Affects: PythonCall
Describe the bug
Because the Python GIL is specific to an OS thread, but Julia tasks can migrate across threads when they're paused/resumed, locking the GIL from Julia on a non-sticky thread leads to undefined behavior (segfaults, deadlocks, etc.).
I think probably the way to handle this is to mark the current task as sticky in GIL.@lock and friends. It's really hard to get this right in libraries that wrap provide python functionality because even if callers know they need to lock the GIL, they may not know about task stickiness.
Your system
We've seen this across a wide variety of environments, Julia 1.10 to 1.12, macOS and ubuntu.
Affects: PythonCall
Describe the bug
Because the Python GIL is specific to an OS thread, but Julia tasks can migrate across threads when they're paused/resumed, locking the GIL from Julia on a non-sticky thread leads to undefined behavior (segfaults, deadlocks, etc.).
I think probably the way to handle this is to mark the current task as sticky in
GIL.@lockand friends. It's really hard to get this right in libraries that wrap provide python functionality because even if callers know they need to lock the GIL, they may not know about task stickiness.Your system
We've seen this across a wide variety of environments, Julia 1.10 to 1.12, macOS and ubuntu.