ENH: Forward & back fill methods

[max-sixty]

I think with np.flip and bn.push, this should be simple. They're both fairly new and so would require version checks / upgrading the minimums.

One small issue, I wonder if anyone has come across this: bottleneck returns the numpy array rather than the DataArray - is that because it's not operating with the correct numpy interface?

Forward fill:

array.values = bn.push(array.values, axis=array.get_axis_num(axis_name))

Backfill:

axis = array.get_axis_num(axis_name)

# reverse for bfill
array = np.flip(array, axis=axis)

# fill 
array.values = bn.push(array.values, axis=axis)

# reverse back to original
result = np.flip(scaling, axis=date_axis)

[shoyer]

One small issue, I wonder if anyone has come across this: bottleneck returns the numpy array rather than the DataArray - is that because it's not operating with the correct numpy interface?

Can you explain? I don't think bottleneck is xarray aware, so I'm not surprised by this. NumPy doesn't have a generic interface for external functions (only ufuncs).

[max-sixty]

Right, but numpy functions return the original type?

In [6]: da=xr.DataArray(np.random.rand(2,4))

In [12]: np.sum(da, axis=1)
Out[12]:
<xarray.DataArray (dim_0: 2)>
array([ 0.880766,  2.058156])
Dimensions without coordinates: dim_0


In [14]: np.flip(da, 1)
Out[14]:
<xarray.DataArray (dim_0: 2, dim_1: 4)>
array([[ 0.279932,  0.307569,  0.291855,  0.001411],
       [ 0.681486,  0.556972,  0.083212,  0.736487]])
Dimensions without coordinates: dim_0, dim_1

[shoyer]

Right, but numpy functions return the original type?

Not quite -- many numpy functions check for a method of the same name on their argument and call it instead of the numpy routine. If you look at the source of np.sum, it first checks for a sum method on its argument and calls that method instead if found. Otherwise, it coerces the argument to a numpy array and calls the numpy method.

[shoyer]

Other numpy functions, e.g., flip check for some attributes and then assume duck type compatibility if found. For example, flip is turned into an indexing call.

[max-sixty]

Ah, that explains a lot. Thanks for the clarification.

[jhamman]

@MaximilianR - I'm a big +1 on these features. Pandas has a missing module. I think these methods, combined with the interpolation methods I'm working on in #1640 would cover a large chunk of our use cases.

[max-sixty]

Closed by #1640