Skew: … if our transformations can be tied to the model or as a standalone function, then we can just reuse the same pipelines to produce the feature values for training and serving.