VictoriaLogs Cluster #

Cluster mode in VictoriaLogs provides horizontal scaling to many nodes when single-node VictoriaLogs reaches vertical scalability limits of a single host. If you have the ability to run a single-node VictoriaLogs on a host with more CPU / RAM / storage space / storage IO, then it is preferred to do this instead of switching to cluster mode, since a single-node VictoriaLogs instance has the following advantages over cluster mode:

• It is easier to configure, manage and troubleshoot, since it consists of a single self-contained component.
• It provides better performance and capacity on the same hardware, since it doesn’t need to transfer data over the network between cluster components.

The migration path from a single-node VictoriaLogs to cluster mode is very easy - just upgrade a single-node VictoriaLogs executable to the latest available release and add it to the list of vlstorage nodes passed via -storageNode command-line flag to vlinsert and vlselect components of the cluster mode. See cluster architecture for more details about VictoriaLogs cluster components.

See quick start guide on how to start working with VictoriaLogs cluster.

Architecture #

VictoriaLogs in cluster mode is composed of three main components: vlinsert, vlselect, and vlstorage.

Ingestion flow:

sequenceDiagram
    participant LS as Log Sources
    participant VI as vlinsert
    participant VS1 as vlstorage-1
    participant VS2 as vlstorage-2

    Note over LS,VI: Log Ingestion Flow
    LS->>VI: Send logs via supported protocols
    VI->>VS1: POST /internal/insert (HTTP)
    VI->>VS2: POST /internal/insert (HTTP)
    Note right of VI: Distributes logs evenly<br/>across vlstorage nodes

Querying flow:

sequenceDiagram
    participant QC as Query Client
    participant VL as vlselect
    participant VS1 as vlstorage-1
    participant VS2 as vlstorage-2

    Note over QC,VL: Query Flow
    QC->>VL: Query via HTTP endpoints
    VL->>VS1: GET /internal/select/* (HTTP)
    VL->>VS2: GET /internal/select/* (HTTP)
    VS1-->>VL: Return local results
    VS2-->>VL: Return local results
    VL->>QC: Processed & aggregated results

• vlinsert handles log ingestion via all supported protocols .
  It distributes (shards) incoming logs evenly across vlstorage nodes, as specified by the -storageNode command-line flag.

• vlselect receives queries through all supported HTTP query endpoints .
  It fetches the required data from the configured vlstorage nodes, processes the queries, and returns the results.

• vlstorage performs two key roles:

  • It stores logs received from vlinsert at the directory defined by the -storageDataPath flag.
    See storage configuration docs for details.
  • It handles queries from vlselect by retrieving and transforming the requested data locally before returning results.

Each vlstorage node operates as a self-contained VictoriaLogs instance.
Refer to the single-node and cluster mode duality documentation for more information.
This design allows you to reuse existing single-node VictoriaLogs instances by listing them in the -storageNode flag for vlselect, enabling unified querying across all nodes.

All VictoriaLogs components are horizontally scalable and can be deployed on hardware best suited to their respective workloads.
vlinsert and vlselect can be run on the same node, which allows the minimal cluster to consist of just one vlstorage node and one node acting as both vlinsert and vlselect.
However, for production environments, it is recommended to separate vlinsert and vlselect roles to avoid resource contention — for example, to prevent heavy queries from interfering with log ingestion.

Communication between vlinsert / vlselect and vlstorage is done via HTTP over the port specified by the -httpListenAddr flag:

• vlinsert sends data to the /internal/insert endpoint on vlstorage.
• vlselect sends queries to endpoints under /internal/select/ on vlstorage.

This HTTP-based communication model allows you to use reverse proxies for authorization, routing, and encryption between components.
Use of vmauth is recommended for managing access control. See Security and Load balancing docs for details.

For advanced setups, refer to the multi-level cluster setup documentation.

High availability #

VictoriaLogs cluster provides high availability for data ingestion path . It continues to accept incoming logs if some of the vlstorage nodes are temporarily unavailable. vlinsert evenly spreads new logs among the remaining available vlstorage nodes in this case, so newly ingested logs are properly stored and are available for querying without any delays. This allows performing maintenance tasks for vlstorage nodes (such as upgrades, configuration updates, etc.) without worrying about data loss. Make sure that the remaining vlstorage nodes have enough capacity for the increased data ingestion workload, in order to avoid availability problems.

