Replace CircleCI with GitHub Action

[janosh]

Continuing first attempt that was aborted in #730.

[janosh]

After a lot of fiddling, openbabel finally successfully installs (sudo ln -s /usr/include/openbabel3 /usr/local/include/openbabel3 as suggested by @fredrikw did the trick) and pytest can actually run. For some reason, GH runs way longer than CircleCI (47 min vs 6 min). Not sure why. Doesn't look like CircleCI does any under the hood caching.

GH action:

= 2 failed, 116 passed, 82 skipped, 224 warnings, 4 errors in 2634.36s (0:43:54) =

CircleCI:

========== 124 passed, 80 skipped, 650 warnings in 202.02s (0:03:22) ===========

In any case, we might be able to bring this down a lot by using something pytest-split.

[janosh]

We need to start pinning pymongo<4.0.0 since on CircleCI, deps were installed with conda whereas on with GH action we're using pip install. On Anaconda, latest pymongo is v3.12.0 whereas on PyPI, it's already v4.0.1 which is why compat issues did not show up before.

atomate is not upwards compatible due to e.g.

/home/runner/work/atomate/atomate/atomate/utils/testing.py:88: in tearDown
    for coll in db.collection_names():
TypeError: 'Collection' object is not callable. If you meant to call the 'collection_names' method on a 'Database' object it is failing because no such method exists.

[janosh]

Looks like there were a few tests in

• atomate/vasp/analysis/linear_response.py
• atomate/vasp/workflows/tests/test_insertion_workflow.py

that would have failed even after the "Fix CI" PR #730 if they hadn't been skipped by pytest. Pinning pymongo<4.0.0 or something else seems to have unskipped them and now they fail. Will fix as part of this PR.

[janosh]

All green! 🎉 🟢

Just a final note that my comment above about GHA being slow was premature.

The much longer run time on GH action was due to me not spinning up a MongoDB server there yet. So every DB connection was timing out which of course took a long time. With that fixed, GH action is still a small amount slower (10m 24s circle vs 12m 25s GHA).

I asked @asottile about that who said

I think their machines might be slightly beefier -- also circle tends to start sooner than GHA -- circle was basically the only one on par with http://pre-commit.ci in start time (not in execution time though1)

So this slight slowdown seems to be expected. Hence, ready for review here.

[janosh]

One more point: I think we should keep the CircleCI workflow running for a while to make sure there are no further discrepancies between CI runs. Maybe after we transition to latest pymatgen with only py38+ support, we can drop CircleCI unless @computron or someone wants to keep it permanently.

[computron]

Merging this prior to new atomate release