github
/
prometheus
mirror of https://github.com/prometheus/prometheus
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '29fff1eca4'
${ noResults }
prometheus/promql/parse_test.go

1932 lines
50 KiB
Raw Normal View History Unescape

Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
// Copyright 2015 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 (
"fmt"
"math"
"reflect"
Fix and improve parsing error output.
10 years ago
"strings"
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
"testing"
"time"
Switch from client_golang/model to common/model
9 years ago
"github.com/prometheus/common/model"
Fix grouping of import blocks
9 years ago
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
"github.com/prometheus/prometheus/storage/metric"
)
var testExpr = []struct {
Fix and improve parsing error output.
10 years ago
input string // The input to be parsed.
expected Expr // The expected expression AST.
fail bool // Whether parsing is supposed to fail.
errMsg string // If not empty the parsing error has to contain this string.
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}{
// Scalars and scalar-to-scalar operations.
{
input: "1",
expected: &NumberLiteral{1},
}, {
input: "+Inf",
Switch from client_golang/model to common/model
9 years ago
expected: &NumberLiteral{model.SampleValue(math.Inf(1))},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "-Inf",
Switch from client_golang/model to common/model
9 years ago
expected: &NumberLiteral{model.SampleValue(math.Inf(-1))},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: ".5",
expected: &NumberLiteral{0.5},
}, {
input: "5.",
expected: &NumberLiteral{5},
}, {
input: "123.4567",
expected: &NumberLiteral{123.4567},
}, {
input: "5e-3",
expected: &NumberLiteral{0.005},
}, {
input: "5e3",
expected: &NumberLiteral{5000},
}, {
input: "0xc",
expected: &NumberLiteral{12},
}, {
input: "0755",
expected: &NumberLiteral{493},
}, {
input: "+5.5e-3",
expected: &NumberLiteral{0.0055},
}, {
input: "-0755",
expected: &NumberLiteral{-493},
}, {
input: "1 + 1",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
expected: &BinaryExpr{itemADD, &NumberLiteral{1}, &NumberLiteral{1}, nil, false},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "1 - 1",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
expected: &BinaryExpr{itemSUB, &NumberLiteral{1}, &NumberLiteral{1}, nil, false},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "1 * 1",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
expected: &BinaryExpr{itemMUL, &NumberLiteral{1}, &NumberLiteral{1}, nil, false},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "1 % 1",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
expected: &BinaryExpr{itemMOD, &NumberLiteral{1}, &NumberLiteral{1}, nil, false},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "1 / 1",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
expected: &BinaryExpr{itemDIV, &NumberLiteral{1}, &NumberLiteral{1}, nil, false},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
input: "1 == bool 1",
expected: &BinaryExpr{itemEQL, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
input: "1 != bool 1",
expected: &BinaryExpr{itemNEQ, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
input: "1 > bool 1",
expected: &BinaryExpr{itemGTR, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
input: "1 >= bool 1",
expected: &BinaryExpr{itemGTE, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
input: "1 < bool 1",
expected: &BinaryExpr{itemLSS, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
}, {
input: "1 <= bool 1",
expected: &BinaryExpr{itemLTE, &NumberLiteral{1}, &NumberLiteral{1}, nil, true},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "+1 + -2 * 1",
expected: &BinaryExpr{
Op: itemADD,
LHS: &NumberLiteral{1},
RHS: &BinaryExpr{
Op: itemMUL, LHS: &NumberLiteral{-2}, RHS: &NumberLiteral{1},
},
},
}, {
input: "1 + 2/(3*1)",
expected: &BinaryExpr{
Op: itemADD,
LHS: &NumberLiteral{1},
RHS: &BinaryExpr{
Op: itemDIV,
LHS: &NumberLiteral{2},
RHS: &ParenExpr{&BinaryExpr{
Op: itemMUL, LHS: &NumberLiteral{3}, RHS: &NumberLiteral{1},
}},
},
},
rewrite operator balancing to be recursive
9 years ago
}, {
input: "1 < bool 2 - 1 * 2",
expected: &BinaryExpr{
Op: itemLSS,
ReturnBool: true,
LHS: &NumberLiteral{1},
RHS: &BinaryExpr{
Op: itemSUB,
LHS: &NumberLiteral{2},
RHS: &BinaryExpr{
Op: itemMUL, LHS: &NumberLiteral{1}, RHS: &NumberLiteral{2},
},
},
},
Implement unary expression for vector types. Closes #956
9 years ago
}, {
input: "-some_metric", expected: &UnaryExpr{
Op: itemSUB,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Implement unary expression for vector types. Closes #956
9 years ago
},
},
},
}, {
input: "+some_metric", expected: &UnaryExpr{
Op: itemADD,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Implement unary expression for vector types. Closes #956
9 years ago
},
},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "",
fail: true,
errMsg: "no expression found in input",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "# just a comment\n\n",
fail: true,
errMsg: "no expression found in input",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1+",
fail: true,
Add missing check for nil expression
9 years ago
errMsg: "no valid expression found",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: ".",
fail: true,
errMsg: "unexpected character: '.'",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "2.5.",
fail: true,
errMsg: "could not parse remaining input \".\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "100..4",
fail: true,
errMsg: "could not parse remaining input \".4\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "0deadbeef",
fail: true,
errMsg: "bad number or duration syntax: \"0de\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 /",
fail: true,
Add missing check for nil expression
9 years ago
errMsg: "no valid expression found",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "*1",
fail: true,
errMsg: "no valid expression found",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "(1))",
fail: true,
errMsg: "could not parse remaining input \")\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "((1)",
fail: true,
errMsg: "unclosed left parenthesis",
Migrate parsing error tests. The promql_test checks failure of various bad syntaxed queries. Those are moved into the parser tests as the new testing language only deals with valid queries.
10 years ago
}, {
input: "999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999",
fail: true,
errMsg: "out of range",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "(",
fail: true,
errMsg: "unclosed left parenthesis",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 and 1",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"and\" not allowed in binary scalar expression",
promql: Remove scalar/scalar comparisons. This change is breaking, use the 'bool' modifier for such comprisons. After this change all comparisons without 'bool' will filter, and all comparisons with 'bool' will return 0/1. This makes the language more consistent and orthogonal, and ultimately easier to learn and use. If we ever figure out sane semantics for filtering scalar/scalar comparisons we can add them in, which will most likely come out of how the new vector() function is used.
