Hey all,
We have been testing and using Dynatrace Davis Anomaly Detection for some weeks now with the goal to automate alerting for our various timeseries data. For detecting changes in slope minute-minute (or just period-period), Davis has proved exceptional, and specifically the Seasonal Baseline Anomaly Detection Model (barring some slight issues with very-high-cardinality data), as well as the Auto-Adaptive Threshold Anomaly Detection for checking for exceptional values in the overall historical timeseries data. That being said, there is one thing we seemingly cannot do, we wish to compare our data against historical data and issue alerts based on deviation from repetitive behavior, something the Seasonal Baseline Anomaly Detection Model seems to do rather poorly overall. For example, the issue that prompted this post, we measure throughputs in Messages Per Second (MPS) as timeseries data, with queries such as "timeseries MPS = avg(MPS_METRIC), by: buckets, interval: 1m". This data traces a daily sinusoidal pattern, as expected. Over the previous blizzard on Sunday in the East Coast, we measured tremendously lower peak MPS during midday than for typical Sundays, as expected from a blizzard, and would want to issue an alert for that after some time. Using the Seasonal Baseline Anomaly Detection Model, the dynamic thresholds for that Sunday peak are lower than they would be, presumably attempting to follow the timeseries data for the anomalous Sunday rather than repeating the historical trend for Sundays in general, such that altering values such as the tolerance or sampling rate can coerce alerts for that Sunday but at the expense of also incurring alerts elsewhere. Is there any way to avoid this problem and to have Davis produce extrapolations and thresholds based on more long-term historical data? For that matter, how would this behavior be altered with increasing cardinalities, where Davis may overflow on training data over longer timeframes? I have attached an image of simulated alerts for reference.
Thank you.
