Dynatrace version 1.253+
A seasonal baseline represents a dynamic approach to baselining where the systems have distinct seasonality patterns. An example of a seasonal pattern is a metric that rises during business hours and lies low outside of them. For such a case, it is difficult to set up an alerting configuration based on a static or auto-adaptive threshold. If you set a fixed threshold based on business hours' behavior, you'll miss an anomaly outside of business hours. Davis® AI learns the seasonal behavior of your metrics and automatically creates a confidence band for it.
When the configuration of a metric event includes multiple entities, each entity receives its own seasonal baseline, and each baseline is evaluated independently. For example, if the scope of an event includes five services, Dynatrace calculates and evaluates five independent baselines.
Let's check an example where a seasonal baseline has an advantage over a non-seasonal threshold. The chart below shows the measurement of the user actions of an application. As you can see, there are daily peaks at about 8 AM. When evaluated by an auto-adaptive threshold, those peaks result in false positives (red marks above the chart).
Davis AI automatically checks if the seasonal model conditions fit and creates a confidence band based on the seasonal pattern. As you can see on the chart below, the model recognizes seasonality and doesn't generate false alarms on the peaks. However, it will still trigger an alert should a spike occur at a different time.
The seasonal baseline model uses the metric values for each minute during the last 14 days as the reference for the baseline calculation. The model learns the usual behavior of the metric and aims to detect regularly repeating patterns. It uses probabilistic prediction to compute the confidence band, which covers the expected range of a future point.
You can adapt the width of the confidence band via the tolerance parameter. A higher tolerance means a broader confidence band, leading to fewer triggered events. The allowed range for the tolerance is from
4 as the default value.
As the name suggests, the seasonal model is beneficial for data with seasonal patterns (for example, daily peaks). As it consumes significant resources and needs more time to learn the behavior, there are some validation checks to ensure optimal performance. The validation evaluates each tuple (unique combinations of metric—dimension—dimension value). If there's no pass, a simple model is used instead, with the respective message appearing in the preview.
Here are some reasons to switch to a simple model:
- the data has many missing values
- the data has many outliers
- there is not enough variability in the data
- there is no seasonality detectable in the data.
The simple model calculates the confidence band based on the robust estimation of the data distribution in the input data. The calculated interval stays constant over time. For the simple model, the tolerance parameter controls the width of the confidence band in the same manner as for the seasonal model.
The model (be it seasonal or simple) is updated daily.
Another important parameter for seasonal baselines is the sliding window that is used to compare current measurements against the confidence band. It defines how often the confidence band must be violated within a sliding window of time to raise an event (violations don't have to be successive). This approach helps to avoid alerting too aggressively on single violations. You can set the sliding window to a maximum of 60 minutes.