9 years ago
}, {
input: "1 == 1",
fail: true,
errMsg: "parse error at char 7: comparisons between scalars must use BOOL modifier",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 or 1",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"or\" not allowed in binary scalar expression",
}, {
input: "1 unless 1",
fail: true,
errMsg: "set operator \"unless\" not allowed in binary scalar expression",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 !~ 1",
fail: true,
errMsg: "could not parse remaining input \"!~ 1\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 =~ 1",
fail: true,
errMsg: "could not parse remaining input \"=~ 1\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Implement unary expression for vector types. Closes #956
9 years ago
input: `-"string"`,
Fix and improve parsing error output.
10 years ago
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: `unary expression only allowed on expressions of type scalar or instant vector, got "string"`,
Fix and improve parsing error output.
10 years ago
}, {
Implement unary expression for vector types. Closes #956
9 years ago
input: `-test[5m]`,
Fix and improve parsing error output.
10 years ago
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: `unary expression only allowed on expressions of type scalar or instant vector, got "range vector"`,
Implement unary expression for vector types. Closes #956
9 years ago
}, {
input: `*test`,
fail: true,
errMsg: "no valid expression found",
Rename matrix to range selector in external error messages The documentation speaks about range vectors and range vector selectors. This change does not fix all issues, we might still expose the term "Matrix" in error messages using %T.
9 years ago
}, {
input: "1 offset 1d",
fail: true,
errMsg: "offset modifier must be preceded by an instant or range selector",
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
}, {
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
input: "a - on(b) ignoring(c) d",
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
fail: true,
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
errMsg: "parse error at char 11: no valid expression found",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
// Vector binary operations.
{
input: "foo * bar",
expected: &BinaryExpr{
Op: itemMUL,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{Card: CardOneToOne},
},
}, {
input: "foo == 1",
expected: &BinaryExpr{
Op: itemEQL,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &NumberLiteral{1},
},
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
}, {
input: "foo == bool 1",
expected: &BinaryExpr{
Op: itemEQL,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
},
},
RHS: &NumberLiteral{1},
ReturnBool: true,
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "2.5 / bar",
expected: &BinaryExpr{
Op: itemDIV,
LHS: &NumberLiteral{2.5},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
},
}, {
input: "foo and bar",
expected: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
}, {
input: "foo or bar",
expected: &BinaryExpr{
Op: itemLOR,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
}, {
input: "foo unless bar",
expected: &BinaryExpr{
Op: itemLUnless,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
// Test and/or precedence and reassigning of operands.
input: "foo + bar or bla and blub",
expected: &BinaryExpr{
Op: itemLOR,
LHS: &BinaryExpr{
Op: itemADD,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{Card: CardOneToOne},
},
RHS: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "bla",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bla"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "blub",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "blub"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
}, {
// Test and/or/unless precedence.
input: "foo and bar unless baz or qux",
expected: &BinaryExpr{
Op: itemLOR,
LHS: &BinaryExpr{
Op: itemLUnless,
LHS: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
RHS: &VectorSelector{
Name: "baz",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "baz"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
RHS: &VectorSelector{
Name: "qux",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "qux"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
VectorMatching: &VectorMatching{Card: CardManyToMany},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
// Test precedence and reassigning of operands.
input: "bar + on(foo) bla / on(baz, buz) group_right(test) blub",
expected: &BinaryExpr{
Op: itemADD,
LHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &BinaryExpr{
Op: itemDIV,
LHS: &VectorSelector{
Name: "bla",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bla"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "blub",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "blub"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardOneToMany,
MatchingLabels: model.LabelNames{"baz", "buz"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Rename On to MatchingLabels
9 years ago
Include: model.LabelNames{"test"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardOneToOne,
MatchingLabels: model.LabelNames{"foo"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "foo * on(test,blub) bar",
expected: &BinaryExpr{
Op: itemMUL,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardOneToOne,
MatchingLabels: model.LabelNames{"test", "blub"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
For many-to-one matches, always copy label from one side. This is a breaking change for everyone using the machine roles labeling approach.
9 years ago
}, {
input: "foo * on(test,blub) group_left bar",
expected: &BinaryExpr{
Op: itemMUL,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
For many-to-one matches, always copy label from one side. This is a breaking change for everyone using the machine roles labeling approach.
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
For many-to-one matches, always copy label from one side. This is a breaking change for everyone using the machine roles labeling approach.
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToOne,
MatchingLabels: model.LabelNames{"test", "blub"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
For many-to-one matches, always copy label from one side. This is a breaking change for everyone using the machine roles labeling approach.
9 years ago
},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "foo and on(test,blub) bar",
expected: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToMany,
MatchingLabels: model.LabelNames{"test", "blub"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
}, {
input: "foo and on() bar",
expected: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
},
},
VectorMatching: &VectorMatching{
Card: CardManyToMany,
MatchingLabels: model.LabelNames{},
On: true,
},
},
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
}, {
input: "foo and ignoring(test,blub) bar",
expected: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToMany,
MatchingLabels: model.LabelNames{"test", "blub"},
Add 'ignoring' as modifier for binops. Where 'on' uses the given labels to match, 'ignoring' uses all other labels to match. group_left/right is not supported yet.
9 years ago
},
},
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
}, {
input: "foo and ignoring() bar",
expected: &BinaryExpr{
Op: itemLAND,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
},
},
VectorMatching: &VectorMatching{
Card: CardManyToMany,
MatchingLabels: model.LabelNames{},
},
},
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
}, {
input: "foo unless on(bar) baz",
expected: &BinaryExpr{
Op: itemLUnless,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
RHS: &VectorSelector{
Name: "baz",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "baz"),
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToMany,
MatchingLabels: model.LabelNames{"bar"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "foo / on(test,blub) group_left(bar) bar",
expected: &BinaryExpr{
Op: itemDIV,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToOne,
MatchingLabels: model.LabelNames{"test", "blub"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Rename On to MatchingLabels
9 years ago
Include: model.LabelNames{"bar"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
}, {
input: "foo / ignoring(test,blub) group_left(blub) bar",
expected: &BinaryExpr{
Op: itemDIV,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToOne,
MatchingLabels: model.LabelNames{"test", "blub"},
Include: model.LabelNames{"blub"},
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
}, {
input: "foo / ignoring(test,blub) group_left(bar) bar",
expected: &BinaryExpr{
Op: itemDIV,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardManyToOne,
MatchingLabels: model.LabelNames{"test", "blub"},
Include: model.LabelNames{"bar"},
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "foo - on(test,blub) group_right(bar,foo) bar",
expected: &BinaryExpr{
Op: itemSUB,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardOneToMany,
MatchingLabels: model.LabelNames{"test", "blub"},
Include: model.LabelNames{"bar", "foo"},
Flip vector matching to be ignoring by default. This is a noop semantically.
