BUG: where behaves badly when dtype of self is datetime or timedelta, and dtype of other is not

[evanpw]

There are a few weird behaviors that this fixes:

1. If other is an int or float, then where throws an Exception while trying to call other.view
2. If other is an np.int64, it works fine
3. If other is an np.float64, there is no error, but the result is bizarre (it reinterprets the bits of the float as an integer, rather than casting it)
4. If other is list-like and has a numerical dtype, then it throws an exception while try to call all on the value False (for some reason, comparing an ndarray with dtype datetime64 and an ndarray with an integer dtype just returns a scalar False rather than a boolean ndarray)

>>> s = Series(date_range('20130102', periods=2))
>>> s.where(s.isnull(), 0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/evanpw/Workspace/pandas/pandas/core/generic.py", line 3399, in where
    try_cast=try_cast)
  File "/home/evanpw/Workspace/pandas/pandas/core/internals.py", line 2469, in where
    return self.apply('where', **kwargs)
  File "/home/evanpw/Workspace/pandas/pandas/core/internals.py", line 2451, in apply
    applied = getattr(b, f)(**kwargs)
  File "/home/evanpw/Workspace/pandas/pandas/core/internals.py", line 1080, in where
    result = func(cond, values, other)
  File "/home/evanpw/Workspace/pandas/pandas/core/internals.py", line 1063, in func
    v, o = self._try_coerce_args(v, o)
  File "/home/evanpw/Workspace/pandas/pandas/core/internals.py", line 1819, in _try_coerce_args
    other = other.view('i8')
AttributeError: 'int' object has no attribute 'view'
>>> s.where(s.isnull(), np.int64(10))
0   1970-01-01 00:00:00.000000010
1   1970-01-01 00:00:00.000000010
dtype: datetime64[ns]
>>> s.where(s.isnull(), np.float64(10))
0   2116-06-17 06:38:37.588971520
1   2116-06-17 06:38:37.588971520
dtype: datetime64[ns]
>>> s.where(s.isnull(), [0, 1])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/evanpw/Workspace/pandas/pandas/core/generic.py", line 3326, in where
    if not (new_other == np.array(other)).all():
AttributeError: 'bool' object has no attribute 'all'

[jreback]

I would like to see if we can consolidate some of this logic to com._infer_fill_value (which may need a bit of logic), but I think should work

[evanpw]

I'm not sure I follow. We already know the source and destination dtypes: datetime64 -> int64 or timedelta64 -> float64. What do we need to infer?

[jreback]

its not that we need to infer, my point is that a very similar code block exists in several places in the code base. needs to be consolidated to avoid repeating is again. Pls look thru and see where it is duplicated (e.g. in internals.py as well)

[evanpw]

I'm sorry, I still don't see the common functionality. You're talking about this function:

def _try_coerce_args(self, values, other):
    """ Coerce values and other to dtype 'i8'. NaN and NaT convert to
        the smallest i8, and will correctly round-trip to NaT if converted
        back in _try_coerce_result. values is always ndarray-like, other
        may not be """
    values = values.view('i8')

    if is_null_datelike_scalar(other):
        other = tslib.iNaT
    elif isinstance(other, datetime):
        other = lib.Timestamp(other).asm8.view('i8')
    elif hasattr(other, 'dtype') and com.is_integer_dtype(other):
        other = other.view('i8')
    else:
        other = np.array(other, dtype='i8')

    return values, other

and this one?

def _infer_fill_value(val):
    """
    infer the fill value for the nan/NaT from the provided scalar/ndarray/list-like
    if we are a NaT, return the correct dtyped element to provide proper block construction

    """

    if not is_list_like(val):
        val = [val]
    val = np.array(val,copy=False)
    if is_datetimelike(val):
        return np.array('NaT',dtype=val.dtype)
    elif is_object_dtype(val.dtype):
        dtype = lib.infer_dtype(_ensure_object(val))
        if dtype in ['datetime','datetime64']:
            return np.array('NaT',dtype=_NS_DTYPE)
        elif dtype in ['timedelta','timedelta64']:
            return np.array('NaT',dtype=_TD_DTYPE)
    return np.nan

I haven't seen any code block that looks similar to _try_coerce_args elsewhere either. Could you give an example?

[jreback]

merged via 1f9b699

thanks!