Thanks

There was https://bugs.python.org/issue24748, it is probably also important in this context.

The official version of load_dynamic is:

    def load_dynamic(name, path, file=None):
        """**DEPRECATED**
        Load an extension module.
        """
        import importlib.machinery
        loader = importlib.machinery.ExtensionFileLoader(name, path)

        # Issue #24748: Skip the sys.modules check in _load_module_shim;
        # always load new extension
        spec = importlib.machinery.ModuleSpec(
            name=name, loader=loader, origin=path)
        return _load(spec)

In pyximport there is also imp imported, i.e. https://github.com/cython/cython/blob/master/pyximport/pyximport.py#L51

However, I'm not sure the cure is not worse than the disease: What is the point in copying the code from imp and then fixing all the copies once a new bug was fixed in imp?

PEP 489 states "All changed loaders (BuiltinImporter and ExtensionFileLoader) will remain backwards-compatible; the load_module method will be replaced by a shim." so I think this change is safe until those shims will be removed by a future PEP. It could be improved to support PEP 489 multi-phase initialization if needed.

The python issue you point to has to do with loading a module when that module already has been loaded. This is not possible here since the module is newly-compiled.

In pyximport there is also imp imported, i.e. https://github.com/cython/cython/blob/master/pyximport/pyximport.py#L51

Interesting, cython is indeed huge. I guess that should be changed too.

The python issue you point to has to do with loading a module when that module already has been loaded. This is not possible here since the module is newly-compiled.

That is true. However, in the context of pyximport, which supports reloading, this probably should be taken into consideration. Having different implementations of load_dynamic might become quite confusing.