Electronic Thesis and Dissertation Repository

Thesis Format

<-- Please Select One -->

Degree

Master of Science

Program

Computer Science

Supervisor

Konstantinos Kontogiannis

Abstract

Context: Continuous Integration (CI) is a DevOps technique which is widely used in practice. Studies show that its adoption rates will increase even further. At the same time, it is argued that maintaining product quality requires extensive and time consuming, testing and code reviews. In this context, if not done properly, shorter sprint cycles and agile practices entail higher risk for the quality of the product. It has been reported in literature [68], that lack of proper test strategies, poor test quality and team dependencies are some of the major challenges encountered in continuous integration and deployment.

Objective: The objective of this thesis, is to bridge the process discontinuity that exists between development teams and testing teams, due to continuous deployments and shorter sprint cycles, by providing a list of potentially buggy or high risk files, which can be used by testers to prioritize code inspection and testing, reducing thus the time between development and release.

Approach: Out approach is based on a five step process. The first step is to select a set of systems, a set of code metrics, a set of repository metrics, and a set of machine learning techniques to consider for training and evaluation purposes. The second step is to devise appropriate client programs to extract and denote information obtained from GitHub repositories and source code analyzers. The third step is to use this information to train the models using the selected machine learning techniques. This step allowed to identify the best performing machine learning techniques out of the initially selected in the first step. The fourth step is to apply the models with a voting classifier (with equal weights) and provide answers to five research questions pertaining to the prediction capability and generality of the obtained fault proneness prediction framework. The fifth step is to select the best performing predictors and apply it to two systems written in a completely different language (C++) in order to evaluate the performance of the predictors in a new environment.

Obtained Results: The obtained results indicate that a) The best models were the ones applied on the same system as the one trained on; b) The models trained using repository metrics outperformed the ones trained using code metrics; c) The models trained using code metrics were proven not adequate for predicting fault prone modules; d) The use of machine learning as a tool for building fault-proneness prediction models is promising, but still there is work to be done as the models show weak to moderate prediction capability.

Conclusion: This thesis provides insights into how machine learning can be used to predict whether a source code file contains one or more faults that may contribute to a major system failure. The proposed approach is utilizing information extracted both from the system’s source code, such as code metrics, and from a series of DevOps tools, such as bug repositories, version control systems and, testing automation frameworks. The study involved five Java and five Python systems and indicated that machine learning techniques have potential towards building models for alerting developers about failure prone code.

Summary for Lay Audience

.

Download

Share

COinS