9 years ago
On: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
}, {
input: "foo - ignoring(test,blub) group_right(bar,foo) bar",
expected: &BinaryExpr{
Op: itemSUB,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
RHS: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
VectorMatching: &VectorMatching{
Rename On to MatchingLabels
9 years ago
Card: CardOneToMany,
MatchingLabels: model.LabelNames{"test", "blub"},
Include: model.LabelNames{"bar", "foo"},
Add support for OneToMany with IGNORING. The labels listed in the group_ modifier will be copied from the one side to the many side. It will be valid to specify no labels. This is intended to replace the existing ON/GROUP_* support.,
9 years ago
},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo and 1",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"and\" not allowed in binary scalar expression",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 and foo",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"and\" not allowed in binary scalar expression",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo or 1",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"or\" not allowed in binary scalar expression",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 or foo",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "set operator \"or\" not allowed in binary scalar expression",
}, {
input: "foo unless 1",
fail: true,
errMsg: "set operator \"unless\" not allowed in binary scalar expression",
}, {
input: "1 unless foo",
fail: true,
errMsg: "set operator \"unless\" not allowed in binary scalar expression",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "1 or on(bar) foo",
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: "vector matching only allowed between instant vectors",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo == on(bar) 10",
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: "vector matching only allowed between instant vectors",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo and on(bar) group_left(baz) bar",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "no grouping allowed for \"and\" operation",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo and on(bar) group_right(baz) bar",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "no grouping allowed for \"and\" operation",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo or on(bar) group_left(baz) bar",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "no grouping allowed for \"or\" operation",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "foo or on(bar) group_right(baz) bar",
fail: true,
Implement relative complement set operator "unless" The `unless` set operator can be used to return all vector elements from the LHS which do not match the elements on the RHS. A use case is to return all metrics for nodes which do not have a specific role: node_load1 unless on(instance) chef_role{role="app"}
9 years ago
errMsg: "no grouping allowed for \"or\" operation",
}, {
input: "foo unless on(bar) group_left(baz) bar",
fail: true,
errMsg: "no grouping allowed for \"unless\" operation",
}, {
input: "foo unless on(bar) group_right(baz) bar",
fail: true,
errMsg: "no grouping allowed for \"unless\" operation",
Migrate parsing error tests. The promql_test checks failure of various bad syntaxed queries. Those are moved into the parser tests as the new testing language only deals with valid queries.
10 years ago
}, {
input: `http_requests{group="production"} + on(instance) group_left(job,instance) cpu_count{type="smp"}`,
fail: true,
Clarify error message around on(x) group_left(x)
9 years ago
errMsg: "label \"instance\" must not occur in ON and GROUP clause at once",
promql: Add 'bool' modifier to comparison functions When doing comparison operations on vectors, filtering sometimes gets in the way and you have to go to a fair bit of effort to workaround it in order to always return a result. The 'bool' modifier instead of filtering returns 0/1 depending on the result of the compairson. This is also a prerequisite to removing plain scalar/scalar comparisons, as it maintains the current behaviour under a new syntax.
9 years ago
}, {
input: "foo + bool bar",
fail: true,
errMsg: "bool modifier can only be used on comparison operators",
}, {
input: "foo + bool 10",
fail: true,
errMsg: "bool modifier can only be used on comparison operators",
}, {
input: "foo and bool 10",
fail: true,
errMsg: "bool modifier can only be used on comparison operators",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
// Test vector selector.
{
input: "foo",
expected: &VectorSelector{
Name: "foo",
Offset: 0,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "foo offset 5m",
expected: &VectorSelector{
Name: "foo",
Offset: 5 * time.Minute,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
Fix single quote parsing, add tests
10 years ago
input: `foo:bar{a="bc"}`,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &VectorSelector{
Name: "foo:bar",
Offset: 0,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "a", "bc"),
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo:bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
Fix single quote parsing, add tests
10 years ago
input: `foo{NaN='bc'}`,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &VectorSelector{
Name: "foo",
Offset: 0,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "NaN", "bc"),
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: `foo{a="b", foo!="bar", test=~"test", bar!~"baz"}`,
expected: &VectorSelector{
Name: "foo",
Offset: 0,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "a", "b"),
mustLabelMatcher(metric.NotEqual, "foo", "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
mustLabelMatcher(metric.RegexMatch, "test", "test"),
mustLabelMatcher(metric.RegexNoMatch, "bar", "baz"),
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
Fix and improve parsing error output.
10 years ago
input: `{`,
fail: true,
errMsg: "unexpected end of input inside braces",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `}`,
fail: true,
errMsg: "unexpected character: '}'",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some{`,
fail: true,
errMsg: "unexpected end of input inside braces",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some}`,
fail: true,
errMsg: "could not parse remaining input \"}\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some_metric{a=b}`,
fail: true,
errMsg: "unexpected identifier \"b\" in label matching, expected string",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some_metric{a:b="b"}`,
fail: true,
errMsg: "unexpected character inside braces: ':'",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo{a*"b"}`,
fail: true,
errMsg: "unexpected character inside braces: '*'",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo{a>="b"}`,
fail: true,
// TODO(fabxc): willingly lexing wrong tokens allows for more precrise error
// messages from the parser - consider if this is an option.
errMsg: "unexpected character inside braces: '>'",
Check for invalid utf-8 in lexer strings. This protects against invalid utf-8 sneaking in via label_replace.
8 years ago
}, {
input: "some_metric{a=\"\xff\"}",
fail: true,
errMsg: "parse error at char 15: invalid UTF-8 rune",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo{gibberish}`,
fail: true,
errMsg: "expected label matching operator but got }",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo{1}`,
fail: true,
errMsg: "unexpected character inside braces: '1'",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `{}`,
fail: true,
errMsg: "vector selector must contain label matchers or metric name",
promql: error if all label matchers are empty.
10 years ago
}, {
input: `{x=""}`,
fail: true,
errMsg: "vector selector must contain at least one non-empty matcher",
}, {
input: `{x=~".*"}`,
fail: true,
errMsg: "vector selector must contain at least one non-empty matcher",
}, {
input: `{x!~".+"}`,
fail: true,
errMsg: "vector selector must contain at least one non-empty matcher",
}, {
input: `{x!="a"}`,
fail: true,
errMsg: "vector selector must contain at least one non-empty matcher",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo{__name__="bar"}`,
fail: true,
errMsg: "metric name must not be set twice: \"foo\" or \"bar\"",
// }, {
// input: `:foo`,
// fail: true,
// errMsg: "bla",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
// Test matrix selector.
