How does Raft achieve consensus compared to classic Multi-Paxos and Viewstamped Replication?

All three protocols solve the problem of State Machine Replication (SMR) over asynchronous, unreliable networks with crash-recovery failures, but they structure their safety invariants differently. Multi-Paxos (Lamport) separates agreement on individual log slots from log ordering: any node can propose an entry for any slot by executing Phase 1 (Prepare/Promise with ballot number n) to learn previously accepted values, followed by Phase 2 (Accept/Accepted). In steady state, a designated master runs Phase 1 once for all future slots, amortizing agreement to a single round-trip (Phase 2). However, log entries can commit out of order, creating "holes" that must be filled before execution. In contrast, Raft (Ongaro & Ousterhout) enforces a strict Leader Completeness Invariant: log entries can only flow in one direction (Leader to Followers), and a candidate can only be elected if its log is at least as up-to-date as any quorum majority (comparing last term, then last index). Raft logs never have holes, and committed entries are strictly contiguous. Viewstamped Replication (VR Revisited, Oki & Liskov / Cowling) is conceptually isomorphic to Raft, using Views instead of Terms and View Changes with strict primary-backup log syncing.

What is Raft Joint Consensus and why is single-step cluster membership change dangerous without constraints?

When reconfiguring a running cluster (adding or removing nodes from configuration C_old to C_new), naive direct transition can cause split-brain: because nodes switch configurations at different times, there exists a window where a subset of nodes in C_old forms a majority (e.g. 2 of 3 in a 3-node cluster) and a disjoint subset in C_new forms a separate majority (e.g. 2 of 3 in an added 3-node group), electing two conflicting leaders simultaneously. To prevent this, Raft specifies Joint Consensus: the leader proposes an intermediate configuration entry C_{old,new}. In this joint configuration, log commitment and leader election require two independent quorums simultaneously: a majority of C_old AND a majority of C_new. Once C_{old,new} is committed across both majorities, the leader commits C_new, which only requires a majority of C_new. Alternatively, Ongaro proved in his dissertation that single-server membership changes (adding or removing exactly one node at a time) guarantee overlapping majorities without joint consensus, provided transitions do not overlap.

How do Leader Leases prevent stale reads in consensus protocols without requiring round-trip quorum reads?

In standard Raft, read requests (queries) must be routed to the leader, but the leader cannot blindly return local state because it might have been partitioned off while a new leader was elected in a different component. Standard linearizable reads require the leader to send a round of Heartbeat/ReadIndex RPCs to a quorum to verify it is still the legitimate leader. To avoid this network round-trip overhead on every read, systems implement Leader Leases: when followers respond to a heartbeat, they grant the leader an exclusive lease for duration T_lease during which they promise not to elect a new leader. As long as the leader local clock has elapsed less than T_lease minus bounded clock drift, the leader can serve reads directly from its local state machine. However, if clock drift exceeds expectations, leader leases can violate linearizability, which is why TrueTime (Google Spanner) or hybrid logical clocks (CockroachDB) are paired with lease renewal.

What is the exact Quorum Intersection Invariant in distributed consensus?

The foundational mathematical requirement of all Paxos/Raft consensus algorithms is the Quorum Intersection Property: in a cluster of N nodes, any two quorums Q1 and Q2 must intersect by at least one node: Q1 ∩ Q2 ≠ ∅. For standard majority quorums where |Q| = ⌊N/2⌋ + 1, two quorums of size (N+1)/2 in an odd-sized cluster (or (N/2)+1 in an even-sized cluster) necessarily share at least (⌊N/2⌋ + 1) + (⌊N/2⌋ + 1) - N ≥ 1 node. In weighted or flexible consensus (Flexible Paxos by Howard et al.), quorums do not need to be simple majorities: as long as every Phase 2 / Replication Quorum Q_write intersects with every Phase 1 / Election Quorum Q_read (Q_read ∩ Q_write ≠ ∅), safety is preserved, enabling small write quorums (e.g. 2 nodes) at the cost of larger election quorums.

Distributed Consensus: Multi-Paxos, Viewstamped Replication & Raft Joint Consensus State Machine Studio

Verified Specification | Standardized Formulas | Instant Precision

Secure & Private (Zero Data Retention) Free Access • No Sign-Up

Distributed Consensus & State Machine Replication Workbench

Explore Multi-Paxos, Viewstamped Replication, and Raft invariants: step through proposal commits, simulate leader failure & split-brain partition, and model Joint Consensus membership reconfigurations.

Term 1

Current Term / Ballot

Node 1

Elected Leader / Primary

Index 3

Quorum Commit Index

3 of 5

Majority Quorum Threshold

C_old (Standard)

Membership Configuration

Cluster Node Replicas & Log State

Green = Committed in Quorum | Amber = Replicating in Transit

Quorum Intersection & Fault Tolerance Theorem

Safety Axiom: Any two majorities Q1 and Q2 intersect by at least one live replica:

|Q| = ⌊N / 2⌋ + 1 = ⌊5 / 2⌋ + 1 = 3 nodes
Max Tolerable Crash Faults (f): f = ⌊(N - 1) / 2⌋ = 2 nodes

With 5 nodes, the cluster withstands 2 simultaneous node crashes or partitions. A minority partition of 2 nodes CANNOT elect a leader or commit entries.

[System] Consensus cluster initialized with 5 nodes in Term 1. Node 1 is leader.

Distributed Consensus & State Machine Replication Workbench

Cluster Node Replicas & Log State

Quorum Intersection & Fault Tolerance Theorem

Production Consensus State Machine Code

🔍 Companion Workbenches & Verification Tools

WebAssembly Linear Memory, Heap Paging & Detached Buffer Architecture Studio

Linux io_uring Advanced: SQPOLL, Registered Buffers & Polled I/O Studio

Database Query Optimizer: Cascades & Volcano Execution Engine Studio

Linux Epoll, Edge-Triggered & Event Loop Architecture Studio

Why Smart Users Are Switching to This Free Toolkit

The Browser Feature Most People Never Use

Stop Paying for Software You Can Get Free

File Conversion Tricks That Save Hours Every Week

This Free Tool Replaced 5 Paid Apps for Our Team

What Your Browser Can Do Without Any Extensions