In fullstack test, unlike in a real deployment, more than one OVS agent is spawned in the same host, reading from the same OVS database. That could lead to isolation errors when receiving notifications. For example, OVS has a thread monitoring the new ports added to the host, using the shell CLI:
"ovsdb-client monitor Interface name,ofport,external_ids"
In a fullstack test, two OVS agents running in the same host will have independent integration bridges (randomly generated names). But when a port is added to one of those integration bridges, both agents will receive the event of a new port (interface) added. The OVS agent where the port is not added, won't find it in its own integration bridge and will treat it as dead port [1], sending a "update_device_down" message to the server.
This behavior could clash with the correct update sent from the other OVS agent, setting the port up. E.g.: in log [2] we can see how the first OVS agent sets the port UP and one second later, the second OVS agent sets the port down [3]. That will make this test fail.
[1]https://opendev.org/openstack/neutron/src/branch/master/neutron/plugins/ml2/drivers/openvswitch/agent/ovs_neutron_agent.py#L1833-L1834
[2]https://4317fd6ef51d22a4668e-8daa7927094e2ef08ab8f0f95a49fc40.ssl.cf5.rackcdn.com/733250/18/check/neutron-fullstack-with-uwsgi/79ec86c/controller/logs/dsvm-fullstack-logs/TestLegacyL3Agent.test_external_subnet_changed/neutron-server--2020-06-26--12-24-50-479143_log.txt
[3]http://paste.openstack.org/show/795317/
Fix proposed to branch: master
Review: https:/