{
input: "test[5s]",
expected: &MatrixSelector{
Name: "test",
Offset: 0,
Range: 5 * time.Second,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "test[5m]",
expected: &MatrixSelector{
Name: "test",
Offset: 0,
Range: 5 * time.Minute,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "test[5h] OFFSET 5m",
expected: &MatrixSelector{
Name: "test",
Offset: 5 * time.Minute,
Range: 5 * time.Hour,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "test[5d] OFFSET 10s",
expected: &MatrixSelector{
Name: "test",
Offset: 10 * time.Second,
Range: 5 * 24 * time.Hour,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: "test[5w] offset 2w",
expected: &MatrixSelector{
Name: "test",
Offset: 14 * 24 * time.Hour,
Range: 5 * 7 * 24 * time.Hour,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: `test{a="b"}[5y] OFFSET 3d`,
expected: &MatrixSelector{
Name: "test",
Offset: 3 * 24 * time.Hour,
Range: 5 * 365 * 24 * time.Hour,
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "a", "b"),
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "test"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[5mm]`,
fail: true,
errMsg: "bad duration syntax: \"5mm\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[0m]`,
fail: true,
errMsg: "duration must be greater than 0",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[5m30s]`,
fail: true,
errMsg: "bad duration syntax: \"5m3\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[5m] OFFSET 1h30m`,
fail: true,
errMsg: "bad number or duration syntax: \"1h3\"",
Migrate parsing error tests. The promql_test checks failure of various bad syntaxed queries. Those are moved into the parser tests as the new testing language only deals with valid queries.
10 years ago
}, {
input: `foo["5m"]`,
fail: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[]`,
fail: true,
errMsg: "missing unit character in duration",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `foo[1]`,
fail: true,
errMsg: "missing unit character in duration",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some_metric[5m] OFFSET 1`,
fail: true,
Consolidate offset modifier parsing Remove duplicated offset modifier parsing and ensure offset can only appear at the end of a selector statement.
9 years ago
errMsg: "unexpected number \"1\" in offset, expected duration",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some_metric[5m] OFFSET 1mm`,
fail: true,
errMsg: "bad number or duration syntax: \"1mm\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `some_metric[5m] OFFSET`,
fail: true,
Consolidate offset modifier parsing Remove duplicated offset modifier parsing and ensure offset can only appear at the end of a selector statement.
9 years ago
errMsg: "unexpected end of input in offset, expected duration",
}, {
input: `some_metric OFFSET 1m[5m]`,
fail: true,
errMsg: "could not parse remaining input \"[5m]\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `(foo + bar)[5m]`,
fail: true,
errMsg: "could not parse remaining input \"[5m]\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
// Test aggregation.
{
input: "sum by (foo)(some_metric)",
expected: &AggregateExpr{
Op: itemSum,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "sum by (foo) keep_common (some_metric)",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &AggregateExpr{
remove keeping_extra because it's replaced with keep_common change all keepExtra label into keepCommon, and move action into removed list change incorrect token list
9 years ago
Op: itemSum,
KeepCommonLabels: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "sum (some_metric) by (foo,bar) keep_common",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &AggregateExpr{
remove keeping_extra because it's replaced with keep_common change all keepExtra label into keepCommon, and move action into removed list change incorrect token list
9 years ago
Op: itemSum,
KeepCommonLabels: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo", "bar"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
input: "avg by (foo)(some_metric)",
expected: &AggregateExpr{
Op: itemAvg,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "COUNT by (foo) keep_common (some_metric)",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &AggregateExpr{
Op: itemCount,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
remove keeping_extra because it's replaced with keep_common change all keepExtra label into keepCommon, and move action into removed list change incorrect token list
9 years ago
Grouping: model.LabelNames{"foo"},
KeepCommonLabels: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "MIN (some_metric) by (foo) keep_common",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expected: &AggregateExpr{
Op: itemMin,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
remove keeping_extra because it's replaced with keep_common change all keepExtra label into keepCommon, and move action into removed list change incorrect token list
9 years ago
Grouping: model.LabelNames{"foo"},
KeepCommonLabels: true,
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
}, {
input: "max by (foo)(some_metric)",
expected: &AggregateExpr{
Op: itemMax,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
Add without aggregator modifier. This has the advantage that the user doesn't need to list all labels they want to keep (as with "by") but without having to worry about inconsistent labels as when there's only one time series (as with "keeping_common"). Almost all aggregation should use this rather than the existing two options as it's much less error prone and easier to maintain due to not having to always add in "job" plus whatever other common job-level labels you have like "region".
9 years ago
}, {
input: "sum without (foo) (some_metric)",
expected: &AggregateExpr{
Op: itemSum,
Without: true,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Add without aggregator modifier. This has the advantage that the user doesn't need to list all labels they want to keep (as with "by") but without having to worry about inconsistent labels as when there's only one time series (as with "keeping_common"). Almost all aggregation should use this rather than the existing two options as it's much less error prone and easier to maintain due to not having to always add in "job" plus whatever other common job-level labels you have like "region".
9 years ago
},
},
Grouping: model.LabelNames{"foo"},
},
}, {
input: "sum (some_metric) without (foo)",
expected: &AggregateExpr{
Op: itemSum,
Without: true,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Add without aggregator modifier. This has the advantage that the user doesn't need to list all labels they want to keep (as with "by") but without having to worry about inconsistent labels as when there's only one time series (as with "keeping_common"). Almost all aggregation should use this rather than the existing two options as it's much less error prone and easier to maintain due to not having to always add in "job" plus whatever other common job-level labels you have like "region".
9 years ago
},
},
Grouping: model.LabelNames{"foo"},
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "stddev(some_metric)",
expected: &AggregateExpr{
Op: itemStddev,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
},
}, {
input: "stdvar by (foo)(some_metric)",
expected: &AggregateExpr{
Op: itemStdvar,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"foo"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
}, {
input: "sum by ()(some_metric)",
expected: &AggregateExpr{
Op: itemSum,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Allow on, ignoring, by and without wit empty laberls. This offers new semantics in allowing on() for matching two single-element vectors with no known common labels. Previosuly this was often done using on(dummy). This also allows making it explict that you meant to do an aggregation without labels via by(). Fixes #1597.
9 years ago
},
},
Grouping: model.LabelNames{},
},
Make topk/bottomk aggregators.
9 years ago
}, {
input: "topk(5, some_metric)",
expected: &AggregateExpr{
Op: itemTopK,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Make topk/bottomk aggregators.
