@kernc, what do you prefer:

• (isinstance(rows, (int, type(Ellipsis))) or len(rows)),
• (not isinstance(rows, (list, np.ndarray)) or len(rows)), or
• (not isinstance(rows, collections.Sized) or len(rows))?

If the method is called with the argument types specified in the docstring, all are equivalent. Also, in cases where they are not equivalent (e.g. if rows is something stupid), the next line will fail anyway.

I already see:

>>> arr = np.ones(5)
>>> empty_list = []
>>> arr[empty_list]
array([], dtype=float64)

I.e. empty input, empty output.

You don't?

I would prefer:

if self.__sortInd is not None:
    if rows is not Ellipsis:
        rows = np.asarray(rows, dtype=int)  # dangerous casting

    new_rows = self.__sortInd[rows]
    ...

But this also relies on indexing (at least) with an empty array working.

I am not aware of this ever not being the case.

I would prefer your first option. Note to also fix .mapFromSourceRows() below.

dtype must not be float; the caller must ensure that. This is similar to __getitem__ which expects ints, and doesn't cast float to int. I made this clear in the docstring.

I fixed mapFromSourceRows().