2023-06-20 05:55:12 +00:00
package logstorage
import (
"fmt"
"math"
"sync"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
2024-05-12 14:33:29 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/slicesutil"
2023-06-20 05:55:12 +00:00
)
// blockHeader contains information about a single block.
//
// blockHeader is stored in the indexFilename file.
type blockHeader struct {
// streamID is a stream id for entries in the block
streamID streamID
// uncompressedSizeBytes is the original (uncompressed) size of log entries stored in the block
uncompressedSizeBytes uint64
// rowsCount is the number of log entries stored in the block
rowsCount uint64
// timestampsHeader contains information about timestamps for log entries in the block
timestampsHeader timestampsHeader
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 14:18:28 +00:00
// columnsHeaderIndexOffset is the offset of columnsHeaderIndex at columnsHeaderIndexFilename
columnsHeaderIndexOffset uint64
// columnsHeaderIndexSize is the size of columnsHeaderIndex at columnsHeaderIndexFilename
columnsHeaderIndexSize uint64
2023-06-20 05:55:12 +00:00
// columnsHeaderOffset is the offset of columnsHeader at columnsHeaderFilename
columnsHeaderOffset uint64
// columnsHeaderSize is the size of columnsHeader at columnsHeaderFilename
columnsHeaderSize uint64
}
// reset resets bh, so it can be re-used.
func ( bh * blockHeader ) reset ( ) {
bh . streamID . reset ( )
bh . uncompressedSizeBytes = 0
bh . rowsCount = 0
bh . timestampsHeader . 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 14:18:28 +00:00
bh . columnsHeaderIndexOffset = 0
bh . columnsHeaderIndexSize = 0
2023-06-20 05:55:12 +00:00
bh . columnsHeaderOffset = 0
bh . columnsHeaderSize = 0
}
func ( bh * blockHeader ) copyFrom ( src * blockHeader ) {
bh . reset ( )
bh . streamID = src . streamID
bh . uncompressedSizeBytes = src . uncompressedSizeBytes
bh . rowsCount = src . rowsCount
bh . timestampsHeader . copyFrom ( & src . timestampsHeader )
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 14:18:28 +00:00
bh . columnsHeaderIndexOffset = src . columnsHeaderIndexOffset
bh . columnsHeaderIndexSize = src . columnsHeaderIndexSize
2023-06-20 05:55:12 +00:00
bh . columnsHeaderOffset = src . columnsHeaderOffset
bh . columnsHeaderSize = src . columnsHeaderSize
}
// marshal appends the marshaled bh to dst and returns the result.
func ( bh * blockHeader ) marshal ( dst [ ] byte ) [ ] byte {
dst = bh . streamID . marshal ( dst )
dst = encoding . MarshalVarUint64 ( dst , bh . uncompressedSizeBytes )
dst = encoding . MarshalVarUint64 ( dst , bh . rowsCount )
dst = bh . timestampsHeader . marshal ( dst )
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 14:18:28 +00:00
dst = encoding . MarshalVarUint64 ( dst , bh . columnsHeaderIndexOffset )
dst = encoding . MarshalVarUint64 ( dst , bh . columnsHeaderIndexSize )
2023-06-20 05:55:12 +00:00
dst = encoding . MarshalVarUint64 ( dst , bh . columnsHeaderOffset )
dst = encoding . MarshalVarUint64 ( dst , bh . columnsHeaderSize )
return dst
}
// unmarshal unmarshals bh from src and returns the remaining tail.
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 14:18:28 +00:00
func ( bh * blockHeader ) unmarshal ( src [ ] byte , partFormatVersion uint ) ( [ ] byte , error ) {
2023-06-20 05:55:12 +00:00
bh . reset ( )
srcOrig := src
// unmarshal bh.streamID
tail , err := bh . streamID . unmarshal ( src )
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal streamID: %w" , err )
}
src = tail
// unmarshal bh.uncompressedSizeBytes
2024-05-13 23:23:44 +00:00
n , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal uncompressedSizeBytes" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
bh . uncompressedSizeBytes = n
// unmarshal bh.rowsCount
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal rowsCount" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n > maxRowsPerBlock {
return srcOrig , fmt . Errorf ( "too big value for rowsCount: %d; mustn't exceed %d" , n , maxRowsPerBlock )
}
bh . rowsCount = n
// unmarshal bh.timestampsHeader
tail , err = bh . timestampsHeader . unmarshal ( src )
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal timestampsHeader: %w" , err )
}
src = tail
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 14:18:28 +00:00
if partFormatVersion >= 1 {
// unmarshal columnsHeaderIndexOffset
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal columnsHeaderIndexOffset" )
}
src = src [ nSize : ]
bh . columnsHeaderIndexOffset = n
// unmarshal columnsHeaderIndexSize
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal columnsHeaderIndexSize" )
}
src = src [ nSize : ]
bh . columnsHeaderIndexSize = n
}
2023-06-20 05:55:12 +00:00
// unmarshal columnsHeaderOffset
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal columnsHeaderOffset" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
bh . columnsHeaderOffset = n
// unmarshal columnsHeaderSize
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal columnsHeaderSize" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n > maxColumnsHeaderSize {
return srcOrig , fmt . Errorf ( "too big value for columnsHeaderSize: %d; mustn't exceed %d" , n , maxColumnsHeaderSize )
}
bh . columnsHeaderSize = n
return src , nil
}
func getBlockHeader ( ) * blockHeader {
v := blockHeaderPool . Get ( )
if v == nil {
return & blockHeader { }
}
return v . ( * blockHeader )
}
func putBlockHeader ( bh * blockHeader ) {
bh . reset ( )
blockHeaderPool . Put ( bh )
}
var blockHeaderPool sync . Pool
// unmarshalBlockHeaders appends unmarshaled from src blockHeader entries to dst and returns the result.
