Writing Python Regression Tests
-------------------------------
Skip Montanaro
(skip@mojam.com)
Introduction
If you add a new module to Python or modify the functionality of an existing
module, you should write one or more test cases to exercise that new
functionality. There are different ways to do this within the regression
testing facility provided with Python; any particular test should use only
one of these options. Each option requires writing a test module using the
conventions of the the selected option:
- PyUnit based tests
- doctest based tests
- "traditional" Python test modules
Regardless of the mechanics of the testing approach you choose,
you will be writing unit tests (isolated tests of functions and objects
defined by the module) using white box techniques. Unlike black box
testing, where you only have the external interfaces to guide your test case
writing, in white box testing you can see the code being tested and tailor
your test cases to exercise it more completely. In particular, you will be
able to refer to the C and Python code in the CVS repository when writing
your regression test cases.
PyUnit based tests
The PyUnit framework is based on the ideas of unit testing as espoused
by Kent Beck and the Extreme Programming (XP) movement. The specific
interface provided by the framework is tightly based on the JUnit
Java implementation of Beck's original SmallTalk test framework. Please
see the documentation of the unittest module for detailed information on
the interface and general guidelines on writing PyUnit based tests.
The test_support helper module provides a single function for use by
PyUnit based tests in the Python regression testing framework:
run_unittest() takes a unittest.TestCase derived class as a parameter
and runs the tests defined in that class. All test methods in the
Python regression framework have names that start with "test_" and use
lower-case names with words separated with underscores.
doctest based tests
Tests written to use doctest are actually part of the docstrings for
the module being tested. Each test is written as a display of an
interactive session, including the Python prompts, statements that would
be typed by the user, and the output of those statements (including
tracebacks, although only the exception msg needs to be retained then).
The module in the test package is simply a wrapper that causes doctest
to run over the tests in the module. The test for the difflib module
provides a convenient example:
import difflib, test_support
test_support.run_doctest(difflib)
If the test is successful, nothing is written to stdout (so you should not
create a corresponding output/test_difflib file), but running regrtest
with -v will give a detailed report, the same as if passing -v to doctest.
A second argument can be passed to run_doctest to tell doctest to search
sys.argv for -v instead of using test_support's idea of verbosity. This
is useful for writing doctest-based tests that aren't simply running a
doctest'ed Lib module, but contain the doctests themselves. Then at
times you may want to run such a test directly as a doctest, independent
of the regrtest framework. The tail end of test_descrtut.py is a good
example:
def test_main(verbose=None):
import test_support, test.test_descrtut
test_support.run_doctest(test.test_descrtut, verbose)
if __name__ == "__main__":
test_main(1)
If run via regrtest, test_main() is called (by regrtest) without specifying
verbose, and then test_support's idea of verbosity is used. But when
run directly, test_main(1) is called, and then doctest's idea of verbosity
is used.
See the documentation for the doctest module for information on
writing tests using the doctest framework.
"traditional" Python test modules
The mechanics of how the "traditional" test system operates are fairly
straightforward. When a test case is run, the output is compared with the
expected output that is stored in .../Lib/test/output. If the test runs to
completion and the actual and expected outputs match, the test succeeds, if
not, it fails. If an ImportError or test_support.TestSkipped error is
raised, the test is not run.
Executing Test Cases
If you are writing test cases for module spam, you need to create a file
in .../Lib/test named test_spam.py. In addition, if the tests are expected
to write to stdout during a successful run, you also need to create an
expected output file in .../Lib/test/output named test_spam ("..."
represents the top-level directory in the Python source tree, the directory
containing the configure script). If needed, generate the initial version
of the test output file by executing:
./python Lib/test/regrtest.py -g test_spam.py
from the top-level directory.
Any time you modify test_spam.py you need to generate a new expected
output file. Don't forget to desk check the generated output to make sure
it's really what you expected to find! All in all it's usually better
not to have an expected-out file (note that doctest- and unittest-based
tests do not).
