Others have pointed out that there are architectural steps you can take to minimize this pain, like keeping all CI operations isolated within scripts that can be run locally (and treating GitHub Actions features purely as progressive enhancements, e.g. only using `GITHUB_STEP_SUMMARY` if actually present).
Another thing that works pretty well to address the feedback loop pain is `workflow_dispatch` + `gh workflow run`: you still need to go through a push cycle, but `gh workflow run` lets you stay in development flow until you actually need to go look at the logs.
(One frustrating limitation with that is that `gh workflow run` doesn't actually spit out the URL of the workflow run it triggers. GitHub claims this is because it's an async dispatch, but I don't see how there can possibly be no context for GitHub to provide here, given that they clearly obtain it later in the web UI.)
2. Don't have logic in your workflows. Workflows should be dumb and simple (KISS) and they should call your scripts.
3. Having standalone scripts will allow you to develop/modify and test locally without having to get caught in a loop of hell.
4. Design your entire CI pipeline for easier debugging, put that print state in, echo out the version of whatever. You don't need it _now_, but your future self will thank you when you do it need it.
5. Consider using third party runners that have better debugging capabilities
Of course, the platforms would rather have you not do that since it nullifies their vendor lock-in.
For me what worked wonders was adopting Nix. Make sure you have a reproducible dev environment and wrap your commands in `nix-shell --run`, or even better `nix develop --command`, or even better your most of your CI tasks derivations that run with `nix build` or `nix flake check`.
Not only does this make it super easy to work with Github Actions, also with your colleagues or other contributors.
Note that I don't really use github actions much but have heard about its architecture
From my understanding, I feel like Github actions should just be a call to some bash or python file. Bash has its issues so I prefer python
I recommend people to take a look at https://paulw.tokyo/standalone-python-script-with-uv/ and please tell me if something like this might be perfect for python scripts in github actions as this script would automatically install uv, get all the dependencies of python and even the runtime I think and then execute the python code all while being very managable usually and it can run locally as well
The only Issue I feel like I might have with this is say why go something with this complex when bash exists or the performance concerns of installing uv but considering its github actions, I feel like the latter is ruled out.
Bash is good as well but bash has some severe limitations. and I feel like Python can be good case for something like this plus its ecosystem is a bit mature and you could even create web servers or have some logs be reported to your custom server or automate just basically everything
To me this script feels like the best of both worlds and something genuinely sane to build upon.
gg watch action
Finds the most recent or currently running action for the branch you have checked out. Among other things.
Sounds strange to say as someone who has a product that is built around making GitHub Actions exponentially faster to close these feedback loops faster.
But I can honestly say it's only really possible because the overall system with GitHub Actions is so poor. We discover new bottlenecks in the runner and control plane weekly. Things that you'd think would be simple are either non-existent, don't work, or have terrible performance.
I'm convinced there are better ways of doing things, and we are actively building ideas in that realm. So if anybody wants to learn more or just have a therapy session about what could be better with GitHub Actions, my email is in my bio.
Looking at these flaws, running workflows from a persistent VM of ur own becomes pretty tempting because you don't need to copy caches around and can easily SSH in.
The most important advice is probably to put as much code as possible into locally runnable scripts written in a cross-platform scripting language (e.g. Python or Node.js) to avoid 'commit-push-ci-failure' roundtrips.
Only use the GH Actions YAML for defining the runtime environment and job dependency tree.
GitHub action is a totally broken piece of s !! I know about that broken loops cause I had to deal with it an incredible number of times.
I very often mention OneDev in my comments, and you know what ? Robin solved this issue 3 years ago : https://docs.onedev.io/tutorials/cicd/diagnose-with-web-term...
You can pause your action, connect through a web terminal, and debug/fix things live until it works. Then, you just patch your action easily.
And that’s just one of the many features that make OneDev superior to pretty much every other CI/CD product out there.
> For the love of all that is holy, don’t let GitHub Actions
> manage your logic. Keep your scripts under your own damn
> control and just make the Actions call them!
