Abstract
Continuous Integration (CI) is a set of software development practices that allow software development teams to generate software builds more quickly and periodically (e.g., daily or even hourly). CI brings many advantages, such as the early identification of errors when integrating code. When builds are generated frequently, a long build duration may hold developers from performing other important tasks. Recent research has shown that a considerable amount of development time is invested on optimizing the generation of builds. However, the reasons behind long build durations are still vague and need an in-depth study. Our initial investigation shows that many projects have build durations that far exceed the acceptable build duration (i.e., 10 minutes) as reported by recent studies. In this paper, we study several characteristics of CI builds that may be associated with the long duration of CI builds. We perform an empirical study on 104,442 CI builds from 67 GitHub projects. We use mixed-effects logistic models to model long build durations across projects. Our results reveal that, in addition to common wisdom factors (e.g., project size, team size, build configuration size, and test density), there are other highly important factors to explain long build durations. We observe that rerunning failed commands multiple times is most likely to be associated with long build durations. We also find that builds may run faster if they are configured (a) to cache content that does not change often or (b) to finish as soon as all the required jobs finish. However, we observe that about 40% of the studied projects do not use or misuse such configurations in their builds. In addition, we observe that triggering builds on weekdays or at daytime is most likely to have a direct relationship with long build durations. Our results suggest that developers should use proper CI build configurations to maintain successful builds and to avoid long build durations. Tool builders should supply development teams with tools to identify cacheable spots of the project in order to accelerate the generation of CI builds.