I was reading through the code and it appears that the killSelection configuration is only considered when scaling, not during restart/redeploy. In a world where 0 downtime redeploys are the goal, I believe it's important to allow teams to make decisions about which nodes are killed first on redeploy as well.
For example, in Akka Cluster where Coordinators/Singletons exist on the oldest node, killing the oldest node will cause the singleton to move on each kill step, whereas if the youngest were killed first then the singletons would only move on the last kill step.
Similar cases may arise where, for example, zookeeper is used for leader election and the oldest ephemeral sequential node is considered to be the leader. If killing the eldest node every time, leadership election could occur on every kill step instead of just once.
This isssue has been created automatically from Marathon GitHub Issue 5369.