VictoriaLogs cluster returns 502 Bad Gateway errors for incoming queries if some of the vlstorage nodes are unavailable. This guarantees consistent query responses (e.g. all the stored logs are taken into account during the query) during maintenance tasks at vlstorage nodes. Note that all the newly incoming logs are properly stored to the remaining vlstorage nodes - see the paragraph above, so they become available for querying immediately after all the vlstorage nodes return back to the cluster.

There are practical cases when it is preferred to return partial responses instead of 502 Bad Gateway errors if some of vlstorage nodes are unavailable. See these docs on how to achieve this.

The real HA scheme for both data ingestion and querying can be built only when copies of logs are sent into independent VictoriaLogs instances (or clusters) located in fully independent availability zones (datacenters).

If an AZ becomes unavailable, then new logs continue to be written to the remaining AZ, while queries return full responses from the remaining AZ. When the AZ becomes available, then the pending buffered logs can be written to it, so the AZ can be used for querying full responses. This HA scheme can be built with the help of vlagent for data replication and buffering, and vmauth for data querying:

flowchart TB
  subgraph haSolution["HA Solution"]
    direction TB

    subgraph ingestion["Ingestion Layer"]
      direction TB
      LS["Log Sources<br/>(Applications)"]
      VLAGENT["Log Collector<br/>• Buffering<br/>• Replication<br/>• Delivery Guarantees"]
      LS --> VLAGENT
    end

    subgraph storage["Storage Layer"]
      direction TB

      subgraph zoneA["Zone A"]
        VLA["VictoriaLogs Cluster A"]
      end

      subgraph zoneB["Zone B"]
        VLB["VictoriaLogs Cluster B"]
      end

      VLAGENT -->|"Replicate logs to<br/>Zone A cluster"| VLA
      VLAGENT -->|"Replicate logs to<br/>Zone B cluster"| VLB
    end

    subgraph query["Query Layer"]
      direction TB
      LB["Load Balancer<br/>(vmauth)<br/>• Health Checks<br/>• Failover<br/>• Query Distribution"]
      QC["Query Clients<br/>(Grafana, API)"]
      VLA -->|"Serve queries from<br/>Zone A cluster"| LB
      VLB -->|"Serve queries from<br/>Zone B cluster"| LB
      LB --> QC
    end
  end

style VLAGENT fill:#9bc7e4
style VLA fill:#ae9be4
style VLB fill:#ae9be4
style LB fill:#9bc7e4
style QC fill:#9fe49b
style LS fill:#9fe49b

• vlagent receives and replicates logs to two VictoriaLogs clusters. If one cluster becomes unavailable, the log shipper continues sending logs to the remaining healthy cluster. It also buffers logs that cannot be delivered to the unavailable cluster. When the failed cluster becomes available again, the log shipper sends the buffered logs and then resumes sending new logs to it. This guarantees that both clusters have full copies of all the ingested logs.
• vmauth routes query requests to healthy VictoriaLogs clusters. If one cluster becomes unavailable, vmauth detects this and automatically redirects all query traffic to the remaining healthy cluster.

There is no magic coordination logic or consensus algorithms in this scheme. This simplifies managing and troubleshooting this HA scheme.

See also replication and Security and Load balancing docs .

Replication #

vlinsert doesn’t replicate incoming logs among vlstorage nodes. Instead, it spreads evenly (shards) incoming logs among vlstorage nodes specified in the -storageNode command-line flag. This provides cost-efficient linear scalability for the cluster capacity, data ingestion performance and querying performance proportional to the number of vlstorage nodes.

It is recommended making regular backups for the data stored across all the vlstorage nodes in order to make sure that the data isn’t lost in case of any disaster (such as accidental data removal because of incorrect config updates or incorrect upgrades, or physical corruption of the data on the persistent storage). See how to backup and restore data for VictoriaLogs - these docs apply to vlstorage nodes .

If you need restoring the data between the backup time and the current time, then it is recommended building HA setup for VictoriaLogs cluster , so you could copy the needed per-day partitions from cluster replica.

Usually the disaster event occurs rarely (e.g. once per year). Every such event has unique preconditions and consequences, so it is impossible to automate recovering from disaster events. These events require human attention and carefully thought manual actions, so there is little practical sense in relying on automatic data recovery from the magically replicated data among storage nodes.

