bot for refreezing

[minrk]

Proposed change

Every once in a while, we want to rebuild the base environments. This is done by running the freeze.py script in the conda buildpack.

Alternative options

keep manually refreezing

Who would use this feature?

Maintainers of repo2docker and folks contributing updates to the base environments who don't want to run refreeze themselves.

One way is to do periodic refreeze, to keep things up to date, as is done with henchbot. Another is to allow pull requests to update the environment spec or other files and then post a comment like @r2d-bot please refreeze.

How much effort will adding it take?

Not sure! Hopefully not too much. Requires writing a bot that talks to the GitHub API. It would need to:

1. respond to a trigger somehow (e.g. a @refreeze-bot please refreeze comment to update a PR or a new please refreeze Issue to trigger a new PR)
2. make a commit in an existing pull request and/or open a new pull request

Who can do this work?

• Interest in / knowledge of the GitHub API
• Interest in / knowledge of writing bots and deploying them somewhere (on heroku, etc.)

References:

• henchbot which does pull requests for updates to repo2docker and binderhub for mybinder.org
• conda-forge-admin bot which does things like add commits to pull requests to run common commands like @conda-forge-admin please rerender

[henchc]

I'd be happy to take this up, I've been wanting to get more into repo2docker but hadn't found a good entry point yet.

I don't see a "refreeze.py" just "freeze py" in the conda and legancy buildpacks. Am I missing something?

Also, does this depend on builds existing outside of the state of the GitHub repo? One of the reasons henchbot just goes every hour is because the images maybe hadn't made it to Dockerhub yet, so it checks there specifically.

[betatim]

I think it is a typo. freeze.py is what we use to (re)freeze the environments.

I think it uses a miniconda docker image so we shouldn't have the problem where we need to wait for an image to build.

[minrk]

Am I missing something?

@betatim is right. I updated the description: We only want to run freeze.py in the conda buildpack. This will regenerate the eight environment.pyX.Y[.frozen].yml files.

does this depend on builds existing outside of the state of the GitHub repo?

Depends, but mostly no in that there's nothing outside to check. There's no outside event to "check if things are up-to-date", since it's all the packages. If we're doing periodic checks, we should maybe perform a diff of the resulting yaml files (the parsed data, not the bytes on disk), though, because the files will change with the Generated on <DATE> comment, even if the content doesn't change.

Unlike mybinder.org, we don't want this to be constantly updating (a PR for every tiny change would be overkill). Rather, prompting to keep it up-to-date every one, two, or even four weeks.

More useful, I think, will be the bot adding the freeze commit to an existing PR, e.g. one that bumps a dependency in environment.yml, letting the bot do the freezing step. See the two commits in #354. For the workflow I have in mind, see this PR on conda-forge where there's one commit by hand changing the input, then a second commit from a bot to regenerate derivative files.

[henchc]

Thanks @minrk ! I will start some work on it this week.

[henchc]

So I've been working on this when I get the chance. The bot can run the freeze and create the PR, I'm still working on the PR text.

My next step question will be if anyone knows how we can run this in an environment that has the required Docker? Can Heroku do this? @betatim ?

[betatim]

I don't think on heroku you can access a docker daemon :-/

[trallard]

Is this still open? If so I would love to give it a go with some Github actions stuff

[betatim]

I think it is!

[trallard]

Cool, so I got a decent start on this - been trying on a repo of my own as I am testing with the PRs and comments.

As I am working this is the workflow I have come up with based on @minrk description/references:

1. Someone submits a PR to repo2docker (change environment.yaml for the conda buildpack)
2. Someone uses trigger phrase in the PR comments to refreeze based on the changes above (only works if the path repo2docker/buildpacks/conda has committed changes (😂 currently using @bot winter-is-coming) )
3. Bot commits the newly generated files (or rather updated from the freeze) on top of the PR

I do have some questions:

• For step 2 do you want to allow anyone to be able to run the command?
• Also, do you want the bot to add a comment with the outcome of the freeze or something similar?
• Do you have any preference regarding the format or message for the bot?

[minrk]

For step 2 do you want to allow anyone to be able to run the command?

Can it be anyone with write access to the PR branch (i.e. author + repo maintainers)?

Also, do you want the bot to add a comment with the outcome of the freeze or something similar?

Maybe only on fail with error messages? At the start, at least, a confirmation message that it has received the request and begun the process might be good, but I suspect we won't want this message once it's running reliably.

Do you have any preference regarding the format or message for the bot?

How about @bot please refreeze?

[manics]

In #958 (WIP) I've got an update.py script that automatically updates one or more dependencies in environment.yml to the latest on conda-forge. I've been playing around with GitHub workflows to try and automate the freeze.py (mostly for my personal curiosity), but if it's useful feel free to take the script.

[trallard]

Right, I was working on this - and got then forgotten as I went on annual leave and been swamped with JupyterCon stuff.

Anyway, I have a working MVP (had to create a separate repo to enable actions though). Example here. The one piece missing is pushing the refreezed yaml files to the PR head instead of master.
Which I can do on Friday as it's my meeting-free day.

Would it be then suitable for adopting @manicsupdate.py for the dependencies update?