Is there an advantage of branching in version-controlled paper writing?

Question

I use a version control system (formerly subversion, now git) to write papers, which for me works very well and is convenient. A few times, I even worked with branches for different logical versions of a paper, such as versions submitted to different journals. However, I found branching of little use for paper writing. When I used different branches for different journal submissions, I found myself working only on the "latest" branch anyway, since I can anyway only submit one version per time, and, if it is rejected, will not continue to work on that particular version at a later time. I didn't find any other good use of branching in paper writing, and later skipped it completely.

Is there any advantage of working with several branches when writing a paper with a version control system? How should one organize the writing process in order to profit from such an advantage?

Answer 1

I use HG, but it is essentially the same as GIT, in that branching is easy. I branch all the time when working with coauthors. The branches tend to be short lived and merged pretty quickly. Basically, I distribute a copy to my coauthors and each set of comments I get back form different branches. This lets me see each authors comments individually. I then work on merging the comments back into the main trunk. Sometimes the merge is easy and the branch is one commit long. Other times it might take longer if I need to work out something (e.g., math, model, or analysis) which might require a couple of revisions before it is ready to go into the main trunk.

Answer 2

Personally, I've never used branches (for papers), and would certainly not use branches for different chapters (WTF? like keeping different branches for different files in a programming project; except for the case when different collaborators edit different parts of documents, then - maybe), but maybe it may work for:

• arXiv/non-arXiv version (or e.g.: working version with notes vs release candidate) of a paper (the exact use-case of projects with some differences, from minor to e.g. also additional sections, see also arXiv preprint and final paper differing by sections and appendices),
• Beamer slideshows for different talks based an similar slides.

Related: git + LaTeX workflow at SO (so, use "advisor" branch):

Branches are also extremely helpful if you are a graduate student. As any grad student will attest, the advisor is bound to have numerous corrections, most of which you don't agree with. Yet, you might be expected to atleast change them for the time being, even if they are reverted later after discussions. So in such cases, you could create a new branch advisor and make changes to their liking, at the same time maintaining your own development branch. You can then merge the two and cherry pick what you need.

;)

Answer 3

In principle, branches seem like a good idea. A potential use case would be:

• working on a draft of a paper
• as we get close to the conference deadline, putting the paper in the conference format and reducing to the prescribed size
• finding typos in this version and fixing them
• merging the typos back into the "main version"

In practice, I've never done this, because I'm not organized enough. Also, svn is not as friendly to this workflow as git is, for other reasons.

Answer 4

An alternative to branching in some cases, especially for short lived branches like for journal submission, is using patch queues. Mercurial has mq. I believe Git has similar things, but I'm not a git user.

As the question says:

When I used different branches for different journal submissions, I found myself working only on the "latest" branch anyway, since I can anyway only submit one version per time, and, if it is rejected, will not continue to work on that particular version at a later time.

I have had similar experiences when using different (named) branches for submission to different journals. The problem is that one ends up having dead branches which correspond to journals the paper was rejected from, or perhaps journals you never ended up applying to. These live forever in history, which is annoying and sub-optimal.

The advantage of using patches, at least with mq, is that one can maintain multiple patch queues, each of these can be managed as a distinct mercurial repository, and they do not become part of the main repository's history unless you want them to, though of course you can push them to remote since they are regular repositories. Also, one can use them with general non-vcs tools like patch and quilt, which is occasionally useful.

However, when applied, these patch queues are a bona-fide part of a mercurial repository, and can be treated as (anonymous) branches. So, suppose one has two patch queues Q1 and Q2. Then if wants to work with both versions simultaneously, one can make a clone of the main repository, thus resulting in two identical copies of the repository. Then one can apply Q1 and Q2 to the copies, and then work with them as one would with regular branches, using mercurial's merge machinery and so forth.

Another use of patch queues which is very useful, and not specific to paper writing, is to queue up small changes in the queue till one is ready to commit them. I usually just stick everything in one patch, though one could divide the changes into multiple patches. Then, when one is ready to apply some or all of the changes, one can do

hg qref -X .
hg qpop --keep-changes

to apply (part of) the patch as local changes, and when done one can resync the patch with

hg qref
hg ci --mq

See the loosely related questions What's the Git approach to publish a patch queue? and git equivalent to hg mq?.