Cloudflare introduces Meerkat, an experimental consensus service using the QuePaxa algorithm for global control planes

Details

Cloudflare developed Meerkat, an experimental distributed consensus service for its internal global control plane, utilizing the QuePaxa algorithm to handle high-latency, intercontinental networks.
QuePaxa removes dependency on a single leader and eliminates fault detection timeouts, allowing the system to continue processing requests during network partitions or leader crashes without stalling.
In global evaluations, Meerkat maintained throughput stability with over 95% of decisions resolved in Phase 1 (leader fast path); when slower paths are required, throughput drops to at most 1/3 of normal but the system does not stop.
The algorithm uses constructive interference where proposers adopt the highest-priority proposal from a majority of recorders, ensuring convergence even if the leader fails, unlike Raft or Paxos which can suffer metastable failures during leader election.
Meerkat operates as a key-value store by agreeing on a sequenced log of events, similar to Consul or etcd, but is designed for environments where network failures are more frequent than server failures.
Future work includes formal verification, extensive simulation, and expanding support from crash-stop to crash recovery, as QuePaxa has fewer validated production cases than Raft or Paxos.

Key point