If there is no user intervention during a cluster split with enabled partition handling there is NO available partition or ONE. It is not possible in this case to have >1 available partitions.
Because of that there is no handling defined for this. As a result the cluster is in an undifined and maybe inconsistent state after merge.
There should be a definition how this situation will be resolved!
- First a manual intervention should be logged in a different way to document it, at the moment the log message is the same no matter whether the state changed by hand or not.
- If there is a majority of nodes (from last stable) this partition should stay active and the other (available) should be erased in the same way as the degraded one.
- If there is a simple majority this should be used
- If there is no majorty ... random?
- In any case a merge of avail-avail should be log a WARN (maybe on each node)
In any case there should be a documented policy to be sure what the ramifications are if there is a manual intervention for the available state.