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 14:18:28 +00:00
func unmarshalBlockHeaders ( dst [ ] blockHeader , src [ ] byte , partFormatVersion uint ) ( [ ] blockHeader , error ) {
dstLen := len ( dst )
2023-06-20 05:55:12 +00:00
for len ( src ) > 0 {
if len ( dst ) < cap ( dst ) {
dst = dst [ : len ( dst ) + 1 ]
} else {
dst = append ( dst , blockHeader { } )
}
bh := & dst [ len ( dst ) - 1 ]
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 14:18:28 +00:00
tail , err := bh . unmarshal ( src , partFormatVersion )
2023-06-20 05:55:12 +00:00
if err != 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 14:18:28 +00:00
return dst , fmt . Errorf ( "cannot unmarshal blockHeader entries: %w" , err )
2023-06-20 05:55:12 +00:00
}
src = tail
}
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 14:18:28 +00:00
if err := validateBlockHeaders ( dst [ dstLen : ] ) ; err != nil {
return dst , err
2023-06-20 05:55:12 +00:00
}
return dst , nil
}
func validateBlockHeaders ( bhs [ ] blockHeader ) error {
for i := 1 ; i < len ( bhs ) ; i ++ {
bhCurr := & bhs [ i ]
bhPrev := & bhs [ i - 1 ]
if bhCurr . streamID . less ( & bhPrev . streamID ) {
return fmt . Errorf ( "unexpected blockHeader with smaller streamID=%s after bigger streamID=%s at position %d" , & bhCurr . streamID , & bhPrev . streamID , i )
}
if ! bhCurr . streamID . equal ( & bhPrev . streamID ) {
continue
}
thCurr := bhCurr . timestampsHeader
thPrev := bhPrev . timestampsHeader
if thCurr . minTimestamp < thPrev . minTimestamp {
return fmt . Errorf ( "unexpected blockHeader with smaller timestamp=%d after bigger timestamp=%d at position %d" , thCurr . minTimestamp , thPrev . minTimestamp , i )
}
}
return nil
}
func resetBlockHeaders ( bhs [ ] blockHeader ) [ ] blockHeader {
for i := range bhs {
bhs [ i ] . reset ( )
}
return bhs [ : 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 14:18:28 +00:00
// columnHeaderRef references column header in the marshaled columnsHeader.
type columnHeaderRef struct {
// columnNameID is the ID of the column name. The column name can be obtained from part.columnNames.
columnNameID uint64
// offset is the offset of the the corresponding columnHeader inside marshaled columnsHeader.
offset uint64
}
// columnsHeaderIndex contains offsets for marshaled column headers.
type columnsHeaderIndex struct {
// columnHeadersRefs contains references to columnHeaders.
columnHeadersRefs [ ] columnHeaderRef
// constColumnsRefs contains references to constColumns.
constColumnsRefs [ ] columnHeaderRef
}
func getColumnsHeaderIndex ( ) * columnsHeaderIndex {
v := columnsHeaderIndexPool . Get ( )
if v == nil {
return & columnsHeaderIndex { }
}
return v . ( * columnsHeaderIndex )
}
func putColumnsHeaderIndex ( cshIndex * columnsHeaderIndex ) {
cshIndex . reset ( )
columnsHeaderIndexPool . Put ( cshIndex )
}
var columnsHeaderIndexPool sync . Pool
func ( cshIndex * columnsHeaderIndex ) reset ( ) {
clear ( cshIndex . columnHeadersRefs )
cshIndex . columnHeadersRefs = cshIndex . columnHeadersRefs [ : 0 ]
clear ( cshIndex . constColumnsRefs )
cshIndex . constColumnsRefs = cshIndex . constColumnsRefs [ : 0 ]
}
func ( cshIndex * columnsHeaderIndex ) resizeConstColumnsRefs ( n int ) [ ] columnHeaderRef {
cshIndex . constColumnsRefs = slicesutil . SetLength ( cshIndex . constColumnsRefs , n )
return cshIndex . constColumnsRefs
}
func ( cshIndex * columnsHeaderIndex ) resizeColumnHeadersRefs ( n int ) [ ] columnHeaderRef {
cshIndex . columnHeadersRefs = slicesutil . SetLength ( cshIndex . columnHeadersRefs , n )
return cshIndex . columnHeadersRefs
}
func ( cshIndex * columnsHeaderIndex ) marshal ( dst [ ] byte ) [ ] byte {
dst = marshalColumnHeadersRefs ( dst , cshIndex . columnHeadersRefs )
dst = marshalColumnHeadersRefs ( dst , cshIndex . constColumnsRefs )
return dst
}
// unmarshalNoArena unmarshals cshIndex from src.
