exception last prims not consistent about galera

[dirtysalt]

I speculate this problem maybe related to early flawed implementation of pc recovery. (I'm not sure in this case, pc.recovery feature is used or not).

For convenience of expression, we label each node
79dcfbbc A
7cfbf591 B
7da57d09 C
7ccb744c D (the log we are reading belongs to this node)

I guess the whole process would be:

• first three nodes A,B,C join and form pc. we can tell that from each node's node_map structure before crash. B and C node_map has the same last_prim view_id(PRIM,79dcfbbc,2)
• then these three nodes join with D and other four nodes. We can tell that from the first view delivery in log.
• then A loses connection to B and C. I guess A should become non-primary component, BC should still be primary component.
• B, C join with D again. We can tell that from the second view delivery in log.
• I guess A should join with other four nodes, because we can tell that from each node's node_map structure before crash because those nodes have same last_prim view_id(PRIM,79dcfbbc,4). __So the question is why A and other four nodes could become primary component, and I guess the one explanation would be pc.recovery helps that. __

Early implementation of pc recovery:

• rst_view is cleared when pc recovery qualification is meet and after calling send_install, not but when primary component is formed.
• state file is not removed event when graceful shutdown.

[ayurchen]

Yan, if you think new implementation fixes this bug, you can probably close it.