mirror of https://github.com/VictoriaMetrics/VictoriaMetrics.git synced 2024-11-21 14:44:00 +00:00

Aliaksandr Valialkin fa566c68a6

lib/mergeset: really limit the number of in-memory parts to 15

It has been appeared that the registration of new time series slows down linearly
with the number of indexdb parts, since VictoriaMetrics needs to check every indexdb part
when it searches for TSID by newly ingested metric name.

The number of in-memory parts grows when new time series are registered
at high rate. The number of in-memory parts grows faster on systems with big number
of CPU cores, because the mergeset maintains per-CPU buffers with newly added entries
for the indexdb, and every such entry is transformed eventually into a separate in-memory part.

The solution has been suggested in https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5212
by @misutoth - to limit the number of in-memory parts with buffered channel.
This solution is implemented in this commit. Additionally, this commit merges per-CPU parts
into a single part before adding it to the list of in-memory parts. This reduces CPU load
when searching for TSID by newly ingested metric name.

The https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5212 recommends setting the limit on the number
of in-memory parts to 100, but my internal testing shows that much lower limit 15 works with the same efficiency
on a system with 16 CPU cores while reducing memory usage for `indexdb/dataBlocks` cache by up to 50%.

Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5190

2024-01-24 03:38:12 +02:00

44 KiB

Raw Blame History

sort

weight

title

aliases

100

CHANGELOG

docs

parent	weight
victoriametrics	100

/CHANGELOG.html

CHANGELOG

The following tip changes can be tested by building VictoriaMetrics components from the latest commits according to the following docs:

Metrics of the latest version of VictoriaMetrics cluster are available for viewing at our sandbox. The sandbox cluster installation is running under the constant load generated by prometheus-benchmark and used for testing latest releases.

