Comment by dranka

A variant of no 3 is used for some large scale telecom equipment supporting multi millions of attached users with serious uptime requirements. State is distributed and handled locally and replicated to a sibling node for fallback protection in case a node goes down. The replication is dynamic so new sibling relationships are established when nodes come and go. There is also some more persistent state (like billing information) that is used to recover in case of total node failure but a lot of the transient state can be rebuilt when users are forced to reattach