Power to the Applications: The Vision of Continuous Decentralized Autoscaling

Autoscaling has been one of the most active research areas since the beginning of the cloud computing era. Nearly all previously proposed approaches focus on decision-making based on averaged monitoring values of many service instances at fixed points in time. This limits responsiveness and can lead to service level objective (SLO) violations when the load suddenly increases. Our vision of continuous decentralized autoscaling avoids these issues by giving individual service instances the power to make scaling decisions in a distributed fashion. Each instance performs self-monitoring and evaluates its state. The service instance initiates upscaling if it detects an overload or downscaling if its load is below a specified threshold. By randomly determining scaling timing, we achieve quasi-continuous scaling behavior when multiple service instances are deployed. We discuss challenges regarding analytical modeling, simulation, and real-world evaluation of this approach.