Cloudflare chooses leaderless QuePaxa for its Meerkat control plane
Original title: Choosing QuePaxa, an unusual consensus algorithm
A 2020 partial switch failure left a leader-based Cloudflare system unable to take writes and about 80 times slower for six hours.
Cloudflare says its experimental Meerkat service, which keeps an ordered log for an internal global control plane, uses a new consensus algorithm called QuePaxa instead of Raft, Multi-Paxos, or Viewstamped Replication. Those leader-based protocols can stall on wide-area links when the leader is slow or disconnected even if replicas have not crashed. The post explains destructive interference in Paxos and why Meerkat needs progress during latency spikes.
- Meerkat replicas agree on a log that can back a key-value store of gets and puts.
- Comparable systems named are Consul, etcd, and Apache ZooKeeper.
- QuePaxa is meant to make progress without a single leader.
- The post says the 2020 incident lost no data but blocked writes until the network issue was fixed.
- Organizations
- Cloudflare
From the original article
14 min read at blog.cloudflare.com
Cloudflare is building an experimental consensus service using a never-before-deployed consensus algorithm, QuePaxa. In this post we explain how QuePaxa works and why we chose it.
By João Pedro Leite · blog.cloudflare.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.