Almost Rerere: Learning to Resolve Conflicts in Distributed Projects

The concurrent development of applications requires reconciling conflicting code updates by different developers. Recent research on the nature of merge conflicts in open source projects shows that a significant fraction of merge conflicts have limited size (one or two lines of code) and are resolved with simple strategies that use code present in the merged versions. Thus the opportunity arises of supporting the resolution of merge conflicts automatically by learning the way in which developers fix them. In this paper we propose a framework for automating the resolution of merge conflicts which learns from the resolutions made by developers and encodes such knowledge into conflict resolution rules applicable to conflicts not seen before. The proposed approach is text-based, does not depend on the programming languages of the merged files and exploits a well-known and general language (search and replacement regular expressions) to encode the conflict resolution rules. Evaluation results on 14,872 conflicts from 25 projects show that the system can synthesize a resolution for
$\approx$
49% of the conflicts occurred during the merge process (
$\approx$
89% if one considers conflicts that have at least one similar conflict in the data set) and can reproduce exactly the same solution that human developers have applied in
$\approx$
55% of the cases (
$\approx$
62% for single line conflicts).