This bug is visible when pluggable IPAM is active. It can be seen with this patch [1]. It does not cause gate failures but it is still something that should be understood. This logstash query [2] seems to find where they occur. It is helpful to look at the DEBUG level logging around the time of the error. For example see this paste [3].
It seems that the session gets broken with an exception that causes a rollback. Then, the IPAM rollback attempts to use the same session for rollback which fails. Should the reference pluggable IPAM driver be using a different session? Or, should it call rollback?
[1] https://review.openstack.org/#/c/181023
[2] http://logstash.openstack.org/#dashboard/file/logstash.json?query=message%3A%5C%22IP%20deallocation%20failed%20on%20external%20system%5C%22
[3] http://paste.openstack.org/show/532891/
Can we write a test that tickles this problem? Could it be as simple as passing two fixed_ips to a port, one that is available and allocation would succeed and then a second that conflicts with an IP that is already allocated? The fixed IPs would get sorted here [1] which could make the test fail or pass intermittently.
[1] https:/