How do I manage and troubleshoot CloudWatch anomaly detection bands?

Resolution

Note: CloudWatch anomaly detection doesn't support alarms across AWS accounts or AWS Regions.

To manage or troubleshoot anomaly detection bands in CloudWatch, see the following scenarios.

Spikes reflected in your prediction band

If spikes are reflected in your prediction band, then check the metrics behavior for the previous 14 days:

• If you see spikes in your metrics, then the anomaly detection model reflects the surge when it trains the model.
• If you don't see spikes in your metrics, then check the metrics for an extended period of time. If the later model and earlier model don't differ significantly, then anomaly detection can keep the earlier model.

To resolve this issue, retrain your model to prevent spikes in your prediction band.

Sensitive or false alarms

If a sudden metric pattern changes, then it can take up to several hours for the anomaly detection bands to adapt to the new level. Because anomalies are still reported as the detection bands are adapting, false alarms can occur. If you have sensitive alarms, then use one of the following methods to adjust the anomaly detection bands:

• Modify the sensitivity to change the band's width, and don't retrain the model.
• Delete and recreate a new anomaly detector to manually retrain the model.

Alarms that don't activate when data points in the Region of a missing band are exceeded

CloudWatch alarms treat the absence of anomaly detection bands as missing data. If you configure an alarm to treat missing data as good or not exceeding, then the alarm maintains its state until the band is accessible.

Related information

New - Amazon CloudWatch anomaly detection

Operationalizing CloudWatch anomaly detection

How to set up CloudWatch anomaly detection to set dynamic alarms, automate actions, and drive online sales