Hi all — has anyone seen

./pants test

install packages that don’t match the lockfile? In my case, the test PEX ends up with two

attrs

versions (

24.2.0

and

25.3.0

). Python imports the older one first, which breaks

cattrs 25.x

with

ImportError: cannot import name 'NothingType' from 'attrs'

. When I run

./pants export ::

, the exported venv correctly has only

attrs==25.3.0

. I’ve cleaned caches, but the duplicate older

attrs

returns during

pants test

. Is this a known issue or a misconfig on my end?

do you happen to have both defined in

python_requirement

targets? maybe the export process can handle the de-duplicating, but the sandboxing for tests has a more difficulr time?

Not explicitly at least. I did not have attrs or cattrs explicitly in any of my requirement files. But is pulled in as a requirement. I have tried to pin these packages to 25.3 and 24.2, but it doesn't change anything 😕

I'll let an actual Pants maintainer wake up and explain this to you, and simply point here to prompt them: https://github.com/pantsbuild/pants/blob/24c4ecae0e84b795fb090c1219ba7dd4aeb0968d/3rdparty/python/pytest.lock#L65-L77 In short, this is a known shortcoming of how Pants Python handling does tools. The JVM side, IIUC handles this well by shading the tool and its internal deps as much as possible whereas the Python backend simply mushes 2 PYTHONPATHs together and prays for the best,

Thank you. Then I at least know why it happens. So as long I pin cattrs and attrs to the same version as used by pants, I should be fine. Does that sound right?