Move pytest entrypoints to own module

[sloretz]

Alternative to #277
Attempts to fix ros2/rclpy#831

This moves the pytest entrypoints to a separate module, and makes them import anything ROS specific as late as possible. This should prevent packages like ament_package importing downstream packages like rclpy.

[sloretz]

This branch is on the commit just before #276, so windows debug CI should show if the issue is fixed or not.

Windows debug: build / test all

Edit: But I forgot CI merges with master - trying again.

[sloretz]

The second commit also reverts #276 for testing the approach purposes

Windows debug CI all:

[jacobperron]

Besides the test failure, the change LGTM.

Do you know if isolating pytest hooks like this is a best practice documented anywhere?
Also, should we consider doing something similar for launch_testing?

[sloretz]

Do you know if isolating pytest hooks like this is a best practice documented anywhere?

I've never heard of it before. IIUC this PR is solving something specific to merged workspaces and how we run CI, and not implementing a general Python workflow. In our CI all the packages are built before any tests are run, which installs them into the install space. If that install space is merged then when tests are run all packages have access to all packages that have been built. Even the lowest level packages (like ament_package) get pytest hooks in higher level ones.

Also, should we consider doing something similar for launch_testing?

Maybe, but I haven't figured out why the _rclpy_pybind11 import failed (it had access to the install space after all) so I don't know if the same issue is likely to apply to that package given its dependencies.

[ivanpauno]

LGTM!

The only weird thing is that the file called hooks.py does not contain the definition of any pytest hook now.
(IIRC pytest_launch_collect_makemodule and pytest_configure are called "hooks" in pytest parlance)

[sloretz]

Rebased to include just the fix, and not a revert revert of #274

[sloretz]

Full CI

• Linux
• Linux-aarch64
• macOS
• Windows

[jacobperron]

I'm not sure why we are seeing those qt_dotgraph test failures. That package does not transitively depend on launch_ros...
I was not able to reproduce the failures locally and haven't seen them appear in any other job. Re-triggering to see if it's still an issue:

• Linux:
• Linux-aarch64: (edit: rclcpp.test_time is an existing failure).
• macOS
• Windows

[ivanpauno]

macOS build failure seems unrelated, running it again:

• macOS