Automatic Performance Baselining of Kubernetes Microservices

When making any change to the IT stack, like deploying a new application release, we must be able to understand its impact on application performance. However, IT systems are exposed to a wide variety of workloads, varying over time both in intensity and typology. This results in wild fluctuations in the measured performance even when we do not make any changes to the IT stack. The goal of the thesis is to develop an automated methodology to derive baseline application performance. More specifically, starting from a dataset containing the behaviour of a real microservice application the methodology will split the performance time series into many different operating regions, essentially reconstructing the workload typologies. Then, for each identified cluster, it will characterize the application performance and derive many Service Level Objectives (SLOs). These are constraints on the application performance, defined by using a proper Site Reliability Engineering (SRE) methodology, which involves defining both a threshold and an error budget for all the relevant time series. The thesis focuses on time series analysis and classification and will provide the student with a solid foundation in performance analysis, using industry-standard methodologies.