State Migration in Styx: Towards Serverless Transactional Functions

Area of Science:

• Computer Science
• Distributed Systems
• Cloud Computing

Background:

• Developing stateful cloud applications with strict consistency is challenging.
• Existing Stateful Functions-as-a-Service (SFaaS) approaches have weak transactional guarantees or high latency.
• Inefficient transactional protocols hinder performance in stateful services.

Purpose of the Study:

• Introduce Styx, a novel dataflow-based SFaaS runtime.
• Provide exactly-once transactional guarantees for stateful functions with arbitrary call-graphs.
• Reduce execution latency and improve scalability for stateful cloud applications.

Main Methods:

• Developed a novel dataflow-based SFaaS runtime named Styx.
• Extended a deterministic transactional protocol with function acknowledgment, caching, and early-commit reply mechanisms.
• Implemented elasticity for state migration to support load balancing.

Main Results:

• Styx achieves at least one order of magnitude higher throughput than state-of-the-art approaches.
• Demonstrated near-linear scalability and low latency across YCSB, TPC-C, and Deathstar benchmarks.
• Styx's state migration approach adapts swiftly to workload changes, outperforming baseline methods.

Conclusions:

• Styx effectively addresses the challenges of developing stateful cloud applications.
• The novel runtime offers significant performance improvements in throughput, scalability, and latency.
• Styx provides a robust and efficient solution for demanding stateful workloads in the cloud.