The functools module imports reduce from _functools, using a guard in case it is not present:

try:
from _functools import reduce
except ImportError:
pass

However, the documentation says nothing about reduce being optional, and it is unconditionally included in the module __all__.

If reduce is guaranteed to be implemented in _functools, then the guard is redundant and should be removed. Otherwise, a pure python fallback should be added.

(The docs for reduce include a pure Python equivalent which might be sufficient.)

See discussion in https://bugs.python.org/issue12428 (especially the last message)

Reviewing the code and the CI test failures on the PR, the trick here is that functools isn't actually *using* functools.reduce, it's just re-exporting it if it's defined.

So if you block importing of "_functools" (which the test suite does in order to test the pure Python fallbacks), then the *only* consequence is that "functools.reduce" will be missing - the module will otherwise be fine.

This isn't at all clear when reading the code though, so I think the simplest resolution here would be to add a comment to the fallback path that says "If _functools.reduce is missing, then functools.reduce will also be missing, but the module will otherwise work".

Alternatively, we could add a fallback implementation based on the recipe in the docs, and adjust the test suite to actually run the reduce tests against the py_functools variant: https://docs.python.org/3/library/functools.html#functools.reduce

New changeset e25d5fc by Victor Stinner (madman-bob) in branch 'master':
bpo-32321: Add pure Python fallback for functools.reduce (GH-8548)
e25d5fc

I just merged PR 8548. See the PR to the discussion. IMHO it's a nice enhancement to have a pure Python implementation:
#8548 (comment)
(and the PR has been approved by 2 other core devs ;-))

But I'm against adding it to Python 3.7 as well, I rejected the backport: PR 10092.

I proposed the author to write another PR to refactor test_functools.py, but I don't think that it's worth it to keep this issue open just for that:
#8548 (comment)

However, the documentation says nothing about reduce being optional, and it is unconditionally included in the module __all__.

Oh, about this specific issue: maybe test___all__ should be fixed to test functools.py with _functools blocked? As done by test_functools.py? But this is a wider change and I'm not sure that it's doable :-( (especially in a generic way!) The PEP-399 requires the same API for the C and the Python implementations.