9 years ago
},
},
Param: &NumberLiteral{5},
},
Add count_values() aggregator. This is useful for counting how many instances of a job are running a particular version/build. Fixes #622
9 years ago
}, {
input: "count_values(\"value\", some_metric)",
expected: &AggregateExpr{
Op: itemCountValues,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Add count_values() aggregator. This is useful for counting how many instances of a job are running a particular version/build. Fixes #622
9 years ago
},
},
Param: &StringLiteral{"value"},
},
Fix parsing of label names which are also keywords The current separation between lexer and parser is a bit fuzzy when it comes to operators, aggregators and other keywords. The lexer already tries to determine the type of a token, even though that type might change depending on the context. This led to the problematic behavior that no tokens known to the lexer could be used as label names, including operators (and, by, ...), aggregators (count, quantile, ...) or other keywords (for, offset, ...). This change additionally checks whether an identifier is one of these types. We might want to check whether the specific item identification should be moved from the lexer to the parser.
8 years ago
}, {
// Test usage of keywords as label names.
input: "sum without(and, by, avg, count, alert, annotations)(some_metric)",
expected: &AggregateExpr{
Op: itemSum,
Without: true,
Expr: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
},
},
Grouping: model.LabelNames{"and", "by", "avg", "count", "alert", "annotations"},
},
}, {
input: "sum without(==)(some_metric)",
fail: true,
errMsg: "unexpected <op:==> in grouping opts, expected label",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `sum some_metric by (test)`,
fail: true,
errMsg: "unexpected identifier \"some_metric\" in aggregation, expected \"(\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `sum (some_metric) by test`,
fail: true,
errMsg: "unexpected identifier \"test\" in grouping opts, expected \"(\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `sum (some_metric) by test`,
fail: true,
errMsg: "unexpected identifier \"test\" in grouping opts, expected \"(\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: `sum () by (test)`,
fail: true,
errMsg: "no valid expression found",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "MIN keep_common (some_metric) by (foo)",
Fix and improve parsing error output.
10 years ago
fail: true,
errMsg: "could not parse remaining input \"by (foo)\"...",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
input: "MIN by(test) (some_metric) keep_common",
Fix and improve parsing error output.
10 years ago
fail: true,
Deprecate `keeping_extra`, rename it to `keep_common`. `keep_common` is more in line with the function name `drop_common_labels()` terminology-wise, and also more in line with `group_left`/`group_right` (no `...ing` verb suffix). We could also go the full way and call it `keep_common_labels`. That would have the benefit of being even more consistent with the function `drop_common_labels()` and would be more explanatory, but it also seems quite long.
10 years ago
errMsg: "could not parse remaining input \"keep_common\"...",
Add without aggregator modifier. This has the advantage that the user doesn't need to list all labels they want to keep (as with "by") but without having to worry about inconsistent labels as when there's only one time series (as with "keeping_common"). Almost all aggregation should use this rather than the existing two options as it's much less error prone and easier to maintain due to not having to always add in "job" plus whatever other common job-level labels you have like "region".
9 years ago
}, {
input: `sum (some_metric) without (test) keep_common`,
fail: true,
errMsg: "cannot use 'keep_common' with 'without'",
}, {
input: `sum (some_metric) without (test) by (test)`,
fail: true,
errMsg: "could not parse remaining input \"by (test)\"...",
}, {
input: `sum without (test) (some_metric) by (test)`,
fail: true,
errMsg: "could not parse remaining input \"by (test)\"...",
Make topk/bottomk aggregators.
9 years ago
}, {
input: `topk(some_metric)`,
fail: true,
errMsg: "parse error at char 17: unexpected \")\" in aggregation, expected \",\"",
Add type check on topk/bottomk parameter.
9 years ago
}, {
input: `topk(some_metric, other_metric)`,
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: "parse error at char 32: expected type scalar in aggregation parameter, got instant vector",
Add count_values() aggregator. This is useful for counting how many instances of a job are running a particular version/build. Fixes #622
9 years ago
}, {
input: `count_values(5, other_metric)`,
fail: true,
errMsg: "parse error at char 30: expected type string in aggregation parameter, got scalar",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
// Test function calls.
{
input: "time()",
expected: &Call{
Func: mustGetFunction("time"),
},
}, {
input: `floor(some_metric{foo!="bar"})`,
expected: &Call{
Func: mustGetFunction("floor"),
Args: Expressions{
&VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.NotEqual, "foo", "bar"),
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
},
},
}, {
input: "rate(some_metric[5m])",
expected: &Call{
Func: mustGetFunction("rate"),
Args: Expressions{
&MatrixSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
Range: 5 * time.Minute,
},
},
},
}, {
input: "round(some_metric)",
expected: &Call{
Func: mustGetFunction("round"),
Args: Expressions{
&VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
},
},
}, {
input: "round(some_metric, 5)",
expected: &Call{
Func: mustGetFunction("round"),
Args: Expressions{
&VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
&NumberLiteral{5},
},
},
}, {
Fix and improve parsing error output.
10 years ago
input: "floor()",
fail: true,
implement label_join issue 1147 (#2806) Replace OptionalArgs int with Variadic int.
8 years ago
errMsg: "expected 1 argument(s) in call to \"floor\", got 0",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "floor(some_metric, other_metric)",
fail: true,
implement label_join issue 1147 (#2806) Replace OptionalArgs int with Variadic int.
8 years ago
errMsg: "expected 1 argument(s) in call to \"floor\", got 2",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "floor(1)",
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: "expected type instant vector in call to function \"floor\", got scalar",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix common english misspellings
8 years ago
input: "non_existent_function_far_bar()",
Fix and improve parsing error output.
10 years ago
fail: true,
Fix common english misspellings
8 years ago
errMsg: "unknown function with name \"non_existent_function_far_bar\"",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
Fix and improve parsing error output.
10 years ago
input: "rate(some_metric)",
fail: true,
Report type names in the form used in documentation
8 years ago
errMsg: "expected type range vector in call to function \"rate\", got instant vector",
Check for invalid utf-8 in lexer strings. This protects against invalid utf-8 sneaking in via label_replace.
8 years ago
}, {
input: "label_replace(a, `b`, `c\xff`, `d`, `.*`)",
fail: true,
errMsg: "parse error at char 23: invalid UTF-8 rune",
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
Add missing check for nil expression
9 years ago
// Fuzzing regression tests.
