Polars is fastest when you avoid eager eval mid-pipeline. If you see a 40x gap it's often from calling .collect() inside a loop or applying Python UDFs row-wise.