An Auto-Scaling Framework for Predictable Open Source FaaS Function Chains

Function as a Service (FaaS) is a novel approach for application virtualization in the cloud. Complex cloud native applications may be implemented as chains of functions. Due to the inter-dependencies between the functions of such chains, their performance is harder to predict. This also means that the design of a resource efficient scaling mechanism is a challenging task, the goal is to avoid wasteful scaling strategies. In this paper, we propose an auto-scaling framework that keeps the amount of resources, allocated for the functions in the chain, at a minimum level, while meeting the latency requirements of an application implemented as a function chain. Then, we show how to integrate our auto-scaling framework with a general open source FaaS system, as well as with, OpenFaaS. We also introduce our modifications on OpenFaaS that enable to use various load-balancing algorithms for distributing the requests between the function instances. We compare our auto-scaling framework to the auto-scaler of OpenFaaS by the completion time of our python3 based image processing test function chain, as well as, by the amount of allocated resources.