Buildbot's Test Suite¶
Buildbot's tests are under buildbot.test and, for the buildslave, buildslave.test. Tests for the slave are similar to the master, although in some cases helpful functionality on the master is not re-implemented on the slave.
Tests are divided into a few suites:
- Unit tests (buildbot.test.unit) - these follow unit-testing practices and attempt to maximally isolate the system under test. Unit tests are the main mechanism of achieving test coverage, and all new code should be well-covered by corresponding unit tests.
- Interface tests (buildbot.test.interface). In many cases, Buildbot has multiple implementations of the same interface -- at least one "real" implementation and a fake implementation used in unit testing. The interface tests ensure that these implementations all meet the same standards. This ensures consistency between implementations, and also ensures that the unit tests are testing against realistic fakes.
- Integration tests (buildbot.test.integration) - these test combinations of multiple units. Of necessity, integration tests are incomplete - they cannot test every condition; difficult to maintain - they tend to be complex and touch a lot of code; and slow - they usually require considerable setup and execute a lot of code. As such, use of integration tests is limited to a few, broad tests to act as a failsafe for the unit and interface tests.
- Regression tests (buildbot.test.regrssions) - these test to prevent re-occurrence of historical bugs. In most cases, a regression is better tested by a test in the other suites, or unlike to recur, so this suite tends to be small.
- Fuzz tests (buildbot.test.fuzz) - these tests run for a long time and apply randomization to try to reproduce rare or unusual failures. The Buildbot project does not currently have a framework to run fuzz tests regularly.
Every code module should have corresponding unit tests. This is not currently true of Buildbot, due to a large body of legacy code, but is a goal of the project. All new code must meet this requirement.
Unit test modules are be named after the package or class they test, replacing . with _ and omitting the buildbot_. For example, test_status_web_authz_Authz.py tests the Authz class in buildbot/status/web/authz.py. Modules with only one class, or a few trivial classes, can be tested in a single test module. For more complex situations, prefer to use multiple test modules.
Interface tests exist to verify that multiple implementations of an interface meet the same requirements. Note that the name 'interface' should not be confused with the sparse use of Zope Interfaces in the Buildbot code -- in this context, an interface is any boundary between testable units.
Ideally, all interfaces, both public and private, should be tested. Certainly, any public interfaces need interface tests.
Interface test modules are named after the interface they are testing, e.g., test_mq.py. They generally begin as follows:
from buildbot.test.util import interfaces from twistd.trial import unittest class Tests(interfaces.InterfaceTests): # define methods that must be overridden per implementation def someSetupMethod(self): raise NotImplementedError # tests that all implementations must pass def test_signature_someMethod(self): @self.assertArgSpecMatches(self.systemUnderTest.someMethod) def someMethod(self, arg1, arg2): pass def test_something(self): pass # ... class RealTests(Tests): # tests that all *real* implementations must pass def test_something_else(self): pass # ...
All of the test methods are defined here, segregated into tests that all implementations must pass, and tests that the fake implementation is not expected to pass. The test_signature_someMethod test above illustrates the assertArgSpecMatches decorator, which can be used to compare the argument specification of a callable with a reference implementation conveniently written as a nested function.
At the bottom of the test module, a subclass is created for each implementation, implementing the setup methods that were stubbed out in the parent classes:
class TestFakeThing(unittest.TestCase, Tests): def someSetupMethod(self): pass # ... class TestRealThing(unittest.TestCase, RealTests): def someSetupMethod(self): pass # ...
For implementations which require optional software, this is the appropriate place to signal that tests should be skipped when their prerequisites are not available.
Integration test modules test several units at once, including their interactions. In general, they serve as a catch-all for failures and bugs that were not detected by the unit and interface tests. As such, they should not aim to be exhaustive, but merely representative.
Integration tests are very difficult to maintain if they reach into the internals of any part of Buildbot. Where possible, try to use the same means as a user would to set up, run, and check the results of an integration test. That may mean writing a master.cfg to be parsed, and checking the results by examining the database (or fake DB API) afterward.
Regression tests are even more rare in Buildbot than integration tests. In many cases, a regression test is not necessary -- either the test is better-suited as a unit or interface test, or the failure is so specific that a test will never fail again.
Regression tests tend to be closely tied to the code in which the error occurred. When that code is refactored, the regression test generally becomes obsolete, and is deleted.
Fuzz tests generally run for a fixed amount of time, running randomized tests against a system. They do not run at all during normal runs of the Buildbot tests, unless BUILDBOT_FUZZ is defined. This is accomplished with something like the following at the end of each test module:
if 'BUILDBOT_FUZZ' not in os.environ: del LRUCacheFuzzer
Buildbot provides a number of purpose-specific mixin classes in master/buildbot/util. These generally define a set of utility functions as well as setUpXxx and tearDownXxx methods. These methods should be called explicitly from your subclass's setUp and tearDown methods. Note that some of these methods return Deferreds, which should be handled properly by the caller.
