Fix an overflow bug in decode_cf_datetime

[fmaussion]

• Closes Unexpected decoded time in xarray >= 0.10.1 #2002
• Tests added
• Tests passed
• Fully documented, including whats-new.rst for all changes and api.rst for new API

Not sure yet if this is the best way to do this.

[fmaussion]

The problem comes from a weird numpy casting behavior. Consider:

In [1]: import numpy as np

In [2]: (np.int32(2) * np.int64(2)).dtype
Out[2]: dtype('int64')

In [3]: (np.array(np.int32(2)) * np.int64(2)).dtype
Out[3]: dtype('int64')

In [4]: (np.array([np.int32(2)]) * np.int64(2)).dtype
Out[4]: dtype('int32')

In the offending line:

flat_num_dates *  _NS_PER_TIME_DELTA[delta]

flat_num_dates is an array of dtype int32 and _NS_PER_TIME_DELTA[delta] is a scalar of dtype int64. Is there a way to make sure this line casts to a secure dtype better than my proposed solution to cast everything to a float64 first?

[fujiisoup]

Thanks @fmaussion .

Is there a way to make sure this line casts to a secure dtype better than my proposed solution to cast everything to a float64 first?

How about np.array([np.int32(2)]) * np.array([np.int64(2)])?

I am not sure whether numpy explains this casting behavior, but I guess this is intentional.
This is consistent with

In [11]: (np.array([1.0], dtype=np.float32) * 1.0).dtype
Out[11]: dtype('float32')

where we expect the array dtype takes precedence.

[fujiisoup]

I am curious why this test is failing in appveyor.
The casting rule of numpy should not depend on the operating system.

[fmaussion]

I am curious why this test is failing in appveyor.
The casting rule of numpy should not depend on the operating system.

I'm curious too, but I don't have a windows machine to test it. The tests are green with the float64 casting though