Why I don’t commit generated files to master
Edited on 12 November 2015
Michael Jackson has created a service called npmcdn.com which covers all the use-cases for committing generated files to source control in the first place. I plan on using that service from now on and will no longer be pushing a latest branch with publish-latest as mentioned at the end of this article. Hopefully this article will still explain the issues involved with committing generated files to source control and encourage you to use npmcdn.com as well :-) Cheers!
Edited on 27 December 2015
If you’re curious, the way to get around the bower requirement is to tell people to do this to install your package with bower (using react as an example):
bower install https://firstname.lastname@example.org/bower.zip
npmcdn even supports semver ranges if you’re into that.
This is an age old question and debate. There are pros and cons to both sides of the argument. In this post, I’ll explain my experience committing generated (built) files to master for my libraries and why/how I don’t do that anymore.
First, generated files?
In the context of a library, the generated files would be what the user of your library includes on their page with something like a script tag. So the concatenated, transpiled, and minified version of your library files. For small libraries this may not be needed, but for multiple-file libraries that need to run in the browser of today which doesn’t support ES6 modules yet, you’re going to want to distribute a single file. This is what I mean when I say “generated files.”
So, why do it?
You might be tempted to commit built files to source for a few reasons:
- Bower needs them there (stop using bower)
- Makes it easier for people just downloading the files (not using npm or bower)
- Enables use of rawgit for examples in tools like jsbin or plunker
For a long time, these are the reasons that I committed my generated files for my libraries to my git repository. Everything worked out great! But I did have a few problems.
These are some of the things that I just dealt with because I hadn’t thought of a way to get those benefits without these issues. Until now.
Exploding git changes
Looks like a huge boost of features right? Nope! that bump is when a build was added to ng-stats (there were a few other significant changes as well). The subsequent bumps are whenever a new version was released (and hence the build re-run and re-committed). This make’s it difficult to get a good read of the activity of a project.
Terrible Git Diffs
When developing the library, you have to make sure that you only commit the built files when doing a release (and not with other file changes), otherwise your git diffs are going to be a huge mess and it’ll be hard to identify when and what changes entered in the codebase.
But even having this separation is frustrating at best because diffing two commits that are separated by a build will result in an impossible git diff as well. 😫
Contributions to dist
If you’re writing an open source library, chances are you want to make it easy to contribute to and welcoming to new contributors. I have often received pull requests that make changes to the generated files rather than the source files (or commit a run of the build script). Their contributions would be overwritten next time the build is run! Others will submit a PR that includes the updated dist (which ruins the diff for that commit as mentioned earlier).
For these kinds of contributions, I always have to (kindly) ask the contributor to resubmit the PR with a force push so that commit doesn’t get into my history. This is not a very good experience for contributors 😿
If you want people to contribute to your library, you want to make it as straightforward as possible. In my experience, this is more easily accomplished when not committing the generated files.
So, how do we get the things we like about committing the generated files without these drawbacks? It’s simple, only push the generated files to a special branch for releases. I call this branch “latest” and I’ve automated this with the package publish-latest.
Basically, you run your build, then run publish-latest which runs this script which will commit your built files and push them to your latest branch automatically. You can totally do this locally, but it actually works best when done with semantic-release and TravisCI (or some other CI service).
How do we still get the benefits?
- Bower -> tag the commit that publish-release makes (semantic-release does this for you)
- Downloading -> Look at the latest branch
- rawgit -> Point to the latest branch
How does this solve the problems?
- Exploding git changes -> Just ignore latest, focus on master (that’s what github’s contributing graph does as well).
- Terrible git diffs -> You never need a git diff on the latest branch. It’s just as generated (useless to git) as the code.
- Contributions to dist -> Most people won’t checkout the latest branch to contribute. I’ve never had that problem.
- Does this work with CDNJS? It appears to work just fine with angular-formly
- What about storing the latest branch in an entirely different repository? This is actually not a bad idea, and with the url flag of publish-latest, you can totally do this.