Re:VACUUM: Nonremovable rows due to wal sender process

Hello
This is exactly the reason why you need to track the age of the oldest transaction on the primary itself and on every replica that has hot_standby_feedback = on. By default hot_standby_feedback is disabled.

> Is there anything I can do short of shutting down and restarting the primary (production system) that would allow the backend_xmin to move forward?

You need to investigate this replica. Not a primary database. What transactions are in progress? Is it reasonable? Is hot_standby_feedback really needed here and is it reasonable to pay for its impact across the entire cluster?
In my practice, hot_standby_feedback = on is only needed on replicas intended for fast OLTP queries. And where any long requests are prohibited. 

regards, Sergei