Predictive Scaling in CI/CD Pipelines Using Machine Learning

Continuous Integration and Continuous Deployment (CI/CD) is widely adopted in modern software development. Current auto-scaling tools incloud environments are based on hardware-level metrics, limiting their ability to anticipate workload changes in a stateful cloud. This thesis investigates the usefulness of applying machine learning to predict the workload of CI/CD pipelines. The models are trained on historical job and pipeline data. Numerous preprocessing and feature techniques are explored to transform the historical data to suitable machine learning input. Two models are tested and evaluated; Random Forest Regression and Long Short-Term Memory (LSTM), with and without continuous learning. The models are evaluated on their ability to predict the workload of all occurring job types in a near future. The results indicate that all models can achieve some form of predictability, but the LSTM continuous learning model suggests to be the most accurate. However, the predictability decreases the longer test data is from the original training interval. The models also show the weakest predictability in less frequently used jobs. The findings suggest that machine learning models can contribute to reduced developer waiting time when predictions are accurate, without increasing a current static rules based solution. Future work includes exploring additional models and features, evaluate the models in a simulated real life scenario and using more computationally capable hardware.