We are probably subject to the issue described here:
http://docs-draft.openstack.org/19/263819/2/check/gate-neutron-specs-docs/1699045//doc/build/html/specs/mitaka/l2-agent-extensions-agent-api.html#problem-description
"One major use case where extensions may need to coordinate with the agent that is running them is the Open vSwitch agent graceful restart feature added in Liberty cycle. For this feature, a per-session agent id is used to distinguish flows that belong to the current agent session from those that belong to the previous one. The problem starts when an extension wants to inject its own flows into the switch. Since those extensions don’t have a way to determine what the current agent id is, they cannot mark their new flows with it, which results in the agent dropping all their flows on restart assuming they are stale."
It is likely that flows to direct traffic to/from patch ports are impacted by this, and dropped on agent restart, and then rebuilt when the bagpipe direct sends RPCs after port is set active by the agent.
This issue remains to be investigated.
WIP in https:/