Davis Anomaly Detector for "slow" metrics

[r_weber]

The old metric events had the same limitation as seems to have the davis anomaly detector:
I. have a "slow" metric/timeseries, where a metric data point is produced only every hour. In a normal Notebook/query I can easily apply the Anomaly Detection (Auto) with whatever window or datapoints I want, and it works fine.

However the Anomaly Detectors expect a metric resolution of 1 minute:

IMO for the anomlay detection it doesn't matter if the datapoints are far apart or not, it is just the amount of datapoints that is relevant, right? So for a "slow" metric I'd like to analyze a longer timeframe of 72h (=72 datapoints) and not just e.g 1h with - in a one minute requirement - 60 datapoints. 

[JeanBlanc]

Hi @r_weber 

Yes, that’s correct — this is indeed a limitation of the Davis Anomaly Detector.

Just like with the old metric events, the detector currently expects a 1-minute metric resolution and only analyzes data within the defined window size (up to 1 hour maximum).

So, for “slow” metrics with hourly data points, the Davis Anomaly Detector won’t perform well, as it doesn’t adapt its analysis window to the lower data frequency.

Best regards,

Jean