Comment 9 for bug 1348115

I saw this in the context of cinder volume service. Basically, if one sets following options in the openstack service config (cinder.conf in this example)
notification_driver=messaging
publish_errors=True

After this, if the rpc connection breaks, the oslo.messaging code invokes LOG.error, which in turn calls oslo.rpc and recursion ensues.
The worst part is that all green threads eventually lock up. The service cannot continue even though rpc connection is restored.