When the statistical properties of input data change over time compared to the training data, potentially degrading model performance. Like a recipe not working well when ingredient quality changes.
A shopping recommendation model experiences data drift when customer behavior shifts during holidays, requiring monitoring and potential retraining.
All four clouds provide managed ML monitoring features that can detect changes in incoming data compared to a baseline (often training data), alert on drift, and support retraining workflows. Names and setup differ, but the goal is the same: catch data distribution shifts before they degrade model quality.