Comments from Andreas,
"in facet_failover() I still don't agree with the inclusion of sync before triggering failover (which is apparently from LU-924). That is not how failover works in real life and I think it will hide bugs during testing. Also, a change like that does not belong in a patch which describes only changes for DNE testing. Per inspections in the LU-924 patch (which is where it should be landed), there are AFAIK only a very small number of tests that may be affected by this problem (basically only after a client first mounts or reconnects after eviction and then does nothing in the few seconds before another eviction) so I'd rather add explicit sync to those tests than for every test, or add some mechanism to only force sync updates to last_rcvd for connections during testing."