//
// cshIndex is valid until src is changed.
func ( cshIndex * columnsHeaderIndex ) unmarshalNoArena ( src [ ] byte ) error {
cshIndex . reset ( )
refs , tail , err := unmarshalColumnHeadersRefsNoArena ( cshIndex . columnHeadersRefs [ : 0 ] , src )
if err != nil {
return fmt . Errorf ( "cannot unmarshal columnHeadersRefs: %w" , err )
}
cshIndex . columnHeadersRefs = refs
src = tail
refs , tail , err = unmarshalColumnHeadersRefsNoArena ( cshIndex . constColumnsRefs [ : 0 ] , src )
if err != nil {
return fmt . Errorf ( "cannot unmarshal constColumnsRefs: %w" , err )
}
cshIndex . constColumnsRefs = refs
if len ( tail ) > 0 {
return fmt . Errorf ( "unexpected non-empty tail left after unmarshaling columnsHeaderIndex; len(tail)=%d" , len ( tail ) )
}
return nil
}
func marshalColumnHeadersRefs ( dst [ ] byte , refs [ ] columnHeaderRef ) [ ] byte {
dst = encoding . MarshalVarUint64 ( dst , uint64 ( len ( refs ) ) )
for _ , r := range refs {
dst = encoding . MarshalVarUint64 ( dst , r . columnNameID )
dst = encoding . MarshalVarUint64 ( dst , r . offset )
}
return dst
}
func unmarshalColumnHeadersRefsNoArena ( dst [ ] columnHeaderRef , src [ ] byte ) ( [ ] columnHeaderRef , [ ] byte , error ) {
srcOrig := src
n , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return dst , srcOrig , fmt . Errorf ( "cannot unmarshal the number of columnHeaderRef items" )
}
src = src [ nSize : ]
for i := uint64 ( 0 ) ; i < n ; i ++ {
columnNameID , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return dst , srcOrig , fmt . Errorf ( "cannot unmarshal column name ID number %d out of %d" , i , n )
}
src = src [ nSize : ]
offset , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return dst , srcOrig , fmt . Errorf ( "cannot unmarshal offset number %d out of %d" , i , n )
}
src = src [ nSize : ]
dst = append ( dst , columnHeaderRef {
columnNameID : columnNameID ,
offset : offset ,
} )
}
return dst , src , nil
}
2023-06-20 05:55:12 +00:00
func getColumnsHeader ( ) * columnsHeader {
v := columnsHeaderPool . Get ( )
if v == nil {
return & columnsHeader { }
}
return v . ( * columnsHeader )
}
func putColumnsHeader ( csh * columnsHeader ) {
csh . reset ( )
columnsHeaderPool . Put ( csh )
}
var columnsHeaderPool sync . Pool
// columnsHeader contains information about columns in a single block.
//
// columnsHeader is stored in the columnsHeaderFilename file.
type columnsHeader struct {
// columnHeaders contains the information about every column seen in the block.
columnHeaders [ ] columnHeader
// constColumns contain fields with constant values across all the block entries.
