Tests style guide

This page documents conventions for our Launchpad tests. Reviewers make sure that code merged is documented and covered by tests. Following the principles outlined in this document will minimize comments related to test style from reviewers.

Reviewers will block merge of code that is under-documented or under-tested. We have two primary means of documentation:

  1. System documentation in doc directories such as lib/lp/<app>/doc.

  2. Page tests in lib/lp/<app>/stories.

  3. View tests in lib/lp/<app>/browser/tests.

While these two types of documentation use the doctest format, which means that they contain testable examples, they are documentation first. So they are not the best place to test many corner cases or various similar possibilities. This is best done in other unit tests or functional tests, which have ensuring complete test coverage as their main objective.

Testable Documentation

Testable documentation includes system documentation doctests and page tests.

What do we mean by “testable documentation”? One rule of thumb is the narrative should be similar to what you’d see in an API reference manual. Another trick to ensure the narrative is appropriately detailed is to mentally remove all of the code snippets and see if the narrative stands by itself.

System Documentation

These are doctests located under lib/lp/<app>/doc. They are used to document the APIs and other internal objects. The documentation should explain to a developer how to use these objects and what purpose they serve.

Each modification to lp.<app>.interfaces should be documented in one of these files.

(Each file in that directory is automatically added to the test suite. If you need to configure the test layer in which the test will be run or need to customize the test fixture, you can add special instructions for the file in the system documentation harness in lib/lp/<app>/tests/test_doc.py.)

Use Cases Documentation: Page Tests

We use page tests to document all the use cases that Launchpad satisfies. The narrative in these files should document the use case. That is, they should explain what the user’s objective is and how he accomplishes it.

The examples in these files uses zope.testbrowser to show how the user would navigate the workflow relevant to the use case.

So each addition to the UI should be covered by an appropriate section in a page test.

The page tests do not need to document and demonstrate each and every possible way to navigate the workflow. This can usually be done in a more direct manner by testing the view object directly.

Browser View Tests

View objects are usually documented that way along other system objects in files named *-pages.txt or in lib/lp/<app>/browser/tests/*-views.txt.

The browser tests directory contains both doctest files for documenting the use of browser view classes and unit tests (e.g. test_*.py) for performing unit tests, including covering corner cases. Currently some apps, registry for example, have a large number of doctests in this location that are not strictly testable documentation. Over time these non-documentation doctest files should be converted to unit tests.

All new browser tests that are not testable documentation should be written as unit tests.

Common Conventions

The basic conventions for testable documentation are:

  • Example code is wrapped at 78 columns, follows the regular Python style guide, and is indented 4 spaces.

  • Narrative text may be wrapped at either 72 or 78 columns.

  • You can use regular Python comments for explanations related to the code and not to the documentation.

  • Doctests use Restructured Text (or “ReST”, see http://docutils.sourceforge.net/docs/user/rst/quickref.html).

    • We use ReST because it’s what the Python community have standardized on and because it makes setting up Sphinx to browse all the doctests simpler.

  • The file should have a first-level title element. An expansion of the filename is usually a good start. For example, the file bugcomment.txt could have this title:

Bug Comments
============
  • Two blank lines are used to separate the start of a new section (a header).

An Example
----------

Launchpad tracks foo and bar elements using the IFooBarSet utility.

    >>> from lp.foo.interfaces.bar import IBar, IFoo, IFooBarSet
    >>> from lp.testing import verifyObject
    >>> foobarset = getUtility(IFooBarSet)

    >>> verifyObject(IFooBarSet, foobarset)
    True

You use the getFoo() method to obtain an IFoo instance by id:

    >>> foo = foobarset.getFoo('aFoo')
    >>> verifyObject(IFoo, foo)
    True

Similarly, you use the getBar() method to retrieve an IBar instance by
id:

    >>> bar = foobarset.getBar('aBar')
    >>> verifyObject(IBar, bar)
    True

Each individual test should be of the form:

>>> do_something()
expected output

This means that something like this isn’t considered a test, but test setup (since it doesn’t produce any output):

>>> do_something()

For the reason above, the assert statement shouldn’t be used in doctests.

Comparing Results

When writing doctests, make sure that if the test fails, the failure message will be helpful to debug the problem. Avoid constructs like:

>>> 'Test' in foo.getText()
True

The failure message for this test will be:

- True
+ False

which isn’t helpful at all in understanding what went wrong. This example is a lot more helpful when it fails:

>>> foo.getText()
'...Test...'

For page tests where the page contains a lot of elements, you should zoom in to the relevant part. You can use the find_main_content(), find_tags_by_class(), find_tag_by_id(), and find_portlet() helper methods. They return BeautifulSoup instances, which makes it easy to access specific elements in the tree.

The new status is displayed in the portlet.

    >>> details_portlet = find_portlet(browser.contents, 'Question details')
    >>> print(details_portlet.find('b', text='Status:').next.strip())
    Needs information

There is also an extract_text() helper that only renders the HTML text:

>>> print(extract_text(
...     find_tag_by_id(browser.contents, 'branchtable')))
main         60 New           firefox
klingon      30 Experimental  gnome-terminal
junk.contrib 60 New 2005-10-31 12:03:57 ... weeks ago

Read Page tests for other tips on writing page tests.

When to print and when to return values

Doctests mimic the Python interactive interpreter, so generally it’s preferred to simply return values and expect to see their string representation. In a few cases though, it’s better to print the results instead of just returning them.

The two most common cases of this are None and strings. The interactive interpreter suppresses None return values, so relying on these means the doctest makes less sense. You could compare against None, but the True or False output isn’t explicit, so it’s almost always better to print values you expect to be None.

Instead of:

>>> should_be_none()
>>> do_something_else()

Use:

>>> print(should_be_none())
None
>>> do_something_else()

For a different reason, it’s also usually better to print string results rather than just returning them. Returning the string causes the quotes to be included in the output, while printing the string does not. Those extra quotes are usually noise.

Instead of:

>>> get_some_text()
'foo'
>>> get_some_string()
"Don't care"

Use:

>>> print(get_some_text())
foo
>>> print(get_some_string())
Don't care

Dictionaries and sets

You can’t just print the value of a dictionary or a set when that collection has more than one element in it, e.g.

>>> print(my_dict)
{'a': 1, 'b': 2}

The reason is that Python does not guarantee the order of its elements in a dictionary or set, so the printed representation of a dictionary is indeterminate. In page tests, there’s a pretty() global which is basically exposing Python’s pretty printer, and you can use it safely:

>>> pretty(my_dict)
{'a': 1, 'b': 2}

Though it’s a bit uglier, you can also print the sorted items of a dictionary:

>>> sorted(my_dict.items())
[('a', 1), ('b', 2)]

Global State

Be especially careful of test code that changes global state. For example, we were recently bitten by code in a test that did this:

socket.setdefaulttimeout(1)

While that may be necessary for the specific test, it’s important to understand that this code changes global state and thus can adversely affect all of our other tests. In fact, this code caused intermittent and very difficult-to-debug failures that mucked up buildbot for many unrelated branches.

The guideline then is this: if code changes global state (for example, by monkey-patching a module’s globals) then the test must be sure to restore the previous state, either in a try-finally clause, or at the end of the doctest, or in the test’s tearDown hook.

Style to Avoid

A very important consideration is that documentation tests are really documentation that happens to be testable. So, the writing style should be appropriate for documentation. It should be affirmative and descriptive. There shouldn’t be any phrases like:

  • Test that…

  • Check that…

  • Verify that…

  • This test…

While these constructs may help the reader understand what is happening, they only have indirect value as documentation. They can usually be replaced by simply stating what the result is.

For example:

Test that the bar was added to the foo's related_bars:

    >>> bar in foo.related_bars
    True

Can be replaced by:

After being linked, the bar is available in the foo's
related_bars attribute:

    >>> bar in foo.related_bars
    True

Also, use of “should” or “will” can usually be replaced by the present tense to make the style affirmative.

For example:

The bar not_a_foo attribute should now be set:

    >>> bar.not_a_foo
    True

Can be replaced by:

The bar not_a_foo attribute is set after this operation:

    >>> bar.not_a_foo
    True

A good rule of thumb to know whether the narrative style works as documentation is to read the narrative as if the code examples were not there. If the text style makes sense, the style is probably good.

