The panic has been introduced in 68f3a02589
While at it, add padding to shard structs in order to avoid false sharing on mordern CPUs
This should improve scalability on systems with many CPU cores
This is OK, since the anchors are implicitly applied to the whole regexp.
This optimization should improve the speed for regexp series filters with explicit $ and ^ anchors.
For example, `{label="^(foo|bar)$"}`
These metrics allow alerting when the number of unique series approach the limit.
For example, the following query alerts when the number of series reaches 90% of the configured limit:
vm_hourly_series_limit_current_series / vm_hourly_series_limit_max_series > 0.9
ioutil.ReadAll is deprecated since Go1.16 - see https://tip.golang.org/doc/go1.16#ioutil
VictoriaMetrics requires at least Go1.18, so it is OK to switch from ioutil.ReadAll to io.ReadAll.
This is a follow-up for 02ca2342ab
The ioutil.{Read|Write}File is deprecated since Go1.16 -
see https://tip.golang.org/doc/go1.16#ioutil
VictoriaMetrics needs at least Go1.18, so it is safe to remove ioutil usage
from source code.
This is a follow-up for 02ca2342ab
* lib/storage: bump max merge concurrency for small parts to 15
The change is based on the feedback from users on github.
Thier examples show, that limit of 8 sometimes become a
bottleneck. Users report that without limit concurrency
can climb up to 15-20 merges at once.
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* Update lib/storage/partition.go
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
* lib/storage: prevent excessive loops when storage is in RO
Returning nil error when storage is in RO mode results
into excessive loops and function calls which could
result into CPU exhaustion. Returning an err instead
will trigger delays in the for loop and save some resources.
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* document the change
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
Previously the time series could be put into dateMetricIDCache without
registering in the per-day inverted index if GetOrCreateTSIDByName
finds TSID entry in the global index. This could lead to missing
series in query results.
The issue has been introduced in the commit 55e7afae3a,
which has been included in VictoriaMetrics v1.78.0
- show dates in human-readable format, e.g. 2022-05-07, instead of a numeric value
- limit the maximum length of queries and filters shown in trace messages
Previously SearchMetricNames was returning unmarshaled metric names.
This wasn't great for vmstorage, which should spend additional CPU time
for marshaling the metric names before sending them to vmselect.
While at it, remove possible duplicate metric names, which could occur when
multiple samples for new time series are ingested via concurrent requests.
Also sort the metric names before returning them to the client.
This simplifies debugging of the returned metric names across repeated requests to /api/v1/series
querytracer has been added to the following storage.Storage methods:
- RegisterMetricNames
- DeleteMetrics
- SearchTagValueSuffixes
- SearchGraphitePaths
The commit 5fb45173ae takes into account only newly registered series
when applying cardinality limits. This means that the cardinality limit could be exceeded with already registered series.
This commit returns back accounting for already registered series when applying cardinality limits.
Previously the creation of per-day indexes and global indexes
for the newly registered time series was decoupled.
Now global indexes and per-day indexes for the current day are created toghether for new time series.
This should speed up registering new time series a bit.
This allows filling the seriesCountByFocusLabelValue list in the /api/v1/status/tsdb response
with label values for the specified focusLabel, which contain the highest number of time series.
TODO: add this to Cardinality explorer at VMUI - https://docs.victoriametrics.com/#cardinality-explorer
Workers count for merges affects the max part size during merges. Such behaviour
protects storage from running out of disk space for scenario when all workers
are merging parts with the max size.
This works very well for most cases. But for systems where high number of CPUs
is allocated for vmstorage components this could significantly impact the max
part size and result in more unmerged parts than expected.
While checking multiple production highly loaded setups it was discovered that
`max_over_time(vm_active_merges{type="storage/big}[1h]}"` rarely exceeds 2,
and `max_over_time(vm_active_merges{type="storage/small}[1h]}"` rarely exceeds 4.
The change in this commit limits the max value for concurrency accordingly.
Signed-off-by: hagen1778 <roman@victoriametrics.com>
The default size of `indexdb/tagFilters` now can be overridden via
`storage.cacheSizeIndexDBTagFilters` flag.
