Anomaly model testing

  • Rversion finale: Australia
  • Mis à jour 12 mars 2026
  • 1 minute de lecture

    • Use anomaly model testing to apply and evaluate anomaly detection for a small set of CIs and metrics, using actual metric data. Compare test results to expected results, then fine-tune the anomaly detection model before enabling anomaly detection for the tested CIs and metrics in the production environment.

    For example, before enabling anomaly detection for Linux Servers, you can use anomaly model testing with actual metric data collected for those Linux Servers. Existing alerts might indicate an anomaly for a specific metric and CI. You can create an anomaly test rule to test anomaly detection for those CIs and metrics, and then check if the expected anomaly is detected. You can also examine the bounds calculated by the anomaly model to see if they match expected bounds. After you fine-tune the test model, you can enable anomaly detection for Linux Servers.

    Anomaly model testing supports up to 20 time series, which can include 20 metric types for the same CI, the same metric type for 20 different CIs, or any combination of those options.

    View test results in the Insights Explorer

    Upon completion of a test run, a URL to an Insights Explorer becomes available. The anomaly model testing results appear in this trimmed-down version of the Insights Explorer. This Insights Explorer is pre-loaded with the charts and scores for the metric series specified in the anomaly test rule. You cannot add any CMDB groups or application services to this Insights Explorer.

    Each chart in the Insights Explorer, contains the Anomaly Test Statistics section in its Chart Settings. This section is specific to anomaly model testing and provides the following options:
    • Show Bounds: Displays a computed upper and lower bound for the metric in an anomaly model testing, based on learning of past metric values. The upper and lower bounds always appear together.
    • Show Anomaly Scores: Displays anomaly scores on a virtual axis of 0-10, for an anomaly model test. Color code is based on the score thresholds that are defined in the Anomaly Score to Event Severity Map [sa_metric_anomaly_score_to_event_severity_map] table.