When a module import fails, other threads can receive a stale module reference that is no longer in sys.modules.

Thread 1 starts importing module X, adds it to sys.modules
Thread 2 checks sys.modules, finds module X, takes the "skip lock" fast path
Thread 1's import fails, removes module X from sys.modules
Thread 2 returns the stale module reference (not in sys.modules)

Affected code paths

Lib/importlib/_bootstrap.py:_find_and_load() lines 1268-1282
Python/import.c:PyImport_GetModule()
Python/import.c:PyImport_ImportModuleLevelObject()

All three have an optimization that skips locking when a module appears already loaded, but none verify the module is still valid after checking _initializing.

Reproducer

import sys
import threading
import tempfile
import os
with tempfile.TemporaryDirectory() as tmpdir:
    sys.path.insert(0, tmpdir)
    # Module that partially initializes then fails
    with open(os.path.join(tmpdir, "failing_mod.py"), "w") as f:
        f.write("import time; time.sleep(0.05); raise RuntimeError('fail')")
    results = []
    def do_import(delay):
        import time; time.sleep(delay)
        try:
            mod = __import__("failing_mod")
            if "failing_mod" not in sys.modules:
                results.append("RACE: got module not in sys.modules")
        except RuntimeError:
            results.append("ok")
    t1 = threading.Thread(target=do_import, args=(0,))
    t2 = threading.Thread(target=do_import, args=(0.01,))
    t1.start(); t2.start(); t1.join(); t2.join()
    print(results)  # May show "RACE" on affected builds

Affected code paths

Reproducer

Linked PRs