VictoriaMetrics/lib/logstorage/block_data.go

399 lines
13 KiB
Go
Raw Normal View History

package logstorage
import (
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/slicesutil"
)
// blockData contains packed data for a single block.
//
// The main purpose of this struct is to reduce the work needed during background merge of parts.
// If the block is full, then the blockData can be written to the destination part
// without the need to unpack it.
type blockData struct {
// streamID is id of the stream for the data
streamID streamID
// uncompressedSizeBytes is the original (uncompressed) size of log entries stored in the block
uncompressedSizeBytes uint64
// rowsCount is the number of log entries in the block
rowsCount uint64
// timestampsData contains the encoded timestamps data for the block
timestampsData timestampsData
lib/logstorage: follow-up for 94627113dbe0b7bed23a7dc4864fc2f2903a819c - Move uniqueFields from rows to blockStreamMerger struct. This allows localizing all the references to uniqueFields inside blockStreamMerger.mustWriteBlock(), which should improve readability and maintainability of the code. - Remove logging of the event when blocks cannot be merged because they contain more than maxColumnsPerBlock, since the provided logging didn't provide the solution for the issue with too many columns. I couldn't figure out the proper solution, which could be helpful for end user, so decided to remove the logging until we find the solution. This commit also contains the following additional changes: - It truncates field names longer than 128 chars during logs ingestion. This should prevent from ingesting bogus field names. This also should prevent from too big columnsHeader blocks, which could negatively affect search query performance, since columnsHeader is read on every scan of the corresponding data block. - It limits the maximum length of const column value to 256. Longer values are stored in an ordinary columns. This helps limiting the size of columnsHeader blocks and improving search query performance by avoiding reading too long const columns on every scan of the corresponding data block. - It deduplicates columns with identical names during data ingestion and background merging. Previously it was possible to pass columns with duplicate names to block.mustInitFromRows(), and they were stored as is in the block. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4762 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/4969
2023-10-02 19:01:17 +02:00
// columnsData contains packed per-column data
columnsData []columnData
lib/logstorage: follow-up for 94627113dbe0b7bed23a7dc4864fc2f2903a819c - Move uniqueFields from rows to blockStreamMerger struct. This allows localizing all the references to uniqueFields inside blockStreamMerger.mustWriteBlock(), which should improve readability and maintainability of the code. - Remove logging of the event when blocks cannot be merged because they contain more than maxColumnsPerBlock, since the provided logging didn't provide the solution for the issue with too many columns. I couldn't figure out the proper solution, which could be helpful for end user, so decided to remove the logging until we find the solution. This commit also contains the following additional changes: - It truncates field names longer than 128 chars during logs ingestion. This should prevent from ingesting bogus field names. This also should prevent from too big columnsHeader blocks, which could negatively affect search query performance, since columnsHeader is read on every scan of the corresponding data block. - It limits the maximum length of const column value to 256. Longer values are stored in an ordinary columns. This helps limiting the size of columnsHeader blocks and improving search query performance by avoiding reading too long const columns on every scan of the corresponding data block. - It deduplicates columns with identical names during data ingestion and background merging. Previously it was possible to pass columns with duplicate names to block.mustInitFromRows(), and they were stored as is in the block. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4762 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/4969
2023-10-02 19:01:17 +02:00
// constColumns contains data for const columns across the block
constColumns []Field
}
// reset resets bd for subsequent re-use
func (bd *blockData) reset() {
bd.streamID.reset()
bd.uncompressedSizeBytes = 0
bd.rowsCount = 0
bd.timestampsData.reset()
cds := bd.columnsData
for i := range cds {
cds[i].reset()
}
bd.columnsData = cds[:0]
ccs := bd.constColumns
for i := range ccs {
ccs[i].Reset()
}
bd.constColumns = ccs[:0]
}
func (bd *blockData) resizeColumnsData(columnsDataLen int) []columnData {
bd.columnsData = slicesutil.SetLength(bd.columnsData, columnsDataLen)
return bd.columnsData
}
// copyFrom copies src to bd.
