I'm surprised they don't have a couple of separate clusters that they roll things out to and monitor. Seems like you could have a very stable "high paying customers" cluster that is at the very end of your deployment cycle after a ton of canary checks on the way get through.