Towards Low-Cost Global Highly Available Large Container-Based Serverless Functions

Serverless computing and the Function-as-a-Service (FaaS) paradigm have transformed how cloud-hosted applications are deployed, providing remarkable scalability and geo-graphical distribution capabilities. AWS Lambda's support for containerized functions has further enhanced the packaging and deployment process, enabling seamless portability and consis-tency across various operational environments. However, a fun-damental challenge remains: the replication of Docker container images across multiple container registries in different regions to ensure high availability of globally deployed functions with low replication latency to support geographically dispersed clients. While multi-region deployment minimizes latency and en-hances service availability for diverse clients, the associated costs and operational complexities are non-trivial. The replication process, particularly for large container images, is both expensive and slow, potentially hindering rapid scalability and respon-siveness. Public cloud container image replication support such as that offered by Amazon's Cross Region Replication (CRR) represents a significant stride forward, automating the replication of container images across regions within container repositories. Despite its advantages, CRR is constrained by several limitations: all images in a container repository are replicated after this feature is enabled, and there is an absence of fine-grained filtering at the image level within repositories. This can result in superfluous replications and increased costs. This paper investigates alternatives for cost-effective and effi-cient global deployment of container-based serverless functions on AWS Lambda. We introduce on-demand container image replication mechanisms with promise to reduce storage costs by avoiding unnecessary replication. We investigate multiple approaches to replicate container images only when needed to minimize data transfer latency and storage overhead. We provide a comprehensive evaluation of these approaches, assessing their impact on cost, replication time, and turnaround time for function availability. Our findings demonstrate potential for substantial cost savings while maintaining high availability and consistent performance for container-based Lambda functions deployed worldwide.