//
// bd is valid until a.reset() is called.
func (bd *blockData) copyFrom(a *arena, src *blockData) {
bd.reset()
bd.streamID = src.streamID
bd.uncompressedSizeBytes = src.uncompressedSizeBytes
bd.rowsCount = src.rowsCount
bd.timestampsData.copyFrom(a, &src.timestampsData)
cdsSrc := src.columnsData
cds := bd.resizeColumnsData(len(cdsSrc))
for i := range cds {
cds[i].copyFrom(a, &cdsSrc[i])
}
bd.columnsData = cds
bd.constColumns = appendFields(a, bd.constColumns[:0], src.constColumns)
}
// unmarshalRows appends unmarshaled from bd log entries to dst.
//
// The unmarshaled log entries are valid until sbu and vd are reset.
func (bd *blockData) unmarshalRows(dst *rows, sbu *stringsBlockUnmarshaler, vd *valuesDecoder) error {
b := getBlock()
defer putBlock(b)
if err := b.InitFromBlockData(bd, sbu, vd); err != nil {
return err
}
lib/logstorage: follow-up for 94627113dbe0b7bed23a7dc4864fc2f2903a819c - Move uniqueFields from rows to blockStreamMerger struct. This allows localizing all the references to uniqueFields inside blockStreamMerger.mustWriteBlock(), which should improve readability and maintainability of the code. - Remove logging of the event when blocks cannot be merged because they contain more than maxColumnsPerBlock, since the provided logging didn't provide the solution for the issue with too many columns. I couldn't figure out the proper solution, which could be helpful for end user, so decided to remove the logging until we find the solution. This commit also contains the following additional changes: - It truncates field names longer than 128 chars during logs ingestion. This should prevent from ingesting bogus field names. This also should prevent from too big columnsHeader blocks, which could negatively affect search query performance, since columnsHeader is read on every scan of the corresponding data block. - It limits the maximum length of const column value to 256. Longer values are stored in an ordinary columns. This helps limiting the size of columnsHeader blocks and improving search query performance by avoiding reading too long const columns on every scan of the corresponding data block. - It deduplicates columns with identical names during data ingestion and background merging. Previously it was possible to pass columns with duplicate names to block.mustInitFromRows(), and they were stored as is in the block. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4762 Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/4969
2023-10-02 19:01:17 +02:00
b.appendRowsTo(dst)
return nil
}
// mustWriteTo writes bd to sw and updates bh accordingly
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
func (bd *blockData) mustWriteTo(bh *blockHeader, sw *streamWriters, g *columnNameIDGenerator) {
// Do not store the version used for encoding directly in the block data, since:
// - all the blocks in the same part use the same encoding
// - the block encoding version can be put in metadata file for the part (aka metadataFilename)
bh.reset()
bh.streamID = bd.streamID
bh.uncompressedSizeBytes = bd.uncompressedSizeBytes
bh.rowsCount = bd.rowsCount
// Marshal timestamps
bd.timestampsData.mustWriteTo(&bh.timestampsHeader, sw)
// Marshal columns
cds := bd.columnsData
csh := getColumnsHeader()
chs := csh.resizeColumnHeaders(len(cds))
for i := range cds {
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
cds[i].mustWriteTo(&chs[i], sw)
}
csh.constColumns = append(csh.constColumns[:0], bd.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 16:18:28 +02:00
csh.mustWriteTo(bh, sw, g)
putColumnsHeader(csh)
}
// mustReadFrom reads block data associated with bh from sr to bd.
//
// The bd is valid until a.reset() is called.
