Five years ago, Cloudflare introduced a method to reduce cold starts in their Workers by pre-warming them during the TLS handshake, leveraging the TLS Server Name Indication (SNI) to preemptively prepare the Workers for incoming requests. Although the initial approach significantly reduced cold starts, the increasing complexity and size of applications deployed on Workers necessitated further innovations. Cloudflare's recent deployment of "Worker sharding" utilizes a consistent hash ring to efficiently manage Worker instances across their global network, routing requests to already active Workers to minimize cold starts. This approach addresses inefficiencies in the previous system, where requests often led to unnecessary cold starts due to suboptimal routing. By coalescing requests on single servers and using techniques like Cap'n Proto RPC for efficient load management and sharding, Cloudflare has managed to reduce the global Worker eviction rate by tenfold, improving memory efficiency and reducing latency. As a result, the warm request rate for enterprise traffic has improved from 99.9% to 99.99%, significantly decreasing the cold start rate. Despite these advancements, Cloudflare acknowledges that there is still potential for further improvement in optimizing their serverless computing platform.