Well, I didn't read the entire article, so maybe there was an answer to this but I'm a bit weary of claims like "can be modified to scale linearly with the number of machines with 1 week of coding".
The main difference in effort between a Scala/lift and Erlang/ErlyWeb implementation is that making the Erlang app distributed and linearly scalable doesn't require any coding. Just store the process IDs in Mnesia and when you add more nodes add those nodes to the Mnesia schema and you're done.
Not having to restart your Erlang server (and therefore lose your client connections) to do code updates is also pretty cool.
I was making a joke ;) However, looking at the scala code it doesn't look like it is written to scale at all. I can't really see any thing in it that indicates it was written in a certain way. Its more a pattern to follow I guess.
This whole thread smells of lack of understanding of software. Hackers should know how inefficient an SQL server would be to handle Twitter-type tasks, versus specialized code. This is supposed to be hacker's news, not small business news.
news.YC is a good example of an efficient architecture. If it was written in Ruby/MySQL, pg would need at least 20 servers to keep up with the current traffic, versus a single box now. And the latency would never be as good as it is now.
