Cloudflare picks QuePaxa consensus algorithm for Meerkat
Cloudflare explains why it chose QuePaxa for Meerkat, its experimental consensus service. QuePaxa avoids leader dependence and kept deciding during a leader failure, with about 65 percent of the normal rate.
By JetherVerse News

Cloudflare published a post on November 11, 2026 explaining why Cloudflare chose QuePaxa, a consensus algorithm, for Meerkat. Meerkat is an experimental distributed consensus service for Cloudflare's internal global control plane. Meerkat keeps an ordered log of get and put operations and the log backs a key-value store.
QuePaxa comes from a 2023 paper by Tennage, Basescu and colleagues. The algorithm is randomized and works in rounds of four phases. Meerkat runs across continents over links with variable latency and connectivity. Cloudflare wanted a system to keep making progress during latency spikes, broken connections and leader failures. In 2020, a partial network switch failure slowed a consensus backed dashboard by 80 times for six hours and blocked writes.
Raft and Multi-Paxos depend on a leader. If the leader fails or slows, the system stalls until a new one is elected. Raft style systems detect failure with timeouts, often set to 5 to 10 times the network round trip time. QuePaxa treats the leader as optional and avoids fault detection. Random priorities let proposers converge on the highest priority proposal instead of competing. Cloudflare gives an example. A 5 ms hedge timeout on a network with 100 ms latency freezes Raft, but in QuePaxa the timeout only adds proposers.
The numbers come from Meerkat. More than 95 percent of decisions finish in the first phase, which costs one round trip. When decisions need extra phases, throughput falls to at most a third of normal. During a leader failure, the decision rate fell to about 65 percent and recovered about a minute later, after the leader was replaced. The system never halted.
Most business owners never run a consensus system. Developers who rely on replicated databases and queues across regions do, through their vendors. The design choice behind those tools decides how your service behaves in a network fault. Ask your vendor what happens when the leader slows down, not only when the leader dies.
For client systems required to stay online, failure behavior deserves as much attention as speed.
Source: Cloudflare Blog
Get the news in your inbox
The tech stories that matter from JetherVerse News, with the source on every one.

