Cluster version


VictoriaMetrics is a fast, cost-effective and scalable time series database. It can be used as a long-term remote storage for Prometheus.

It is recommended using single-node version instead of cluster version for ingestion rates lower than a million of data points per second. Single-node version scales perfectly with the number of CPU cores, RAM and available storage space. Single-node version is easier to configure and operate comparing to cluster version, so think twice before sticking to cluster version.

Join our Slack or contact us with consulting and support questions.

Prominent features

Architecture overview

VictoriaMetrics cluster consists of the following services:

  • vmstorage - stores the data
  • vminsert - proxies the ingested data to vmstorage shards using consistent hashing
  • vmselect - performs incoming queries using the data from vmstorage

Each service may scale independently and may run on the most suitable hardware. vmstorage nodes don't know about each other, don't communicate with each other and don't share any data. This is shared nothing architecture. It increases cluster availability, simplifies cluster maintenance and cluster scaling.


VictoriaMetrics cluster supports multiple isolated tenants (aka namespaces). Tenants are identified by accountID or accountID:projectID, which are put inside request urls. See these docs for details. Some facts about tenants in VictoriaMetrics:

  • Each accountID and projectID is identified by an arbitrary 32-bit integer in the range [0 .. 2^32). If projectID is missing, then it is automatically assigned to 0. It is expected that other information about tenants such as auth tokens, tenant names, limits, accounting, etc. is stored in a separate relational database. This database must be managed by a separate service sitting in front of VictoriaMetrics cluster such as vmauth or vmgateway. Contact us if you need assistance with such service.

  • Tenants are automatically created when the first data point is written into the given tenant.

  • Data for all the tenants is evenly spread among available vmstorage nodes. This guarantees even load among vmstorage nodes when different tenants have different amounts of data and different query load.

  • The database performance and resource usage doesn't depend on the number of tenants. It depends mostly on the total number of active time series in all the tenants. A time series is considered active if it received at least a single sample during the last hour or it has been touched by queries during the last hour.

  • VictoriaMetrics doesn't support querying multiple tenants in a single request.


Compiled binaries for cluster version are available in the assets section of releases page. See archives containing cluster word.

Docker images for cluster version are available here:

  • vminsert -
  • vmselect -
  • vmstorage -

Building from sources

Source code for cluster version is available at cluster branch.

Production builds

There is no need in installing Go on a host system since binaries are built inside the official docker container for Go. This makes reproducible builds. So install docker and run the following command:

make vminsert-prod vmselect-prod vmstorage-prod

Production binaries are built into statically linked binaries. They are put into bin folder with -prod suffixes:

$ make vminsert-prod vmselect-prod vmstorage-prod
$ ls -1 bin

Development Builds

  1. Install go. The minimum supported version is Go 1.15.
  2. Run make from the repository root. It should build vmstorage, vmselect and vminsert binaries and put them into the bin folder.

Building docker images

Run make package. It will build the following docker images locally:

  • victoriametrics/vminsert:<PKG_TAG>
  • victoriametrics/vmselect:<PKG_TAG>
  • victoriametrics/vmstorage:<PKG_TAG>

<PKG_TAG> is auto-generated image tag, which depends on source code in the repository. The <PKG_TAG> may be manually set via PKG_TAG=foobar make package.

By default images are built on top of alpine image in order to improve debuggability. It is possible to build an image on top of any other base image by setting it via <ROOT_IMAGE> environment variable. For example, the following command builds images on top of scratch image:

ROOT_IMAGE=scratch make package


Cluster setup

A minimal cluster must contain the following nodes:

  • a single vmstorage node with -retentionPeriod and -storageDataPath flags
  • a single vminsert node with -storageNode=<vmstorage_host>:8400
  • a single vmselect node with -storageNode=<vmstorage_host>:8401

It is recommended to run at least two nodes for each service for high availability purposes.

An http load balancer such as vmauth or nginx must be put in front of vminsert and vmselect nodes:

  • requests starting with /insert must be routed to port 8480 on vminsert nodes.
  • requests starting with /select must be routed to port 8481 on vmselect nodes.

Ports may be altered by setting -httpListenAddr on the corresponding nodes.

It is recommended setting up monitoring for the cluster.