{
input: "-=",
fail: true,
errMsg: `no valid expression found`,
Implement unary expression for vector types. Closes #956
9 years ago
}, {
Add missing check for nil expression
9 years ago
input: "++-++-+-+-<",
fail: true,
errMsg: `no valid expression found`,
Implement unary expression for vector types. Closes #956
9 years ago
}, {
Add missing check for nil expression
9 years ago
input: "e-+=/(0)",
fail: true,
errMsg: `no valid expression found`,
Implement unary expression for vector types. Closes #956
9 years ago
}, {
Add missing check for nil expression
9 years ago
input: "-If",
fail: true,
errMsg: `no valid expression found`,
},
Support escape sequences in strings and add raw strings. This adapts some functionality from the Go standard library for string literal lexing and unquoting/unescaping. The following string types are now supported: Double- or single-quoted strings: These support all escape sequences that Go supports in double-quoted string literals. The difference is that Prometheus also has single-quoted strings (instead of single-quoted runes in Go). Raw newlines are not allowed. Backtick-quoted raw strings: Strings quoted in backticks are treated as raw strings just like in Go and may contain raw newlines and other special characters directly. Fixes https://github.com/prometheus/prometheus/issues/1122 Fixes https://github.com/prometheus/prometheus/issues/1121
9 years ago
// String quoting and escape sequence interpretation tests.
{
input: `"double-quoted string \" with escaped quote"`,
expected: &StringLiteral{
Val: "double-quoted string \" with escaped quote",
},
}, {
input: `'single-quoted string \' with escaped quote'`,
expected: &StringLiteral{
Val: "single-quoted string ' with escaped quote",
},
}, {
input: "`backtick-quoted string`",
expected: &StringLiteral{
Val: "backtick-quoted string",
},
}, {
input: `"\a\b\f\n\r\t\v\\\" - \xFF\377\u1234\U00010111\U0001011111☺"`,
expected: &StringLiteral{
Val: "\a\b\f\n\r\t\v\\\" - \xFF\377\u1234\U00010111\U0001011111☺",
},
}, {
input: `'\a\b\f\n\r\t\v\\\' - \xFF\377\u1234\U00010111\U0001011111☺'`,
expected: &StringLiteral{
Val: "\a\b\f\n\r\t\v\\' - \xFF\377\u1234\U00010111\U0001011111☺",
},
}, {
input: "`" + `\a\b\f\n\r\t\v\\\"\' - \xFF\377\u1234\U00010111\U0001011111☺` + "`",
expected: &StringLiteral{
Val: `\a\b\f\n\r\t\v\\\"\' - \xFF\377\u1234\U00010111\U0001011111☺`,
},
}, {
input: "`\\``",
fail: true,
errMsg: "could not parse remaining input",
}, {
input: `"\`,
fail: true,
errMsg: "escape sequence not terminated",
}, {
input: `"\c"`,
fail: true,
errMsg: "unknown escape sequence U+0063 'c'",
}, {
input: `"\x."`,
fail: true,
errMsg: "illegal character U+002E '.' in escape sequence",
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}
func TestParseExpressions(t *testing.T) {
for _, test := range testExpr {
Remove `name` arg from `Parse*` functions, enhance parsing errors.
10 years ago
parser := newParser(test.input)
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expr, err := parser.parseExpr()
forbid unexpected (runtime) errors in parse tests
9 years ago
// Unexpected errors are always caused by a bug.
if err == errUnexpected {
t.Fatalf("unexpected error occurred")
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
if !test.fail && err != nil {
t.Errorf("error in input '%s'", test.input)
t.Fatalf("could not parse: %s", err)
}
if test.fail && err != nil {
Fix and improve parsing error output.
10 years ago
if !strings.Contains(err.Error(), test.errMsg) {
t.Errorf("unexpected error on input '%s'", test.input)
t.Fatalf("expected error to contain %q but got %q", test.errMsg, err)
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
continue
}
err = parser.typecheck(expr)
if !test.fail && err != nil {
t.Errorf("error on input '%s'", test.input)
t.Fatalf("typecheck failed: %s", err)
}
if test.fail {
if err != nil {
Fix and improve parsing error output.
10 years ago
if !strings.Contains(err.Error(), test.errMsg) {
t.Errorf("unexpected error on input '%s'", test.input)
t.Fatalf("expected error to contain %q but got %q", test.errMsg, err)
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
continue
}
t.Errorf("error on input '%s'", test.input)
t.Fatalf("failure expected, but passed with result: %q", expr)
}
if !reflect.DeepEqual(expr, test.expected) {
t.Errorf("error on input '%s'", test.input)
t.Fatalf("no match\n\nexpected:\n%s\ngot: \n%s\n", Tree(test.expected), Tree(expr))
}
}
}
// NaN has no equality. Thus, we need a separate test for it.
func TestNaNExpression(t *testing.T) {
Remove `name` arg from `Parse*` functions, enhance parsing errors.
10 years ago
parser := newParser("NaN")
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
expr, err := parser.parseExpr()
if err != nil {
t.Errorf("error on input 'NaN'")
t.Fatalf("coud not parse: %s", err)
}
nl, ok := expr.(*NumberLiteral)
if !ok {
t.Errorf("error on input 'NaN'")
t.Fatalf("expected number literal but got %T", expr)
}
if !math.IsNaN(float64(nl.Val)) {
t.Errorf("error on input 'NaN'")
Fix unkeyed field errors.
9 years ago
t.Fatalf("expected 'NaN' in number literal but got %v", nl.Val)
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}
}
var testStatement = []struct {
input string
expected Statements
fail bool
}{
{
// Test a file-like input.
input: `
# A simple test recording rule.
dc:http_request:rate5m = sum(rate(http_request_count[5m])) by (dc)
Consolidate offset modifier parsing Remove duplicated offset modifier parsing and ensure offset can only appear at the end of a selector statement.
9 years ago
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
# A simple test alerting rule.
Change WITH keyword to LABELS
9 years ago
ALERT GlobalRequestRateLow IF(dc:http_request:rate5m < 10000) FOR 5m
LABELS {
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
service = "testservice"
# ... more fields here ...
}
Implement new alerting rule syntax
9 years ago
ANNOTATIONS {
summary = "Global request rate low",
description = "The global request rate is low"
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
foo = bar{label1="value1"}
promql: add runbook to alert statement. This commit adds the RUNBOOK keyword to alert statements. The field is optional and expected to be a link.
