:LastChangedDate: $LastChangedDate$ :LastChangedRevision: $LastChangedRevision$ :LastChangedBy: $LastChangedBy$ Test-driven development with Twisted ==================================== Writing good code is hard, or at least it can be. A major challenge is to ensure that your code remains correct as you add new functionality. `Unit testing `_ is a modern, light-weight testing methodology in widespread use in many programming languages. Development that relies on unit tests is often referred to as Test-Driven Development (`TDD `_ ). Most Twisted code is tested using TDD. To gain a solid understanding of unit testing in Python, you should read the `unittest --Unit testing framework chapter `_ of the `Python LibraryReference `_ . There is a lot of information available online and in books. Introductory example of Python unit testing ------------------------------------------- This document is principally a guide to Trial, Twisted's unit testing framework. Trial is based on Python's unit testing framework. While we do not aim to give a comprehensive guide to general Python unit testing, it will be helpful to consider a simple non-networked example before expanding to cover networking code that requires the special capabilities of Trial. If you are already familiar with unit test in Python, jump straight to the section specific to :ref:`testing Twisted code ` . .. note:: In what follows we will make a series of refinements to some simple classes. In order to keep the examples and source code links complete and to allow you to run Trial on the intermediate results at every stage, I add ``_N`` (where the ``N`` are successive integers) to file names to keep them separate. This is a minor visual distraction that should be ignored. Creating an API and writing tests --------------------------------- We'll create a library for arithmetic calculation. First, create a project structure with a directory called ``calculus`` containing an empty ``__init__.py`` file. Then put the following simple class definition API into ``calculus/base_1.py`` : :download:`base_1.py ` .. literalinclude:: listings/trial/calculus/base_1.py (Ignore the ``test-case-name`` comment for now. You'll see why that's useful :ref:`below ` .) We've written the interface, but not the code. Now we'll write a set of tests. At this point of development, we'll be expecting all tests to fail. Don't worry, that's part of the point. Once we have a test framework functioning, and we have some decent tests written (and failing!), we'll go and do the actual development of our calculation API. This is the preferred way to work for many people using TDD - write tests first, make sure they fail, then do development. Others are not so strict and write tests after doing the development. Create a ``test`` directory beneath ``calculus`` , with an empty ``__init__.py`` file. In a ``calculus/test/test_base_1.py`` , put the following: :download:`test_base_1.py ` .. literalinclude:: listings/trial/calculus/test/test_base_1.py You should now have the following 4 files: .. code-block:: console calculus/__init__.py calculus/base_1.py calculus/test/__init__.py calculus/test/test_base_1.py To run the tests, there are two things you must set up. Make sure you get both done - nothing below will work unless you do. First, make sure that the directory that *contains* your ``calculus`` directory is in your Python load path. If you're using the Bash shell on some form of unix (e.g., Linux, Mac OS X), run ``PYTHONPATH="$PYTHONPATH:`pwd`/.."`` at the command line in the ``calculus`` directory. Once you have your Python path set up correctly, you should be able to run Python from the command line and ``import calculus`` without seeing an import error. Second, make sure you can run the ``trial`` command. That is, make sure the directory containing the ``trial`` program on you system is in your shell's ``PATH`` . The easiest way to check if you have this is to try running ``trial --help`` at the command line. If you see a list of invocation options, you're in business. If your shell reports something like ``trial: command not found`` , make sure you have Twisted installed properly, and that the Twisted ``bin`` directory is in your ``PATH`` . If you don't know how to do this, get some local help, or figure it out by searching online for information on setting and changing environment variables for you operating system. With those (one-time) preliminary steps out of the way, let's perform the tests. Run ``trial calculus.test.test_base_1`` from the command line when you are in the directory containing the ``calculus`` directory. You should see the following output (though your files are probably not in ``/tmp`` ): .. code-block:: console $ trial calculus.test.test_base_1 calculus.test.test_base_1 CalculationTestCase test_add ... [FAIL] test_divide ... [FAIL] test_multiply ... [FAIL] test_subtract ... [FAIL] =============================================================================== [FAIL] Traceback (most recent call last): File "/tmp/calculus/test/test_base_1.py", line 8, in test_add self.assertEqual(result, 11) twisted.trial.unittest.FailTest: not equal: a = None b = 11 calculus.test.test_base_1.CalculationTestCase.test_add =============================================================================== [FAIL] Traceback (most recent call last): File "/tmp/calculus/test/test_base_1.py", line 23, in test_divide self.assertEqual(result, 2) twisted.trial.unittest.FailTest: not equal: a = None b = 2 calculus.test.test_base_1.CalculationTestCase.test_divide =============================================================================== [FAIL] Traceback (most recent call last): File "/tmp/calculus/test/test_base_1.py", line 18, in test_multiply self.assertEqual(result, 60) twisted.trial.unittest.FailTest: not equal: a = None b = 60 calculus.test.test_base_1.CalculationTestCase.test_multiply =============================================================================== [FAIL] Traceback (most recent call last): File "/tmp/calculus/test/test_base_1.py", line 13, in test_subtract self.assertEqual(result, 4) twisted.trial.unittest.FailTest: not equal: a = None b = 4 calculus.test.test_base_1.CalculationTestCase.test_subtract ------------------------------------------------------------------------------- Ran 4 tests in 0.042s FAILED (failures=4) How to interpret this output? You get a list of the individual tests, each followed by its result. By default, failures are printed at the end, but this can be changed with the ``-e`` (or ``--rterrors`` ) option. One very useful thing in this output is the fully-qualified name of the failed tests. This appears at the bottom of each =-delimited area of the output. This allows you to copy and paste it to just run a single test you're interested in. In our example, you could run ``trial calculus.test.test_base_1.CalculationTestCase.test_subtract`` from the shell. Note that trial can use different reporters to modify its output. Run ``trial --help-reporters`` to see a list of reporters. The tests can be run by ``trial`` in multiple ways: - ``trial calculus`` : run all the tests for the calculus package. - ``trial calculus.test`` : run using Python's ``import`` notation. - ``trial calculus.test.test_base_1`` : as above, for a specific test module. You can follow that logic by putting your class name and even a method name to only run those specific tests. - .. _core-howto-trial-comment: ``trial --testmodule=calculus/base_1.py`` : use the ``test-case-name`` comment in the first line of ``calculus/base_1.py`` to find the tests. - ``trial calculus/test`` : run all the tests in the test directory (not recommended). - ``trial calculus/test/test_base_1.py`` : run a specific test file (not recommended). The first 3 versions using full qualified names are strongly encouraged: they are much more reliable and they allow you to easily be more selective in your test runs. You'll notice that Trial creates a ``_trial_temp`` directory in the directory where you run the tests. This has a file called ``test.log`` which contains the log output of the tests (created using ``log.msg`` or ``log.err`` functions). Examine this file if you add logging to your tests. Making the tests pass --------------------- Now that we have a working test framework in place, and our tests are failing (as expected) we can go and try to implement the correct API. We'll do that in a new version of the above base_1 module, ``calculus/base_2.py`` : :download:`base_2.py ` .. literalinclude:: listings/trial/calculus/base_2.py We'll also create a new version of test_base_1 which imports and tests this new implementation, in ``calculus/test_base_2.py`` : :download:`test_base_2.py ` .. literalinclude:: listings/trial/calculus/test/test_base_2.py is a copy of test_base_1, but with the import changed. Run ``trial`` again as above, and your tests should now pass: .. code-block:: console $ trial calculus.test.test_base_2 Running 4 tests. calculus.test.test_base CalculationTestCase test_add ... [OK] test_divide ... [OK] test_multiply ... [OK] test_subtract ... [OK] ------------------------------------------------------------------------------- Ran 4 tests in 0.067s PASSED (successes=4) Factoring out common test logic ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ You'll notice that our test file contains redundant code. Let's get rid of that. Python's unit testing framework allows your test class to define a ``setUp`` method that is called before *each* test method in the class. This allows you to add attributes to ``self`` that can be used in test methods. We'll also add a parameterized test method to further simplify the code. Note that a test class may also provide the counterpart of ``setUp`` , named ``tearDown`` , which will be called after *each* test (whether