2019-05-22 23:16:55 +02:00
package promql
import (
2019-05-24 15:12:31 +02:00
"flag"
2019-05-22 23:16:55 +02:00
"fmt"
"math"
"sort"
2020-11-11 11:38:44 +01:00
"strings"
2019-05-22 23:16:55 +02:00
"sync"
"sync/atomic"
2019-05-24 15:12:31 +02:00
"time"
2019-05-22 23:16:55 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/netstorage"
2020-12-25 15:44:26 +01:00
"github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/querystats"
2023-01-10 07:19:53 +01:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
2021-03-15 11:35:44 +01:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
2022-06-01 01:29:19 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/querytracer"
2021-09-24 00:19:59 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/storage"
2019-05-22 23:16:55 +02:00
"github.com/VictoriaMetrics/metrics"
2020-04-28 14:28:22 +02:00
"github.com/VictoriaMetrics/metricsql"
2019-05-22 23:16:55 +02:00
)
2020-11-11 11:38:44 +01:00
var (
treatDotsAsIsInRegexps = flag . Bool ( "search.treatDotsAsIsInRegexps" , false , "Whether to treat dots as is in regexp label filters used in queries. " +
` For example, foo { bar=~"a.b.c"} will be automatically converted to foo { bar=~"a\\.b\\.c"}, i.e. all the dots in regexp filters will be automatically escaped ` +
` in order to match only dot char instead of matching any char. Dots in ".+", ".*" and ". { n}" regexps aren't escaped. ` +
2021-02-03 19:35:45 +01:00
` This option is DEPRECATED in favor of { __graphite__="a.*.c"} syntax for selecting metrics matching the given Graphite metrics filter ` )
2020-11-11 11:38:44 +01:00
)
2019-05-24 15:12:31 +02:00
2022-08-15 12:50:14 +02:00
// UserReadableError is a type of error which supposed to be returned to the user without additional context.
2022-08-15 12:38:47 +02:00
type UserReadableError struct {
2022-08-15 12:50:14 +02:00
// Err is the error which needs to be returned to the user.
2022-08-15 12:38:47 +02:00
Err error
}
2022-08-15 12:50:14 +02:00
// Unwrap returns ure.Err.
//
// This is used by standard errors package. See https://golang.org/pkg/errors
func ( ure * UserReadableError ) Unwrap ( ) error {
return ure . Err
}
2022-08-15 12:38:47 +02:00
// Error satisfies Error interface
2022-08-15 12:50:14 +02:00
func ( ure * UserReadableError ) Error ( ) string {
2022-08-15 12:38:47 +02:00
return ure . Err . Error ( )
}
2019-07-01 16:14:49 +02:00
// Exec executes q for the given ec.
2022-06-01 01:29:19 +02:00
func Exec ( qt * querytracer . Tracer , ec * EvalConfig , q string , isFirstPointOnly bool ) ( [ ] netstorage . Result , error ) {
2020-12-25 15:44:26 +01:00
if querystats . Enabled ( ) {
startTime := time . Now ( )
2020-12-27 11:06:04 +01:00
defer querystats . RegisterQuery ( q , ec . End - ec . Start , startTime )
2020-12-25 15:42:05 +01:00
}
2019-05-24 15:12:31 +02:00
2019-05-22 23:16:55 +02:00
ec . validate ( )
e , err := parsePromQLWithCache ( q )
if err != nil {
2020-07-08 17:55:25 +02:00
return nil , err
2019-05-22 23:16:55 +02:00
}
2020-07-08 17:55:25 +02:00
qid := activeQueriesV . Add ( ec , q )
2022-06-01 01:29:19 +02:00
rv , err := evalExpr ( qt , ec , e )
2020-07-08 17:55:25 +02:00
activeQueriesV . Remove ( qid )
2019-05-22 23:16:55 +02:00
if err != nil {
2020-07-08 17:55:25 +02:00
return nil , err
2019-05-22 23:16:55 +02:00
}
2019-07-01 16:14:49 +02:00
if isFirstPointOnly {
// Remove all the points except the first one from every time series.
