5352: Set node size on records returned from libcloud create_node
[arvados.git] / services / nodemanager / arvnodeman / computenode / dispatch / slurm.py
index b4ed088e2169adbb1496f160539bf50194057e2d..919b57f42c8973bab91de742d1fee48598296f35 100644 (file)
@@ -34,8 +34,16 @@ class ComputeNodeShutdownActor(ShutdownActorBase):
     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.
+
     @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
-    def cancel_shutdown(self):
+    def cancel_shutdown(self, reason):
         if self._nodename:
             if self._get_slurm_state() in self.SLURM_DRAIN_STATES:
                 # Resume from "drng" or "drain"
@@ -44,7 +52,7 @@ class ComputeNodeShutdownActor(ShutdownActorBase):
                 # 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()
+        return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
 
     @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
     @ShutdownActorBase._stop_if_window_closed