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
func (bd *blockData) mustReadFrom(a *arena, bh *blockHeader, sr *streamReaders, partFormatVersion uint, columnNames []string) {
bd.reset()
bd.streamID = bh.streamID
bd.uncompressedSizeBytes = bh.uncompressedSizeBytes
bd.rowsCount = bh.rowsCount
// Read timestamps
bd.timestampsData.mustReadFrom(a, &bh.timestampsHeader, sr)
// Read columns
if bh.columnsHeaderOffset != sr.columnsHeaderReader.bytesRead {
logger.Panicf("FATAL: %s: unexpected columnsHeaderOffset=%d; must equal to the number of bytes read: %d",
sr.columnsHeaderReader.Path(), bh.columnsHeaderOffset, sr.columnsHeaderReader.bytesRead)
}
columnsHeaderSize := bh.columnsHeaderSize
if columnsHeaderSize > maxColumnsHeaderSize {
logger.Panicf("BUG: %s: too big columnsHeaderSize: %d bytes; mustn't exceed %d bytes", sr.columnsHeaderReader.Path(), columnsHeaderSize, maxColumnsHeaderSize)
}
bb := longTermBufPool.Get()
bb.B = bytesutil.ResizeNoCopyMayOverallocate(bb.B, int(columnsHeaderSize))
sr.columnsHeaderReader.MustReadFull(bb.B)
csh := getColumnsHeader()
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
if err := csh.unmarshalNoArena(bb.B, partFormatVersion); err != nil {
logger.Panicf("FATAL: %s: cannot unmarshal columnsHeader: %s", sr.columnsHeaderReader.Path(), err)
}
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
if partFormatVersion >= 1 {
readColumnNamesFromColumnsHeaderIndex(bh, sr, csh, columnNames)
}
chs := csh.columnHeaders
cds := bd.resizeColumnsData(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 16:18:28 +02:00
cds[i].mustReadFrom(a, &chs[i], sr, partFormatVersion)
}
bd.constColumns = appendFields(a, bd.constColumns[:0], csh.constColumns)
putColumnsHeader(csh)
longTermBufPool.Put(bb)
}
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
func readColumnNamesFromColumnsHeaderIndex(bh *blockHeader, sr *streamReaders, csh *columnsHeader, columnNames []string) {
bb := longTermBufPool.Get()
defer longTermBufPool.Put(bb)
n := bh.columnsHeaderIndexSize
if n > maxColumnsHeaderIndexSize {
logger.Panicf("BUG: %s: too big columnsHeaderIndexSize: %d bytes; mustn't exceed %d bytes", sr.columnsHeaderIndexReader.Path(), n, maxColumnsHeaderIndexSize)
}
bb.B = bytesutil.ResizeNoCopyMayOverallocate(bb.B, int(n))
sr.columnsHeaderIndexReader.MustReadFull(bb.B)
cshIndex := getColumnsHeaderIndex()
if err := cshIndex.unmarshalNoArena(bb.B); err != nil {
logger.Panicf("FATAL: %s: cannot unmarshal columnsHeaderIndex: %s", sr.columnsHeaderIndexReader.Path(), err)
}
if err := csh.setColumnNames(cshIndex, columnNames); err != nil {
logger.Panicf("FATAL: %s: %s", sr.columnsHeaderIndexReader.Path(), err)
}
putColumnsHeaderIndex(cshIndex)
}
// timestampsData contains the encoded timestamps data.
type timestampsData struct {
// data contains packed timestamps data.
data []byte
// marshalType is the marshal type for timestamps
marshalType encoding.MarshalType
// minTimestamp is the minimum timestamp in the timestamps data
minTimestamp int64
// maxTimestamp is the maximum timestamp in the timestamps data
maxTimestamp int64
}
// reset resets td for subsequent re-use
func (td *timestampsData) reset() {
td.data = nil
td.marshalType = 0
td.minTimestamp = 0
td.maxTimestamp = 0
}
// copyFrom copies src to td.
//
// td is valid until a.reset() is called.
func (td *timestampsData) copyFrom(a *arena, src *timestampsData) {
td.reset()
td.data = a.copyBytes(src.data)
td.marshalType = src.marshalType
td.minTimestamp = src.minTimestamp
td.maxTimestamp = src.maxTimestamp
}
// mustWriteTo writes td to sw and updates th accordingly
func (td *timestampsData) mustWriteTo(th *timestampsHeader, sw *streamWriters) {
th.reset()
th.marshalType = td.marshalType
th.minTimestamp = td.minTimestamp
th.maxTimestamp = td.maxTimestamp
th.blockOffset = sw.timestampsWriter.bytesWritten
th.blockSize = uint64(len(td.data))
if th.blockSize > maxTimestampsBlockSize {
logger.Panicf("BUG: too big timestampsHeader.blockSize: %d bytes; mustn't exceed %d bytes", th.blockSize, maxTimestampsBlockSize)
}
sw.timestampsWriter.MustWrite(td.data)
}
// mustReadFrom reads timestamps data associated with th from sr to td.
