Expose PEP 740 attestations functionality

[woodruffw]

WIP, still experimenting here. Not ready for review 🙂

This adds PEP 740 attestation generation to the workflow: when the Trusted Publishing flow is used, this will generate a publish attestation for each distribution being uploaded. These generated attestations are then fed into twine, which newly supports them via --attestations.

Breakdown:

• Generate attestations for each dist
• Actually hook up twine --attestations
• Documentation

See: pypi/warehouse#15871

[woodruffw]

I've confirmed that the basic version of this works as expected (attestations is currently ignored by PyPI and TestPyPI): https://github.com/woodruffw-experiments/gha-pep740-experiments/actions/runs/9008318313/job/24750093954

[webknjaz]

@woodruffw I just bumped Twine FYI. And pre-commit is fine now. Rebasing should get the blockers out of the way.

[woodruffw]

@woodruffw I just bumped Twine FYI. And pre-commit is fine now. Rebasing should get the blockers out of the way.

Good timing, so did @facutuesca 😅

[facutuesca]

I'm squashing and rebasing this branch now

[facutuesca]

@woodruffw @webknjaz The remaining lint failure is due to an error message string (_TOKEN_RETRIEVAL_FAILED_MESSAGE) being duplicated in attestations.py and oidc-exchange.py.
Would you prefer de-duplicating it by moving it into a new common.py file, or another approach?

[woodruffw]

Would you prefer de-duplicating it by moving it into a new common.py file, or another approach?

That's what I was thinking originally, although common.py won't be on sys.path so importing it won't be straightforward 🙁

I think the "clean" thing to do here would be to turn the Python files here into a project structure that gets installed as part of the Docker image's build. But that's a little heavyweight, so @webknjaz may have another idea 🙂

[webknjaz]

I'm leaning towards just being available on $PYTHONPATH, not pip installed. The way I did it @ https://github.com/re-actors/alls-green/blob/223e4bb/action.yml#L50-L53. However, such a restructuring should be separate and would probably include a src/pypi_publish/__main__.py entrypoint invoking the current main script through subprocess as the first step. I don't want a huge PR since big ones tend to cause long reviews and merge delays. Another thing to do would be to start to use our own managed virtualenv in the container rather than installing the deps globally into the shared userspace.

[webknjaz]

@woodruffw hey, it looks like GitHub rolled out their own attestations in beta: https://github.com/actions/attest-build-provenance / https://github.com/pypa/gh-action-pypi-publish/attestations / https://github.com/orgs/community/discussions/122028. I wonder if we could somehow integrate with that...

And it seems like there's even a new privilege attestations: write introduced for that on the platform: https://docs.github.com/en/actions/security-guides/using-artifact-attestations-to-establish-provenance-for-builds / https://docs.github.com/en/actions/using-workflows/workflow-syntax-for-github-actions#jobsjob_idpermissions.

[woodruffw]

(Sorry, was on PTO -- catching up on mentions now)

@woodruffw hey, it looks like GitHub rolled out their own attestations in beta: https://github.com/actions/attest-build-provenance / https://github.com/pypa/gh-action-pypi-publish/attestations / https://github.com/orgs/community/discussions/122028. I wonder if we could somehow integrate with that...

Yep, I've been thinking about how to make integrate the two -- the last comment on the PEP discussion thread suggests an approach that would allow GitHub-generated attestations to be compatible with this PEP.

TL;DR: I think our options here are:

1. Move forward with the current "version 1" specified in the PEP. This would mean no further changes are required on the PEP side. However, it would also mean that "version 1" isn't compatible with the default attestation format generated by GitHub, and that we would have to create a "version 2" that is compatible.
2. Edit the PEP in its current form and make it compatible with GitHub's attestation format. This would be a small change to the PEP overall, but has knock-on consequences (e.g. conceptually unbounded JSON payloads being signed over).

[woodruffw]

Just to update here: I've begun updating PEP 740 to be compatible with what GitHub's artifact attestations feature is doing, and it looks like it'll be pretty smooth.

That then leaves the question of how to approach attestation generation here 🙂:

1. Go forward with what's currently in this PR (a small script that uses the id-token permission + pypi_attestation_models to manually generate the attestation)
2. Move to a more discrete/black-boxed approach, e.g. something like this:

steps:
  - name: attest
    # does not exist yet!
    uses: pypa/gh-action-pypi-attest
    with:
      # this would be the default; just for illustration
      packages-dir: dist/

  - name: publish
    uses: pypa/gh-action-pypi-publish@release/v1

The benefit to approach (2) is that it's more devolved and can be built up on top of actions/attest@v1. OTOH it requires users to add another workflow step and get the order right, versus doing it implicitly in this action. Thoughts @webknjaz?

