What happened
Cloudflare’s announcement says Meerkat is a distributed system in which replicas agree on an ordered log of events. Applications can use that log to build shared state, including a key-value store; Cloudflare compares this class of system with Consul, etcd, and Apache ZooKeeper. The company says Meerkat is designed for global deployment, where links between continents can have high and variable latency.
Cloudflare chose QuePaxa, a consensus algorithm published in 2023, over established options such as Raft, Paxos, and Viewstamped Replication. The post argues that leader-based algorithms can become slow or unavailable when a leader is delayed, disconnected, or mistakenly suspected of failure. Electing a replacement can itself be disruptive: multiple candidates may interfere with one another, and a transient network issue can trigger a prolonged recovery process. Cloudflare cites a 2020 partial failure of a core network switch, during which a consensus-powered system made its dashboard 80 times slower for six hours; the company says no data was lost, but writes became unavailable until the issue was resolved.
QuePaxa uses proposers and recorders. Proposers submit candidate values; recorders retain proposals and relay information among proposers. Each proposal has a priority, and proposers adopt higher-priority proposals they learn about. The aim is to make proposers converge on a value instead of blocking one another through competing ballots. The algorithm proceeds through four phases in a round: propose and adopt, spread, check and decide, then carry forward a proposal if no decision was reached. Cloudflare says that, given enough rounds, proposers are guaranteed to agree on a proposal.
The proposed approach is deliberately different from diagnosing a failed leader and then starting an election. Cloudflare says a process in a distributed system cannot directly distinguish a crashed remote machine from a late message or a temporarily broken connection. In leader-based designs, that uncertainty can start recovery even when the leader is still functioning. Detection timeouts also have to account for network round-trip time; Cloudflare notes that Raft is commonly configured with election timeouts of five to ten times the round-trip time. Across long network paths, that can turn an accurate failure decision into a noticeable period in which the system cannot progress.
The algorithm has an optional leader with the highest priority. When conditions are favorable, that leader can decide in a single network round trip. But clients are not required to send requests only to the leader: Cloudflare says clients hedge proposals to every replica, while non-leader replicas wait briefly to give the leader the chance to use the fast path. Cloudflare’s implementation puts this hedge wait on clients rather than replicas, which it says reduces message volume. Lowering the hedge timeout can make the system react faster to leader changes, but can increase traffic and make a decision take longer. The announcement describes Meerkat as experimental; it does not present an external product, general-availability date, or price.
What the archive adds
New Runtime’s earlier coverage of Cloudflare’s post-quantum privacy protocols for Privacy Pass is a separate Cloudflare infrastructure story (The post-quantum future of web privacy, November 6, 2026). It does not cover Meerkat or consensus. This is a new event and opens a distinct thread about global coordination systems.
Trend signal
This does not yet fit a pattern New Runtime tracks. Cloudflare’s account is a design rationale for an experimental consensus service, not evidence of a broader shift in how AI agents are built or operated.
Connections to other events
New Runtime’s earlier coverage noted Cloudflare’s Workers KV Instant work: a globally replicated key-value service built on Quicksilver, with the company reporting sub-two-millisecond p99 reads and 250-millisecond global replication across more than 300 edge locations. Meerkat addresses a different layer: agreeing on the ordered events from which a shared state can be constructed. Together, the two efforts point to Cloudflare treating globally distributed state as an infrastructure problem with separate read, replication, and coordination concerns.
Why it matters
For teams building systems that need consistent shared state across regions, the announcement is a reminder that network behavior is part of the failure model. A protocol that handles a crashed server but stalls on a delayed message may be a poor fit for a wide-area deployment. QuePaxa’s optional leader and cooperative proposals illustrate one alternative to making progress depend entirely on a reliably reachable coordinator.
That distinction matters for AI systems as well as conventional infrastructure. An agent workflow that coordinates leases, job ownership, rate limits, durable task state, or tool-side effects across regions may eventually need a strongly consistent control-plane component. The model does not remove the need to decide who owns a task, whether an action has already been committed, or which state transition is authoritative. Those requirements turn latency spikes and partitions into product behavior, not merely an operations concern.
The trade-off is operational, not just theoretical. Cloudflare explicitly describes a relationship between the hedge timeout, message volume, and decision time. Teams evaluating similar designs would need to test those variables against their own latency distribution, failure patterns, and write-availability requirements. The post is not a benchmark showing that QuePaxa outperforms Raft or other mature implementations in general, nor does an experimental internal system establish that the algorithm is ready for broad adoption.
Limits and uncertainty
Cloudflare characterizes Meerkat as experimental and describes its intended design, but this announcement does not establish production maturity or independently verified performance. No comparative benchmark figures for Meerkat versus Raft, Paxos, or other systems appear in the account summarized here. The stated guarantee of eventual agreement is distinct from a claim about bounded latency or availability under every network condition.
The practical costs of the hedge strategy will depend on workload and deployment details; Cloudflare acknowledges additional traffic and possible slower decisions when more proposers participate. The announcement also does not establish how Meerkat behaves under a particular production topology, request mix, replica count, or sequence of correlated network faults. Those are the conditions that would determine whether its design trade-offs are favorable for another operator.
Skill to build now
Write down the coordination path for one distributed workflow this week. Identify its authoritative state, the component that acts as coordinator, the timeout that changes behavior, and what happens if messages are delayed rather than lost. Then run a fault-injection test that adds latency without crashing a process. The useful outcome is not adopting QuePaxa; it is learning whether a short network disturbance triggers retries, leadership churn, duplicate work, or a write outage in the system you already operate.
Reader takeaway
Cloudflare’s Meerkat is an experiment in making global consensus less dependent on a single healthy, reachable leader. The transferable lesson is to match a consensus protocol to the network faults a system actually faces—and to treat faster recovery, traffic overhead, and decision latency as linked trade-offs, not isolated tuning knobs.