constColumns [ ] Field
}
func ( csh * columnsHeader ) reset ( ) {
chs := csh . columnHeaders
for i := range chs {
chs [ i ] . reset ( )
}
csh . columnHeaders = chs [ : 0 ]
ccs := csh . constColumns
for i := range ccs {
ccs [ i ] . Reset ( )
}
csh . constColumns = ccs [ : 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 14:18:28 +00:00
func ( csh * columnsHeader ) resizeConstColumns ( n int ) [ ] Field {
csh . constColumns = slicesutil . SetLength ( csh . constColumns , n )
2024-05-12 14:33:29 +00:00
return csh . constColumns
2023-06-20 05:55:12 +00:00
}
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 14:18:28 +00:00
func ( csh * columnsHeader ) resizeColumnHeaders ( n int ) [ ] columnHeader {
csh . columnHeaders = slicesutil . SetLength ( csh . columnHeaders , n )
2024-05-12 14:33:29 +00:00
return csh . columnHeaders
2023-06-20 05:55:12 +00:00
}
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 14:18:28 +00:00
func ( csh * columnsHeader ) setColumnNames ( cshIndex * columnsHeaderIndex , columnNames [ ] string ) error {
if len ( cshIndex . columnHeadersRefs ) != len ( csh . columnHeaders ) {
return fmt . Errorf ( "unpexected number of column headers; got %d; want %d" , len ( cshIndex . columnHeadersRefs ) , len ( csh . columnHeaders ) )
}
for i := range csh . columnHeaders {
columnNameID := cshIndex . columnHeadersRefs [ i ] . columnNameID
if columnNameID >= uint64 ( len ( columnNames ) ) {
return fmt . Errorf ( "unexpected columnNameID=%d in columnHeadersRef; len(columnNames)=%d; columnNames=%v" , columnNameID , len ( columnNames ) , columnNames )
}
csh . columnHeaders [ i ] . name = columnNames [ columnNameID ]
}
if len ( cshIndex . constColumnsRefs ) != len ( csh . constColumns ) {
return fmt . Errorf ( "unexpected number of const columns; got %d; want %d" , len ( cshIndex . constColumnsRefs ) , len ( csh . constColumns ) )
}
for i := range csh . constColumns {
columnNameID := cshIndex . constColumnsRefs [ i ] . columnNameID
if columnNameID >= uint64 ( len ( columnNames ) ) {
return fmt . Errorf ( "unexpected columnNameID=%d in constColumnsRefs; len(columnNames)=%d; columnNames=%v" , columnNameID , len ( columnNames ) , columnNames )
}
csh . constColumns [ i ] . Name = columnNames [ columnNameID ]
}
return nil
}
func ( csh * columnsHeader ) mustWriteTo ( bh * blockHeader , sw * streamWriters , g * columnNameIDGenerator ) {
bb := longTermBufPool . Get ( )
defer longTermBufPool . Put ( bb )
cshIndex := getColumnsHeaderIndex ( )
bb . B = csh . marshal ( bb . B , cshIndex , g )
columnsHeaderData := bb . B
bb . B = cshIndex . marshal ( bb . B )
columnsHeaderIndexData := bb . B [ len ( columnsHeaderData ) : ]
putColumnsHeaderIndex ( cshIndex )
bh . columnsHeaderIndexOffset = sw . columnsHeaderIndexWriter . bytesWritten
bh . columnsHeaderIndexSize = uint64 ( len ( columnsHeaderIndexData ) )
if bh . columnsHeaderIndexSize > maxColumnsHeaderIndexSize {
logger . Panicf ( "BUG: too big columnsHeaderIndexSize: %d bytes; mustn't exceed %d bytes" , bh . columnsHeaderIndexSize , maxColumnsHeaderIndexSize )
}
sw . columnsHeaderIndexWriter . MustWrite ( columnsHeaderIndexData )
bh . columnsHeaderOffset = sw . columnsHeaderWriter . bytesWritten
bh . columnsHeaderSize = uint64 ( len ( columnsHeaderData ) )
if bh . columnsHeaderSize > maxColumnsHeaderSize {
logger . Panicf ( "BUG: too big columnsHeaderSize: %d bytes; mustn't exceed %d bytes" , bh . columnsHeaderSize , maxColumnsHeaderSize )
}
sw . columnsHeaderWriter . MustWrite ( columnsHeaderData )
}
func ( csh * columnsHeader ) marshal ( dst [ ] byte , cshIndex * columnsHeaderIndex , g * columnNameIDGenerator ) [ ] byte {
dstLen := len ( dst )
2023-06-20 05:55:12 +00:00
chs := csh . columnHeaders
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 14:18:28 +00:00
chsRefs := cshIndex . resizeColumnHeadersRefs ( len ( chs ) )
2023-06-20 05:55:12 +00:00
dst = encoding . MarshalVarUint64 ( dst , uint64 ( len ( chs ) ) )
for i := range chs {
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 14:18:28 +00:00
columnNameID := g . getColumnNameID ( chs [ i ] . name )
offset := len ( dst ) - dstLen
2023-06-20 05:55:12 +00:00
dst = chs [ i ] . marshal ( dst )
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 14:18:28 +00:00
chsRefs [ i ] = columnHeaderRef {
columnNameID : columnNameID ,
offset : uint64 ( offset ) ,
}
2023-06-20 05:55:12 +00:00
}
ccs := csh . constColumns
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 14:18:28 +00:00
ccsRefs := cshIndex . resizeConstColumnsRefs ( len ( ccs ) )
2023-06-20 05:55:12 +00:00
dst = encoding . MarshalVarUint64 ( dst , uint64 ( len ( ccs ) ) )
for i := range ccs {
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 14:18:28 +00:00
columnNameID := g . getColumnNameID ( ccs [ i ] . Name )
offset := len ( dst ) - dstLen
dst = ccs [ i ] . marshal ( dst , false )
ccsRefs [ i ] = columnHeaderRef {
columnNameID : columnNameID ,
offset : uint64 ( offset ) ,
}
2023-06-20 05:55:12 +00:00
}
return dst
}
2024-10-13 11:25:36 +00:00
// unmarshalNoArena unmarshals csh from src.
