2023-06-20 07:55:12 +02:00
package logstorage
import (
2024-10-29 11:28:41 +01:00
"math/bits"
2023-06-20 07:55:12 +02:00
"path/filepath"
"sync"
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
"github.com/cespare/xxhash/v2"
2023-06-20 07:55:12 +02:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
2024-10-29 11:28:41 +01:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/slicesutil"
2023-06-20 07:55:12 +02:00
)
// writerWithStats writes data to w and tracks the total amounts of data written at bytesWritten.
type writerWithStats struct {
w filestream . WriteCloser
bytesWritten uint64
}
func ( w * writerWithStats ) reset ( ) {
w . w = nil
w . bytesWritten = 0
}
func ( w * writerWithStats ) init ( wc filestream . WriteCloser ) {
w . reset ( )
w . w = wc
}
func ( w * writerWithStats ) Path ( ) string {
return w . w . Path ( )
}
func ( w * writerWithStats ) MustWrite ( data [ ] byte ) {
fs . MustWriteData ( w . w , data )
w . bytesWritten += uint64 ( len ( data ) )
}
// MustClose closes the underlying w.
func ( w * writerWithStats ) MustClose ( ) {
w . w . MustClose ( )
}
// streamWriters contain writers for blockStreamWriter
type streamWriters struct {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
columnNamesWriter writerWithStats
2023-06-20 07:55:12 +02:00
metaindexWriter writerWithStats
indexWriter writerWithStats
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
columnsHeaderIndexWriter writerWithStats
2023-06-20 07:55:12 +02:00
columnsHeaderWriter writerWithStats
timestampsWriter writerWithStats
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
messageBloomValuesWriter bloomValuesWriter
2024-10-29 11:28:41 +01:00
bloomValuesShards [ ] bloomValuesWriter
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
}
type bloomValuesWriter struct {
bloom writerWithStats
values writerWithStats
}
func ( w * bloomValuesWriter ) reset ( ) {
w . bloom . reset ( )
w . values . reset ( )
}
func ( w * bloomValuesWriter ) init ( sw bloomValuesStreamWriter ) {
w . bloom . init ( sw . bloom )
w . values . init ( sw . values )
}
func ( w * bloomValuesWriter ) totalBytesWritten ( ) uint64 {
return w . bloom . bytesWritten + w . values . bytesWritten
}
func ( w * bloomValuesWriter ) MustClose ( ) {
w . bloom . MustClose ( )
w . values . MustClose ( )
}
type bloomValuesStreamWriter struct {
bloom filestream . WriteCloser
values filestream . WriteCloser
2023-06-20 07:55:12 +02:00
}
func ( sw * streamWriters ) reset ( ) {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnNamesWriter . reset ( )
2023-06-20 07:55:12 +02:00
sw . metaindexWriter . reset ( )
sw . indexWriter . reset ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnsHeaderIndexWriter . reset ( )
2023-06-20 07:55:12 +02:00
sw . columnsHeaderWriter . reset ( )
sw . timestampsWriter . reset ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . messageBloomValuesWriter . reset ( )
2024-10-29 11:28:41 +01:00
for i := range sw . bloomValuesShards {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . bloomValuesShards [ i ] . reset ( )
}
2024-10-29 11:28:41 +01:00
sw . bloomValuesShards = sw . bloomValuesShards [ : 0 ]
2023-06-20 07:55:12 +02:00
}
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
func ( sw * streamWriters ) init ( columnNamesWriter , metaindexWriter , indexWriter , columnsHeaderIndexWriter , columnsHeaderWriter , timestampsWriter filestream . WriteCloser ,
2024-10-29 11:28:41 +01:00
messageBloomValuesWriter bloomValuesStreamWriter , bloomValuesShards [ ] bloomValuesStreamWriter ,
2023-06-20 07:55:12 +02:00
) {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnNamesWriter . init ( columnNamesWriter )
2023-06-20 07:55:12 +02:00
sw . metaindexWriter . init ( metaindexWriter )
sw . indexWriter . init ( indexWriter )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnsHeaderIndexWriter . init ( columnsHeaderIndexWriter )
2023-06-20 07:55:12 +02:00
sw . columnsHeaderWriter . init ( columnsHeaderWriter )
sw . timestampsWriter . init ( timestampsWriter )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . messageBloomValuesWriter . init ( messageBloomValuesWriter )
2024-10-29 11:28:41 +01:00
sw . bloomValuesShards = slicesutil . SetLength ( sw . bloomValuesShards , len ( bloomValuesShards ) )
