Avoid the limitation of not being able to import rclpy C extensions at module level

[ivanpauno]

Kind of reverting the old #147, plus solving the "not nice grouping" problem in another way.

See discussion: #413 (comment)

---

There's some new nightly failures that are somewhat related with this problem, eg:
https://ci.ros2.org/view/nightly/job/nightly_win_rel/1302/testReport/junit/(root)/projectroot/test_client/

Basically, a pytest plugin is trying to import the C extensions before we "hack" the import in __init__.py file of the test directory.
This PR doesn't solve that problem.
I'm not to sure why this is only hitting as on Windows.
There are some possible solutions:

• avoiding importing rclpy at module level in the plugin
• Figuring out why we're loading that pytest plugin on Windows and not in the other os (rclpy doesn't depend on the package that provides the plugin).
• Change how we install python packages, to avoid doing weird imports while testing.

[dirk-thomas]

There's some new nightly failures that are somewhat related with this problem, eg:
https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/test_client/

The referenced test passed?

[ivanpauno]

• Linux
• Linux-aarch64
• macOS
• Windows

[ivanpauno]

There's some new nightly failures that are somewhat related with this problem, eg:
https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/test_client/

The referenced test passed?

Sorry, wrong link. Updated post.

[dirk-thomas]

https://ci.ros2.org/view/nightly/job/nightly_win_rel/1302/testReport/junit/(root)/projectroot/test_client/

Still referencing a passed test.

[hidmic]

The original link shows the regression https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/test_client/, but it can also be seen everywhere in CI. There's ros2/build_farmer#235 tracking it.

[ivanpauno]

There's something weird in the links, when I enter the first time, they show the regression and the error message.
When I enter again a time later, they show Passed.
If I enter to the job status, navigate through the failing tests and enter again, I see the correct regression message ... and the same link as before.

It happened again with @hidmic link ....

[ivanpauno]

The grouping of the tests is looking well after this patch, see:
https://ci.ros2.org/job/ci_linux/8072/testReport/

[hidmic]

@ivanpauno using that envvar is so shady... there has to be a natural way to do this. I wonder what setuptools does under the hood when installing a regular Python package.

[dirk-thomas]

There's something weird in the links

If you look at the list of tests (https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/) you will find that there are three test results named test_client.

[ivanpauno]

If you look at the list of tests (https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/) you will find that there are three test results named test_client.

That's really inconvenient. This other doesn't have a repeated name:
https://ci.ros2.org/view/nightly/job/nightly_win_deb/1366/testReport/junit/(root)/projectroot/test_parameters_callback/

@ivanpauno using that envvar is so shady... there has to be a natural way to do this. I wonder what setuptools does under the hood when installing a regular Python package.

The current solution is also shady ...

[dirk-thomas]

That's really inconvenient.

More like not smart naming of our tests. All three tests have the same name:

• rclcpp test client
• rclcpp_action test client
• rclpy test client

Giving them more precise and therefore unique names would resolve this.

[ivanpauno]

Checking than this still have failures on windows when launch_testing_ros is present:

• Windows

[ivanpauno]

Closing, because this is outdated.
We should still investigate ways of solving this issue.