2019-05-22 21:16:55 +00:00
package storage
import (
"bytes"
"fmt"
"math/rand"
"os"
"reflect"
2020-05-28 08:58:47 +00:00
"regexp"
2020-11-04 22:15:43 +00:00
"sort"
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
"sync/atomic"
2019-05-22 21:16:55 +00:00
"testing"
"time"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
2019-10-08 13:25:24 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
2021-06-11 09:42:26 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
2021-02-21 20:06:45 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/mergeset"
2020-11-02 17:11:48 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/uint64set"
2019-08-13 18:35:19 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
"github.com/VictoriaMetrics/fastcache"
2019-05-22 21:16:55 +00:00
)
2022-10-23 09:15:24 +00:00
func TestMarshalUnmarshalMetricIDs ( t * testing . T ) {
f := func ( metricIDs [ ] uint64 ) {
t . Helper ( )
2024-01-23 14:09:52 +00:00
// Try marshaling and unmarshaling to an empty dst
2022-10-23 09:15:24 +00:00
data := marshalMetricIDs ( nil , metricIDs )
2024-01-23 14:09:52 +00:00
result := mustUnmarshalMetricIDs ( nil , data )
2022-10-23 09:15:24 +00:00
if ! reflect . DeepEqual ( result , metricIDs ) {
t . Fatalf ( "unexpected metricIDs after unmarshaling;\ngot\n%d\nwant\n%d" , result , metricIDs )
}
2024-01-23 14:09:52 +00:00
// Try marshaling and unmarshaling to non-empty dst
dataPrefix := [ ] byte ( "prefix" )
data = marshalMetricIDs ( dataPrefix , metricIDs )
if len ( data ) < len ( dataPrefix ) {
t . Fatalf ( "too short len(data)=%d; must be at least len(dataPrefix)=%d" , len ( data ) , len ( dataPrefix ) )
}
if string ( data [ : len ( dataPrefix ) ] ) != string ( dataPrefix ) {
t . Fatalf ( "unexpected prefix; got %q; want %q" , data [ : len ( dataPrefix ) ] , dataPrefix )
}
data = data [ len ( dataPrefix ) : ]
resultPrefix := [ ] uint64 { 889432422 , 89243 , 9823 }
result = mustUnmarshalMetricIDs ( resultPrefix , data )
if len ( result ) < len ( resultPrefix ) {
t . Fatalf ( "too short result returned; len(result)=%d; must be at least len(resultPrefix)=%d" , len ( result ) , len ( resultPrefix ) )
}
if ! reflect . DeepEqual ( result [ : len ( resultPrefix ) ] , resultPrefix ) {
t . Fatalf ( "unexpected result prefix; got %d; want %d" , result [ : len ( resultPrefix ) ] , resultPrefix )
}
result = result [ len ( resultPrefix ) : ]
if ( len ( metricIDs ) > 0 || len ( result ) > 0 ) && ! reflect . DeepEqual ( result , metricIDs ) {
t . Fatalf ( "unexpected metricIDs after unmarshaling from prefix;\ngot\n%d\nwant\n%d" , result , metricIDs )
}
2022-10-23 09:15:24 +00:00
}
2024-01-23 14:09:52 +00:00
2022-10-23 09:15:24 +00:00
f ( nil )
2024-01-23 14:09:52 +00:00
f ( [ ] uint64 { 0 } )
2022-10-23 09:15:24 +00:00
f ( [ ] uint64 { 1 } )
f ( [ ] uint64 { 1234 , 678932943 , 843289893843 } )
2024-01-23 14:09:52 +00:00
f ( [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 8989898 , 823849234 , 1 << 64 - 1 , 1 << 32 - 1 , 0 } )
2022-10-23 09:15:24 +00:00
}
func TestMergeSortedMetricIDs ( t * testing . T ) {
f := func ( a , b [ ] uint64 ) {
t . Helper ( )
m := make ( map [ uint64 ] bool )
var resultExpected [ ] uint64
for _ , v := range a {
if ! m [ v ] {
m [ v ] = true
resultExpected = append ( resultExpected , v )
}
}
for _ , v := range b {
if ! m [ v ] {
m [ v ] = true
resultExpected = append ( resultExpected , v )
}
}
sort . Slice ( resultExpected , func ( i , j int ) bool {
return resultExpected [ i ] < resultExpected [ j ]
} )
result := mergeSortedMetricIDs ( a , b )
if ! reflect . DeepEqual ( result , resultExpected ) {
t . Fatalf ( "unexpected result for mergeSortedMetricIDs(%d, %d); got\n%d\nwant\n%d" , a , b , result , resultExpected )
}
result = mergeSortedMetricIDs ( b , a )
if ! reflect . DeepEqual ( result , resultExpected ) {
t . Fatalf ( "unexpected result for mergeSortedMetricIDs(%d, %d); got\n%d\nwant\n%d" , b , a , result , resultExpected )
}
}
f ( nil , nil )
f ( [ ] uint64 { 1 } , nil )
f ( nil , [ ] uint64 { 23 } )
f ( [ ] uint64 { 1234 } , [ ] uint64 { 0 } )
f ( [ ] uint64 { 1 } , [ ] uint64 { 1 } )
f ( [ ] uint64 { 1 } , [ ] uint64 { 1 , 2 , 3 } )
f ( [ ] uint64 { 1 , 2 , 3 } , [ ] uint64 { 1 , 2 , 3 } )
f ( [ ] uint64 { 1 , 2 , 3 } , [ ] uint64 { 2 , 3 } )
f ( [ ] uint64 { 0 , 1 , 7 , 8 , 9 , 13 , 20 } , [ ] uint64 { 1 , 2 , 7 , 13 , 15 } )
f ( [ ] uint64 { 0 , 1 , 2 , 3 , 4 } , [ ] uint64 { 5 , 6 , 7 , 8 } )
f ( [ ] uint64 { 0 , 1 , 2 , 3 , 4 } , [ ] uint64 { 4 , 5 , 6 , 7 , 8 } )
f ( [ ] uint64 { 0 , 1 , 2 , 3 , 4 } , [ ] uint64 { 3 , 4 , 5 , 6 , 7 , 8 } )
f ( [ ] uint64 { 2 , 3 , 4 } , [ ] uint64 { 1 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 } , [ ] uint64 { 1 , 2 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 } , [ ] uint64 { 1 , 2 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 , 6 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 , 6 , 7 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 , 6 , 7 , 8 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 2 , 3 , 4 , 6 , 7 , 8 , 9 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 1 , 2 , 3 , 4 , 6 , 7 , 8 , 9 } , [ ] uint64 { 1 , 2 , 3 , 4 , 5 , 6 , 7 } )
f ( [ ] uint64 { 1 , 2 , 3 , 4 , 6 , 7 , 8 , 9 } , [ ] uint64 { 2 , 3 , 4 , 5 , 6 , 7 } )
2022-12-19 19:56:46 +00:00
f ( [ ] uint64 { } , [ ] uint64 { 1 , 2 , 3 } )
f ( [ ] uint64 { 0 } , [ ] uint64 { 1 , 2 , 3 } )
f ( [ ] uint64 { 1 } , [ ] uint64 { 1 , 2 , 3 } )
f ( [ ] uint64 { 1 , 2 } , [ ] uint64 { 3 , 4 } )
2022-10-23 09:15:24 +00:00
}
2020-05-27 18:35:58 +00:00
func TestReverseBytes ( t * testing . T ) {
f := func ( s , resultExpected string ) {
t . Helper ( )
result := reverseBytes ( nil , [ ] byte ( s ) )
if string ( result ) != resultExpected {
t . Fatalf ( "unexpected result for reverseBytes(%q); got %q; want %q" , s , result , resultExpected )
}
}
f ( "" , "" )
f ( "a" , "a" )
f ( "av" , "va" )
f ( "foo.bar" , "rab.oof" )
}
2019-10-08 13:25:24 +00:00
func TestMergeTagToMetricIDsRows ( t * testing . T ) {
f := func ( items [ ] string , expectedItems [ ] string ) {
t . Helper ( )
var data [ ] byte
2021-02-21 20:06:45 +00:00
var itemsB [ ] mergeset . Item
2019-10-08 13:25:24 +00:00
for _ , item := range items {
data = append ( data , item ... )
2021-02-21 20:06:45 +00:00
itemsB = append ( itemsB , mergeset . Item {
Start : uint32 ( len ( data ) - len ( item ) ) ,
End : uint32 ( len ( data ) ) ,
} )
2019-10-08 13:25:24 +00:00
}
2021-02-21 20:06:45 +00:00
if ! checkItemsSorted ( data , itemsB ) {
2019-11-06 12:24:48 +00:00
t . Fatalf ( "source items aren't sorted; items:\n%q" , itemsB )
2019-10-08 13:25:24 +00:00
}
resultData , resultItemsB := mergeTagToMetricIDsRows ( data , itemsB )
if len ( resultItemsB ) != len ( expectedItems ) {
t . Fatalf ( "unexpected len(resultItemsB); got %d; want %d" , len ( resultItemsB ) , len ( expectedItems ) )
}
2021-02-21 20:06:45 +00:00
if ! checkItemsSorted ( resultData , resultItemsB ) {
2019-11-06 12:24:48 +00:00
t . Fatalf ( "result items aren't sorted; items:\n%q" , resultItemsB )
2019-10-08 13:25:24 +00:00
}
2021-02-21 20:06:45 +00:00
buf := resultData
for i , it := range resultItemsB {
item := it . Bytes ( resultData )
if ! bytes . HasPrefix ( buf , item ) {
t . Fatalf ( "unexpected prefix for resultData #%d;\ngot\n%X\nwant\n%X" , i , buf , item )
2019-10-08 13:25:24 +00:00
}
2021-02-21 20:06:45 +00:00
buf = buf [ len ( item ) : ]
2019-10-08 13:25:24 +00:00
}
2021-02-21 20:06:45 +00:00
if len ( buf ) != 0 {
t . Fatalf ( "unexpected tail left in resultData: %X" , buf )
2019-10-08 13:25:24 +00:00
}
var resultItems [ ] string
2021-02-21 20:06:45 +00:00
for _ , it := range resultItemsB {
resultItems = append ( resultItems , string ( it . Bytes ( resultData ) ) )
2019-10-08 13:25:24 +00:00
}
if ! reflect . DeepEqual ( expectedItems , resultItems ) {
t . Fatalf ( "unexpected items;\ngot\n%X\nwant\n%X" , resultItems , expectedItems )
}
}
2019-11-09 21:17:42 +00:00
xy := func ( nsPrefix byte , accountID , projectID uint32 , key , value string , metricIDs [ ] uint64 ) string {
dst := marshalCommonPrefix ( nil , nsPrefix , accountID , projectID )
if nsPrefix == nsPrefixDateTagToMetricIDs {
dst = encoding . MarshalUint64 ( dst , 1234567901233 )
}
2019-10-08 13:25:24 +00:00
t := & Tag {
Key : [ ] byte ( key ) ,
Value : [ ] byte ( value ) ,
}
dst = t . Marshal ( dst )
for _ , metricID := range metricIDs {
dst = encoding . MarshalUint64 ( dst , metricID )
}
return string ( dst )
}
2019-11-09 21:17:42 +00:00
x := func ( accountID , projectID uint32 , key , value string , metricIDs [ ] uint64 ) string {
return xy ( nsPrefixTagToMetricIDs , accountID , projectID , key , value , metricIDs )
}
y := func ( accountID , projectID uint32 , key , value string , metricIDs [ ] uint64 ) string {
return xy ( nsPrefixDateTagToMetricIDs , accountID , projectID , key , value , metricIDs )
}
2019-10-08 13:25:24 +00:00
f ( nil , nil )
f ( [ ] string { } , nil )
f ( [ ] string { "foo" } , [ ] string { "foo" } )
f ( [ ] string { "a" , "b" , "c" , "def" } , [ ] string { "a" , "b" , "c" , "def" } )
f ( [ ] string { "\x00" , "\x00b" , "\x00c" , "\x00def" } , [ ] string { "\x00" , "\x00b" , "\x00c" , "\x00def" } )
f ( [ ] string {
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
} , [ ] string {
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
} )
2019-11-09 21:17:42 +00:00
f ( [ ] string {
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
} , [ ] string {
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 0 , 0 , "" , "" , [ ] uint64 { 0 } ) ,
} )
2019-10-08 13:25:24 +00:00
f ( [ ] string {
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
"xyz" ,
} , [ ] string {
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
"xyz" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
