Currently if an array is empty, we just return np.empty(0), which has a dtype f8. This means that dtype of the output won't match for other dtypes. Further, you can have "empty" arrays with shape != () as long as one of the dimensions is 0 (e.g. shape=(0, 10, 10)).

The da.Array objects have the necessary metadata to create empty arrays with the appropriate dtype/shape, but these aren't passed on to _finalize. One solution for this would be to make _finalize a property, and return a partial with the metadata if the output is known to be empty. This slightly complicates the implementation, but isn't too bad. I'll do it here:

def finalize_empty(shape, dtype, res):
    return np.empty(shape, dtype=dtype)

class Array(Base):
    ...
    @property
    def _finalize(self):
        if self.shape or 0 not in self._shape:
            return finalize
        return partial(finalize_empty, self.shape, self.dtype)

This would solve the issue and is (I think) pretty simple, but I'll open it here for discussion. This also might be a non-issue, as we might not care if the metadata matches for empty arrays.