@rpkilby OK, too early... Tests pass but if __getattribute__ raises the AttributeError, why is __getattr__ not then (re-)called?

I might have this completely wrong, as I don't know much about the cpython internals, but... my understanding is that the underlying c code calls __getattribute__ and if an AttributeError is raised, it then calls __getattr__ (if defined). The default implementation of __getattribute__ isn't directly calling __getattr__, which is why we don't see an infinite loop.

In essence:

• Underlying C code calls Request.__getattribute__ which fails and raises an AttributeError.
• Underlying C code calls Request.__getattr__, which checks the WSGIRequest.
• WSGIRequest check fails, so Request.__getattr__ calls Request.__getattribute__ to raise the exception.

I just did a simple timeit test, and the PR does seem to be faster (which makes sense, since there are less caught exceptions. I used the following test:

class Perf(TestCase):
    def test_perf(self):
        from timeit import repeat

        wsgi_request = factory.get('/')
        request = Request(wsgi_request)

        # non-existent attribute
        def foo():
            try:
                request.foo
            except AttributeError:
                pass

        # attribute exists on WSGIRequest
        def body():
            request.body

        # attribute exists on DRF Request
        def _request():
            request._request

        print(min(repeat(foo, number=100000)))
        print(min(repeat(body, number=100000)))
        print(min(repeat(_request, number=100000)))

        raise Exception('done')

OK. It does read that way. Good linking skills.