tip

SECURITY: upgrade Go builder from Go1.21.5 to Go1.21.6. See the list of issues addressed in Go1.21.6.
FEATURE: improve new time series registration speed on systems with high number of CPU cores. Thanks to @misutoth for the initial idea and implementation.
FEATURE: vmagent: add support for discovering Hetzner Cloud and Hetzner Robot scrape targets. See this feature request and these docs.
FEATURE: graphite: add support for negative index in groupByNode and aliasByNode functions. Thanks to @rbizos for the pull request.
FEATURE: vmagent: add support for DataDog v2 data ingestion protocol. See these docs and this feature request.
FEATURE: vmagent: expose ability to set OAuth2 endpoint parameters per each -remoteWrite.url via the command-line flag -remoteWrite.oauth2.endpointParams. See these docs. Thanks to @mhill-holoplot for the pull request.
FEATURE: vmagent: add ability to set attach_metadata.node=true option for all the kubernetes_sd_configs defined at -promscrape.config via -promscrape.kubernetes.attachNodeMetadataAll command-line flag. See this feature request. Thanks to @wasim-nihal for the initial implementation.
FEATURE: vmalert: expose ability to set OAuth2 endpoint parameters via the following command-line flags:
- -datasource.oauth2.endpointParams for -datasource.url
- -notifier.oauth2.endpointParams for -notifier.url
- -remoteRead.oauth2.endpointParams for -remoteRead.url
- -remoteWrite.oauth2.endpointParams for -remoteWrite.url
FEATURE: vmauth: add ability to proxy incoming requests to different backends based on the requested host via src_hosts option at url_map. See these docs.
FEATURE: vmauth: expose vmauth_user_request_backend_errors_total and vmauth_unauthorized_user_request_backend_errors_total metrics, which track the number of failed requests because of backend errors. See this feature request.
FEATURE: vmauth: add an ability to specify additional labels for per-user metrics via metric_labels section. See this feature request.
FEATURE: all VictoriaMetrics components: break HTTP client connection if an error occurs after the server at -httpListenAddr already sent response status code. Previously such an error couldn't be detected at client side. Now the client will get an error about invalid chunked response. The error message is simultaneously written to the server log and in the last line of the response. This should help detecting errors when migrating data between VictoriaMetrics instances by vmctl. See this issue.
FEATURE: all VictoriaMetrics components: add ability to specify arbitrary HTTP headers to send with every request to -pushmetrics.url. See push metrics docs.
FEATURE: all VictoriaMetrics components: add -metrics.exposeMetadata command-line flag, which allows displaying TYPE and HELP metadata at /metrics page exposed at -httpListenAddr. This may be needed when the /metrics page is scraped by collector, which requires the TYPE and HELP metadata such as Google Cloud Managed Prometheus.
FEATURE: all VictoriaMetrics components: add ability to dynamically re-read auth keys and passwords from files and urls when using file:///path/to/file or http://host/path syntax for the following command-line flags: -configAuthKey, -deleteAuthKey, -flagsAuthKey, -forceMergeAuthKey, -forceFlushAuthKey, -httpAuth.password, -metricsAuthKey, -pprofAuthKey, -reloadAuthKey, -search.resetCacheAuthKey, -snapshotAuthKey. For example, -httpAuth.password=file:///path/to/password. See these docs for details.
FEATURE: dashboards/cluster: add panels for detailed visualization of traffic usage between vmstorage, vminsert, vmselect components and their clients. New panels are available in the rows dedicated to specific components.
FEATURE: dashboards/cluster: update "Slow Queries" panel to show percentage of the slow queries to the total number of read queries served by vmselect. The percentage value should make it more clear for users whether there is a service degradation.
FEATURE: dashboards/single: change dashboard title from VictoriaMetrics to VictoriaMetrics - single-node. The new title should provide better understanding of this dashboard purpose.
FEATURE: vmctl: rename cmd-line flag vm-native-disable-retries to vm-native-disable-per-metric-migration to better reflect its meaning.
FEATURE: vmctl: add -vm-native-src-insecure-skip-verify and -vm-native-dst-insecure-skip-verify command-line flags for native protocol. It can be used for skipping TLS certificate verification when connecting to the source or destination addresses.
FEATURE: Alerting rules for VictoriaMetrics: add job label to DiskRunsOutOfSpace alerting rule, so it is easier to understand to which installation the triggered instance belongs.
FEATURE: vmstorage: add tenant identifier for log messages regarding dropping excessive labels due to limits defined by -maxLabelsPerTimeseries or -maxLabelValueLen command-line flags. Previously, it was hard to understand to which tenant the dropped labels belong.
FEATURE: vmui: add the ability to export and import query reports:
- add a Query Analyzer page that allows you to build graphs from JSON data containing the results of executing a query request.
- add an Export query button to the graph that saves the result of executing the query in JSON. See this pull request.
FEATURE: vmui: add -vmui.defaultTimezone flag to set a default timezone. See this issue and these docs.
FEATURE: vmui: include UTC in the timezone selection dropdown for standardized time referencing. See this issue.
FEATURE: add VictoriaMetrics datasource to docker compose environment. See this pull request.
FEATURE: VictoriaMetrics cluster: disable compression for RPC responses sent from vmstorage to vmselect. This improves query performance at the cost of increased network bandwidth usage. The compression can be enabled back by passing -rpc.disableCompression=false command-line flag to vmstorage. The compression can be enabled back in multi-level cluster setup by passing -clusternative.disableCompression=false command-line flag to lower-level vmselect nodes.
BUGFIX: properly return errors from export APIs. Previously these errors were silently suppressed. See this pull request.
BUGFIX: VictoriaMetrics cluster: properly return full results when -search.skipSlowReplicas command-line flag is passed to vmselect and when vmstorage groups are in use. Previously partial results could be returned in this case.
BUGFIX: vminsert: properly accept samples via OpenTelemetry data ingestion protocol when these samples have no resource attributes. Previously such samples were silently skipped.
BUGFIX: vmstorage: added missing -inmemoryDataFlushInterval command-line flag, which was missing in VictoriaMetrics cluster after implementing this feature in v1.85.0.
BUGFIX: vmstorage: properly expire storage/prefetchedMetricIDs cache. Previously this cache was never expired, so it could grow big under high churn rate. This could result in increasing CPU load over time.
BUGFIX: vmalert: check -external.url schema when starting vmalert, must be http or https. Before, alertmanager could reject alert notifications if -external.url contained no or wrong schema.
BUGFIX: vmalert: automatically add exported_ prefix for original evaluation result label if it's conflicted with external or reserved one, previously it was overridden. See this issue.
BUGFIX: MetricsQL: consistently sort results for q1 or q2 query, so they do not change colors with each refresh in Grafana. See this issue.
BUGFIX: MetricsQL: properly return results from bottomk and bottomk_*() functions when some of these results contain NaN values. See this issue. Thanks to @xiaozongyang for the fix.
BUGFIX: MetricsQL: properly handle queries, which wrap rollup functions with multiple arguments without explicitly specified lookbehind window in square brackets into aggregate functions. For example, sum(quantile_over_time(0.5, process_resident_memory_bytes)) was resulting to expecting at least 2 args to ...; got 1 args error. Thanks to @atykhyy for the pull request.
BUGFIX: vmctl: retry on import errors in vm-native mode. Before, retries happened only on writes into a network connection between source and destination. But errors returned by server after all the data was transmitted were logged, but not retried.
BUGFIX: vmagent: properly assume role with AWS IRSA authorization. Previously role chaining was not supported. See this issue for details.
BUGFIX: vmagent: exit if there is config syntax error in scrape_config_files when -promscrape.config.strictParse=true. See this issue.
BUGFIX: vmagent: properly discover targets for role: endpoints and role: endpointslice in kubernetes_sd_configs. Previously some endpoints and endpointslice targets could be left undiscovered or some targets could have missing __meta_* labels when performing service discovery in busy Kubernetes clusters with large number of pods. See this pull request.
BUGFIX: vmagent: respect explicitly set series_limit: 0 in scrape_config. This allows removing series_limit restriction on a per-scrape_config basis when global limit is set via -promscrape.seriesLimitPerTarget. Previously, 0 value was ignored in favor of -promscrape.seriesLimitPerTarget.
BUGFIX: vmui: fix a link for the statistic inaccuracy explanation in the cardinality explorer tool. See this issue.
BUGFIX: vmui: fix the display of autocomplete results and cache the results. See this issue and this issue.
BUGFIX: vmui: send step param for instant queries. The change reverts this issue due to reasons explained in this comment.
BUGFIX: all: fix potential panic during components shutdown when metrics push is configured. See this issue. Thanks to @zhdd99 for the pull request.
BUGFIX: MetricsQL: properly process queries with too big lookbehind window such as foo[100y]. Previously, such queries could return empty responses even if foo is present in database. See this issue.
BUGFIX: MetricsQL: properly handle possible negative results caused by float operations precision error in rollup functions like rate() or increase(). See this issue.

