Raft
Also known as:
Raft Consensus Algorithm
Definition
Raft is a consensus algorithm from 2014 that gives the same fault tolerance as Paxos but is much easier to read and write. It splits the problem into three smaller pieces. Leader election picks a single leader. Log replication copies entries from the leader to followers. Safety rules make sure no two leaders disagree.
Key Takeaways
- Raft elects exactly one leader per term. It uses random election timeouts so two candidates rarely start at once.
- Every client write goes to the leader, which appends it to the log and copies it to followers. The entry is committed once a majority quorum has stored it.
- Followers reject log entries that conflict with their own log. The leader retries until the logs match.
- Raft powers etcd, Consul, CockroachDB, and TiKV, mostly because it is far easier to get right than classic Paxos.
How It Works
- A follower whose election timer fires becomes a candidate, bumps the term, votes for itself, and asks the others for votes.
- If it gets votes from a majority, it becomes leader and sends heartbeats to keep the others quiet.
- The leader appends client commands to its log and ships them to followers using AppendEntries calls.
- Once an entry is on a majority of nodes, the leader marks it committed and tells followers to apply it to their state.
Where It Is Used
- etcd uses Raft to back the Kubernetes control plane.
- CockroachDB and TiDB run thousands of independent Raft groups, one per data range, so consensus scales out.
- Consul and Nomad use Raft for service discovery and orchestrator state.