Environment variables

Each flag values can be set thru environment variables by following these rules:

  • The -envflag.enable flag must be set
  • Each . in flag names must be substituted by _ (for example -insert.maxQueueDuration <duration> will translate to insert_maxQueueDuration=<duration>)
  • For repeating flags, an alternative syntax can be used by joining the different values into one using , as separator (for example -storageNode <nodeA> -storageNode <nodeB> will translate to storageNode=<nodeA>,<nodeB>)
  • It is possible setting prefix for environment vars with -envflag.prefix. For instance, if -envflag.prefix=VM_, then env vars must be prepended with VM_


All the cluster components expose various metrics in Prometheus-compatible format at /metrics page on the TCP port set in -httpListenAddr command-line flag. By default the following TCP ports are used:

  • vminsert - 8480
  • vmselect - 8481
  • vmstorage - 8482

It is recommended setting up vmagent or Prometheus to scrape /metrics pages from all the cluster components, so they can be monitored and analyzed with the official Grafana dashboard for VictoriaMetrics cluster or an alternative dashboard for VictoriaMetrics cluster.

It is recommended setting up alerts in vmalert or in Prometheus from this config.

URL format

  • URLs for data ingestion: http://<vminsert>:8480/insert/<accountID>/<suffix>, where:
    • <accountID> is an arbitrary 32-bit integer identifying namespace for data ingestion (aka tenant). It is possible to set it as accountID:projectID, where projectID is also arbitrary 32-bit integer. If projectID isn't set, then it equals to 0.
    • <suffix> may have the following values:
      • prometheus and prometheus/api/v1/write - for inserting data with Prometheus remote write API
      • influx/write and influx/api/v2/write - for inserting data with Influx line protocol.
      • opentsdb/api/put - for accepting OpenTSDB HTTP /api/put requests. This handler is disabled by default. It is exposed on a distinct TCP address set via -opentsdbHTTPListenAddr command-line flag. See these docs for details.
      • prometheus/api/v1/import - for importing data obtained via api/v1/export on vmselect (see below).
      • prometheus/api/v1/import/native - for importing data obtained via api/v1/export/native on vmselect (see below).
      • prometheus/api/v1/import/csv - for importing arbitrary CSV data. See these docs for details.
      • prometheus/api/v1/import/prometheus - for importing data in Prometheus text exposition format and in OpenMetrics format. See these docs for details.
  • URLs for Prometheus querying API: http://<vmselect>:8481/select/<accountID>/prometheus/<suffix>, where:
    • <accountID> is an arbitrary number identifying data namespace for the query (aka tenant)
    • <suffix> may have the following values:
      • api/v1/query - performs PromQL instant query.
      • api/v1/query_range - performs PromQL range query.
      • api/v1/series - performs series query.
      • api/v1/labels - returns a list of label names.
      • api/v1/label/<label_name>/values - returns values for the given <label_name> according to API.
      • federate - returns federated metrics.
      • api/v1/export - exports raw data in JSON line format. See this article for details.
      • api/v1/export/native - exports raw data in native binary format. It may be imported into another VictoriaMetrics via api/v1/import/native (see above).
      • api/v1/export/csv - exports data in CSV. It may be imported into another VictoriaMetrics via api/v1/import/csv (see above).
      • api/v1/series/count - returns the total number of series.
      • api/v1/status/tsdb - for time series stats. See these docs for details.
      • api/v1/status/active_queries - for currently executed active queries. Note that every vmselect maintains an independent list of active queries, which is returned in the response.
      • api/v1/status/top_queries - for listing the most frequently executed queries and queries taking the most duration.
  • URLs for Graphite Metrics API: http://<vmselect>:8481/select/<accountID>/graphite/<suffix>, where:
    • <accountID> is an arbitrary number identifying data namespace for query (aka tenant)
    • <suffix> may have the following values:
      • render - implements Graphite Render API. See these docs. This functionality is available in Enterprise package.
      • metrics/find - searches Graphite metrics. See these docs.
      • metrics/expand - expands Graphite metrics. See these docs.
      • metrics/index.json - returns all the metric names. See these docs.
      • tags/tagSeries - registers time series. See these docs.
      • tags/tagMultiSeries - register multiple time series. See these docs.
      • tags - returns tag names. See these docs.
      • tags/<tag_name> - returns tag values for the given <tag_name>. See these docs.
      • tags/findSeries - returns series matching the given expr. See these docs.
      • tags/autoComplete/tags - returns tags matching the given tagPrefix and/or expr. See these docs.
      • tags/autoComplete/values - returns tag values matching the given valuePrefix and/or expr. See these docs.
      • tags/delSeries - deletes series matching the given path. See these docs.
  • URL for query stats across all tenants: http://<vmselect>:8481/api/v1/status/top_queries. It lists with the most frequently executed queries and queries taking the most duration.

