Wow, it seems a good find for a potential reason!
I'd totally subscribe to the cause, if it only affected my Atom HCI clusters, who have
had timing issues so bad, even the initial 3 node HCI setup had to be run on a 5 GHz Kaby
Lake to succeed, because my J5005 Atoms evidently were too slow.
But I see the very same effect also on another farm, not using quite the newest generation
of hardware, but even a 28 core Broadwell Xeon, so if timing is the issue, somebody from
QA needs a serious performance review.
Did you get a chance to test the OVA export/import cycle?