for _ , ts := range rv {
ts . Values = ts . Values [ : 1 ]
ts . Timestamps = ts . Timestamps [ : 1 ]
}
2022-06-01 01:29:19 +02:00
qt . Printf ( "leave only the first point in every series" )
2019-07-01 16:14:49 +02:00
}
2019-05-22 23:16:55 +02:00
maySort := maySortResults ( e , rv )
result , err := timeseriesToResult ( rv , maySort )
if err != nil {
2020-07-08 17:55:25 +02:00
return nil , err
2019-05-22 23:16:55 +02:00
}
2022-06-01 01:29:19 +02:00
if maySort {
qt . Printf ( "sort series by metric name and labels" )
} else {
qt . Printf ( "do not sort series by metric name and labels" )
}
2021-03-15 11:35:44 +01:00
if n := ec . RoundDigits ; n < 100 {
for i := range result {
values := result [ i ] . Values
for j , v := range values {
values [ j ] = decimal . RoundToDecimalDigits ( v , n )
}
}
2022-06-01 01:29:19 +02:00
qt . Printf ( "round series values to %d decimal digits after the point" , n )
2021-03-15 11:35:44 +01:00
}
2022-08-15 12:38:47 +02:00
return result , nil
2019-05-22 23:16:55 +02:00
}
2019-12-25 20:35:47 +01:00
func maySortResults ( e metricsql . Expr , tss [ ] * timeseries ) bool {
2021-04-07 23:09:34 +02:00
switch v := e . ( type ) {
case * metricsql . FuncExpr :
switch strings . ToLower ( v . Name ) {
case "sort" , "sort_desc" ,
2022-09-14 16:41:09 +02:00
"sort_by_label" , "sort_by_label_desc" ,
"sort_by_label_numeric" , "sort_by_label_numeric_desc" :
2021-04-07 23:09:34 +02:00
return false
}
case * metricsql . AggrFuncExpr :
switch strings . ToLower ( v . Name ) {
case "topk" , "bottomk" , "outliersk" ,
2021-09-30 12:22:52 +02:00
"topk_max" , "topk_min" , "topk_avg" , "topk_median" , "topk_last" ,
"bottomk_max" , "bottomk_min" , "bottomk_avg" , "bottomk_median" , "bottomk_last" :
2021-04-07 23:09:34 +02:00
return false
}
2019-05-22 23:16:55 +02:00
}
2021-04-07 23:09:34 +02:00
return true
2019-05-22 23:16:55 +02:00
}
func timeseriesToResult ( tss [ ] * timeseries , maySort bool ) ( [ ] netstorage . Result , error ) {
2022-04-20 18:53:24 +02:00
tss = removeEmptySeries ( tss )
2019-05-22 23:16:55 +02:00
result := make ( [ ] netstorage . Result , len ( tss ) )
2019-08-20 21:51:45 +02:00
m := make ( map [ string ] struct { } , len ( tss ) )
2019-05-22 23:16:55 +02:00
bb := bbPool . Get ( )
for i , ts := range tss {
2021-09-24 00:19:59 +02:00
bb . B = marshalMetricNameSorted ( bb . B [ : 0 ] , & ts . MetricName )
2023-01-10 07:19:53 +01:00
k := bytesutil . InternBytes ( bb . B )
if _ , ok := m [ k ] ; ok {
2019-11-23 23:02:18 +01:00
return nil , fmt . Errorf ( ` duplicate output timeseries: %s ` , stringMetricName ( & ts . MetricName ) )
2019-05-22 23:16:55 +02:00
}
2023-01-10 07:19:53 +01:00
m [ k ] = struct { } { }
2019-05-22 23:16:55 +02:00
rs := & result [ i ]
2023-01-10 07:38:31 +01:00
rs . MetricName . MoveFrom ( & ts . MetricName )
rs . Values = ts . Values
ts . Values = nil
rs . Timestamps = ts . Timestamps
ts . Timestamps = nil
2019-05-22 23:16:55 +02:00
}
bbPool . Put ( bb )
if maySort {
sort . Slice ( result , func ( i , j int ) bool {
2021-09-24 00:19:59 +02:00
return metricNameLess ( & result [ i ] . MetricName , & result [ j ] . MetricName )
2019-05-22 23:16:55 +02:00
} )
}
return result , nil
}
2021-09-24 00:19:59 +02:00
func metricNameLess ( a , b * storage . MetricName ) bool {
if string ( a . MetricGroup ) != string ( b . MetricGroup ) {
return string ( a . MetricGroup ) < string ( b . MetricGroup )
}
// Metric names for a and b match. Compare tags.