//
// td is valid until a.reset() is called.
func (td *timestampsData) mustReadFrom(a *arena, th *timestampsHeader, sr *streamReaders) {
td.reset()
td.marshalType = th.marshalType
td.minTimestamp = th.minTimestamp
td.maxTimestamp = th.maxTimestamp
timestampsReader := &sr.timestampsReader
if th.blockOffset != timestampsReader.bytesRead {
logger.Panicf("FATAL: %s: unexpected timestampsHeader.blockOffset=%d; must equal to the number of bytes read: %d",
timestampsReader.Path(), th.blockOffset, timestampsReader.bytesRead)
}
timestampsBlockSize := th.blockSize
if timestampsBlockSize > maxTimestampsBlockSize {
logger.Panicf("FATAL: %s: too big timestamps block with %d bytes; the maximum supported block size is %d bytes",
timestampsReader.Path(), timestampsBlockSize, maxTimestampsBlockSize)
}
td.data = a.newBytes(int(timestampsBlockSize))
timestampsReader.MustReadFull(td.data)
}
// columnData contains packed data for a single column.
type columnData struct {
// name is the column name
name string
// valueType is the type of values stored in valuesData
valueType valueType
// minValue is the minimum encoded uint* or 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 uint* or 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
// valuesData contains packed values data for the given column
valuesData []byte
// bloomFilterData contains packed bloomFilter data for the given column
bloomFilterData []byte
}
// reset rests cd for subsequent re-use
func (cd *columnData) reset() {
cd.name = ""
cd.valueType = 0
cd.minValue = 0
cd.maxValue = 0
cd.valuesDict.reset()
cd.valuesData = nil
cd.bloomFilterData = nil
}
// copyFrom copies src to cd.
//
// cd is valid until a.reset() is called.
func (cd *columnData) copyFrom(a *arena, src *columnData) {
cd.reset()
cd.name = a.copyString(src.name)
cd.valueType = src.valueType
cd.minValue = src.minValue
cd.maxValue = src.maxValue
cd.valuesDict.copyFrom(a, &src.valuesDict)
cd.valuesData = a.copyBytes(src.valuesData)
cd.bloomFilterData = a.copyBytes(src.bloomFilterData)
}
// mustWriteTo writes cd to sw and updates ch accordingly.
//
// ch is valid until cd 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 16:18:28 +02:00
func (cd *columnData) mustWriteTo(ch *columnHeader, sw *streamWriters) {
ch.reset()
ch.name = cd.name
ch.valueType = cd.valueType
ch.minValue = cd.minValue
ch.maxValue = cd.maxValue
ch.valuesDict.copyFromNoArena(&cd.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 16:18:28 +02:00
bloomValuesWriter := sw.getBloomValuesWriterForColumnName(ch.name)
// marshal values
ch.valuesSize = uint64(len(cd.valuesData))
if ch.valuesSize > maxValuesBlockSize {
logger.Panicf("BUG: too big valuesSize: %d bytes; mustn't exceed %d bytes", ch.valuesSize, maxValuesBlockSize)
}
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
ch.valuesOffset = bloomValuesWriter.values.bytesWritten
bloomValuesWriter.values.MustWrite(cd.valuesData)
// marshal bloom filter
ch.bloomFilterSize = uint64(len(cd.bloomFilterData))
if ch.bloomFilterSize > maxBloomFilterBlockSize {
logger.Panicf("BUG: too big bloomFilterSize: %d bytes; mustn't exceed %d bytes", ch.bloomFilterSize, maxBloomFilterBlockSize)
}
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
ch.bloomFilterOffset = bloomValuesWriter.bloom.bytesWritten
bloomValuesWriter.bloom.MustWrite(cd.bloomFilterData)
}
// mustReadFrom reads columns data associated with ch from sr to cd.
//
// cd is valid until a.reset() is called.