2024-05-12 14:33:29 +00:00
//
2024-10-13 11:25:36 +00:00
// csh is valid until src is changed.
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 14:18:28 +00:00
func ( csh * columnsHeader ) unmarshalNoArena ( src [ ] byte , partFormatVersion uint ) error {
2023-06-20 05:55:12 +00:00
csh . reset ( )
// unmarshal columnHeaders
2024-05-13 23:23:44 +00:00
n , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return fmt . Errorf ( "cannot unmarshal columnHeaders len" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n > maxColumnsPerBlock {
return fmt . Errorf ( "too many column headers: %d; mustn't exceed %d" , n , maxColumnsPerBlock )
}
2024-05-13 23:23:44 +00:00
2023-06-20 05:55:12 +00:00
chs := csh . resizeColumnHeaders ( int ( n ) )
for i := range chs {
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 14:18:28 +00:00
tail , err := chs [ i ] . unmarshalNoArena ( src , partFormatVersion )
2023-06-20 05:55:12 +00:00
if err != nil {
return fmt . Errorf ( "cannot unmarshal columnHeader %d out of %d columnHeaders: %w" , i , len ( chs ) , err )
}
src = tail
}
csh . columnHeaders = chs
// unmarshal constColumns
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return fmt . Errorf ( "cannot unmarshal constColumns len" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n + uint64 ( len ( csh . columnHeaders ) ) > maxColumnsPerBlock {
return fmt . Errorf ( "too many columns: %d; mustn't exceed %d" , n + uint64 ( len ( csh . columnHeaders ) ) , maxColumnsPerBlock )
}
2024-05-13 23:23:44 +00:00
2023-06-20 05:55:12 +00:00
ccs := csh . resizeConstColumns ( int ( n ) )
for i := range ccs {
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 14:18:28 +00:00
tail , err := ccs [ i ] . unmarshalNoArena ( src , partFormatVersion < 1 )
2023-06-20 05:55:12 +00:00
if err != nil {
return fmt . Errorf ( "cannot unmarshal constColumn %d out of %d columns: %w" , i , len ( ccs ) , err )
}
src = tail
}
// Verify that the src is empty
if len ( src ) > 0 {
return fmt . Errorf ( "unexpected non-empty tail left after unmarshaling columnsHeader: len(tail)=%d" , len ( src ) )
}
return nil
}
// columnHeaders contains information for values, which belong to a single label in a single block.
//
// The main column with an empty name is stored in messageValuesFilename,
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 14:18:28 +00:00
// while the rest of columns are stored in smallValuesFilename or bigValuesFilename depending
// on the block size (see maxSmallValuesBlockSize).
2023-06-20 05:55:12 +00:00
// This allows minimizing disk read IO when filtering by non-message columns.
//
// Every block column contains also a bloom filter for all the tokens stored in the column.
// This bloom filter is used for fast determining whether the given block may contain the given tokens.
//
// Tokens in bloom filter depend on valueType:
//
2024-05-12 14:33:29 +00:00
// - valueTypeString stores tokens seen in all the values
2023-06-20 05:55:12 +00:00
// - valueTypeDict doesn't store anything in the bloom filter, since all the encoded values
// are available directly in the valuesDict field
// - valueTypeUint8, valueTypeUint16, valueTypeUint32 and valueTypeUint64 stores encoded uint values
// - valueTypeFloat64 stores encoded float64 values
// - valueTypeIPv4 stores encoded into uint32 ips
// - valueTypeTimestampISO8601 stores encoded into uint64 timestamps
//
// Bloom filters for main column with an empty name is stored in messageBloomFilename,
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 14:18:28 +00:00
// while the rest of columns are stored in smallBloomFilename or bigBloomFilename depending on their size
// (see maxSmallBloomFilterBlockSize).
