API/ENH: support datetime.time in a Timedelta

[jreback]

So this as expected. Though I its pretty straightforward to add support for this directly to a Timedelta, after the notion is the same, though a time object can support a time-zone so not sure if this is odd (or maybe we should add an optional tz to Timedelta).

In [5]: from datetime import time

In [6]: time(secIn [8]: pd.Timedelta(time(second=1))
ValueError: Value must be Timedelta, string, integer, float, timedelta or convertible

In [9]: pd.Timedelta('1s')          
Out[9]: Timedelta('0 days 00:00:01')

[jreback]

cc @rockg

[shoyer]

Hmm... not so sure this is desirable. This would be assuming that midnight is always zero? And that times are always positive? I think it would probably be better to require users to do this conversion explicitly.

side note: adding time-zones to Timedelta sounds awful!

[jreback]

@shoyer hah! just putting it out there if anyone sees a need for this. I added a tests to test that this is not supported, otherwise leaving for now.

[jorisvandenbossche]

What is the best way to convert a column of times to a timedelta? As

pd.to_timedelta(hours=df['time'].dt.hour, minutes=df['time'].dt.minute, seconds=df['time'].dt.second)

or

df['time'].apply(lambda x: datetime.timedelta(hours=x.hour, minutes=x.minute, seconds=x.second))

seems not very concise

[jreback]

actually this is a dupe of #8343

[jreback]

as I said, I think this type is useless. Those solns's are fine for the very limited amount this comes up. An in-between soln is to allow conversion directly to Timedelta which is not that hard.

[jorisvandenbossche]

@jreback I think it are two different things, how I understand it.

In this issue, it is proposed that you can provide a datetime.time object to pd.to_timedelta to convert it to a timedelta (just as you can also provide strings, ints, ..). But so, always something a user explicitly does.
#8343 is about integrating datetime.time in the pandas internals as a timedelta (which will lead to automatic conversion of datetime.time to timedeltas), which is a much more profound question (and not something I am in favor of)