Unit Tests in Twisted

Each unit test tests one bit of functionality in the software. Unit tests are entirely automated and complete quickly. Unit tests for the entire system are gathered into one test suite, and may all be run in a single batch. The result of a unit test is simple: either it passes, or it doesn't. All this means you can test the entire system at any time without inconvenience, and quickly see what passes and what fails.

Unit Tests in the Twisted Philosophy

The Twisted development team adheres to the practice of Extreme Programming (XP), and the usage of unit tests is a cornerstone XP practice. Unit tests are a tool to give you increased confidence. You changed an algorithm -- did you break something? Run the unit tests. If a test fails, you know where to look, because each test covers only a small amount of code, and you know it has something to do with the changes you just made. If all the tests pass, you're good to go, and you don't need to second-guess yourself or worry that you just accidentally broke someone else's program.

What to Test, What Not to Test

You don't have to write a test for every single method you write, only production methods that could possibly break.

-- Kent Beck, Extreme Programming Explained, p. 58.

Running the Tests

How

From the root of the Twisted source tree, run Trial:

$ bin/trial twisted
    

You'll find that having something like this in your emacs init files is quite handy:

(defun runtests () (interactive)
  (compile "python /somepath/Twisted/bin/trial /somepath/Twisted"))

(global-set-key [(alt t)] 'runtests)

When

Always, always, always be sure all the tests pass before committing any code. If someone else checks out code at the start of a development session and finds failing tests, they will not be happy and may decide to hunt you down.

Since this is a geographically dispersed team, the person who can help you get your code working probably isn't in the room with you. You may want to share your work in progress over the network, but you want to leave the main Subversion tree in good working order. So use a branch, and merge your changes back in only after your problem is solved and all the unit tests pass again.

Adding a Test

Please don't add new modules to Twisted without adding tests for them too. Otherwise we could change something which breaks your module and not find out until later, making it hard to know exactly what the change that broke it was, or until after a release, and nobody wants broken code in a release.

Tests go into dedicated test packages such as twisted/test/ or twisted/conch/test/, and are named test_foo.py, where foo is the name of the module or package being tested. Extensive documentation on using the PyUnit framework for writing unit tests can be found in the links section below.

One deviation from the standard PyUnit documentation: To ensure that any variations in test results are due to variations in the code or environment and not the test process itself, Twisted ships with its own, compatible, testing framework. That just means that when you import the unittest module, you will from twisted.trial import unittest instead of the standard import unittest.

As long as you have followed the module naming and placement conventions, trial will be smart enough to pick up any new tests you write.

PyUnit provides a large number of assertion methods to be used when writing tests. Many of these are redundant. For consistency, Twisted unit tests should use the assert forms rather than the fail forms. Also, use assertEqual, assertNotEqual, and assertAlmostEqual rather than assertEquals, assertNotEquals, and assertAlmostEquals. assertTrue is also preferred over assert_. You may notice this convention is not followed everywhere in the Twisted codebase. If you are changing some test code and notice the wrong method being used in nearby code, feel free to adjust it.

When you add a unit test, make sure all methods have docstrings specifying at a high level the intent of the test. That is, a description that users of the method would understand.

Test Implementation Guidelines

Here are some guidelines to follow when writing tests for the Twisted test suite. Many tests predate these guidelines and so do not follow them. When in doubt, follow the guidelines given here, not the example of old unit tests.

Real I/O

Most unit tests should avoid performing real, platform-implemented I/O operations. Real I/O is slow, unreliable, and unwieldy. When implementing a protocol, twisted.test.proto_helpers.StringTransport can be used instead of a real TCP transport. StringTransport is fast, deterministic, and can easily be used to exercise all possible network behaviors.

Real Time

Most unit tests should also avoid waiting for real time to pass. Unit tests which construct and advance a twisted.internet.task.Clock are fast and deterministic.

The Global Reactor

Since unit tests are avoiding real I/O and real time, they can usually avoid using a real reactor. The only exceptions to this are unit tests for a real reactor implementation. Unit tests for protocol implementations or other application code should not use a reactor. Unit tests for real reactor implementations should not use the global reactor, but should instead use twisted.internet.test.reactormixins.ReactorBuilder so they can be applied to all of the reactor implementations automatically. In no case should new unit tests use the global reactor.

Skipping tests, TODO items

Trial, the Twisted unit test framework, has some extensions which are designed to encourage developers to add new tests. One common situation is that a test exercises some optional functionality: maybe it depends upon certain external libraries being available, maybe it only works on certain operating systems. The important common factor is that nobody considers these limitations to be a bug.

