Data and Model Drift Detection

A model can look great offline, say 95% accuracy, and still fall to 80% just weeks after deployment, and nothing in the infrastructure metrics will tell you why. The serving latency is fine, the error rate is zero, the feature pipeline is running on schedule. But the data the model sees in production no longer looks like the data it trained on.

Monitoring catches the symptom (accuracy dropped). Drift detection diagnoses the cause (the input distribution shifted, or the relationship between features and labels changed, or both).