I mean your problem was not `build.rs` here and Makefiles did not solve it, was your logic not already in `build.rs` which was called by Cargo via GitHub Actions?The problem was the environment setup? You couldn't get CUE on Linux ARM and I am assuming when you moved to Makefiles you removed the need for CUE or something? So really the solution was something like Nix or Mise to install the tooling, so you have the same tooling/version locally & on CI?
I was convinced GH actions was best practice and it was normal to waste hours on try-and-pray build debugging, until one day GH actions went down and I ran deploys from my laptop and remembered how much better life can be without it..
(Solo dev here - but opensource CI on an EC2 instance can be just as nice)
I write a piece of code, however small the change. Then run the proc, first it takes ~20 mins to compile. Then since it is ML, it can only run on a remote server. It takes an easy 20 more minutes to start running in the remote server. Then after another 40 minutes I can confirm the job failed. The only way to debug is to read through a massive log file, for which we have an in built log reader lol. The log file will have thousands of errors that literally don’t matter, and you’ll never have enough context to know which errors don’t matter. You can simply ping someone and ask, does this error matter, could this be the reason my entire proc failed, oh no this is just a useless log that fails all the time, I shouldn’t have been wasting a day digging into this, ok thanks bye.
But this isn’t it, not even close lol, we in fact have a custom DSL to define computational graphs, which of course does not have any linter, or even any compiler and a very broken visualizer but our entire org runs on this. Syntax errors, logic errors, actual race conditions are all caught the exact same way —- as the process dying after trying to run on a remote server for 2+ hours with no useful error log. So our workflow is to just get a cup of coffee and stare at the graph which can get thousands and thousands of lines big to find at times completely trivial bugs that any half decent language would have caught as a syntax error.
My exp in BigTechCo makes me completely understand GitHub actions, my guess is many big companies have equally janky tools that harm dev productivity but still somehow take absolutely massive workloads. GitHub just thought of sharing it to the public.
Who here has been thinking about this problem? Have you come up with any interesting ideas? What's the state of the art in this space?
GHA was designed in ~2018. What would it look like if you designed it today, with all we know now?
Also, GitHub actions itself just breaks sometimes, which is super annoying. A couple of weeks ago, half of all the macOS runner images broke when using GitHub's caching system. Seems like that would have been caught on GitHub's side before that happened, but what do I know!
There was a time I wanted our GH actions to be more capable, but now I just want them to do as little as possible. I've got a Cloudflare worker receiving the GitHub webhooks firehose, storing metadata about each push and each run so I don't have to pass variables between workflows (which somehow is a horrible experience), and any long-running task that should run in parallel (like evaluations) happens on a Hetzner machine instead.
I'm very open to hear of nice alternatives that integrate well with GitHub, but are more fun to configure.
or try something like https://docs.dagger.io/use-cases#portable-ci
I like being able to run self-hosted runners, that is a very cool part of GitHub Actions/Workflow.
I appreciate all the other advice about limit my yamls to: 1) checkout, 2) call a script to do the entire task. I am already half-way there, just need to knuckle-down and do the work.
I was dismayed that parallel tasks aren't really a thing in the yaml, I wanted to fanout a bunch of parallel tasks and I found I couldn't do it. Now that I'm going to consolidate my build process into a single script I own, I can do the fanout myself.
> For the love of all that is holy, don’t let GitHub Actions
> manage your logic. Keep your scripts under your own damn
> control and just make the Actions call them!
The pain is real. I think everyone that's ever used GitHub actions has come to this conclusion. An ideal action has 2 steps: (1) check out the code, (2) invoke a sane script that you can test locally.Honestly, I wonder if a better workflow definition would just have a single input: a single command to run. Remove the temptation to actually put logic in the actions workflow.
