Re-run CI

[mattwthompson]

CI is failing on feature branches for reasons that are not clear to me. I'm re-running CI here to see if the failures are genuine and on the main branch.

[j-wags]

Oh, wow, thanks for catching this.

It looks like 4 errors have started to happen:

• The importlib thing which you caught here (and are fixing other places too)
• A bunch errors around the pint 0.20 release which got resolved 15 days ago by the openff-units release
• A test which has been flaky for a while openff/bespokefit/tests/executor/services/qcgenerator/test_worker.py::test_compute_torsion_drive where an optimization doesn't converge and raises geometric.errors.GeomOptNotConvergedError: Optimizer.optimizeGeometry() failed to converge.
• Whatever this new thing is, in test_get_optimization which manifests as HTTPStatusError. I agree with you that it only seems to manifests on PRs from branches in the last few days.

So I'll keep looking into this last one, and I may work on that flaky convergence one as well.

[j-wags]

Fastapi is a possible culprit here, though they have a really fast release pace so maybe it's nothing. https://anaconda.org/conda-forge/fastapi/files

[mattwthompson]

Here's the most recent nightly CI run that passed: https://github.com/openforcefield/openff-bespokefit/actions/runs/3231103232/jobs/5290258911

[j-wags]

Looks like that was it. Possible even rootier root issue is the fastapi conda package changing from requiring starlette 0.20 to 0.21. Rabbit hole begins here: tiangolo/fastapi#5646

[mattwthompson]

That's exactly it; the currently-failing tests (👁️=false) are on Starlette 0.21 and the passing tests (👁️=true). I think that dependency update was intentional: https://github.com/tiangolo/fastapi/blob/4638b2c64e259b90bef6a44748e00e405825a111/docs/en/docs/release-notes.md#0870

FastAPI is known for being relatively unstable as a framework, which may or may not be fair since Django and Flask are a decade or so older and more hardened.

Let's see a different pin helps ...

[mattwthompson]

^ above failures vs. passing tests is because I forgot to update both environments, not necessarily because of some mystical relationship between OpenEye and Starlette

[mattwthompson]

If this pin works, my current plan is:

• Slim down this PR to relevant changes, merge
• Update Use micromamba in CI #202, merge
• Merge changes into Support OpenFF Toolkit v0.11+ #198

I can also push the pin into the conda builds if you think that's worthwhile

[codecov]

Codecov Report

Merging #203 (19c090a) into main (5663b52) will not change coverage.
The diff coverage is n/a.

Additional details and impacted files

[j-wags]

That plan sounds good to me, and I agree we should pin the conda builds as well!

I'd like to gain a better understanding of where the rift in our fitting stack is going to open up - There's some range of packages that work with OFFTK 0.10, and another that works with OFFTK 0.11.

Also I'm a little bothered by the sporadic GeomeTRIC convergence errors, but I'll open a separate issue for that.

[mattwthompson]

Merging these fixes as part of #202 since these environments are pretty slow to solve (one took 50 minutes)