Previously Block columns wasn't properly limited by maxColumnsPerBlock.
And it was possible a case, when more columns per block added than
expected.
For example, if ingested log stream has many unuqie fields
and it's sum exceed maxColumnsPerBlock.
We only enforce fieldsPerBlock limit during row parsing, which limits
isn't enough to mitigate this issue. Also it
would be very expensive to apply maxColumnsPerBlock limit during
ingestion, since it requires to track all possible field tags
combinations.
This commit adds check for maxColumnsPerBlock limit during
MustInitFromRows function call. And it returns offset of the rows and
timestamps added to the block.
Function caller must create another block and ingest remaining rows
into it.
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7568
### Describe Your Changes
Please provide a brief description of the changes you made. Be as
specific as possible to help others understand the purpose and impact of
your modifications.
### Checklist
The following checks are **mandatory**:
- [ ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: f41gh7 <nik@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
Move the repeated check for an empty value into statsUniqValuesProcessor.updateState() function.
This allow removing duplicate code for this check from statsUniqValuesProcessor.updateState() call sites.
This eliminates possible bugs related to forgotten Query.Optimize() calls.
This also allows removing optimize() function from pipe interface.
While at it, drop filterNoop inside filterAnd.
(cherry picked from commit 66b2987f49)
This msy be useful when ingesting logs from different sources, which store the log message in different fields.
For example, `_msg_field=message,event.data,some_field` will get log message from the first non-empty field:
`message`, `event.data` and `some_field`.
(cherry picked from commit ed73f8350b)
If the number of output (bloom, values) shards is zero, then this may lead to panic
as shown at https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7391 .
This panic may happen when parts with only constant fields with distinct values are merged into
output part with non-constant fields, which should be written to (bloom, values) shards.
(cherry picked from commit 102e9d4f4e)
### Describe Your Changes
fix function name
### Checklist
The following checks are **mandatory**:
- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
(cherry picked from commit 45896fb477)
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)
This improves performance of `field_values` pipe when it is applied to large number of data blocks.
This also improves performance of /select/logsql/field_values HTTP API.
(cherry picked from commit 8d968acd0a)
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7182
- add a separate index cache for searches which might read through large
amounts of random entries. Primary use-case for this is retention and
downsampling filters, when applying filters background merge needs to
fetch large amount of random entries which pollutes an index cache.
Using different caches allows to reduce effect on memory usage and cache
efficiency of the main cache while still having high cache hit rate. A
separate cache size is 5% of allowed memory.
- reduce size of indexdb/dataBlocks cache in order to free memory for
new sparse cache. Reduced size by 5% and moved this to a separate cache.
- add a separate metricName search which does not cache metric names -
this is needed in order to allow disabling metric name caching when
applying downsampling/retention filters. Applying filters during
background merge accesses random entries, this fills up cache and does
not provide an actual improvement due to random access nature.
Merge performance and memory usage stats before and after the change:
- before
- after
---------
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
(cherry picked from commit 837d0d136d)
These caches aren't expected to grow big, so it is OK to use the most simplest cache based on sync.Map.
The benefit of this cache compared to workingsetcache is better scalability on systems with many CPU cores,
since it doesn't use mutexes at fast path.
An additional benefit is lower memory usage on average, since the size of in-memory cache equals
working set for the last 3 minutes.
The downside is that there is no upper bound for the cache size, so it may grow big during workload spikes.
But this is very unlikely for typical workloads.
(cherry picked from commit 0f24078146)
Partition directories can be manually deleted and copied from another sources such as backups or other VitoriaLogs instances.
In this case the persisted cache becomes out of sync with partitions. This can result in missing index entries
during data ingestion or in incorrect results during querying. So it is better to do not persist caches.
This shouldn't hurt VictoriaLogs performance just after the restart too much, since its caches usually contain
small amounts of data, which can be quickly re-populated from the persisted data.
(cherry picked from commit 8aa144fa74)
Unpack the full columnsHeader block instead of unpacking meta-information per each individual column
when the query, which selects all the columns, is executed. This improves performance when scanning
logs with big number of fields.
(cherry picked from commit 2023f017b1)
- Parallelize merging of per-CPU `stats by(...)` result shards.
- Parallelize writing `stats by(...)` results to the next pipe.
(cherry picked from commit c4b2fdff70)
- Use parallel merge of per-CPU shard results. This improves merge performance on multi-CPU systems.
- Use topN heap sort of per-shard results. This improves performance when results contain millions of entries.
(cherry picked from commit 192c07f76a)
1. Verify if field in [fields
pipe](https://docs.victoriametrics.com/victorialogs/logsql/#fields-pipe)
exists. If not, it generates a metric with illegal float value "" for
prometheus metrics protocol.
2. check if multiple time range filters produce conflicted query time
range, for instance:
```
query: _time: 5m | stats count(),
start:2024-10-08T10:00:00.806Z,
end: 2024-10-08T12:00:00.806Z,
time: 2024-10-10T10:02:59.806Z
```
must give no result due to invalid final time range.
---------
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
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 .
Empty fields are treated as non-existing fields by VictoriaLogs data model.
So there is no sense in returning empty fields in query results, since they may mislead and confuse users.
(cherry picked from commit bac193e50b)
s.partitions can be changed when new partition is registered or when old partition is dropped.
This could lead to data races and panics when s.partitions slice is accessed by concurrently executed queries.
The fix is to make a copy of the selected partitions under s.partitionsLock before performing the query.
(cherry picked from commit b4b79a4961)
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)
This should improve performance when blockSearch.getColumnsHeader() is called multiple times
from different places of the code.
(cherry picked from commit 9e48074b59)
br.bs may be nil when br contains the block with additional filters applied during pipe calculations.
For example, `* | count() if (error) errors`.
(cherry picked from commit 867f671cc4)
Improperly written pipes could be silently parsed as filter pipe.
For example, the following query:
* | by (x)
was silently parsed to:
* | filter "by" x
It is better to return error, so the user could identify and fix invalid pipe
instead of silently executing invalid query with `filter` pipe.
(cherry picked from commit 7b475ed95d)
