Indicators for Merge Conflicts in the Wild: Survey and Empirical Study
Abstract
With modern version-control systems, while the creation
of new branches and forks is easy and fast, merging is
often time-consuming. Especially, when dealing with
many branches or forks, a prediction of merge costs
based on proper indicators would be desirable to help
developers recognize problematic merging scenarios
before potential conflicts become too severe in the
evolution of a software project.
We analyze the predictive power of several indicators,
such as the number, size, or scattering degree of
commits in each branch, derived either from the
version-control system or directly from the source
code. Based on a survey of 41 developers, we expect 7
indicators being feasible to predict the number of
merge conflicts. We test our hypotheses in a study of
155 open-source projects including 32,579 merge
scenarios and comprising 50 million lines of code.
A notable (“negative”) result is that none of 7
indicators suggested by the developer survey has a
predictive power concerning the frequency of merge
conflicts. We discuss this and other findings as well
as perspectives thereof.
Hypotheses and Results
Hypothesis H1
Active, diverted branches are more likely to result
in conflicts than inactive branches that remain
close to each other.
Scatter plot: number of commits / number of conflictsRejected
Number of commits do not correlate with the number of merge conflicts.
Hypothesis H2
Many commits within a small time span are more
likely to produce conflicts than the same number of
commits over longer time spans.
Scatter plot: commits last week / number of conflictsRejected
Commit density does not correlate with the number of conflicts.
Hypothesis H3
The more files are changed by both branches, the
more likely a conflict occurs.
Scatter plot: files changed by both branches / number of
conflictsRejected
The number of files changed in both branches does not correlate with the number of conflicts.
Hypothesis H4
Larger changes that modify more lines of code are
more likely to cause conflicts than smaller
changes.
Scatter plot: change size / number of conflictsRejected
Large changes are not more likely to lead to conflicts than smaller changes.
Hypothesis H5
More code fragmentation of the committed changes
results in more conflicts than lesser code
fragmentation.
Scatter plot: number of chunks / number of conflictsRejected
Fragmentation of changes is not more likely to result in conflicts.
Hypothesis H6
Scattered changes (across classes or methods) are
more likely to spawn conflicts than cohesive
changes.
Scatter plot: scattering over classes / number of conflictsScatter plot: scattering over methods / number of conflictsRejected
Scattered changes across classes or methods are not related to the number of conflicts.
Hypothesis H7
Changes above the level of class declarations
(which are typically inserted and maintained
automatically) are more likely to lead to merge
conflicts than changes inside class declarations
(introduced by human developers).
Scatter plot: rate of changes above class level / number of
conflictsScatter plot: rate of changes in classes / number of conflictsRejected
Granularity of changes (above or within class declarations) has no influence on merge conflicts.
