Instances with delete_on_terminate volumes do not reschedule

[Vish Ishaya]

If you create an instance with attached delete_on_terminate volumes and it fails to launch on a host (due to a hardware or hypervisor issue) it will be rescheduled but the volumes will have already been deleted. The volumes should only be deleted if the instance is not going to be rescheduled (or perhaps not at all).

[Vish Ishaya (vishvananda)]

I'm thinking something like the following

diff --git a/nova/compute/manager.py b/nova/compute/manager.py
index 6f756e4..2e7aebb 100644
--- a/nova/compute/manager.py
+++ b/nova/compute/manager.py
@@ -1533,6 +1533,10 @@ class ComputeManager(manager.Manager):
                         reason=six.text_type(exc_info[1]))
             else:
                 # not re-scheduling, go to error:
+ if bdms is None:
+ bdms = objects.BlockDeviceMappingList.get_by_instance_uuid(
+ context, instance.uuid)
+ self._cleanup_volumes(context, instance.uuid, bdms)
                 raise exc_info[0], exc_info[1], exc_info[2]

         # spawn success
@@ -1569,7 +1573,6 @@ class ComputeManager(manager.Manager):

             self._shutdown_instance(context, instance,
                                     bdms, requested_networks)
- self._cleanup_volumes(context, instance.uuid, bdms)
         except Exception:
             # do not attempt retry if clean up failed:
             with excutils.save_and_reraise_exception():

[Andrew Laski (alaski)]

This was fixed with https://review.openstack.org/#/c/106477/, though that only applies to deployments who are using the new build_and_run_instance method in manager.py. A new fix would be necessary in order to backport it to deployments using run_instance.

[Vish Ishaya (vishvananda)]

Thanks andrew, marked this as a duplicate