BUG: timedelta64 dtype inference in Series

[cancan101]

related #5690

This works:

pd.Series(np.array([np.timedelta64(300000000)]))

0   00:00:00.300000
dtype: timedelta64

but this does not:

pd.Series([np.timedelta64(300000000)])

on Numpy 1.7:

TypeError: don't know how to convert scalar number to float

and on Numpy 1.8:

0    300000000
dtype: int64

[jreback]

on master/numpy 1.7

inference could be improvied in the non-dtype specified case

In [23]: pd.Series([np.timedelta64(300000000)])
Out[23]: 
0    300000000 generic time units
dtype: object

In [24]: pd.Series([np.timedelta64(300000000)],dtype='m8[ns]')
Out[24]: 
0   00:00:00.300000
dtype: timedelta64[ns]

In [25]: pd.to_timedelta([np.timedelta64(300000000)])
Out[25]: 
0   00:00:00.300000
dtype: timedelta64[ns]

[cancan101]

I am not sure if this is a related issue or not, but the notion of nullness is not handled well for timedelta64 columns when performing a left join:

In [194]:
pd.DataFrame(pd.Series([np.timedelta64(300000000),np.timedelta64(300000000)],dtype='m8[ns]',index=["A","B"])).join(
     pd.DataFrame(pd.Series([np.timedelta64(300000000)],dtype='m8[ns]',index=["A"])),rsuffix='r', how="left").info()

Out [194]:
<class 'pandas.core.frame.DataFrame'>
Index: 2 entries, A to B
Data columns (total 2 columns):
0     2  non-null values
0r    1  non-null values
dtypes: float64(1), timedelta64[ns](1)

The column with a mix of timedelta64 and nulls gets cast to a float64.

This seems incorrect since NaT should be usable to indicate the null:

In [196]:
pd.Series([np.timedelta64(300000000), pd.NaT],dtype='m8[ns]')

Out[196]:
0   00:00:00.300000
1               NaT
dtype: timedelta64[ns]

[jreback]

that's unrelated I think, so maybe open a different issue for it