Merging variables with overlapping not but not conflicting values

[shoyer]

It should be possible to safely merge together variables with values [NaN, 1, 2] and [0, 1, NaN] by using methods such as combine_first (which should also be OK with conflicting values, like the pandas method) and merge (which should raise if values conflict).

See this stackoverflow post for a merge example: http://stackoverflow.com/questions/36731870/how-to-merge-xarray-datasets-with-conflicting-coordinates

[shoyer]

I'm currently in the process of refactoring the merge code so this should be pretty easy to add in.

A few snippets that should come in handy:

def combine_first(arrays):
    notnulls = [ops.notnull(array) for array in arrays]
    first_notnull = ops.argmax(ops.stack(notnulls), axis=0)
    return ops.choose(first_notnull, arrays)

def combine_if_equal(arrays):
    combined = combine_first(arrays)
    combined_null = ops.isnull(combined)
    if not all(((combined == array) | combined_null).all() for array in arrays):
        raise ValueError
    return combined

[shoyer]

Actually, we can just write:

def combine_first(left, right):
    return ops.where(ops.notnull(left), left, right)

[shoyer]

Fixed by #1204