Regression in ma.median handling of nan values, and calculates incorrect value

The handling of nan values in masked arrays passed in to ma.median changed some time between 1.10.2 and 1.11.2, and the value returned in 1.11.2 is incorrect. In 1.10.2 nan was returned in this case. In 1.11.2 an (incorrect) numerical result is returned.

This example demonstrates both problems:

import numpy as np
dm = np.ma.array([1, np.nan, 3])
np.ma.median(dm)

In numpy 1.10.2 the output is as expected: nan

In numpy 1.11.2 the output is incorrect (not nan and not the correct numerical value ignoring the nan):

masked_array(data = 3.0,
             mask = False,
       fill_value = 1e+20)

Based on test failures we are seeing in https://github.com/astropy/astropy/pull/5232, the issue continues in 1.12.0b1.

[charris]

I’m going to do a 11.3 for the file append issue, so If you fix it up for 1.12 I might put it in 1.11.3 also.

[juliantaylor]

I think it would be good to squeeze it into 1.12, I’ll have a look today.