Timestamp rounding wrong implementation

[Tux1]

Hi,
I was looking at the recent ENH #4314 and there is a wrong implementation of round method for datetime.

    def round(self, freq):
            """
            return a new Timestamp rounded to this resolution
            Parameters
            ----------
            freq : a freq string indicating the rouding resolution
            """
            cdef int64_t unit
            cdef object result, value

            from pandas.tseries.frequencies import to_offset
            unit = to_offset(freq).nanos
            if self.tz is not None:
                value = self.tz_localize(None).value
            else:
                value = self.value
            result = Timestamp(unit*np.floor(value/unit),unit='ns')
            if self.tz is not None:
                result = result.tz_localize(self.tz)
            return result

You are flooring the nano timestamp instead of rounding. You should replace np.floor by np.round.
Also I propose to add floor and ceil method to Timestamp in order to have the same behavior that float. (It is very usefull to get the upper/lower bound for a specific freq, likely more usefull than round)

Do you agree @jreback ?

[jreback]

hmm, though looks right. Will need some specific tests that catch this.

I would add addtl functions .ceil & .floor (with the same signature). You can of course re-use the impl completely. (just subst the functions). (maybe move the impl to a private ._round which takes the rounding function).

This is used in DatetimeIndex/TimedeltaIndex, Timestamp, Timedelta (the scalars have slightly diff impl).

Just add the this issue number onto the existing note in whatsnew.

[Tux1]

Yes, I will try to do it

[jreback]

@Tux1 gr8!

also need to add for .dt as well (this is trivial, just add to the accessors list, and test in test_series)

[Tux1]

yeah, I have looked at your previous commit on that