It’s grown in a way that degraded it and that required actual effort. For example:
- The fancy new diff viewer frontend that barely works. Someone wrote that code — it didn’t happen by itself.
- The unbelievably buggy and slow code review frontend (which is surely related to the diff frontend) was added complexity that did not need to happen. Its badness has nothing to do with how many users use it. It’s just bad in a no-scaling-involved way.
- GitHub actions. It’s … bad. I suppose there wasn’t a predecessor that was better.
> And it's had to expand into the AI field, which is not an incremental thing like "hey let's launch a new feature or better dashboards."
No, it did not have to expand into the AI field. A competent AI-free GitHub Core that could have an optional AI layer on top would have worked just fine if not dramatically better than the current mess.
(I say this as a paying user who will probably cancel soon. The Copilot reviews are kind of nice, but they’re not any better than a third-party system, and I’m getting sick of GitHub not working. Plus, the repos I’ve already migrated off of GitHub get to have nice non-AI things like gasp service accounts.)
Im an outsider and a layman, so this might be totally off base, but...
The way I hear people talking about github reliability doesnt sound like scaling problems to me. If you drive 20 miles every day but then decide to drive 2000 miles and run out of gas, thats a problem of scale. If you drive 2000 miles and your engine explodes, thats a problem of design.
Maybe their design problems are being made evident because of sudden scale, but they're still design problems.
Car analogies are typical, so I'll add in there.
My car can take the four of us, and we can load it up with things from the shops. I can put a bunch of heavy tins of food in there, or some DIY things, but if I put several tons of stones in the boot it'll totally fuck it up.
Is that a design problem?
Not really, it's a relatively cheap regular car, and it failed at a certain scale.
It would be a design problem if it were a flatbed truck, despite it being the same scaling that showed the problem.
Making my car resilient enough to take that weight would require tradeoffs that would either make it worse for other jobs I want it to do or at least add significantly to the cost.
This is similar in engineering software systems too, you can make it handle scaling up better, but this can require a much more complex architecture that may make it slower at smaller scales. It can make it more complicated to work with, add additional risks of failure as well.
There might not have been a predecessor, but it's been obvious for at least a decade that GHA are a very poorly designed programming language, yet nothing was done to improve. They introduced Github Apps that solve many of the issues with Actions, but that requires deploying a service and aren't anywhere near the ease of use of Actions.
Maybe a hot take: no, it didn't.
I think it had all the pieces (api,cli,etc.) already that it would've still be very useful in an AI world without deeply integrating AI things (copilot, etc.). I'd take higher availability over AI features any day.
Yes, and what Github wants public github.com to be is free QA for Github Enterprise. My company is a paying customer with 200 engineers and it's pretty clear we're just Guinea pigs for the Enterprise product.
Isn't this the massive problem? You're trying to do everything, and you can't, and you're trying to do it for everyone all at once, and have tied it all together so much that scaling up gets worse. If it's more than twice as hard to cope with twice the use, then you have to charge a bunch more to customers as you grow - and that's for your customers to get no actual benefit.
> GitHub is not perfect but I don't think it's "degraded faster" at all. It's _grown_ faster.
The experience has degraded. It's really, really bad. I've seen companies spending thousands and thousands of dollars weekly in developer time *hitting rerun on broken actions*. It's so expensive to start with then so expensive in how awful it is to use.
Something I really don't get I guess is what out of all of this actually needs to be cross-project. How much of my github use needs access to something that isn't running on the same machine? I worked with a team building things actively, maybe 20 devs? That's not really a large set of users. Let's say 10 devs with the workload of 20, the cheapest plan would be $40/mo, enterprise would be ~$200. Would ten heavy users really max out a 64GB ram, 6+8 core new i5 with dual nvme drives, a gigabit connection and unlimited traffic? That's about $40 at hetzner for a box.
I'm not arguing a big federated position, I just don't really get why some of these enormous companies need to be so centralised. It feels like the problem is trying to be a big interlinked thing, and failing at it. The only things I can think of are
1. Links between issues
2. Accounts
3. Search
The first is mostly solved with literally just links, accounts isn't a huge problem and search is fair enough - but search is utterly awful and I cannot find things within one single repo or organisation reliably. So global stuff is irrelevant.
> And it's had to expand into the AI field, which is not an incremental thing like "hey let's launch a new feature or better dashboards." Nobody knows what AI wants to be when it grows up
If github persists in being utterly shit for developers, it won't be around to find out. I'm not sure at all what part of the AI stuff needs to make everything else bad, and I'm extremely bullish on AI and agentic coding.
To really hammer this last point home, as agentic coding means we can do a lot more and faster - the unreliability of github has become much more apparent and impactful. Unreliable tests, unreliable code pulling and pushing, unreliable diffs. You're making the agents jobs harder, making the devs jobs harder exactly in the place they now spend much more time.
It makes github dramatically more expensive as a place to work. Also just really fucking annoying.
I think the general answer is that it would take real development effort to make federation work, and having to have compatibilty with other installations slows down your own pace of possible features -- I think these things are undeniable. Arguably worth it for society/the community (I wish we had more open standard federation and less centralization), but from the point of view of the company will it actually lead to increased profits sufficient to justify? In fact, it may do the opposite, if you are one of the largest, then lock-in is better for your profitability. Compatibility with other services is only important for the small upstarts trying to get customers from the largest.
I don't like it, but I think we will get proprietary centralization as long as we have capitalism of the sort we have.