I've seen this over and over again over the years in projects where things like ant, maven, gradle, puppet, ansible, etc. Invariably somebody tries to do something really complicated/clever in a convoluted way. They'll add plugins, more plugins, all sorts of complex configuration, etc. Your script complexity explodes. And then it doesn't work and you spend hours/days trying to make it do the right thing and fighting a tool that just wasn't designed to do what you are doing well. The problem is using the wrong tool for the job. All these tools have the tendency to evolve into everything tools. And they just aren't good at everything. Just because it has some feature doesn't mean it's a good idea to use it.
The author actually calls this out. Just write a bash script and run that instead. If that gets too complicated pick something else more appropriate to the job. Python, whatever you like. The point here is to pick something that's easy for you to run, test, and debug locally. Obviously people have different preferences here. If you are shoe horning complex fork/join behavior, conditional logic, etc. into a Yaml / CI build, maybe simplify your build. Yaml just isn't suitable as a general purpose programming language.
Externalizing the complex stuff to some script also has the benefit of that stuff still working if you ever decide to switch CI provider. Maybe you want to move to Gitlab. Or somebody decides Jenkins wasn't so bad after all. The same scripts will probably be easy to adapt to those.
One of my current github actions basically starts a vm, runs a build.sh script, stops the vm. I don't need a custom runner for that. I get to pick the vm type. I can do whatever I need to in my build.sh. I have one project with an actual matrix build. But that's about the most complex thing I do with github actions. A lot of my build steps are just inline shell commands.
And obligatory AI comment here, if you are doing all this manually, having scripts that run locally also means you can put claude code/codex/whatever to work fixing them for you. I've been working on some ansible scripts with codex today. Works great. It produces better ansible scripts than me. These tools work better if they can test/run what they are working on.
i just checked and in 2025 there was at least 2 outages a month every month https://x.com/swyx/status/2011463717683118449?s=20 . not quite 3 nines.
How about writing a separate repo and testing it separately
Keywords: reusable workflow/actions
For the script getting run, there's one other thing. I build my containers locally, test the scripts thoroughly, and those scripts and container are what are then used in the build and deploy via Action. As the entire environment is the same, I haven't encountered many issues at all.
As soon as I need more than two tries to get some workflow working, I set up a tmate session and debug things using a proper remote shell. It doesn't solve all the pain points, but it makes things a lot better.
This way we can test it on local machine before deployment.
Also as other commenters have said - bash is not a good option - Use Python or some other language and write reusabe scripts. If not for this then for the off chance that it'll be migrated to some other cicd platform
Being able to run your entire "pipeline" locally with breakpoints is much more productive than whatever the hell goes on in GH Actions these days.
Doc'd it here: https://revolveteam.com/blog/goa-radicle-ci/
Like most of the glaring nonsense that costs people time when using msft, this is financially beneficial to msft in that each failed run counts against paid minutes. It's a racket from disgusting sleaze scum who literally hold meetings dedicated to increasing user pain because otherwise the bottom line will slip fractionally and no one in redmond has a single clue how to make money without ripping off the userbase.
I prefer an api with documented contracts between its abstractions
Main issue is Rust. Writing catchy headlines about hating something may feel good, but a lot of people could avoid these pains if
- zig cc gets support for new linker flag that Rust requires https://codeberg.org/ziglang/zig/pulls/30628 - rust-lang/libc gets to 1.0 which removes iconv issues for macos https://github.com/rust-lang/libc/issues/3248
But almost every company uses GitHub, and changing to Bitbucket isn't usually viable.
And that’s where there’s a Mac Studio that sits sadly in the corner, waiting for a new check in so it has something to do.
Granted, if you are working on "Windows 12", you won't be building, installing, testing, and deploying that locally. I understand and acknowledge that "as tight as possible" will still sometimes push you into remote services or heavyweight processes that can't be pushed towards you locally. This is an ideal to strive for, but not one that can always be accomplished.
However, I see people surrender the ability to work locally much sooner than they should, and implement massively heavyweight processes without any thought for whether you could have gotten 90% of the result of that process with a bit more thought and kept it local and fast.
