#cluster #distributed-system #raft #consensus

almost-raft

An implementation of Raft leader election

3 unstable releases

0.3.0 Aug 10, 2021
0.2.1 Apr 18, 2021
0.2.0 Apr 17, 2021

#881 in Network programming

32 downloads per month
Used in cluster-mode

MIT license

40KB
832 lines

Build Crates.io MIT licensed

almost-raft

Consensus or agreeing on some value is a fundamental issue in a distributed system. While there are algorithms like Paxos exists since long back, the complexity of those make implementation complicated.

So Raft was designed to solve the problem while keeping the algorithm understandable.

Raft tackles the problem in two steps -

  • Leader Election - Elect a node as a leader on startup or when the existing one fails
  • Log Replication - Maintain the log consistency among nodes

For more on Raft https://raft.github.io.

Why almost-raft

There is already an implementation of Raft in Rust by awesome people at tikv. If the complete raft is the requirement, check out the tikv/raft-rs.

almost-raft was written with a few things in mind -

  • Personal necessity, of course
  • Minimalism
  • Offloading the inter-node communication mechanism to implementing crates
  • Users don't need to know or care about the inner workings of Raft

almost-raft handles only the first step of the algorithm, the election, hence the name. It's up to the developer on how to handle the log replication, if necessary.

Usage

almost-raft uses a closed loop, the only way to communicate is to use mpsc channel and control messages.

First step is to implement trait Node. For example - a simple node that uses mpsc channel to communicate with others

use tokio::sync::mpsc::Sender;
use almost_raft::{Message, Node};
#[derive(Debug, Clone)]
struct NodeMPSC {
    id: String,
    sender: Sender<Message<NodeMPSC>>,
}

#[async_trait]
impl Node for NodeMPSC {
    type NodeType = NodeMPSC;
    async fn send_message(&self, msg: Message<Self::NodeType>) {
        self.sender.send(msg).await;
    }

    fn node_id(&self) -> &String {
        &self.id
    }
}

To initiate RaftElectionState

let (heartbeat_interval, message_timeout, timeout, max_node, min_node) =
    (1000, 20, 5000, 5, 3);
let (tx, mut from_raft) = mpsc::channel(10);
let self_id = uuid::Uuid::new_v4().to_string();
let nodes = vec![]; // we'll add node later
let (state, tx_to_raft) = RaftElectionState::init(
    self_id,
    timeout,
    heartbeat_interval,
    message_timeout,
    nodes,
    tx.clone(),
    max_node,
    min_node,
);

Now we can start the election process using the state. But this will not necessarily start the election, it'll wait as long as there isn't enough node (min_node).

tokio::spawn(raft_election(state));

Let's add nodes

let (tx,rx) = mpsc::channel(10);
tx_to_raft
    .send(Message::ControlAddNode(NodeMPSC {
        id: uuid::Uuid::new_v4().to_string(),
        sender: tx,
    }))
    .await;

Raft will notify through mpsc channel if there's any change in leadership. To receive the event

// let (tx, mut from_raft) = mpsc::channel(10);
// tx was used to initialize RaftElectionState
from_raft.recv().await;

Dependencies

~3.5–5MB
~92K SLoC