For RHOSO control plane, we are running ovs-vswitchd inside pods. And using net-attach-def we attach secondary NICs to these pods which are used for creating genev tunnel(genev_sys_6081).

When Centralized Floating IPs are used, we see connectivity loss[1] to the FIPs when these OVS pods are updated or restarted.

This is happening as when pod is deleted/recreated network namespaces bound to the pod also gets deleted along with tunnel interfaces. This triggers router ports to fail over and thus some loss to FIP traffic.

When there is only one gateway chassis connectivity is restored when the ovs pods are full restarted(tunnels are recreated).

In previous OSP release ovs-vswitchd used to run on host itself so tunnel interfaces were not touched during update/restarts.

This spike is to look for options if and how this can be avoided.

Running pods with HostNetwork=true could have helped but with that it will conflict with ovs/ovn running on host for OCP workloads.

So if there is some way with which we can run multiple ovs on same namespace that could help here. or if option of using pre created network namespace can be used for pods so those don't get removed with recreate, or if netdev/dpdk datapath is an option or something else?

[1] https://issues.redhat.com/browse/OSPRH-10821