// Tags must be already sorted by the caller, so just compare them.
ats := a . Tags
bts := b . Tags
for i := range ats {
if i >= len ( bts ) {
// a contains more tags than b and all the previous tags were identical,
// so a is considered bigger than b.
return false
}
at := & ats [ i ]
bt := & bts [ i ]
if string ( at . Key ) != string ( bt . Key ) {
return string ( at . Key ) < string ( bt . Key )
}
if string ( at . Value ) != string ( bt . Value ) {
return string ( at . Value ) < string ( bt . Value )
}
}
return len ( ats ) < len ( bts )
}
2022-04-20 18:53:24 +02:00
func removeEmptySeries ( tss [ ] * timeseries ) [ ] * timeseries {
2019-05-22 23:16:55 +02:00
rvs := tss [ : 0 ]
for _ , ts := range tss {
2019-07-12 18:56:23 +02:00
allNans := true
2019-05-22 23:16:55 +02:00
for _ , v := range ts . Values {
2019-07-12 18:56:23 +02:00
if ! math . IsNaN ( v ) {
allNans = false
break
2019-05-22 23:16:55 +02:00
}
}
2019-07-12 18:56:23 +02:00
if allNans {
2019-05-22 23:16:55 +02:00
// Skip timeseries with all NaNs.
continue
}
rvs = append ( rvs , ts )
}
2019-07-12 18:56:23 +02:00
for i := len ( rvs ) ; i < len ( tss ) ; i ++ {
// Zero unused time series, so GC could reclaim them.
tss [ i ] = nil
}
2019-05-22 23:16:55 +02:00
return rvs
}
2020-08-06 22:18:03 +02:00
func adjustCmpOps ( e metricsql . Expr ) metricsql . Expr {
metricsql . VisitAll ( e , func ( expr metricsql . Expr ) {
be , ok := expr . ( * metricsql . BinaryOpExpr )
if ! ok {
return
}
if ! metricsql . IsBinaryOpCmp ( be . Op ) {
return
}
2020-12-02 11:08:47 +01:00
if isNumberExpr ( be . Right ) || ! isScalarExpr ( be . Left ) {
2020-08-06 22:18:03 +02:00
return
}
// Convert 'num cmpOp query' expression to `query reverseCmpOp num` expression
2020-12-02 11:08:47 +01:00
// like Prometheus does. For instance, `0.5 < foo` must be converted to `foo > 0.5`
2020-08-06 22:18:03 +02:00
// in order to return valid values for `foo` that are bigger than 0.5.
be . Right , be . Left = be . Left , be . Right
be . Op = getReverseCmpOp ( be . Op )
} )
return e
}
2020-12-02 11:08:47 +01:00
func isNumberExpr ( e metricsql . Expr ) bool {
_ , ok := e . ( * metricsql . NumberExpr )
return ok
}
func isScalarExpr ( e metricsql . Expr ) bool {
if isNumberExpr ( e ) {
return true
}
if fe , ok := e . ( * metricsql . FuncExpr ) ; ok {
// time() returns scalar in PromQL - see https://prometheus.io/docs/prometheus/latest/querying/functions/#time
return strings . ToLower ( fe . Name ) == "time"
}
return false
}
2020-08-06 22:18:03 +02:00
func getReverseCmpOp ( op string ) string {
switch op {
case ">" :
return "<"
case "<" :
return ">"
case ">=" :
return "<="
case "<=" :
return ">="
default :
// there is no need in changing `==` and `!=`.
