This is bug is very similar to https://bugzilla.redhat.com/show_bug.cgi?id=2160767, with some differences.

In this case, the issue has been reproduced using the NB DB driver.

It can be reproduced with the test `test_dvr_vip_failover_basic`:
https://opendev.org/x/whitebox-neutron-tempest-plugin/src/commit/182582a0d55d9ca131a459962616a092f05a16ed/whitebox_neutron_tempest_plugin/tests/scenario/test_dvr_ovn.py#L888

This test creates two VMs (VM0 and VM1) on different computes (C0 and C1). Then it creates a port (VIP) and adds a FIP to that port (VIP_FIP). The VIP's IP is configured on both VM ports as an allowed-address-pair.

1. The test adds the VIP's IP to VM0  (ip a add <IP> dev eth0) and sends an arping.
2. Then, it checks connectivity to the VIP_FIP and it checks traffic is routed through C0.
3. The test removes the VIP's IP from VM0, adds it to VM1 and sends an arping from there.
4. Finally, it checks connectivity to the VIP_FIP and it checks traffic is routed through C1.

Steps 3 and 4 sometimes fail because the VIP_FIP is not immediately exposed from C0 and C1 respectively, but they are exposed when the ovn-bgp-agent's periodic sync is run.

I will provide some logs in the next comments.