Training-serving skew
Also written: skew
Data prepared differently for training and for predictions, such as filling missing ratings with the median in one and 0 in the other. Nothing crashes, but scores are slightly wrong; one shared `prepare` function prevents it.