Selftest

1. Introduction

A Simple Pythonic Self Test Tool

Selftest is a simple and extensible test tool for Python. Selftest was first known as Autotest where auto means self. The key features are:

• tests are ordinary functions, anywhere in the application code,
• gathering and running tests is automatic and follows the structure of your code,
• testing stops on first failure with a standard Python stack trace,

The benefit of this approach are:

• there is no separate test tree to maintain,
• dependencies are automatically tested,
• quicker cycles, more focus.

The core is extremely small and support for async, filters, operators, fixtures, diffs, wildcards, guards, etc are all implemented in hooks.

In-source testing

Selttest runs tests within the source code alongside the implementation, similar to Rust's module tests and Vitest in-source testing.

This makes the tests share closure with the implementation and enables it to test against private state without exporting. Meanwhile, it also brings a closer feedback loop for development.

History

Selftest began, as a recalcitrant move away from the frameworks, with the following decorator above my tests:

def test(f):                            # the test decorator
    print(f.__qualname__)
    f()

def i_need_testing(x):
    return x

@test                                   # marks and runs the test
def a_test():
    assert 42 == i_need_testing(42)

It would just run the test every time I imported it. That turned out to work so well that it grew out to what we have here today.

Features

Meanwhile selftest gained some features. It

1. works in a familiar Pythonic way, no magic,
2. has one simple API via the test object,
3. is based on standard modules operator, pdb, logger, difflib, inspect, etc,
4. seamlessly scales from microtests to systemtests,
5. discovers tests automatically through the importing mechanism,
6. crosses module, package and project boundaries easily,
7. makes refactoring easier, even across projects, tests just move with code,
8. executes tests immediately after discovery, in well-defined order,
9. stops on first failure, fits into a fast REPL-style way of working,
10. supports levels: unit, integration, performance etc.,
11. there are fixtures (context managers) like in other test tools,
12. async tests and fixtures are fully supported,
13. most functionality is in hooks which you can extend easily,
14. there is a root tester which can have subtesters,
15. output is send to a logger.

Although selftest promotes an agile, rigorous and Pythonic way of testing, since there is little magic and tests are just functions, you are free to organise them as you wish. You can even do it the Python unittest way, if you want.

An example

Selftest has a global root tester that can have an arbitrarily deep and wide tree of child testers. A typical module uses it as follows:

import selftest
test = selftest.get_tester(__name__)

def area(w, h):
    return w * h

@test
def area_basics():
    assert 9 == area(3, 3)
    assert 6 == area(2, 3)

Its creates a subtester using get_tester(). The resulting tester object is the main access point to all functionality of selftest. In this case, it is used as a decorator to mark and execute a test function.

More on assert later.

2. Basic API

General

Tester objects

Selftest has a hierarchical tree of test runners called tester objects, a bit like Pythons logging facility. The main program is supposed to configure the root (although it doesn't have to) with various options. Testers lower in the tree can override these options.

Hooks

Tester objects only mark and execute tests, optionally with options. Any functionality beyond that is provided by hooks, which themselves are just options that can be overridden. There are hooks for:

• fixtures
• operators
• async support
• filters
• diffs
• wildcards
• levels

These hooks introduce new functionality by extending the API of the tester object.

API

The API falls apart into five categories:

• a module level API,
• a tester object API,
• core options,
• hooks API.
• APIs introduced by hooks

Module Level API

The selftest core consist of two module level functions:

basic_config(**options)

Sets options for the root tester. This can be called only once, before get_tester(). If not called, default options are used. This typicalliy happens in the main of an application or in a program for running tests.

get_tester(name=None)

When name is None returns the root tester. Otherwise it returns a named child of the root. Name is a potentially hierarchical name separated by dots. Each level in this hierarchy becomes a child of the one preceding it. The last tester object is returned. Thus, get_tester("main.sub") creates a child main of the root and a child sub of the child main. It returns the latter.

Testers created this way become globally available. A call to get_tester() with the same name repeatedly will return the same tester.

Recommended is to use test = get_tester(__name__) at the start of your module. Using subtesters is a powerful way of organising tests. See the source code of selftest for many examples.

Tester Objects API

A tester object as returned from get_tester() support the following methods:

__call__(func)

A decorator for marking functions as tests:

@test
def function_marked_as_test():
    pass

This runs the given function and returns None. Thus, function_marked_as_test becomes None and the function is garbage collected subsequently. Keeping the test is possible with an option.

__call__(**options)

A way for setting options:

@test(keep=True, my_option=42)
def function_marked_as_test():
    pass

Important: This creates an anonymous child tester with given options. This tester is only used once and then discarded. There is a fixtures that gives tests access to it, see Fixtures below.

