Messing around with data in preparation for ML can be very time consuming, and avoiding introducing slowdowns is most useful.