Design Goals: ~10^6 to 10^7 machines, ~10^13 directories, ~10^18 bytes of storage (that's 888 petabytes), spread at 100s to 100s of location around the world, ~10^9 client machines.
Simply staggering. I guess now we also have a formal definition of "Google Scale" which the Google PR people always go on about.
It's one machine per 100 to 1000 clients. There's many ways to interpret that. On one hand, it's a measure of how much computing Google is doing for clients. On the other, it's an indication of the level of Google's costs per client, and indirectly how much revenue they make from those clients.
It's almost government-scale infrastructure. It's approaching the point where Google is running a machine per street. I wouldn't be surprised if a government made a grab for the local section of this infrastructure in some country within the next couple of decades, whether it's to control their citizens, or to wrest control from the corporation on behalf of their citizens.
Jeff Dean is like a god around here. Fact: Jeff Dean was forced to invent asynchronous APIs one day when he optimized a function so that it returned before it was invoked. Also, gcc -O4 sends your code directly to Jeff Dean for a rewrite.
Simply staggering. I guess now we also have a formal definition of "Google Scale" which the Google PR people always go on about.