Using Sample Data

If possible, avoid using the existing sample data in tests, apart from some basic objects, like users. Sample data is good for demonstrating the UI, but it can make tests harder to understand, since it requires knowledge of the properties of the used sample data. Using sample data in tests also makes it harder to modify the data.

If you do use sample data in the test, assert your expectations to avoid subtle errors if someone modifies it. For example:

Anonymous users can't see a private bug's description.

    >>> private_bug = getUtility(IBugSet).get(5)
    >>> private_bug.private
    True

    >>> login(ANONYMOUS)
    >>> private_bug.description
    Traceback (most recent call last):
    ...
    Unauthorized:...

When using fake domains and especially fake email addresses, wherever possible use the example.{com,org,net} domains, e.g. aperson@example.com. These are guaranteed by Internet standard never to exist, so it can’t be possible to accidentally spam them if something goes wrong on our end.

Fixtures and Helpers

Sometimes a lot of code is needed to set up a test, or to extract the relevant information in the examples. It is usually a good idea to factor this code into functions that can be documented in the file itself (when the function will only be used in that file), or even better, moved into a test helper module from which you can import.

These helpers currently live in lib/lp/testing. New helpers should go there, unless they’re very specific to a particular corner of the application; in such cases you can use something like lp.foo.testing.

Functional and Unit Tests

Complete test coverage without impairing documentation often requires dedicated functional or unit tests. In Launchpad, Python test cases are used for these types of tests. You may encounter legacy code that uses doctests for functional testing. If you do, please consider converting it to a Python test case.

Functional tests are found in the tests subdirectory of each directory containing code under test.

Python Test Cases

Although Python test cases are not documentation they must still be human-readable. So:

  • Keep the test short and concise.

  • Stick to the “setup, exercise, assert” pattern, especially avoid “exercise, assert, exercise some more, assert”.

  • Put into the docstring of each test case what is being tested. As a special case for test methods, a comment block at the beginning of the method is considered an acceptable substitute to a docstring. Please observe “Style to avoid”, as explained above.

  • Organize related test cases in the same class. Explain test objectives in the class docstring.

  • When asserting for equality use the form assertEqual(expected_results, actual_results, "...") (the third argument is optional, for use if failure messages would otherwise be particularly unclear).

  • Make sure that each assert fails with an appropriate error message explaining what is expected. lp.testing.TestCase and TestCaseWithFactory are derived from testtools.TestCase and therefore produce good error messages. Only some cases may warrant an explicit error message. For example, this:

self.assertTrue('aString' in result)

could be replaced by:

self.assertIn('aString', result)
  • Consider using testtools matchers where reasonable. These can often improve failure messages so that they show more information in one go, which can be useful when debugging mysterious failures. For example, instead of this:

self.assertEqual('expected', obj.foo)
self.assertEqual('more expected', obj.bar)

prefer this:

self.assertThat(obj, MatchesStructure.byEquality(
    foo='expected',
    bar='more expected'))

In general, you should follow Launchpad coding conventions (see Python style guide), however when naming test methods:

  • Use PEP 8 names, e.g. test_for_my_feature()

  • When testing a specific Launchpad method, a mix of PEP 8 and camel case is used, e.g. test_fooBarBaz()

  • When testing alternatives for a LP method, use this style: test_fooBarBaz_with_first_alternative(), test_fooBarBaz_with_second_alternative(), etc.

How To Use the Correct Test URL

When tests run, and need to connect to the application server instance under test, you need to ensure that a URL with the correct port for that test instance is used. Here’s how to do that.

The config instance has an API which allows the correct URL to be determined. The API is defined in LaunchpadConfig and as a convenience is available as a class method on BaseLayer.

def appserver_root_url(self, facet='mainsite', ensureSlash=False):
    """Return the correct app server root url for the given facet."""

Code snippets for a number of scenarios are as follows.

Doc Tests

>>> from lp.testing.layers import BaseLayer
>>> root_url = BaseLayer.appserver_root_url()
>>> browser.open(root_url)

Unit Tests

class TestOpenIDReplayAttack(TestCaseWithFactory):
    layer = AppServerLayer

    def test_replay_attacks_do_not_succeed(self):
        browser = Browser(mech_browser=MyMechanizeBrowser())
        browser.open('%s/+login' % self.layer.appserver_root_url())