How to write modern Python tests?

What is the latest way to write Python 开发者_如何学Gotests? What modules/frameworks to use?

And another question: are doctest tests still of any value? Or should all the tests be written in a more modern testing framework?

Thanks, Boda Cydo.

The usual way is to use the builtin unittest module for creating unit tests and bundling them together to test suites which can be run independently. unittest is very similar to (and inspired by) jUnit and thus very easy to use.

If you're interested in the very latest changes, take a look at the new PyCon talk by Michael Foord:

PyCon 2010: New and Improved: Coming changes to unittest

Using the built-in unittest module is as relevant and easy as ever. The other unit testing options, py.test,nose, and twisted.trial are mostly compatible with unittest.

Doctests are of the same value they always were—they are great for testing your documentation, not your code. If you are going to put code examples in your docstrings, doctest can assure you keep them correct and up to date. There's nothing worse than trying to reproduce an example and failing, only to later realize it was actually the documentation's fault.

I don't know much about doctests, but at my university, nose testing is taught and encouraged.

Nose can be installed by following this procedure (I'm assuming you're using a PC - Windows OS):

1. install setuptools
2. Run DOS Command Prompt (Start -> All Programs -> Accessories -> Command Prompt)
3. For this step to work, you must be connected to the internet. In DOS, type: C:\Python25\Scripts\easy_install nose

If you are on a different OS, check this site

EDIT:

It's been two years since I originally wrote this post. Now, I've learned of this programming principle called Designing by Contract. This allows a programmer to define preconditions, postconditions and invariants (called contracts) for all functions in their code. The effect is that an error is raised if any of these contracts are violated.

The DbC framework that I would recommend for python is called PyContract I have successfully used it in my evolutionary programming framework

In my current project I'm using unittest, minimock, nose. In the past I've made heavy use of doctests, but in a large projects some tests can get kinda unwieldy, so I tend to reserve usage of doctests for simpler functions.

If you are using setuptools or distribute (you should be switching to distribute), you can set up nose as the default test collector so that you can run your tests with "python setup.py test"

setup(name='foo',
      ...
      test_suite='nose.collector',
      ...

Now running "python setup.py test" will invoke nose, which will crawl your project for things that look like tests and run them, accumulating the results. If you also have doctests in your project, you can run nosetests with the --with-doctest option to enable the doctest plugin.

nose also has integration with coverage

nosetests --with-coverage.

You can also use the --cover-html --cover-html-dir options to generate an HTML coverage report for each module, with each line of code that is not under test highlighted. I wouldn't get too obsessed with getting coverage to report 100% test coverage for all modules. Some code is better left for integration tests, which I'll cover at the end.

I have become a huge fan of minimock, as it makes testing code with a lot of external dependencies really easy. While it works really well when paired with doctest, it can be used with any testing framework using the unittest.TraceTracker class. I would encourage you to avoid using it to test all of your code though, since you should still try to write your code so that each translation unit can be tested in isolation without mocking. Sometimes that's not possible though.

Here is an (untested) example of such a test using minimock and unittest:

# tests/test_foo.py
import minimock
import unittest

import foo

class FooTest(unittest2.TestCase):
    def setUp(self):
        # Track all calls into our mock objects. If we don't use a TraceTracker
        # then all output will go to stdout, but we want to capture it.
        self.tracker = minimock.TraceTracker()

    def tearDown(self):
        # Restore all objects in global module state that minimock had
        # replaced.
        minimock.restore()

    def test_bar(self):
        # foo.bar invokes urllib2.urlopen, and then calls read() on the
        # resultin file object, so we'll use minimock to create a mocked
        # urllib2.
        urlopen_result = minimock.Mock('urlobject', tracker=self.tracker)
        urlopen_result.read = minimock.Mock(
            'urlobj.read', tracker=self.tracker, returns='OMG')
        foo.urllib2.urlopen = minimock.Mock(
            'urllib2.urlopen', tracker=self.tracker, returns=urlopen_result)

        # Now when we call foo.bar(URL) and it invokes
        # *urllib2.urlopen(URL).read()*, it will not actually send a request
        # to URL, but will instead give us back the dummy response body 'OMG',
        # which it then returns.
        self.assertEquals(foo.bar('http://example.com/foo'), 'OMG')

        # Now we can get trace info from minimock to verify that our mocked
        # urllib2 was used as intended. self.tracker has traced our calls to
        # *urllib2.urlopen()*
        minimock.assert_same_trace(self.tracker, """\
Called urllib2.urlopen('http://example.com/foo)
Called urlobj.read()
Called urlobj.close()""")

Unit tests shouldn't be the only kinds of tests you write though. They are certainly useful and IMO extremely important if you plan on maintaining this code for any extended period of time. They make refactoring easier and help catch regressions, but they don't really test the interaction between various components and how they interact (if you do it right).

When I start getting to the point where I have a mostly finished product with decent test coverage that I intend to release, I like to write at least one integration test that runs the complete program in an isolated environment.

I've had a lot of success with this on my current project. I had about 80% unit test coverage, and the rest of the code was stuff like argument parsing, command dispatch and top level application state, which is difficult to cover in unit tests. This program has a lot of external dependencies, hitting about a dozen different web services and interacting with about 6,000 machines in production, so running this in isolation proved kinda difficult.

I ended up writing an integration test which spawns a WSGI server written with eventlet and webob that simulates all of the services my program interacts with in production. Then the integration test monkey patches our web service client library to intercept all HTTP requests and send them to the WSGI application. After doing that, it loads a state file that contains a serialized snapshot of the state of the cluster, and invokes the application by calling it's main() function. Now all of the external services my program interacts with are simulated, so that I can run my program as it would be run in production in a repeatable manner.

The important thing to remember about doctests is that the tests are based on string comparisons, and the way that numbers are rendered as strings will vary on different platforms and even in different python interpreters.

Most of my work deals with computations, so I use doctests only to test my examples and my version string. I put a few in the __init__.py since that will show up as the front page of my epydoc-generated API documentation.

I use nose for testing, although I'm very interested in checking out the latest changes to py.test.