I wish there was more data offered!
Request-response, XHR, long-polling, "COMET" (blarg, I hate that moniker) when talking to the backend?
He said that they're using LAMP. I'd like to hear more about how they deal with schema updates and how they model their data in general.
They use cache components to scale their writes, since they have so many writes. I assume this means that instead of hitting disk for every write, data in the cache is updated instead, and asynchronously saved to disk. Really, this pattern should be generalized, in some sort of memcached extension/fork or something.