Skip to content

CONCEPT Cited by 1 source

Leaderless consensus

Leaderless consensus is a consensus protocol design where no single replica is required to drive progress. Any replica can propose and drive a decision to completion — a leader may exist for efficiency but the system never becomes unavailable due to a single replica (the leader) being down, degraded, or unreachable.

Definition

In a leaderless consensus protocol:

  1. Any replica can accept client writes and drive consensus for the current slot
  2. No timeout-triggered leader election blocks progress
  3. Concurrent proposals cooperate rather than destructively interfering (contrast with Raft's competing leadership campaigns)
  4. A leader, if present, is an optimization — it can decide in fewer round-trips (1 vs 3+) but is not on the critical path for availability

Motivation

Leader-based protocols (Raft, Multi-Paxos) suffer two problems on wide-area networks:

  1. Leader failure → total write unavailability until a new leader is elected via timeout
  2. Timeout tuning is intractable — too short causes false elections under normal WAN jitter; too long means slow recovery after real failures; concurrent campaigns destructively interfere

These problems have caused real production incidents at Cloudflare.

Trade-offs

Property Leader-based Leaderless
Latency (best case) 1 RT (leader) 1 RT (if leader used)
Latency (non-leader write) Forward to leader + 1 RT 3+ RT
Availability under leader failure Blocked until election Unaffected
Complexity Lower (easier to reason about) Higher (subtle protocol)
Concurrent write throughput Single bottleneck Distributed

Seen in

Last updated · 573 distilled / 1,747 read