This allows reducing the amounts of data, which must be read during queries over logs with big number of fields (aka "wide events").
This, in turn, improves query performance when the data, which needs to be scanned during the query, doesn't fit OS page cache.
(cherry picked from commit 7a62eefa34)
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 .
This localizes blockSearch.getColumnsHeader() call at block_search.go .
This call is going to be optimized in the next commits in order to avoid
unmarshaling of header data for unneeded columns, which weren't requested
by getConstColumnValue() / getColumnHeader().
(cherry picked from commit 507b206a7d)
Refer the original byte slice with the marshaled columnsHeader for columns names and dictionary-encoded column values.
This improves query performance a bit when big number of blocks with big number of columns are scanned during the query.
(cherry picked from commit 279e25e7c8)
Change the return values for these functions - now they return the unmarshaled result plus
the size of the unmarshaled result in bytes, so the caller could re-slice the src for further unmarshaling.
This improves performance of these functions in hot loops of VictoriaLogs a bit.