Chris Morgan

I’m a software developer, dealing mostly in web things.

I’m also (more importantly) a committed Christian—please talk to me at any time about that.

Elsewhere

GitHub: a case study in link maintenance and 404 pages

GitHub is remarkably successful. I certainly owe that much to them. They’ve done a lot of things right and there’s a lot to learn from them—but they’re still not perfect. Here’s a lesson that can be learned from how they’ve done something badly, and what they can do to fix it.

Definitions and assumptions

The topic here is link maintenance. By link maintenance, I mean keeping links working: if http://example.com/example exists now, it should always exist. If for some reason it can’t exist any more, don’t just let it go: it should show a useful error.

I’m going to focus on the most frequent failure case: the HTTP status code 404 Not Found; it’s the most obvious failure case. My feeling on the matter is that the fact that the number 404 is well known among non‐technical users shows that we as web developers and maintainers have failed in our duty.

I also start with the assumption that you agree with W3C that cool URIs don’t change. If you don’t, I’m afraid I can’t help you yet; come back when you do.

The problem

You know GitHub’s 404 Not Found page? If you’ve used GitHub much at all, you’re almost certain to have come across it. My opinion over the past few years is that GitHub has been vying with Microsoft for the “worst maintainer of links on the entire Web” trophy. I’m not sure who’s winning it at present.

Screenshot of the GitHub 404 page: “This is not the web page you are looking for.”
You’ve seen this too often.

Linkrot is unfortunate but inevitable. Certainly for a website like GitHub’s they cannot prevent it in all cases. But if you look at that 404 page, you’ll see that it is almost supremely useless. It’s absolutely generic and tells you nothing about what or why the problem is.

What belongs on an error page?

The most important thing is an indication that something is wrong. Few people make a mess of this. GitHub don’t do this particularly badly.

Unfortunately, that’s where most 404 pages stop. There seems to be an assumption that 404 pages must be useless. Well, I’m here to tell you that that’s not the case. You are allowed to make a useful 404 page.

How about we add these elements to the 404 page:

The GitHub example

How could these apply to GitHub? Let’s start by assuming I’m looking at the URL https://github.com/user/repo/blob/branch/path/file. Uh oh, it’s not working. Now I want to know: what is the actual problem? As it stands, in such a case if I really want to know (and I normally will if I’ve come there, in GitHub’s case—if you’re a different site, you’re much more likely to just luck out and lose me as a visitor), I’ll start going up a directory at a time until I stop getting a 404 page. This is very inefficient for both GitHub and me; the 404 page could do this for me. Here are some of the things that could be wrong with a given path and what could be done about them:

I’m using the example of GitHub in this article. But remember: this article is not about GitHub. Your 404 page is very unlikely to be without spot or blemish—statistically, it’s as generic as GitHub’s, though perhaps a little more useful; you too can improve your 404 page.

Avoiding 404 Not Found where possible

I’ve covered above some of the ways that GitHub’s 404 pages could be made better, while still being a 404 page. Now let’s take it to an even higher level of usefulness: a lot of GitHub’s 404 pages actually shouldn’t be returning 404 Not Found in the first place. And they have to hand a lot of extra data which they can use to achieve this. This section is more specific to GitHub; for many websites it won’t be feasible—but do think about it.

Things get moved

Suppose you have a link to https://github.com/chris-morgan/rust-http/blob/master/src/libhttp/request.rs. Uh oh—that link is broken. How did that happen? It worked a month ago. The answer is that I moved the file src/libhttp/request.rs to src/libhttp/server/request.rs. Translate the move into HTTP parlance and the answer is that the first‐mentioned URL should probably redirect with 303 See Other (it should be a temporary move, not a permanent one—the original file name may come into existence again) to server/request.rs. I don’t say it should just redirect you and not tell you what it’s done; displaying a message that the file got moved in such‐and‐such a commit would be necessary for sanity, but my key point is this: 404 Not Found is probably the wrong behaviour.

Here’s another demonstration that this is the wrong behaviour: what happens when file A is moved to B and a new file is placed at A?

You see, most of the problem here is that GitHub’s link management scheme is not designed for permanence. The links it gives you by default are to the branch, which may be updated, rather than to the changeset which is (hopefully) permanent, but at least never wrong. There’s a conflict of interests here; consistency demands links to immutable changesets. Convenience and keeping up to date demand links to the branch.

This actually reflects Git’s overall philosophy; permanence is strangely something it’s not particularly interested in. Mercurial, on the other hand (my preferred version control system, though I’m fairly comfortable in Git), is designed for complete durability: once something’s in, it’s in for good.