  • URL for time series deletion: http://<vmselect>:8481/delete/<accountID>/prometheus/api/v1/admin/tsdb/delete_series?match[]=<timeseries_selector_for_delete>. Note that the delete_series handler should be used only in exceptional cases such as deletion of accidentally ingested incorrect time series. It shouldn't be used on a regular basis, since it carries non-zero overhead.

  • vmstorage nodes provide the following HTTP endpoints on 8482 port:
    • /internal/force_merge - initiate forced compactions on the given vmstorage node.
    • /snapshot/create - create instant snapshot, which can be used for backups in background. Snapshots are created in <storageDataPath>/snapshots folder, where <storageDataPath> is the corresponding command-line flag value.
    • /snapshot/list - list available snasphots.
    • /snapshot/delete?snapshot=<id> - delete the given snapshot.
    • /snapshot/delete_all - delete all the snapshots.

    Snapshots may be created independently on each vmstorage node. There is no need in synchronizing snapshots' creation across vmstorage nodes.

Cluster resizing and scalability

Cluster performance and capacity scales with adding new nodes.

  • vminsert and vmselect nodes are stateless and may be added / removed at any time. Do not forget updating the list of these nodes on http load balancer. Adding more vminsert nodes scales data ingestion rate. See this comment about ingestion rate scalability. Adding more vmselect nodes scales select queries rate.
  • vmstorage nodes own the ingested data, so they cannot be removed without data loss. Adding more vmstorage nodes scales cluster capacity.

Steps to add vmstorage node:

  1. Start new vmstorage node with the same -retentionPeriod as existing nodes in the cluster.
  2. Gradually restart all the vmselect nodes with new -storageNode arg containing <new_vmstorage_host>:8401.
  3. Gradually restart all the vminsert nodes with new -storageNode arg containing <new_vmstorage_host>:8400.

Updating / reconfiguring cluster nodes

All the node types - vminsert, vmselect and vmstorage - may be updated via graceful shutdown. Send SIGINT signal to the corresponding process, wait until it finishes and then start new version with new configs.

Cluster should remain in working state if at least a single node of each type remains available during the update process. See cluster availability section for details.

Cluster availability

  • HTTP load balancer must stop routing requests to unavailable vminsert and vmselect nodes.
  • The cluster remains available if at least a single vmstorage node exists:

    • vminsert re-routes incoming data from unavailable vmstorage nodes to healthy vmstorage nodes
    • vmselect continues serving partial responses if at least a single vmstorage node is available. If consistency over availability is preferred, then either pass -search.denyPartialResponse command-line flag to vmselect or pass deny_partial_response=1 query arg in requests to vmselect.

vmselect doesn't serve partial responses for API handlers returning raw datapoints - /api/v1/export* endpoints, since users usually expect this data is always complete.

Data replication can be used for increasing storage durability. See these docs for details.

Capacity planning

Each instance type - vminsert, vmselect and vmstorage - can run on the most suitable hardware.


  • The recommended total number of vCPU cores for all the vminsert instances can be calculated from the ingestion rate: vCPUs = ingestion_rate / 150K.
  • The recommended number of vCPU cores per each vminsert instance should equal to the number of vmstorage instances in the cluster.
  • The amount of RAM per each vminsert instance should be 1GB or more. RAM is used as a buffer for spikes in ingestion rate. The maximum amount of used RAM per vminsert node can be tuned with -memory.allowedPercent or -memory.allowedBytes command-line flags. For instance, -memory.allowedPercent=20 limits the maximum amount of used RAM to 20% of the available RAM on the host system.
  • Sometimes -rpc.disableCompression command-line flag on vminsert instances could increase ingestion capacity at the cost of higher network bandwidth usage between vminsert and vmstorage.