v1.96.0

See changes here

v1.95.1

See changes here

v1.95.0

See changes here

v1.94.0

See changes here

v1.93.10

Released at 2024-01-17

v1.93.x is a line of LTS releases (e.g. long-time support). It contains important up-to-date bugfixes. The v1.93.x line will be supported for at least 12 months since v1.93.0 release

SECURITY: upgrade Go builder from Go1.21.5 to Go1.21.6. See the list of issues addressed in Go1.21.6.
BUGFIX: vminsert: properly accept samples via OpenTelemetry data ingestion protocol when these samples have no resource attributes. Previously such samples were silently skipped.
BUGFIX: vmstorage: added missing -inmemoryDataFlushInterval command-line flag, which was missing in VictoriaMetrics cluster after implementing this feature in v1.85.0.
BUGFIX: vmstorage: properly expire storage/prefetchedMetricIDs cache. Previously this cache was never expired, so it could grow big under high churn rate. This could result in increasing CPU load over time.
BUGFIX: vmalert: check -external.url schema when starting vmalert, must be http or https. Before, alertmanager could reject alert notifications if -external.url contained no or wrong schema.
BUGFIX: MetricsQL: properly return results from bottomk and bottomk_*() functions when some of these results contain NaN values. See this issue. Thanks to @xiaozongyang for the fix.
BUGFIX: MetricsQL: properly handle queries, which wrap rollup functions with multiple arguments without explicitly specified lookbehind window in square brackets into aggregate functions. For example, sum(quantile_over_time(0.5, process_resident_memory_bytes)) was resulting to expecting at least 2 args to ...; got 1 args error. Thanks to @atykhyy for the pull request.
BUGFIX: vmagent: properly assume role with AWS IRSA authorization. Previously role chaining was not supported. See this issue for details.
BUGFIX: all: fix potential panic during components shutdown when metrics push is configured. See this issue. Thanks to @zhdd99 for the pull request.
BUGFIX: vmctl: check for Error field in response from influx client during migration. Before, only network errors were checked. Thanks to @wozz for the pull request.
BUGFIX: vmctl: retry on import errors in vm-native mode. Before, retries happened only on writes into a network connection between source and destination. But errors returned by server after all the data was transmitted were logged, but not retried.

SECURITY: upgrade Go builder from Go1.21.5 to Go1.21.6. See the list of issues addressed in Go1.21.6.
BUGFIX: vmstorage: added missing -inmemoryDataFlushInterval command-line flag, which was missing in VictoriaMetrics cluster after implementing this feature in v1.85.0.
BUGFIX: MetricsQL: properly handle queries, which wrap rollup functions with multiple arguments without explicitly specified lookbehind window in square brackets into aggregate functions. For example, sum(quantile_over_time(0.5, process_resident_memory_bytes)) was resulting to expecting at least 2 args to ...; got 1 args error. Thanks to @atykhyy for the pull request.
BUGFIX: vmstorage: properly expire storage/prefetchedMetricIDs cache. Previously this cache was never expired, so it could grow big under high churn rate. This could result in increasing CPU load over time.
BUGFIX: MetricsQL: properly return results from bottomk and bottomk_... functions when some of these results contain NaN values. See this issue. Thanks to @xiaozongyang for the fix.
BUGFIX: all: fix potential panic during components shutdown when metrics push is configured. See this issue. Thanks to @zhdd99 for the pull request.