Predictive Detection learns the normal shape of every metric, log volume, and trace pattern in your system. When the shape starts to drift toward a previously-seen incident pattern, the model fires an early-warning signal. Median lead time across customers: 4 hours and 12 minutes before the symptoms hit your alerts.
For each service, Nova learns a per-window baseline of latency, error rate, log volume, and queue depth. A drift score is computed every minute against the baseline. Separately, Nova carries a library of 120+ failure-mode signatures (memory leak, connection-pool saturation, cache stampede, GC death spiral) and matches the live shape against them. A prediction fires when both layers agree.
Predictions ship with a confidence score (0-100). You set the threshold for "page on-call" vs "post in slack" vs "log only". Default thresholds favor lead time: Slack at 60, page at 80. Tighten them if your team finds the early-warning signal noisy. The threshold is per-service, so tier-0 services can be more sensitive than tier-2.
Models drift. Yours did, mine does, ours does. Nova monitors prediction accuracy on a 7-day rolling window. When precision falls below your threshold, the model is flagged for retraining. The previous version stays active during the retrain so coverage never drops. You see the drift before customers do.
Open the page. The top panel shows live predictions across all your services, ranked by confidence. Click any prediction to see the drift chart, the matching signature, the suspect changes from Nova Rewind, and the suggested runbook. Approve "auto-remediate" on a prediction to let the agent fleet act before the page fires.
Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.
Predictive Detection turns operations from reactive to proactive. Most of the time, you fix the problem before anyone has a reason to file a ticket.