
Cloudflare has introduced Meerkat, an internal experimental consensus service built to manage control-plane state across its global network of more than 330 data centers. The project, developed over the past year by Cloudflare’s Research team, addresses a specific and recurring problem: leader-based consensus algorithms failing under the unpredictable conditions of a wide-area network.
The Problem with Leader-Based Consensus
Many distributed systems rely on consensus algorithms such as Raft to coordinate reads and writes across replicas. In Raft, a single elected leader is the only replica permitted to accept writes. If that leader crashes or becomes unreachable, the system stalls until a timeout expires and a new leader is elected. Cloudflare engineers note that configuring timeout values is difficult in networks with highly variable latencies, and the company has experienced multiple real incidents caused by unavailable leaders in consensus-driven systems.
QuePaxa as the Foundation
Meerkat is powered by QuePaxa, a consensus algorithm published in 2023 by researchers at EPFL. The key difference from Raft is that QuePaxa allows all replicas to perform writes at any time, and the system never halts progress because of a timeout. Cloudflare believes this will be the first industrial deployment of QuePaxa at global scale.
What Meerkat Is Built to Do
Meerkat provides a consensus log on top of which Cloudflare layers applications. Two examples already in development are a transactional key-value store and a leasing system. The kinds of control-plane state Meerkat is intended to manage include:
- Placement information, such as where a specific AI model instance is stored across the network.
- Leadership information, meaning which machine is currently authorised to write to a replicated database.
The system is designed to provide linearizability, the strongest consistency guarantee available, meaning all reads after a write will reflect that write. Cloudflare also notes Meerkat’s key-value store provides serializability, a property they plan to cover in a future post.
Current Status and Scope
Meerkat is described as experimental and still in active development. For now, it is being kept internal and is scoped to managing small pieces of control-plane state, such as leadership coordination for replicated databases. Cloudflare has indicated this post is the first in a series covering the project.
For hosting and infrastructure operators, Meerkat represents a practical bet that leaderless consensus can outperform timeout-based approaches in environments where network conditions are unpredictable and global availability is non-negotiable.