Toward a Smell-aware Prediction Model for CI Build Failures

Abstract

During the last years, researchers have explored the potential factors behind Continuous integration (CI) build failures focusing mainly on metrics related to code changes, statistics about the project etc. However, code quality indicators such as the presence of bad smells have been rarely discussed in the context of CI. In this paper, we aim at investigating the extent to which CI build failures prediction can be improved by the detection of bad smells. Specifically, we evaluate the contribution of 28 well-known bad smells when added to BF-DETECTOR, an existing tool for CI build failures prediction. We conduct a case study on a dataset of 15,041 Travis CI builds extracted from five GitHub projects. The obtained results demonstrate the efficiency of the smell-aware prediction to improve the F1-score of BF-DETECTOR by 4% on average. In particular, we found that Excessive Parameter List (EPL), Sensitive Equality (SE) and Lazy Test (LT) are the most contributing to the prediction.