If I look into perf, it seems mostly parsing overhead. I can saturate a create newtable as select from oldtable. Unfortunately, CSV seems still the lingua franca for transport between DBs. Maybe some day a more binary oriented transport protocol will appear( e.g parquet?)