Training-Serving Skew

A performance degradation problem in machine learning systems caused by inconsistency between feature definitions in offline training and online serving. It is a common hidden failure source in real-time risk control systems.