Hey EKS, it would be really cool of you if I didn't have to play Russian roulette with you when I try to scale down a node.
I cordoned a drained a node for you to notice, wouldn't it be too hard to let me select which one to put down instead of doing whatever you want?
Thanks.

Lol, my old company had a similar issue with ECS. We wanted to scale down (or up) processing as needed, but not to kill a node that was in the middle of a task. We ended up doing some really weird hacky workarounds to ensure that ECS wouldn't take down the wrong nodes.

With kubernetes you should not have this requirement. If your workload cannot be terminated (in a timely fashion) at any time; it's not suited for k8s.
This is the inherent design. It will schedule and reschedule workloads, it will only guarantee that an pod is running, either minimum replicas or restarting/retrying.