[BUG] test failure: test_unicode.UnicodeTest.test_raiseMemError on 32bit in Debian CI

[stefanor]

Describe the bug

test_unicode.UnicodeTest.test_raiseMemError is failing on i386, armhf and armel (the 32 bit ARM ports) in Debian's CI.

Logs: https://ci.debian.net/packages/c/cython/

The cause isn't obvious so we'll skip the test, but I'm filing this so that someone can investigate more deeply.

Code to reproduce the behaviour:

No response

Expected behaviour

No response

OS

Linux (Debian)

Python version

3.12.1

Cython version

3.0.6

Additional context

======================================================================
FAIL: test_raiseMemError (test_unicode.UnicodeTest.test_raiseMemError)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "tests/run/test_unicode.pyx", line 2355, in test_unicode.UnicodeTest.test_raiseMemError (test_unicode.c:144223)
    self.assertRaises(MemoryError, alloc)
AssertionError: MemoryError not raised by lambda

======================================================================
FAIL: test_raiseMemError (test_unicode.UnicodeTest.test_raiseMemError)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "tests/run/test_unicode.pyx", line 2355, in test_unicode.UnicodeTest.test_raiseMemError (test_unicode.cpp:144235)
    self.assertRaises(MemoryError, alloc)
AssertionError: MemoryError not raised by lambda

[stefanor]

I'm guessing is is because we're running in a 32bit container on a 64bit kernel, so processes can allocate 4GiB

This is non-trivial to detect from userspace, I can't think of a reliable way to do so offhand.

[da-woods]

Is Python itself passing it's test-suite?

This is a test that's ultimately copied from CPython (https://github.com/python/cpython/blob/f108468970bf4e70910862476900f924fb701399/Lib/test/test_str.py#L2451) so it'd be interesting to know if that version of the same test is passing/failing/being skipped.

[stefanor]

Yep, seems to be passing: https://ci.debian.net/packages/p/python3.12/unstable/i386/40342642/#L2130 (until 3.13, it's called test_unicode)

I will say that I couldn't reproduce this one locally (an i386 container on an amd64 host), but it seems to be failing reliably on our CI infra.

[da-woods]

Thanks for confirming it.

I'm not too worried about this test since it's mostly testing Cpython internals rather than anything we're doing so I think it's fine to ignore.

This kind of bug (where it's very very platform specific) always ends up fairly hard for us to reproduce unfortunately.

[paulgevers]

One of the maintainers of ci.debian.net here. If there is a particular command you'd like me to run in the testbed where the test failed, let me know and I'll see what I can do.

[da-woods]

My view is we should just skip this test.

The original test I think was designed to allocate a string that's just too big and is thus guaranteed to fail. However, it's based on struct sizes that are version dependent and have shrunk a little since we copied the test. So it's possible that the test is now testing something that's probably too big but might just is possible.

It's testing a CPython implementation detail that's tested properly as part of CPython. I don't think there's much value to Cython in having this test and I don't think it's worth you spend time debugging it.

I'll make a PR to remove it shortly