I see. That does sound like it should work. My best guess is that you’re scaling down to zero virtual machines and what you see is the cached site in your browser. Or some other way the kids balancer isn’t connecting to the host. Or discourse isn’t getting the real ip and it’s rate limiting (but usually you would see an error).
But your high availability features are providing low availability. Unless you’re going from having tens of users most of the time to thousands some others (as for a sports site) then scaling is likely to cause more problems than it solves.
So the first thing I’d do is get rid of the load balancer and see if that fixes it. Then decide what to do from there. If it happens once a month it won’t be easy to diagnose.