Running Python test under pyinstrument or other custom executable

[eguiraud]

Hello,

I have some Python tests (typically run with pants test ...) that I sometimes would like to run under a profiler such as pyinstrument, or cProfile, and some other times even under a C/C++ debugger like gdb in order to break into problematic logic in a given extension module.

What is the simplest way to do so?

What I am doing at the moment:

• change BUILD.pants file to add a pex_binary rule for the python_test rule that I want to instrument/debug
• change the test to make it run the test what it's run as a standalone Python executable by adding an invocation to the test function under if __name__ == "__main__"
• run pants package for the pex_binary target
• gdb --args python ./dist/path/to/target.pex now seems to work

[huonw]

I think there's potentially two other ways to do this:

1. allow running the code outside pants, along the lines of https://www.pantsbuild.org/docs/setting-up-an-ide: export a venv for third-party code, set up PYTHONPATH for first-party code, and then run "as normal"
2. jump into the sandbox pants is running, e.g. use pants --keep-sandboxes=always test path/to:target, then go to the path that's printed and use __run.sh as the starting point for what you want to do

1 is more structured/reusable, whereas 2 may be simpler/faster for one-off things.

[eguiraud]

Thanks for the suggestions!

set up PYTHONPATH for first-party code

I'm having some trouble there, feels like pants should take care of doing (the moral equivalent of) pip install -e .?

After export PYTHONPATH=$PWD (where $PWD there is the repository root), running pytest some/path/to/a/test.py in the exported environment still fails to find the project's modules.

[huonw]

What sort of artifacts are you building? Of interest is whether you have any libraries (python_distribution) or only "binaries" like pex_binary and docker images.

In any case, PYTHONPATH needs to point to the directories that contain the importable modules, e.g. if you have src/foo and src/bar that are imported like import foo, import bar, PYTHONPATH=src. Pants calls these "source roots": https://www.pantsbuild.org/2.18/docs/using-pants/key-concepts/source-roots and those IDE docs suggest some commands to help set them up.

So, PYTHONPATH=$PWD makes sense if your modules are in directories immediately under the repo root. Is that the case?

[eguiraud]

yes that's the case, and pants root prints . so PYTHONPATH=$PWD should be equivalent to the command suggested in the link.

I'll debug 😬

[eguiraud-pf]

Found out why exporting PYTHONPATH doesn't work: the pytest executable in the env created by pants export explicitly discards PYTHONPATH!

$ bat -p /..snip../dist/export/python/virtualenvs/python-default/3.10.13/bin/pytest
#!/..snip../dist/export/python/virtualenvs/python-default/3.10.13/bin/python3.10 -sE
# -*- coding: utf-8 -*-
import re
import sys
from pytest import console_main
import os
print(os.getenv("PYTHONPATH"))
if __name__ == '__main__':
    sys.argv[0] = re.sub(r'(-script\.pyw|\.exe)?$', '', sys.argv[0])
    sys.exit(console_main())

As you can see (if you stare at it long enough) it's run under python3.10 -sE and -E discards PYTHONPATH and other PYTHON* env variables.

So setting PYTHONPATH to make first-party sources available doesn't work when running the pytest that comes with the environment generated by pants export.