} , [ ] string {
"\x00asdf" ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
} )
f ( [ ] string {
"\x00asdf" ,
2019-11-09 21:17:42 +00:00
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
} , [ ] string {
"\x00asdf" ,
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 1 , 2 , "" , "" , [ ] uint64 { 0 } ) ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-10-08 13:25:24 +00:00
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-11-09 21:17:42 +00:00
"xyz" ,
} , [ ] string {
"\x00asdf" ,
2019-10-08 13:25:24 +00:00
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-11-09 21:17:42 +00:00
"xyz" ,
} )
f ( [ ] string {
"\x00asdf" ,
2019-10-08 13:25:24 +00:00
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-11-09 21:17:42 +00:00
y ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
y ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-10-08 13:25:24 +00:00
"xyz" ,
} , [ ] string {
"\x00asdf" ,
x ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-11-09 21:17:42 +00:00
y ( 3 , 1 , "" , "" , [ ] uint64 { 0 } ) ,
2019-10-08 13:25:24 +00:00
"xyz" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 4 , 2 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 4 , 2 , "" , "" , [ ] uint64 { 2 } ) ,
x ( 4 , 2 , "" , "" , [ ] uint64 { 3 } ) ,
x ( 4 , 2 , "" , "" , [ ] uint64 { 4 } ) ,
"xyz" ,
} , [ ] string {
"\x00asdf" ,
x ( 4 , 2 , "" , "" , [ ] uint64 { 1 , 2 , 3 , 4 } ) ,
"xyz" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 2 } ) ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 3 } ) ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 4 } ) ,
} , [ ] string {
"\x00asdf" ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 1 , 2 , 3 } ) ,
x ( 1 , 1 , "" , "" , [ ] uint64 { 4 } ) ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 2 , 2 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 2 , 2 , "" , "" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 2 , 2 , "" , "" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 2 , 2 , "" , "" , [ ] uint64 { 3 , 5 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 2 , 2 , "" , "" , [ ] uint64 { 1 , 2 , 3 , 4 , 5 } ) ,
"foo" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 3 , 3 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 3 , 3 , "" , "a" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 3 , 3 , "" , "a" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 3 , 3 , "" , "b" , [ ] uint64 { 3 , 5 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 3 , 3 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 3 , 3 , "" , "a" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 3 , 3 , "" , "b" , [ ] uint64 { 3 , 5 } ) ,
"foo" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 2 , 4 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 2 , 4 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 2 , 4 , "y" , "" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 2 , 4 , "y" , "x" , [ ] uint64 { 3 , 5 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 2 , 4 , "" , "" , [ ] uint64 { 1 } ) ,
x ( 2 , 4 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 2 , 4 , "y" , "" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 2 , 4 , "y" , "x" , [ ] uint64 { 3 , 5 } ) ,
"foo" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 2 , 4 , "x" , "a" , [ ] uint64 { 1 } ) ,
x ( 2 , 5 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 3 , 4 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 3 , 4 , "x" , "b" , [ ] uint64 { 3 , 5 } ) ,
x ( 3 , 4 , "x" , "b" , [ ] uint64 { 5 , 6 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 2 , 4 , "x" , "a" , [ ] uint64 { 1 } ) ,
x ( 2 , 5 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 } ) ,
x ( 3 , 4 , "x" , "a" , [ ] uint64 { 2 , 3 , 4 , 5 } ) ,
x ( 3 , 4 , "x" , "b" , [ ] uint64 { 3 , 5 , 6 } ) ,
"foo" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 2 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 1 , 3 } ) ,
x ( 2 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 2 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 2 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 2 , 3 } ) ,
"foo" ,
} )
f ( [ ] string {
"\x00asdf" ,
x ( 3 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 2 , 2 , 4 } ) ,
x ( 3 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 2 , 3 } ) ,
"foo" ,
} , [ ] string {
"\x00asdf" ,
x ( 3 , 2 , "sdf" , "aa" , [ ] uint64 { 1 , 2 , 3 , 4 } ) ,
"foo" ,
} )
// Construct big source chunks
var metricIDs [ ] uint64
metricIDs = metricIDs [ : 0 ]
for i := 0 ; i < maxMetricIDsPerRow - 1 ; i ++ {
metricIDs = append ( metricIDs , uint64 ( i ) )
}
f ( [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
2019-11-09 21:17:42 +00:00
y ( 2 , 3 , "foo" , "bar" , metricIDs ) ,
y ( 2 , 3 , "foo" , "bar" , metricIDs ) ,
2019-10-08 13:25:24 +00:00
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
2019-11-09 21:17:42 +00:00
y ( 2 , 3 , "foo" , "bar" , metricIDs ) ,
2019-10-08 13:25:24 +00:00
"x" ,
} )
metricIDs = metricIDs [ : 0 ]
for i := 0 ; i < maxMetricIDsPerRow ; i ++ {
metricIDs = append ( metricIDs , uint64 ( i ) )
}
f ( [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} )
metricIDs = metricIDs [ : 0 ]
for i := 0 ; i < 3 * maxMetricIDsPerRow ; i ++ {
metricIDs = append ( metricIDs , uint64 ( i ) )
}
f ( [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} )
f ( [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , [ ] uint64 { 0 , 0 , 1 , 2 , 3 } ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 3 , 2 , "foo" , "bar" , [ ] uint64 { 0 , 1 , 2 , 3 } ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 3 , 2 , "foo" , "bar" , metricIDs ) ,
"x" ,
} )
// Check for duplicate metricIDs removal
metricIDs = metricIDs [ : 0 ]
for i := 0 ; i < maxMetricIDsPerRow - 1 ; i ++ {
metricIDs = append ( metricIDs , 123 )
}
f ( [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
2019-11-09 21:17:42 +00:00
y ( 1 , 1 , "foo" , "bar" , metricIDs ) ,
2019-10-08 13:25:24 +00:00
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 } ) ,
2019-11-09 21:17:42 +00:00
y ( 1 , 1 , "foo" , "bar" , [ ] uint64 { 123 } ) ,
2019-10-08 13:25:24 +00:00
"x" ,
} )
// Check fallback to the original items after merging, which result in incorrect ordering.
metricIDs = metricIDs [ : 0 ]
for i := 0 ; i < maxMetricIDsPerRow - 3 ; i ++ {
metricIDs = append ( metricIDs , uint64 ( 123 ) )
}
f ( [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
"x" ,
} , [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
"x" ,
} )
f ( [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
2019-11-09 21:17:42 +00:00
y ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
2019-10-08 13:25:24 +00:00
} , [ ] string {
"\x00aa" ,
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
2019-11-09 21:17:42 +00:00
y ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
2019-10-08 13:25:24 +00:00
} )
f ( [ ] string {
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
} , [ ] string {
x ( 1 , 2 , "foo" , "bar" , metricIDs ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 123 , 125 } ) ,
x ( 1 , 2 , "foo" , "bar" , [ ] uint64 { 123 , 124 } ) ,
} )
}
2019-09-25 14:55:13 +00:00
func TestRemoveDuplicateMetricIDs ( t * testing . T ) {
f := func ( metricIDs , expectedMetricIDs [ ] uint64 ) {
2019-10-08 13:25:24 +00:00
t . Helper ( )
2019-09-25 14:55:13 +00:00
a := removeDuplicateMetricIDs ( metricIDs )
if ! reflect . DeepEqual ( a , expectedMetricIDs ) {
t . Fatalf ( "unexpected result from removeDuplicateMetricIDs:\ngot\n%d\nwant\n%d" , a , expectedMetricIDs )
}
}
f ( nil , nil )
f ( [ ] uint64 { 123 } , [ ] uint64 { 123 } )
f ( [ ] uint64 { 123 , 123 } , [ ] uint64 { 123 } )
f ( [ ] uint64 { 123 , 123 , 123 } , [ ] uint64 { 123 } )
2019-09-25 15:23:13 +00:00
f ( [ ] uint64 { 123 , 1234 , 1235 } , [ ] uint64 { 123 , 1234 , 1235 } )
2019-09-25 14:55:13 +00:00
f ( [ ] uint64 { 0 , 1 , 1 , 2 } , [ ] uint64 { 0 , 1 , 2 } )
f ( [ ] uint64 { 0 , 0 , 0 , 1 , 1 , 2 } , [ ] uint64 { 0 , 1 , 2 } )
f ( [ ] uint64 { 0 , 1 , 1 , 2 , 2 } , [ ] uint64 { 0 , 1 , 2 } )
f ( [ ] uint64 { 0 , 1 , 2 , 2 } , [ ] uint64 { 0 , 1 , 2 } )
}
2019-05-22 21:16:55 +00:00
func TestIndexDBOpenClose ( t * testing . T ) {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
var s Storage
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
tableName := nextIndexDBTableName ( )
2019-05-22 21:16:55 +00:00
for i := 0 ; i < 5 ; i ++ {
2024-02-23 21:29:23 +00:00
var isReadOnly atomic . Bool
2023-07-22 22:20:21 +00:00
db := mustOpenIndexDB ( tableName , & s , & isReadOnly )
2019-05-22 21:16:55 +00:00
db . MustClose ( )
}
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
if err := os . RemoveAll ( tableName ) ; err != nil {
2019-05-22 21:16:55 +00:00
t . Fatalf ( "cannot remove indexDB: %s" , err )
}
}
func TestIndexDB ( t * testing . T ) {
const accountsCount = 3
const projectsCount = 2
const metricGroups = 10
t . Run ( "serial" , func ( t * testing . T ) {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
const path = "TestIndexDB-serial"
2023-09-01 07:27:51 +00:00
s := MustOpenStorage ( path , retentionMax , 0 , 0 )
2019-05-22 21:16:55 +00:00
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db := s . idb ( )
2022-11-25 18:32:45 +00:00
mns , tsids , tenants , err := testIndexDBGetOrCreateTSIDByName ( db , accountsCount , projectsCount , metricGroups )
2019-05-22 21:16:55 +00:00
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
2022-11-25 18:32:45 +00:00
if err := testIndexDBCheckTSIDByName ( db , mns , tsids , tenants , false ) ; err != nil {
2019-05-22 21:16:55 +00:00
t . Fatalf ( "unexpected error: %s" , err )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Re-open the storage and verify it works as expected.
