Distributed Consensus Protocols — Raft and Paxos — Distributed consensus protocols enable a group of networked nodes to agree on a single value or sequence of values despite node failures and network partitions, forming the foundation of fault-tolerant replicated systems.

Paxos Protocol Fundamentals — The original consensus algorithm defines three roles:

• Proposers — initiate consensus by sending prepare requests with unique proposal numbers, competing to have their proposed values accepted by the cluster
• Acceptors — respond to prepare and accept requests, promising not to accept proposals with lower numbers and eventually accepting the highest-numbered proposal they have seen
• Learners — discover the chosen value by observing which proposal has been accepted by a majority of acceptors, enabling them to apply the decision to their state
• Two-Phase Protocol — the prepare phase establishes a proposal's priority and discovers any previously accepted values, while the accept phase commits the value once a majority agrees

Raft Protocol Design — Raft was designed explicitly for understandability:

• Leader Election — nodes start as followers and transition to candidates after an election timeout, requesting votes from peers and becoming leader upon receiving a majority of votes
• Log Replication — the leader receives client requests, appends entries to its log, and replicates them to followers through AppendEntries RPCs, committing entries once a majority acknowledges
• Term-Based Ordering — monotonically increasing term numbers partition time into leadership periods, with any node rejecting messages from leaders with stale terms
• Safety Guarantees — Raft ensures that committed entries are never lost by requiring leaders to have all committed entries in their logs before being elected

Leader Election Mechanics — Establishing leadership is critical for progress:

• Randomized Timeouts — Raft uses randomized election timeouts to reduce the probability of split votes where multiple candidates compete simultaneously
• Pre-Vote Extension — a pre-vote phase allows candidates to check if they would win before incrementing their term, preventing unnecessary term inflation from network-partitioned nodes
• Heartbeat Mechanism — leaders send periodic heartbeat messages to maintain authority and prevent followers from starting unnecessary elections
• Leadership Transfer — graceful leadership transfer allows a leader to hand off responsibility to a specific follower, useful for planned maintenance or load balancing

Fault Tolerance and Safety Properties — Consensus protocols guarantee correctness under failures:

• Majority Quorums — requiring agreement from a majority of nodes ensures that any two quorums overlap, preventing conflicting decisions even during network partitions
• Persistence Requirements — nodes must persist their current term, voted-for state, and log entries to stable storage before responding, ensuring correctness across crash-recovery scenarios
• Linearizability — properly implemented consensus provides linearizable semantics where operations appear to execute atomically at some point between invocation and response
• Liveness Considerations — consensus protocols guarantee safety always but can only guarantee progress when a majority of nodes are operational and can communicate

Raft and Paxos underpin virtually all modern distributed databases, coordination services, and replicated state machines, with Raft's clarity making it the preferred choice for new implementations while Paxos variants continue to power large-scale production systems.