Cold/warm Kafka partition assignment

A rebalancing algorithm that cut P99 latency spikes during scale-out from ~20 s to 300 ms. Patented as US-12309441.

Built at Amazon for Ring’s internal service-communication platform, which runs on MSK (Kafka) and was showcased at AWS re:Invent. When consumers scaled out, the standard partition rebalancing handed traffic to subscribers that weren’t ready for it, producing ~20-second latency spikes at P99.

The algorithm distinguishes between cold and warm subscribers and assigns partitions accordingly, so new capacity is brought in without stalling traffic. Latency spikes during scale-out dropped to around 300 ms.