More generally I should be able to choose to run an interruptible workload that I know to leak memory. I should expect that if I don’t, one of my coworkers will, and the node will stay up. Not leaving enough RAM for the node’s core resources is a mistake, but far from the worst thing in the world.