10 years ago
ALERT BazAlert IF foo > 10
Implement new alerting rule syntax
9 years ago
ANNOTATIONS {
description = "BazAlert",
runbook = "http://my.url",
summary = "Baz",
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
`,
expected: Statements{
&RecordStmt{
Name: "dc:http_request:rate5m",
Expr: &AggregateExpr{
Op: itemSum,
Switch from client_golang/model to common/model
9 years ago
Grouping: model.LabelNames{"dc"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
Expr: &Call{
Func: mustGetFunction("rate"),
Args: Expressions{
&MatrixSelector{
Name: "http_request_count",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "http_request_count"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
Range: 5 * time.Minute,
},
},
},
},
Labels: nil,
},
&AlertStmt{
Name: "GlobalRequestRateLow",
Expr: &ParenExpr{&BinaryExpr{
Op: itemLSS,
LHS: &VectorSelector{
Name: "dc:http_request:rate5m",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "dc:http_request:rate5m"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &NumberLiteral{10000},
}},
Implement new alerting rule syntax
9 years ago
Labels: model.LabelSet{"service": "testservice"},
Duration: 5 * time.Minute,
Annotations: model.LabelSet{
"summary": "Global request rate low",
"description": "The global request rate is low",
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
&RecordStmt{
Name: "foo",
Expr: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "label1", "value1"),
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Labels: nil,
},
&AlertStmt{
Name: "BazAlert",
Expr: &BinaryExpr{
Op: itemGTR,
LHS: &VectorSelector{
Name: "foo",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "foo"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &NumberLiteral{10},
},
Implement new alerting rule syntax
9 years ago
Labels: model.LabelSet{},
Annotations: model.LabelSet{
"summary": "Baz",
"description": "BazAlert",
"runbook": "http://my.url",
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: `foo{x="", a="z"} = bar{a="b", x=~"y"}`,
expected: Statements{
&RecordStmt{
Name: "foo",
Expr: &VectorSelector{
Name: "bar",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, "a", "b"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
mustLabelMatcher(metric.RegexMatch, "x", "y"),
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "bar"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Switch from client_golang/model to common/model
9 years ago
Labels: model.LabelSet{"x": "", "a": "z"},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
Support old alerting rule syntax
9 years ago
}, {
input: `ALERT SomeName IF some_metric > 1
LABELS {}
Implement new alerting rule syntax
9 years ago
ANNOTATIONS {
Support old alerting rule syntax
9 years ago
summary = "Global request rate low",
description = "The global request rate is low",
Implement new alerting rule syntax
9 years ago
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
`,
expected: Statements{
&AlertStmt{
Name: "SomeName",
Expr: &BinaryExpr{
Op: itemGTR,
LHS: &VectorSelector{
Name: "some_metric",
LabelMatchers: metric.LabelMatchers{
storage: improve index lookups tl;dr: This is not a fundamental solution to the indexing problem (like tindex is) but it at least avoids utilizing the intersection problem to the greatest possible amount. In more detail: Imagine the following query: nicely:aggregating:rule{job="foo",env="prod"} While it uses a nicely aggregating recording rule (which might have a very low cardinality), Prometheus still intersects the low number of fingerprints for `{__name__="nicely:aggregating:rule"}` with the many thousands of fingerprints matching `{job="foo"}` and with the millions of fingerprints matching `{env="prod"}`. This totally innocuous query is dead slow if the Prometheus server has a lot of time series with the `{env="prod"}` label. Ironically, if you make the query more complicated, it becomes blazingly fast: nicely:aggregating:rule{job=~"foo",env=~"prod"} Why so? Because Prometheus only intersects with non-Equal matchers if there are no Equal matchers. That's good in this case because it retrieves the few fingerprints for `{__name__="nicely:aggregating:rule"}` and then starts right ahead to retrieve the metric for those FPs and checking individually if they match the other matchers. This change is generalizing the idea of when to stop intersecting FPs and go into "retrieve metrics and check them individually against remaining matchers" mode: - First, sort all matchers by "expected cardinality". Matchers matching the empty string are always worst (and never used for intersections). Equal matchers are in general consider best, but by using some crude heuristics, we declare some better than others (instance labels or anything that looks like a recording rule). - Then go through the matchers until we hit a threshold of remaining FPs in the intersection. This threshold is higher if we are already in the non-Equal matcher area as intersection is even more expensive here. - Once the threshold has been reached (or we have run out of matchers that do not match the empty string), start with "retrieve metrics and check them individually against remaining matchers". A beefy server at SoundCloud was spending 67% of its CPU time in index lookups (fingerprintsForLabelPairs), serving mostly a dashboard that is exclusively built with recording rules. With this change, it spends only 35% in fingerprintsForLabelPairs. The CPU usage dropped from 26 cores to 18 cores. The median latency for query_range dropped from 14s to 50ms(!). As expected, higher percentile latency didn't improve that much because the new approach is _occasionally_ running into the worst case while the old one was _systematically_ doing so. The 99th percentile latency is now about as high as the median before (14s) while it was almost twice as high before (26s).
9 years ago
mustLabelMatcher(metric.Equal, model.MetricNameLabel, "some_metric"),
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
RHS: &NumberLiteral{1},
},
Implement new alerting rule syntax
9 years ago
Labels: model.LabelSet{},
Annotations: model.LabelSet{
"summary": "Global request rate low",
"description": "The global request rate is low",
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
},
},
}, {
input: `
# A simple test alerting rule.
Change WITH keyword to LABELS
9 years ago
ALERT GlobalRequestRateLow IF(dc:http_request:rate5m < 10000) FOR 5
LABELS {
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
service = "testservice"
Consolidate offset modifier parsing Remove duplicated offset modifier parsing and ensure offset can only appear at the end of a selector statement.
9 years ago
# ... more fields here ...
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}
Implement new alerting rule syntax
9 years ago
ANNOTATIONS {
summary = "Global request rate low"
description = "The global request rate is low"
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
`,
fail: true,
}, {
input: "",
expected: Statements{},
}, {
input: "foo = time()",
rules: Allow recorded rules expressions to be scalars. This is useful if you want to build up a constant metric, such as a set of alert thresholds that vary by label value.
9 years ago
expected: Statements{
&RecordStmt{
Name: "foo",
Expr: &Call{Func: mustGetFunction("time")},
Labels: nil,
}},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "foo = 1",
rules: Allow recorded rules expressions to be scalars. This is useful if you want to build up a constant metric, such as a set of alert thresholds that vary by label value.
9 years ago
expected: Statements{
&RecordStmt{
Name: "foo",
Expr: &NumberLiteral{1},
Labels: nil,
}},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}, {
input: "foo = bar[5m]",
fail: true,
}, {
input: `foo = "test"`,
fail: true,
}, {
input: `foo = `,
fail: true,
}, {
input: `foo{a!="b"} = bar`,
fail: true,
}, {
input: `foo{a=~"b"} = bar`,
fail: true,
}, {
input: `foo{a!~"b"} = bar`,
fail: true,
},
forbid unexpected (runtime) errors in parse tests
9 years ago
// Fuzzing regression tests.
{
input: `I=-/`,
fail: true,
},
{
input: `I=3E8/-=`,
fail: true,
},
{
input: `M=-=-0-0`,
fail: true,
},
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
}
func TestParseStatements(t *testing.T) {
for _, test := range testStatement {
Remove `name` arg from `Parse*` functions, enhance parsing errors.
