Bug report
test.support.warnings_helper.check_warnings() raises KeyError: 'warnings'
under -X lazy_imports=all, taking down every test that uses it.
Reproducer:
./python -X lazy_imports=all -m test test_sundry test_time
Lib/test/support/warnings_helper.py:176:
# Because test_warnings swap the module, we need to look up in the
# sys.modules dictionary.
wmod = sys.modules['warnings']
warnings is not in sys.modules at interpreter startup, with or without lazy
imports, so that key is populated solely by the module-level import warnings
at the top of this same file. Reading the module out of sys.modules by
subscript means the global name is never touched, so under lazy imports the
import is never resolved and the key is never created.
41 files under Lib/test/ use check_warnings() / check_no_warnings().
One-line fix — the default argument is evaluated on every call, which resolves
the lazy import before the lookup, and the module swap done by test_warnings
still wins:
wmod = sys.modules.get('warnings', warnings)
Verified on main (d915492): with that change test_sundry, test_time
and test_warnings pass under -X lazy_imports=all; test___all__,
test_sundry, test_time, test_warnings and test_support still pass
without the flag.
test_sundry and test_time are not in the exclusion list proposed in
GH-151105, so this is a failure mode the planned CI would not have covered.
The same pattern appears in the tests themselves. After the helper is fixed,
test___all__ still fails at Lib/test/test___all__.py:49:
if not hasattr(sys.modules[modname], "__all__"):
with KeyError: '_osx_support' — the module is imported by
exec("import %s" % modname) a few lines above, which under lazy imports
leaves sys.modules untouched. Lines 71 and 76 do the same thing.
grep -rn 'sys\.modules\[' Lib/test/ reports 126 subscript reads. Most are
legitimate identity checks after an eager import; the ones worth auditing are
those where the key is expected to be filled by an import in the same file, or
by an execed import statement.
Related to #149640.
CPython versions tested on:
CPython main branch (d915492)
Operating systems tested on:
Linux
🤖 Investigated and drafted with Claude Code,
model Claude Opus 5 (1M context).
Linked PRs
Bug report
test.support.warnings_helper.check_warnings()raisesKeyError: 'warnings'under
-X lazy_imports=all, taking down every test that uses it.Reproducer:
Lib/test/support/warnings_helper.py:176:warningsis not insys.modulesat interpreter startup, with or without lazyimports, so that key is populated solely by the module-level
import warningsat the top of this same file. Reading the module out of
sys.modulesbysubscript means the global name is never touched, so under lazy imports the
import is never resolved and the key is never created.
41 files under
Lib/test/usecheck_warnings()/check_no_warnings().One-line fix — the default argument is evaluated on every call, which resolves
the lazy import before the lookup, and the module swap done by
test_warningsstill wins:
Verified on
main(d915492): with that changetest_sundry,test_timeand
test_warningspass under-X lazy_imports=all;test___all__,test_sundry,test_time,test_warningsandtest_supportstill passwithout the flag.
test_sundryandtest_timeare not in the exclusion list proposed inGH-151105, so this is a failure mode the planned CI would not have covered.
The same pattern appears in the tests themselves. After the helper is fixed,
test___all__still fails atLib/test/test___all__.py:49:with
KeyError: '_osx_support'— the module is imported byexec("import %s" % modname)a few lines above, which under lazy importsleaves
sys.modulesuntouched. Lines 71 and 76 do the same thing.grep -rn 'sys\.modules\[' Lib/test/reports 126 subscript reads. Most arelegitimate identity checks after an eager import; the ones worth auditing are
those where the key is expected to be filled by an import in the same file, or
by an
execed import statement.Related to #149640.
CPython versions tested on:
CPython main branch (d915492)
Operating systems tested on:
Linux
🤖 Investigated and drafted with Claude Code,
model Claude Opus 5 (1M context).
Linked PRs