With a very brief look, it appears that maybe some flows are missing. I
only checked on the control node and compared against a job that was
passing [0]. the failing job is here [1]

The FIP in the failing job is 10.10.10.10 and it's mac is fa:16:3e:87:dd:47.
That mac has one flow in table 21 on the control node and no other flows
using that mac. In the passing case, the FIP is 10.10.10.12 and has a mac
of fa:16:3e:ed:89:c3. That mac has flows in table 21, 50 and 51.

Unsure if that is the root cause or not, but the fact that these jobs are
running identical test cases and have differing flow tables pertaining to
the FIP seems to be some clue.

[0] https://logs.opendaylight.org/releng/jenkins092/netvirt-csit-1node-openstack-newton-nodl-v2-upstream-stateful-snat-conntrack-carbon/110/log.html.gz
[1] https://logs.opendaylight.org/releng/jenkins092/netvirt-csit-1node-openstack-newton-nodl-v2-upstream-stateful-snat-conntrack-carbon/111/log.html.gz