Interesting - curious about what the actual failures were as I've never seen failures like that with fairly large Varnish caches. nginx wasn't an option for us until they made it handle the HTTP Vary header correctly so I don't have recent comparative experience with it as a cache.