On the Efficient Gathering of Sufficient Statistics for Classification from Large SQL Databases