Cloudflare's Meerkat: Global Consensus Without Raft's Pain

Beyond Raft: Meerkat's Global Consensus

Cloudflare's introduction of Meerkat, powered by the QuePaxa consensus algorithm, presents a compelling solution for achieving strong consistency and high availability in wide-area networks, directly addressing limitations of traditional Raft-based systems. The core innovation lies in QuePaxa's ability to allow all replicas to perform writes concurrently and avoid leader-election timeouts, which have historically plagued distributed systems operating across unpredictable internet conditions. By layering applications like a transactional key-value store atop Meerkat's consensus log, Cloudflare aims to simplify development for its numerous internal services that require robust control-plane state management across its global data center footprint. The emphasis on linearizability for developers is a significant benefit, abstracting away complex distributed system behaviors and allowing them to reason about data as if it were in a single-threaded machine.

However, as an experimental service, Meerkat's primary limitation is its current internal-only status. While the promise of global-scale deployment is evident, the practical implications for external developers and the broader ecosystem are yet to be seen. The article acknowledges that Meerkat does not handle Byzantine faults, which is a standard limitation for many consensus algorithms but worth noting for highly sensitive applications. Furthermore, while QuePaxa's avoidance of timeouts is a strength, the article doesn't deeply explore the potential trade-offs in terms of complexity or the specific mechanisms it employs to ensure progress and safety under extreme network partitions beyond a simple majority. The long-term maintainability and operational characteristics of QuePaxa at Cloudflare's scale will be crucial to monitor.

Key Points

• Cloudflare introduces Meerkat, an experimental consensus service utilizing the QuePaxa algorithm.
• Meerkat aims to provide strong consistency (linearizability) and high availability for internal control-plane state management across its global data centers.
• QuePaxa differs from Raft by allowing all replicas to perform writes concurrently and avoiding leader-election timeouts, making it better suited for unpredictable wide-area networks.
• The system abstracts distributed complexity, allowing developers to reason about data as in a single-threaded machine.
• Meerkat's architecture involves a distributed log of events, with applications building state from these events.
• The service guarantees that reads will always see up-to-date information after writes, even with requests hitting different replicas globally.
• Current deployment is internal-only, focusing on small pieces of control plane state.

---

📖 Source: Introducing Meerkat: an experiment in global consensus