See: https://github.com/pandas-dev/pandas/issues/21137 The question is: how exactly do we replicate the functionality?