Buildbot provides a number of pre-defined fake implementations of internal interfaces, in master/buildbot/fake. These are designed to be used in unit tests to limit the scope of the test. For example, the fake DB API eliminates the need to create a real database when testing code that uses the DB API, and isolates bugs in the system under test from bugs in the real DB implementation.
The danger of using fakes is that the fake interface and the real interface can differ. The interface tests exist to solve this problem. All fakes should be fully tested in an integration test, so that the fakes pass the same tests as the "real" thing. It is particularly important that the method signatures be compared.
Bad tests are worse than no tests at all, since they waste developers' time wondering "was that a spurious failure?" or "what the heck is this test trying to do?" Buildbot needs good tests. So what makes a good test?
Independent of Time¶
Tests that depend on wall time will fail. As a bonus, they run very slowly. Do not use reactor.callLater to wait "long enough" for something to happen.
For testing things that themselves depend on time, consider using twisted.internet.tasks.Clock. This may mean passing a clock instance to the code under test, and propagating that instance as necessary to ensure that all of the code using callLater uses it. Refactoring code for testability is difficult, but worthwhile.
For testing things that do not depend on time, but for which you cannot detect the "end" of an operation: add a way to detect the end of the operation!
Make your tests readable. This is no place to skimp on comments! Others will attempt to learn about the expected behavior of your class by reading the tests. As a side note, if you use a Deferred chain in your test, write the callbacks as nested functions, rather than using methods with funny names:
def testSomething(self): d = doThisFirst() def andThisNext(res): pass # ... d.addCallback(andThisNext) return d
This isolates the entire test into one indented block. It is OK to add methods for common functionality, but give them real names and explain in detail what they do.
Test method names should follow the pattern test_METHOD_CONDITION where METHOD is the method being tested, and CONDITION is the condition under which it's tested. Since we can't always test a single method, this is not a hard-and-fast rule.
Assert Only One Thing¶
Where practical, each test should have a single assertion. This may require a little bit of work to get several related pieces of information into a single Python object for comparison. The problem with multiple assertions is that, if the first assertion fails, the remainder are not tested. The test results then do not tell the entire story.
Prefer Fakes to Mocks¶
Mock objects are too "compliant", and this often masks errors in the system under test. For example, a mis-spelled method name on a mock object will not raise an exception.
Where possible, use one of the pre-written fake objects (see Fakes) instead of a mock object. Fakes themselves should be well-tested using interface tests.
Where they are appropriate, Mock objects can be constructed easily using the aptly-named mock module, which is a requirement for Buildbot's tests.
The shorter each test is, the better. Test as little code as possible in each test.
It is fine, and in fact encouraged, to write the code under test in such a way as to facilitate this. As an illustrative example, if you are testing a new Step subclass, but your tests require instantiating a BuildMaster, you're probably doing something wrong!
This also applies to test modules. Several short, easily-digested test modules are preferred over a 1000-line monster.
Each test should be maximally independent of other tests. Do not leave files laying around after your test has finished, and do not assume that some other test has run beforehand. It's fine to use caching techniques to avoid repeated, lengthy setup times.
Tests should be as robust as possible, which at a basic level means using the available frameworks correctly. All Deferreds should have callbacks and be chained properly. Error conditions should be checked properly. Race conditions should not exist (see Independent of Time, above).
Note that tests will pass most of the time, but the moment when they are most useful is when they fail.
When the test fails, it should produce output that is helpful to the person chasing it down. This is particularly important when the tests are run remotely, in which case the person chasing down the bug does not have access to the system on which the test fails. A test which fails sporadically with no more information than "AssertionFailed" is a prime candidate for deletion if the error isn't obvious. Making the error obvious also includes adding comments describing the ways a test might fail.
Python does not allow assignment to anything but the innermost local scope or the global scope with the global keyword. This presents a problem when creating nested functions:
def test_localVariable(self): cb_called = False def cb(): cb_called = True cb() self.assertTrue(cb_called) # will fail!
The cb_called = True assigns to a different variable than cb_called = False. In production code, it's usually best to work around such problems, but in tests this is often the clearest way to express the behavior under test.
The solution is to change something in a common mutable object. While a simple list can serve as such a mutable object, this leads to code that is hard to read. Instead, use State:
from buildbot.test.state import State def test_localVariable(self): state = State(cb_called=False) def cb(): state.cb_called = True cb() self.assertTrue(state.cb_called) # passes
This is almost as readable as the first example, but it actually works.