Toward an Edge-Friendly Distributed Object Store for Serverless Functions

Serverless computing is changing the way in which we structure and deploy computations in Internet-scale edge systems. This paper presents Capybara, a new scalable and programmable distributed object store for storing and sharing serverless function data objects (state) on edge infrastructures. The key innovations here are (1) achieving scalability and avoiding the significant DRAM cost through a consistent DHT-based P2P architecture; and (2) providing a programmable handler abstraction to customize state management policies (e.g., data caching policies, container "keep-alive" times, access control methods, and data replication policies). We implement Capybara prototype on the Pastry DHT, deploy it on 150 Amazon EC2 nodes, and evaluate it by building several use cases to conduct real-world experiments, demonstrating its significant gains in data locality, state management customization, and scalability compared to the state-of-the-art.