Quick Start

For a broader overview please visit the navigation page.

How to install and run vmanomaly

To run vmanomaly, you need to have VictoriaMetrics Enterprise license. You can get a trial license key here.

The following options are available:

• To run Docker image
• To run in Kubernetes with Helm charts
• To run with VM Operator

Anomaly detection models can be kept {{% available_from "v1.13.0" anomaly %}} on host filesystem after fit stage (instead of default in-memory option); This will drastically reduce RAM for larger configurations. Similar optimization {{% available_from "v1.16.0" anomaly %}} can be applied to data read from VictoriaMetrics TSDB. See instructions of how to enable it here.

Command-line arguments

The vmanomaly service supports a set of command-line arguments to configure its behavior, including options for licensing, logging levels, and more.

vmanomaly supports {{% available_from "v1.18.5" anomaly %}} running on config directories, see the config positional arg description in help message below.

Single-dashed command-line argument {{% available_from "v1.23.3" anomaly %}} format can be used, e.g. -license.forceOffline in addition to --license.forceOffline. This aligns better with other VictoriaMetrics ecosystem components. Mixing the two styles is also supported, e.g. -license.forceOffline --loggerLevel INFO.

usage: vmanomaly.py [-h] [--license STRING | --licenseFile PATH] [--license.forceOffline] [--loggerLevel {DEBUG,WARNING,FATAL,ERROR,INFO}] [--watch] [--dryRun] [--outputSpec PATH] config [config ...]

VictoriaMetrics Anomaly Detection Service

positional arguments:
  config                YAML config file(s) or directories containing YAML files. Multiple files will recursively merge each other values so multiple configs can be combined. If a directory is provided,
                        all `.yaml` files inside will be merged, without recursion. Default: vmanomaly.yaml is expected in the current directory.

options:
  -h                    Show this help message and exit
  --license STRING      License key for VictoriaMetrics Enterprise. See https://victoriametrics.com/products/enterprise/trial/ to obtain a trial license.
  --licenseFile PATH    Path to file with license key for VictoriaMetrics Enterprise. See https://victoriametrics.com/products/enterprise/trial/ to obtain a trial license.
  --license.forceOffline
                        Whether to force offline verification for VictoriaMetrics Enterprise license key, which has been passed either via -license or via -licenseFile command-line flag. The issued
                        license key must support offline verification feature. Contact [email protected] if you need offline license verification.
  --loggerLevel {DEBUG,WARNING,FATAL,ERROR,INFO}
                        Minimum level to log. Possible values: DEBUG, INFO, WARNING, ERROR, FATAL.
  --watch               Watch config files for changes and trigger hot reloads. Watches the specified config file or directory for modifications, deletions, or additions. Upon detecting changes,
                        triggers config reload. If new config validation fails, continues with previous valid config and state.
  --dryRun              Validate only: parse + merge all YAML(s) and run schema checks, then exit. Does not require a license to run. Does not expose metrics, or launch vmanomaly service(s).
  --outputSpec PATH     Target location of .yaml output spec.

You can specify these options when running vmanomaly to fine-tune logging levels or handle licensing configurations, as per your requirements.

Licensing

The license key can be specified with the help of the following command-line arguments: --license, --licenseFile, --license.forceOffline

In order to make it easier to monitor the license expiration date, the following metrics are exposed (see Monitoring section for details on how to scrape them):

# HELP vm_license_expires_at When the license expires as a Unix timestamp in seconds
# TYPE vm_license_expires_at gauge
vm_license_expires_at 1.6963776e+09
# HELP vm_license_expires_in_seconds Amount of seconds until the license expires
# TYPE vm_license_expires_in_seconds gauge
vm_license_expires_in_seconds 4.886608e+06

Example alerts for vmalert:

groups:
  - name: vm-license
    # note the `job` label and update accordingly to your setup
    rules:
      - alert: LicenseExpiresInLessThan30Days
        expr: vm_license_expires_in_seconds < 30 * 24 * 3600
        labels:
          severity: warning
        annotations:
          summary: "{{ $labels.job }} instance {{ $labels.instance }} license expires in less than 30 days"
          description: "{{ $labels.instance }} of job {{ $labels.job }} license expires in {{ $value | humanizeDuration }}.
            Please make sure to update the license before it expires."