Edit: for context, the reason we can't use actions/attest-build-provenance or actions/attest directly is because (1) we have our own custom predicate, and (2) our attestation format requires a transformation step from the output bundle produced by those actions, so we need to wrap them in a "composite" or similar action to get everything into the shape PyPI expects.

[ofek]

Awesome stuff! I would much prefer option 1, we should prioritize UX.

[sethmlarson]

I agree with @ofek that ergonomics is the most important part here, requiring users add another action step means we will see delayed rollout of the feature compared to adding it on to the existing workflow.

[woodruffw]

SG, I'll keep going with the current approach then 🙂

[woodruffw]

This is good for an initial review! To summarize:

1. By default, nothing changes.
2. If the user sets attestations: true and is in a Trusted Publishing flow, this workflow will now transparently generate a "PyPI Publish" attestation for them, for each distribution being uploaded.
3. This attestation will be transparently stabled to each distribution, thanks to twine --attestations.

[sethmlarson]

@woodruffw Just double checking long-term strategy, is the attestations: true setting only there because this is an experimental feature and the goal is that one day a new version of the workflow will automatically do this without the attestations: true setting?

[woodruffw]

Just double checking long-term strategy, is the attestations: true setting only there because this is an experimental feature and the goal is that one day a new version of the workflow will automatically do this without the attestations: true setting?

Yes, exactly -- my thinking is that it's already non-default due to the tail of people using older versions of this action, so having it be (temporarily) opt-in makes experimenting easier while giving us the ability to flip the switch once everything is stable.

[webknjaz]

Sorry I didn't react earlier. I also think that UX is important. Not sure how well it plays with uploading attestations to GitHub, though.

Can we attempt relying on the attestations: write privilege instead of the action input? Can we feature-detect that it's enabled and use that as a flag? It feels like better ergonomics + potentially better integration with GH IIUC.

Another thought: why do we need to make attestations opt-in even? If it's supported in the Twine version and PyPI side is a no-op (still?), what's the harm in doing it unconditionally whenever we have OIDC privilege? If you're worried about crashes in our own code, we could just stick overly generic exception handling for the experimentation period.

[woodruffw]

Gentle ping for review 🙂 -- we're getting close to enabling attestation persistence on PyPI, so being able to roll this out (with the current toggle) will help us smoke test the functionality there.

[webknjaz]

I was hoping the refactoring PR would be completed by now, but it looks like that ain't happening... I suppose, this is a good patch to accept. Could you address a few comments below before I merge?

[di]

This LGTM, one thoughts arises about future work: We should eventually support attestations: false if/when this becomes the defualt so users can easily turn this off if they need to.

[di]

Is this necessary given pypa/twine#1099?

[woodruffw]

It's not necessary, since twine will catch this case and produce an appropriate error. OTOH catching it here might make the error slightly more intelligible, since the twine error doesn't get propagated into a ::warning annotation like this one does.

This is similar to how the action pre-validates the dist directory, even though twine also validates it. But if you think it makes it makes things more confusing than necessary, I can drop this check 🙂

[di]

I think it's probably fine for this to catch it before twine and give a more tailored error message, I'm mostly just thinking about the long-term pain of maintaining/updating this list of supported indexes in multiple places, but this is fine for now.

[webknjaz]

Thanks! Love the attention to detail in your PRs, as always :)

[webknjaz]

v1.10.0 is now live. Appologies that it took so long to merge this.

[woodruffw]

No problem whatsoever! Thank you for your diligent reviews and for seeing this through!

[webknjaz]

UPD: Upon seeing Dependabot sending bumps of this version to repositories, I improved the release notes to include a call to try out attestations with that code example so it's discoverable more easily.

[webknjaz]

@woodruffw looks like the sanity check is broken — it exploded in attrs..

[webknjaz]

UPD: v1.10.1 fixed that bug via 0ab0b79 and Hugo reported that it works.

[webknjaz]

This is where the bug snuck in that needed fixing by 0ab0b79.

[woodruffw]

🤦 thanks for fixing that.

[webknjaz]

@woodruffw re: GH Attestations API — it seems pretty straightforward and doesn't really need going the composite action route, it seems. See https://github.com/actions/toolkit/blob/77f247b/packages/attest/src/store.ts#L5-L44.

[webknjaz]

xref #288