for i := range sw . bloomValuesShards {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . bloomValuesShards [ i ] . init ( bloomValuesShards [ i ] )
}
2023-06-20 07:55:12 +02:00
}
func ( sw * streamWriters ) totalBytesWritten ( ) uint64 {
n := uint64 ( 0 )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
n += sw . columnNamesWriter . bytesWritten
2023-06-20 07:55:12 +02:00
n += sw . metaindexWriter . bytesWritten
n += sw . indexWriter . bytesWritten
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
n += sw . columnsHeaderIndexWriter . bytesWritten
2023-06-20 07:55:12 +02:00
n += sw . columnsHeaderWriter . bytesWritten
n += sw . timestampsWriter . bytesWritten
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
n += sw . messageBloomValuesWriter . totalBytesWritten ( )
2024-10-29 11:28:41 +01:00
for i := range sw . bloomValuesShards {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
n += sw . bloomValuesShards [ i ] . totalBytesWritten ( )
}
2023-06-20 07:55:12 +02:00
return n
}
func ( sw * streamWriters ) MustClose ( ) {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnNamesWriter . MustClose ( )
2023-06-20 07:55:12 +02:00
sw . metaindexWriter . MustClose ( )
sw . indexWriter . MustClose ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . columnsHeaderIndexWriter . MustClose ( )
2023-06-20 07:55:12 +02:00
sw . columnsHeaderWriter . MustClose ( )
sw . timestampsWriter . MustClose ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . messageBloomValuesWriter . MustClose ( )
2024-10-29 11:28:41 +01:00
for i := range sw . bloomValuesShards {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
sw . bloomValuesShards [ i ] . MustClose ( )
}
}
func ( sw * streamWriters ) getBloomValuesWriterForColumnName ( name string ) * bloomValuesWriter {
if name == "" {
return & sw . messageBloomValuesWriter
}
2024-10-29 11:28:41 +01:00
n := len ( sw . bloomValuesShards )
idx := uint64 ( 0 )
if n > 1 {
h := xxhash . Sum64 ( bytesutil . ToUnsafeBytes ( name ) )
idx = h % uint64 ( n )
}
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
return & sw . bloomValuesShards [ idx ]
2023-06-20 07:55:12 +02:00
}
// blockStreamWriter is used for writing blocks into the underlying storage in streaming manner.
type blockStreamWriter struct {
// streamWriters contains writer for block data
streamWriters streamWriters
// sidLast is the streamID for the last written block
sidLast streamID
// sidFirst is the streamID for the first block in the current indexBlock
sidFirst streamID
2024-10-29 11:28:41 +01:00
// bloomValuesFieldsCount is the number of fields with (bloom, values) pairs in the output part.
bloomValuesFieldsCount uint64
2023-06-20 07:55:12 +02:00
// minTimestampLast is the minimum timestamp seen for the last written block
minTimestampLast int64
// minTimestamp is the minimum timestamp seen across written blocks for the current indexBlock
minTimestamp int64
// maxTimestamp is the maximum timestamp seen across written blocks for the current indexBlock
maxTimestamp int64
// hasWrittenBlocks is set to true if at least a single block is written to the current indexBlock
hasWrittenBlocks bool
// globalUncompressedSizeBytes is the total size of all the log entries written via bsw
globalUncompressedSizeBytes uint64
// globalRowsCount is the total number of log entries written via bsw
globalRowsCount uint64
// globalBlocksCount is the total number of blocks written to bsw
globalBlocksCount uint64
// globalMinTimestamp is the minimum timestamp seen across all the blocks written to bsw
globalMinTimestamp int64
// globalMaxTimestamp is the maximum timestamp seen across all the blocks written to bsw
globalMaxTimestamp int64
// indexBlockData contains marshaled blockHeader data, which isn't written yet to indexFilename
indexBlockData [ ] byte
// metaindexData contains marshaled indexBlockHeader data, which isn't written yet to metaindexFilename
metaindexData [ ] byte
// indexBlockHeader is used for marshaling the data to metaindexData
indexBlockHeader indexBlockHeader
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
// columnNameIDGenerator is used for generating columnName->id mapping for all the columns seen in bsw
columnNameIDGenerator columnNameIDGenerator
2023-06-20 07:55:12 +02:00
}
// reset resets bsw for subsequent re-use.
