Dynamic Storage Selection for Mitigating Tail Latency in Serverless Pipelines

In the rapidly evolving landscape of cloud computing, serverless computing has seen significant growth, offering a user-friendly programming interface for building scalable microservices at reduced costs. However, due to its stateless nature, managing intermediate results in complex application pipelines is challenging due to the need for distributed object storage like Azure Blob Storage, AWS S3, or Redis. To address this challenge, we propose HyperStore, an ephemeral storage management framework that features a smart function container enabling seamless and real-time interaction with various storage providers, without altering the function code. Furthermore, HyperStore dynamically determines and selects the optimal storage for each pipeline function to meet Service Level Objectives (SLOs) at minimal cost and mitigate tail latency. Our framework can be integrated with existing Serverless systems such as OpenFaas, allowing function containers to communicate via a shared host volume for reduced latency in case of function co-location. Our detailed experimental results demonstrate the efficiency and advantages of our approach.