Node drain
Taking a server out of rotation on purpose: stop sending it new work, let the requests and jobs it already holds finish, then shut it down or upgrade it. A load balancer's connection draining and Kubernetes' `kubectl drain` both work this way. Killing a busy server instead drops whatever it was doing.
Reliability
In Quorum#
Health checks research adds a Drain button to a Node's card. The Node takes no new traffic at once, and its share goes to the other Nodes on the path. It finishes what it holds: a Queue hands its backlog to its Workers. Ten sim seconds later, empty, it rejoins the rotation. Draining the Node an Incident is hitting cuts that Incident to the 10 s drain window. A Node with no other copy to take the load strands every request sent to it.