  • The recommended total number of vCPU cores for all the vmstorage instances can be calculated from the ingestion rate: vCPUs = ingestion_rate / 150K.
  • The recommended total amount of RAM for all the vmstorage instances can be calculated from the number of active time series: RAM = 2 * active_time_series * 1KB. Time series is active if it received at least a single data point during the last hour or if it has been queried during the last hour. The required RAM per each vmstorage should be multiplied by -replicationFactor if replication is enabled. Additional RAM can be required for query processing. Calculated RAM requrements may differ from actual RAM requirements due to various factors:
    • The average number of labels per time series. More labels require more RAM.
    • The average length of label names and label values. Longer labels require more RAM.
    • The type of queries. Heavy queries that scan big number of time series over long time ranges require more RAM.
  • The recommended total amount of storage space for all the vmstorage instances can be calculated from the ingestion rate and retention: storage_space = ingestion_rate * retention_seconds.


The recommended hardware for vmselect instances highly depends on the type of queries. Lightweight queries over small number of time series usually require small number of vCPU cores and small amount of RAM on vmselect, while heavy queries over big number of time series (>10K) usually require bigger number of vCPU cores and bigger amounts of RAM.

In general it is recommended increasing the number of vCPU cores and RAM per vmselect node for higher query performance, while adding new vmselect nodes only when old nodes are overloaded with incoming query stream.

High availability

It is recommended to run all the components for a single cluster in the same subnetwork with high bandwidth, low latency and low error rates. This improves cluster performance and availability. It isn't recommended spreading components for a single cluster across multiple availability zones, since cross-AZ network usually has lower bandwidth, higher latency and higher error rates comparing the network inside AZ.

If you need multi-AZ setup, then it is recommended running independed clusters in each AZ and setting up vmagent in front of these clusters, so it could replicate incoming data into all the cluster. Then promxy could be used for querying the data from multiple clusters.

Another solution is to use multi-level cluster setup, where the top level of vminsert nodes replicate data among the lower level of vminsert nodes located at different availability zones. These vminsert nodes then spread the data among vmstorage nodes in each AZ. See these docs for more details.

Multi-level cluster setup

vminsert nodes can accept data from another vminsert nodes starting from v1.60.0 if -clusternativeListenAddr command-line flag is set. For example, if vminsert is started with -clusternativeListenAddr=:8400 command-line flag, then it can accept data from another vminsert nodes at TCP port 8400 in the same way as vmstorage nodes do. This allows chaining vminsert nodes and building multi-level cluster topologies with flexible configs. For example, the top level of vminsert nodes can replicate data among the second level of vminsert nodes located in distinct availability zones (AZ), while the second-level vminsert nodes can spread the data among vmstorage nodes located in the same AZ. Such setup guarantees cluster availability if some AZ becomes unavailable. The data from all the vmstorage nodes in all the AZs can be read via vmselect nodes, which are configured to query all the vmstorage nodes in all the availability zones (e.g. all the vmstorage addresses are passed via -storageNode command-line flag to vmselect nodes). Additionally, -replicationFactor=k+1 must be passed to vmselect nodes, where k is the lowest number of vmstorage nodes in a single AZ. See replication docs for more details.


Helm chart simplifies managing cluster version of VictoriaMetrics in Kubernetes. It is available in the helm-charts repository.

Kubernetes operator

K8s operator simplifies managing VictoriaMetrics components in Kubernetes.

Replication and data safety

By default VictoriaMetrics offloads replication to the underlying storage pointed by -storageDataPath.

The replication can be enabled by passing -replicationFactor=N command-line flag to vminsert. This guarantees that all the data remains available for querying if up to N-1 vmstorage nodes are unavailable. The cluster must contain at least 2*N-1 vmstorage nodes, where N is replication factor, in order to maintain the given replication factor for newly ingested data when N-1 of storage nodes are lost. For example, when -replicationFactor=3 is passed to vminsert, then it replicates all the ingested data to 3 distinct vmstorage nodes, so up to 2 vmstorage nodes can be lost without data loss. The minimum number of vmstorage nodes should be equal to 2*3-1 = 5, so when 2 vmstorage nodes are lost, the remaining 3 vmstorage nodes could provide the -replicationFactor=3 for newly ingested data.

