What is feature drift vs label drift in production ML systems?

Feature drift refers to changes in input data distribution, while label drift refers to changes in target distribution over time.

Feature drift occurs when statistical properties of input variables change, such as user behavior shifts. Label drift occurs when the distribution of outcomes changes, such as increased fraud rates or seasonal demand changes. Label drift is harder to detect because labels are often delayed or partially available. Both require different monitoring strategies and retraining triggers.