If you get creative, you could also run:

def function_not_marked():
    pass

test(keep=True, my_option=42)(function_not_marked)    # or
test(function_not_marked, keep=True, my_option=42)

All methods are 100% equivalent. In fact, the full signature is:

__call__(*funcs, **options)

So you can run multiple test functions with the given options at once.

get_child(**options)

This function is an alias for __call__(**options). It does exactly the same.

child(**options)

This creates a child and returns a context manager.

test = selftest.get_tester(__name__)
with test.child(level=CRITICAL) as crit:
    @crit
    def a_critical_test_function():
        pass

fail(*args, **kwargs)

Use as guard in tests. Raises AssertionError with the given args, appending kwargs to args.

log_stats()

Log the current value of the statistics to the configured output. The actual log record contains lots of data, but by default only the message is printed. See __main__.py for how to configure loggers.

Core Options

The core knows four options. Hooks may support additional options. Options can be given to any of these calls:

• basic_config(**options),
• __call__(**options),
• get_child(**options),
• child(**options).

Child testers inherit options from their parents and can override them.

option	type	default	Explanation
keep	boolean	False	Keep the function instead of discarding it.
run	boolean	True	Run immediately.
hooks	list	[]	List of hooks that are invoked in order.
subprocess	boolean	False	Runs test when inside a subprocess.

Normally, selftest runs a test as soon as it discovers it and then discards it. The example below show how tests can be run later by keeping and invoking them.

@test
def this_test_runs_immediately():
  pass

assert this_test_runs_immediately is None

@test(keep=True, run=False)
def another_test_for_running_later():
  pass

another_test_for_running_later()

Tests do not run in subprocesses which are spawned because the spawn method reimports all needed modules causing all tests to run again and in an endless loop. This only happens when a test spawns of course, but is is disabled by default because these tests run in practically the same context, which add little value.

Hooks API

Hooks are callable objects, optionally also implementing lookup(). Selftest core only dispatches to the hooks and most useful functionality is implemented in standaard hooks.

Installing a hook is done with the hooks option.

with test.child(hooks=[my_hook]) as hooked:
    @hooked
    def some_test():
        pass

__call__(tester, func)

A hook is an ordinary function accepting arguments tester and func. It is called when a test function is discovered, usually when the tester is used as decorator. The tester argument supports the Options API so hooks can manipulate options in the current tester. It should return the same func or a wrapper. If it returns None evaluating stops completely.

As an example, here is the complete hook for filtering:

def filter_hook(runner, func):
    f = runner.option_get('filter', '')
    if f in func.__qualname__:
        return func

Note that all hooks get to process func in turn, so be nice to them an use functools.wraps when you wrap.

lookup(tester, name)

Implemented by a hook that wants to intercept attribute lookups on the tester object. The hook can no longer be a simple function but must be an object understanding both __call__(tester, func) and lookup(tester, name). It is called when an attribute lookup takes place on the tester. When it returns a value, lookup stops. When it raises AttributeError, it continues with the next hook.

As an example, here is the hook for diffs, implementing both __call__ and lookup (references to diff and print functions omitted for clarity):

class DiffHook:

    def __call__(self, runner, func):
        return func

    def lookup(self, runner, name):
        if name == 'diff':
            return diff
        if name == 'diff2':
            return diff2
        if name == 'prrint':
            return prrint
        raise AttributeError

logrecord(tester, func, record)

Implemented by a hook when it wants to change something in the logrecord before it is emitted to the logger. The levels hook uses this to add the level of the test to the logrecord as an extra attribute, for example.

Options API

The Options API is meant for hooks manipulating options. Options are hierarchically registered, that is, each tester can have local values for options, and looks up missing ones in its parent.

option_get(name, default=None)

Returns the value for the option with given name for this tester or its closest parent.

option_setdefault(name, default)

Get option with name, searching all parents. When missing, sets the option on this tester with default as value and return it.

option_enumerate(name)

Enumerates all values for the option with the given name, starting with this tester, up to all its parent. List and tuple values are reversed and flattened (concatenated).

3) APIs from Hooks

Operators

Hook operator.py introduces the possibility to use various builtin operators instead of the assert statement. As a last resort, it looks up methods of the first argument to use as asserting statement. For example:

@test
def another_test():
    test.all(x > 1 for x in [1,2,3])      # use builtin all()
    test.startswith("rumbush", "rum")     # use method of first argument

When the given operator returns False according to bool() it raises AssertionError with the actual values of the arguments.

This shows how selftest stays close to Python as we know it. It does nothing more than looking up the given attribute in four places:

1. module operator, e.g.: test.gt(2, 1),
2. module builtins, e.g.: test.isinstance('aa', str),
3. module inspect, e.g.: test.isfunction(len),
4. the first argument, e.g.: test.isupper("ABC").

The benefits of this is that we do not have to learn new methods, that the assert functions are not limited, and that selftest can print the arguments for us on failure.

diff

All operators obtained this way support a keyword diff=<function> that, when present, is invoked with the actual arguments. The result is then given to the AssertionError instead of the actual arguments.

@test
def another_test():
    a = {7, 1, 2, 8, 3, 4}
    b = {1, 2, 9, 3, 4, 6}
    test.eq(a, b, diff=set.symmetric_difference)

The code above will raise AssertionError with as argument: {6, 7, 8, 9}.