func ( bsw * blockStreamWriter ) reset ( ) {
bsw . streamWriters . reset ( )
bsw . sidLast . reset ( )
bsw . sidFirst . reset ( )
2024-10-29 11:28:41 +01:00
bsw . bloomValuesFieldsCount = 0
2023-06-20 07:55:12 +02:00
bsw . minTimestampLast = 0
bsw . minTimestamp = 0
bsw . maxTimestamp = 0
bsw . hasWrittenBlocks = false
bsw . globalUncompressedSizeBytes = 0
bsw . globalRowsCount = 0
bsw . globalBlocksCount = 0
bsw . globalMinTimestamp = 0
bsw . globalMaxTimestamp = 0
bsw . indexBlockData = bsw . indexBlockData [ : 0 ]
if len ( bsw . metaindexData ) > 1024 * 1024 {
// The length of bsw.metaindexData is unbound, so drop too long buffer
// in order to conserve memory.
bsw . metaindexData = nil
} else {
bsw . metaindexData = bsw . metaindexData [ : 0 ]
}
bsw . indexBlockHeader . reset ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bsw . columnNameIDGenerator . reset ( )
2023-06-20 07:55:12 +02:00
}
2024-04-08 10:11:12 +02:00
// MustInitForInmemoryPart initializes bsw from mp
2023-06-20 07:55:12 +02:00
func ( bsw * blockStreamWriter ) MustInitForInmemoryPart ( mp * inmemoryPart ) {
bsw . reset ( )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
messageBloomValues := mp . messageBloomValues . NewStreamWriter ( )
2024-10-29 11:28:41 +01:00
bloomValuesShards := [ ] bloomValuesStreamWriter {
mp . fieldBloomValues . NewStreamWriter ( ) ,
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
}
bsw . streamWriters . init ( & mp . columnNames , & mp . metaindex , & mp . index , & mp . columnsHeaderIndex , & mp . columnsHeader , & mp . timestamps , messageBloomValues , bloomValuesShards )
2023-06-20 07:55:12 +02:00
}
// MustInitForFilePart initializes bsw for writing data to file part located at path.
//
// if nocache is true, then the written data doesn't go to OS page cache.
2024-10-29 11:28:41 +01:00
func ( bsw * blockStreamWriter ) MustInitForFilePart ( path string , nocache bool , bloomValuesShardsCount uint64 ) {
2023-06-20 07:55:12 +02:00
bsw . reset ( )
fs . MustMkdirFailIfExist ( path )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
columnNamesPath := filepath . Join ( path , columnNamesFilename )
2023-06-20 07:55:12 +02:00
metaindexPath := filepath . Join ( path , metaindexFilename )
indexPath := filepath . Join ( path , indexFilename )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
columnsHeaderIndexPath := filepath . Join ( path , columnsHeaderIndexFilename )
2023-06-20 07:55:12 +02:00
columnsHeaderPath := filepath . Join ( path , columnsHeaderFilename )
timestampsPath := filepath . Join ( path , timestampsFilename )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
// Always cache columnNames files, since it is re-read immediately after part creation
columnNamesWriter := filestream . MustCreate ( columnNamesPath , false )
// Always cache metaindex file, since it is re-read immediately after part creation
2023-06-20 07:55:12 +02:00
metaindexWriter := filestream . MustCreate ( metaindexPath , false )
indexWriter := filestream . MustCreate ( indexPath , nocache )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
columnsHeaderIndexWriter := filestream . MustCreate ( columnsHeaderIndexPath , nocache )
2023-06-20 07:55:12 +02:00
columnsHeaderWriter := filestream . MustCreate ( columnsHeaderPath , nocache )
timestampsWriter := filestream . MustCreate ( timestampsPath , nocache )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
messageBloomFilterPath := filepath . Join ( path , messageBloomFilename )
messageValuesPath := filepath . Join ( path , messageValuesFilename )
messageBloomValuesWriter := bloomValuesStreamWriter {
bloom : filestream . MustCreate ( messageBloomFilterPath , nocache ) ,
values : filestream . MustCreate ( messageValuesPath , nocache ) ,
}
2024-10-29 11:28:41 +01:00
bloomValuesShardsCount = adjustBloomValuesShardsCount ( bloomValuesShardsCount )
bloomValuesShards := make ( [ ] bloomValuesStreamWriter , bloomValuesShardsCount )
for i := range bloomValuesShards {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
shard := & bloomValuesShards [ i ]
bloomPath := getBloomFilePath ( path , uint64 ( i ) )
shard . bloom = filestream . MustCreate ( bloomPath , nocache )
valuesPath := getValuesFilePath ( path , uint64 ( i ) )
shard . values = filestream . MustCreate ( valuesPath , nocache )
}
bsw . streamWriters . init ( columnNamesWriter , metaindexWriter , indexWriter , columnsHeaderIndexWriter , columnsHeaderWriter , timestampsWriter , messageBloomValuesWriter , bloomValuesShards )
2023-06-20 07:55:12 +02:00
}
2024-10-29 11:28:41 +01:00
func adjustBloomValuesShardsCount ( n uint64 ) uint64 {
if n == 0 {
2024-10-30 13:37:27 +01:00
// At least a single shard is needed for writing potential non-const fields,
// which can appear after merging of const fields.
