Cold Start Latency Optimization Strategies for Function as a Service Platforms
Tingjie Chen, Jianbo Ding
Abstract
Function as a Service (FaaS) has established itself as the dominant delivery model within the serverless computing (SC) ecosystem, enabling developers to deploy stateless, event-driven workloads without provisioning or managing any underlying server infrastructure. Despite substantial operational advantages in cost granularity and scaling automation, FaaS platforms are subject to a persistent performance bottleneck known as cold start latency, which occurs whenever a new execution environment must be initialized from scratch before an incoming function invocation can be served. Cold start penalties range from tens of milliseconds for lightweight runtimes to several seconds for applications executing on the Java Virtual Machine (JVM), producing direct violations of service-level objectives (SLOs) in latency-sensitive production deployments. This paper reviews optimization strategies for cold start latency across four principal categories: pre-warming and keep-alive policies, snapshot-based checkpoint-restore techniques, lightweight virtualization and isolation mechanisms including WebAssembly (Wasm), and scheduling and resource management strategies. Machine learning (ML) is examined as a cross-cutting enabler for predictive and adaptive mitigation. This review synthesizes the current state of understanding, characterizes trade-offs among competing strategies, and identifies open challenges including snapshot staleness management, isolation-speed tensions, and edge deployment constraints.