successful or not). ``tearDown`` is mainly used for post-test cleanup purposes. We will not use ``tearDown`` until later. Create ``calculus/test/test_base_2b.py`` as follows: :download:`test_base_2b.py ` .. literalinclude:: listings/trial/calculus/test/test_base_2b.py Much cleaner, isn't it? We'll now add some additional error tests. Testing just for successful use of the API is generally not enough, especially if you expect your code to be used by others. Let's make sure the ``Calculation`` class raises exceptions if someone tries to call its methods with arguments that cannot be converted to integers. We arrive at ``calculus/test/test_base_3.py`` : :download:`test_base_3.py ` .. literalinclude:: listings/trial/calculus/test/test_base_3.py We've added four new tests and one general-purpose function, ``_test_error`` . This function uses the ``assertRaises`` method, which takes an exception class, a function to run and its arguments, and checks that calling the function on the arguments does indeed raise the given exception. If you run the above, you'll see that not all tests fail. In Python it's often valid to add and multiply objects of different and even differing types, so the code in the add and multiply tests does not raise an exception and those tests therefore fail. So let's add explicit type conversion to our API class. This brings us to ``calculus/base_3.py`` : :download:`base_3.py ` .. literalinclude:: listings/trial/calculus/base_3.py Here the ``_make_ints`` helper function tries to convert a list into a list of equivalent integers, and raises a ``TypeError`` in case the conversion goes wrong. .. note:: The ``int`` conversion can also raise a ``TypeError`` if passed something of the wrong type, such as a list. We'll just let that exception go by, as ``TypeError`` is already what we want in case something goes wrong. .. _core-howto-trial-twisted: Twisted specific testing ------------------------ Up to this point we've been doing fairly standard Python unit testing. With only a few cosmetic changes (most importantly, directly importing ``unittest`` instead of using Twisted's :api:`twisted.trial.unittest ` version) we could make the above tests run using Python's standard library unit testing framework. Here we will assume a basic familiarity with Twisted's network I/O, timing, and Deferred APIs. If you haven't already read them, you should read the documentation on :doc:`Writing Servers ` , :doc:`Writing Clients ` , and :doc:`Deferreds ` . Now we'll get to the real point of this tutorial and take advantage of Trial to test Twisted code. Testing a protocol ------------------ We'll now create a custom protocol to invoke our class from a telnet-like session. We'll remotely call commands with arguments and read back the response. The goal will be to test our network code without creating sockets. Creating and testing the server ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ First we'll write the tests, and then explain what they do. The first version of the remote test code is: :download:`test_remote_1.py ` .. literalinclude:: listings/trial/calculus/test/test_remote_1.py To fully understand this client, it helps a lot to be comfortable with the Factory/Protocol/Transport pattern used in Twisted. We first create a protocol factory object. Note that we have yet to see the ``RemoteCalculationFactory`` class. It is in ``calculus/remote_1.py`` below. We call ``buildProtocol`` to ask the factory to build us a protocol object that knows how to talk to our server. We then make a fake network transport, an instance of ``twisted.test.proto_helpers.StringTransport`` class (note that test packages are generally not part of Twisted's public API;``twisted.test.proto_helpers`` is an exception). This fake transport is the key to the communications. It is used to emulate a network connection without a network. The address and port passed to ``buildProtocol`` are typically used by the factory to choose to immediately deny remote connections; since we're using a fake transport, we can choose any value that will be acceptable to the factory. In this case the factory just ignores the address, so we don't need to pick anything in particular. Testing protocols without the use of real network connections is both simple and recommended when testing Twisted code. Even though there are many tests in Twisted that use the network, most good tests don't. The problem with unit tests and networking is that networks aren't reliable. We cannot know that they will exhibit reasonable behavior all the time. This creates intermittent test failures due to network vagaries. Right now we're trying to test our Twisted code, not network reliability. By setting up and using a fake transport, we can write 100% reliable tests. We can also test network failures in a deterministic manner, another important part of your complete test suite. The final key to understanding