Chunks for interpolation methods other than "conservative"

[TimBusker]

We try to use linear and nearest interpolation to regrid a large xarray dataset with x and y as (only) dimensions. We would like to use the chunks as we define in the input data, and get the same chunks as output after the operation. This is supported by using the "conservative" method, but not for the other interpolation methods. Are there any plans to include this?

[slevang]

The linear and cubic methods are based on xarray's core interpolation routines, which are currently not that friendly to chunking along the interpolation dimensions. See this issue for more context. There have been some recent improvements in dask that help to some extent so you may want to check your dask version. Other approaches like a version of interp that uses map_overlap have been discussed and could be helpful.

I'm less familiar with how the flox-based statistical reductions behave. I think they should handle this well but don't actually know.

[TimBusker]

Thank you for the clarification. We will take a look at map_overlap and see if it can help.