Anomaly detection
Isolation Forest on delta features. A score threshold, plus a rule that ignores combined anomalies while CPU and RAM stay under 80%, keeps false positives down.
Last 30–500 points, retrained before every detection (every 30 s by default). The model is not stored.