Single-node and cluster mode duality #

Every vlstorage node can be used as a single-node VictoriaLogs instance:

• It can accept logs via all the supported data ingestion protocols .
• It can accept select queries via all the supported HTTP querying endpoints .

A single-node VictoriaLogs instance can be used as vlstorage node in VictoriaLogs cluster:

• It accepts data ingestion requests from vlinsert via /internal/insert HTTP endpoint at the TCP port specified via -httpListenAddr command-line flag.
• It accepts queries from vlselect via /internal/select/* HTTP endpoints at the TCP port specified via -httpListenAddr command-line flag.

See also security docs .

Multi-level cluster setup #

• vlinsert can send the ingested logs to other vlinsert nodes if they are specified via -storageNode command-line flag. This allows building multi-level data ingestion schemes when top-level vlinsert spreads the incoming logs evenly among multiple lower-level clusters of VictoriaLogs.

• vlselect can send queries to other vlselect nodes if they are specified via -storageNode command-line flag. This allows building multi-level cluster schemes when top-level vlselect queries multiple lower-level clusters of VictoriaLogs.

See security docs on how to protect communications between multiple levels of vlinsert and vlselect nodes.

Security #

All the VictoriaLogs cluster components must run in a protected internal network without direct access from the Internet. vlstorage must have no access from the Internet. HTTP authorization proxies such as vmauth must be used in front of vlinsert and vlselect for authorizing access to these components from the Internet. See Security and Load balancing docs .

It is possible to disallow access to /internal/insert and /internal/select/* endpoints at a single-node VictoriaLogs instance by running it with -internalinsert.disable and -internalselect.disable command-line flags. Note that vlagent sends the collected logs to the /internal/insert endpoint, so it should be available for data ingestion if you use vlagent.

TLS #

By default, vlinsert and vlselect communicate with vlstorage via unencrypted HTTP. This is OK if all these components are located in the same protected internal network. This isn’t OK if these components communicate over the Internet, since a third party can intercept or modify the transferred data. It is recommended to switch to HTTPS in this case:

• Specify -tls, -tlsCertFile and -tlsKeyFile command-line flags at vlstorage, so it accepts incoming requests over HTTPS instead of HTTP at the corresponding -httpListenAddr:

        ./victoria-logs-prod -httpListenAddr=... -storageDataPath=... -tls -tlsCertFile=/path/to/certfile -tlsKeyFile=/path/to/keyfile
      

• Specify -storageNode.tls command-line flag at vlinsert and vlselect, which communicate with the vlstorage over untrusted networks such as the Internet:

        ./victoria-logs-prod -storageNode=... -storageNode.tls
      

It is also recommended to authorize HTTPS requests to vlstorage via Basic Auth:

• Specify -httpAuth.username and -httpAuth.password command-line flags at vlstorage, so it verifies the Basic Auth username + password in HTTPS requests received via -httpListenAddr:

        ./victoria-logs-prod -httpListenAddr=... -storageDataPath=... -tls -tlsCertFile=... -tlsKeyFile=... -httpAuth.username=... -httpAuth.password=...
      

• Specify -storageNode.username and -storageNode.password command-line flags at vlinsert and vlselect, which communicate with the vlstorage over untrusted networks:

        ./victoria-logs-prod -storageNode=... -storageNode.tls -storageNode.username=... -storageNode.password=...
      

Another option is to use third-party HTTP proxies such as vmauth , nginx, etc. to authorize and encrypt communications between VictoriaLogs cluster components over untrusted networks.

By default, all the components (vlinsert, vlselect, vlstorage) support all the HTTP endpoints including /insert/* and /select/*. It is recommended to disable select endpoints on vlinsert and insert endpoints on vlselect:

      # Disable select endpoints on vlinsert
./victoria-logs-prod -storageNode=... -select.disable

# Disable insert endpoints on vlselect
./victoria-logs-prod -storageNode=... -insert.disable
    

This helps prevent sending select requests to vlinsert nodes or insert requests to vlselect nodes in case of a misconfiguration in the authorization proxy in front of the vlinsert and vlselect nodes.