return op
}
}
2019-12-25 20:35:47 +01:00
func parsePromQLWithCache ( q string ) ( metricsql . Expr , error ) {
2019-05-22 23:16:55 +02:00
pcv := parseCacheV . Get ( q )
if pcv == nil {
2019-12-25 20:35:47 +01:00
e , err := metricsql . Parse ( q )
2020-08-06 22:18:03 +02:00
if err == nil {
2020-10-07 20:15:06 +02:00
e = metricsql . Optimize ( e )
2020-08-06 22:18:03 +02:00
e = adjustCmpOps ( e )
2020-11-11 11:38:44 +01:00
if * treatDotsAsIsInRegexps {
e = escapeDotsInRegexpLabelFilters ( e )
}
2020-08-06 22:18:03 +02:00
}
2019-05-22 23:16:55 +02:00
pcv = & parseCacheValue {
e : e ,
err : err ,
}
parseCacheV . Put ( q , pcv )
}
if pcv . err != nil {
return nil , pcv . err
}
return pcv . e , nil
}
2020-11-11 11:38:44 +01:00
func escapeDotsInRegexpLabelFilters ( e metricsql . Expr ) metricsql . Expr {
metricsql . VisitAll ( e , func ( expr metricsql . Expr ) {
me , ok := expr . ( * metricsql . MetricExpr )
if ! ok {
return
}
2023-07-16 08:48:21 +02:00
for _ , lfs := range me . LabelFilterss {
for i := range lfs {
f := & lfs [ i ]
if f . IsRegexp {
f . Value = escapeDots ( f . Value )
}
2020-11-11 11:38:44 +01:00
}
}
} )
return e
}
func escapeDots ( s string ) string {
dotsCount := strings . Count ( s , "." )
if dotsCount <= 0 {
return s
}
result := make ( [ ] byte , 0 , len ( s ) + 2 * dotsCount )
for i := 0 ; i < len ( s ) ; i ++ {
if s [ i ] == '.' && ( i == 0 || s [ i - 1 ] != '\\' ) && ( i + 1 == len ( s ) || i + 1 < len ( s ) && s [ i + 1 ] != '*' && s [ i + 1 ] != '+' && s [ i + 1 ] != '{' ) {
// Escape a dot if the following conditions are met:
// - if it isn't escaped already, i.e. if there is no `\` char before the dot.
// - if there is no regexp modifiers such as '+', '*' or '{' after the dot.
result = append ( result , '\\' , '.' )
} else {
result = append ( result , s [ i ] )
}
}
return string ( result )
}
2019-05-22 23:16:55 +02:00
var parseCacheV = func ( ) * parseCache {
pc := & parseCache {
m : make ( map [ string ] * parseCacheValue ) ,
}
metrics . NewGauge ( ` vm_cache_requests_total { type="promql/parse"} ` , func ( ) float64 {
return float64 ( pc . Requests ( ) )
} )
metrics . NewGauge ( ` vm_cache_misses_total { type="promql/parse"} ` , func ( ) float64 {
return float64 ( pc . Misses ( ) )
} )
metrics . NewGauge ( ` vm_cache_entries { type="promql/parse"} ` , func ( ) float64 {
return float64 ( pc . Len ( ) )
} )
return pc
} ( )
const parseCacheMaxLen = 10e3
type parseCacheValue struct {
2019-12-25 20:35:47 +01:00
e metricsql . Expr
2019-05-22 23:16:55 +02:00
err error
}
type parseCache struct {
2019-10-17 17:22:56 +02:00
// Move atomic counters to the top of struct for 8-byte alignment on 32-bit arch.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212
2019-05-22 23:16:55 +02:00
requests uint64
misses uint64
2019-10-17 17:22:56 +02:00
m map [ string ] * parseCacheValue
mu sync . RWMutex
2019-05-22 23:16:55 +02:00
}
func ( pc * parseCache ) Requests ( ) uint64 {
return atomic . LoadUint64 ( & pc . requests )
}
func ( pc * parseCache ) Misses ( ) uint64 {
return atomic . LoadUint64 ( & pc . misses )
}
func ( pc * parseCache ) Len ( ) uint64 {
pc . mu . RLock ( )
n := len ( pc . m )
pc . mu . RUnlock ( )
return uint64 ( n )
}
func ( pc * parseCache ) Get ( q string ) * parseCacheValue {
atomic . AddUint64 ( & pc . requests , 1 )
pc . mu . RLock ( )
pcv := pc . m [ q ]
pc . mu . RUnlock ( )
if pcv == nil {
atomic . AddUint64 ( & pc . misses , 1 )
}
return pcv
}
func ( pc * parseCache ) Put ( q string , pcv * parseCacheValue ) {
pc . mu . Lock ( )
overflow := len ( pc . m ) - parseCacheMaxLen
if overflow > 0 {
// Remove 10% of items from the cache.
overflow = int ( float64 ( len ( pc . m ) ) * 0.1 )
for k := range pc . m {
delete ( pc . m , k )
overflow --
if overflow <= 0 {
break
}
}
}
pc . m [ q ] = pcv
pc . mu . Unlock ( )
}