Safety and MLOps
Model monitoring
Model monitoring measures deployed model inputs, outputs, quality, latency, and failures over time.
How it works
Operational metrics show whether the service works, while statistical and outcome metrics show whether the model remains useful. Alerts need owners and response procedures.
A concrete example
Illustrative: alert when p95 latency exceeds 800 ms for 10 minutes or null outputs exceed 0.5%.
Do not confuse it with Logging
logging records events; monitoring aggregates signals and triggers action.