Actually we have notice that Auto vacuum in PG10 keeps vacuuming the master tables and that takes a lot of time and Don't go the child table to remove the dead tuples.
What do the logs say actually got done during these long running autovacuums? Is it feasible to increase the work allowed before autovacuum stops (autovacuum_vacuum_cost_limit) or perhaps increase the number of workers? What is the update/deletes workload balance? That is, would it make sense to decrease the fillfactor on these tables such that you get more HOT (heap only tuple) updates and indexes are less bloated to get better performance that way? How often are you manually vacuuming?
