On Jan 18, 2008 9:53 PM, Steve Summit <scs(a)eskimo.com> wrote:
Do the load balancers in WMF's server farms
attempt to associate
sessions with servers? At first I thought this unlikely, but I
guess it could significantly improve the hit rate.
Most likely you are seeing the effect of caching in the second tier
squids. LVS distributes connections to the front end squids, front
end squids do a tiny amount of memory only caching, then distribute
requests to back end squids *based on the hash of the URL*.
This provides a tremendous gain in effective memory and disk storage
(not storing hundreds of copies of every object), and the skinny cache
on the front ends prevent single object demand spikes from taking out
the site.
[snip; out of order]
When we were just having those stale-cache issues,
with
redirected pages and the like, it seemed that you got a
consistently stale page, not the randomly stale page you'd expect
Also, on a frequently requested object if multiple caches get ahold of
the object they'll all likely get purged or all not be purged, because
the purging is triggered directly from mediawiki (via the spiffy
multicast purger) and the caching lifetime is very long for objects
which have not been purged.