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
|
|
|
package logstorage
|
|
|
|
|
|
|
|
import (
|
|
|
|
"fmt"
|
|
|
|
"io"
|
|
|
|
"strings"
|
|
|
|
|
|
|
|
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
|
|
|
|
"github.com/VictoriaMetrics/VictoriaMetrics/lib/filestream"
|
|
|
|
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
|
|
|
|
)
|
|
|
|
|
|
|
|
func mustWriteColumnNames(w *writerWithStats, columnNames []string) {
|
|
|
|
data := marshalColumnNames(nil, columnNames)
|
|
|
|
w.MustWrite(data)
|
|
|
|
}
|
|
|
|
|
|
|
|
func mustReadColumnNames(r filestream.ReadCloser) ([]string, map[string]uint64) {
|
|
|
|
src, err := io.ReadAll(r)
|
|
|
|
if err != nil {
|
|
|
|
logger.Panicf("FATAL: %s: cannot read colum names: %s", r.Path(), err)
|
|
|
|
}
|
|
|
|
|
|
|
|
columnNames, err := unmarshalColumnNames(src)
|
|
|
|
if err != nil {
|
|
|
|
logger.Panicf("FATAL: %s: %s", r.Path(), err)
|
|
|
|
}
|
|
|
|
|
|
|
|
columnNameIDs, err := getColumnNameIDs(columnNames)
|
|
|
|
if err != nil {
|
|
|
|
logger.Panicf("BUG: %s: %s; columnNames=%v", r.Path(), err, columnNameIDs)
|
|
|
|
}
|
|
|
|
|
|
|
|
return columnNames, columnNameIDs
|
|
|
|
}
|
|
|
|
|
|
|
|
func getColumnNameIDs(columnNames []string) (map[string]uint64, error) {
|
|
|
|
m := make(map[uint64]string, len(columnNames))
|
|
|
|
columnNameIDs := make(map[string]uint64, len(columnNames))
|
|
|
|
for i, name := range columnNames {
|
|
|
|
id := uint64(i)
|
|
|
|
if prevName, ok := m[id]; ok {
|
|
|
|
return nil, fmt.Errorf("duplicate column name id=%d for columns %q and %q", id, prevName, name)
|
|
|
|
}
|
|
|
|
m[id] = name
|
|
|
|
columnNameIDs[name] = id
|
|
|
|
}
|
|
|
|
|
|
|
|
return columnNameIDs, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func marshalColumnNames(dst []byte, columnNames []string) []byte {
|
|
|
|
data := encoding.MarshalVarUint64(nil, uint64(len(columnNames)))
|
2024-11-06 13:22:13 +00:00
|
|
|
data = marshalStrings(data, columnNames)
|
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.CompressZSTDLevel(dst, data, 1)
|
|
|
|
|
|
|
|
return dst
|
|
|
|
}
|
|
|
|
|
|
|
|
func unmarshalColumnNames(src []byte) ([]string, error) {
|
|
|
|
data, err := encoding.DecompressZSTD(nil, src)
|
|
|
|
if err != nil {
|
|
|
|
return nil, fmt.Errorf("cannot decompress column names from len(src)=%d: %w", len(src), err)
|
|
|
|
}
|
|
|
|
src = data
|
|
|
|
|
|
|
|
n, nBytes := encoding.UnmarshalVarUint64(src)
|
|
|
|
if nBytes <= 0 {
|
|
|
|
return nil, fmt.Errorf("cannot parse the number of column names for len(src)=%d", len(src))
|
|
|
|
}
|
|
|
|
src = src[nBytes:]
|
|
|
|
|
|
|
|
m := make(map[string]uint64, n)
|
|
|
|
columnNames := make([]string, n)
|
|
|
|
for id := uint64(0); id < n; id++ {
|
|
|
|
name, nBytes := encoding.UnmarshalBytes(src)
|
|
|
|
if nBytes <= 0 {
|
|
|
|
return nil, fmt.Errorf("cannot parse colum name number %d out of %d", id, n)
|
|
|
|
}
|
|
|
|
src = src[nBytes:]
|
|
|
|
|
|
|
|
if idPrev, ok := m[string(name)]; ok {
|
|
|
|
return nil, fmt.Errorf("duplicate ids for column name %q: %d and %d", name, idPrev, id)
|
|
|
|
}
|
|
|
|
|
|
|
|
m[string(name)] = id
|
|
|
|
columnNames[id] = string(name)
|
|
|
|
}
|
|
|
|
|
|
|
|
if len(src) > 0 {
|
|
|
|
return nil, fmt.Errorf("unexpected non-empty tail left after unmarshaling column name ids; len(tail)=%d", len(src))
|
|
|
|
}
|
|
|
|
|
|
|
|
return columnNames, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
type columnNameIDGenerator struct {
|
|
|
|
// columnNameIDs contains columnName->id mapping for already seen columns
|
|
|
|
columnNameIDs map[string]uint64
|
|
|
|
|
|
|
|
// columnNames contains id->columnName mapping for already seen columns
|
|
|
|
columnNames []string
|
|
|
|
}
|
|
|
|
|
|
|
|
func (g *columnNameIDGenerator) reset() {
|
|
|
|
g.columnNameIDs = nil
|
|
|
|
g.columnNames = nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (g *columnNameIDGenerator) getColumnNameID(name string) uint64 {
|
|
|
|
id, ok := g.columnNameIDs[name]
|
|
|
|
if !ok {
|
|
|
|
if g.columnNameIDs == nil {
|
|
|
|
g.columnNameIDs = make(map[string]uint64)
|
|
|
|
}
|
|
|
|
id = uint64(len(g.columnNames))
|
|
|
|
nameCopy := strings.Clone(name)
|
|
|
|
g.columnNameIDs[nameCopy] = id
|
|
|
|
g.columnNames = append(g.columnNames, nameCopy)
|
|
|
|
}
|
|
|
|
return id
|
|
|
|
}
|