this client code is the ``_test`` method. The call to ``dataReceived`` simulates data arriving on the network transport. But where does it arrive? It's handed to the ``lineReceived`` method of the protocol instance (in ``calculus/remote_1.py`` below). So the client is essentially tricking the server into thinking it has received the operation and the arguments over the network. The server (once again, see below) hands over the work to its ``CalculationProxy`` object which in turn hands it to its ``Calculation`` instance. The result is written back via ``sendLine`` (into the fake string transport object), and is then immediately available to the client, who fetches it with ``tr.value()`` and checks that it has the expected value. So there's quite a lot going on behind the scenes in the two-line ``_test`` method above. *Finally* , let's see the implementation of this protocol. Put the following into ``calculus/remote_1.py`` : :download:`remote_1.py ` .. literalinclude:: listings/trial/calculus/remote_1.py As mentioned, this server creates a protocol that inherits from :api:`twisted.protocols.basic.LineReceiver ` , and then a factory that uses it as protocol. The only trick is the ``CalculationProxy`` object, which calls ``Calculation`` methods through ``remote_*`` methods. This pattern is used frequently in Twisted, because it is very explicit about what methods you are making accessible. If you run this test (``trial calculus.test.test_remote_1`` ), everything should be fine. You can also run a server to test it with a telnet client. To do that, call ``python calculus/remote_1.py`` . You should have the following output: .. code-block:: console 2008-04-25 10:53:27+0200 [-] Log opened. 2008-04-25 10:53:27+0200 [-] __main__.RemoteCalculationFactory starting on 46194 2008-04-25 10:53:27+0200 [-] Starting factory <__main__.RemoteCalculationFactory instance at 0x846a0cc> 46194 is replaced by a random port. You can then call telnet on it: :: $ telnet localhost 46194 Trying 127.0.0.1... Connected to localhost. Escape character is '^]'. add 4123 9423 13546 It works! Creating and testing the client ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Of course, what we build is not particularly useful for now: we'll now build a client for our server, to be able to use it inside a Python program. And it will serve our next purpose. Create ``calculus/test/test_client_1.py`` : :download:`test_client_1.py ` .. literalinclude:: listings/trial/calculus/test/test_client_1.py It's really symmetric to the server test cases. The only tricky part is that we don't use a client factory. We're lazy, and it's not very useful in the client part, so we instantiate the protocol directly. Incidentally, we have introduced a very important concept here: the tests now return a Deferred object, and the assertion is done in a callback. When a test returns a Deferred, the reactor is run until the Deferred fires and its callbacks run. The important thing to do here is to **not forget to return the Deferred** . If you do, your tests will pass even if nothing is asserted. That's also why it's important to make tests fail first: if your tests pass whereas you know they shouldn't, there is a problem in your tests. We'll now add the remote client class to produce ``calculus/client_1.py`` : :download:`client_1.py ` .. literalinclude:: listings/trial/calculus/client_1.py More good practices ------------------- Testing scheduling ~~~~~~~~~~~~~~~~~~ When testing code that involves the passage of time, waiting e.g. for a two hour timeout to occur in a test is not very realistic. Twisted provides a solution to this, the :api:`twisted.internet.task.Clock ` class that allows one to simulate the passage of time. As an example we'll test the code for client request timeout: since our client uses TCP it can hang for a long time (firewall, connectivity problems, etc...). So generally we need to implement timeouts on the client side. Basically it's just that we send a request, don't receive a response and expect a timeout error to be triggered after a certain duration. :download:`test_client_2.py ` .. literalinclude:: listings/trial/calculus/test/test_client_2.py What happens here? We instantiate our protocol as usual, the only trick is to create the clock, and assign ``proto.callLater`` to ``clock.callLater`` . Thus, every ``callLater`` call in the protocol will finish before ``clock.advance()`` returns. In the new test (test_timeout), we call ``clock.advance`` , that simulates an advance in time (logically it's similar to a ``time.sleep`` call). And we just have to verify that our Deferred got a timeout error. Let's implement that in our code. :download:`client_2.py ` .. literalinclude:: listings/trial/calculus/client_2.py If everything completed successfully, it is important to remember to cancel the ``DelayedCall`` returned by ``callLater`` . Cleaning up after tests ~~~~~~~~~~~~~~~~~~~~~~~ This chapter is mainly intended for people who want to have sockets or processes created in their tests. If it's still not obvious, you must try to avoid using them, because it ends up with a lot of problems, one of them being intermittent failures. And intermittent failures are the plague of automated tests. To actually test that, we'll launch a server with our protocol. :download:`test_remote_2.py ` .. literalinclude:: listings/trial/calculus/test/test_remote_2.py Recent versions of trial will fail loudly if you remove the ``stopListening`` call, which is good. Also, you should be aware that ``tearDown`` will be called in any case, after success or failure. So don't expect every object you created in the test method to be present, because your tests may have failed in the middle. Trial also has a ``addCleanup`` method, which makes these kind of cleanups easy and removes the need for ``tearDown`` . For example, you could remove the code in ``_test`` this way: .. code-block:: python def setUp(self): factory = RemoteCalculationFactory() self.port = reactor.listenTCP(0, factory, interface="127.0.0.1") self.addCleanup(self.port.stopListening) def _test(self, op, a, b, expected): creator = protocol.ClientCreator(reactor, RemoteCalculationClient) def cb(client): self.addCleanup(self.client.transport.loseConnection) return getattr(client, op)(a, b).addCallback(self.assertEqual, expected) return creator.connectTCP('127.0.0.1', self.port.getHost().port).addCallback(cb) This removes the need of a ``tearDown`` method, and you don't have to check for the value of self.client: you only call addCleanup when the client is created. Handling logged errors ~~~~~~~~~~~~~~~~~~~~~~ Currently, if you send an invalid command or invalid arguments to our server, it logs an exception and closes the connection. This is a perfectly valid behavior, but for the sake of this tutorial, we want to return an error to the user if they send invalid operators, and log any errors on server side. So we'll want a test like this: .. code-block:: python def test_invalidParameters(self): self.proto.dataReceived('add foo bar\r\n') self.assertEqual(self.tr.value(), "error\r\n") :download:`remote_2.py ` .. literalinclude:: listings/trial/calculus/remote_2.py If you try something like that, it will not work. Here is the output you should have: .. code-block:: console trial calculus.test.test_remote_3.RemoteCalculationTestCase.test_invalidParameters calculus.test.test_remote_3 RemoteCalculationTestCase test_invalidParameters ... [ERROR] =============================================================================== [ERROR]: calculus.test.test_remote_3.RemoteCalculationTestCase.test_invalidParameters Traceback (most recent call last): File "/tmp/calculus/remote_2.py", line 27, in lineReceived result = op(a, b) File "/tmp/calculus/base_3.py", line 11, in add a, b = self._make_ints(a, b) File "/tmp/calculus/base_3.py", line 8, in _make_ints raise TypeError exceptions.TypeError: ------------------------------------------------------------------------------- Ran 1 tests in 0.004s FAILED (errors=1) At first, you could think there is a problem, because you catch this exception. But in fact trial doesn't let you do that without controlling it: you must expect logged errors and clean them. To do that, you have to use the ``flushLoggedErrors`` method. You call it with the exception you expect, and it returns the list of exceptions logged since the start of the test. Generally, you'll want to check that this list has the expected length, or possibly that each exception has an expected message. We do the former in our test: :download:`test_remote_3.py ` .. literalinclude:: listings/trial/calculus/test/test_remote_3.py Resolve a bug ------------- A bug was left over during the development of the timeout (probably several bugs, but that's not the point), concerning the reuse of the protocol when you got a timeout: the connection is not dropped, so you can get timeout forever. Generally a user will come to you saying "I have this strange problem on my crappy network. It seems you could solve it with doing XXX at YYY." Actually, this bug can be corrected several ways. But if you correct it without adding tests, one day you'll face a big problem: regression. So the first step is adding a failing test. :download:`test_client_3.py ` .. literalinclude:: listings/trial/calculus/test/test_client_3.py What have we done here ? - We switched to StringTransportWithDisconnection. This transport manages ``loseConnection`` and forwards it to its protocol. - We assign the protocol to the transport via the ``protocol`` attribute. - We check that after a timeout our connection has closed. For doing that, we then use the ``TimeoutMixin`` class, that does almost everything we want. The great thing is that it almost changes nothing to our class. :download:`client_3.py ` .. literalinclude:: listings/trial/calculus/client_3.py