Is it possible to install a snapshot built by the client, outside of the state machine?

[maackle]

First of all, thank you for the great work on openraft, it has been really wonderful to work with :)

I am trying something unusual. I'm trying to use openraft in a situation where the client knows how to create snapshots, but the raft state machine does not. In my use case, essentially, the state machine logic lives in the client (in a JS library), not in the Rust raft implementation.

Thus, I would like clients to be able to instruct the raft leader to purge logs, install a constructed snapshot and propagate that. The client would be responsible for ensuring that the snapshot represents the logs up to that point.

I have no problem getting the leader to install the snapshot and purge its logs, but I am not seeing it replicate the new snapshot to the members. This is what I'm asking about: how can I get the cluster to fully accept this external snapshot as the current state?

I have tried the following, but only the leader updates its state:

        raft.with_state_machine(
            move |s: &mut Arc<StateMachineStore<memstore::TypeConfig>>| {
                async move {
                    s.clone()
                        .install_snapshot(&meta.clone(), snapshot)
                        .await
                        .unwrap();
                    // s.build_snapshot().await.unwrap();
                }
                .boxed()
            },
        )
        .await
        .unwrap()
        .unwrap();

        let trigger = raft.trigger();
        trigger
            .purge_log(smd.last_applied.map(|l| l.index).unwrap_or_default())
            .await
            .unwrap();
        trigger.snapshot().await.unwrap();
        trigger.heartbeat().await.unwrap();

Is it possible?

[github-actions]

👋 Thanks for opening this issue!

Get help or engage by:

• /help : to print help messages.
• /assignme : to assign this issue to you.

[drmingdrmer]

This functionality isn't implemented yet. Currently, snapshot transmission from Leader to Follower only occurs when the Follower requests log entries that have already been purged from the Leader's storage.

Adding an API to manually trigger snapshot transmission would be beneficial. I'll implement this feature.

[maackle]

Thank you for considering this as a feature. In your last message you implied that snapshot transmission could be triggered by making each Follower request purged log entries from the Leader. This could be a good workaround for me for now.

However I don't know how I would make those requests, I only see ways to read the local storage as a Follower, and it seems that all log data I can read depends on the Leader replicating it to me first. So how can I as a Follower ask the Leader for log data to trigger the snapshot propagation?

[drmingdrmer]

Thank you for considering this as a feature. In your last message you implied that snapshot transmission could be triggered by making each Follower request purged log entries from the Leader. This could be a good workaround for me for now.

Yes. If the follower doesn't have log entries that are already purged on the leader, the leader will transmit a snapshot.

But how do you implement this? The approach I can think of is: in your application service layer on the follower, don't pass the AppendEntriesRequest request to the Raft implementation. Instead, return a fabricated conflict response directly to the leader. After several retry rounds, the leader will gradually move its nextIndex cursor backward into the purged range. Once the leader tries to send entries that fall entirely within the purged range, it will trigger the snapshot transmission mechanism.

It's still a very tricky way :(