2023-06-20 05:55:12 +00:00
type columnHeader struct {
// name contains column name aka label name
name string
// valueType is the type of values stored in the block
valueType valueType
// minValue is the minimum encoded value for uint*, ipv4, timestamp and float64 value in the columnHeader
//
// It is used for fast detection of whether the given columnHeader contains values in the given range
minValue uint64
// maxValue is the maximum encoded value for uint*, ipv4, timestamp and float64 value in the columnHeader
//
// It is used for fast detection of whether the given columnHeader contains values in the given range
maxValue uint64
// valuesDict contains unique values for valueType = valueTypeDict
valuesDict valuesDict
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 14:18:28 +00:00
// valuesOffset contains the offset of the block in either messageValuesFilename, smallValuesFilename or bigValuesFilename
2023-06-20 05:55:12 +00:00
valuesOffset uint64
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 14:18:28 +00:00
// valuesSize contains the size of the block in either messageValuesFilename, smallValuesFilename or bigValuesFilename
2023-06-20 05:55:12 +00:00
valuesSize uint64
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 14:18:28 +00:00
// bloomFilterOffset contains the offset of the bloom filter in messageBloomFilename, smallBloomFilename or bigBloomFilename
2023-06-20 05:55:12 +00:00
bloomFilterOffset uint64
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 14:18:28 +00:00
// bloomFilterSize contains the size of the bloom filter in messageBloomFilename, smallBloomFilename or bigBloomFilename
2023-06-20 05:55:12 +00:00
bloomFilterSize uint64
}
// reset resets ch
func ( ch * columnHeader ) reset ( ) {
ch . name = ""
ch . valueType = 0
ch . minValue = 0
ch . maxValue = 0
ch . valuesDict . reset ( )
ch . valuesOffset = 0
ch . valuesSize = 0
ch . bloomFilterOffset = 0
ch . bloomFilterSize = 0
}
// marshal appends marshaled ch to dst and returns the result.
func ( ch * columnHeader ) marshal ( dst [ ] byte ) [ ] byte {
// check minValue/maxValue
if ch . valueType == valueTypeFloat64 {
minValue := math . Float64frombits ( ch . minValue )
maxValue := math . Float64frombits ( ch . maxValue )
if minValue > maxValue {
2024-05-12 14:33:29 +00:00
logger . Panicf ( "BUG: minValue=%g must be smaller than maxValue=%g for valueTypeFloat64" , minValue , maxValue )
2023-06-20 05:55:12 +00:00
}
2024-05-12 14:33:29 +00:00
} else if ch . valueType == valueTypeTimestampISO8601 {
minValue := int64 ( ch . minValue )
maxValue := int64 ( ch . maxValue )
if minValue > maxValue {
logger . Panicf ( "BUG: minValue=%g must be smaller than maxValue=%g for valueTypeTimestampISO8601" , minValue , maxValue )
2023-06-20 05:55:12 +00:00
}
2024-05-12 14:33:29 +00:00
} else if ch . minValue > ch . maxValue {
logger . Panicf ( "BUG: minValue=%d must be smaller than maxValue=%d for valueType=%d" , ch . minValue , ch . maxValue , ch . valueType )
2023-06-20 05:55:12 +00:00
}
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 14:18:28 +00:00
// Do not encode ch.name, since it should be encoded at columnsHeaderIndex.columnHeadersRefs
// Encode common field - ch.valueType
2023-06-20 05:55:12 +00:00
dst = append ( dst , byte ( ch . valueType ) )
// Encode other fields depending on ch.valueType
switch ch . valueType {
case valueTypeString :
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeDict :
dst = ch . valuesDict . marshal ( dst )
dst = ch . marshalValues ( dst )
case valueTypeUint8 :
dst = append ( dst , byte ( ch . minValue ) )
dst = append ( dst , byte ( ch . maxValue ) )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeUint16 :
dst = encoding . MarshalUint16 ( dst , uint16 ( ch . minValue ) )
dst = encoding . MarshalUint16 ( dst , uint16 ( ch . maxValue ) )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeUint32 :
dst = encoding . MarshalUint32 ( dst , uint32 ( ch . minValue ) )
dst = encoding . MarshalUint32 ( dst , uint32 ( ch . maxValue ) )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeUint64 :
dst = encoding . MarshalUint64 ( dst , ch . minValue )
dst = encoding . MarshalUint64 ( dst , ch . maxValue )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeFloat64 :
// float64 values are encoded as uint64 via math.Float64bits()
dst = encoding . MarshalUint64 ( dst , ch . minValue )
dst = encoding . MarshalUint64 ( dst , ch . maxValue )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeIPv4 :
dst = encoding . MarshalUint32 ( dst , uint32 ( ch . minValue ) )
dst = encoding . MarshalUint32 ( dst , uint32 ( ch . maxValue ) )
dst = ch . marshalValuesAndBloomFilters ( dst )
case valueTypeTimestampISO8601 :
// timestamps are encoded in nanoseconds
dst = encoding . MarshalUint64 ( dst , ch . minValue )
dst = encoding . MarshalUint64 ( dst , ch . maxValue )
dst = ch . marshalValuesAndBloomFilters ( dst )
default :
logger . Panicf ( "BUG: unknown valueType=%d" , ch . valueType )
}
return dst
}
func ( ch * columnHeader ) marshalValuesAndBloomFilters ( dst [ ] byte ) [ ] byte {
dst = ch . marshalValues ( dst )
dst = ch . marshalBloomFilters ( dst )
return dst
}
func ( ch * columnHeader ) marshalValues ( dst [ ] byte ) [ ] byte {
dst = encoding . MarshalVarUint64 ( dst , ch . valuesOffset )
dst = encoding . MarshalVarUint64 ( dst , ch . valuesSize )
return dst
}
func ( ch * columnHeader ) marshalBloomFilters ( dst [ ] byte ) [ ] byte {
dst = encoding . MarshalVarUint64 ( dst , ch . bloomFilterOffset )
dst = encoding . MarshalVarUint64 ( dst , ch . bloomFilterSize )
return dst
}
2024-10-13 11:25:36 +00:00
// unmarshalNoArena unmarshals ch from src and returns the tail left after unmarshaling.
