consensus liveness of the implementation

[Lbqds]

Hi, I'm diving into the implementation of the rapid membership protocol, and I have a question about the consensus liveness. As described below:

There are 9 nodes in the cluster, the FastPaxos need at least 7 votes, let's assume that 3 nodes are unreachable, which means every consensus will fallback to classic paxos. Now:

1. one of the node(let's say node A) start the classic paxos because of fast round timeout, and assume that node A have
   largest node index in the remain 6 nodes
2. node A send phase1a messages to other nodes
3. other nodes handle phase1a message and set the rnd to Rank(2, A.NodeIndex)
4. now node A crashed because of unknown reason, and we have 5 nodes in the cluster
5. the remain 5 nodes will continue the classic paxos, but no one accept the phase1a message at here because of the node have larger rnd, and the consensus get stuck

Am I misunderstanding something?

[lalithsuresh]

@Lbqds good catch! You are correct that the retry forever part isn't implemented yet.

This piece of code here is where a process would need to keep retrying with new rounds until a decision is reached. Right now, it starts just one round and hopes that it will succeed:

rapid/rapid/src/main/java/com/vrg/rapid/FastPaxos.java

Line 189 in 0aef896

void startClassicPaxosRound() {