s . MustClose ( )
2023-09-01 07:27:51 +00:00
s = MustOpenStorage ( path , retentionMax , 0 , 0 )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db = s . idb ( )
2022-11-25 18:32:45 +00:00
if err := testIndexDBCheckTSIDByName ( db , mns , tsids , tenants , false ) ; err != nil {
2019-05-22 21:16:55 +00:00
t . Fatalf ( "unexpected error: %s" , err )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
s . MustClose ( )
fs . MustRemoveAll ( path )
2019-05-22 21:16:55 +00:00
} )
t . Run ( "concurrent" , func ( t * testing . T ) {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
const path = "TestIndexDB-concurrent"
2023-09-01 07:27:51 +00:00
s := MustOpenStorage ( path , retentionMax , 0 , 0 )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db := s . idb ( )
2019-05-22 21:16:55 +00:00
ch := make ( chan error , 3 )
for i := 0 ; i < cap ( ch ) ; i ++ {
go func ( ) {
2022-11-25 18:32:45 +00:00
mns , tsid , tenants , err := testIndexDBGetOrCreateTSIDByName ( db , accountsCount , projectsCount , metricGroups )
2019-05-22 21:16:55 +00:00
if err != nil {
ch <- err
return
}
2022-11-25 18:32:45 +00:00
if err := testIndexDBCheckTSIDByName ( db , mns , tsid , tenants , true ) ; err != nil {
2019-05-22 21:16:55 +00:00
ch <- err
return
}
ch <- nil
} ( )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
deadlineCh := time . After ( 30 * time . Second )
2019-05-22 21:16:55 +00:00
for i := 0 ; i < cap ( ch ) ; i ++ {
select {
case err := <- ch :
if err != nil {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
t . Fatalf ( "unexpected error: %s" , err )
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
case <- deadlineCh :
2019-05-22 21:16:55 +00:00
t . Fatalf ( "timeout" )
}
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
s . MustClose ( )
fs . MustRemoveAll ( path )
2019-05-22 21:16:55 +00:00
} )
}
2022-11-25 18:32:45 +00:00
func testIndexDBGetOrCreateTSIDByName ( db * indexDB , accountsCount , projectsCount , metricGroups int ) ( [ ] MetricName , [ ] TSID , [ ] string , error ) {
2023-01-24 04:10:29 +00:00
r := rand . New ( rand . NewSource ( 1 ) )
2019-05-22 21:16:55 +00:00
// Create tsids.
var mns [ ] MetricName
var tsids [ ] TSID
2022-11-25 18:32:45 +00:00
tenants := make ( map [ string ] struct { } )
2019-05-22 21:16:55 +00:00
2020-07-23 21:30:33 +00:00
is := db . getIndexSearch ( 0 , 0 , noDeadline )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
date := uint64 ( timestampFromTime ( time . Now ( ) ) ) / msecPerDay
2019-05-22 21:16:55 +00:00
2019-09-23 17:40:38 +00:00
var metricNameBuf [ ] byte
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
for i := 0 ; i < 401 ; i ++ {
2019-05-22 21:16:55 +00:00
var mn MetricName
2019-05-22 21:23:23 +00:00
mn . AccountID = uint32 ( ( i + 2 ) % accountsCount )
mn . ProjectID = uint32 ( ( i + 1 ) % projectsCount )
2022-11-25 18:32:45 +00:00
tenant := fmt . Sprintf ( "%d:%d" , mn . AccountID , mn . ProjectID )
tenants [ tenant ] = struct { } { }
2019-05-22 21:16:55 +00:00
// Init MetricGroup.
2020-05-27 18:35:58 +00:00
mn . MetricGroup = [ ] byte ( fmt . Sprintf ( "metricGroup.%d\x00\x01\x02" , i % metricGroups ) )
2019-05-22 21:16:55 +00:00
// Init other tags.
2023-01-24 04:10:29 +00:00
tagsCount := r . Intn ( 10 ) + 1
2019-05-22 21:16:55 +00:00
for j := 0 ; j < tagsCount ; j ++ {
key := fmt . Sprintf ( "key\x01\x02\x00_%d_%d" , i , j )
value := fmt . Sprintf ( "val\x01_%d\x00_%d\x02" , i , j )
mn . AddTag ( key , value )
}
mn . sortTags ( )
2019-09-23 17:40:38 +00:00
metricNameBuf = mn . Marshal ( metricNameBuf [ : 0 ] )
2019-05-22 21:16:55 +00:00
// Create tsid for the metricName.
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
var genTSID generationTSID
if ! is . getTSIDByMetricName ( & genTSID , metricNameBuf , date ) {
generateTSID ( & genTSID . TSID , & mn )
2023-07-14 06:13:21 +00:00
createAllIndexesForMetricName ( is , & mn , & genTSID . TSID , date )
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if genTSID . TSID . AccountID != mn . AccountID {
return nil , nil , nil , fmt . Errorf ( "unexpected TSID.AccountID; got %d; want %d; mn:\n%s\ntsid:\n%+v" , genTSID . TSID . AccountID , mn . AccountID , & mn , & genTSID . TSID )
2019-05-22 21:23:23 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if genTSID . TSID . ProjectID != mn . ProjectID {
return nil , nil , nil , fmt . Errorf ( "unexpected TSID.ProjectID; got %d; want %d; mn:\n%s\ntsid:\n%+v" , genTSID . TSID . ProjectID , mn . ProjectID , & mn , & genTSID . TSID )
2019-05-22 21:23:23 +00:00
}
2019-05-22 21:16:55 +00:00
mns = append ( mns , mn )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
tsids = append ( tsids , genTSID . TSID )
2019-06-25 09:55:27 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db . putIndexSearch ( is )
2019-09-23 17:40:38 +00:00
// Flush index to disk, so it becomes visible for search
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db . s . DebugFlush ( )
2019-06-25 09:55:27 +00:00
2022-11-25 18:32:45 +00:00
var tenantsList [ ] string
for tenant := range tenants {
tenantsList = append ( tenantsList , tenant )
}
sort . Strings ( tenantsList )
return mns , tsids , tenantsList , nil
2019-09-23 17:40:38 +00:00
}
2022-11-25 18:32:45 +00:00
func testIndexDBCheckTSIDByName ( db * indexDB , mns [ ] MetricName , tsids [ ] TSID , tenants [ ] string , isConcurrent bool ) error {
2022-06-12 01:32:13 +00:00
hasValue := func ( lvs [ ] string , v [ ] byte ) bool {
for _ , lv := range lvs {
if string ( v ) == lv {
2019-05-22 21:16:55 +00:00
return true
}
}
return false
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
currentTime := timestampFromTime ( time . Now ( ) )
2022-06-12 01:32:13 +00:00
allLabelNames := make ( map [ accountProjectKey ] map [ string ] bool )
2019-05-22 21:23:23 +00:00
timeseriesCounters := make ( map [ accountProjectKey ] map [ uint64 ] bool )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
var genTSID generationTSID
2019-05-22 21:16:55 +00:00
var metricNameCopy [ ] byte
for i := range mns {
mn := & mns [ i ]
tsid := & tsids [ i ]
2019-05-22 21:23:23 +00:00
apKey := accountProjectKey {
AccountID : tsid . AccountID ,
ProjectID : tsid . ProjectID ,
}
tc := timeseriesCounters [ apKey ]
if tc == nil {
tc = make ( map [ uint64 ] bool )
timeseriesCounters [ apKey ] = tc
}
2019-05-22 21:16:55 +00:00
tc [ tsid . MetricID ] = true
mn . sortTags ( )
metricName := mn . Marshal ( nil )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
is := db . getIndexSearch ( 0 , 0 , noDeadline )
if ! is . getTSIDByMetricName ( & genTSID , metricName , uint64 ( currentTime ) / msecPerDay ) {
return fmt . Errorf ( "cannot obtain tsid #%d for mn %s" , i , mn )
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db . putIndexSearch ( is )
2019-05-22 21:16:55 +00:00
if isConcurrent {
// Copy tsid.MetricID, since multiple TSIDs may match
// the same mn in concurrent mode.
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
genTSID . TSID . MetricID = tsid . MetricID
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if ! reflect . DeepEqual ( tsid , & genTSID . TSID ) {
return fmt . Errorf ( "unexpected tsid for mn:\n%s\ngot\n%+v\nwant\n%+v" , mn , & genTSID . TSID , tsid )
2019-05-22 21:16:55 +00:00
}
// Search for metric name for the given metricID.
2023-09-22 09:32:59 +00:00
var ok bool
metricNameCopy , ok = db . searchMetricNameWithCache ( metricNameCopy [ : 0 ] , genTSID . TSID . MetricID , genTSID . TSID . AccountID , genTSID . TSID . ProjectID )
if ! ok {
return fmt . Errorf ( "cannot find metricName for metricID=%d; i=%d" , genTSID . TSID . MetricID , i )
2019-05-22 21:16:55 +00:00
}
if ! bytes . Equal ( metricName , metricNameCopy ) {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
return fmt . Errorf ( "unexpected mn for metricID=%d;\ngot\n%q\nwant\n%q" , genTSID . TSID . MetricID , metricNameCopy , metricName )
2019-05-22 21:16:55 +00:00
}
// Try searching metric name for non-existent MetricID.
