the hard part (and why it hasn't been done yet) is doing so in a way that scales to very large datasets