// This fixes https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7391
return 1
2024-10-29 11:28:41 +01:00
}
n = 1 << bits . Len64 ( n - 1 )
if n > bloomValuesMaxShardsCount {
n = bloomValuesMaxShardsCount
}
return n
}
2023-06-20 07:55:12 +02:00
// MustWriteRows writes timestamps with rows under the given sid to bsw.
//
// timestamps must be sorted.
// sid must be bigger or equal to the sid for the previously written rs.
func ( bsw * blockStreamWriter ) MustWriteRows ( sid * streamID , timestamps [ ] int64 , rows [ ] [ ] Field ) {
if len ( timestamps ) == 0 {
return
}
b := getBlock ( )
b . MustInitFromRows ( timestamps , rows )
bsw . MustWriteBlock ( sid , b )
putBlock ( b )
}
// MustWriteBlockData writes bd to bsw.
//
// The bd.streamID must be bigger or equal to the streamID for the previously written blocks.
func ( bsw * blockStreamWriter ) MustWriteBlockData ( bd * blockData ) {
if bd . rowsCount == 0 {
return
}
bsw . mustWriteBlockInternal ( & bd . streamID , nil , bd )
}
// MustWriteBlock writes b under the given sid to bsw.
//
// The sid must be bigger or equal to the sid for the previously written blocks.
// The minimum timestamp in b must be bigger or equal to the minimum timestamp written to the same sid.
func ( bsw * blockStreamWriter ) MustWriteBlock ( sid * streamID , b * block ) {
rowsCount := b . Len ( )
if rowsCount == 0 {
return
}
bsw . mustWriteBlockInternal ( sid , b , nil )
}
func ( bsw * blockStreamWriter ) mustWriteBlockInternal ( sid * streamID , b * block , bd * blockData ) {
if sid . less ( & bsw . sidLast ) {
logger . Panicf ( "BUG: the sid=%s cannot be smaller than the previously written sid=%s" , sid , & bsw . sidLast )
}
hasWrittenBlocks := bsw . hasWrittenBlocks
if ! hasWrittenBlocks {
bsw . sidFirst = * sid
bsw . hasWrittenBlocks = true
}
isSeenSid := sid . equal ( & bsw . sidLast )
bsw . sidLast = * sid
bh := getBlockHeader ( )
2024-10-29 11:28:41 +01:00
columnsLen := 0
2023-06-20 07:55:12 +02:00
if b != nil {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
b . mustWriteTo ( sid , bh , & bsw . streamWriters , & bsw . columnNameIDGenerator )
2024-10-29 11:28:41 +01:00
columnsLen = len ( b . columns )
2023-06-20 07:55:12 +02:00
} else {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bd . mustWriteTo ( bh , & bsw . streamWriters , & bsw . columnNameIDGenerator )
2024-10-29 11:28:41 +01:00
columnsLen = len ( bd . columnsData )
2023-06-20 07:55:12 +02:00
}
2024-10-29 11:28:41 +01:00
if bsw . bloomValuesFieldsCount < uint64 ( columnsLen ) {
bsw . bloomValuesFieldsCount = uint64 ( columnsLen )
}
2023-06-20 07:55:12 +02:00
th := & bh . timestampsHeader
if bsw . globalRowsCount == 0 || th . minTimestamp < bsw . globalMinTimestamp {
bsw . globalMinTimestamp = th . minTimestamp
}
if bsw . globalRowsCount == 0 || th . maxTimestamp > bsw . globalMaxTimestamp {
bsw . globalMaxTimestamp = th . maxTimestamp
}
if ! hasWrittenBlocks || th . minTimestamp < bsw . minTimestamp {
bsw . minTimestamp = th . minTimestamp
}
if ! hasWrittenBlocks || th . maxTimestamp > bsw . maxTimestamp {
bsw . maxTimestamp = th . maxTimestamp
}
if isSeenSid && th . minTimestamp < bsw . minTimestampLast {
logger . Panicf ( "BUG: the block for sid=%s cannot contain timestamp smaller than %d, but it contains timestamp %d" , sid , bsw . minTimestampLast , th . minTimestamp )
}
bsw . minTimestampLast = th . minTimestamp
bsw . globalUncompressedSizeBytes += bh . uncompressedSizeBytes
bsw . globalRowsCount += bh . rowsCount
bsw . globalBlocksCount ++
// Marshal bh
bsw . indexBlockData = bh . marshal ( bsw . indexBlockData )
putBlockHeader ( bh )
if len ( bsw . indexBlockData ) > maxUncompressedIndexBlockSize {
bsw . mustFlushIndexBlock ( bsw . indexBlockData )
bsw . indexBlockData = bsw . indexBlockData [ : 0 ]
}
}
func ( bsw * blockStreamWriter ) mustFlushIndexBlock ( data [ ] byte ) {
if len ( data ) > 0 {
bsw . indexBlockHeader . mustWriteIndexBlock ( data , bsw . sidFirst , bsw . minTimestamp , bsw . maxTimestamp , & bsw . streamWriters )
bsw . metaindexData = bsw . indexBlockHeader . marshal ( bsw . metaindexData )
}
bsw . hasWrittenBlocks = false
bsw . minTimestamp = 0
bsw . maxTimestamp = 0
bsw . sidFirst . reset ( )
}
// Finalize() finalizes the data write process and updates ph with the finalized stats
//
// It closes the writers passed to MustInit().
