Arvados

As of this writing, the Node Manager's ComputeNodeMonitorActor and friends look at the node record's info['slurm_state'] string to decide whether or not the node is eligible for shutdown.

But the API server is responsible for knowing its dispatch method and translating between that and a common string. It exposes this as crunch_worker_state, which can be one of 'busy', 'idle', or 'down'. Node Manager should use this field to make shutdown decisions instead.

Note that I'm only talking about making a change when Node Manager is looking at a node record to make shutdown decisions. Code that talks to SLURM directly, like the ComputeNodeShutdownActor in the SLURM dispatch module, doesn't need to be changed.