More scaling options than just memory and CPU would be great. This could be as simple as polling a /metrics endpoint with custom logic that triggers a 503 or something like that. I'm thinking of redis queue depth for instance