Does Code Decay? Assessing the Evidence from Change Management Data

Reviewer: Andrew Brooks

Change management data from a fifteen-plus year change history on millions of lines of code is analyzed to determine whether code decay (the increasing difficulty over time of making changes) is real. Figure 3 presents a convincing visualization that the probability of a modification request touching more than one file increases from a low of less than 2 percent in 1989 to more than 5 percent in 1996. Whilst the authors recognize the alternative interpretation that the inherent difficulty of the desired changes is increasing, the modification request data is not examined independently from this perspective. Figure 4 presents a convincing visualization of the 'breaking down' of functional separation between two clusters of modules meaning some modification requests repeatedly touched on the same group of modules. A discussion on the nature of these particular modification requests might have provided insights as to whether or not the inherent difficulty of the desired changes was increasing. As further evidence of code decay, the authors provide models for fault prediction and state the best models predict numbers of faults using numbers of changes to the module in the past. Forming an appreciation of the strength of this claim is, however, not possible without referring to an earlier publication. This paper represents a major contribution to the understanding of code decay and is highly recommended to specialists in software maintenance with access to change management data who are looking for ideas on metrics and visualizations of change.