Looks like the exception type is wrong in the test, could you have another look?

On Mac with Python 3.9.17 I can't reproduce it

On a Linux docker container:
(all packages in requirements-dev.txt have been reinstalled prior to running this)

root@f6dc42ebe522:/app/duckdb# python3 -m pytest tools/pythonpkg/tests/fast/test_import_without_pyarrow_dataset.py 
============================= test session starts ==============================
platform linux -- Python 3.9.12, pytest-7.4.2, pluggy-1.3.0
rootdir: /app/duckdb/tools/pythonpkg/tests
configfile: pytest.ini
collected 1 item                                                               

tools/pythonpkg/tests/fast/test_import_without_pyarrow_dataset.py .      [100%]

============================== 1 passed in 0.73s ===============================
root@f6dc42ebe522:/app/duckdb# uname -a 
Linux f6dc42ebe522 5.15.49-linuxkit #1 SMP PREEMPT Tue Sep 13 07:51:32 UTC 2022 aarch64 aarch64 aarch64 GNU/Linux

Think I found the culprit, the test that's failing is actually underspecified, which I now fixed.
It uses arrow_dataset().IsLoaded() which will return true if it was loaded correctly by us before.

The test however monkeypatches the pyarrow dataset to not be present, so in the sys modules it's no longer present.
We don't check for this so the VerifyDatasetLoaded method succeeds, while later on where the attributes are imported this will fail.

I don't see why that's a huge deal though, in both cases it throws an exception, the one we throw is just a little more descriptive, none the less I changed and hopefully fixed it now :)

This doesn't fix it either, unless we actually make an environment where pyarrow.dataset is not available we can't really test this, because monkeypatching just makes it as if it wasn't imported

If we import pyarrow.dataset from within duckdb, I believe that just undoes all of the monkeypatching
If I add a test above this one which doesn't monkeypatch, then the second test will fail because it was already imported and probably cached.

I am tempted to remove the failing test
It's testing a scenario that we can not reliably, deterministically replicate

If the test is run standalone it passes, when I run it together with other tests, it doesn't raise an exception (because pyarrow is loaded before it was monkeypatched out)
And I'm guessing the order tests are run is platform dependent, which makes the exception we get on Linux different because of the sequence of tests that are run before the offending test.

Thanks!