In 2 GB of memory.
Technically speaking, I find hacker news persistent strategy one of the most interesting things about the implementation.
I use a similar strategy for my blog and just playing with larger datasets I've certainly run into hard limits on what seems to be acceptable.