Please, be careful with changing default size since it may
lead to inefficient work of the vmstorage or OOM exceptions.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2663
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Nikolay <nik@victoriametrics.com>
* lib/{storage,flagutil} - Add option for snapshot autoremoval
- add prometheus-like duration as command flag
- add option to delete stale snapshots
- update duration.go flag to re-use own code
* wip
* lib/flagutil: re-use Duration.Set() call in NewDuration
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
* lib/{storage,regexpcache}: replaces regexpCacheMap with LRU cache
It should decrease memory usage for regexp caching
with storing cacheEntry by pointer - golang map should be able to effectivly shrink it's size
original issue with this case - unexpected map grows and storage OOM
Apply suggestions from code review
Co-authored-by: Roman Khavronenko <roman@victoriametrics.com>
Adds missing metrics for regexp cache and regexpPrefixes cache
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
* fix for issue 2255 - matchTagFilters for positive empty-match filters
* add example to comments
* formatting
* add test for positive empty match
* formatting
- Postpone the pre-poulation to the last hour of the current day. This should reduce the number
of useless entries in the next per-day index, which shouldn't be created there,
when the corresponding time series are stopped to be pushed during the current day.
- Make the pre-population more smooth in time by using the hash of MetricID instead of MetricID itself
when calculating the need for for the given MetricID pre-population.
- Sync the logic for pre-population of the next day inverted index with the logic of pre-populating tsid cache
after indexdb rotation. This should improve code maintainability.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/430
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
* lib/index: reduce read/write load after indexDB rotation
IndexDB in VM is responsible for storing TSID - ID's used for identifying
time series. The index is stored on disk and used by both ingestion and read path.
IndexDB is stored separately to data parts and is global for all stored data.
It can't be deleted partially as VM deletes data parts. Instead, indexDB is
rotated once in `retention` interval.
The rotation procedure means that `current` indexDB becomes `previous`,
and new freshly created indexDB struct becomes `current`. So in any time,
VM holds indexDB for current and previous retention periods.
When time series is ingested or queried, VM checks if its TSID is present
in `current` indexDB. If it is missing, it checks the `previous` indexDB.
If TSID was found, it gets copied to the `current` indexDB. In this way
`current` indexDB stores only series which were active during the retention
period.
To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both
write and read path consult `tsidCache` and on miss the relad lookup happens.
When rotation happens, VM resets the `tsidCache`. This is needed for ingestion
path to trigger `current` indexDB re-population. Since index re-population
requires additional resources, every index rotation event may cause some extra
load on CPU and disk. While it may be unnoticeable for most of the cases,
for systems with very high number of unique series each rotation may lead
to performance degradation for some period of time.
This PR makes an attempt to smooth out resource usage after the rotation.
The changes are following:
1. `tsidCache` is no longer reset after the rotation;
2. Instead, each entry in `tsidCache` gains a notion of indexDB to which
they belong;
3. On ingestion path after the rotation we check if requested TSID was
found in `tsidCache`. Then we have 3 branches:
3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID.
3.2 Slow path. It wasn't found, so we generate it from scratch,
add to `current` indexDB, add it to `tsidCache`.
3.3 Smooth path. It was found but does not belong to the `current` indexDB.
In this case, we add it to the `current` indexDB with some probability.
The probability is based on time passed since the last rotation with some threshold.
The more time has passed since rotation the higher is chance to re-populate `current` indexDB.
The default re-population interval in this PR is set to `1h`, during which entries from
`previous` index supposed to slowly re-populate `current` index.
The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs
were moved from `previous` indexDB to the `current` indexDB. This metric supposed to
grow only during the first `1h` after the last rotation.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
* lib/promscrape: support prometheus-like duration in scrape configs
The change allows to specify duration values like `1d`, `1w`
for fields `scrape_interval`, `scrape_timeout`, etc.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/817#issuecomment-1033384766
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* lib/blockcache: make linter happy
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* lib/promscrape: support prometheus-like duration in scrape configs
* add support for extra fields `scrape_align_interval` and `scrape_offset`;
* support Prometheus duration parsing for `__scrape_interval__`
and `__scrape_duration__` labels;
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
* wip
* docs/CHANGELOG.md: document the feature
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
* optimized code ,because only the first error,so no need var errors []error
* optimized code ,because only the first error,so no need var errors []error
Co-authored-by: lirenzuo <lirenzuo@shein.com>
Previously bytesutil.Resize() was copying the original byte slice contents to a newly allocated slice.
This wasted CPU cycles and memory bandwidth in some places, where the original slice contents wasn't needed
after slize resizing. Switch such places to bytesutil.ResizeNoCopy().
Rename the original bytesutil.Resize() function to bytesutil.ResizeWithCopy() for the sake of improved readability.
