Unexpected FinalizeMembershipChange entries in commited entries after finalize

[haraldng]

I assume from the documentation that begin_membership_change should be called once by the leader and finalize_membership_change should be called once by all nodes. Please let me know if that is not the intended usage :)

Following the documentation of membership_change and example of five_mem_node I am doing the following:

// in function on_ready()
        if let Some(committed_entries) = ready.committed_entries.take() {
            for entry in &committed_entries {
                if entry.data.is_empty() {
                    // From new elected leaders.
                    continue;
                }
                if let EntryType::EntryConfChange = entry.get_entry_type() {
                    // For conf change messages, make them effective.
                    let mut cc = ConfChange::default();
                    cc.merge_from_bytes(&entry.data).unwrap();
                    let change_type = cc.get_change_type();
                    match &change_type {
                        ConfChangeType::BeginMembershipChange => {
                            let reconfig = cc.get_configuration();
                            let start_index = cc.get_start_index();
                            raft_node
                                .raft
                                .begin_membership_change(&cc)
                                .expect("Failed to begin reconfiguration");

                            assert!(raft_node.raft.is_in_membership_change());
                        }
                        ConfChangeType::FinalizeMembershipChange => {
                                raft_node
                                    .raft
                                    .finalize_membership_change(&cc)
                                    .expect("Failed to finalize reconfiguration");
                        }
          }

The function finalize_membership_change is successful the first time but then the FinalizeMembershipChange entries reouccur in the commited entries which causes the program to panic as next_configuration is None. FinalizeMembershipChange entries occur both multiple times in ready.commited_entries and later when on_ready() is called again.

[hicqu]

@haraldng thanks for your report. It's expected that FinalizeMembershipChange occurs multi times, so you needs to handle the error returned by finalize_membership_change by yourself. The reason is FinalizeMembershipChange is appended in commit_apply, so if a peer restarts with such an entry which is not applied, it could append it again.

BTW it's better to not use the joint-consensus implemantation. For now just use normal configuration change is ok. Later we will change the implementation to etcd/raft's style. Please take a look at https://github.com/tikv/raft-rs/pull/317. Would you like to contribute to this?

[haraldng]

@hicqu Thank you for your reply. If I am using AddNode and RemoveNode variants instead, will that be batched as one configuration change? Or will they be handled separately as two configuration changes? How does it affect the performance?