2023-09-22 09:32:59 +00:00
buf , found := db . searchMetricNameWithCache ( nil , 1 , mn . AccountID , mn . ProjectID )
if found {
return fmt . Errorf ( "unexpected metricName found for non-existing metricID; got %X" , buf )
2019-05-22 21:16:55 +00:00
}
if len ( buf ) > 0 {
return fmt . Errorf ( "expecting empty buf when searching for non-existent metricID; got %X" , buf )
}
2022-06-12 01:32:13 +00:00
// Test SearchLabelValuesWithFiltersOnTimeRange
lvs , err := db . SearchLabelValuesWithFiltersOnTimeRange ( nil , mn . AccountID , mn . ProjectID , "__name__" , nil , TimeRange { } , 1e5 , 1e9 , noDeadline )
2019-05-22 21:16:55 +00:00
if err != nil {
2022-06-12 01:32:13 +00:00
return fmt . Errorf ( "error in SearchLabelValuesWithFiltersOnTimeRange(labelName=%q): %w" , "__name__" , err )
2019-05-22 21:16:55 +00:00
}
2022-06-12 01:32:13 +00:00
if ! hasValue ( lvs , mn . MetricGroup ) {
return fmt . Errorf ( "SearchLabelValuesWithFiltersOnTimeRange(labelName=%q): couldn't find %q; found %q" , "__name__" , mn . MetricGroup , lvs )
2019-05-22 21:16:55 +00:00
}
2022-06-12 01:32:13 +00:00
labelNames := allLabelNames [ apKey ]
if labelNames == nil {
labelNames = make ( map [ string ] bool )
allLabelNames [ apKey ] = labelNames
2019-05-22 21:23:23 +00:00
}
2019-05-22 21:16:55 +00:00
for i := range mn . Tags {
tag := & mn . Tags [ i ]
2022-06-12 01:32:13 +00:00
lvs , err := db . SearchLabelValuesWithFiltersOnTimeRange ( nil , mn . AccountID , mn . ProjectID , string ( tag . Key ) , nil , TimeRange { } , 1e5 , 1e9 , noDeadline )
2019-05-22 21:16:55 +00:00
if err != nil {
2022-06-12 01:32:13 +00:00
return fmt . Errorf ( "error in SearchLabelValuesWithFiltersOnTimeRange(labelName=%q): %w" , tag . Key , err )
2019-05-22 21:16:55 +00:00
}
2022-06-12 01:32:13 +00:00
if ! hasValue ( lvs , tag . Value ) {
return fmt . Errorf ( "SearchLabelValuesWithFiltersOnTimeRange(labelName=%q): couldn't find %q; found %q" , tag . Key , tag . Value , lvs )
2019-05-22 21:16:55 +00:00
}
2022-06-12 01:32:13 +00:00
labelNames [ string ( tag . Key ) ] = true
2019-05-22 21:16:55 +00:00
}
}
2022-06-12 01:32:13 +00:00
// Test SearchLabelNamesWithFiltersOnTimeRange (empty filters, global time range)
for k , labelNames := range allLabelNames {
lns , err := db . SearchLabelNamesWithFiltersOnTimeRange ( nil , k . AccountID , k . ProjectID , nil , TimeRange { } , 1e5 , 1e9 , noDeadline )
2019-05-22 21:23:23 +00:00
if err != nil {
2022-11-25 18:32:45 +00:00
return fmt . Errorf ( "error in SearchLabelNamesWithFiltersOnTimeRange: %w" , err )
2019-05-22 21:23:23 +00:00
}
2022-06-12 01:32:13 +00:00
if ! hasValue ( lns , [ ] byte ( "__name__" ) ) {
return fmt . Errorf ( "cannot find __name__ in %q" , lns )
2019-05-22 21:23:23 +00:00
}
2022-06-12 01:32:13 +00:00
for labelName := range labelNames {
if ! hasValue ( lns , [ ] byte ( labelName ) ) {
return fmt . Errorf ( "cannot find %q in %q" , labelName , lns )
2019-05-22 21:23:23 +00:00
}
2019-05-22 21:16:55 +00:00
}
}
2022-11-25 18:32:45 +00:00
// Test SearchTenants on global time range
tenantsGot , err := db . SearchTenants ( nil , TimeRange { } , noDeadline )
if err != nil {
return fmt . Errorf ( "error in SearchTenants: %w" , err )
}
sort . Strings ( tenantsGot )
if ! reflect . DeepEqual ( tenants , tenantsGot ) {
return fmt . Errorf ( "unexpected tenants got when searching in global time range;\ngot\n%s\nwant\n%s" , tenantsGot , tenants )
}
// Test SearchTenants on specific time range
tr := TimeRange {
MinTimestamp : currentTime - msecPerDay ,
MaxTimestamp : currentTime + msecPerDay ,
}
tenantsGot , err = db . SearchTenants ( nil , tr , noDeadline )
if err != nil {
return fmt . Errorf ( "error in SearchTenants: %w" , err )
}
sort . Strings ( tenantsGot )
if ! reflect . DeepEqual ( tenants , tenantsGot ) {
return fmt . Errorf ( "unexpected tenants got when searching in global time range;\ngot\n%s\nwant\n%s" , tenantsGot , tenants )
}
2019-06-10 09:36:42 +00:00
// Check timerseriesCounters only for serial test.
// Concurrent test may create duplicate timeseries, so GetSeriesCount
// would return more timeseries than needed.
if ! isConcurrent {
for k , tc := range timeseriesCounters {
2020-07-23 17:42:57 +00:00
n , err := db . GetSeriesCount ( k . AccountID , k . ProjectID , noDeadline )
2019-06-10 09:36:42 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "unexpected error in GetSeriesCount(%v): %w" , k , err )
2019-06-10 09:36:42 +00:00
}
if n != uint64 ( len ( tc ) ) {
return fmt . Errorf ( "unexpected GetSeriesCount(%v); got %d; want %d" , k , n , uint64 ( len ( tc ) ) )
}
}
}
2019-05-22 21:16:55 +00:00
// Try tag filters.
for i := range mns {
mn := & mns [ i ]
tsid := & tsids [ i ]
// Search without regexps.
2019-05-22 21:23:23 +00:00
tfs := NewTagFilters ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , mn . MetricGroup , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for MetricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
for j := 0 ; j < len ( mn . Tags ) ; j ++ {
t := & mn . Tags [ j ]
if err := tfs . Add ( t . Key , t . Value , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for tag: %w" , err )
2019-05-22 21:16:55 +00:00
}
}
if err := tfs . Add ( nil , [ ] byte ( "foobar" ) , true , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add negative filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
if err := tfs . Add ( nil , nil , true , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add no-op negative filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err := searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by exact tag filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
if ! testHasTSID ( tsidsFound , tsid ) {
2019-09-23 17:40:38 +00:00
return fmt . Errorf ( "tsids is missing in exact tsidsFound\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s\ni=%d" , tsid , tsidsFound , tfs , mn , i )
2019-05-22 21:16:55 +00:00
}
// Verify tag cache.
2022-10-23 09:15:24 +00:00
tsidsCached , err := searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by exact tag filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
if ! reflect . DeepEqual ( tsidsCached , tsidsFound ) {
return fmt . Errorf ( "unexpected tsids returned; got\n%+v; want\n%+v" , tsidsCached , tsidsFound )
}
// Add negative filter for zeroing search results.
if err := tfs . Add ( nil , mn . MetricGroup , true , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add negative filter for zeroing search results: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by exact tag filter with full negative: %w" , err )
2019-05-22 21:16:55 +00:00
}
if testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "unexpected tsid found for exact negative filter\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
2020-05-28 08:58:47 +00:00
// Search for Graphite wildcard
tfs . Reset ( mn . AccountID , mn . ProjectID )
n := bytes . IndexByte ( mn . MetricGroup , '.' )
if n < 0 {
return fmt . Errorf ( "cannot find dot in MetricGroup %q" , mn . MetricGroup )
}
re := "[^.]*" + regexp . QuoteMeta ( string ( mn . MetricGroup [ n : ] ) )
if err := tfs . Add ( nil , [ ] byte ( re ) , false , true ) ; err != nil {
return fmt . Errorf ( "cannot create regexp tag filter for Graphite wildcard" )
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2020-05-28 08:58:47 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by regexp tag filter for Graphite wildcard: %w" , err )
2020-05-28 08:58:47 +00:00
}
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing in regexp for Graphite wildcard tsidsFound\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
2021-09-09 18:09:18 +00:00
// Search with a filter matching empty tag (a single filter)
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1601
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , mn . MetricGroup , false , false ) ; err != nil {
return fmt . Errorf ( "cannot create tag filter for MetricGroup: %w" , err )
}
if err := tfs . Add ( [ ] byte ( "non-existent-tag" ) , [ ] byte ( "foo|" ) , false , true ) ; err != nil {
return fmt . Errorf ( "cannot create regexp tag filter for non-existing tag: %w" , err )
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2021-09-09 18:09:18 +00:00
if err != nil {
return fmt . Errorf ( "cannot search with a filter matching empty tag: %w" , err )
}
2021-09-11 07:57:13 +00:00
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing when matching a filter with empty tag tsidsFound\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
2021-09-09 18:09:18 +00:00
// Search with filters matching empty tags (multiple filters)
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1601
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , mn . MetricGroup , false , false ) ; err != nil {
return fmt . Errorf ( "cannot create tag filter for MetricGroup: %w" , err )
}
if err := tfs . Add ( [ ] byte ( "non-existent-tag1" ) , [ ] byte ( "foo|" ) , false , true ) ; err != nil {
return fmt . Errorf ( "cannot create regexp tag filter for non-existing tag1: %w" , err )
}
if err := tfs . Add ( [ ] byte ( "non-existent-tag2" ) , [ ] byte ( "bar|" ) , false , true ) ; err != nil {
return fmt . Errorf ( "cannot create regexp tag filter for non-existing tag2: %w" , err )
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2021-09-09 18:09:18 +00:00
if err != nil {
return fmt . Errorf ( "cannot search with multipel filters matching empty tags: %w" , err )
}
2021-09-11 07:57:13 +00:00
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing when matching multiple filters with empty tags tsidsFound\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
2021-09-09 18:09:18 +00:00
2019-05-22 21:16:55 +00:00
// Search with regexps.
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , mn . MetricGroup , false , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create regexp tag filter for MetricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
for j := 0 ; j < len ( mn . Tags ) ; j ++ {
t := & mn . Tags [ j ]
if err := tfs . Add ( t . Key , append ( t . Value , "|foo*." ... ) , false , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create regexp tag filter for tag: %w" , err )
2019-05-22 21:16:55 +00:00
}
if err := tfs . Add ( t . Key , append ( t . Value , "|aaa|foo|bar" ... ) , false , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create regexp tag filter for tag: %w" , err )
2019-05-22 21:16:55 +00:00
}
}
if err := tfs . Add ( nil , [ ] byte ( "^foobar$" ) , true , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add negative filter with regexp: %w" , err )
2019-05-22 21:16:55 +00:00
}
if err := tfs . Add ( nil , nil , true , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add no-op negative filter with regexp: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by regexp tag filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing in regexp tsidsFound\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
if err := tfs . Add ( nil , mn . MetricGroup , true , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add negative filter for zeroing search results: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by regexp tag filter with full negative: %w" , err )
2019-05-22 21:16:55 +00:00
}
if testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "unexpected tsid found for regexp negative filter\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
// Search with filter matching zero results.
