en.Wedoany.com Reported - Cloudflare has launched a consensus service called Meerkat, helping services across more than 330 global data centers read and modify a consistent control plane state, allowing the system to remain writable even when some data centers or links fail.

Due to the internet's susceptibility to outages and queue delays, building a globally consistent and available data system poses challenges. The Meerkat service employs a consensus algorithm called QuePaxa, which, similar to other algorithms of its kind, enables a group of machines to agree on the same sequence of values as long as a majority remain operational and can communicate with each other.

QuePaxa does not require a single authoritative leader to ensure progress. During what Cloudflare calls the "tyranny of timeouts," any replica or non-leader can continuously perform write operations. Architecturally, developers request a set of Meerkat replicas and specify data center locations, with Meerkat handling placement. Clients send application requests to any replica, which converts the request into events in a shared consensus log and returns a response. Each replica maintains an identical order of decided log entries, though some replicas may temporarily lag. Applications built on Meerkat, such as key-value stores or lease systems, read the log and construct their own state accordingly.

Cloudflare explains that without a required leader, the system does not become unavailable or degraded due to a single replica, nor does it require leader elections. Proposals made simultaneously by different replicas constructively interfere with each other, which is ideal for Cloudflare's network where latency can vary dramatically. The company claims to have run multiple proof-of-concept tests involving up to 50 globally distributed replicas, where leaders repeatedly failed, but the cluster continued operating without an increase in error rates.

Meerkat is not yet in production and remains an internal-only service. The company emphasizes its limitations, noting it is not designed for creating general-purpose data systems like databases. All consensus algorithms come with costs; the latency for proposal decisions is proportional to the delay among the majority of replicas. If replicas are far apart, latency increases, which is unavoidable.