And even once you pass the event horizon where the system as a whole can't be feasibly built/tested/whatever on anything but a CI system, I see them surrendering the ability to at least run the part of the thing you're working on locally.
I know it's a bit more work, building sufficient mocks and stubs for expensive remote services that you can feasibly run things locally, but the payoff for putting a bit of work into having it run locally for testing and development purposes is just huge, really huge, the sort of huge you should not be ignoring.
"Locally" here does not mean "on your local machine" per se, though that is a pretty good case, but more like, in an environment that you have sole access to, where you're not constantly fighting with latency, and where you have full control. Where if you're debugging even a complex orchestration between internal microservices, you have enough power to crank them all up to "don't ever timeout" and attach debuggers to all of them simultaneously, if you want to. Where you can afford to log every message in the system, interrupt any process, run any test, and change any component in the system in any manner necessary for debugging or development without having to coordinate with anyone. The more only the CI system can do by basically mailing it a PR, and the harder it is to convince it to do just the thing you need right now rather than the other 45 minutes of testing it's going to run before running the 10 second test you actually need, the worse your development speed is going to be.
Fortunately, and I don't even how exactly the ratio between sarcasm and seriousness here (but I'm definitely non-zero serious), this is probably going to fix itself in the next decade or so... because while paying humans to sit there and wait for CI and get sidetracked and distracted is just Humans Doing Work and after all what else are we paying them for, all of this stuff is going to be murder on AI-centric workflows, which need tight testing cycles to work at their best. Can't afford to have AI waiting for 30 minutes to find out that its PR is syntactically invalid, and can't afford for the invalid syntax to come back with bad error messages that leave it baffled as to what the actual problem is. If we won't do it for the humans, we'll do it for the AIs. This is definitely not something AI fixes, despite the fact they are way more patient than us and much less prone to distraction in the meantime since from their "lived experience" they don't experience the time taken for things to build and test, it is made much worse and more obvious that this is a real problem and not just humans being whiny and refusing to tough it through.
Ansible. Same reasons but 100x worse.
Jira, Microsoft Teams, HP Printers, Ticketmaster, …
No. It's cargo cult science.
Original plan: You have an oven in another building that automatically bakes your cake. But the oven needs a mixer, and every time you bake, it has to wait 2-3 minutes for someone to bring the mixer from storage.
Problem: For one specific oven (Linux ARM), the mixer never arrives. So your cake fails. You keep trying different ways to get the mixer delivered. Each attempt: 2-3 minutes wait.
What you finally do: Stop waiting for the mixer to be delivered. Just mix the batter at home where you already have a mixer. Send the pre-mixed batter to the other building. Now the oven just bakes it - no waiting for the mixer.
Translation: Stop trying to generate files in GitHub Actions (where it takes 2-3 minutes each time). Generate them locally on your computer where you already have the tools. Upload the finished files. GitHub Actions just uses them.
Sometimes "pre-mix the batter at home" beats "wait for the mixer every single time."
This is why I started DSCI - https://github.com/melezhik/DSCI
One can replace GitHub actions yaml pipeline runner by general purpose languages/scripts but still using all other standard features of GitHub server
1. Here's your goal "...", fix it, jj squash and git push, run gh pr checks --watch --fail-fast pr-link
Not by GitHub, but isn't act supposed to be that?
Took me a while to figure that out. While I appreciate occasional banters in blog articles, this one seems to diverge into rant a bit too much, and could have made its point much clearer, with, for example, meaningful section headers.
> Sure, I still make fun of the PHP I remember from the days of PHP 4.1, but even then I didn’t hate it. (ref. And “PHP: Training Wheels Without a Bike” is still in the Top 10 of my favorite memes.)
I still use it to build my latest projects, and for me it’s like breathing. Simple and functional. It’s not perfect, but no language is.
> Linux ARM
> macOS ARM
> Linux x86_64
> macOS x86_64
Oh, here we go again. Java was invented to solve that, "Write once run everywhere" :] I.e. `int` means `i32` on all platforms, no `usize`.