df_trans['DELTA'].dt.days should work, in both pandas in xarray.

Pardon me for extending this discussion.

I encountered the same problem when calculating timedelta in a dataframe. It even ended with an error when I tried to call the days attribute. I am using Numpy 1.6.1

AttributeError: 'Series' object has no attribute 'days'

Problem

df_trans['DELTA'] = df_trans['DATE2'] - df_trans['DATE1']

print df_trans['DELTA'].dtype

timedelta64[ns]

print df_trans['DELTA']

0 8 days, 00:00:00 1 15 days, 00:00:00 2 5 days, 00:00:00

df_trans['DELTA'] = df_trans['DELTA'].astype('timedelta64[D]') print df_trans['DELTA'].dtype

Name: DELTA, dtype: timedelta64[D]

print df_trans['DELTA']

0 8 days, 00:00:00 1 15 days, 00:00:00 2 5 days, 00:00:00 Nothing changed at all

print df_trans['DELTA'].days

AttributeError: 'Series' object has no attribute 'days'

I get rid of the problem by putting it in to a list for the conversion.

            Ss_timedelta = df_trans['DATE2'] - df_trans['DATE1']
            ls_timedelta = Ss_timedelta.values.astype('timedelta64[D]').tolist()
            for i in range(0, len(ls_timedelta)):
                    ls_timedelta[i] = ls_timedelta[i].days / 1000                        
            df_trans['HOLDDAYS'] = pd.Series(ls_timedelta)

The pandas docs do seem to say that conversion to timedelta64[D] (or other frequencies) is possible - see: http://pandas.pydata.org/pandas-docs/stable/timedeltas.html#frequency-conversion

Also here's a more realistic example of why this is problematic for me - I have a sequence of dates and I want to calculate the difference between them in days: possible in pandas, but not possible in xarray without first reverting to pandas/numpy types

``` dates = pandas.Series([datetime.date(2016, 01, 10), datetime.date(2016, 01, 20), datetime.date(2016, 01, 25)]).astype('datetime64[ns]')

dates.diff().astype('timedelta64[D]').astype(float)

returns

0 NaN

1 10.0

2 5.0

dtype: float6

xarray.DataArray(dates).diff(dim = 'dim_0').astype('timedelta64[D]').astype(float)

returns

<xarray.DataArray (dim_0: 2)>

array([ 8.64000000e+14, 4.32000000e+14])

Coordinates:

* dim_0 (dim_0) int64 1 2

```

Again the xarray result is in ns rather than days.

Thanks

Interesting. Pandas always uses nanosecond precision for TimedeltaIndex but not Series: ``` In [13]: s = pandas.Series([1,2,3,4]).astype('timedelta64[D]')

In [14]: s Out[14]: 0 1 days 1 2 days 2 3 days 3 4 days dtype: timedelta64[D]

In [16]: pandas.Index(s) Out[16]: TimedeltaIndex(['1 days', '2 days', '3 days', '4 days'], dtype='timedelta64[ns]', freq=None) ```

This might actually be a pandas bug -- as far as I recall, this goes against the documented behavior.

The conversion to timedelta64[ns] is done on this line of code: https://github.com/pydata/xarray/blob/d66f673ab25fe0fc0483bd5d67479fc94a14e46d/xarray/core/variable.py#L169

Is there a reason behind the conversion, or could it be removed?