Currently, messages sent from one cluster to another (e.g. A:sfc -> B:nyc) are unreliable; they are not retransmitted when lost.

The reason is that RELAY3 is the top protocol; retransmission protocols such as UNICAST3 are below it.

If we moved RELAY3 further down, e.g. below UNICAST3, these messages would get retransmitted.

We would probably not want to do this for multicast messages, as NAKACK2 also requires STABLE, and doing this across clusters would be slow (but this should also be investigated!)...

Example for unicast messages:

• A:sfc sends message M to B:nyc
• The site master of sfc crashes
• A new site master of sfc is picked
• UNICAST3 on A retransmits M
• M is sent via the new site master to B:nyc

When site nyc goes down, retransmission should stop. RELAY3 probably needs to tell UNICAST3 to stop retransmission of M.

It should be able to enable/disable this behavior, either via a message flag or an attribute.