A longitudinal study of programmers' backtracking

Programming often involves reverting source code to an earlier state, which we call backtracking. We performed a longitudinal study of programmers' backtracking, analyzing 1,460 hours of fine-grained code editing logs collected from 21 people. Our analysis method keeps track of the change history of each abstract syntax tree node and looks for backtracking instances within each node. Using this method, we detected a total of 15,095 backtracking instances, which gives an average backtracking rate of 10.3/hour. The size of backtracking varied con-siderably, ranging from a single character to thousands of char-acters. 34% of the backtracking was performed by manually deleting or typing the desired code, and 9.5% of all backtracking was selective, meaning that it could not have been performed using the conventional undo command present in the IDE. The study results show that programmers need better backtracking tools, and also provide design implications for such tools.