Timestamp .round precision for ns

[jreback]

xref https://github.com/pandas-dev/pandas/pull/15568/files#r104304074

In [7]: pd.DatetimeIndex(['2016-10-17 12:00:00.0015']).round('ns')
Out[7]: DatetimeIndex(['2016-10-17 12:00:00.001499904'], dtype='datetime64[ns]', freq=None)

This is the unfortunate problem of converting from int64->float64->int64. I am not sure there is an easy way to fix this. We could:

• ignore (maybe doc it)
• raise/warn

[jreback]

so actually this is 'easy'; I think for values of round < 1us (e.g. ns values), you can simply do something like this.

simulating .round('10ns')

In [48]: i
Out[48]: DatetimeIndex(['2016-10-17 12:00:00.001501031'], dtype='datetime64[ns]', freq=None)

In [49]: DatetimeIndex(1000000 * (i.asi8 // 1000000) + 10*(np.round((i.asi8 % 1000000)/float(10))).astype('i8'))
Out[49]: DatetimeIndex(['2016-10-17 12:00:00.001501030'], dtype='datetime64[ns]', freq=None)

IOW, we avoid the float precision problem by only working with the last 6 digits, then adding back the original portion.

cc @mroeschke