GitHub’s approach has had the additional unfortunate side‐effect that links to a particular line number or range of a file on a branch are of very limited value; they will rapidly become incorrect, and incorrect pages and references are just as bad, often worse, than returning 404 Not Foundreturning the wrong content is still a broken link. (This is a tricky point in HTTP, because the server doesn’t even get sent the hash. I’ll write more about the hash on a later occasion, with a case study of how Stack Exchange got it right.)

This is also reflected in a comparison of GitHub’s and hgweb’s link generation policy:

So, the basic link architecture that GitHub has chosen (and, I presume, will not change) will lead to inevitable and frequent broken links. But there’s still more that can be done to improve the situation for the hapless user.

GitHub has improved their 404 page a little from what it was a few years ago by the inclusion of the search box, but I doubt people actually use it much. No, we need more drastic improvements.

Why would you not make a better 404 page?

I hope I’ve convinced you by now that the humble 404 page can be improved and made useful. If not, I really want to know why—please contact me about it; there is, however, one fairly solid reason which I’m sure is why GitHub don’t have a good 404 page.

Here, you see, is the rub: serving a static 404 page is cheap and easy. Serving a really useful 404 page is probably computationally expensive. It’s also handy to treat them as throw‐away responses which won’t trigger any more work for the server as it’s common for them to be generated and looked at by misbehaving things, and you don’t want to burden your site still more because of the 404 page—you just want to get it out of the way quickly.

GitHub know this, as you’ll notice if you look at the source of their 404 page:

<!--
 
Hello future GitHubber! I bet you're here to remove those nasty inline styles,
DRY up these templates and make 'em nice and re-usable, right?
 
Please, don't. https://github.com/styleguide/templates/2.0
 
-->

Looking at the cited style guide indicates that for error pages they do not allow error pages to make any requests. (Not quite true—it makes an AJAX call to check if you’re logged in so it can let you log in to access hidden repositories.) You see, it’s a good idea to make generating error pages cheap as they will often happen quite a bit.

The end result is that 404.html is 221KB and almost entirely self‐contained. This is a pretty large response (considerably larger than I would consider acceptable for myself) but can be served very cheaply as it’s precomputed. Still more, it won’t trigger more than one extra request for the server when rendered.

It’s fairly obvious that checking for would be more expensive. Still, I think simple analysis would expose a variety of easy‐to‐pick cases where it’ll be worth while doing more checking to give a good result to the user.

OK then, where’s the balance?

You want to make really good and useful error pages—even to fix errors where possible—but you also don’t want to waste money on serving 404s.

My own conclusion is that GitHub would do well to pay more attention to their link maintenance scheme. The way in which they allow links to break with utter abandon distresses me.

As I noted earlier, GitHub added repository redirects earlier this year; that’s a start, but there’s still a lot more that needs to be done. Also, the way that they treat some of this as a “special bonus” is not the attitude I want to see on the web: I want keeping links alive to be the norm rather than the exception. (Yes, I’m idealistic.)

As a counter‐point to the “serving a static 404 page is easy” I mentioned above, I will declare that the benefits of serving a single 404 page are overrated. Remember, any URL resulting in 404 Not Found has already had to check in the database to find that fact out; the server already knows whether the problem was that the repository didn’t exist or that the file didn’t exist. Providing even a slightly more helpful 404 page is unlikely to be a great burden there. Caching of results and common redirects would further reduce the burden of that issue to what I think might actually be almost zero.

Then again, on the matter of a 404 page not being permitted to make further requests: this I feel to be misguided. I have no statistics which could indicate how frequently they serve error pages, but I think there would be various heuristics which could be fairly simply applied to determine whether making a more useful 404 page is worth the effort. As it stands, broken links are likely to make people either keep loading more pages and get mildly miffed, or to get understandably upset.

Conclusion

Link maintenance is hard; the web doesn’t just automatically stay intact; it requires effort on your part. Some types of breakages are unavoidable, but you can make more useful error pages (not just 404) to mitigate the harmful effects. Better still: prevent links from breaking in the first place by monitoring the site. (If you don’t have the time to spare to do a good job of checking your links to prevent breakages, I’m working on something for you—see below.)

This is the end of the matter; all has been heard. Check your logs and take a look at your 404 page. See if you can make common patterns of errors more useful. Better still, prevent links from breaking in the first place.

Incidentally, if you work for GitHub, you might like to share this with your developers mailing list. I’d love to hear thoughts from you in particular.

Comments? Questions? Corrections? If you want to contact me about anything in this post, write to me at @__chrismorgan or email me.