Introducing Meerkat: an experiment in global consensus

Cloudflare Research introduces Meerkat, a consensus service built on the QuePaxa algorithm (Tennage & Băsescu et al., 2023). Unlike Raft, where only the leader can write and leader failure blocks progress until a timeout-driven election completes, QuePaxa lets all replicas propose writes at all times. A leader exists but is optional — its only benefit is fewer round trips.

Meerkat targets small control-plane state (e.g., leadership for replicated databases) across Cloudflare's 330+ data centers. It provides linearizability and serializability, tolerating crash and network failures so long as a majority of replicas remain reachable. Under adversarial network conditions the algorithm reportedly sustains ~10× the throughput of Raft or Multi-Paxos.

The trade-off is latency: consensus requires 1–3 round trips to a majority (1 for the leader, 3 for a non-leader, more under contention). Mitigations include batching writes, stale local reads, and bundling multiple operations into one proposal. These characteristics suit infrequently-written control plane data that must stay consistent.

Meerkat is not yet in production. In proof-of-concept tests with up to 50 globally distributed replicas, leader failures were constant yet the cluster continued operating with no increase in error rate. The team plans further posts on QuePaxa internals, formal verification of their Rust implementation, and deterministic simulation testing.