PATTERN Cited by 1 source
Leaderless consensus for WAN¶
Pattern¶
Choose a leaderless consensus protocol (e.g., QuePaxa) over leader-based alternatives (e.g., Raft) when the deployment environment is a wide-area network with unpredictable latencies and no tolerance for single-point-of-failure unavailability.
When to apply¶
- Replicas span multiple geographic regions connected by the public Internet
- Network latencies vary wildly and unpredictably
- Any single replica failure must not cause write unavailability (no leader dependency)
- The workload is write-infrequent, consistency-critical control-plane data
- You have experienced production incidents from leader unavailability or election storms
Trade-offs¶
Gains: - System never becomes unavailable due to a single replica failure - No leader elections that degrade performance during transitions - Concurrent proposals cooperate rather than destructively interfering - No timeout tuning headaches
Costs: - Higher per-write latency for non-leader proposals (3+ round-trips vs 1) - More complex protocol implementation and formal verification - Less mature ecosystem and operational tooling compared to Raft
Cloudflare instance¶
Cloudflare chose QuePaxa for Meerkat after experiencing multiple production incidents from unavailable leaders in Raft-based systems. Their 330+ data center WAN has latencies that "can and do vary wildly," making timeout configuration intractable.
Seen in¶
- sources/2026-07-08-cloudflare-introducing-meerkat-global-consensus — Cloudflare's architectural decision rationale
Related¶
- concepts/leaderless-consensus — the underlying concept
- concepts/wide-area-consensus — the deployment environment
- concepts/leader-timeout-problem — the failure mode this pattern addresses