It is weird to say both that "CSV files have terrible compression" and then that the proposed format, Apache Parquet, has "Really good compression properties, competitive with .csv.gz". I think what's meant here is that csv compresses really well but you loose the ability to "seek" inside the file.