- @ShutdownActorBase._retry((subprocess.CalledProcessError,))
- def cancel_shutdown(self):
- self._set_node_state('RESUME')
- return super(ComputeNodeShutdownActor, self).cancel_shutdown()
+ def _get_slurm_state(self):
+ return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
+
+ # The following methods retry on OSError. This is intended to mitigate bug
+ # #6321 where fork() of node manager raises "OSError: [Errno 12] Cannot
+ # allocate memory" resulting in the untimely death of the shutdown actor
+ # and tends to result in node manager getting into a wedged state where it
+ # won't allocate new nodes or shut down gracefully. The underlying causes
+ # of the excessive memory usage that result in the "Cannot allocate memory"
+ # error are still being investigated.
+
+ @RetryMixin._retry((subprocess.CalledProcessError, OSError))
+ def cancel_shutdown(self, reason):
+ if self._nodename:
+ if self._get_slurm_state() in self.SLURM_DRAIN_STATES:
+ # Resume from "drng" or "drain"
+ self._set_node_state('RESUME')
+ else:
+ # Node is in a state such as 'idle' or 'alloc' so don't
+ # try to resume it because that will just raise an error.
+ pass
+ return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)