Understanding the Raft Consensus Algorithm: States, Leader Election, Consistency, and Split‑Brain Handling

1: Raft States

Raft clusters have three possible roles for each node: follower (initial state), candidate (temporary role during elections), and leader (the sole node that handles client requests and replicates logs to followers). The term (or election term) is a monotonically increasing integer that all nodes track.

2: Leader Election

When a Raft cluster starts or when the current leader fails, nodes initiate an election. Each node starts as a follower, sets its term to 0, and begins a random election timeout (150‑300 ms). The first node whose timeout expires becomes a candidate, increments its term, votes for itself, and requests votes from other nodes. If it receives votes from a majority, it becomes the leader and begins sending periodic heartbeats (AppendEntries RPC) to maintain authority.

If two candidates time out simultaneously, they may split votes; the election repeats with new random timeouts until a leader is chosen.

Election Scenarios

Leader crash : Followers that miss heartbeats start a new election.

Follower crash : The leader continues sending heartbeats; the follower rejoins as a follower when it recovers.

3: Ensuring Cluster Consistency

Log replication is driven by the leader. When a client submits a command, the leader creates an uncommitted log entry, replicates it to a majority of followers, and waits for acknowledgments. Once a majority have stored the entry, the leader marks it committed, replies to the client, and notifies followers to commit.

The protocol defines several failure cases (leader crash before/after replication, partial replication, etc.) and guarantees that a new leader will always have the most up‑to‑date log, preventing data loss and preserving consistency.

4: Handling Split‑Brain (Network Partition)

During a network partition, separate sub‑clusters may each elect a leader, creating a split‑brain. The partition with the majority of nodes continues to accept writes, while the minority cannot achieve a quorum.

When the network heals, the minority leader steps down (its term is lower), rolls back divergent logs, and synchronizes with the majority leader, restoring a single authoritative leader and consistent state across the cluster.

5: Summary

The article outlines Raft’s node roles, leader election mechanics, log replication for consistency, and the resolution of split‑brain situations, offering a high‑level understanding of how Raft achieves high availability and strong consistency in distributed systems.