Currently the DB query used by the instance usage audit, instance_get_active_by_window_joined() uses the terminated_at timestamp.
terminated_at is normally set as part of the instance deletion processing, however there are cases where an exception at the wrong time could prevent terminated_at from being set. Also the recent bug fixed by this change https://review.openstack.org/#/c/42534/ missed out this update altogether, so instances created by a system with that bug in-situ will now have large numbers of instances that will be continually reported as existing even though the entry in the DB is deleted.
Given that instance_usage_audit is meant to report on instances that from the DB perspective existed on a host in the previous audit period it would be more consistent to change instance_get_active_by_window_joined() to use deleted_at - which is set directly by the DB layer when the entry is deleted.
This would mean instances which have terminated_at set but not deleted would be reported as existing - which is also more consistent with the intended behviour of the audit.
The only case where terminated_at is not set as part of deletion is already reported as a separate bug: https://bugs.launchpad.net/nova/+bug/1189554
Good catch Phil. This one doesn't look too hard to fix, so I'm going to try to squash it for havana-rc1