Stateful Serverless Computing : A Comparative Analysis of Orchestrator-based and Actor-like Models

Sammanfattning

Serverless cloud computing, also mostly known as Function-as-a-Service (FaaS) is a popular cloud model where cloud users write a piece of code, or, namely, functions, and cloud providers take the responsibility of the underlying operational aspects, such as execution, auto-scaling, and monitoring of these functions, instead of the cloud users themselves. While this model provides benefits and is widely adopted, the stateless nature of FaaS led to the evolution of initial serverless functions to more advanced and architectural patterns that we refer to as a stateful serverless system (SSS) in this thesis. SSS provides abstractions that hide the challenges of managing and persisting state while keeping the benefits of serverless functions, such as auto-scaling and usage-based billing. In this thesis, we conducted a comparative analysis of stateful serverless architectures, where we specifically focused on two different architectural approaches: orchestrator-based and actor-like models. For this purpose, we chose Azure Durable Functions (ADF) and Apache Flink Statefun (FSF) to compare their architectural differences and performance characteristics, such as latency and throughput, through quantitative benchmarking. In our benchmarking scenarios, FSF consistently achieved the highest throughput and lowest latency in all tested experiments, while ADF with the Default backend configuration showed modest performance due to its I/O bounded state access pattern. Our benchmarking results provide quantitative metrics about the performance characteristics and architectural attributes of stateful serverless approaches, contributing to the understanding of these systems. The results and contributions of this thesis are valuable for system architects, developers, and researchers.