Rationale to change the Python C API¶
To be able to introduce backward incompatible changes to the C API without breaking too many C extensions, this project proposes two things:
- design a helper layer providing removed functions;
- a new Python runtime which is only usable with C extensions compiled with the new stricter and smaller C API (and the new stable ABI) for Python 3.8 and newer, whereas the existing “regular python” becomes the “regular runtime” which provides maximum backward compatibility with Python 3.7 and older.
The current C API has multiple issues:
- The Python lifecycle is shorter than the lifecycle of some operating systems: how to get the latest Python on an “old” but stable operating system?
- Python debug build is currently mostly unusable in practice, making development of C extension harder, especially debugging.
PyPy cpyext is slow¶
The PyPy cpyext is slow because there is no efficient way to write a correct implementation of the current C API. For example, borrowed references is hard to optimize since the runtime cannot track the lifetime of a borrowed object.
The stable ABI is not usable¶
The current Python C API produces an ABI which is far from stable. The API leaks many implementation details into the ABI.
For example, using the
PyTuple_GET_ITEM() macro produces machine code which
uses an hardcoded offset. If the PyObject structure changes, the offset
changes, and so the ABI changes.
See Relationship between the C API and the ABI.
Issues with an unstable ABI:
- Packaging. Developers have to publish one binary package per Python version (3.6, 3.7, 3.8, etc.). Upgrading Python 3 to a newer version is very hard because of this issue: each binary package must be distributed in many versions, one per Python version. A stable ABI would allow to build a package for Python 3.6 and use the same binary for Python 3.7, 3.8, 3.9, etc.
- Experiment issue. It is not possible to change the base Python structures like PyObject or PyTupleObject to reduce the memory footprint or implement new optimizations. Experiment such changes requires to rebuild C extensions.
- Debug builds. Most Linux distributions provide a debug build of Python
python3-dbg), but such build has a different ABI and so all C extensions must be recompiled to be usable on such flavor of Python. Being able to use debug builds would ease debugging since the debug builds add many debug checks which are too expensive (CPU/memory) to be enabled by default in a release build.
Performance problem: How can we make Python 2x faster?¶
“Leaking” implementation details into the ABI prevents many optimizations opportunities. See Optimization ideas.
Keep backward compatibility¶
Existing C extensions will still be supported and will not have to be modified. The old C API is not deprecated and there is no plan penalize users of the old C API.