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-key" ) , [ ] byte ( "foobar" ) , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot add non-existing key: %w" , err )
2019-05-22 21:16:55 +00:00
}
if err := tfs . Add ( nil , mn . MetricGroup , false , true ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for MetricGroup matching zero results: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search by non-existing tag filter: %w" , err )
2019-05-22 21:16:55 +00:00
}
if len ( tsidsFound ) > 0 {
return fmt . Errorf ( "non-zero tsidsFound for non-existing tag filter: %+v" , tsidsFound )
}
if isConcurrent {
// Skip empty filter search in concurrent mode, since it looks like
// it has a lag preventing from correct work.
continue
}
2022-10-23 09:15:24 +00:00
// Search with empty filter. It should match all the results.
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search for common prefix: %w" , err )
2019-05-22 21:16:55 +00:00
}
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing in common prefix\ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
// Search with empty metricGroup. It should match zero results.
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , nil , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for empty metricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search for empty metricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
if len ( tsidsFound ) != 0 {
return fmt . Errorf ( "unexpected non-empty tsids found for empty metricGroup: %v" , tsidsFound )
}
// Search with multiple tfss
2019-05-22 21:23:23 +00:00
tfs1 := NewTagFilters ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs1 . Add ( nil , nil , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for empty metricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
2019-05-22 21:23:23 +00:00
tfs2 := NewTagFilters ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs2 . Add ( nil , mn . MetricGroup , false , false ) ; err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot create tag filter for MetricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs1 , tfs2 } , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search for empty metricGroup: %w" , err )
2019-05-22 21:16:55 +00:00
}
if ! testHasTSID ( tsidsFound , tsid ) {
return fmt . Errorf ( "tsids is missing when searching for multiple tfss \ntsid=%+v\ntsidsFound=%+v\ntfs=%s\nmn=%s" , tsid , tsidsFound , tfs , mn )
}
// Verify empty tfss
2022-10-23 09:15:24 +00:00
tsidsFound , err = searchTSIDsInTest ( db , nil , tr )
2019-05-22 21:16:55 +00:00
if err != nil {
2020-06-30 19:58:18 +00:00
return fmt . Errorf ( "cannot search for nil tfss: %w" , err )
2019-05-22 21:16:55 +00:00
}
if len ( tsidsFound ) != 0 {
2019-05-22 21:23:23 +00:00
return fmt . Errorf ( "unexpected non-empty tsids fround for nil tfss" )
2019-05-22 21:16:55 +00:00
}
}
return nil
}
2022-10-23 09:15:24 +00:00
func searchTSIDsInTest ( db * indexDB , tfss [ ] * TagFilters , tr TimeRange ) ( [ ] TSID , error ) {
metricIDs , err := db . searchMetricIDs ( nil , tfss , tr , 1e5 , noDeadline )
if err != nil {
return nil , err
}
if len ( tfss ) == 0 {
if len ( metricIDs ) > 0 {
return nil , fmt . Errorf ( "expecting empty metricIDs for non-empty tfss; got %d metricIDs" , len ( metricIDs ) )
}
return nil , nil
}
accountID := tfss [ 0 ] . accountID
projectID := tfss [ 0 ] . projectID
return db . getTSIDsFromMetricIDs ( nil , accountID , projectID , metricIDs , noDeadline )
}
2019-05-22 21:16:55 +00:00
func testHasTSID ( tsids [ ] TSID , tsid * TSID ) bool {
for i := range tsids {
if tsids [ i ] == * tsid {
return true
}
}
return false
}
func TestMatchTagFilters ( t * testing . T ) {
var mn MetricName
2019-05-22 21:23:23 +00:00
mn . AccountID = 123
mn . ProjectID = 456
2019-05-22 21:16:55 +00:00
mn . MetricGroup = append ( mn . MetricGroup , "foobar_metric" ... )
for i := 0 ; i < 5 ; i ++ {
key := fmt . Sprintf ( "key %d" , i )
value := fmt . Sprintf ( "value %d" , i )
mn . AddTag ( key , value )
}
var bb bytesutil . ByteBuffer
2019-05-22 21:23:23 +00:00
// Verify tag filters for different account / project
tfs := NewTagFilters ( mn . AccountID , mn . ProjectID + 1 )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter: %s" , err )
}
ok , err := matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
2019-05-22 21:23:23 +00:00
if ok {
t . Fatalf ( "Tag filters shouldn't match for invalid projectID" )
}
tfs . Reset ( mn . AccountID + 1 , mn . ProjectID )
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Tag filters shouldn't match for invalid accountID" )
}
// Correct AccountID , ProjectID
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
2019-05-22 21:16:55 +00:00
if ! ok {
t . Fatalf ( "should match" )
}
// Empty tag filters should match.
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "empty tag filters should match" )
}
// Negative match by MetricGroup
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foobar" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "obar.+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foob.+metric" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2020-04-01 14:40:18 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , [ ] byte ( ".+" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:16:55 +00:00
// Positive match by MetricGroup
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "foobar.+etric" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "obar_metric" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( nil , [ ] byte ( "ob.+metric" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2020-04-01 14:40:18 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , [ ] byte ( ".+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, positive filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:16:55 +00:00
2022-03-18 10:58:22 +00:00
// Positive empty match by non-existing tag
2022-03-18 14:15:01 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2022-03-18 10:58:22 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( "foobar|" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, positive filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:16:55 +00:00
// Negative match by non-existing tag
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( "foobar" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( "obar.+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( "foobar_metric" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
2019-07-30 12:14:09 +00:00
if ! ok {
t . Fatalf ( "Should match" )
2019-05-22 21:16:55 +00:00
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( "foob.+metric" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
2019-07-30 12:14:09 +00:00
if ! ok {
t . Fatalf ( "Should match" )
}
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( ".+" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
2019-05-22 21:16:55 +00:00
}
2020-04-01 14:40:18 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "non-existing-tag" ) , [ ] byte ( ".+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, non-negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:16:55 +00:00
// Negative match by existing tag
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 0" ) , [ ] byte ( "foobar" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 1" ) , [ ] byte ( "obar.+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "value 2" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "v.+lue 3" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2020-04-01 14:40:18 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( ".+" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:16:55 +00:00
2020-06-10 15:40:00 +00:00
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/546
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "|value 3" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "|value 2" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:16:55 +00:00
// Positive match by existing tag
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 0" ) , [ ] byte ( "value 0" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 1" ) , [ ] byte ( ".+lue 1" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "value 3" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2020-04-01 14:40:18 +00:00
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "v.+lue 2|" ) , true , true ) ; err != nil {
2019-05-22 21:16:55 +00:00
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
2019-07-30 12:14:09 +00:00
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
2019-05-22 21:16:55 +00:00
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2020-04-01 14:40:18 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( ".+" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, non-negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
2019-05-22 21:16:55 +00:00
// Positive match by multiple tags and MetricGroup
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
if err := tfs . Add ( [ ] byte ( "key 0" ) , [ ] byte ( "value 0" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "value [0-9]" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "value 23" ) , true , false ) ; err != nil {
t . Fatalf ( "cannt add no regexp, negative filter: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "lue.+43" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "foo.+metric" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "sdfdsf" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "o.+metr" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! ok {
t . Fatalf ( "Should match" )
}
// Negative match by multiple tags and MetricGroup
2019-05-22 21:23:23 +00:00
tfs . Reset ( mn . AccountID , mn . ProjectID )
2019-05-22 21:16:55 +00:00
// Positive matches
if err := tfs . Add ( [ ] byte ( "key 0" ) , [ ] byte ( "value 0" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "value [0-9]" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "key 3" ) , [ ] byte ( "value 23" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
// Negative matches
if err := tfs . Add ( [ ] byte ( "key 2" ) , [ ] byte ( "v.+2" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "obar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, no negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "oo.+metric" ) , false , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, no negative filter: %s" , err )
}
// Positive matches
if err := tfs . Add ( nil , [ ] byte ( "sdfdsf" ) , true , false ) ; err != nil {
t . Fatalf ( "cannot add no regexp, negative filter: %s" , err )
}
if err := tfs . Add ( nil , [ ] byte ( "o.+metr" ) , true , true ) ; err != nil {
t . Fatalf ( "cannot add regexp, negative filter: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2020-04-01 14:40:18 +00:00
// Negative match for multiple non-regexp positive filters
tfs . Reset ( mn . AccountID , mn . ProjectID )
if err := tfs . Add ( nil , [ ] byte ( "foobar_metric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add non-regexp positive filter for MetricGroup: %s" , err )
}
if err := tfs . Add ( [ ] byte ( "non-existing-metric" ) , [ ] byte ( "foobar" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add non-regexp positive filter for non-existing tag: %s" , err )
}
ok , err = matchTagFilters ( & mn , toTFPointers ( tfs . tfs ) , & bb )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ok {
t . Fatalf ( "Shouldn't match" )
}
2019-05-22 21:16:55 +00:00
}
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
func TestIndexDBRepopulateAfterRotation ( t * testing . T ) {
2023-01-24 04:10:29 +00:00
r := rand . New ( rand . NewSource ( 1 ) )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
path := "TestIndexRepopulateAfterRotation"
2023-09-03 08:33:37 +00:00
s := MustOpenStorage ( path , retention31Days , 1e5 , 1e5 )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
db := s . idb ( )
if db . generation == 0 {
t . Fatalf ( "expected indexDB generation to be not 0" )
}
const metricRowsN = 1000
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
currentDayTimestamp := ( time . Now ( ) . UnixMilli ( ) / msecPerDay ) * msecPerDay
timeMin := currentDayTimestamp - 24 * 3600 * 1000
timeMax := currentDayTimestamp + 24 * 3600 * 1000
mrs := testGenerateMetricRows ( r , metricRowsN , timeMin , timeMax )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
if err := s . AddRows ( mrs , defaultPrecisionBits ) ; err != nil {
t . Fatalf ( "unexpected error when adding mrs: %s" , err )
}
s . DebugFlush ( )
// verify the storage contains rows.
