Cloudflare is developing Meerkat, a globally consistent control-plane service built on the asynchronous QuePaxa consensus algorithm. Unlike traditional Raft-based systems, Meerkat eliminates reliance on leader nodes and timeouts, allowing all replicas to accept writes independently. This design aims to maintain strong consistency while preventing availability losses during network disruptions, according to Cloudflare engineers James Larisch, Bob Halley, and João Pedro Leite.
Moving Beyond Raft: Addressing Wide-Area Network Constraints
Traditional consensus algorithms such as Raft depend on a designated leader to process writes. If that leader crashes or experiences network degradation, the system enters a period of unavailability while nodes wait for timeouts and elect a successor. In a global network like Cloudflare’s, these latency spikes and timeouts can trigger significant service interruptions.
According to Larisch, Halley, and Leite, the primary limitation of Raft in distributed environments is its dependency on partially synchronous behavior. These systems only make progress if message delays remain low relative to timeout durations. By adopting QuePaxa, the Meerkat project seeks to remove these bottlenecks entirely, ensuring the system remains functional even when message delays fluctuate wildly across the globe.
Did you know?
QuePaxa is an asynchronous consensus algorithm. Unlike Paxos or Raft, it does not rely on timeouts to make progress, which allows it to handle the unpredictable nature of wide-area network communication more effectively.
How the Meerkat Log Operates
The Meerkat service utilizes a log structured as a sequence of slots. Each slot acts as a container for an event. A slot is defined as “decided” once the system reaches a consensus on its contents. Once a value is placed in a slot, the system enforces a strict invariant: no two replicas will ever disagree on the value assigned to that slot.
This structure supports critical infrastructure, including transactional key-value stores and leasing systems. While the service is not intended for general-purpose database applications, it provides a robust foundation for distributed coordination. Cloudflare reports that the project has successfully completed multiple proof-of-concept tests involving up to 50 replicas distributed across its global network.
Performance Trade-offs in Distributed Consensus
Consensus overhead remains a significant challenge for any distributed system. Aniket Ray, a software engineer at KPIT, notes that consensus mechanisms typically introduce 40–60% additional latency in multi-region deployments. Because QuePaxa requires between one and three round trips between an initial proposer and a majority of replicas to finalize a proposal, performance must be carefully managed.
The engineering team at Cloudflare acknowledges these limitations. While QuePaxa offers superior availability, it does not eliminate the physical reality of round-trip costs. The team is currently evaluating whether the performance remains competitive enough for standard operational use compared to existing, leader-based alternatives.
When evaluating distributed systems, consider whether your workload requires strong consistency or if eventual consistency can satisfy your latency requirements. Asynchronous algorithms like QuePaxa are specifically designed for environments where network partition and high latency are expected.
Frequently Asked Questions
- What is the main difference between Raft and QuePaxa?
Raft relies on a single leader and timeouts to maintain consensus, which can cause downtime during network issues. QuePaxa is asynchronous and allows any replica to accept writes, avoiding leader-related availability loss. - Is Meerkat currently in production?
No, Meerkat is not yet deployed to production. Cloudflare has completed multiple proofs of concept with up to 50 global replicas to test its viability. - Can Meerkat be used as a general-purpose database?
No. According to Cloudflare, Meerkat is designed for control-plane services such as key-value stores and leasing systems, not as a general-purpose database.
Stay updated on the latest developments in distributed systems engineering. Subscribe to our newsletter for deep dives into infrastructure technology, or browse our archive for more articles on network consensus and global architecture.
Keep reading