Let a control loop add and remove capacity in response to live demand — so you pay for load you actually have, not load you might one day see, and the system rides out spikes without a human awake to notice.
Run two identical production environments, deploy to the idle one, and cut over by flipping a router — so release and rollback are a single instant switch, and the database migration is the part that actually bites.
Ship a new version to a small slice of real traffic, compare it against a baseline on live signals, and promote or roll back on evidence — progressive delivery where the release decision is made by data, not by nerve.
Spreading traffic across replicas is easy; choosing well under real conditions is not. Algorithms, L4 vs L7, health checks, session affinity, and why the best-informed strategy often fails worst.
Hand the provider your code and the event that triggers it: they run it on demand, bill per invocation, and scale to zero when idle — so you own the business logic and none of the servers under it.