skiff

An embedded Raft consensus library backed by sled. Skiff lets you add replicated, persistent key-value storage directly to your Rust application without running a separate consensus service.

Features

• Leader election and re-election
• Log replication with majority-commit semantics
• Hard state persistence (survives restarts)
• Dynamic cluster membership (add_server / remove_server)
• Follower request forwarding (connect to any node)
• Change subscriptions (watch prefix for streaming updates)

Not yet implemented: log compaction / snapshotting.

Installation

[dependencies]
skiff-rs = "0.1"
tokio = { version = "1", features = ["full"] }

Note: skiff-rs requires protoc to be installed at build time. On Debian/Ubuntu: sudo apt install protobuf-compiler On macOS: brew install protobuf

Quick start

use skiff_rs::{Builder, Client};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Build and start a single-node cluster.
    let node = Builder::new()
        .set_dir("/tmp/my-skiff-node")
        .bind("127.0.0.1".parse()?)
        .build()?;

    let node_ref = node.clone();
    tokio::spawn(async move { node_ref.start().await });

    // Block until a leader is elected before connecting a client.
    node.wait_for_leader(std::time::Duration::from_secs(2)).await?;

    // Connect a client and perform some operations.
    let mut client = Client::new(vec!["127.0.0.1".parse()?]);
    client.connect().await?;

    client.insert("greeting", "hello world").await?;
    let value: Option<String> = client.get("greeting").await?;
    println!("{:?}", value); // Some("hello world")

    Ok(())
}

Multi-node cluster

Pass existing node addresses to join_cluster when constructing additional nodes. The new node contacts a peer and registers itself via the Raft add_server RPC.

use skiff_rs::Builder;
use std::time::Duration;

// Node 1 — bootstraps a new single-node cluster.
let node1 = Builder::new()
    .set_dir("/tmp/node1")
    .bind("127.0.0.1".parse()?)
    .build()?;

// Node 2 — joins the cluster through node 1.
let node2 = Builder::new()
    .set_dir("/tmp/node2")
    .bind("127.0.0.2".parse()?)
    .join_cluster(vec!["127.0.0.1".parse()?])
    .build()?;

let node1_ref = node1.clone();
tokio::spawn(async move { node1_ref.start().await });

// Wait for node1 to elect itself leader before node2 tries to join.
node1.wait_for_leader(Duration::from_secs(2)).await?;

tokio::spawn(async move { node2.start().await });

Hierarchical keys

Keys use / as a path separator, providing a simple namespace hierarchy:

client.insert("users/alice", alice_data).await?;
client.insert("users/bob", bob_data).await?;

// List all keys under "users/"
let keys = client.list_keys("users/").await?;

// Get all top-level prefixes
let prefixes = client.get_prefixes().await?; // ["users"]

Change subscriptions

let mut sub = client.watch("users/").await?;
loop {
    let (key, value): (String, MyType) = sub.recv().await?;
    println!("updated: {} = {:?}", key, value);
}

Shutdown

Call shutdown() before dropping a node to allow background tasks and the sled database lock to be released cleanly:

node.shutdown();
drop(node);

Architecture

Each Skiff node runs two background tasks:

• Cluster-join task — on startup, contacts a peer and calls add_server to register the node if the cluster has more than one known member.
• Election manager — drives the Raft state machine: sends heartbeats as leader (every 75 ms) or fires a randomised election timeout as follower (150–300 ms).

All inter-node communication uses gRPC (via tonic). Persistent state is stored in sled named trees:

Status

v0.1 — suitable for experimentation and projects that can tolerate the missing features noted above. The on-disk format is not yet considered stable.