My research paper just got published 🎉, and we were able to improve the cache-hit ratio for nginx workloads by 146%.
Last week, Aayush and I were brainstorming an idea to improve the cache hit ratio, and that’s when we applied Deep Q-Networks to see if we could get some major benefits.
The core idea is - instead of evicting the least recently used key, we trained a small model that, given various features of an NGINX request, chooses the best key to evict among the coldest k.
Give it a read to understand our thesis and nuances. The link is in the comment.
The paper is titled “Cold-RL: Learning Cache Eviction with Offline Reinforcement Learning for NGINX”.