BUG: Series.median() treats NaT as INT_MIN

[ischwabacher]

Fixing this will be easier once numpy/numpy#5222 is fixed.

In [1]: import pandas as pd

In [2]: pd.Series([0, pd.tslib.iNaT], dtype='m8[ns]')
Out[2]: 
0   0 days
1      NaT
dtype: timedelta64[ns]

In [3]: _.median()
Out[3]: 
0   -53375 days, 23:53:38.427388
dtype: timedelta64[ns]

[jreback]

nothing to do with numpy, this is an error here: https://github.com/pydata/pandas/blob/master/pandas/core/nanops.py#L281(just take out the mask and it should work), the mask is getting overwritten

[ischwabacher]

Yeah, it's not the fix that's the problem (though you do have to carefully propagate the reshaping of the input to the mask; you can't just close it into get_median and expect things to work). The reason I care about the upstream problems is that when I added a bunch of all- and partial NaT arrays to test_nanops.py and extended check_funs to use them appropriately, I got a pile of tests that failed because the numpy ufuncs we're testing against mishandle NaT.

[jreback]

@ischwabacher I fixed all of this for 0.15.0 (aside from this one), which I guess was just not completely tested. You can fix upstream if you want but tests don't rely on numpy very much for these tests.

[ischwabacher]

It looks like max and min with skipna=False are still treating NaT as INT_MIN as well, and argmax and argmin are returning nan. Numpy behavior with floats (which appears to be intended) is to return the index of the first nan, so that arr[argmax(arr)].equals(max(arr)). Or would if np.ndarray had an equals method.

In [81]: pd.Series([0, pd.tslib.iNaT], dtype='M8[ns]').max(skipna=False)
Out[81]: Timestamp('1970-01-01 00:00:00')

In [82]: pd.Series([0, pd.tslib.iNaT], dtype='m8[ns]').max(skipna=False)
Out[82]: Timedelta('0 days 00:00:00')

In [83]: pd.Series([0, pd.tslib.iNaT], dtype='M8[ns]').min(skipna=False)
Out[83]: NaT

In [84]: pd.Series([0, pd.tslib.iNaT], dtype='m8[ns]').min(skipna=False)
Out[84]: NaT

[jreback]

for these dtypes skipna=False is not well defined

by definition you are not using a mask so the values are the values
the above look right 2 me (the min/max) for skipna=False

[ischwabacher]

Really? Why aren't we treating NaT the same as NaN in that case?

Also, if NDFrame behaves differently from ndarray it makes things harder to test. But I'll look at the tests for floats to see how they're handled, since currently pandas treats NaN differently from the way numpy treats it.

[jreback]

so the min/max for skipna=False are wrong above (81,82).

pandas is NOT numpy. so treatment can be different, though not w/o a good reason. (like numpy doesn't handle anything NaT related correctly). usually pretty good with NaN.

Don't test directly with numpy. Write a function that returns the correct result.

[ischwabacher]

Don't test directly with numpy. Write a function that returns the correct result.

And then merge it into numpy. :)

I really think that getting the NaT handling working correctly in numpy is the first step to implementing NA_integer_, since that's effectively what NaT is.

[jreback]

well you can touch numpy if you would like :).

We are going to bypass numpy to support integer NA, using libdynd, essentially a much better numpy (with the same exact API).

[ischwabacher]

Welp, I guess that's a thing.