🌐 Distributed Systems
📄 Article
🗞️ Cloudflare Blog, 2026-11-11 19:54 IST
⏱️ ~15 min
Meerkat is Cloudflare's experimental, globally replicated consensus log (a KV store like etcd, Consul or ZooKeeper) for its internal control plane. The post explains why Raft, Multi-Paxos and VR degrade on volatile WAN links. Destructive interference: competing proposers, and competing Raft candidates during elections, keep blocking each other. Metastable failures: a brief network blip triggers elections that keep themselves going. Fault detection: timeouts mistake a slow leader for a dead one. Cloudflare cites a 2020 switch partial failure that made a consensus-backed dashboard 80× slower for six hours. QuePaxa keeps a leader fast path that takes one round trip (over 95% of Meerkat decisions), then falls back to randomized proposer priorities and proposal adoption, so proposers converge instead of competing and no failure detection is needed to make progress. In Cloudflare's tests, a failed leader dropped throughput to about 65% rather than stalling. Cloudflare is extending QuePaxa from crash-stop to crash-recovery and formally verifying the Rust implementation, with follow-up posts promised.
Why you care: Distributed Systems is your #2 weight. This is a clear, practical explanation of why leader-based consensus has availability limits on a WAN, which is useful for system-design interviews and for reasoning about etcd/KRaft/JetStream behaviour under partial network failure.