We never published most of the things we did in order to accomplish this scaling. Lots of cool things here like storing data on outside of disk platters for better read bandwidth, better index compression, to then storing the entire web index in RAM (in each serving cluster!)

→ View original post on X — @jeffdean