To run a single test after modifying a module, simply run regrtest.py
without the -g flag:
./python Lib/test/regrtest.py test_spam.py
While debugging a regression test, you can of course execute it
independently of the regression testing framework and see what it prints:
./python Lib/test/test_spam.py
To run the entire test suite:
[UNIX, + other platforms where "make" works] Make the "test" target at the
top level:
make test
{WINDOWS] Run rt.bat from your PCBuild directory. Read the comments at
the top of rt.bat for the use of special -d, -O and -q options processed
by rt.bat.
[OTHER] You can simply execute the two runs of regrtest (optimized and
non-optimized) directly:
./python Lib/test/regrtest.py
./python -O Lib/test/regrtest.py
But note that this way picks up whatever .pyc and .pyo files happen to be
around. The makefile and rt.bat ways run the tests twice, the first time
removing all .pyc and .pyo files from the subtree rooted at Lib/.
Test cases generate output based upon values computed by the test code.
When executed, regrtest.py compares the actual output generated by executing
the test case with the expected output and reports success or failure. It
stands to reason that if the actual and expected outputs are to match, they
must not contain any machine dependencies. This means your test cases
should not print out absolute machine addresses (e.g. the return value of
the id() builtin function) or floating point numbers with large numbers of
significant digits (unless you understand what you are doing!).
Test Case Writing Tips
Writing good test cases is a skilled task and is too complex to discuss in
detail in this short document. Many books have been written on the subject.
I'll show my age by suggesting that Glenford Myers' "The Art of Software
Testing", published in 1979, is still the best introduction to the subject
available. It is short (177 pages), easy to read, and discusses the major
elements of software testing, though its publication predates the
object-oriented software revolution, so doesn't cover that subject at all.
Unfortunately, it is very expensive (about $100 new). If you can borrow it
or find it used (around $20), I strongly urge you to pick up a copy.
The most important goal when writing test cases is to break things. A test
case that doesn't uncover a bug is much less valuable than one that does.
In designing test cases you should pay attention to the following:
* Your test cases should exercise all the functions and objects defined
in the module, not just the ones meant to be called by users of your
module. This may require you to write test code that uses the module
in ways you don't expect (explicitly calling internal functions, for
example - see test_atexit.py).
* You should consider any boundary values that may tickle exceptional
conditions (e.g. if you were writing regression tests for division,
you might well want to generate tests with numerators and denominators
at the limits of floating point and integer numbers on the machine
performing the tests as well as a denominator of zero).
* You should exercise as many paths through the code as possible. This
may not always be possible, but is a goal to strive for. In
particular, when considering if statements (or their equivalent), you
want to create test cases that exercise both the true and false
branches. For loops, you should create test cases that exercise the
loop zero, one and multiple times.
* You should test with obviously invalid input. If you know that a
function requires an integer input, try calling it with other types of
objects to see how it responds.
* You should test with obviously out-of-range input. If the domain of a
function is only defined for positive integers, try calling it with a
negative integer.
* If you are going to fix a bug that wasn't uncovered by an existing
test, try to write a test case that exposes the bug (preferably before
fixing it).
* If you need to create a temporary file, you can use the filename in
test_support.TESTFN to do so. It is important to remove the file
when done; other tests should be able to use the name without cleaning
up after your test.
Regression Test Writing Rules
Each test case is different. There is no "standard" form for a Python
regression test case, though there are some general rules (note that
these mostly apply only to the "classic" tests; unittest- and doctest-
based tests should follow the conventions natural to those frameworks):
* If your test case detects a failure, raise TestFailed (found in
test_support).
* Import everything you'll need as early as possible.
* If you'll be importing objects from a module that is at least
partially platform-dependent, only import those objects you need for
the current test case to avoid spurious ImportError exceptions that
prevent the test from running to completion.