See also mTLS .

mTLS #

Enterprise version of VictoriaLogs supports the ability to verify client TLS certificates at the vlstorage side for TLS connections established from vlinsert and vlselect nodes (aka mTLS ). See TLS docs for details on how to set up TLS communications between VictoriaLogs cluster nodes.

mTLS authentication can be enabled by passing the -mtls command-line flag to the vlstorage node in addition to the -tls command-line flag. In this case it verifies TLS client certificates for connections from vlinsert and vlselect at the address specified via -httpListenAddr command-line flag.

The client TLS certificate must be specified at vlinsert and vlselect nodes via -storageNode.tlsCertFile and -storageNode.tlsKeyFile command-line flags.

By default, the system-wide root CA certificates are used for verifying client TLS certificates. The -mtlsCAFile command-line flag can be used at vlstorage for pointing to custom root CA certificates.

See also generic mTLS docs for VictoriaLogs .

Enterprise version of VictoriaLogs can be downloaded and evaluated for free from the releases page . See how to request a free trial license .

Rebalancing #

Every vlinsert node spreads evenly (shards) incoming logs among vlstorage nodes specified in the -storageNode command-line flag according to the VictoriaLogs cluster architecture . This guarantees that the data is spread evenly among vlstorage nodes. When new vlstorage nodes are added to the -storageNode list at vlinsert, then all the newly ingested logs are spread evenly among old and new vlstorage nodes, while historical data remains on the old vlstorage nodes. This improves data ingestion performance and querying performance for typical production workloads, since newly ingested logs are spread evenly across all the vlstorage nodes, while typical queries are performed over the newly ingested logs, which are already present among all the vlstorage nodes. This also provides the following benefits comparing to the scheme with automatic data rebalancing:

• Cluster performance remains reliable just after adding new vlstorage nodes, since network bandwidth, disk IO and CPU resources aren’t spent on automatic data rebalancing, which may take days for re-balancing of petabytes of data.
• This eliminates the whole class of hard-to-troubleshoot and resolve issues, which may happen with the cluster during automatic data rebalancing. For example, what happens if some of vlstorage nodes become unavailable during the re-balancing? Or what happens if new vlstorage nodes are added while the previous data re-balancing isn’t finished yet?
• This allows building flexible cluster schemes where distinct subsets of vlinsert nodes spread incoming logs among different subsets of vlstorage nodes with different configs and different hardware resources.

The following approaches exist for manual data re-balancing among old and new vlstorage nodes if it is really needed:

• To wait until historical data is automatically deleted from old vlstorage nodes according to the configured retention . Then old and new vlstorage nodes will have equal amounts of data.
• To configure vlinsert to write newly ingested logs only to new vlstorage nodes, while vlselect nodes should continue querying data from all the vlstorage nodes. Then wait until the data size on the new vlstorage nodes becomes equal to the data size on the old vlstorage nodes, and return back old vlstorage nodes to -storageNode list at vlinsert.
• To manually move historical per-day partitions from old vlstorage nodes to new vlstorage nodes. VictoriaLogs provides the functionality, which simplifies doing this work without the need to stop or restart vlstorage nodes - see partitions lifecycle docs .

Quick start #

The following topics for are covered below:

• How to download the VictoriaLogs executable.
• How to start a VictoriaLogs cluster, which consists of two vlstorage nodes, a single vlinsert node and a single vlselect node running on localhost according to cluster architecture .
• How to ingest logs into the cluster.
• How to query the ingested logs.

If you want running VictoriaLogs cluster in Kubernetes, then please read these docs .

Download and unpack the latest VictoriaLogs release:

      curl -L -O https://github.com/VictoriaMetrics/VictoriaLogs/releases/download/v1.38.0/victoria-logs-linux-amd64-v1.38.0.tar.gz
tar xzf victoria-logs-linux-amd64-v1.38.0.tar.gz
    

Start the first vlstorage node , which accepts incoming requests at the port 9491 and stores the ingested logs in the victoria-logs-data-1 directory:

      ./victoria-logs-prod -httpListenAddr=:9491 -storageDataPath=victoria-logs-data-1 &
    

This command and all the following commands start cluster components as background processes. Use jobs, fg, bg commands for manipulating the running background processes. Use the kill command and/or Ctrl+C to stop running processes when they are no longer needed. See these docs for details.

