CI: nighlty numpy broke ci

[phofl]

It looks like the new numpy nightly wheel broke our ci.

https://github.com/pandas-dev/pandas/runs/6919550144?check_suite_focus=true First failing build today

[phofl]

Probably due to numpy/numpy#21437

[mroeschke]

I opened up numpy/numpy#21784 for an example in our CI where this is failing cc @seberg

[seberg]

Sorry, I guess I did not check pandas. There are two changes:

1. arr.fill() is a bit different, this causes ValueError when assigning NaN into an integer array.
2. There will be a lot of warnings about invalid casts, there are two categories (probably):
   • Casting nan or inf (or out of bounds values) to integer will cause a RuntimeWarning("invalid value...") (this may not always be the case for out of bounds values).
   • Casting for example 2**300 to a float32, will cause a RuntimeWarning("overflow ...")

My hope is that the second one is a bit of a chore, but requires only a few tests to be updated or np.errstate(invalid="ignore") calls to be added.
The first (arr.fill change)... While intentional, maybe we have to work around it for pandas?

[mroeschke]

At least in the example I posted in numpy/numpy#21784, this was being hit in a code block that also has

pandas/pandas/core/dtypes/cast.py

Lines 1699 to 1701 in 89578fe

	if length and is_integer_dtype(dtype) and isna(value):
	# coerce if we have nan for an integer dtype
	dtype = np.dtype("float64")

So not sure why length is a necessary condition but I can try dropping it to see what happens to work around the arr.fill change

[seberg]

Yeah, that condition looks strange, why fill an empty array? But it does looks like logic to decide whether the result should be int, or needs to be float64 (to hold NaNs)?
For example if pandas is to fill an empty integer array with NaN/NA it can maybe reasonably decide to avoid upcasting to float64.