Lazy rectilinear interpolator, with sparse

[fnattino]

🚀 Pull Request

Description

Enable the rectilinear interpolator to run lazily #6002.

The original implementation makes use of a CSR matrix from scipy.sparse to represent the weights. However, this does not seem to work well with Dask arrays, since it does not adhere (completely) to the array interface. Switching to sparse arrays from sparse seems to work well, allowing the sparse matrix to be wrapped in a Dask array.

Would this be an acceptable change? On the long run, it looks like sparse will replace scipy.sparse in the PyData ecosystem..

[fnattino]

Additional note on the use of sparse: this is also the library used by xESMF. The drawback is that it would add numba as a dependency, which I realise it might not be ideal..

[fnattino]

Marking it as ready for review to hear your thoughts on this. What is in here for now seems to work (see e.g. code snippet below), provided that sparse is installed.

Test code snippet

import iris

from iris.analysis import RectilinearInterpolator


LATITUDE = [16., 16.1]
LONGITUDE = 226.


def get_cube():
    filename = iris.sample_data_path('E1_north_america.nc')
    return iris.load_cube(filename, 'air_temperature')


def interpolate(method):
    assert method in ('linear', 'nearest')
    cube = get_cube()
    coords = ('latitude', 'longitude')
    points = (LATITUDE, LONGITUDE)
    interpolator = RectilinearInterpolator(cube, coords, method, "mask")
    print('Cube is lazy: ', cube.has_lazy_data())
    # Cube is lazy:  True
    result = interpolator(points, collapse_scalar=True)
    print('Result is lazy: ', result.has_lazy_data())
    # Result is lazy:  True


if __name__ == '__main__':
    interpolate('linear')
    interpolate('nearest')

[trexfeathers]

Hi @fnattino, thanks for your hard work. Just to warn you I expect several weeks' delay before we can get to this as we are accumulating a backlog of ESMValTool changes while we prepare GeoVista for SciPy 2024 and complete the mesh-focused Iris 3.10 release.

[fnattino]

I have worked on a different approach that makes the intepolator lazy by using a similar approach as used in the regridder, see #6084. This has the disadvantage that it requires merging the chunks along the interpolating dimensions, but it is not as "disruptive" as in it does not add new dependencies, so maybe worth to have a look at #6084 first keeping this for the longer run? In the meantime I can also try to run some benchmarks to compare the performance of sparse vs scipy.sparse.

[HGWright]

From discussion at the @SciTools/peloton. Is this still required since #6084 has been completed and merged in? @fnattino

[fnattino]

Hi @HGWright, this was indeed initial work for an alternative to #6084 -- it can now be closed.