BLD: pandas-wheels are failing

[jreback]

https://travis-ci.org/MacPython/pandas-wheels/builds/406975380

looks like #21879 somehow broke this. The error messages I think are red herrings for somehow picking up a wrong version of numpy when things are compiled.

[jreback]

@pandas-dev/pandas-core

[TomAugspurger]

Will take a look today.

[TomAugspurger]

Side-note, does anyone know why we use numpy 1.9 for the 3.6 builds as well? Should that be using NumPy 1.11?

[TomAugspurger]

The issue seems to be pandas_dtype calling np.dtype with certain arguments.

In [1]: import numpy as np

In [2]: import warnings

In [3]: warnings.simplefilter("always", DeprecationWarning)

Some things are fine,

In [4]: np.dtype(('b', 'c'))
Out[4]: dtype('int8')

others are not

In [5]: np.dtype(('b1', 'b2'))
/Users/taugspurger/miniconda3/envs/pandas-debug/bin/ipython:1: DeprecationWarning: Specified size is invalid for this data type.
Size will be ignored in NumPy 1.7 but may throw an exception in future versions.
  #!/Users/taugspurger/miniconda3/envs/pandas-debug/bin/python
Out[5]: dtype('bool')

We could quickly work around this by just calling doing this inside a warnings filter, but it hints at a deeper problem: we shouldn't be passing data to np.dtype in the first place. E.g. this causes the warning: pd.Index(('b1', 'b2')), which is silly.

This is coming through is_extension_array_dtype (mea culpa).

I'll spend some time today and tomorrow trying to clean this up.

[TomAugspurger]

I'm no longer working on this, if someone else wants to.