Skip to content- Sources: primary
- Summary: Cloudflare is building Meerkat, an experimental consensus service for its global internal control plane, on QuePaxa, an algorithm it says has not been deployed before. The post argues that leader-based protocols stall when the leader is slow or cut off, and that Raft elections can interfere with each other and turn a short network fault into a metastable failure. It cites a 2020 partial switch failure that made a consensus-backed dashboard 80 times slower for six hours.
- Why it matters: Teams running etcd, Consul or ZooKeeper across regions get a concrete account of how timeout-based leader detection fails on volatile links.
- Background: Consensus algorithms let replicas agree on an ordered log so their state never diverges. Raft, Multi-Paxos and Viewstamped Replication route all proposals through one leader to avoid competing proposers, which is fast on a healthy network but needs failure detection and re-election when the leader looks unreachable. Over wide-area links, false suspicions from latency spikes can trigger repeated elections that block progress even though no replica crashed.