Detailed 4.0 project planning for release

[mrshirts]

Moving from the general task #255 to specific tasks

• [x] Decide what to do about deprecation for old names and results structures. Do we just support the new ones? Do we add specific errors for the old names? Do we do a 3.0 release that deprecates names? (Discussion question!)
• [x] Add in PR #444 (approval from others?) (Functionality)
• [x] Address pytest issues with flaky? (see conversation below)
• [x] Linting issues leading to test fails (Levi?)
• [x] Move #442 into 4.0 (MRS, approval from others)
• [x] Add JAX support, including if it is not installed (#402 and related) (MRS) (Functionality) now in process at #447
• [x] add conda jax support (Levi?) (in #447)
• [x] Adding bootstrap uncertainty support (like #302, implemented in #448) (MRS)(Functionality)
• [x] Address #338 (MRS) (done in #457)
• [ ] Address #366 (MRS)
• [ ] Address #379 (MRS)
• [ ] Address #382 (MRS)
• [x] Address #391 (MRS)
• [x] Coordinate with alchemlyb (in process at https://github.com/alchemistry/alchemlyb/issues/204).
• [x] Update documentation (https://pymbar.readthedocs.io/en/master/), MRS: add in all the text, somebody else: explain to MRS what to do to retrigger/rebuild - starting to address this in #455)

[Lnaden]

It looks like Conda JAX is usable as is with some small tweaking: https://github.com/google/jax/issues/189. There is some odd nuance with a couple of caveates:

• Add a jaxlib==0.3.10=*cuda* to the install (we can do that with the meta file)
• you need a CONDA_OVERRIDE_CUDA="11.2" as part of the invocation of conda install if you dont have a GPU.
• We may want to look at the OpenMM feedstock for the multiple conditions of gpus/cpus conda/pip

[Lnaden]

@mrshirts can you make a nudge or re-trigger #402 so we can see what the errors are exactly?

[Lnaden]

Address pytest issues with flaky? (Levi?)

We do have the flaky plugin setup to be used in one test already. I'm going to keep triggering tests locally and on my fork to see if I can figure out which tests are problematic.

[mrshirts]

For #402, that will take a couple of days, I need to revisit what is actually the right way to do this.

[mrshirts]

The specific test failure I got was:

test_mbar_solvers.py:20: PytestUnknownMarkWarning: Unknown pytest.mark.flaky - is this a typo? You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html @pytest.mark.flaky(max_runs=2) # Uses flaky plugin for pytest

But that was running pytest v 7.1.1 locally, so I guess the issue was that I was running w/o the plugin?

[Lnaden]

The specific test failure I got was:

That's the generic error message from pytest if you try to use a mark it doesn't recognize (or don't have the plugin to recognize). The test will run fine without the flaky plugin, so its a relatively harmless warning. I have that one test marked flaky because it can stochastically fail rarely. Post 4.0, it may not even be a valid test anymore since we are changing the way solvers work. For now, the GitHub Actions CI installs the plugin so it should not throw the warning in online testing.

[mattwthompson]

I've had pretty good success pulling down JAX from conda-forge, with the caveat that I haven't needed CUDA acceleration and for a brief period of time all of the builds were broken on macOS because the feedstock was more or less only maintained by automation. But it's generally available and not been a pain to install lately.

[mattwthompson]

And a totally trivial suggestion, but you might want to remove devtools/conda-recipe now that the conda builds happen in the feedstock.