Comment 6 for bug 1491276

terminate_connection is not sure to fail because it it more common to have the cinder-volume node as a separate machine than the nova-cpu node.
In this scenario, the nova-compute host is down but the cinder-volume node is up, and all proceeds as designed.
If the original connection is still present (because terminate_connection is never called), it will be picked up by the original nova-compute hosts when it restarts and does an iscsi rescan (in the case of Iscsi) and can lead to data corruption.