Automatic Invariant Selection for Online Anomaly Detection

Invariants are stable relationships among system metrics expected to hold during normal operating conditions. The violation of such relationships can be used to detect anomalies at runtime. However, this approach does not scale to large systems, as the number of invari- ants quickly grows with the number of considered metrics. The resulting “background noise” for the invariant-based detection system hinders its effectiveness. In this paper we propose a general and automatic approach for identifying a subset of mined invariants that properly model system runtime behavior with a reduced amount of background noise. This translates into better overall performance (i.e., less false positives).