Cloudflare Meerkat - Globally distributed consensus

4 hours ago

Cloudflare needs a distributed consensus service for global control-plane state management, ensuring strong consistency and availability despite network unpredictability.
Raft-based systems suffer in wide-area networks due to leader dependency and timeout issues, leading to availability problems during leader failures or network degradation.
Meerkat is Cloudflare's new consensus service using the QuePaxa algorithm, allowing all replicas to perform writes without timeouts, improving fault tolerance and availability.
Meerkat guarantees linearizability for applications like a key-value store, ensuring all reads see prior writes even across distributed replicas, using a consensus log to synchronize events.
QuePaxa provides higher availability than Raft by eliminating required leaders, avoiding leader election delays, and handling concurrent proposals constructively in unreliable networks.
Meerkat has latency limitations due to consensus round-trips but optimizes performance through batching, replica placement control, and optional stale reads for non-critical data.
The service is experimental and internal-only for now, designed for infrequently written control-plane data, with future blog posts planned on its implementation and verification.

Hasty Briefsbeta