issue_comments

Thanks @headtr1ck !

Forcing the user to be explicit reduces bugs and user support requests :) so we like to do that.

Thanks again @headtr1ck this is a great PR!

Nice this looks like an improvement for everything other than dask arrays with only 100 elements, which is not a good use-case for dask.

I was slightly concerned that the recrusive algorithm wouldn't work well with dask but it does seem to work better.

``` python def other_polyval(coord, coeffs, degree_dim="degree"): x = coord.data

deg_coord = coeffs[degree_dim]
N = int(deg_coord.max()) + 1

lhs = xr.DataArray(
    np.stack([x ** (N - 1 - i) for i in range(N)], axis=1),
    dims=(coord.name, degree_dim),
    coords={
        coord.name: coord.data,
        degree_dim: np.arange(deg_coord.max() + 1)[::-1],
    },
)
return xr.dot(lhs, coeffs, dims=degree_dim)

coeffs = xr.DataArray(np.random.randn(2), dims="degree") da = xr.DataArray(dask.array.random.random((10**6), chunks=(10000)), dims=("x")) print(len(da.data.dask)) print(len(xr.polyval(da, coeffs).data.dask)) print(len(other_polyval(da, coeffs).data.dask)) 100 502 1005 ```

Benchmark did not succeed since the inputs are not compatible with the old algorith... Do we change it such that it is compatible?

If you're confident this is faster and it's not a trivial amount of work to adjust them, I would leave it...

This looks like excellent code @headtr1ck , especially so for a first PR. Welcome!

so I used a "hack" of adding an 0-valued DataArray/Dataset.

I think it's fine to do the hack and reference that issue; we can clean it up then (though if others have ideas then great).

I had to add a copy(deep=True) since the broadcast returned a read-only DataArray. Any better ideas?

IIRC we generally leave this up to the caller (and generally discourage mutation).

Someone who knows better should review the algo & opine on the datetime issue; I agree I'm not sure whether we need to support them.