2024-05-12 14:33:29 +00:00
//
2024-10-13 11:25:36 +00:00
// ch is valid until src is changed.
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 14:18:28 +00:00
func ( ch * columnHeader ) unmarshalNoArena ( src [ ] byte , partFormatVersion uint ) ( [ ] byte , error ) {
2023-06-20 05:55:12 +00:00
ch . reset ( )
srcOrig := src
// Unmarshal column name
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 14:18:28 +00:00
if partFormatVersion < 1 {
data , nSize := encoding . UnmarshalBytes ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal column name" )
}
src = src [ nSize : ]
ch . name = bytesutil . ToUnsafeString ( data )
2023-06-20 05:55:12 +00:00
}
// Unmarshal value type
if len ( src ) < 1 {
return srcOrig , fmt . Errorf ( "cannot unmarshal valueType from 0 bytes for column %q; need at least 1 byte" , ch . name )
}
ch . valueType = valueType ( src [ 0 ] )
src = src [ 1 : ]
// Unmarshal the rest of data depending on valueType
switch ch . valueType {
case valueTypeString :
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeString for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeDict :
2024-10-13 11:25:36 +00:00
tail , err := ch . valuesDict . unmarshalNoArena ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal dict at valueTypeDict for column %q: %w" , ch . name , err )
}
src = tail
tail , err = ch . unmarshalValues ( src )
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values at valueTypeDict for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeUint8 :
if len ( src ) < 2 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeUint8 from %d bytes for column %q; need at least 2 bytes" , len ( src ) , ch . name )
}
ch . minValue = uint64 ( src [ 0 ] )
ch . maxValue = uint64 ( src [ 1 ] )
src = src [ 2 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeUint8 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeUint16 :
if len ( src ) < 4 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeUint16 from %d bytes for column %q; need at least 4 bytes" , len ( src ) , ch . name )
}
ch . minValue = uint64 ( encoding . UnmarshalUint16 ( src ) )
ch . maxValue = uint64 ( encoding . UnmarshalUint16 ( src [ 2 : ] ) )
src = src [ 4 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeUint16 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeUint32 :
if len ( src ) < 8 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeUint32 from %d bytes for column %q; need at least 8 bytes" , len ( src ) , ch . name )
}
ch . minValue = uint64 ( encoding . UnmarshalUint32 ( src ) )
ch . maxValue = uint64 ( encoding . UnmarshalUint32 ( src [ 4 : ] ) )
src = src [ 8 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeUint32 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeUint64 :
if len ( src ) < 16 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeUint64 from %d bytes for column %q; need at least 16 bytes" , len ( src ) , ch . name )
}
ch . minValue = encoding . UnmarshalUint64 ( src )
ch . maxValue = encoding . UnmarshalUint64 ( src [ 8 : ] )
src = src [ 16 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeUint64 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeFloat64 :
if len ( src ) < 16 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeFloat64 from %d bytes for column %q; need at least 16 bytes" , len ( src ) , ch . name )
}
// min and max values must be converted to real values with math.Float64frombits() during querying.