var m Metrics
s . UpdateMetrics ( & m )
2022-12-05 23:27:57 +00:00
if rowsCount := m . TableMetrics . TotalRowsCount ( ) ; rowsCount < uint64 ( metricRowsN ) {
t . Fatalf ( "expecting at least %d rows in the table; got %d" , metricRowsN , rowsCount )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
}
// check new series were registered in indexDB
2024-02-23 22:15:21 +00:00
added := db . s . newTimeseriesCreated . Load ( )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
if added != metricRowsN {
t . Fatalf ( "expected indexDB to contain %d rows; got %d" , metricRowsN , added )
}
// check new series were added to cache
var cs fastcache . Stats
s . tsidCache . UpdateStats ( & cs )
if cs . EntriesCount != metricRowsN {
t . Fatalf ( "expected tsidCache to contain %d rows; got %d" , metricRowsN , cs . EntriesCount )
}
// check if cache entries do belong to current indexDB generation
var genTSID generationTSID
for _ , mr := range mrs {
s . getTSIDFromCache ( & genTSID , mr . MetricNameRaw )
if genTSID . generation != db . generation {
t . Fatalf ( "expected all entries in tsidCache to have the same indexDB generation: %d;" +
"got %d" , db . generation , genTSID . generation )
}
}
prevGeneration := db . generation
// force index rotation
2023-07-29 02:47:02 +00:00
s . mustRotateIndexDB ( time . Now ( ) )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
// check tsidCache wasn't reset after the rotation
var cs2 fastcache . Stats
s . tsidCache . UpdateStats ( & cs2 )
if cs . EntriesCount != metricRowsN {
t . Fatalf ( "expected tsidCache after rotation to contain %d rows; got %d" , metricRowsN , cs2 . EntriesCount )
}
dbNew := s . idb ( )
if dbNew . generation == 0 {
t . Fatalf ( "expected new indexDB generation to be not 0" )
}
if dbNew . generation == prevGeneration {
t . Fatalf ( "expected new indexDB generation %d to be different from prev indexDB" , dbNew . generation )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Re-insert rows again and verify that all the entries belong to new generation
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
if err := s . AddRows ( mrs , defaultPrecisionBits ) ; err != nil {
t . Fatalf ( "unexpected error when adding mrs: %s" , err )
}
s . DebugFlush ( )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
for _ , mr := range mrs {
s . getTSIDFromCache ( & genTSID , mr . MetricNameRaw )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if genTSID . generation != dbNew . generation {
t . Fatalf ( "unexpected generation for data after rotation; got %d; want %d" , genTSID . generation , dbNew . generation )
}
2023-05-18 18:28:51 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
s . MustClose ( )
if err := os . RemoveAll ( path ) ; err != nil {
t . Fatalf ( "cannot remove %q: %s" , path , err )
lib/index: reduce read/write load after indexDB rotation (#2177)
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
}
}
2019-11-12 13:09:33 +00:00
func TestSearchTSIDWithTimeRange ( t * testing . T ) {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
const path = "TestSearchTSIDWithTimeRange"
2023-09-01 07:27:51 +00:00
s := MustOpenStorage ( path , retentionMax , 0 , 0 )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db := s . idb ( )
2019-11-12 13:09:33 +00:00
// Create a bunch of per-day time series
const accountID = 12345
const projectID = 85453
2020-07-23 21:30:33 +00:00
is := db . getIndexSearch ( accountID , projectID , noDeadline )
2019-11-12 13:09:33 +00:00
const days = 5
const metricsPerDay = 1000
theDay := time . Date ( 2019 , time . October , 15 , 5 , 1 , 0 , 0 , time . UTC )
now := uint64 ( timestampFromTime ( theDay ) )
baseDate := now / msecPerDay
var metricNameBuf [ ] byte
2020-11-02 17:11:48 +00:00
perDayMetricIDs := make ( map [ uint64 ] * uint64set . Set )
var allMetricIDs uint64set . Set
2022-06-12 01:32:13 +00:00
labelNames := [ ] string {
2023-01-07 08:50:14 +00:00
"__name__" , "constant" , "day" , "UniqueId" , "some_unique_id" ,
2020-11-04 22:15:43 +00:00
}
2022-06-12 01:32:13 +00:00
labelValues := [ ] string {
2020-11-04 22:15:43 +00:00
"testMetric" ,
}
2022-06-12 01:32:13 +00:00
sort . Strings ( labelNames )
2024-05-29 12:07:44 +00:00
newMN := func ( name string , day , metric int ) MetricName {
var mn MetricName
mn . AccountID = accountID
mn . ProjectID = projectID
mn . MetricGroup = [ ] byte ( name )
mn . AddTag (
"constant" ,
"const" ,
)
mn . AddTag (
"day" ,
fmt . Sprintf ( "%v" , day ) ,
)
mn . AddTag (
"UniqueId" ,
fmt . Sprintf ( "%v" , metric ) ,
)
mn . AddTag (
"some_unique_id" ,
fmt . Sprintf ( "%v" , day ) ,
)
mn . sortTags ( )
return mn
}
2019-11-12 13:09:33 +00:00
for day := 0 ; day < days ; day ++ {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
date := baseDate - uint64 ( day )
var metricIDs uint64set . Set
2019-11-12 13:09:33 +00:00
for metric := 0 ; metric < metricsPerDay ; metric ++ {
2024-05-29 12:07:44 +00:00
mn := newMN ( "testMetric" , day , metric )
2019-11-12 13:09:33 +00:00
metricNameBuf = mn . Marshal ( metricNameBuf [ : 0 ] )
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
var genTSID generationTSID
if ! is . getTSIDByMetricName ( & genTSID , metricNameBuf , date ) {
generateTSID ( & genTSID . TSID , & mn )
2023-07-14 06:13:21 +00:00
createAllIndexesForMetricName ( is , & mn , & genTSID . TSID , date )
2019-11-12 13:09:33 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if genTSID . TSID . AccountID != accountID {
t . Fatalf ( "unexpected accountID; got %d; want %d" , genTSID . TSID . AccountID , accountID )
2019-11-12 13:09:33 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if genTSID . TSID . ProjectID != projectID {
t . Fatalf ( "unexpected accountID; got %d; want %d" , genTSID . TSID . ProjectID , projectID )
2019-11-12 13:09:33 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
metricIDs . Add ( genTSID . TSID . MetricID )
2019-11-12 13:09:33 +00:00
}
2020-11-02 17:11:48 +00:00
allMetricIDs . Union ( & metricIDs )
perDayMetricIDs [ date ] = & metricIDs
2019-11-12 13:09:33 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db . putIndexSearch ( is )
2019-11-12 13:09:33 +00:00
// Flush index to disk, so it becomes visible for search
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
s . DebugFlush ( )
2019-11-12 13:09:33 +00:00
2020-11-02 17:11:48 +00:00
is2 := db . getIndexSearch ( accountID , projectID , noDeadline )
// Check that all the metrics are found for all the days.
for date := baseDate - days + 1 ; date <= baseDate ; date ++ {
metricIDs , err := is2 . getMetricIDsForDate ( date , metricsPerDay )
if err != nil {
t . Fatalf ( "unexpected error: %s" , err )
}
if ! perDayMetricIDs [ date ] . Equal ( metricIDs ) {
t . Fatalf ( "unexpected metricIDs found;\ngot\n%d\nwant\n%d" , metricIDs . AppendTo ( nil ) , perDayMetricIDs [ date ] . AppendTo ( nil ) )
}
}
2021-07-30 05:37:10 +00:00
// Check that all the metrics are found in global index
metricIDs , err := is2 . getMetricIDsForDate ( 0 , metricsPerDay * days )
if err != nil {
2020-11-02 17:11:48 +00:00
t . Fatalf ( "unexpected error: %s" , err )
}
2021-07-30 05:37:10 +00:00
if ! allMetricIDs . Equal ( metricIDs ) {
2020-11-02 17:11:48 +00:00
t . Fatalf ( "unexpected metricIDs found;\ngot\n%d\nwant\n%d" , metricIDs . AppendTo ( nil ) , allMetricIDs . AppendTo ( nil ) )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
db . putIndexSearch ( is2 )
2020-11-02 17:11:48 +00:00
2024-05-29 12:07:44 +00:00
// add a metric that will be deleted shortly
is3 := db . getIndexSearch ( accountID , projectID , noDeadline )
day := days
date := baseDate - uint64 ( day )
mn := newMN ( "deletedMetric" , day , 999 )
mn . AddTag (
"labelToDelete" ,
fmt . Sprintf ( "%v" , day ) ,
)
mn . sortTags ( )
metricNameBuf = mn . Marshal ( metricNameBuf [ : 0 ] )
var genTSID generationTSID
if ! is3 . getTSIDByMetricName ( & genTSID , metricNameBuf , date ) {
generateTSID ( & genTSID . TSID , & mn )
createAllIndexesForMetricName ( is3 , & mn , & genTSID . TSID , date )
}
// delete the added metric. It is expected it won't be returned during searches
deletedSet := & uint64set . Set { }
deletedSet . Add ( genTSID . TSID . MetricID )
s . setDeletedMetricIDs ( deletedSet )
db . putIndexSearch ( is3 )
s . DebugFlush ( )
2022-06-12 01:32:13 +00:00
// Check SearchLabelNamesWithFiltersOnTimeRange with the specified time range.
tr := TimeRange {
2020-11-04 22:15:43 +00:00
MinTimestamp : int64 ( now ) - msecPerDay ,
MaxTimestamp : int64 ( now ) ,
2022-06-12 01:32:13 +00:00
}
lns , err := db . SearchLabelNamesWithFiltersOnTimeRange ( nil , accountID , projectID , nil , tr , 10000 , 1e9 , noDeadline )
2020-11-04 22:15:43 +00:00
if err != nil {
2022-06-12 01:32:13 +00:00
t . Fatalf ( "unexpected error in SearchLabelNamesWithFiltersOnTimeRange(timeRange=%s): %s" , & tr , err )
2020-11-04 22:15:43 +00:00
}
2022-06-12 01:32:13 +00:00
sort . Strings ( lns )
if ! reflect . DeepEqual ( lns , labelNames ) {
t . Fatalf ( "unexpected labelNames; got\n%s\nwant\n%s" , lns , labelNames )
2020-11-04 22:15:43 +00:00
}
2022-06-12 01:32:13 +00:00
// Check SearchLabelValuesWithFiltersOnTimeRange with the specified time range.
lvs , err := db . SearchLabelValuesWithFiltersOnTimeRange ( nil , accountID , projectID , "" , nil , tr , 10000 , 1e9 , noDeadline )
2020-11-04 22:15:43 +00:00
if err != nil {
2022-06-12 01:32:13 +00:00
t . Fatalf ( "unexpected error in SearchLabelValuesWithFiltersOnTimeRange(timeRange=%s): %s" , & tr , err )
2020-11-04 22:15:43 +00:00
}
2022-06-12 01:32:13 +00:00
sort . Strings ( lvs )
if ! reflect . DeepEqual ( lvs , labelValues ) {
t . Fatalf ( "unexpected labelValues; got\n%s\nwant\n%s" , lvs , labelValues )
2020-11-04 22:15:43 +00:00
}
2019-11-12 13:09:33 +00:00
// Create a filter that will match series that occur across multiple days
tfs := NewTagFilters ( accountID , projectID )
if err := tfs . Add ( [ ] byte ( "constant" ) , [ ] byte ( "const" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter: %s" , err )
}
2024-03-11 23:43:27 +00:00
tfsMetricName := NewTagFilters ( accountID , projectID )
if err := tfsMetricName . Add ( [ ] byte ( "constant" ) , [ ] byte ( "const" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter on label: %s" , err )
}
if err := tfsMetricName . Add ( nil , [ ] byte ( "testMetric" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter on metric name: %s" , err )
}
2019-11-12 13:09:33 +00:00
2020-10-01 16:03:34 +00:00
// Perform a search within a day.
