Predictive Auto-scaling in Kubernetes - Thesis

Over the course of my final year at Williams College, I spent a lot of time working on a distributed systems thesis with Professor Jeannie Albrecht. My thesis is entitled “Predictive Pod Auto-scaling in the Kubernetes Container Cluster Manager”, and it is entirely open-source. The written portion, presentation slides, and evaluation code can be found on my Github. Additionally, our contributions to Google’s open-source cluster container manager Kubernetes can be found on my fork. These changes will hopefully be merged into Kubernetes master branch soon.

In short, my thesis focused on adding predictive auto-scaling to Kubernetes. Auto-scaling simply means allocating different amounts of resources to an application as its external load changes. Previously, Kubernetes implemented horizontal, reactive auto-scaling, meaning containerized applications were replicated or destroyed based on the current resource utilization of the application. The auto-scaler created and destroyed replica pods so as to ensure each application maintained a certain desired level of resource utilization (i.e. 60% CPU utilization). Our addition of predictive auto-scaling follows a similar pattern, yet creates or destroys replica applications based on predicted resource utilization. The interval of time required for a replica pod to share in the computational work determines the interval into the future for which we predict.

Working on this project was incredibly fulfilling and I’m so grateful to those who made it possible.