      - alert: LicenseExpiresInLessThan7Days
        expr: vm_license_expires_in_seconds < 7 * 24 * 3600
        labels:
          severity: critical
        annotations:
          summary: "{{ $labels.job }} instance {{ $labels.instance }} license expires in less than 7 days"
          description: "{{ $labels.instance }} of job {{ $labels.job }} license expires in {{ $value | humanizeDuration }}.
            Please make sure to update the license before it expires."

Docker

To run vmanomaly, you need to have VictoriaMetrics Enterprise license. You can get a trial license key here.

Due to the upcoming DockerHub pull limits, an additional image registry, Quay.io, has been introduced for VictoriaMetrics images, including vmanomaly. If you encounter pull rate limits, switch from:

Copy

docker pull victoriametrics/vmanomaly:vX.Y.Z

to:

Copy

docker pull quay.io/victoriametrics/vmanomaly:vX.Y.Z

Below are the steps to get vmanomaly up and running inside a Docker container:

1. Pull Docker image:

docker pull victoriametrics/vmanomaly:v1.29.3

2. Create the license file with your license key.

export LICENSE_KEY=YOUR_LICENSE_KEY
echo $LICENSE_KEY > license

3. Create and modify your config.yaml file to your liking. An example can be found here

4. Start the vmanomaly Docker container with a license file, use the command below. Make sure to replace YOUR_LICENSE_FILE_PATH, and YOUR_CONFIG_FILE_PATH with your specific details:

docker run -it \
    -v ./license:/license \
    -v ./config.yaml:/config.yaml \
    -p 8490:8490 \
    victoriametrics/vmanomaly:v1.29.3 \
    /config.yaml \
    --licenseFile=/license \
    --loggerLevel=INFO \
    --watch

Use the below configuration if settings.restore_state is True (vmanomaly runs in stateful mode) or on-disk mode is preferred over in-memory.

docker run -it \
    -v ./license:/license \
    -v ./config.yaml:/config.yaml \
    -v vmanomaly_data:/tmp/vmanomaly \
    -e VMANOMALY_DATA_DUMPS_DIR=/tmp/vmanomaly/data \
    -e VMANOMALY_MODEL_DUMPS_DIR=/tmp/vmanomaly/models \
    -p 8490:8490 \
    victoriametrics/vmanomaly:v1.29.3 \
    /config.yaml \
    --licenseFile=/license \
    --loggerLevel=INFO \
    --watch

# docker-compose.yml file
services:
  # ...
  vmanomaly:
    container_name: vmanomaly
    image: victoriametrics/vmanomaly:v1.29.3
    # ...
    restart: always
    volumes:
      - ./config.yaml:/config.yaml
      - ./license:/license
      # Enable if settings.restore_state is True
      # - vmanomaly_data:/tmp/vmanomaly
    environment:
      # Enable if on-disk mode over in-memory is preferred
      # Required, if settings.restore_state is True
      - VMANOMALY_MODEL_DUMPS_DIR=/tmp/vmanomaly/models
      - VMANOMALY_DATA_DUMPS_DIR=/tmp/vmanomaly/data
    ports:
      - "8490:8490"
    command:
      - "/config.yaml"
      - "--licenseFile=/license"
      - "--loggerLevel=INFO"
      - "--watch"

volumes:
  # ...
  # Enable if on-disk mode over in-memory is preferred
  # Required, if settings.restore_state is True
  vmanomaly_data: {}

For a complete docker-compose example please refer to our alerting guide, chapter docker-compose

See also:

• Verify the license online OR offline. See the details here.
• How to configure vmanomaly

Kubernetes with Helm charts

To run vmanomaly, you need to have VictoriaMetrics Enterprise license. You can get a trial license key here.

With the forthcoming DockerHub pull limits additional image registry was introduced (quay.io) for VictoriaMetric images, vmanomaly images in particular. If hitting pull limits, try switching your docker pull victoriametrics/vmanomaly:vX.Y.Z to docker pull quay.io/victoriametrics/vmanomaly:vX.Y.Z

You can run vmanomaly in Kubernetes environment with these Helm charts.

VM Operator

If you are using VM Operator to manage your Kubernetes cluster, vmanomaly can be deployed using the following custom resource guide.

How to configure vmanomaly

To run vmanomaly, use YAML files or directories containing YAML files. The configuration files support shallow merge, allowing splitting the configuration into multiple files for better organization.

