Data and Model Drift Detection

Model monitoring tells you that performance changed. Drift detection helps explain whether the data, labels, or feature-label relationship changed.

A model can serve with normal latency and zero errors while its input population moves away from the training distribution. If labels are delayed, drift may be the first useful warning that the model is operating outside the conditions it was validated on.

The useful output is a decision: investigate, retrain, change thresholds, fix a pipeline, or roll back.