Additionally, allocate new slice with `make()` instead of `append()`. This guarantees that the capacity of the allocated slice
exactly matches the requested size. The `append()` could return a slice with bigger capacity as an optimization for further `append()` calls.
This could result in excess memory usage when the returned byte slice was cached (for instance, in lib/blockcache).
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
Previously these caches could exceed limits set via `-memory.allowedPercent` and/or `-memory.allowedBytes`,
since limits were set independently per each data part. If the number of data parts was big, then limits could be exceeded,
which could result to out of memory errors.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2007
For example, `{__graphite__=~"foo.(bar|baz)"}` is automatically converted to `{__graphite__=~"foo.{bar,baz}"}` before execution.
This allows using multi-value Grafana template variables such as `{__graphite__=~"foo.($app)"}`.
The vm_cache_size_max_bytes metric can be used for determining caches which reach their capacity via the following query:
vm_cache_size_bytes / vm_cache_size_max_bytes > 0.9
* adds read-only mode for vmstorage
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/269
* changes order a bit
* moves isFreeDiskLimitReached var to storage struct
renames functions to be consistent
change protoparser api - with optional storage limit check for given openned storage
* renames freeSpaceLimit to ReadOnly
The number of series per target can be limited with the following options:
* Global limit with `-promscrape.maxSeriesPerTarget` command-line option.
* Per-target limit with `max_series: N` option in `scrape_config` section.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1561
* Rename -search.maxMetricsPointSearch to -search.maxSamplesPerQuery, so it is more consistent with the existing -search.maxSamplesPerSeries
* Move the -search.maxSamplesPerQuery from vmstorage to vmselect, so it could effectively limit the number of raw samples obtained from all the vmstorage nodes
* Document the -search.maxSamplesPerQuery in docs/CHANGELOG.md
Previously needsDedup() could return true if the de-duplication wasn't needed for the following case:
d < interval
/ \
| v | v |
interval interval
Now it properly returns false for this case
This reverts commit 7c6d3981bf.
Reason for revert: high contention at bucket16Pool on systems with big number of CPU cores.
This slows down query processing significantly.
This should reduce memory usage on systems with big number of CPU cores,
since every inmemoryPart object occupies at least 64KB of memory and sync.Pool maintains
a separate pool inmemoryPart objects per each CPU core.
Though the new scheme for the pool worsens per-cpu cache locality, this should be amortized
by big sizes of inmemoryPart objects.
Previously the stats for cache misses could be improperly counted, because it had inflated cache misses
if the entry was missing in the curr cache, but was existing in the prev cache.
The same applies to cache requests - they were inflated if the entry was missing in the curr cache.
One minute cache timeout result in slower queries in some production workloads where the interval
between query execution is in the range 1 minute - 2 minutes.
This should reduce memory usage on a system with high number of active time series and a high churn rate.
One minute is enough for caching the blocks needed for repeated queries (e.g. alerting rules, recording rules and dashboard refreshes).
These functions are non-trivial, while their code has minimal differences.
It is better from maintainability PoV to merge these functions into a single function.
It must match all the time series on the given time range.
Previously it was matched to all the time series without the restriction on the given time range.
It should be faster querying all the labels and/or all the values instead of querying per-day labels/values on time ranges exceeding maxDaysForPerDaySearch
This option can be useful when samples for the same time series are ingested with distinct order of labels.
For example, metric{k1="v1",k2="v2"} and metric{k2="v2",k1="v1"}.
Do not pass filter metric ids to getMetricIDsForTagFilter, since it has been appeared that this slows down
the function by multiple times when it finds big number of metricIDs (tens of millions).
While this may increase CPU and disk IO usage needed for background merge,
this also recudes CPU usage during queries in production. This is because
such queries tend to read recently added data and it is better to have lower number
of parts for such data in order to reduce CPU usage.
This partially reverts ebf8da3730
The filter arg has been removed in the commit c7ee2fabb8
because it was preventing from caching the number of matching time series per each tf.
Now the cache contains duration for tf execution, so the filter shouldn't break such caching.
For example `{label=~"foo\.bar"}` should be converted to `{label="foo.bar"}`. Previously it has was mistakenly conveted to `{label="foo\.bar"}` .
This could result in missing time series for such tag filters.
The `filter` arg breaks the logic for sorting tag filters by the matching metrics,
which may result in non-optimal performance during time series search.
Production workloads show that indexdb blocks must be cached unconditionally for reducing CPU usage.
This shouldn't increase memory usage too much, since unused blocks are removed from the cache every two minutes.
