Skew: we may have different pipelines for generating features for training and serving which can introduce skew through the subtle differences. Solution: create features using a unified pipeline and store them in a central location that the training and serving pipelines pull from.