I think each document is a row and it parses it to native structures on load. I think I saw something on the dev list about that being improved - but its something to check if it matters.