Function as a Service (FaaS) is a cloud computing model that relieves application developers from the responsibility of managing infrastructure tasks like resource provisioning and scaling. Serverless functions, however, have specific and limited execution times, making effective auto-scaling decisions for these services particularly challenging. To ensure proper configuration and scaling of resources, it is essential to have a thorough understanding of environmental changes and dynamic factors that influence system performance, alongside considering function specifications and user needs. To tackle this issue, we introduce FaaScaler, a reinforcement learning-based approach for scaling functions and containers, aimed at meeting function deadlines and enhancing processor efficiency. FaaScaler models the scaling challenge as a Markov Decision Process (MDP) and utilizes the Proximal Policy Optimization (PPO) learning algorithm to train an agent to make both horizontal and vertical scaling decisions. The proposed system is evaluated using the OpenFaaS platform, with results indicating that FaaScaler effectively meets execution deadlines for most requests while optimizing processor utilization.

Keywords