Testing Deferreds without the reactor ------------------------------------- Above we learned about returning Deferreds from test methods in order to make assertions about their results, or side-effects that only happen after they fire. This can be useful, but we don't actually need the feature in this example. Because we were careful to use ``Clock`` , we don't need the global reactor to run in our tests. Instead of returning the Deferred with a callback attached to it which performs the necessary assertions, we can use a testing helper, :api:`twisted.trial.unittest.SynchronousTestCase.successResultOf ` (and the corresponding error-case helper :api:`twisted.trial.unittest.SynchronousTestCase.failureResultOf ` ), to extract its result and make assertions against it directly. Compared to returning a Deferred, this avoids the problem of forgetting to return the Deferred, improves the stack trace reported when the assertion fails, and avoids the complexity of using global reactor (which, for example, may then require cleanup). :download:`test_client_4.py ` .. literalinclude:: listings/trial/calculus/test/test_client_4.py This version of the code makes the same assertions, but no longer returns any Deferreds from any test methods. Instead of making assertions about the result of the Deferred in a callback, it makes the assertions as soon as it *knows* the Deferred is supposed to have a result (in the ``_test`` method and in ``test_timeout`` and ``test_timeoutConnectionLost`` ). The possibility of *knowing* exactly when a Deferred is supposed to have a test is what makes ``successResultOf`` useful in unit testing, but prevents it from being applicable to non-testing purposes. ``successResultOf`` will raise an exception (failing the test) if the ``Deferred`` passed to it does not have a result, or has a failure result. Similarly, ``failureResultOf`` will raise an exception (also failing the test) if the ``Deferred`` passed to it does not have a result, or has a success result. There is a third helper method for testing the final case, :api:`twisted.trial.unittest.SynchronousTestCase.assertNoResult ` , which only raises an exception (failing the test) if the ``Deferred`` passed to it *has* a result (either success or failure). Dropping into a debugger ------------------------ In the course of writing and running your tests, it is often helpful to employ the use of a debugger. This can be particularly helpful in tracking down where the source of a troublesome bug is in your code. Python's standard library includes a debugger in the form of the `pdb `_ module. Running your tests with ``pdb`` is as simple as invoking twisted with the ``--debug`` option, which will start ``pdb`` at the beginning of the execution of your test suite. Trial also provides a ``--debugger`` option which can run your test suite using another debugger instead. To specify a debugger other than ``pdb`` , pass in the fully-qualified name of an object that provides the same interface as ``pdb`` . Most third-party debuggers tend to implement an interface similar to ``pdb`` , or at least provide a wrapper object that does. For example, invoking trial with the line ``trial --debug --debugger pudb`` will open the `PuDB `_ debugger instead, provided it is properly installed. Code coverage ------------- Code coverage is one of the aspects of software testing that shows how much your tests cross (cover) the code of your program. There are different kinds of measures: path coverage, condition coverage, statement coverage... We'll only consider statement coverage here, whether a line has been executed or not. Trial has an option to generate the statement coverage of your tests. This option is --coverage. It creates a coverage directory in _trial_temp, with a file .cover for every module used during the tests. The ones interesting for us are calculus.base.cover and calculus.remote.cover. Each line starts with a counter showing how many times the line was executed during the tests, or the marker '>>>>>>' if the line was not covered. If you went through the whole tutorial to this point, you should have complete coverage :). Again, this is only another useful pointer, but it doesn't mean your code is perfect: your tests should consider every possible input and output, to get **full** coverage (condition, path, etc.) as well . Conclusion ---------- So what did you learn in this document? - How to use the trial command-line tool to run your tests - How to use string transports to test individual clients and servers without creating sockets - If you really want to create sockets, how to cleanly do it so that it doesn't have bad side effects - And some small tips you can't live without. If one of the topics still looks cloudy to you, please give us your feedback! You can file tickets to improve this document - learn how to contribute `on the Twisted web site `_.