ssi.floats) && ssi.iHistograms >= len(ssi.histograms) {
return chunkenc.ValNone
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
for ssi.currT < t {
if ssi.Next() == chunkenc.ValNone {
return chunkenc.ValNone
}
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
if ssi.currH != nil {
return chunkenc.ValFloatHistogram
}
return chunkenc.ValFloat
}
func (ssi *storageSeriesIterator) At() (t int64, v float64) {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
return ssi.currT, ssi.currF
}
func (ssi *storageSeriesIterator) AtHistogram() (int64, *histogram.Histogram) {
panic(errors.New("storageSeriesIterator: AtHistogram not supported"))
}
func (ssi *storageSeriesIterator) AtFloatHistogram() (int64, *histogram.FloatHistogram) {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
return ssi.currT, ssi.currH
}
func (ssi *storageSeriesIterator) AtT() int64 {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
return ssi.currT
}
func (ssi *storageSeriesIterator) Next() chunkenc.ValueType {
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
if ssi.currH != nil {
ssi.iHistograms++
} else {
ssi.iFloats++
}
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
var (
pickH, pickF = false, false
floatsExhausted = ssi.iFloats >= len(ssi.floats)
histogramsExhausted = ssi.iHistograms >= len(ssi.histograms)
)
switch {
case floatsExhausted:
if histogramsExhausted { // Both exhausted!
return chunkenc.ValNone
}
pickH = true
case histogramsExhausted: // and floats not exhausted.
pickF = true
// From here on, we have to look at timestamps.
case ssi.histograms[ssi.iHistograms].T < ssi.floats[ssi.iFloats].T:
// Next histogram comes before next float.
pickH = true
default:
// In all other cases, we pick float so that we first iterate
// through floats if the timestamp is the same.
pickF = true
}
switch {
case pickF:
p := ssi.floats[ssi.iFloats]
ssi.currT = p.T
ssi.currF = p.F
ssi.currH = nil
return chunkenc.ValFloat
case pickH:
p := ssi.histograms[ssi.iHistograms]
ssi.currT = p.T
ssi.currF = 0
ssi.currH = p.H
return chunkenc.ValFloatHistogram
promql: Separate `Point` into `FPoint` and `HPoint` In other words: Instead of having a “polymorphous” `Point` that can either contain a float value or a histogram value, use an `FPoint` for floats and an `HPoint` for histograms. This seemingly small change has a _lot_ of repercussions throughout the codebase. The idea here is to avoid the increase in size of `Point` arrays that happened after native histograms had been added. The higher-level data structures (`Sample`, `Series`, etc.) are still “polymorphous”. The same idea could be applied to them, but at each step the trade-offs needed to be evaluated. The idea with this change is to do the minimum necessary to get back to pre-histogram performance for functions that do not touch histograms. Here are comparisons for the `changes` function. The test data doesn't include histograms yet. Ideally, there would be no change in the benchmark result at all. First runtime v2.39 compared to directly prior to this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10) ``` And then runtime v2.39 compared to after this commit: ``` name old time/op new time/op delta RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8) RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10) RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10) RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8) RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10) RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9) RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10) RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10) RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9) ``` In summary, the runtime doesn't really improve with this change for queries with just a few steps. For queries with many steps, this commit essentially reinstates the old performance. This is good because the many-step queries are the one that matter most (longest absolute runtime). In terms of allocations, though, this commit doesn't make a dent at all (numbers not shown). The reason is that most of the allocations happen in the sampleRingIterator (in the storage package), which has to be addressed in a separate commit. Signed-off-by: beorn7 <beorn@grafana.com>
2022-10-28 14:58:40 +00:00
default:
panic("storageSeriesIterater.Next failed to pick value type")
}
}
func (ssi *storageSeriesIterator) Err() error {
return nil
}