If you are using directories, all .yaml files inside will be shallow merged, without deeper recursion. If you want to merge multiple YAML files, you can specify them as separate arguments, e.g.

shellhelpCopy

    vmanomaly config1.yaml config2.yaml ./config_dir/

Before deploying, check the correctness of your configuration validate config file(s) with --dryRun command-line flag for chosen deployment method (Docker, Kubernetes, etc.). This will parse and merge all YAML files, run schema checks, log errors and warnings (if found) and then exit without starting the service or requiring a license. {{% available_from "v1.27.0" anomaly %}} it can be also used to check for migration compatibility issues when upgrading to a newer version of vmanomaly. See Migration section for more details.

Example

Here is an example of a config file that will run the Prophet model on vm_cache_entries metric, with periodic scheduler that runs inference every minute and fits the model every day. The model will be trained on the last 2 weeks of data each time it is (re)fitted. The model will produce anomaly_score, yhat, yhat_lower, and yhat_upper series for debugging purposes. The model will be timezone-aware and will use cyclical encoding for the hour of the day and day of the week seasonality.

settings:
  # https://docs.victoriametrics.com/anomaly-detection/components/settings/
  n_workers: 2  # number of workers to run workload in parallel, set to 0 or negative number to use all available CPU cores
  anomaly_score_outside_data_range: 5.0  # default anomaly score for anomalies outside expected data range
  restore_state: true  # restore state from previous run, available since v1.24.0
  # https://docs.victoriametrics.com/anomaly-detection/components/settings/#logger-levels
  # to override service-global logger levels, use the `logger_levels` section
  logger_levels:
    # vmanomaly: INFO
    # scheduler: INFO
    # reader: INFO
    # writer: INFO
    model.prophet: WARNING

schedulers:
  1d_5m:
    # https://docs.victoriametrics.com/anomaly-detection/components/scheduler/#periodic-scheduler
    class: 'periodic'
    infer_every: '5m'
    fit_every: '1d'
    fit_window: '4w'

models:
  # https://docs.victoriametrics.com/anomaly-detection/components/models/#prophet
  prophet_model:
    class: 'prophet'
    provide_series: ['anomaly_score', 'yhat', 'yhat_lower', 'yhat_upper']  # for debugging
    tz_aware: True  # set to True if your data is timezone-aware, to deal with DST changes correctly
    tz_use_cyclical_encoding: True
    tz_seasonalities: # intra-day + intra-week seasonality
      - name: 'hod'  # intra-day seasonality, hour of the day
        fourier_order: 4  # keep it 3-8 based on intraday pattern complexity
        prior_scale: 10
      - name: 'dow'  # intra-week seasonality, time of the week
        fourier_order: 2  # keep it 2-4, as dependencies are learned separately for each weekday
    compression:  # available since v1.28.1
      window: "30m"               # downsample 5m data into 30m intervals before fitting
      agg_method: "mean"          # use mean aggregation within each window
      adjust_boundaries: true     # adjust confidence intervals after downsampling
    # inner model args (key-value pairs) accepted by
    # https://facebook.github.io/prophet/docs/quick_start#python-api
    args:
      interval_width: 0.98  # see https://facebook.github.io/prophet/docs/uncertainty_intervals

reader:
  class: 'vm'  # use VictoriaMetrics as a data source
  # https://docs.victoriametrics.com/anomaly-detection/components/reader/#vm-reader
  datasource_url: "https://play.victoriametrics.com/" # [YOUR_DATASOURCE_URL]
  tenant_id: '0:0'
  sampling_period: "5m"
  queries:
    # define your queries with MetricsQL - https://docs.victoriametrics.com/victoriametrics/metricsql/
    cpu_user:
      expr: 'sum(rate(node_cpu_seconds_total{mode=~"user"}[10m])) by (container)'
      max_datapoints_per_query: 15000  # to deal with longer queries hitting search.MaxPointsPerTimeseries
    # other queries ...

writer:
  class: 'vm'  # use VictoriaMetrics as a data destination
  # https://docs.victoriametrics.com/anomaly-detection/components/writer/#vm-writer
  datasource_url:  "http://victoriametrics:8428/" # [YOUR_DATASOURCE_URL]
  # optional tenant ID
  # tenant_id: "0:0"

UI

{{% available_from "v1.26.0" anomaly %}} vmanomaly's built-in web UI can be used for prototyping and interactive experimenting to produce vmanomaly's and vmalert's configuration files. Please refer to the UI documentation for detailed instructions and examples. {{% available_from "v1.29.0" anomaly %}} Connect MCP server to the UI to benefit from better response quality and tool access in the UI Copilot, which provides AI-assisted configuration generation and debugging capabilities. See the UI documentation for instructions on how to set it up.

