Replace old nose tests with pytest based tests

[Zsailer]

Okay, I've gone through every test and rewritten it using only pytest's fixture design/syntax.

This also uses the pytest-tornasync to enable an async/await syntax for tests that hit the server REST API.

One cool thing about these pytest fixtures is that the conftest.py offers some really nice fixtures that we could generalize for anyone writing a server extension (i.e. jupyterlab, voila, etc.). These fixtures can be used to easily spin up a configurable test ServerApp. Some helpful fixtures include:

• serverapp : a vanilla running server app.
• configurable_serverapp: a configurable (via argv or a config dict) server that will start running after configurable.
• app: the server web application
• fetch: a function that hits the serverapp's REST API.

@kevin-bates, @echarles, and @maartenbreddels

[Zsailer]

This is a massive PR. If the tests pass, I propose we merge and address problems as they arise.

[Zsailer]

@kevin-bates One of the gateway tests is failing. Here's the error I get. Do you have any immediate ideas?

______________________ test_gateway_get_named_kernelspec _______________________

init_gateway = None
fetch = <function fetch.<locals>.client_fetch at 0x1078de0d0>

    async def test_gateway_get_named_kernelspec(init_gateway, fetch):
        # Validate that a specific kernelspec can be retrieved from gateway (and an invalid spec can't)
        with mocked_gateway:
            r = await fetch(
                'api', 'kernelspecs', 'kspec_foo',
                method='GET'
            )
            assert r.code == 200
            kspec_foo = json.loads(r.body.decode('utf-8'), encoding='utf-8')
            assert kspec_foo.get('name') == 'kspec_foo'
    
            with pytest.raises(tornado.httpclient.HTTPClientError) as e:
                await fetch(
                    'api', 'kernelspecs', 'no_such_spec',
                    method='GET'
                )
>           assert expected_http_error(e, 404)
E           assert False
E            +  where False = expected_http_error(<ExceptionInfo HTTP 500: Internal Server Error tblen=1>, 404)

[Zsailer]

The 3.5 tests are failing due to a pathlib error.

[kevin-bates]

@Zsailer - regarding the gateway test failure. I need to look into this. This doesn't occur when the tests are run from a separate location (as we were doing with the juptyer_server_tests repo). Only when collocated.

[kevin-bates]

The gateway test failure was due to the fact that the test rewrite was based on the test code prior to PR #139. Since we were running tests against an older branch of jupyter_server, all was good. Once this PR was created, however, it introduced a new server code vs. old test code. The changes are only to the import statements surrounding the HTTPError (mess). Once I updated the reworked test with the updated imports from #139 the test passed. The updates to be merged into the pytest branch can be found here: Zsailer#1

Note that in either case (before or after my fix), the travis tests aren't running at all. With the change, all github actions pass (except for 3.8 on Windows - which we know about) and travis (where no tests are running).

[Zsailer]

• Appveyor and travis are failing because of the nbconvert problem mentioned in test(fix): make tests run with nbconvert >=6,<7 jupyter/notebook#5092. I don't think we should block on this test. I think we should fix this by pinning nbconvert to <6 until that PR lands.

• Windows 3.8 tests are failing because of workaround tornado py38 compatibility issue jupyter/notebook#5047.

• Windows 3.6 in Github Actions is failing for unknown reasons. I'm working on that next.

[kevin-bates]

@echarles and I just finished looking into the Windows 3.8 issues and the issue appears to be related to using the pytest-tornasync plugin. In there, it sets up an io_loop fixture so that it can ensure a new loop is setup for each test.

Having added print statements, it was determined that the patch in jupyter/notebook#5047 was establishing the correct event loop, even within each test, but the pytest-tornasync fixtures must be coming into play after that. Since we had already setup a asyncio_patch fixture, we found that overriding the plugin's io_loop fixture with a dependency on the asyncio_patch fixture proper behavior was exhibited. Here's the change...

@pytest.fixture
def asyncio_patch():
    _init_asyncio_patch()


@pytest.fixture
def io_loop(asyncio_patch):
    """
    Create new io loop for each test, and tear it down after.
    """
    loop = tornado.ioloop.IOLoop()
    loop.make_current()
    yield loop
    loop.clear_current()
    loop.close(all_fds=True)

I'm sure there's probably a better way to do this, but I also suspect this is the idea behind fixtures (last override wins). Still not sure why this was necessary since that patch's installment of the SelectorEventLoopPolicy should have taken place already.

[kevin-bates]

I think it might be a good idea to add a comment here that this fixture (sans the asyncio_patch reference) comes directly from plugin pytest-tornasync and that we need to override fixture io_loop in order to ensure the appropriate event loop policy is in place and that these (asyncio_patch and this io_loop) fixtures can be removed once the ProactorEventLoop works correctly on Windows (and pytlhon 3.8+).

[Zsailer]

👍

[Zsailer]

ALL GREEN. 🎉

@echarles or @kevin-bates you do the honors!

[Zsailer]

@echarles told me to merge it!

[echarles]

Congrats @Zsailer