Abstract

Many systems must run without interruptions, yet they must upgrade to address issues, such as fixing bugs or adding new functionalities. Delivering uninterrupted services is not just the responsibility of the system providing the services, but also the environment hosting the system. So, it is essential to understand and analyze the service availability guaranteed during an upgrade by the orchestration platform that hosts the system.
Kubernetes is a popular orchestrator of containerized workloads and services. It is essential to understand and analyze the impact of upgrades in a Kubernetes cluster: the effect of an upgrade on the service availability, how a failure during an upgrade is taken care of by Kubernetes and by the tools managing the Kubernetes cluster; and the effects of an upgrade process failure.
This thesis investigates, quantifies these impacts, and analyzes the causes. This thesis identifies three upgrade levels in a Kubernetes cluster: Kubernetes cluster version upgrade, Kubernetes application upgrade, and container runtime upgrade. We evaluate and analyze the state of the practice of upgrades for each level by performing various experiments under different (failure) scenarios. For each experiment, the manual collection of event timestamps and then the calculation of evaluation metrics (using collected timestamps) is a tedious and time-consuming task. To tackle this issue, we devise and implement an “Auto-Metric collector” tool that automates this process of event collection and metric calculation. The results of our experiments and analysis highlight the shortcoming of Kubernetes in identifying upgrade process failure and taking remediation measures. We propose potential solutions for some of the identified shortcomings.