Thanks

You have a point there regarding parquet. Just checked, it does read a separate row group in each task, the thing is that the row groups are unbalanced, so the second taks gets just some KB of data.

Case closed, kudos @Werner Stinckens​