Cloudflare cut 100TB of RAM by deriving the coefficient of variation of its own load distribution.

Kevin Guthrie and Mariia Iurchenko showed Pingora Backend Router's use of pingora-ketama consistent hashing had a load-distribution CV of sqrt((N-1)/(N*k+1)), which meant the final 90,000 hashes bought a 0.7% improvement and 32-bit hash collisions make anything past about 10,000 hashes per server counterproductive. They cut hashes per server 90% and shrank each entry from an 8-byte struct to a 6-byte array, for 75% and 25% reductions, totaling 100TB globally. Cloudflare Rollout ran both rings simultaneously and migrated datacenter by datacenter to avoid cache invalidation spiking origin traffic.