Drop setting dl open flags in dolfinx/__init__.py

[francesco-ballarin]

Opening this to assess question #2835 on CI

Closes #2835

[francesco-ballarin]

All 20 CI runs were successful. Is there any additional CI workflow that would be useful to test, but should be triggered manually?

[garth-wells]

Could you test without 'no-binary' in the pip install of NumPy, see

dolfinx/docker/Dockerfile.test-env

Line 151 in cc47953

# Install numpy via pip. Exclude binaries to avoid conflicts with libblas

,

[francesco-ballarin]

OK, so there are four scenarios:

1. fiddling with dl open flags in dolfinx/__init__.py, and installing numpy with no-binary: this is current main branch, and therefore there is not need to test it here;
2. not fiddling with dl open flags in dolfinx/__init__.py, and installing numpy with no-binary: tested in 9f5804d;
3. fiddling with dl open flags in dolfinx/__init__.py, and installing numpy without no-binary: tested in 92a5640;
4. not fiddling with dl open flags in dolfinx/__init__.py, and installing numpy without no-binary: tested in 8f92275

All four scenarios are successful on CI, so I would conclude that fiddling with dl open flags in dolfinx/__init__.py is not necessary, and that installing numpy with no-binary is not necessary either.

Todo list after this PR is approved, and before merge:

• rebase so that 9f5804d and 57b03f7 are kept, and 8673643, 92a5640 and 8f92275 are dropped
• rerun again https://github.com/FEniCS/dolfinx/actions/workflows/docker-redhat-test-env.yml and https://github.com/FEniCS/dolfinx/actions/workflows/docker-dev-test-env.yml from this branch, but with the default docker tag (rather than the temporary one pr-2836)
• ask @jhale to clean up the temporary docker tags pr-2836 from the registries (or grant me permissions to do so)
• ask @drew-parsons and @minrk if they want to test this on ubuntu and conda, respectively?

[jhale]

This fix/hack has been in a very long time! I'm relatively persuaded by the above experiments: it's (probably) no longer an issue.

If we get reports of issues we can quickly back it out.

[jhale]

Don't worry about conda and debian, my memory is that this is related to numpy binary wheels and openmpi

[drew-parsons]

Agreed about debian. If it's passing github CI then it's safe for the purposes of this PR to assume that tests will pass in Debian builds.

[francesco-ballarin]

I've carried out 3 of the 4 tasks in my todo list (the remaining one is to clean up the docker repositories together with @jhale). This is now ready to merge.