Use pytest.fixtures for starting servers, use importlib*metadata

[bollwyvl]

References

• fixes Make launching lab/notebook server under test more automated #309
• fixes DistributionNotFound(Requirement.parse('webcolors>=1.11; extra == "format-nongpl"'), #378

Code Changes

• add conftest.py
  • add a number of @fixtures to support starting a server
• remove use of hard-coded PORT and TOKEN
  • ensure coverage is captured
• remove pytest.ini
• move pytest config to pyproject.toml
• squash some test warnings
  • use importlib(.|_)metadata instead of pkg_resources (undeclared setuptools dep)
• remove bit about starting a server before testing in CONTRIBUTING.md

Future Work

• further isolate the test environment with e.g. custom $HOME (tricky on windows)
• split up existing test_proxies.py into smaller files
• potentially move some more repetitive things to fixtures
• add parameterization of --tornado_settings base_url=/foo/bar/ to test nested paths
• fix more test warnings
  • hoist new test warnings to errors
• improve speed
  • starting a sever takes about 1.5s
  • can be challenging to get the scope correct with e.g. tmp_path
  • the widely-used pytest-xdist can run tests in parallel for a moderate speedup, but carries some additional costs
  • some ports, etc. in the jupyter_server_config.py, and tests, still hard-coded, and would collide

[Finesim97]

This will close #378.

[consideRatio]

Looks great, thank you @bollwyvl!!!

[consideRatio]

Is this to be listed under the changelog as a bugfix and/or maintenance? Typically I've avoided having it listed twice, but I figure it should be seen under bugfix since its fixes a reported bug etc, but the title is all about maintenance.

Let's make it list under both this time.

[Finesim97]

Thank you both!

[bollwyvl]

Cheers. Love entry_points, one of the most powerful things built into the python packaging system, but even this transition has been far from smooth, as the backport and stdlib APIs still don't work the same way... Until 3.10, I guess. a little aside: we actually found an (already-fixed) bug in the stdlib/backport behavior when looking into entry_points for replacing data_files in jupyter_core, hence the version detection. Older stdlib loads (and does not cache) _all_ eps, which goes straight to performance hell once you have hundreds of them. It may even be worth considering loading them lazily, if possible.