HyDuo: A Low-Latency and Cost-Efficient Storage System for Stateful Serverless Applications

Abstract

More applications are adopting serverless architectures for elasticity and fine-grained billing. However, stateful workloads suffer performance degradation due to frequent cross-stage data exchanges that rely on expensive remote storage access. Existing work rarely considers both data access patterns and heterogeneous storage media, limiting its ability to reduce this overhead. Modern storage spans multiple tiers, from fast but expensive DRAM and NVMe to slower, cheaper SSDs and HDDs. We also observe that stateful serverless workloads exhibit clear patterns, including read-heavy behavior, high access intensity, periodicity, and strong inter-stage correlation. We model intermediate data placement as a generalized assignment problem by capturing both storage heterogeneity and access patterns. Based on this model, we design a dynamic migration algorithm that promotes and demotes data across tiers to reduce latency while controlling cost. Experiments show that our approach achieves latency close to an in-memory SOTA system for serverless analytics workloads, while reducing storage cost by about 95%.