Discuss your project
Web architecture

Commit configuration changes when one network site slows down

Three sites need to apply changes in the same order. Learn why Cloudflare is experimenting with QuePaxa and how to measure the network’s effect on a small test setup.

Valider une configuration quand un site du réseau ralentit

Three sites receive configuration updates. Each must publish the same sequence of changes. Consensus lets replicas agree on that sequence despite delays and certain failures.

Why Cloudflare chose QuePaxa

On November 11, Cloudflare explained its choice for Meerkat, an experimental global consensus service. The QuePaxa algorithm comes from a 2023 paper. In favorable conditions a designated proposer can reach a decision in one round trip. Other replicas can propose when it takes too long. The design combines delayed proposals with random priorities. Results depend on assumptions such as crash-stop failures and whether a quorum, the group required to decide, remains available.

Replay comparable network conditions

  1. Set up three test replicas and a synthetic registry with a request ID. Compare two instrumented consensus implementations at the same message size, load, and placement.
  2. Measure stable links, then degrade the link to the designated proposer. Stop one replica and test recovery. Repeat with the same fixtures.
  3. Record 95th-percentile decision time, message volume, and pending requests. The 95th percentile is the duration below which about 95% of decisions complete. Compare committed order on every replica after recovery. Check that lost quorum cannot produce apparent successful writes.

Balance gains under disruption against traffic cost, operational complexity, and steady-state behavior. The post describes an experiment, and your measurements apply to your setup.

Share this article