A follow-up after 9bb5ba5d2f
that impacted compression ratio for data compressed with native GO zstd lib (`make test-pure`).
Signed-off-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit 38bded4e58)
It has been appeared these optimizatios do not give measurable performance improvements,
while they complicate the code too much and may result in slowdown when the ingested logs have
different sets of fields.
This is a follow-up for 630601488e
(cherry picked from commit dce5eb88d3)
It has been appeared that 256 files increase RAM usage too much comparing to 128 files
when ingesting logs with hundreds of fields (aka wide events). So let's return back 128 files
limit for now.
This is a follow-up for 9bb5ba5d2f
(cherry picked from commit a50ab10998)
This should improve query performance for logs with hundreds of fields (aka wide events).
Previously there was a high chance that the data for multiple log fields is stored in the same file.
This could result in query performance slowdown and/or increased disk read IO,
since the operating system could read unnecessary data for the fields, which aren't used in the query.
Now log fields are guaranteed to be stored in separate files until the number of fields exceeds 256.
After that multiple log fields start sharing files.
(cherry picked from commit 9bb5ba5d2f)
- Re-use column names and values from the previously added rows if possible.
This increases locality of reference for field names and values, while improving
access speed for the field names and values.
- Postpone sorting fields in the added rows until creating inmemory part from them.
This allows optimizing the sorting for log fields with the same set of fields.
This is usually the case for logs, which belong to the same logs stream.
(cherry picked from commit 630601488e)
This eliminates a class of potential bugs with incorrect stats calculations when an additional filter
is applied to the blockResult before passing it to the stats function, and this filter removes
all the rows from blockResult.
1. Use distinct code paths for blockResult.getValues() and blockResult.getValuesBucketed().
This should simplify debugging and maintenance of the resulting code.
2. Do not load column values if all the values in the block fit the same bucket.
Use blockResultColumn.minValue and blockResultColumn.maxValue for determining whether
column values must be loaded via blockResultColumn.getValuesEncoded().
This signiciantly improves performance for big buckets, which cover all the column
values in a block.
3. Properly calculate buckets for negative values.
4. Properly adjust weekly buckets by Monday.
This improves performance for queries, which use `sort by (...) limit N` without mentioning _time field.
For example, the following query must work faster now
_time:1d | rm _time | sort by (request_duration desc) limit 10
(cherry picked from commit 422caf6bd7)
- Add a fast path for timestamps ending with 'Z'
- Use strings.LastIndexAny instead of strings.IndexAny for searching
for timezone offset at the end of the string. This works faster
for timestamps with sub-second precision.
(cherry picked from commit 335071cf3d)
Update the pipe state only once per each series of repeated strings, uint8 values and tuples.
This improves performance a bit for the following `top` pipes:
- top (string_field)
- top (uint8_field)
- top (field1, ..., fieldN)
Do not apply the optimization for uint16, uint32, uint64 and int64 fields, since they
usually contain big number of unique values, which do not repeat most of the time.
Previously RFC3339 timestamps with sub-second precision could be incorrectly compared by lessString().
For example, 2025-01-20T10:20:30.1Z was incorrectly treated as smaller than 2025-01-20T10:20:30.09Z,
because the first timestamp has smaller decimal number after the last dot than the second timestamp.
(cherry picked from commit 81d359507d)
Split unique values (groups) into shards according to the configured concurrency
during processing of the matching rows if the number of unique values exceeds the hardcoded threshold.
Previously this splitting was performed unconditionally at the merge stage when merging independently
calculated per-CPU states into a single state. It is faster to perform the split during rows processing
if the number of unique values is big.
This gives up to 30% perfromance improvements when these pipes are applied to big number of unique values (groups).
(cherry picked from commit 48602a1ae8)
The number of worker shards per each pipe processor is created during query initialization.
This number equals to the `options(concurrency=N)` if this option is set or to the number of available CPU cores.
This means that all the pipes must adhere the given concurrency when passing data blocks
to the next pipe.
See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8201
The bug has been introduced in 0214aa328e
It is better to use the AppendFieldsToJSON function directly
instead of hiding it under RowsFormatter abstraction.
(cherry picked from commit 95f182053b)
This should reduce the time needed for opening the storage with retentions exceeding a few months.
While at at, limit the concurrency of opening partitions in parallel to the number of available CPU cores,
since higher concurrency may increase RAM usage and CPU usage without performance improvements
if opening a single partition is CPU-bound task.
This is a follow-up for 17988942ab
The purpose of extra filters ( https://docs.victoriametrics.com/victorialogs/querying/#extra-filters )
is to limit the subset of logs, which can be queried. For example, it is expected that all the queries
with `extra_filters={tenant=123}` can access only logs, which contain `123` value for the `tenant` field.
Previously this wasn't the case, since the provided extra filters weren't applied to subqueries.
For example, the following query could be used to select all the logs outside `tenant=123`, for any `extra_filters` arg:
* | union({tenant!=123})
This commit fixes this by propagating extra filters to all the subqueries.
While at it, this commit also properly propagates [start, end] time range filter from HTTP querying APIs
into all the subqueries, since this is what most users expect. This behaviour can be overriden on per-subquery
basis with the `options(ignore_global_time_filter=true)` option - see https://docs.victoriametrics.com/victorialogs/logsql/#query-options
Also properly apply apply optimizations across all the subqueries. Previously the optimizations at Query.optimize()
function were applied only to the top-level query.
logger.Fatalf("BUG: ...") complicates investigating the bug, since it doesn't show the call stack,
which led to the bug. So it is better to consistently use logger.Panicf("BUG: ...") for logging programming bugs.
This is done via 'options(concurrency=N)' prefix for the query.
For example, the following query is executed on at most 4 CPU cores:
options(concurrency=4) _time:1d | count_uniq(user_id)
This allows reducing RAM and CPU usage at the cost of longer query execution times,
since by default every query is executed in parallel on all the available CPU cores.
See https://docs.victoriametrics.com/victorialogs/logsql/#query-options
Also always initialize Query.timestamp with the timestamp from the lexer.
This should avoid potential problems with relative timestamps inside inner queries.
For example, the `_time:1h` filter in the following query is correctly executed
relative to the current timestamp:
foo:in(_time:1h | keep foo)
