Repository navigation
Segfault on re-entrant lazy import pandas #159408
Description
Crash report
With PYTHON_LAZY_IMPORTS=all, I get a segfault with pandas being initialized inside a freezegun, likely because GC frees it somehow.
Tested with freezegun==1.5.5 and pandas==3.0.6, require 2 files below
lib.py
import pandasrepro.py
import lib
from freezegun import freeze_time
for _ in range(2): # the first call is fine, the second crashes
with freeze_time("2024-01-01"):
passCalling it
$ PYTHON_LAZY_IMPORTS=all python repro.py
Fatal Python error: Segmentation fault
Current thread ... (most recent call first):
Garbage-collecting
File "<frozen importlib._bootstrap_external>", line 500 in _compile_bytecode
...
freezegun runs dir() and getattr() over every module. That getattr resolves the lazy pandas reference, but if i try to reproduce without freezegun, it gives an error but no segfault
repro-python.py
import sys
import pandas
for _ in range(2):
for m in list(sys.modules.values()):
for n in dir(m):
try: getattr(m, n)
except ModuleNotFoundError: passRunning
$ PYTHON_LAZY_IMPORTS=all python repro-python.py
File "pandas/_libs/tslibs/np_datetime.pyx", line 33, in init pandas._libs.tslibs.np_datetime
File ".../numpy/__init__.cython-30.pxd", line 1081, in numpy.import_array
ImportError: numpy._core.multiarray failed to import
Possibly related to #158402 or #159266 but very different crashes
CPython versions tested on:
3.15
Operating systems tested on:
Linux
Output from running 'python -VV' on the command line:
Python 3.15.0 (main, Oct 9 2026, 15:24:40) [Clang 22.1.3 ]
Activity
zolaski333 commentedon Oct 14, 2026
The GC crash also reproduces without lazy imports on Python 3.12, and without freezegun by replacing datetime classes before importing a small Cython extension. The native stack recurses between the extension traversal callback and subtype_traverse().
Cython's native datetime declarations currently resolve mutable module attributes. The proposed fix in cython/cython#8059 now supports C type objects in extern declarations generically. datetime.pxd obtains the native types from PyDateTime_CAPI; the compiler has no datetime-specific handling. The interface remains under review. Downstream extensions would need regeneration and rebuilding.
There is also a separate pandas initialization problem: its option-registration and datetime/parser capsule imports can be deferred in global lazy mode. pandas-dev/pandas#71173 tracks that, and pandas-dev/pandas#71178 proposes explicit initialization imports.
With both proposed changes and rebuilt pandas, the reworked fix passes the original two-file scenario in five fresh Linux processes with empty stderr and no CPython source changes. The reduced C/C++ regressions pass on Windows and Linux, including free-threaded CPython. This does not establish general pandas compatibility with global lazy imports: a separate Timestamp import cycle remains.
A verified workaround is to initialize pandas with normal imports before entering freeze_time: