Towards quantifiable boundaries for elastic horizontal scaling of microservices

One of the most useful features of a microservices architecture is its versatility to scale horizontally. However, not all services scale in or out uniformly. The performance of an application composed of microservices depends largely on a suitable combination of replica count and resource capacity. In practice, this implies limitations to the efficiency of autoscalers which often overscale based on an isolated consideration of single service metrics. Consequently, application providers pay more than necessary despite zero gain in overall performance. Solving this issue requires an application-specific determination of scaling limits due to the general infeasibility of an application-agnostic solution. In this paper, we study microservices scalability, the auto-scaling of containers as microservice implementations and the relation between the number of replicas and the resulting application task performance. We contribute a replica count determination solution with a mathematical approach. Furthermore, we offer a calibration software tool which places scalability boundaries into declarative composition descriptions of applications ready to be consumed by cloud platforms