They denormalize them most certainly (lots of data duplication in the data models) which is fast but relatively complex to maintain.