Abstract

Pull requests are a critical approach for developers to collaborate in software development, which also initiates the following code review and integration. However, tangled pull requests can be introduced into version control systems via committing unrelated or multi-purposes code changes in one single pull request, which have been found to bring a negative impact on code recommendation systems and bug prediction models in the previous research. In this thesis, we conduct a case study on 640 pull requests among 8 popular open-source Java projects from GitHub. Through manual analysis, we find that 47% of the pull requests are tangled. In order to further understand the characteristics of tangled pull requests, we perform a qualitative annotation and classify the reasons for tangled pull requests. We find that 75% are tangled because bug fixing, feature improvement, or new feature adding are committed with an update in test code. The remaining 25% of pull requests are tangled because developers commit two or more unrelated bug fixing, feature improvement, test code adding or modifying, new feature adding, feature improvement and bug fixing, and other combinations inside one pull request. We also propose an approach to predict whether a pull request is tangled or not with an AUC of 0.87. Furthermore, we also predict whether two lines of code changes belong to the same task, which achieves an AUC of 0.74.