10 years ago
parser := newParser(test.input)
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
stmts, err := parser.parseStmts()
forbid unexpected (runtime) errors in parse tests
9 years ago
// Unexpected errors are always caused by a bug.
if err == errUnexpected {
t.Fatalf("unexpected error occurred")
}
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
if !test.fail && err != nil {
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("could not parse: %s", err)
}
if test.fail && err != nil {
continue
}
err = parser.typecheck(stmts)
if !test.fail && err != nil {
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("typecheck failed: %s", err)
}
if test.fail {
if err != nil {
continue
}
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("failure expected, but passed")
}
if !reflect.DeepEqual(stmts, test.expected) {
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("no match\n\nexpected:\n%s\ngot: \n%s\n", Tree(test.expected), Tree(stmts))
}
}
}
Switch from client_golang/model to common/model
9 years ago
func mustLabelMatcher(mt metric.MatchType, name model.LabelName, val model.LabelValue) *metric.LabelMatcher {
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
m, err := metric.NewLabelMatcher(mt, name, val)
if err != nil {
panic(err)
}
return m
}
func mustGetFunction(name string) *Function {
Refactor query evaluation. This copies the evaluation logic from the current rules/ package. The new engine handles the execution process from query string to final result. It provides query timeout and cancellation and general flexibility for future changes. functions.go: Add evaluation implementation. Slight changes to in/out data but not to the processing logic. quantile.go: No changes. analyzer.go: No changes. engine.go: Actually new part. Mainly consists of evaluation methods which were not changed. setup_test.go: Copy of rules/helpers_test.go to setup test storage. promql_test.go: Copy of rules/rules_test.go.
10 years ago
f, ok := getFunction(name)
Create promql package with lexer/parser. This commit creates a (so far unused) package. It contains the a custom lexer/parser for the query language. ast.go: New AST that interacts well with the parser. lex.go: Custom lexer (new). lex_test.go: Lexer tests (new). parse.go: Custom parser (new). parse_test.go: Parser tests (new). functions.go: Changed function type, dummies for parser testing (barely changed/dummies). printer.go: Adapted from rules/ and adjusted to new AST (mostly unchanged, few additions).
10 years ago
if !ok {
panic(fmt.Errorf("function %q does not exist", name))
}
return f
}
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
var testSeries = []struct {
input string
Switch from client_golang/model to common/model
9 years ago
expectedMetric model.Metric
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
expectedValues []sequenceValue
fail bool
}{
{
input: `{} 1 2 3`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{},
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
expectedValues: newSeq(1, 2, 3),
}, {
input: `{a="b"} -1 2 3`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
"a": "b",
},
expectedValues: newSeq(-1, 2, 3),
}, {
input: `my_metric 1 2 3`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
},
expectedValues: newSeq(1, 2, 3),
}, {
input: `my_metric{} 1 2 3`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
},
expectedValues: newSeq(1, 2, 3),
}, {
input: `my_metric{a="b"} 1 2 3`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
"a": "b",
},
expectedValues: newSeq(1, 2, 3),
}, {
input: `my_metric{a="b"} 1 2 3-10x4`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
"a": "b",
},
expectedValues: newSeq(1, 2, 3, -7, -17, -27, -37),
promql: fix parsing ambiguity for serial repetitions
10 years ago
}, {
input: `my_metric{a="b"} 1 2 3-0x4`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
promql: fix parsing ambiguity for serial repetitions
10 years ago
"a": "b",
},
expectedValues: newSeq(1, 2, 3, 3, 3, 3, 3),
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
}, {
input: `my_metric{a="b"} 1 3 _ 5 _x4`,
Switch from client_golang/model to common/model
9 years ago
expectedMetric: model.Metric{
model.MetricNameLabel: "my_metric",
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
"a": "b",
},
expectedValues: newSeq(1, 3, none, 5, none, none, none, none),
}, {
input: `my_metric{a="b"} 1 3 _ 5 _a4`,
fail: true,
},
}
Add parsing and execution of new test format. This commit adds a new test structure that parses and executes the new testing language.
10 years ago
// For these tests only, we use the smallest float64 to signal an omitted value.
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
const none = math.SmallestNonzeroFloat64
func newSeq(vals ...float64) (res []sequenceValue) {
for _, v := range vals {
if v == none {
res = append(res, sequenceValue{omitted: true})
} else {
Switch from client_golang/model to common/model
9 years ago
res = append(res, sequenceValue{value: model.SampleValue(v)})
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
}
}
return res
}
func TestParseSeries(t *testing.T) {
for _, test := range testSeries {
parser := newParser(test.input)
parser.lex.seriesDesc = true
metric, vals, err := parser.parseSeriesDesc()
forbid unexpected (runtime) errors in parse tests
9 years ago
// Unexpected errors are always caused by a bug.
if err == errUnexpected {
t.Fatalf("unexpected error occurred")
}
Add time series description parsing. This commit adds parsing of time series description to the exisiting query language parser. Time series descriptions are defined by a metric followed by a sequence of values.
10 years ago
if !test.fail && err != nil {
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("could not parse: %s", err)
}
if test.fail && err != nil {
continue
}
if test.fail {
if err != nil {
continue
}
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("failure expected, but passed")
}
if !reflect.DeepEqual(vals, test.expectedValues) || !reflect.DeepEqual(metric, test.expectedMetric) {
t.Errorf("error in input: \n\n%s\n", test.input)
t.Fatalf("no match\n\nexpected:\n%s %s\ngot: \n%s %s\n", test.expectedMetric, test.expectedValues, metric, vals)
}
}
}
Handle parser runtime panics gracefully
9 years ago
Log runtime errors during query evaluation instead of panicking.
9 years ago
func TestRecoverParserRuntime(t *testing.T) {
Handle parser runtime panics gracefully
9 years ago
var p *parser
var err error
defer p.recover(&err)
// Cause a runtime panic.
var a []int
a[123] = 1
forbid unexpected (runtime) errors in parse tests
9 years ago
if err != errUnexpected {
t.Fatalf("wrong error message: %q, expected %q", err, errUnexpected)
Handle parser runtime panics gracefully
9 years ago
}
}
Log runtime errors during query evaluation instead of panicking.
9 years ago
func TestRecoverParserError(t *testing.T) {
Handle parser runtime panics gracefully
9 years ago
var p *parser
var err error
e := fmt.Errorf("custom error")
Fix "go vet" errors. I ignored all errors of the type "composite literal uses unkeyed fields". Most of them are wrong because of https://github.com/golang/go/issues/9171.
9 years ago
defer func() {
if err.Error() != e.Error() {
t.Fatalf("wrong error message: %q, expected %q", err, e)
}
}()
defer p.recover(&err)
panic(e)
Handle parser runtime panics gracefully
9 years ago
}