To make it easy to test as much as possible, some tests may be skipped in certain situations. Individual test cases can raise the SkipTest exception to indicate that they should be skipped, and the remainder of the test is not run. In the summary (the very last thing printed, at the bottom of the test output) the test is counted as a skip instead of a success or fail. This should be used inside a conditional which looks for the necessary prerequisites:

class SSHClientTests(unittest.TestCase):
    def test_sshClient(self):
        if not ssh_path:
            raise unittest.SkipTest("cannot find ssh, nothing to test")
        foo() # do actual test after the SkipTest

You can also set the .skip attribute on the method, with a string to indicate why the test is being skipped. This is convenient for temporarily turning off a test case, but it can also be set conditionally (by manipulating the class attributes after they've been defined):

class SomeThingTests(unittest.TestCase):
    def test_thing(self):
        dotest()
    test_thing.skip = "disabled locally"
class MyTestCase(unittest.TestCase):
    def test_one(self):
        ...
    def test_thing(self):
        dotest()

if not haveThing:
    MyTestCase.test_thing.im_func.skip = "cannot test without Thing"
    # but test_one() will still run

Finally, you can turn off an entire TestCase at once by setting the .skip attribute on the class. If you organize your tests by the functionality they depend upon, this is a convenient way to disable just the tests which cannot be run.

class TCPTestCase(unittest.TestCase):
    ...
class SSLTestCase(unittest.TestCase):
    if not haveSSL:
        skip = "cannot test without SSL support"
    # but TCPTestCase will still run
    ...

.todo and Testing New Functionality

Two good practices which arise from the XP development process are sometimes at odds with each other:

These two goals will sometimes conflict. The unit tests that are written first, before any implementation has been done, are certain to fail. We want developers to commit their code frequently, for reliability and to improve coordination between multiple people working on the same problem together. While the code is being written, other developers (those not involved in the new feature) should not have to pay attention to failures in the new code. We should not dilute our well-indoctrinated Failing Test Horror Syndrome by crying wolf when an incomplete module has not yet started passing its unit tests. To do so would either teach the module author to put off writing or committing their unit tests until after all the functionality is working, or it would teach the other developers to ignore failing test cases. Both are bad things.

.todo is intended to solve this problem. When a developer first starts writing the unit tests for functionality that has not yet been implemented, they can set the .todo attribute on the test methods that are expected to fail. These methods will still be run, but their failure will not be counted the same as normal failures: they will go into an expected failures category. Developers should learn to treat this category as a second-priority queue, behind actual test failures.

As the developer implements the feature, the tests will eventually start passing. This is surprising: after all those tests are marked as being expected to fail. The .todo tests which nevertheless pass are put into a unexpected success category. The developer should remove the .todo tag from these tests. At that point, they become normal tests, and their failure is once again cause for immediate action by the entire development team.

The life cycle of a test is thus:

  1. Test is created, marked .todo. Test fails: expected failure.
  2. Code is written, test starts to pass. unexpected success.
  3. .todo tag is removed. Test passes. success.
  4. Code is broken, test starts to fail. failure. Developers spring into action.
  5. Code is fixed, test passes once more. success.

Any test which remains marked with .todo for too long should be examined. Either it represents functionality which nobody is working on, or the test is broken in some fashion and needs to be fixed. Generally, .todo may be of use while you are developing a feature, but by the time you are ready to commit anything, all the tests you have written should be passing. In other words, you should rarely, if ever, feel the need to add a test marked todo to trunk. When you do, consider whether a ticket in the issue tracker would be more useful.

Line Coverage Information

Trial provides line coverage information, which is very useful to ensure old code has decent coverage. Passing the --coverage option to to Trial will generate the coverage information in a file called coverage which can be found in the _trial_temp folder. This option requires Python 2.3.3 or newer.

Associating Test Cases With Source Files

Please add a test-case-name tag to the source file that is covered by your new test. This is a comment at the beginning of the file which looks like one of the following:

# -*- test-case-name: twisted.test.test_defer -*-

or

#!/usr/bin/env python
# -*- test-case-name: twisted.test.test_defer -*-

This format is understood by emacs to mark File Variables. The intention is to accept test-case-name anywhere emacs would on the first or second line of the file (but not in the File Variables: block that emacs accepts at the end of the file). If you need to define other emacs file variables, you can either put them in the File Variables: block or use a semicolon-separated list of variable definitions:

# -*- test-case-name: twisted.test.test_defer; fill-column: 75; -*-

If the code is exercised by multiple test cases, those may be marked by using a comma-separated list of tests, as follows: (NOTE: not all tools can handle this yet.. trial --testmodule does, though)

# -*- test-case-name: twisted.test.test_defer,twisted.test.test_tcp -*-

The test-case-name tag will allow trial --testmodule twisted/dir/myfile.py to determine which test cases need to be run to exercise the code in myfile.py. Several tools (as well as 's twisted-dev.el's F9 command) use this to automatically run the right tests.

See also Tips for writing tests for Twisted code.