that's x[:, None].ndim

I would rather check the shape as that is what needs to be correct and do not want to risk ndim also being out of sync.

If you think than .ndim may be returning something wrong then .shape could also be wrong, no?

In the pandas issue they pointed out that this changes was due to an optimization (see pandas-dev/pandas#27775 (comment)) on Index and the current implementation of ndim is:

@property
def ndim(self):
    """
    Number of dimensions of the underlying data, by definition 1.
    """
    return 1

https://github.com/pandas-dev/pandas/blob/8b6942f2876264b745529a3c3a7cf410ee096c34/pandas/core/base.py#L694-L699 which confirmed my suspicioun that shape and ndim were out of sync.

However, digging through this convinced me that we I should check that x[:, None].ndim >= 2 as this will fail on older version of pandas and stop relying on an unintentional quirk of Index classes almost ducktyping as arrays.

you don't want to do a np.asarray here first? That should make it work with older pandas versions, but also robust to potential future changes (eg if pandas starts raising for that)

we do not want to asarray things we don't have to (in particular because in some cases that can strip units off of things). So long as future versions raises one of the two exceptions caught below it will "just work" and as soon as you decide what exception you are going to raise we can add it.

The point of this is to check if we need to asarray (which in done implicitly in atleast_1d).

OK, I see. I suppose if we raise an error it will be an IndexError, so that would already be covered then