COMPAT: Floating point fractional hours for timedeltas gives unexpected precision issue

[ahjulstad]

Creating a timedelta with fractional hours gives an unexpected large round-off error.

>>> to_timedelta(1.0/3, unit='h')
Timedelta('0 days 00:19:59.998800')

This is indicates a precision around 2x10^-7 ( 1/(1000_60_60) ), which is less than a single precision float.

[ahjulstad]

Test-case code:

        # Tests with fractional hours as input:
        expected = np.array([0, 20*60*1000000000, 30*60*1000000000, 40*60*1000000000], dtype='timedelta64[ns]')
        result = to_timedelta([0., 1.0/3, 0.5, 2.0/3], unit='h', box=False)
        tm.assert_almost_equal(expected, result)

[jreback]

Here is where to change: https://github.com/pydata/pandas/blob/master/pandas/tslib.pyx#L2448

prob need a few tests for this though (and this also affects datetimes as well, e.g. you can use to_datetime in a similar manner). can prob change for h/m I suspect.

[ahjulstad]

I have had a look at it, but for some reason am struggling with recompiling.
I am using anaconda, with a conda environment for pandas, and pandas installed with
pip install -e ., but python just hangs on import pandas.
Pointers to detailed instructions appreciated.

[jreback]

don't use pip.

This gets you all the deps
conda install pandas

This gets you the build deps
conda install cython libpython nose

To build
python setup.py build

[ahjulstad]

It builds, but I cannot load the compiled modules within the git directory without some more magic. After build_ext –inplace, import pandas hangs (eating a lot of memory). I am on Windows using VS2010 compiler. I have limited time the next days to look at this, and I have a full reinstall coming up soon, so I won’t worry too much about it now.

[jreback]

you cannot use the VS 2010 compiler as python doesn't support it (except for >= 3.3.). use the recipe I show above. which works on all pythons and doesn't depend on your environment.