TL;DR
Cloudflare Research has introduced Meerkat, an experimental internal consensus service built on the QuePaxa algorithm. The company says it is aimed at small pieces of strongly consistent control-plane state across its global network, but it is not in production and no customer release has been announced.
Cloudflare Research has introduced Meerkat, an experimental internal consensus service powered by QuePaxa, in a July 8 Cloudflare Blog post. The project matters because Cloudflare says it is designed to keep control-plane state consistent across its 330-plus data centers even when machines, data centers or network links fail.
Cloudflare said Meerkat is not in production and will remain internal-only for the near future. The service is being built for small, high-value pieces of control-plane data, such as database leadership or resource-placement information, rather than as a general-purpose database.
The company said Meerkat uses a shared consensus log: replicas agree on the same ordered sequence of events, and applications such as a transactional key-value store or leasing system build state from that log. Cloudflare says each replica can receive reads and writes, and developers can choose which data centers may host replicas.
Cloudflare is basing the service on QuePaxa, a consensus algorithm published in 2023. According to Cloudflare, QuePaxa differs from common Raft-based systems because all replicas can help drive writes; a leader can improve speed, but is not required for progress if another healthy replica can reach a majority.
Leader Bottlenecks Drive Meerkat Bet
The project targets a hard operational problem for large distributed networks: keeping state strongly consistent without making one leader machine the temporary choke point for writes. Cloudflare said it has seen incidents tied to unavailable leaders in existing consensus-driven systems.
If Meerkat performs as Cloudflare expects, it could give internal services a way to keep linearizable control data available during some failures that disrupt leader-based systems. That would matter most for services where a stale or split view of who owns a lock, where a resource lives, or which node may write to a database can cause broader reliability problems.
Think Distributed Systems: Mental models of reliable and scalable software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Raft Limits Shaped The Project
Cloudflare framed Meerkat as a response to the limits of leader-based consensus in wide-area networks. In systems such as Raft, the leader handles writes; if it fails or becomes slow, writes can block until other replicas detect the problem and choose a new leader.
The QuePaxa paper says the protocol was designed to avoid dependence on timeouts for liveness while retaining efficient normal operation. Cloudflare said tests of the underlying algorithm by its authors showed roughly 10x higher throughput than Raft and Multi-Paxos under adverse conditions, but that figure comes from research evaluation, not Cloudflare production use.
“experimental consensus service”
— Cloudflare Blog authors James Larisch, Bob Halley and João Pedro Leite
Mastering PostgreSQL for Data Engineering and Cloud Deployment: Design, Optimize, and Manage PostgreSQL Databases for Data Engineering, High … Engineering — SQL & PostgreSQL Path)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Production Timing Remains Open
Several details remain unconfirmed. Cloudflare has not announced a production launch date, customer availability, pricing, or a public API. It also has not published production performance data for Meerkat inside Cloudflare’s live systems.
The company also said Meerkat is not meant to handle Byzantine faults, meaning it assumes no actor inside the replica set is actively malicious. Its near-term use case is infrequently written control-plane state, not broad database replacement.
Cloud Native Data Center Networking: Architecture, Protocols, and Tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
More Meerkat Tests And Papers
Cloudflare said it plans more Meerkat posts over the next year covering QuePaxa, formal verification of its Rust implementation, bootstrapping, cluster management, replica placement and deterministic simulation testing. The company also said it is preparing a peer-review manuscript.
transactional key-value store
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Cloudflare Meerkat?
Meerkat is an experimental Cloudflare Research service for distributed consensus, aimed first at internal control-plane data that needs strong consistency across data centers.
Is Meerkat available to Cloudflare customers?
No. Cloudflare says Meerkat is internal-only for the near future and not deployed to production.
What makes Meerkat different from Raft systems?
Cloudflare says Meerkat’s QuePaxa-based design lets any healthy replica help drive consensus, while Raft systems depend on a single active leader for writes.
What problems is Meerkat meant to solve?
It is meant to keep small pieces of control-plane state strongly consistent and available during some machine, link or data-center failures.
What remains unknown about Meerkat?
Cloudflare has not given a production timeline, customer roadmap or live performance results. The next concrete milestone is more technical disclosure and a planned peer-review manuscript.
Source: hn
