For the migration test failures Jakub has already created this etherpad https://etherpad.openstack.org/p/neutron-dvr-multinode-scenario-gate-failures
My analysis is this -
DVR and HA migration tempest scenario tests are failing(or passing) intermittently. In the existing tests, immediately after the port update API is returned we are trying ssh connectivity, without checking the dependent resources (like below) created or updated properly.
1) new interfaces are created
2) existing interfaces updated
3) interfaces bound to agents
4) interfaces status updated
5) agents creates namespaces etc
For example, during DVR to HA migration, as soon as the router update api is returned, ssh test might try to use old data plane created with DVR router, as agents might have not synced(removed namespaces, ovs flows and ip routes) with server. If the ssh reply packets arrived back before the old data plane is removed, then ssh can be succesful. If this data path is reconstructed(because of the migration) before the packet arrived, then ssh can fail. Though ssh can retry, it may use existing conection track and try to follow the same old data path(just my assumption)
When I updated tests to check for the dependent resources before trying for ssh, tests are passing reliably. So we can have these checks before we try for ssh connectivity.
Fix proposed to branch: master
Review: https:/