[!TIP] Public playgrounds with pre-configured vmanomaly instances and VictoriaMetrics/VictoriaLogs/VictoriaTraces datasources are available for interactive experimenting without the need to set up your own instance or getting an enterprise license. You can find them in the UI documentation or access them directly via the links - metrics, logs, traces - or embedded versions in the collapsible blocks.

{{% collapse name="Playground on VictoriaMetrics Datasource" %}}

{{% /collapse %}}

{{% collapse name="Playground on VictoriaLogs Datasource" %}}

{{% /collapse %}}

{{% collapse name="Playground on VictoriaTraces Datasource" %}}

{{% /collapse %}}

Recommended steps

For optimal service behavior, consider the following tweaks when configuring vmanomaly:

• Set settings.n_workers {{% available_from "v1.23.0" anomaly %}} arg > 1 to utilize more of available CPU cores for parallel workload processing. This can significantly improve performance, especially on larger datasets with a lot of reader.queries and longer scheduler.fit_window intervals. Setting it to zero or negative number will enable using all available CPU cores.

• Set up on-disk mode {{% available_from "v1.13.0" anomaly %}} to reduce RAM usage, especially for larger datasets. This mode allows vmanomaly to keep models and the data on the host filesystem after the fit stage, rather than in memory.

• Set up state restoration {{% available_from "v1.24.0" anomaly %}} to resume from the last known state for long-term stability. This is controlled by the settings.restore_state boolean arg.

• Set up config hot-reloading {{% available_from "v1.25.0" anomaly %}} to automatically reload configurations on config files changes. This can be enabled via the --watch CLI argument and allows for configuration updates without explicit service restarts.

Schedulers:

• Configure the inference frequency in the scheduler section of the configuration file.
• Ensure that infer_every aligns with your minimum required alerting frequency.
  • For example, if receiving alerts every 15 minutes is sufficient (when anomaly_score > 1), set infer_every to match reader.sampling_period or override it per query via reader.queries.query_xxx.step for an optimal setup.

Reader:

• Setup the datasource to read data from in the reader section. Include tenant ID if using a cluster version of VictoriaMetrics (multitenant value {{% available_from "v1.16.2" anomaly %}} can be also used here).
• Define queries for input data using MetricsQL under reader.queries section. Note, it's possible to override reader-level arguments at query level for increased flexibility, e.g. specifying per-query timezone or sampling period.
• For longer fit_window intervals in scheduler, consider splitting queries into smaller time ranges to avoid excessive memory usage, timeouts and hitting server-side constraints, so they can be queried separately and reconstructed on vmanomaly side. Please refer to this example for more details.

If applicable - consider VLogsReader {{% available_from "v1.26.0" anomaly %}} to perform anomaly detection on log-derived metrics. This is particularly useful for scenarios where log data needs to be analyzed for unusual patterns or behaviors, such as error rates or request latencies.

Writer:

• Specify where and how to store anomaly detection metrics in the writer section.
• Include tenant ID if using a cluster version of VictoriaMetrics for writing the results.
• Adding for label to metric_format argument is recommended for smoother visual experience in the anomaly score dashboard. Please refer to metric_format argument description here.

Models:

• Configure built-in models hyperparameters according to your needs in the models section. Where possible, incorporate domain knowledge for optimal results.
• (Optional) Develop or integrate your custom models with vmanomaly.
• Adding y to provide_series argument values is recommended for smoother visual experience in the anomaly score dashboard. Also, other vmanomaly output series can be specified in provide_series, such as yhat, yhat_lower, yhat_upper, etc. This will allow you to visualize the expected values and their confidence intervals in the dashboard.

  Only univariate models support the generation of such output. Other models, such as multivariate or custom, may not support this feature.

Visualization:

• Set up anomaly score dashboard to visualize the results of anomaly detection.
• Set up self-monitoring dashboard to monitor the health of vmanomaly service and its components.

Logging:

• Tune logging levels in the settings.logger_levels section to control the verbosity of logs. This can help in debugging and monitoring the service behavior, as well as in disabling excessive logging for production environments.

Check also

Please refer to the following links for a deeper understanding of Anomaly Detection and vmanomaly:

• High Availability and Horizontal Scalability
• State Restoration
• Guide: Anomaly Detection and Alerting Setup
• FAQ
• Migration Guide
• CHANGELOG
• UI CHANGELOG
• Anomaly Detection Blog