In order to know how much money to throw at hardware for kinja we occasionally benchmark it to see how much headroom we have left. Only instead of coming up with an artificial benchmark that tries to accurately represent our millions of daily users, we do the inverse and *remove* servers from live traffic to see just how few we can run on without collapsing.

Today: eleven. sixteen. (see comments)

appserver cpu utlization

browser impact according to newrelic

mantle response times


core response times

profile response times


- even just going 90/10 (effectively 22 servers) shows an impact in every component except overall browser render time (do we trust that?)
- between 22 and 11 servers things got incrementally worse but kept working
- at 10 servers everything blew up (4:15)