Thanks for submitting your first pull request! You are awesome! 🤗

If you haven't done so already, check out Jupyter's Code of Conduct. Also, please make sure you followed the pull request template, as this will help us review your contribution more quickly.

You can meet the other Jovyans by joining our Discourse forum. There is also a intro thread there where you can stop by and say Hi! 👋

Welcome to the Jupyter community! 🎉

Thanks for this PR! I was surprised to see that we don't already have this, which means this is for sure useful :D

Looking and thinking about the code now.

Some immediate thoughts:

1. will this work when BinderHub is not running on kubernetes? Recently @manics and/or @yuvipanda did some work on making that possible. Would be good to know how this interacts. If it would require extra work it doesn't need to be addressed here (can be a new PR) xref Allow running dev setup without kubernetes #1323
2. do we need to have a new traitlet type? The reason I'm asking is because that adds a lot of code for which we'd need some tests, etc. We did the same for the memory spec, so we probably need it. But who knows maybe we got smarter since then and now no longer need it?
3. there are no tests :( can we copy from/be inspired by the memory spec test suite?

I think it'll be fine for non-k8s systems, we can ignore the limits 😄.

However if we want to make it easier to implement limits in non-k8s backends it might be nice to tighten the specification.
e.g. CPU is currently "Can be an integer (1), fraction (0.5) or millicore specification (100m)", whereas we could choose to only allow only of those (either 0-1 or millicore).

@betatim, @manics, thanks for feedback

Tbh I was confused about units in JH/BH, seems like different approach from k8s is used (M, G, K instead of Mi, Gi, Ki), so I was not sure what is "right" here. Due to time pressure, just ended with k8s convention, I'm open to adjust. On one hand currently it feels unnatural to use convention different from k8s, otoh in this case probably is bad idea to use convention that differs from rest of JH ecosystem.

My honest feeling is the whole idea of having these traits for extra validation/translation for different convention breaks least surprise principle, and is if not already a source of some confusion.

If i'm an operator and deploy it to k8s, expectation is that requests and limits follow k8s conventions and be passed verbatim as specified in config to the pod spec, rather than being transformed via traits to some magic unexpected common denominators. Applies to JH as well.

As for tests, I'd be more than glad to write some, as this PR is driven by some internal requirements and we're interested in merging it to upstream to avoid the cost of maintaining fork. Will raise this tomorrow internally.

@noroutine I understand your point! There's no single right answer, some people expect things to follow a standard format that is independent of the underlying implementation, some expect it to follow a low level standard (K8s spec), others expect it to follow a different low-level standard (e.g. systemd uses percentages).

I think the decision should come down to whether we'd like to support limits in future alternative backends. If we decide only K8s matters then using the K8s format is fine, otherwise we should consider how other developers could implement it.

Regardless of the format I think it'd be nice to include a description of it in the help as well as a URL. That guarantees it's always known, even if the URL changes in future.

The BytesSpecification comes from JupyterHub which runs on many platforms, not just kubernetes. I think that helps explain why it isn't "just like k8s". BinderHub re-used it probably because it was easy to reuse it (even though back in the day there was no plan for BH to ever work without k8s.) and because it is consistent within the Jupyter(Hub) eco-system.

For CPU limits: the ones implemented in this PR follow the format kubernetes uses as well as what the docker CLI "natively" understands. This is also the format repo2docker uses.

So unless someone has strong opinions on an alternative format I'd say we are consistent within our ecosystem and with kubernetes. So lets stick with this for now?

It would be great to have some tests, even just simple ones that check that the basics aren't broken. Without this something will sooner or later break this feature and no one will notice.

Hey guys

I've just pushed some tests to cover that trait, let me know if i can do any more from my side to get this through!

Kind regards

@manics @betatim tests has been added and the code looks good to me.

I also think that the config is reasonable in a non-k8s context, it could be relevant to have such requests/limits there as well even though maybe it doesn't - and then it can be documented they will be ignored for such environments later.

Go for merge?

Not sure why the tests are failing :(

If you go to the logs for the failing check: https://github.com/jupyterhub/binderhub/pull/1348/checks?check_run_id=3697325231
and expand Show hub logs you'll see the Hub errors:

[E 210924 07:41:15 web:1789] Uncaught exception GET /services/binder/build/gh/binderhub-ci-repos/cached-minimal-dockerfile/HEAD (::1)
    HTTPServerRequest(protocol='http', host='localhost:8000', method='GET', uri='/services/binder/build/gh/binderhub-ci-repos/cached-minimal-dockerfile/HEAD', version='HTTP/1.1', remote_ip='::1')
    Traceback (most recent call last):
      File "/opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/tornado/web.py", line 1704, in _execute
        result = await result
      File "/opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/binderhub/builder.py", line 402, in get
        self.build = build = BuildClass(
    TypeError: __init__() got an unexpected keyword argument 'cpu_limit'

Work on #1318 went a lot faster than expected! The new parameter needs to be added to

The alternative is to add this as a configurable Trailets on the Build class when it's refactored as discussed in #1318 (comment). That way you can avoid adding a parameter to BinderHub, the configurable should be passed down to the class, and since it can be a configurable on e.g. KubernetesBuilder it can use Kubernetes specific config. The downside is that'll delay the addition of this new feature.

However there's another issue. The build pod runs repo2docker, which does the build in another Docker container (by communicating directly with the Docker daemon through a socket). Since the Docker build takes place outside K8S the K8S pod limits won't apply, they somehow need to be passed to repo2docker.

Based on this comment:

This means cpu_request will effectively reserve some CPU for the build, but cpu_limit will have no effect unless it's also added to repo2docker.

Hi gentlemen

Unfortunate it failed again.

I was a bit out on leave, and checking now above, trying to understand what best action on making this work again.
Will dive into this again this week, but have to understand what shall I do regarding repo2docker code

Thank you

@noroutine You'll need to add cpu limit support to repo2docker, then you can come back here and pass the appropriate flags when repo2docker is run.

Thank you @manics