//
// bsw can be re-used after calling Finalize().
func ( bsw * blockStreamWriter ) Finalize ( ph * partHeader ) {
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
ph . FormatVersion = partFormatLatestVersion
2023-06-20 07:55:12 +02:00
ph . UncompressedSizeBytes = bsw . globalUncompressedSizeBytes
ph . RowsCount = bsw . globalRowsCount
ph . BlocksCount = bsw . globalBlocksCount
ph . MinTimestamp = bsw . globalMinTimestamp
ph . MaxTimestamp = bsw . globalMaxTimestamp
2024-10-29 11:28:41 +01:00
ph . BloomValuesShardsCount = uint64 ( len ( bsw . streamWriters . bloomValuesShards ) )
ph . BloomValuesFieldsCount = bsw . bloomValuesFieldsCount
2023-06-20 07:55:12 +02:00
bsw . mustFlushIndexBlock ( bsw . indexBlockData )
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
// Write columnNames data
mustWriteColumnNames ( & bsw . streamWriters . columnNamesWriter , bsw . columnNameIDGenerator . columnNames )
2023-06-20 07:55:12 +02:00
// Write metaindex data
lib/logstorage: refactor storage format to be more efficient for querying wide events
It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields.
For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log.
Such logs are also known as "wide events".
The previous storage format was optimized for logs with a few fields. When at least a single field
was referenced in the query, then the all the meta-information about all the log fields was unpacked
and parsed per each scanned block during the query. This could require a lot of additional disk IO
and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset)
index per each field in every data block. This index allows reading and extracting only the needed
metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ).
This allows increasing performance for queries over wide events by 10x and more.
Another issue was that the data for bloom filters and field values across all the log fields except of _msg
was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ).
This could result in huge disk read IO overhead when some small field was referred in the query,
since the Operating System usually reads more data than requested. It reads the data from disk
in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB).
So, if 512-byte bloom filter or values' block is read from the file, then the Operating System
reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible
for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache),
but this overhead may become very annoying when performing the query over large volumes of data
which isn't present in OS page cache.
The solution for this issue is to split bloom filters and field values across multiple shards.
This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards,
while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N.
Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases
performance for queries over large volumes of newly ingested data by up to 1000x.
The new storage format is versioned as v1, while the old storage format is version as v0.
It is stored in the partHeader.FormatVersion.
Parts with the old storage format are converted into parts with the new storage format during background merge.
It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
mustWriteIndexBlockHeaders ( & bsw . streamWriters . metaindexWriter , bsw . metaindexData )
2023-06-20 07:55:12 +02:00
ph . CompressedSizeBytes = bsw . streamWriters . totalBytesWritten ( )
bsw . streamWriters . MustClose ( )
bsw . reset ( )
}
var longTermBufPool bytesutil . ByteBufferPool
// getBlockStreamWriter returns new blockStreamWriter from the pool.
//
// Return back the blockStreamWriter to the pool when it is no longer needed by calling putBlockStreamWriter.
func getBlockStreamWriter ( ) * blockStreamWriter {
v := blockStreamWriterPool . Get ( )
if v == nil {
return & blockStreamWriter { }
}
return v . ( * blockStreamWriter )
}
// putBlockStreamWriter returns bsw to the pool.
func putBlockStreamWriter ( bsw * blockStreamWriter ) {
bsw . reset ( )
blockStreamWriterPool . Put ( bsw )
}
var blockStreamWriterPool sync . Pool
