The problem is that when they don't talk to each other that certain data can remain stale. Especially roles.

I have seen on some systems that the bus will fail and then a few seconds later it works fine. In the latest 7202 testfix and in 7502 and later when this condition has been detected the server will clear its caches. This way it will then sync up with the database instead of using in-memory cache.

Consistent failure is a different problem. In those cases other servers won't see changes to roles such as those assigned to All users because those are kept permanent in an in-memory cache. If it never sees a clear cache request it doesn't change.

Just looking at a role is not using any cache information so that is always going against the database. What is cached is the roles within logged in users. You could change the role and if the bus is not working the logged in users on another server in the cluster would not see the change and they would still have the older entitlements.

So intermittent short term failures are very tolerable. What is not tolerable is a long term failure.