2 Data Mining Version Histories 2.1 Learning from History 2.2 Mining Rules

Program analysis long has been understood as the analysis of source code alone. A modern software product, though, is more than just program code; it contains documentation, interface descriptions, resource data—all of which must be maintained and organized. In this paper, we propose a novel approach to maintain such non-program entities: By learning from the development history of the product, we can determine coupling between entities: “Programmers who changed ComparePreferencePage.java typically also changed plugin.properties”. As a first proof of concept, our ROSE plug-in for ECLIPSE automatically guides the programmer along related changes. 2.1 Learning from History Shopping for a book at Amazon.com, you may have come across a section that reads “Customers who bought this book also bought. . . ”, listing other books that were typically included in the same purchase. Such information is gathered by data mining— the automated extraction of hidden predictive information from large data sets. We have applied such data mining to the version histories of large open-source software systems. This results in rules like the following: Coupling between entities: “Programmers who changed the fkeys[] field always also changed the initDefaults() function”. The initDefaults() function initializes new elements of the fkeys[] field; whenever fkeys[] was extended by a new element, initDefaults() was extended by a statement that initialized the element. Coupling between programs and documentation: “In 8 out of 10 cases, Programmers who changed the embedded SQL statement in line 47 of status.py changed the JPEG image igordb.jpg”. The JPEG image is part of the product documentation and is a view of the database schema; whenever the schema changed, the SQL statements were changed, too, and the documentation was updated. Such rules can reveal invariants of the development process (such as updating documentation); they can reveal factual coupling through common changes; and they can be put to use for actual programmers. Figure 2.1 shows our ROSE plug-in for the ECLIPSE programming environment, actually working on the ECLIPSE source code: As soon as the programmer makes a change to fkeys[], ROSE suggests further related changes as listed above.