2019-11-12 13:09:33 +00:00
// This should return the metrics for the day
2022-06-12 01:32:13 +00:00
tr = TimeRange {
2019-11-12 13:09:33 +00:00
MinTimestamp : int64 ( now - 2 * msecPerHour - 1 ) ,
MaxTimestamp : int64 ( now ) ,
}
2022-10-23 09:15:24 +00:00
matchedTSIDs , err := searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-11-12 13:09:33 +00:00
if err != nil {
t . Fatalf ( "error searching tsids: %v" , err )
}
if len ( matchedTSIDs ) != metricsPerDay {
2020-10-01 16:03:34 +00:00
t . Fatalf ( "expected %d time series for current day, got %d time series" , metricsPerDay , len ( matchedTSIDs ) )
2019-11-12 13:09:33 +00:00
}
2022-06-12 01:32:13 +00:00
// Check SearchLabelNamesWithFiltersOnTimeRange with the specified filter.
lns , err = db . SearchLabelNamesWithFiltersOnTimeRange ( nil , accountID , projectID , [ ] * TagFilters { tfs } , TimeRange { } , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelNamesWithFiltersOnTimeRange(filters=%s): %s" , tfs , err )
}
sort . Strings ( lns )
if ! reflect . DeepEqual ( lns , labelNames ) {
t . Fatalf ( "unexpected labelNames; got\n%s\nwant\n%s" , lns , labelNames )
}
// Check SearchLabelNamesWithFiltersOnTimeRange with the specified filter and time range.
lns , err = db . SearchLabelNamesWithFiltersOnTimeRange ( nil , accountID , projectID , [ ] * TagFilters { tfs } , tr , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelNamesWithFiltersOnTimeRange(filters=%s, timeRange=%s): %s" , tfs , & tr , err )
}
sort . Strings ( lns )
if ! reflect . DeepEqual ( lns , labelNames ) {
t . Fatalf ( "unexpected labelNames; got\n%s\nwant\n%s" , lns , labelNames )
}
2024-03-11 23:43:27 +00:00
// Check SearchLabelNamesWithFiltersOnTimeRange with filters on metric name and time range.
lns , err = db . SearchLabelNamesWithFiltersOnTimeRange ( nil , accountID , projectID , [ ] * TagFilters { tfsMetricName } , tr , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelNamesWithFiltersOnTimeRange(filters=%s, timeRange=%s): %s" , tfs , & tr , err )
}
sort . Strings ( lns )
if ! reflect . DeepEqual ( lns , labelNames ) {
t . Fatalf ( "unexpected labelNames; got\n%s\nwant\n%s" , lns , labelNames )
}
2022-06-12 01:32:13 +00:00
// Check SearchLabelValuesWithFiltersOnTimeRange with the specified filter.
lvs , err = db . SearchLabelValuesWithFiltersOnTimeRange ( nil , accountID , projectID , "" , [ ] * TagFilters { tfs } , TimeRange { } , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelValuesWithFiltersOnTimeRange(filters=%s): %s" , tfs , err )
}
sort . Strings ( lvs )
if ! reflect . DeepEqual ( lvs , labelValues ) {
t . Fatalf ( "unexpected labelValues; got\n%s\nwant\n%s" , lvs , labelValues )
}
// Check SearchLabelValuesWithFiltersOnTimeRange with the specified filter and time range.
lvs , err = db . SearchLabelValuesWithFiltersOnTimeRange ( nil , accountID , projectID , "" , [ ] * TagFilters { tfs } , tr , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelValuesWithFiltersOnTimeRange(filters=%s, timeRange=%s): %s" , tfs , & tr , err )
}
sort . Strings ( lvs )
if ! reflect . DeepEqual ( lvs , labelValues ) {
t . Fatalf ( "unexpected labelValues; got\n%s\nwant\n%s" , lvs , labelValues )
2024-03-11 23:43:27 +00:00
}
// Check SearchLabelValuesWithFiltersOnTimeRange with filters on metric name and time range.
lvs , err = db . SearchLabelValuesWithFiltersOnTimeRange ( nil , accountID , projectID , "" , [ ] * TagFilters { tfsMetricName } , tr , 10000 , 1e9 , noDeadline )
if err != nil {
t . Fatalf ( "unexpected error in SearchLabelValuesWithFiltersOnTimeRange(filters=%s, timeRange=%s): %s" , tfs , & tr , err )
}
sort . Strings ( lvs )
if ! reflect . DeepEqual ( lvs , labelValues ) {
t . Fatalf ( "unexpected labelValues; got\n%s\nwant\n%s" , lvs , labelValues )
2022-06-12 01:32:13 +00:00
}
2019-11-12 13:09:33 +00:00
// Perform a search across all the days, should match all metrics
tr = TimeRange {
MinTimestamp : int64 ( now - msecPerDay * days ) ,
MaxTimestamp : int64 ( now ) ,
}
2022-10-23 09:15:24 +00:00
matchedTSIDs , err = searchTSIDsInTest ( db , [ ] * TagFilters { tfs } , tr )
2019-11-12 13:09:33 +00:00
if err != nil {
t . Fatalf ( "error searching tsids: %v" , err )
}
if len ( matchedTSIDs ) != metricsPerDay * days {
2020-10-01 16:03:34 +00:00
t . Fatalf ( "expected %d time series for all days, got %d time series" , metricsPerDay * days , len ( matchedTSIDs ) )
2019-11-12 13:09:33 +00:00
}
2020-04-22 16:57:36 +00:00
2022-06-14 14:46:16 +00:00
// Check GetTSDBStatus with nil filters.
status , err := db . GetTSDBStatus ( nil , accountID , projectID , nil , baseDate , "day" , 5 , 1e6 , noDeadline )
2020-04-22 16:57:36 +00:00
if err != nil {
2022-06-14 14:46:16 +00:00
t . Fatalf ( "error in GetTSDBStatus with nil filters: %s" , err )
2020-04-22 16:57:36 +00:00
}
if ! status . hasEntries ( ) {
t . Fatalf ( "expecting non-empty TSDB status" )
}
expectedSeriesCountByMetricName := [ ] TopHeapEntry {
{
Name : "testMetric" ,
Count : 1000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByMetricName , expectedSeriesCountByMetricName ) {
t . Fatalf ( "unexpected SeriesCountByMetricName;\ngot\n%v\nwant\n%v" , status . SeriesCountByMetricName , expectedSeriesCountByMetricName )
}
2022-06-14 13:32:38 +00:00
expectedSeriesCountByLabelName := [ ] TopHeapEntry {
2023-01-07 08:50:14 +00:00
{
Name : "UniqueId" ,
Count : 1000 ,
} ,
2022-06-14 13:32:38 +00:00
{
Name : "__name__" ,
Count : 1000 ,
} ,
{
Name : "constant" ,
Count : 1000 ,
} ,
{
Name : "day" ,
Count : 1000 ,
} ,
{
2023-01-07 08:50:14 +00:00
Name : "some_unique_id" ,
2022-06-14 13:32:38 +00:00
Count : 1000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByLabelName , expectedSeriesCountByLabelName ) {
t . Fatalf ( "unexpected SeriesCountByLabelName;\ngot\n%v\nwant\n%v" , status . SeriesCountByLabelName , expectedSeriesCountByLabelName )
}
2022-06-14 14:46:16 +00:00
expectedSeriesCountByFocusLabelValue := [ ] TopHeapEntry {
{
Name : "0" ,
Count : 1000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByFocusLabelValue , expectedSeriesCountByFocusLabelValue ) {
t . Fatalf ( "unexpected SeriesCountByFocusLabelValue;\ngot\n%v\nwant\n%v" , status . SeriesCountByFocusLabelValue , expectedSeriesCountByFocusLabelValue )
}
2020-04-22 16:57:36 +00:00
expectedLabelValueCountByLabelName := [ ] TopHeapEntry {
{
2023-01-07 08:50:14 +00:00
Name : "UniqueId" ,
2020-04-22 16:57:36 +00:00
Count : 1000 ,
} ,
{
Name : "__name__" ,
Count : 1 ,
} ,
{
Name : "constant" ,
Count : 1 ,
} ,
{
Name : "day" ,
Count : 1 ,
} ,
2023-01-07 08:50:14 +00:00
{
Name : "some_unique_id" ,
Count : 1 ,
} ,
2020-04-22 16:57:36 +00:00
}
if ! reflect . DeepEqual ( status . LabelValueCountByLabelName , expectedLabelValueCountByLabelName ) {
t . Fatalf ( "unexpected LabelValueCountByLabelName;\ngot\n%v\nwant\n%v" , status . LabelValueCountByLabelName , expectedLabelValueCountByLabelName )
}
expectedSeriesCountByLabelValuePair := [ ] TopHeapEntry {
{
Name : "__name__=testMetric" ,
Count : 1000 ,
} ,
{
Name : "constant=const" ,
Count : 1000 ,
} ,
{
Name : "day=0" ,
Count : 1000 ,
} ,
{
2023-01-07 08:50:14 +00:00
Name : "some_unique_id=0" ,
Count : 1000 ,
2020-04-22 16:57:36 +00:00
} ,
{
2023-01-07 08:50:14 +00:00
Name : "UniqueId=1" ,
2020-04-22 16:57:36 +00:00
Count : 1 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByLabelValuePair , expectedSeriesCountByLabelValuePair ) {
t . Fatalf ( "unexpected SeriesCountByLabelValuePair;\ngot\n%v\nwant\n%v" , status . SeriesCountByLabelValuePair , expectedSeriesCountByLabelValuePair )
}
2022-06-08 15:43:05 +00:00
expectedTotalSeries := uint64 ( 1000 )
if status . TotalSeries != expectedTotalSeries {
t . Fatalf ( "unexpected TotalSeries; got %d; want %d" , status . TotalSeries , expectedTotalSeries )
}
2023-01-07 08:50:14 +00:00
expectedLabelValuePairs := uint64 ( 5000 )
2022-06-08 15:43:05 +00:00
if status . TotalLabelValuePairs != expectedLabelValuePairs {
t . Fatalf ( "unexpected TotalLabelValuePairs; got %d; want %d" , status . TotalLabelValuePairs , expectedLabelValuePairs )
}
2021-05-12 12:18:45 +00:00
2022-06-14 14:46:16 +00:00
// Check GetTSDBStatus with non-nil filter, which matches all the series
2021-05-12 12:18:45 +00:00
tfs = NewTagFilters ( accountID , projectID )
if err := tfs . Add ( [ ] byte ( "day" ) , [ ] byte ( "0" ) , false , false ) ; err != nil {
t . Fatalf ( "cannot add filter: %s" , err )
}
2022-06-14 14:46:16 +00:00
status , err = db . GetTSDBStatus ( nil , accountID , projectID , [ ] * TagFilters { tfs } , baseDate , "" , 5 , 1e6 , noDeadline )
2021-05-12 12:18:45 +00:00
if err != nil {
2022-06-14 14:46:16 +00:00
t . Fatalf ( "error in GetTSDBStatus: %s" , err )
2021-05-12 12:18:45 +00:00
}
if ! status . hasEntries ( ) {
t . Fatalf ( "expecting non-empty TSDB status" )
}
expectedSeriesCountByMetricName = [ ] TopHeapEntry {
{
Name : "testMetric" ,
Count : 1000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByMetricName , expectedSeriesCountByMetricName ) {
t . Fatalf ( "unexpected SeriesCountByMetricName;\ngot\n%v\nwant\n%v" , status . SeriesCountByMetricName , expectedSeriesCountByMetricName )
}
