Consensus Algorithms: From Paxos to Modern Raft Variants

The Consensus Problem

In distributed systems, consensus ensures all nodes agree on a single value, even when some nodes fail. This is the bedrock of replicated state machines, distributed databases, and blockchain networks.

Raft: Understandable Consensus

Raft was designed for understandability. It decomposes consensus into three sub-problems:

Leader Election — One node becomes the leader

Log Replication — The leader replicates entries to followers

Safety — Only nodes with complete logs can become leader

Term 1:    [S1: Leader] → [S2: Follower] → [S3: Follower]
                
           AppendEntries   AppendEntries    AppendEntries
                
           [Log: A,B,C]   [Log: A,B,C]    [Log: A,B,C]

Modern Variants

Multi-Raft partitions state across multiple Raft groups, enabling horizontal scaling. CockroachDB and TiKV use this approach — each data range gets its own Raft group.

Flexible Paxos relaxes quorum requirements, allowing different quorum sizes for different phases. This enables configurations optimized for read-heavy or write-heavy workloads.

Understanding these algorithms is essential for anyone designing systems that must remain correct and available in the face of failures.