Misleading test coverage report

[fjetter]

As a dask maintainer, I want to trust the code coverage report.

Our coverage badge is a bit misleading showing coverage below 90%. This is due to us not collecting coverage in a few places. Also, we simply have a few modules which are only there for debugging and/or historical reasons

The most relevant parts (scheduler, worker, etc.) do have quite good coverage. I believe the <90% batch doesn't reflect well on the project and the wrong configuration creates a lot of noise making it harder to see relevant gaps (there are quite a few)

Things I spot immediately wehich we should exclude/ignore (non exhaustive)

• Exclude nvml calls
• Exclude UCX
• Exclude LOG_PDB brances
• Ignore pytest_resourceleaks
• Ignore _version.py (vendored)

Pragma no cover

• We have cases with relatively long switch-like statements ending in a final exception, e.g. ValueError "unknown value"
• Conditional imports may need be covered with pragma statements and/or a dedicated job

Relevant documentation https://coverage.readthedocs.io/en/coverage-4.3.3/excluding.html I would hope us being able to do this without code modification, if possible. A proper coverage.rc should be sufficient

There are other modules which we are not entirely sure. If it is not straight forward, this can be discussed in the PR review, e.g.

• distributed/deploy/old_ssh.py
• distributed/cli/dask_ssh.py
• distributed/nanny.py

Long term goal:

• We should have 100% test coverage
• If there are weird edge cases where an extremely complicated test would be required to test an unimportant test case. For these cases we want to use deliberate nopragma instructions

AC

• All very obvious cases should have a pragma statement and/or exclude configuration option
• Debatable cases should still be left uncovered and be handled in a follow up ticket
• Developer documentation is updated with a guideline

[fjetter]

The most important files are

• node.py
• core.py
• scheduler.py
• worker.py
• client.py
• stealing.py