BUG: incorrect usage of pchip for timedelta

[jreback]

This just started breaking on scipy master:
https://dev.azure.com/pandas-dev/pandas/_build/results?buildId=10858

xfailing in #26190

> /Users/jreback/pandas/pandas/tests/series/test_missing.py(1446)test_interpolate_timedelta_index()
-> result = df[0].interpolate(method=method, **kwargs)
(Pdb) n
> /Users/jreback/pandas/pandas/tests/series/test_missing.py(1447)test_interpolate_timedelta_index()
-> expected = pd.Series([0.0, 1.0, 2.0, 3.0], name=0, index=ind)
(Pdb) p result
0 days 00:00:00.000000    0.0
1 days 00:00:00.000000    1.0
2 days 00:00:00.000000    2.0
3 days 00:00:00.000000    3.0
Freq: D, Name: 0, dtype: float64
(Pdb) p df[0]
0 days 00:00:00.000000    0.0
1 days 00:00:00.000000    1.0
2 days 00:00:00.000000    NaN
3 days 00:00:00.000000    3.0
Freq: D, Name: 0, dtype: float64

I think we are passing invalid inputs

> /Users/jreback/pandas/pandas/core/missing.py(311)_interpolate_scipy_wrapper()
-> new_y = method(x, y, new_x, **kwargs)
(Pdb) p method
<function pchip_interpolate at 0x81bb9e9d8>
(Pdb) p x
array([              1,  86400000000001, 259200000000001],
      dtype='timedelta64[ns]')
(Pdb) p y
array([0., 1., 3.])

[jreback]

cc @rgommers

[rgommers]

thanks @jreback looking into it now

[rgommers]

I'll add the full failure from the log here to save some clicking:

The key part is

>       if np.any(dx <= 0):
E       numpy.core._exceptions.UFuncTypeError: Cannot cast ufunc 'less_equal' input 0 from dtype('<m8[ns]') to dtype('<m8') with casting rule 'same_kind'

I think we are passing invalid inputs

It's one of those things - we never test with timedelta or datetime anywhere in scipy and in principle we'd indeed consider it invalid input, however it is a numpy builtin dtype and it mostly just works. This failure can perhaps be easily worked around.

``` [gw1] linux -- Python 3.7.3 /home/vsts/miniconda3/envs/pandas-dev/bin/python

self = <pandas.tests.series.test_missing.TestSeriesInterpolateData object at 0x7f4a444c1ba8>
interp_methods_ind = ('pchip', {})

def test_interpolate_timedelta_index(self, interp_methods_ind):
    """
    Tests for non numerical index types  - object, period, timedelta
    Note that all methods except time, index, nearest and values
    are tested here.
    """
    # gh 21662
    ind = pd.timedelta_range(start=1, periods=4)
    df = pd.DataFrame([0, 1, np.nan, 3], index=ind)

    method, kwargs = interp_methods_ind
    if method == "pchip":
        _skip_if_no_pchip()

    if method in {"linear", "pchip"}:

      result = df[0].interpolate(method=method, **kwargs)

pandas/tests/series/test_missing.py:1445:

---

pandas/core/generic.py:6884: in interpolate
**kwargs)
pandas/core/internals/managers.py:520: in interpolate
return self.apply('interpolate', **kwargs)
pandas/core/internals/managers.py:396: in apply
applied = getattr(b, f)(**kwargs)
pandas/core/internals/blocks.py:1127: in interpolate
downcast=downcast, **kwargs)
pandas/core/internals/blocks.py:1191: in _interpolate
interp_values = np.apply_along_axis(func, axis, data)
../../../miniconda3/envs/pandas-dev/lib/python3.7/site-packages/numpy/core/overrides.py:150: in public_api
implementation, public_api, relevant_args, args, kwargs)
../../../miniconda3/envs/pandas-dev/lib/python3.7/site-packages/numpy/lib/shape_base.py:379: in apply_along_axis
res = asanyarray(func1d(inarr_view[ind0], *args, **kwargs))
pandas/core/internals/blocks.py:1188: in func
bounds_error=False, **kwargs)
pandas/core/missing.py:237: in interpolate_1d
order=order, **kwargs)
pandas/core/missing.py:310: in _interpolate_scipy_wrapper
new_y = method(x, y, new_x, **kwargs)
../../../miniconda3/envs/pandas-dev/lib/python3.7/site-packages/scipy/interpolate/_cubic.py:343: in pchip_interpolate
P = PchipInterpolator(xi, yi, axis=axis)
../../../miniconda3/envs/pandas-dev/lib/python3.7/site-packages/scipy/interpolate/_cubic.py:235: in init
x, _, y, axis, _ = prepare_input(x, y, axis)

---

x = array([ 1, 86400000000001, 259200000000001],
dtype='timedelta64[ns]')
y = array([0., 1., 3.]), axis = 0, dydx = None

def prepare_input(x, y, axis, dydx=None):
    """Prepare input for cubic spline interpolators.

    All data are converted to numpy arrays and checked for correctness.
    Axes equal to `axis` of arrays `y` and `dydx` are rolled to be the 0-th
    axis. The value of `axis` is converted to lie in
    [0, number of dimensions of `y`).
    """

    x, y = map(np.asarray, (x, y))

    if np.issubdtype(x.dtype, np.complexfloating):
        raise ValueError("`x` must contain real values.")

    if np.issubdtype(y.dtype, np.complexfloating):
        dtype = complex
    else:
        dtype = float

    if dydx is not None:
        dydx = np.asarray(dydx)
        if y.shape != dydx.shape:
            raise ValueError("The shapes of `y` and `dydx` must be identical.")
        if np.issubdtype(dydx.dtype, np.complexfloating):
            dtype = complex
        dydx = dydx.astype(dtype, copy=False)

    y = y.astype(dtype, copy=False)
    axis = axis % y.ndim
    if x.ndim != 1:
        raise ValueError("`x` must be 1-dimensional.")
    if x.shape[0] < 2:
        raise ValueError("`x` must contain at least 2 elements.")
    if x.shape[0] != y.shape[axis]:
        raise ValueError("The length of `y` along `axis`={0} doesn't "
                         "match the length of `x`".format(axis))

    if not np.all(np.isfinite(x)):
        raise ValueError("`x` must contain only finite values.")
    if not np.all(np.isfinite(y)):
        raise ValueError("`y` must contain only finite values.")

    if dydx is not None and not np.all(np.isfinite(dydx)):
        raise ValueError("`dydx` must contain only finite values.")

    dx = np.diff(x)

  if np.any(dx <= 0):

E numpy.core._exceptions.UFuncTypeError: Cannot cast ufunc 'less_equal' input 0 from dtype('<m8[ns]') to dtype('<m8') with casting rule 'same_kind'

../../../miniconda3/envs/pandas-dev/lib/python3.7/site-packages/scipy/interpolate/_cubic.py:65: UFuncTypeError

</details>

[rgommers]

Should work again with scipy master after scipy/scipy#10090