Train-Serve Skew

Train-serve skew refers to the mismatch between the data structure a model sees during training and the data it receives in production. If feature computations differ between offline training and live serving, model behavior can degrade significantly. This issue often arises when data science and engineering workflows evolve separately. Shared feature definitions, feature stores, and serving-aligned preprocessing pipelines help reduce the risk. A good model is not only well trained, but also consistently served.