Fork in the Road - Reflections and Optimizations for Cold Start Latency in Production Serverless Systems

Learn about cutting-edge optimizations for serverless computing cold start latency through this 16-minute conference presentation from OSDI '25. Explore an in-depth investigation of existing cold start optimization techniques and their real-world effectiveness in large-scale industrial deployments at Ant Group. Discover three previously overlooked sources of latency: control path latency from serverless runtime interactions, resource contention latency under high concurrency, and user code initialization latency reflecting the trade-off between resource efficiency and startup performance. Examine how simplified research assumptions often overlook the complexities of production systems, why isolated component optimization fails to address end-to-end workflow interactions, and how concurrent execution environments introduce unique challenges. Understand the novel suite of techniques proposed by researchers from Tsinghua University, Ant Group, and Shanghai Jiao Tong University to overcome these limitations, designed for real-world workload adaptability and large-scale deployment scalability. See how their AFaaS (Ant FaaS) system reduces cold start latency to millisecond levels and has maintained stable performance in production for over 18 months, addressing the persistent performance bottleneck that affects serverless computing adoption in public cloud environments.