What happens?
If Python is embedded in another application and the process terminates without a chance for Py_Finalize to be called, there's a DuckDBPyConnection object that gets destroyed by the C runtime when the process ends, and that causes the process to crash.
In ~DuckDBPyConnection it's assumed the GIL is already held, but this isn't true when the process is being terminated (gil_scoped_release calls PyEval_SaveThread, and PyEval_SaveThread crashes because the GIL is not held). This object is cleaned up if Py_Finalize is called via a capsule, but that shouldn't be relied on.
A quick fix would be to add Py_IsInitialized and PyGILState_Check checks to ~DuckDBPyConnection, but a better fix would be to not do this cleanup in the destructor of a static and instead to move this to module state so Python can manage its lifecycle. I see there's a comment about PEP 489 already in the code suggesting the same.
To Reproduce
# Importing duckdb is enough to cause this script to crash.
# Comment this line out and it will work fine.
import duckdb
import ctypes
# ctypes releases the GIL around this call (standard behavior for foreign DLL calls,
# unlike ctypes.pythonapi calls). ExitProcess never returns, so the GIL stays released
# from here on - (no Python code running, GIL not held).
# Py_Finalize is never called. ExitProcess still triggers DLL_PROCESS_DETACH for every
# loaded DLL, so _duckdb.pyd's static destructor runs with no GIL held and no valid
# thread state for this thread.
ctypes.windll.kernel32.ExitProcess(0)
OS:
Windows
DuckDB Package Version:
1.5.4
Python Version:
3.14.5
Full Name:
Tony Roberts
Affiliation:
PyXLL Ltd
What is the latest build you tested with? If possible, we recommend testing with the latest nightly build.
I have tested with a source build
Did you include all relevant data sets for reproducing the issue?
Not applicable - the reproduction does not require a data set
Did you include all code required to reproduce the issue?
Did you include all relevant configuration to reproduce the issue?
What happens?
If Python is embedded in another application and the process terminates without a chance for Py_Finalize to be called, there's a DuckDBPyConnection object that gets destroyed by the C runtime when the process ends, and that causes the process to crash.
In ~DuckDBPyConnection it's assumed the GIL is already held, but this isn't true when the process is being terminated (gil_scoped_release calls PyEval_SaveThread, and PyEval_SaveThread crashes because the GIL is not held). This object is cleaned up if Py_Finalize is called via a capsule, but that shouldn't be relied on.
A quick fix would be to add Py_IsInitialized and PyGILState_Check checks to ~DuckDBPyConnection, but a better fix would be to not do this cleanup in the destructor of a static and instead to move this to module state so Python can manage its lifecycle. I see there's a comment about PEP 489 already in the code suggesting the same.
To Reproduce
OS:
Windows
DuckDB Package Version:
1.5.4
Python Version:
3.14.5
Full Name:
Tony Roberts
Affiliation:
PyXLL Ltd
What is the latest build you tested with? If possible, we recommend testing with the latest nightly build.
I have tested with a source build
Did you include all relevant data sets for reproducing the issue?
Not applicable - the reproduction does not require a data set
Did you include all code required to reproduce the issue?
Did you include all relevant configuration to reproduce the issue?