
You've shipped a model that looked stable in offline testing, but production performance has started to move. The team wants to know whether the issue is a change in the input data, a change in the target relationship, or both.
How do you detect data drift and concept drift in production?