2022-06-08 15:43:05 +00:00
expectedTotalSeries = 1000
if status . TotalSeries != expectedTotalSeries {
t . Fatalf ( "unexpected TotalSeries; got %d; want %d" , status . TotalSeries , expectedTotalSeries )
}
2023-01-07 08:50:14 +00:00
expectedLabelValuePairs = 5000
2022-06-08 15:43:05 +00:00
if status . TotalLabelValuePairs != expectedLabelValuePairs {
t . Fatalf ( "unexpected TotalLabelValuePairs; got %d; want %d" , status . TotalLabelValuePairs , expectedLabelValuePairs )
}
2022-06-12 11:17:44 +00:00
2022-06-14 14:46:16 +00:00
// Check GetTSDBStatus with non-nil filter, which matches all the series on a global time range
status , err = db . GetTSDBStatus ( nil , accountID , projectID , nil , 0 , "day" , 5 , 1e6 , noDeadline )
2022-06-12 11:17:44 +00:00
if err != nil {
2022-06-14 14:46:16 +00:00
t . Fatalf ( "error in GetTSDBStatus: %s" , err )
2022-06-12 11:17:44 +00:00
}
if ! status . hasEntries ( ) {
t . Fatalf ( "expecting non-empty TSDB status" )
}
expectedSeriesCountByMetricName = [ ] TopHeapEntry {
{
Name : "testMetric" ,
Count : 5000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByMetricName , expectedSeriesCountByMetricName ) {
t . Fatalf ( "unexpected SeriesCountByMetricName;\ngot\n%v\nwant\n%v" , status . SeriesCountByMetricName , expectedSeriesCountByMetricName )
}
expectedTotalSeries = 5000
if status . TotalSeries != expectedTotalSeries {
t . Fatalf ( "unexpected TotalSeries; got %d; want %d" , status . TotalSeries , expectedTotalSeries )
}
2023-01-07 08:50:14 +00:00
expectedLabelValuePairs = 25000
2022-06-12 11:17:44 +00:00
if status . TotalLabelValuePairs != expectedLabelValuePairs {
t . Fatalf ( "unexpected TotalLabelValuePairs; got %d; want %d" , status . TotalLabelValuePairs , expectedLabelValuePairs )
}
2022-06-14 14:46:16 +00:00
expectedSeriesCountByFocusLabelValue = [ ] TopHeapEntry {
{
Name : "0" ,
Count : 1000 ,
} ,
{
Name : "1" ,
Count : 1000 ,
} ,
{
Name : "2" ,
Count : 1000 ,
} ,
{
Name : "3" ,
Count : 1000 ,
} ,
{
Name : "4" ,
Count : 1000 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByFocusLabelValue , expectedSeriesCountByFocusLabelValue ) {
t . Fatalf ( "unexpected SeriesCountByFocusLabelValue;\ngot\n%v\nwant\n%v" , status . SeriesCountByFocusLabelValue , expectedSeriesCountByFocusLabelValue )
}
2022-06-12 11:17:44 +00:00
2022-06-14 14:46:16 +00:00
// Check GetTSDBStatus with non-nil filter, which matches only 3 series
2022-06-08 15:43:05 +00:00
tfs = NewTagFilters ( accountID , projectID )
2023-01-07 08:50:14 +00:00
if err := tfs . Add ( [ ] byte ( "UniqueId" ) , [ ] byte ( "0|1|3" ) , false , true ) ; err != nil {
2022-06-08 15:43:05 +00:00
t . Fatalf ( "cannot add filter: %s" , err )
}
2022-06-14 14:46:16 +00:00
status , err = db . GetTSDBStatus ( nil , accountID , projectID , [ ] * TagFilters { tfs } , baseDate , "" , 5 , 1e6 , noDeadline )
2022-06-08 15:43:05 +00:00
if err != nil {
2022-06-14 14:46:16 +00:00
t . Fatalf ( "error in GetTSDBStatus: %s" , err )
2022-06-08 15:43:05 +00:00
}
if ! status . hasEntries ( ) {
t . Fatalf ( "expecting non-empty TSDB status" )
}
expectedSeriesCountByMetricName = [ ] TopHeapEntry {
{
Name : "testMetric" ,
Count : 3 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByMetricName , expectedSeriesCountByMetricName ) {
t . Fatalf ( "unexpected SeriesCountByMetricName;\ngot\n%v\nwant\n%v" , status . SeriesCountByMetricName , expectedSeriesCountByMetricName )
}
expectedTotalSeries = 3
if status . TotalSeries != expectedTotalSeries {
t . Fatalf ( "unexpected TotalSeries; got %d; want %d" , status . TotalSeries , expectedTotalSeries )
}
2023-01-07 08:50:14 +00:00
expectedLabelValuePairs = 15
2022-06-08 15:43:05 +00:00
if status . TotalLabelValuePairs != expectedLabelValuePairs {
t . Fatalf ( "unexpected TotalLabelValuePairs; got %d; want %d" , status . TotalLabelValuePairs , expectedLabelValuePairs )
}
2022-06-12 11:17:44 +00:00
2022-06-14 14:46:16 +00:00
// Check GetTSDBStatus with non-nil filter on global time range, which matches only 15 series
status , err = db . GetTSDBStatus ( nil , accountID , projectID , [ ] * TagFilters { tfs } , 0 , "" , 5 , 1e6 , noDeadline )
2022-06-12 11:17:44 +00:00
if err != nil {
2022-06-14 14:46:16 +00:00
t . Fatalf ( "error in GetTSDBStatus: %s" , err )
2022-06-12 11:17:44 +00:00
}
if ! status . hasEntries ( ) {
t . Fatalf ( "expecting non-empty TSDB status" )
}
expectedSeriesCountByMetricName = [ ] TopHeapEntry {
{
Name : "testMetric" ,
Count : 15 ,
} ,
}
if ! reflect . DeepEqual ( status . SeriesCountByMetricName , expectedSeriesCountByMetricName ) {
t . Fatalf ( "unexpected SeriesCountByMetricName;\ngot\n%v\nwant\n%v" , status . SeriesCountByMetricName , expectedSeriesCountByMetricName )
}
expectedTotalSeries = 15
if status . TotalSeries != expectedTotalSeries {
t . Fatalf ( "unexpected TotalSeries; got %d; want %d" , status . TotalSeries , expectedTotalSeries )
}
2023-01-07 08:50:14 +00:00
expectedLabelValuePairs = 75
2022-06-12 11:17:44 +00:00
if status . TotalLabelValuePairs != expectedLabelValuePairs {
t . Fatalf ( "unexpected TotalLabelValuePairs; got %d; want %d" , status . TotalLabelValuePairs , expectedLabelValuePairs )
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping
Previously all the newly ingested time series were registered in global `MetricName -> TSID` index.
This index was used during data ingestion for locating the TSID (internal series id)
for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names).
The `MetricName -> TSID` index is stored on disk in order to make sure that the data
isn't lost on VictoriaMetrics restart or unclean shutdown.
The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding
data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache,
and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics
uses in-memory cache for speeding up the lookup for active time series.
This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested
active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk.
VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases:
- If `storage/tsid` cache capacity isn't enough for active time series.
Then just increase available memory for VictoriaMetrics or reduce the number of active time series
ingested into VictoriaMetrics.
- If new time series is ingested into VictoriaMetrics. In this case it cannot find
the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index,
since it doesn't know that the index has no the corresponding entry too.
This is a typical event under high churn rate, when old time series are constantly substituted
with new time series.
Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index,
are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics.
Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName`
for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod.
This index can become very large under high churn rate and long retention. VictoriaMetrics
caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups.
The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing
recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series.
This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly
reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics
consults only the index for the current day when new time series is ingested into it.
The downside of this change is increased indexdb size on disk for workloads without high churn rate,
e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store
identical `MetricName -> TSID` entries for static time series for every day.
This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation,
since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 .
At the same time the change fixes the issue, which could result in lost access to time series,
which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698
The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed
in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685
This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
s . MustClose ( )
fs . MustRemoveAll ( path )
2019-11-12 13:09:33 +00:00
}
2019-05-22 21:16:55 +00:00
func toTFPointers ( tfs [ ] tagFilter ) [ ] * tagFilter {
tfps := make ( [ ] * tagFilter , len ( tfs ) )
for i := range tfs {
tfps [ i ] = & tfs [ i ]
}
return tfps
}
2021-06-11 09:42:26 +00:00
func newTestStorage ( ) * Storage {
2021-06-15 11:56:51 +00:00
s := & Storage {
2021-06-11 09:42:26 +00:00
cachePath : "test-storage-cache" ,
2022-07-05 11:49:03 +00:00
metricIDCache : workingsetcache . New ( 1234 ) ,
metricNameCache : workingsetcache . New ( 1234 ) ,
tsidCache : workingsetcache . New ( 1234 ) ,
dateMetricIDCache : newDateMetricIDCache ( ) ,
2023-09-01 07:27:51 +00:00
retentionMsecs : retentionMax . Milliseconds ( ) ,
2021-06-11 09:42:26 +00:00
}
2021-06-15 11:56:51 +00:00
s . setDeletedMetricIDs ( & uint64set . Set { } )
return s
2021-06-11 09:42:26 +00:00
}
func stopTestStorage ( s * Storage ) {
s . metricIDCache . Stop ( )
s . metricNameCache . Stop ( )
s . tsidCache . Stop ( )
2022-09-13 12:48:20 +00:00
fs . MustRemoveDirAtomic ( s . cachePath )
2021-06-11 09:42:26 +00:00
}