GitHub dropped Pygments

:: Racket, software

My first-ever open source contribution, a couple years ago, was to a project called Pygments. My motivation? GitHub was displaying Racket source code poorly. Pygments didn’t have a Racket lexer. GitHub was using a Scheme lexer for Racket code. The Scheme lexer was highlighting square brackets in red as an “error”. This was really distracting and ugly.

I contributed a new Racket lexer to Pygments, and waited for that to roll into a Pygments release and in turn be deployed on GitHub. Finally Racket code looked good! Later Dave Corbett substantially improved the Racket lexer beyond my small start.

A few days ago, I was confused to see that Racket code was displaying poorly again on GitHub. The square brackets were highlighted in red as errors — again??

Cartoon-me’s thought balloons: WAT, OMFG, FML, &c. Why are we going in circles?

Someone submitted an issue against pygments.rb, which is GitHub’s Ruby library to use the Python Pygments library. I commented on that issue, while it began to dawn on me that GitHub wasn’t using Pygments anymore. Which led to me submitting this issue against linguist.

The GitHub folks responded quickly, especially during the weekend. They proposed a reasonable way to at least improve the status quo — use an existing TextMate lexer.

In issue 1717’s comments, it didn’t seem appropriate for me to editorialize beyond the issue itself. But I can editorialize here on my blog.

I’m sure GitHub had what they felt were compelling reasons for abandoning Pygments. My guesses:

Again, these guesses — speculation based on my experience with software projects, but speculation nonetheless. I don’t have any insider information. Hopefully at some point GitHub will share their thinking. So far they haven’t, which leads to the next point.

Regardless of the reasons, the change was deployed without advance notice, as far as I can tell. The approach seemed to be, if things break, people will report it and we’ll fix it. In other words, we’ll fix on-demand what people care about, rather than proactively worry about things maybe no ones cares about.

That is actually a very valid and reasonable approach in some situations — particularly older, “legacy” applications with a huge surface area and user attrition. Although that’s not what GitHub seems like to me on the outside, maybe that’s how it feels to them internally, I don’t know.

But in any case, here’s the problem. Although people use GitHub to collaborate1, many also use it as a portfolio. It is part of how people hope to get a job, or persuade developers to try a new language, or whatever. To suddenly doink the appearance of people’s portfolios is unfortunate.

In dropping Pygments, I’m sure GitHub realized they’d lose dozens of lexers that many people have worked to contribute over the years. That negative must have been outweighed, in their opinion, by positives.

I’m not sure they realized that they’re also losing a well-documented, relatively easy process for contributing lexers. The Pygments project makes it clear and simple for people who want to add or improve a lexer. Documentation and help are available. Lexers are written in simple Python, which is arguably easier than wrangling plist XML or even JSON.2 There are probably a hundred example lexers to learn from. And finally, there is a process to review and vet the lexers, before they reach GitHub production. Well, past-tense: “reached”.

Effectively GitHub has opted to take all this upon themselves, now. If they do it as well as Pygments, great, but it’s a lot of work for them and a lot of redundancy. On the other hand if they don’t do it as well, it’s bad for all the not-top-twenty, not-hip-this-year languages.

Anyway, that’s my first reaction. Maybe this decision makes sense in a way that I just don’t understand yet.

  1. In which case you could argue that, outside of pull requests, people mostly view the project code locally, so the code formatting doesn’t matter quite so much. 

  2. The Racket lexer eventually grew to nearly 1,000 lines of Pyton. Granted the bulk is lists of keyword and built-ins. Even so, it seems that converting it to a plist or cson representation is going to multiply the line count by 3, 5, maybe 10. If such a conversion were straightforward, perhaps GitHub would have implemented that?