This was for a high volume retailer. A Proxy / caching solution was supposed to be provided and TW would do the app. At t-4, it turned out that TW would also have to provide the proxy / cache, so this was a hasty investigation into Squid.
1 hardware LB
2 Squid boxes - 8 core machine
1 carp process
lots of child processes
going to 16 app servers.
16,000 requests per second = 5% of traffic
TTL for items ranged from 5 minutes to 1 hour for stable furniture.
whole site does 250 million requests per day
peak 24,000 requests per second
importance of good HTTP Caching directives. Discussion of making the application (and by implication, the application developers) aware of considering ETag / Expires / Vary for all parts of the application, versus just adding it via apache ReWrite or similar. Most people in the room (including me, very strongly; RFC2616 is my favourite RFC) were in the former camp.
heap LFUDA was a good change to make.
Tried Varnish, but couldn't get good numbers out of it, in the timescales available. Artur opined that Varnish should provide better numbers than Squid; he's arguably conflicted, but seemed pretty convincing! Another factor in that was very likely that they were running RHEL old shit, and Varnish works best with a shiny new kernel (cite?)