Quantitative modeling of software reviews in an industrial setting

Abstract

Technical reviews are a cost effective method commonly used to detect software defects early. To exploit their full potential, it is necessary to collect measurement data to constantly monitor and improve the implemented review procedure. This paper postulates a model of the factors that affect the number of defects detected during a technical review, and tests the model empirically using data from a large software development organization. The data set comes from more than 300 specification, design, and code reviews that were performed at Lucent's Product Realization Center for Optical Networking (PRC-ON) in Nuernberg, Germany. Since development projects within PRC-ON usually spend between 12% and 18% of the total development effort on reviews, it is essential to understand the relationships among the factors that determine review success. One major finding of this study is that the number of detected defects is primarily determined by the preparation effort of reviewers rather than the size of the reviewed artifact. In addition, the size of the reviewed artifact has only limited influence on review effort. Furthermore, we identified consistent ceiling effects in the relationship between size and effort with the number of defects detected. These results suggest that managers at PRC-ON must consider adequate preparation effort in their review planning to ensure high quality artifacts as well as a mature review process.