* Print all your test case results using the print statement. For
non-fatal errors, print an error message (or omit a successful
completion print) to indicate the failure, but proceed instead of
raising TestFailed.
* Use "assert" sparingly, if at all. It's usually better to just print
what you got, and rely on regrtest's got-vs-expected comparison to
catch deviations from what you expect. assert statements aren't
executed at all when regrtest is run in -O mode; and, because they
cause the test to stop immediately, can lead to a long & tedious
test-fix, test-fix, test-fix, ... cycle when things are badly broken
(and note that "badly broken" often includes running the test suite
for the first time on new platforms or under new implementations of
the language).
Miscellaneous
There is a test_support module you can import from your test case. It
provides the following useful objects:
* TestFailed - raise this exception when your regression test detects a
failure.
* TestSkipped - raise this if the test could not be run because the
platform doesn't offer all the required facilities (like large
file support), even if all the required modules are available.
* verbose - you can use this variable to control print output. Many
modules use it. Search for "verbose" in the test_*.py files to see
lots of examples.
* verify(condition, reason='test failed'). Use this instead of
assert condition[, reason]
verify() has two advantages over assert: it works even in -O mode,
and it raises TestFailed on failure instead of AssertionError.
* TESTFN - a string that should always be used as the filename when you
need to create a temp file. Also use try/finally to ensure that your
temp files are deleted before your test completes. Note that you
cannot unlink an open file on all operating systems, so also be sure
to close temp files before trying to unlink them.
* sortdict(dict) - acts like repr(dict.items()), but sorts the items
first. This is important when printing a dict value, because the
order of items produced by dict.items() is not defined by the
language.
* findfile(file) - you can call this function to locate a file somewhere
along sys.path or in the Lib/test tree - see test_linuxaudiodev.py for
an example of its use.
* use_large_resources - true iff tests requiring large time or space
should be run.
* fcmp(x,y) - you can call this function to compare two floating point
numbers when you expect them to only be approximately equal withing a
fuzz factor (test_support.FUZZ, which defaults to 1e-6).
NOTE: Always import something from test_support like so:
from test_support import verbose
or like so:
import test_support
... use test_support.verbose in the code ...
Never import anything from test_support like this:
from test.test_support import verbose
"test" is a package already, so can refer to modules it contains without
"test." qualification. If you do an explicit "test.xxx" qualification, that
can fool Python into believing test.xxx is a module distinct from the xxx
in the current package, and you can end up importing two distinct copies of
xxx. This is especially bad if xxx=test_support, as regrtest.py can (and
routinely does) overwrite its "verbose" and "use_large_resources"
attributes: if you get a second copy of test_support loaded, it may not
have the same values for those as regrtest intended.
Python and C statement coverage results are currently available at
http://www.musi-cal.com/~skip/python/Python/dist/src/
As of this writing (July, 2000) these results are being generated nightly.
You can refer to the summaries and the test coverage output files to see
where coverage is adequate or lacking and write test cases to beef up the
coverage.
Some Non-Obvious regrtest Features
* Automagic test detection: When you create a new test file
test_spam.py, you do not need to modify regrtest (or anything else)
to advertise its existence. regrtest searches for and runs all
modules in the test directory with names of the form test_xxx.py.
* Miranda output: If, when running test_spam.py, regrtest does not
find an expected-output file test/output/test_spam, regrtest
pretends that it did find one, containing the single line
test_spam
This allows new tests that don't expect to print anything to stdout
to not bother creating expected-output files.
* Two-stage testing: To run test_spam.py, regrtest imports test_spam
as a module. Most tests run to completion as a side-effect of
getting imported. After importing test_spam, regrtest also executes
test_spam.test_main(), if test_spam has a "test_main" attribute.
This is rarely needed, and you shouldn't create a module global
with name test_main unless you're specifically exploiting this
gimmick. In such cases, please put a comment saying so near your
def test_main, because this feature is so rarely used it's not
obvious when reading the test code.
cc91ac09ef