And this isn’t even the interesting part of the presentation. By putting new versions of FreeBSD-current into production every few days, Netflix spotted a perf regression quickly and was able to “easily” bisect a few days worth of changes to find the offending commit, then get it fixed—because the context was fresh in the dev’s mind.
Now think about how much PR-based workflows in your org delay your changes getting into prod, and what that really costs.
https://mastodon.online/@karppinen/112672855308433453
This is a good point, but also this bisection in and of itself is pretty amazing to be able to do, when each bisection test takes 4 hours to figure out if the commit is good or bad!
RE: https://mastodon.social/@ratkins/112675470604166076
@alpha I cannot tell you how valuable a four hour production-load perf test would be. Especially one sensitive enough to catch an 8% cpu variance.
@josephholsten @alpha Yeah, this is one hell of an engineering flex to have this infrastructure already in place and be able to use it to conclusively track down problems like that.