The script below creates a basic PEP 420 style package with a single module in it ('package.sub') and tries to import that module With 3.5 the script runs without problems and prints 42 (as expected). With a 3.5 (fresh checkout as of 2015-05-14) I get an SystemError: $ python3.5 demo.py Traceback (most recent call last): File "demo.py", line 10, in <module> import package.mod File "<frozen importlib._bootstrap>", line 958, in _find_and_load File "<frozen importlib._bootstrap>", line 947, in _find_and_load_unlocked File "<frozen importlib._bootstrap>", line 657, in _load_unlocked File "<frozen importlib._bootstrap>", line 575, in module_from_spec File "<frozen importlib._bootstrap>", line 519, in _init_module_attrs SystemError: Parent module '' not loaded, cannot perform relative import ####### import os os.mkdir('path1') os.mkdir('path1/package') with open('path1/package/mod.py', 'w') as fp: fp.write('print(42)\n') import site site.addsitedir('path1') import package.mod

I presume you meant that it works with 3.4?

The problem is right where the traceback says. Apparently there is a gap in the namespace package tests that I slipped through with my recent work to split out path-based import. I'll work up a patch.

Hmm, look like the test suite masks the issue due to the fact that importlib gets imported before running the applicable tests in test_namespace_pkgs.py. This causes _frozen_importlib.__package__ to get set properly, thus masking the problem. The problem is the use of relative imports in importlib._bootstrap. The solution is to accomplish this in a different way.

Here's a fix. If I don't hear from anyone right away I'll push it in a few hours (or tomorrow morning).

New changeset 46b2c99121f5 by Eric Snow in branch 'default': Issue #24192: Fix namespace package imports. https://hg.python.org/cpython/rev/46b2c99121f5

I can confirm that this fixes the problem I was having, both for the sample code I posted and for the real codebase I extracted this from. Thanks.

Great! The buildbots are happy too. :)