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
func (cd *columnData) mustReadFrom(a *arena, ch *columnHeader, sr *streamReaders, partFormatVersion uint) {
cd.reset()
cd.name = a.copyString(ch.name)
cd.valueType = ch.valueType
cd.minValue = ch.minValue
cd.maxValue = ch.maxValue
cd.valuesDict.copyFrom(a, &ch.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 16:18:28 +02:00
bloomValuesReader := sr.getBloomValuesReaderForColumnName(ch.name, partFormatVersion)
// read values
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
if ch.valuesOffset != bloomValuesReader.values.bytesRead {
logger.Panicf("FATAL: %s: unexpected columnHeader.valuesOffset=%d; must equal to the number of bytes read: %d",
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bloomValuesReader.values.Path(), ch.valuesOffset, bloomValuesReader.values.bytesRead)
}
valuesSize := ch.valuesSize
if valuesSize > maxValuesBlockSize {
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
logger.Panicf("FATAL: %s: values block size cannot exceed %d bytes; got %d bytes", bloomValuesReader.values.Path(), maxValuesBlockSize, valuesSize)
}
cd.valuesData = a.newBytes(int(valuesSize))
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bloomValuesReader.values.MustReadFull(cd.valuesData)
// read bloom filter
// bloom filter is missing in valueTypeDict.
if ch.valueType != valueTypeDict {
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
if ch.bloomFilterOffset != bloomValuesReader.bloom.bytesRead {
logger.Panicf("FATAL: %s: unexpected columnHeader.bloomFilterOffset=%d; must equal to the number of bytes read: %d",
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bloomValuesReader.bloom.Path(), ch.bloomFilterOffset, bloomValuesReader.bloom.bytesRead)
}
bloomFilterSize := ch.bloomFilterSize
if bloomFilterSize > maxBloomFilterBlockSize {
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
logger.Panicf("FATAL: %s: bloom filter block size cannot exceed %d bytes; got %d bytes", bloomValuesReader.bloom.Path(), maxBloomFilterBlockSize, bloomFilterSize)
}
cd.bloomFilterData = a.newBytes(int(bloomFilterSize))
lib/logstorage: refactor storage format to be more efficient for querying wide events It has been appeared that VictoriaLogs is frequently used for collecting logs with tens of fields. For example, standard Kuberntes setup on top of Filebeat generates more than 20 fields per each log. Such logs are also known as "wide events". The previous storage format was optimized for logs with a few fields. When at least a single field was referenced in the query, then the all the meta-information about all the log fields was unpacked and parsed per each scanned block during the query. This could require a lot of additional disk IO and CPU time when logs contain many fields. Resolve this issue by providing an (field -> metainfo_offset) index per each field in every data block. This index allows reading and extracting only the needed metainfo for fields used in the query. This index is stored in columnsHeaderIndexFilename ( columns_header_index.bin ). This allows increasing performance for queries over wide events by 10x and more. Another issue was that the data for bloom filters and field values across all the log fields except of _msg was intermixed in two files - fieldBloomFilename ( field_bloom.bin ) and fieldValuesFilename ( field_values.bin ). This could result in huge disk read IO overhead when some small field was referred in the query, since the Operating System usually reads more data than requested. It reads the data from disk in at least 4KiB blocks (usually the block size is much bigger in the range 64KiB - 512KiB). So, if 512-byte bloom filter or values' block is read from the file, then the Operating System reads up to 512KiB of data from disk, which results in 1000x disk read IO overhead. This overhead isn't visible for recently accessed data, since this data is usually stored in RAM (aka Operating System page cache), but this overhead may become very annoying when performing the query over large volumes of data which isn't present in OS page cache. The solution for this issue is to split bloom filters and field values across multiple shards. This reduces the worst-case disk read IO overhead by at least Nx where N is the number of shards, while the disk read IO overhead is completely removed in best case when the number of columns doesn't exceed N. Currently the number of shards is 8 - see bloomValuesShardsCount . This solution increases performance for queries over large volumes of newly ingested data by up to 1000x. The new storage format is versioned as v1, while the old storage format is version as v0. It is stored in the partHeader.FormatVersion. Parts with the old storage format are converted into parts with the new storage format during background merge. It is possible to force merge by querying /internal/force_merge HTTP endpoint - see https://docs.victoriametrics.com/victorialogs/#forced-merge .
2024-10-16 16:18:28 +02:00
bloomValuesReader.bloom.MustReadFull(cd.bloomFilterData)
}
}