That's exactly what we are thinking about. This would give you overnight access to all the goodness of DuckDB and Polars with respect to scalar functions and the Map part of series aggregations. Also, we realized that most windowed operations are done with groupby dimensions. Therefore, if you were to align your partition dimensions with your windowed groupby dimensions ahead of the windowed operations (modulo some upstream Ray-powered repartitioning), you could use DuckDB's and Polar's
window
to execute them, without requiring any reductions and without having to worry about cross-partition spillovers.