A question that's been on my mind: Why did Facebook scale so well and Twitter scale so poorly? Is there really any way to learn how to cope with huge load before you actually hit the point where you need it?
Facebook have an order of magnitude more employees than Twitter which probably makes those kind of bespoke uber-scaling advancements feasible.
They use memcached with some nasty kernel hacks.
(Caused some friction when they forked an old version, added new code which was then suddenly unleashed on the community without warning. Each new feature then had to be manually added to the current version)
(highscalability.com is down. perhaps they need highavailability.com?)