When the replication is enabled, -replicationFactor=N and -dedup.minScrapeInterval=1ms command-line flag must be passed to vmselect nodes. The -replicationFactor=N improves query performance when up to N-1 vmstorage nodes respond slowly and/or temporarily unavailable. Sometimes -replicationFactor at vmselect nodes can result in partial responses. See this issues for details. The -dedup.minScrapeInterval=1ms de-duplicates replicated data during queries. It is OK if -dedup.minScrapeInterval exceeds 1ms when deduplication is used additionally to replication.

Note that replication doesn't save from disaster, so it is recommended performing regular backups. See these docs for details.

Note that the replication increases resource usage - CPU, RAM, disk space, network bandwidth - by up to -replicationFactor times. So it may be worth offloading the replication to underlying storage pointed by -storageDataPath such as Google Compute Engine persistent disk, which is protected from data loss and data corruption. It also provide consistently high performance and may be resized without downtime. HDD-based persistent disks should be enough for the majority of use cases.

It is recommended using durable replicated persistent volumes in Kubernetes.


It is recommended performing periodical backups from instant snapshots for protecting from user errors such as accidental data deletion.

The following steps must be performed for each vmstorage node for creating a backup:

  1. Create an instant snapshot by navigating to /snapshot/create HTTP handler. It will create snapshot and return its name.
  2. Archive the created snapshot from <-storageDataPath>/snapshots/<snapshot_name> folder using vmbackup. The archival process doesn't interfere with vmstorage work, so it may be performed at any suitable time.
  3. Delete unused snapshots via /snapshot/delete?snapshot=<snapshot_name> or /snapshot/delete_all in order to free up occupied storage space.

There is no need in synchronizing backups among all the vmstorage nodes.

Restoring from backup:

  1. Stop vmstorage node with kill -INT.
  2. Restore data from backup using vmrestore into -storageDataPath directory.
  3. Start vmstorage node.


All the cluster components provide the following handlers for profiling:

  • http://vminsert:8480/debug/pprof/heap for memory profile and http://vminsert:8480/debug/pprof/profile for CPU profile
  • http://vmselect:8481/debug/pprof/heap for memory profile and http://vmselect:8481/debug/pprof/profile for CPU profile
  • http://vmstorage:8482/debug/pprof/heap for memory profile and http://vmstorage:8482/debug/pprof/profile for CPU profile

Example command for collecting cpu profile from vmstorage:

curl -s http://vmstorage:8482/debug/pprof/profile > cpu.pprof

Example command for collecting memory profile from vminsert:

curl -s http://vminsert:8480/debug/pprof/heap > mem.pprof

Community and contributions

We are open to third-party pull requests provided they follow the KISS design principle:

  • Prefer simple code and architecture.
  • Avoid complex abstractions.
  • Avoid magic code and fancy algorithms.
  • Avoid big external dependencies.
  • Minimize the number of moving parts in the distributed system.
  • Avoid automated decisions, which may hurt cluster availability, consistency or performance.

Adhering to the KISS principle simplifies the resulting code and architecture, so it can be reviewed, understood and verified by many people.

Due to KISS, cluster version of VictoriaMetrics has no the following "features" popular in distributed computing world:

  • Fragile gossip protocols. See failed attempt in Thanos.
  • Hard-to-understand-and-implement-properly Paxos protocols.
  • Complex replication schemes, which may go nuts in unforeseen edge cases. See replication docs for details.
  • Automatic data reshuffling between storage nodes, which may hurt cluster performance and availability.
  • Automatic cluster resizing, which may cost you a lot of money if improperly configured.
  • Automatic discovering and addition of new nodes in the cluster, which may mix data between dev and prod clusters :)
  • Automatic leader election, which may result in split brain disaster on network errors.

Reporting bugs

Report bugs and propose new features here.

Zip contains three folders with different image orientation (main color and inverted version).

Files included in each folder:

  • 2 JPEG Preview files
  • 2 PNG Preview files with transparent background
  • 2 EPS Adobe Illustrator EPS10 files

Logo Usage Guidelines

Font used:

  • Lato Black
  • Lato Regular

Color Palette:

We kindly ask:

  • Please don't use any other font instead of suggested.
  • There should be sufficient clear space around the logo.
  • Do not change spacing, alignment, or relative locations of the design elements.
  • Do not change the proportions of any of the design elements or the design itself. You may resize as needed but must retain all proportions.