pd.to_timedelta(single_string) returns a Series, which ruins broadcasting.

[danielballan]

Currently

In [10]: pd.to_timedelta(Series(['00:03:37', '00:05:05'])) - pd.to_timedelta('00:03:00')
Out[10]: 
0   00:00:37
1        NaT
dtype: timedelta64[ns]

because

In [13]: pd.to_timedelta('00:03:00')
Out[13]: 
0   00:03:00
dtype: timedelta64[ns]

More expected and consistent behavior, in my opinion, would be

In [12]: pd.to_timedelta(Series(['00:03:37', '00:05:05'])) - pd.to_timedelta('00:03:00')
Out[12]: 
0   00:00:37
1   00:02:05
dtype: timedelta64[ns]

If, instead, pd.to_timedelta('00:03:00') gave numpy.timedelta64(180000000000,'ns'), it would broadcast properly, as demonstrated by this work-around

In [12]: pd.to_timedelta(Series(['00:03:37', '00:05:05'])) - pd.to_timedelta('00:03:00')[0]
Out[12]: 
0   00:00:37
1   00:02:05
dtype: timedelta64[ns]

I have not regularly used timedeltas until now. Am I misjudging the expected usage?

Update, OK, one can also get the expected result by setting box=False. But box=False should be default for a scalar, no?

[jreback]

this is correct (and on purpose); we don't have a timedelta scalar ATM (like Timestamp), so only option is to return a timedelta64, which works but is not friendly

[danielballan]

Understood.

[jreback]

this particular usage is easy to fix though (this is a_bug_ I guess in the timedelta subtraction) operation....marking for 0.14

[danielballan]

So, until Timedelta scalars come along, worth submitting a PR that returns timedelta64 when input is not list-like?

[jreback]

or...could just have a single element pd.to_timedelta(...) return the underlying value...

[danielballan]

Yeah, I think we are saying the same thing. I'll make a PR.

[jreback]

you can try it and see what breaks...it should work (I think)....and I think that is right actually...(to return a single-element as a timedelta64, bascially just don't box it), to_datetime DOES return a Timestamp for a single element