You put nginx or Haproxy in front of the hosts, drop the one that needs maint from the pool, and re-add once it’s ready.

Depending on the deployment and any SLAs, I either don't worry about it (just do a late night rollout when nobody is on the system) or rely on my deployment architecture's sibling checks (I can see when a given machine is still versioning and requeue subsequent rollouts to other machines).

I’m pretty sure for most systems that does not matter in the slightest.

You spin up a second host and load balance

How is this an issue in a world where load balancers exist? I was part of a Unicorn that ran prod on 8 boxes and literally never had customer facing outages due to infrastructure updates.