we redid the network as intended. it now looks roughly like this:
the 10 broken servers were found to have faulty CMOS batteries. still rather annoying, but easy enough to fix.
we're discussion hosting proposals with Kennisnet and Belnet in the Netherlands - that'll hopefully be the first proper off-site cluster (rather than just squids).
a while ago i posted about caching pages for logged-in users. brion's recently committed some changes to remove preferences options' effects on the parser HTML output, which is one step closer to actually realising that. gwicke wants to do this with ESI, but i'm wary of such an intimate dependence on squid, particularly as it's only supported on squid 3, which has been in development for about 50 years. we've also started removing the LIMIT/offset based queries for things like page history & user contributions, and replacing them with LIMIT + page_timestamp. for long histories, that's much faster (should be pretty much constant time however far back you go in the history).
we're now back to Perlbal (again) as squid is still pretty useless at load distribution, even with icpagent.
an installation of FreeBSD on one of the idle machines is planned to test it for squid performance. we also have shiny new squid performance statistics courtesy of dammit.