Análise de desempenho de estratégias de autoscaling do kubernetes visando disponibilidade e tolerância a falhas

Abstract

With the evolution of application architectures, particularly with the adoption of the microservices paradigm and advances in server virtualization through cloud computing providers, it has become possible to dynamically allocate computing resources according to the demand. This flexibility has enabled more effective and accurate approaches to handle unexpected application behaviors, improving resource utilization by minimizing waste and preventing shortages. As a result, applications are able to maintain performance and availability even under unpredictable conditions. Application scaling refers to the ability of software to automatically adapt to variations in workload and demand, ensuring operational stability. In this context, container technologies play a key role in cloud environments and are widely used for deploying microservices. Containers can be managed manually by human operators or, more efficiently and automatically, by container orchestration platforms such as Kubernetes. In this study, it was used Kubernetes, the most widely adopted orchestrator nowadays. Within Kubernetes, containers are grouped into units called Pods, and the platform provides several Pod scaling options, known as autoscaling mechanisms. Therefore, this thesis implements different Kubernetes Pod autoscaling strategies to perform a comparative analysis of their performance under various configurations and software scenarios, highlighting the impact of each approach with a focus on availability and fault tolerance. The results showed that strategies based purely on computational resources can be effective for synchronous response applications but proved insufficient for asynchronous and event-driven applications. In event-driven scenarios, only mechanisms based on custom metrics related to specific events (such as queue size) were able to ensure performance, stability, and efficient resource management, demonstrating that the choice of autoscaling depends directly on the architecture of the software employed.