I've heard that they have used clusters of servers (100's) that keep a large chunk of the most recent data in RAM. This would be a huge accomplishment, but I don't think that it's possible from a business side.

I would say that they separate all of the data into small clusters and sub-clusters, then those could be organized by date, predicted/actual traffic, type of data, ect.. Those sub-clusters would then have a reasonable amount of data to sift through (maybe only a few million rows), this would speed up the response and retrieval.

Again, I don't know, I'm just guessing.

__________________
My Site