How Daft optimizes the reading of real-world data which is often a mix of "many small files" and "few large files"