X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/2c69d491fbbd1461bb5a6dd9c0f0dd493081b45e..de4df7f80c531ab16e59ea36671a8efa9e6ff33d:/services/nodemanager/arvnodeman/daemon.py

diff --git a/services/nodemanager/arvnodeman/daemon.py b/services/nodemanager/arvnodeman/daemon.py
index cb11521ef9..f23b2615e2 100644
--- a/services/nodemanager/arvnodeman/daemon.py
+++ b/services/nodemanager/arvnodeman/daemon.py
@@ -232,7 +232,7 @@ class NodeManagerDaemonActor(actor_class):
     def try_pairing(self):
         for record in self.cloud_nodes.unpaired():
             for arv_rec in self.arvados_nodes.unpaired():
-                if record.actor.offer_arvados_pair(arv_rec.arvados_node).get():
+                if record.actor is not None and record.actor.offer_arvados_pair(arv_rec.arvados_node).get():
                     self._pair_nodes(record, arv_rec.arvados_node)
                     break
 
@@ -243,16 +243,15 @@ class NodeManagerDaemonActor(actor_class):
         return s
 
     def _node_states(self, size):
-        states = pykka.get_all(rec.actor.get_state()
-                               for rec in self.cloud_nodes.nodes.itervalues()
-                               if ((size is None or rec.cloud_node.size.id == size.id) and
-                                   rec.shutdown_actor is None and
-                                   rec.actor is not None))
-        states += ['shutdown' for rec in self.cloud_nodes.nodes.itervalues()
-                   if ((size is None or rec.cloud_node.size.id == size.id) and
-                       (rec.shutdown_actor is not None or
-                        rec.actor is None))]
-        return states
+        proxy_states = []
+        states = []
+        for rec in self.cloud_nodes.nodes.itervalues():
+            if size is None or rec.cloud_node.size.id == size.id:
+                if rec.shutdown_actor is None and rec.actor is not None:
+                    proxy_states.append(rec.actor.get_state())
+                else:
+                    states.append("shutdown")
+        return states + pykka.get_all(proxy_states)
 
     def _state_counts(self, size):
         states = self._node_states(size)
@@ -427,16 +426,25 @@ class NodeManagerDaemonActor(actor_class):
 
     @_check_poll_freshness
     def node_can_shutdown(self, node_actor):
-        if self._nodes_excess(node_actor.cloud_node.get().size) > 0:
-            self._begin_node_shutdown(node_actor, cancellable=True)
-        elif self.cloud_nodes.nodes.get(node_actor.cloud_node.get().id).arvados_node is None:
-            # Node is unpaired, which means it probably exceeded its booting
-            # grace period without a ping, so shut it down so we can boot a new
-            # node in its place.
-            self._begin_node_shutdown(node_actor, cancellable=False)
-        elif node_actor.in_state('down').get():
-            # Node is down and unlikely to come back.
-            self._begin_node_shutdown(node_actor, cancellable=False)
+        try:
+            if self._nodes_excess(node_actor.cloud_node.get().size) > 0:
+                self._begin_node_shutdown(node_actor, cancellable=True)
+            elif self.cloud_nodes.nodes.get(node_actor.cloud_node.get().id).arvados_node is None:
+                # Node is unpaired, which means it probably exceeded its booting
+                # grace period without a ping, so shut it down so we can boot a new
+                # node in its place.
+                self._begin_node_shutdown(node_actor, cancellable=False)
+            elif node_actor.in_state('down').get():
+                # Node is down and unlikely to come back.
+                self._begin_node_shutdown(node_actor, cancellable=False)
+        except pykka.ActorDeadError as e:
+            # The monitor actor sends shutdown suggestions every time the
+            # node's state is updated, and these go into the daemon actor's
+            # message queue.  It's possible that the node has already been shut
+            # down (which shuts down the node monitor actor).  In that case,
+            # this message is stale and we'll get ActorDeadError when we try to
+            # access node_actor.  Log the error.
+            self._logger.debug("ActorDeadError in node_can_shutdown: %s", e)
 
     def node_finished_shutdown(self, shutdown_actor):
         try: