Hmmm, maybe we can use type(dtype)._is_numeric and then set NPY_DT_NUMERIC flag over there that is a bit private, and it would be nice to expose that as something like isinstance(dtype, numpy.dtypes.NumericDType) or so, but that is fine.

That might be a bit more robust (i.e. should work even if you use a random character) -- although, I am not sure if you need something around isnan here too...

Hmmm, maybe we can use type(dtype)._is_numeric and then set NPY_DT_NUMERIC flag over there that is a bit private

@hawkinsp FWIW I added NPY_DT_NUMERIC for uses like this. I also agree with sebastian it would be really nice to have a better story for registering dtypes in an ontology or hierarchy that's centrally managed by NumPy.

This is very load-bearing code in the ecosystem so I'm a little nervous about changing behavior for legacy dtypes. But we can always add a check for an _is_numeric attribute, since I don't think this is performance-sensitive code, at least not at the level of one getattr call.

Thanks, let's try with this, since Nathan also thought it should be a good solution.