BUG: np.timedelta64 arithmetic appears buggy

[jreback]

Buggy timedelta arithmetic with np.timedelta64 in np 1.7.1

In [41]: s = Series([Timestamp('20130301'),Timestamp('20130228 23:00:00'),Timestamp('20130228 22:00:00'),Timestamp('20130228 21:00:00')])

In [42]: s
Out[42]: 
0   2013-03-01 00:00:00
1   2013-02-28 23:00:00
2   2013-02-28 22:00:00
3   2013-02-28 21:00:00
dtype: datetime64[ns]

In [43]: s + timedelta(hours=1)
Out[43]: 
0   2013-03-01 01:00:00
1   2013-03-01 00:00:00
2   2013-02-28 23:00:00
3   2013-02-28 22:00:00
dtype: datetime64[ns]

Should be the same as above

In [45]: s + np.timedelta64(1,'h')
Out[45]: 
0   2013-03-01 00:00:00.000000001
1   2013-02-28 23:00:00.000000001
2   2013-02-28 22:00:00.000000001
3   2013-02-28 21:00:00.000000001
dtype: datetime64[ns]

[cpcloud]

does pandas handle timedelta and timedelta64 differently?

In [12]: (s + timedelta(hours=1)).values
Out[12]:
array(['2013-02-28T20:00:00.000000000-0500',
       '2013-02-28T19:00:00.000000000-0500',
       '2013-02-28T18:00:00.000000000-0500',
       '2013-02-28T17:00:00.000000000-0500'], dtype='datetime64[ns]')

In [13]: (s + timedelta64(1, 'h')).values
Out[13]:
array(['2013-02-28T19:00:00.000000001-0500',
       '2013-02-28T18:00:00.000000001-0500',
       '2013-02-28T17:00:00.000000001-0500',
       '2013-02-28T16:00:00.000000001-0500'], dtype='datetime64[ns]')

In [14]: (s.values + timedelta64(1, 'h'))
Out[14]:
array(['2013-02-28T20:00:00.000000000-0500',
       '2013-02-28T19:00:00.000000000-0500',
       '2013-02-28T18:00:00.000000000-0500',
       '2013-02-28T17:00:00.000000000-0500'], dtype='datetime64[ns]')

[cpcloud]

things are not being converted to ns. i got this...

[jreback]

@cpcloud #4534 fixes the compat with 1.6 in a different way that I originally suggest to you

seem that
s.astype('O') + timedelta(...) when s is a timedelta64[ns] series works in 1.6 (and then is converted back to timedelta64[ns] when created

can you you rebase off of this and see if you tests work out? (also I think your compat checking (the _is_numpy_less_17) stuff.. is a bit cleaner than mine)....