What are the Factors Impacting Build Breakage?

Abstract

Continuous Integration (CI) has become a good practice of software development in recent years. As an essential part of CI, build creates software from source code. Predicting build outcome help developers to review and fix bugs before building to save time. However, we are missing objective evidence of practical factors affecting build result. Travis CI provides a hosted, distributed continuous integration service used to build and test software projects hosted at GitHub. The TravisTorrent is a dataset which deeply analyzes source code, process and dependency status of projects hosting on Travis CI. We use this dataset to investigate which factors may impact a build result. We first preprocess TravisTorrent data to extract 27 features. We then analyze the correlation between these features and the result of a build. Finally, we build four prediction models to predict the result of a build and perform a horizontal analysis. We found that in our study, the number of commits in a build (git_num_all_built_commits) is the most import factor that has significant impact on the build result, and SVM performs best in the four of the prediction models we used.