This is usually referred to as "high availability", where you'd have a hot failover to swap to in case the main server goes down. This is usually implemented with a load balancer that checks if the upstream server is alive before sending requests to it. If the upstream server isn't responding, switch to the other one.

A load balancer could also spread the load evenly across multiple machines, at least for reads. Generally there's far more reads than writes, and reads are more easily scalable since you can have database replicas that just need to sync in one direction.

I don't think Lemmy supports any of this yet though.

The other approach is to split the large instances into multiple smaller instances. For Fediverse stuff, I don't know which approach is considered "better".