Comment on page
Autonomous Anomaly Detector
StackState Self-hosted v5.0.x
Anomaly detection identifies abnormal behavior in your fast-changing IT environment. This helps direct the attention of IT operators to the root cause of problems or can provide an early warning. The Autonomous Anomaly Detector (AAD) requires zero configuration. It is fully autonomous in selecting both the metric streams it will apply anomaly detection to, and the appropriate machine learning algorithms to use for each metric stream.
The Autonomous Anomaly Detector (AAD) is enabled as soon as the AAD StackPack has been installed in StackState. When the AAD has been enabled, metric streams are identified and analyzed in search of any anomalous behavior based on their past. After the initial training period, detected anomalies will be reported in the following way:
- The anomaly and time period during which anomalous behaviour was detected are shown on the associated metric stream chart. The color indicates the anomaly severity.
Each identified anomaly is given a severity. This can be HIGH, MEDIUM or LOW. The severity shows how far a metric point has deviated from the expected model and the length of time for which anomalous data has been observed.
HIGH, MEDIUM and LOW severity anomalies
When a HIGH severity anomaly is detected on a metric stream, a
Metric Stream Anomalyevent is generated. Anomaly events are listed on the Events Perspective and will also be reported as one of the Probable Causes for any associated problem. Select an event to display detailed information about it in the right panel Selection details tab.
Metric stream anomaly event detailed information
- Metric Stream - The name of the metric stream on which the anomaly was detected.
- Metric chart - A chart with an extract from the metric stream centered around the detected anomaly.
- Anomaly interval - The time period during which anomalous behaviour was detected. This is also shaded on the metric chart.
- Description - A description of the observed anomaly.
- Elements - The name of the element (or elements) on which the metric stream is attached
Note that feedback is not used to train the running instance of the AAD.
Models are selected by the AAD and optimized for each metric stream. The quality of the anomalies reported is determined to a large extent by how well the selected model describes the stream that it runs on. The StackState team works with representative datasets to develop new models and optimize the hyperparameters used for model selection and training the AAD.
To enable improvement of the AAD, users can add feedback to reported anomalies. This feedback can then be used by StackState to assist in the ongoing development of the AAD.
The feedback sent to StackState consists of:
- Thumbs-up, Thumbs-down votes - Each user can cast one vote per reported anomaly.
- Comments - Free-form text entered by users. Note that any comments added to an anomaly will be included in the feedback sent to StackState. Take care not to include sensitive data in comments.
- Anomaly details - The description, interval, severity (score), model information, metric query and element, stream names.
- Metric data - Data from the metric stream leading up to the anomaly.
Add feedback to an anomaly
The AAD will need to train on your data before it can begin reporting anomalies. With data collected in 1 minute buckets, the AAD requires a 2 hour training period. If historic data exists for relevant metric streams, this will also be used for training the AAD. In this case, the first results can be expected within an hour. Up to a day of data is used for training. After the initial training, the AAD will continuously refine its model and adapt to any changes in the data.
The AAD scales to large environments by autonomously prioritizing metric streams based on its knowledge of the 4T data model and the stream priority defined by users. The metric stream selection algorithm ranks metric streams based on the criteria below:
- Components in views that have the most stars by the most users are ranked highest.
- Anomaly detection will be disabled on streams if more than 20% of their time is flagged as anomalous.
You cannot directly control the stream selected, but you can steer the metric stream selection of the AAD by manipulating the above-mentioned factors.
After an initial training period, the AAD ensures that prioritized metric streams are checked for anomalies in a timely fashion. Anomalies occurring in the highest prioritized metric streams are detected within about 5 minutes.
To uninstall the AAD StackPack, simply press the UNINSTALL button. No other actions need to be taken.
Autonomous Anomaly Detector StackPack v0.9.2 (02-04-2021)
- Common version bumped from 2.4.3 to 3.0.0
- StackState min version bumped to 4.3.0