ch . minValue = encoding . UnmarshalUint64 ( src )
ch . maxValue = encoding . UnmarshalUint64 ( src [ 8 : ] )
src = src [ 16 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeFloat64 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeIPv4 :
if len ( src ) < 8 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeIPv4 from %d bytes for column %q; need at least 8 bytes" , len ( src ) , ch . name )
}
ch . minValue = uint64 ( encoding . UnmarshalUint32 ( src ) )
ch . maxValue = uint64 ( encoding . UnmarshalUint32 ( src [ 4 : ] ) )
src = src [ 8 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeIPv4 for column %q: %w" , ch . name , err )
}
src = tail
case valueTypeTimestampISO8601 :
if len ( src ) < 16 {
return srcOrig , fmt . Errorf ( "cannot unmarshal min/max values at valueTypeTimestampISO8601 from %d bytes for column %q; need at least 16 bytes" ,
len ( src ) , ch . name )
}
ch . minValue = encoding . UnmarshalUint64 ( src )
ch . maxValue = encoding . UnmarshalUint64 ( src [ 8 : ] )
src = src [ 16 : ]
2024-05-13 23:23:44 +00:00
tail , err := ch . unmarshalValuesAndBloomFilters ( src )
2023-06-20 05:55:12 +00:00
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values and bloom filters at valueTypeTimestampISO8601 for column %q: %w" , ch . name , err )
}
src = tail
default :
return srcOrig , fmt . Errorf ( "unexpected valueType=%d for column %q" , ch . valueType , ch . name )
}
return src , nil
}
func ( ch * columnHeader ) unmarshalValuesAndBloomFilters ( src [ ] byte ) ( [ ] byte , error ) {
srcOrig := src
tail , err := ch . unmarshalValues ( src )
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal values: %w" , err )
}
src = tail
tail , err = ch . unmarshalBloomFilters ( src )
if err != nil {
return srcOrig , fmt . Errorf ( "cannot unmarshal bloom filters: %w" , err )
}
src = tail
return src , nil
}
func ( ch * columnHeader ) unmarshalValues ( src [ ] byte ) ( [ ] byte , error ) {
srcOrig := src
2024-05-13 23:23:44 +00:00
n , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal valuesOffset" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
ch . valuesOffset = n
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal valuesSize" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n > maxValuesBlockSize {
return srcOrig , fmt . Errorf ( "too big valuesSize: %d bytes; mustn't exceed %d bytes" , n , maxValuesBlockSize )
}
ch . valuesSize = n
return src , nil
}
func ( ch * columnHeader ) unmarshalBloomFilters ( src [ ] byte ) ( [ ] byte , error ) {
srcOrig := src
2024-05-13 23:23:44 +00:00
n , nSize := encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal bloomFilterOffset" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
ch . bloomFilterOffset = n
2024-05-13 23:23:44 +00:00
n , nSize = encoding . UnmarshalVarUint64 ( src )
if nSize <= 0 {
return srcOrig , fmt . Errorf ( "cannot unmarshal bloomFilterSize" )
2023-06-20 05:55:12 +00:00
}
2024-05-13 23:23:44 +00:00
src = src [ nSize : ]
2023-06-20 05:55:12 +00:00
if n > maxBloomFilterBlockSize {
return srcOrig , fmt . Errorf ( "too big bloomFilterSize: %d bytes; mustn't exceed %d bytes" , n , maxBloomFilterBlockSize )
}
ch . bloomFilterSize = n
return src , nil
}
// timestampsHeader contains the information about timestamps block.
type timestampsHeader struct {
// blockOffset is an offset of timestamps block inside timestampsFilename file
blockOffset uint64
// blockSize is the size of the timestamps block inside timestampsFilename file
blockSize uint64
2024-05-12 14:33:29 +00:00
// minTimestamp is the mimumum timestamp seen in the block in nanoseconds
2023-06-20 05:55:12 +00:00
minTimestamp int64
2024-05-12 14:33:29 +00:00
// maxTimestamp is the maximum timestamp seen in the block in nanoseconds
2023-06-20 05:55:12 +00:00
maxTimestamp int64
// marshalType is the type used for encoding the timestamps block
marshalType encoding . MarshalType
}
// reset resets th, so it can be reused
func ( th * timestampsHeader ) reset ( ) {
th . blockOffset = 0
th . blockSize = 0
th . minTimestamp = 0
th . maxTimestamp = 0
th . marshalType = 0
}
func ( th * timestampsHeader ) copyFrom ( src * timestampsHeader ) {
th . blockOffset = src . blockOffset
th . blockSize = src . blockSize
th . minTimestamp = src . minTimestamp
th . maxTimestamp = src . maxTimestamp
th . marshalType = src . marshalType
}
// marshal appends marshaled th to dst and returns the result.
func ( th * timestampsHeader ) marshal ( dst [ ] byte ) [ ] byte {
dst = encoding . MarshalUint64 ( dst , th . blockOffset )
dst = encoding . MarshalUint64 ( dst , th . blockSize )
dst = encoding . MarshalUint64 ( dst , uint64 ( th . minTimestamp ) )
dst = encoding . MarshalUint64 ( dst , uint64 ( th . maxTimestamp ) )
dst = append ( dst , byte ( th . marshalType ) )
return dst
}
// unmarshal unmarshals th from src and returns the tail left after the unmarshaling.
func ( th * timestampsHeader ) unmarshal ( src [ ] byte ) ( [ ] byte , error ) {
th . reset ( )
if len ( src ) < 33 {
return src , fmt . Errorf ( "cannot unmarshal timestampsHeader from %d bytes; need at least 33 bytes" , len ( src ) )
}
th . blockOffset = encoding . UnmarshalUint64 ( src )
th . blockSize = encoding . UnmarshalUint64 ( src [ 8 : ] )
th . minTimestamp = int64 ( encoding . UnmarshalUint64 ( src [ 16 : ] ) )
th . maxTimestamp = int64 ( encoding . UnmarshalUint64 ( src [ 24 : ] ) )
th . marshalType = encoding . MarshalType ( src [ 32 ] )
return src [ 33 : ] , nil
}