Start the second vlstorage node, which accepts incoming requests at the port 9492 and stores the ingested logs in the victoria-logs-data-2 directory:

      ./victoria-logs-prod -httpListenAddr=:9492 -storageDataPath=victoria-logs-data-2 &
    

Start the vlinsert node, which accepts logs at the port 9481 and spreads them evenly across the two vlstorage nodes started above:

      ./victoria-logs-prod -httpListenAddr=:9481 -storageNode=localhost:9491,localhost:9492 &
    

Start the vlselect node, which accepts incoming queries at the port 9471 and requests the needed data from vlstorage nodes started above:

      ./victoria-logs-prod -httpListenAddr=:9471 -storageNode=localhost:9491,localhost:9492 &
    

Note that all the VictoriaLogs cluster components - vlstorage, vlinsert and vlselect - share the same executable - victoria-logs-prod. Their roles depend on whether the -storageNode command-line flag is set - if this flag is set, then the executable runs in vlinsert and vlselect modes. Otherwise, it runs in vlstorage mode, which is identical to a single-node VictoriaLogs mode .

Let’s ingest some logs (aka wide events ) from GitHub archive into the VictoriaLogs cluster with the following command:

      curl -s https://data.gharchive.org/$(date -d '2 days ago' '+%Y-%m-%d')-10.json.gz \
        | curl -T - -X POST -H 'Content-Encoding: gzip' 'http://localhost:9481/insert/jsonline?_time_field=created_at&_stream_fields=type'
    

Let’s query the ingested logs via /select/logsql/query HTTP endpoint . For example, the following command returns the number of stored logs in the cluster:

      curl http://localhost:9471/select/logsql/query -d 'query=* | count()'
    

See these docs for details on how to query logs from the command line.

Logs can also be explored and queried via the built-in Web UI . Open http://localhost:9471/select/vmui/ in the web browser, select last 7 days time range in the top right corner and explore the ingested logs. See LogsQL docs to familiarize yourself with the query language.

Every vlstorage node can be queried individually because it is equivalent to a single-node VictoriaLogs . For example, the following command returns the number of stored logs at the first vlstorage node started above:

      curl http://localhost:9491/select/logsql/query -d 'query=* | count()'
    

We recommend reading key concepts before you start working with VictoriaLogs.

See also security docs .

Capacity planning #

It is recommended leaving the following amounts of spare resource across all the components of VictoriaLogs cluster :

• 50% of free RAM for reducing the probability of OOM (out of memory) crashes and slowdowns during temporary spikes in workload.
• 50% of spare CPU for reducing the probability of slowdowns during temporary spikes in workload.
• At least 20% of free storage space at vlstorage nodes at the directory pointed by the -storageDataPath command-line flag. Too small amounts of free disk space may result in significant slowdown for both data ingestion and querying because of inability to merge newly created smaller data parts into bigger data parts.

Performance tuning #

Cluster components of VictoriaLogs automatically adjust their settings for the best performance and the lowest resource usage on the given hardware. So there is no need for any tuning of these components in general. The following options can be used for achieving higher performance / lower resource usage on systems with constrained resources:

• vlinsert limits the number of concurrent requests to every vlstorage node. The default concurrency works great in most cases. Sometimes it can be increased via -insert.concurrency command-line flag at vlinsert in order to achieve higher data ingestion rate at the cost of higher RAM usage at vlinsert and vlstorage nodes.

• vlinsert compresses the data sent to vlstorage nodes in order to reduce network bandwidth usage at the cost of slightly higher CPU usage at vlinsert and vlstorage nodes. The compression can be disabled by passing -insert.disableCompression command-line flag to vlinsert. This reduces CPU usage at vlinsert and vlstorage nodes at the cost of significantly higher network bandwidth usage.

• vlselect requests compressed data from vlstorage nodes in order to reduce network bandwidth usage at the cost of slightly higher CPU usage at vlselect and vlstorage nodes. The compression can be disabled by passing -select.disableCompression command-line flag to vlselect. This reduces CPU usage at vlselect and vlstorage nodes at the cost of significantly higher network bandwidth usage.

Advanced usage #

Cluster components of VictoriaLogs provide various settings, which can be configured via command-line flags if needed. Default values for all the command-line flags work great in most cases, so it isn’t recommended tuning them without the real need. See the list of supported command-line flags at VictoriaLogs .