For more general purpose diff functions, see the hook diffs.py.

Operators is included in the default root tester.

Fixtures

Hook fixtures.py introduces fixtures as seen in other test tools. The test.fixture attribute registers the next function as a fixture. A fixture is a Python contextmanager and can be used as such, or it can be specified as a formal argument to a test function. Fixtures accept arguments themselves by using the ':' notation.

@test.fixture
def answer(a=42):
    yield a

with test.answer as p:               # as a context manager
    test.eq(42, p)

@test
def probe_the_universe(answer):      # as a formal argument
    test.eq(42, answer)

@test
def something_wrong(answer:43):      # with argument 43
    test.ne(42, answer)
    test.eq(43, answer)

There are standard fixtures for:

1. test - gives access to the current tester, useful for anonymous subtesters,
2. stdout - captures sys.stdout, including that of subprocesses, in a StringIO,
3. stderr - captures sys.stderr, including that of subprocesses, in a StringIO,
4. stdin - replaces sys.stdin temporarily with a StringIO. Make sure to seek(0) after writing.
5. argv - saves sys.argv and restores it afterwards.
6. tmp_path:subpath - creates a temporary pathlib.Path object, optionally with a subpath,
7. raises:(Exception, message) - raises AssertionError if given code does not raise given exception with given message; sends an object as target from which the exception can be retieved for further testing. See example.
8. guard - isolates tests by saving and restoring sys.path, sys.meta_path, and sys.modules.

An example for using raises() in different ways:

@test
def should_raise_error(raises:(AttributeError, "'list' object has no attribute 'a'")):
    [].a

with test.raises(AttributeError, "'list' object has no attribute 'a'"):
    [].a

with test.raises(ImportError) as e:
    import worldpeace
test.eq('worldpeace', e.exception.name)
test.eq(None, e.exception.path)

Fixtures can be async (async def) but async fixtures can only be used in async tests. Async fixtures are executed in the event loop of the async test they are declared for.

Fixtures is included in the default root tester.

Filtering

Hook filter.py supports the option filter=<str> and only executes test whose qualified name includes the given <str>.

with test(hooks=[filter_hook]) as ftest:
    with ftest(filter='moon') as moon:
        @moon
        def phase_of_the_moon_bug():
            pass

Filtering is included in the default root tester.

Diffs

Hook diffs.py provides the attributes:

• test.diff(a b) -- a Python pprint + difflib based general purpose diff for use with the operator hook,
• test.diff2(a, b) -- an Selftest prrint + difflib based diff for Plain Old Data (POD) objects,
• test.prrint(a) -- a pretty printer for POD objects. Use instead of Pythons pprint().

Diffs is included in the default root tester.

Async

Hook asyncer.py supports asyncio tests defined with async def. Async tests can contain other async tests, however due to limitations in Python (being that async is partially a syntax feature and not fully dynamic) this forces nested async tests to be executed in a separate event loop in a separate thread.

Non-nested async tests are run in a new event loop in de current thread.

The option timeout=<time in s> specifies the maximum amount in seconds a test can run. After that an TimeoutError is raised.

The option slow_callback_duration=<time in s> specifies the time after which asyncio begins to emit warnings about tests running too long.

Async is included in the default root tester.

Wildcards

Hook wildcard.py introduces the attribute test.any which can be used in structured data comparisons as a wildcard. Its matching can optionally be limited using a function as argument. It is nice to combine this with the operator hook:

test.eq([4, test.any,           42], [4, -3, 42])               # succeeds
test.ne([4, test.any(test.pos), 42], [4, -3, 42])               # fails

Wildcards is included in the default root tester.

Levels

Hook levels.py introduces test levels such as unit, integration etcetera, together with two options, level and threshold, to control it. It is meant to run only certain tests depending on the context. During development for example, for reasons of speed, integration and performance tests can be skipped. The levels are just numbers and a number functions as a level/threshold, much like as in Pythons logging.

The levels are:

level	value
critical	50
unit	40
integration	30
performance	20

The default level is unit. Test levels are provided as attributes on the tester:

@test.critical
def a_critical_test():
    pass

Tests can also be put at a certain level with an option:

@test(level='critical')
def a_critical_test():
    pass

with test.child(level='critical') as critical:
     @critical
     def one():
         pass

The default threshold is integration.

The Levels hook adds the following information to any logrecord being generated:

• the number of the test level (20..50) , as testlevel
• the name of the test level ('unit', 'integration', etc) as testlevelname
• the name of the test level is also prepended to the msg attribute of the logrecord

A parent can have the threshold option set to one of the levels. This will block execution of tests in these children of levels lower than threshold. Note that tests do not have levels, only Testers have. NB: the highest threshold in chain of parent-children determines the tests to run. This means a child tester can run integration tests (threshold='integration'), but the root tester can overrule this to run only unit tests.

Levels is included in the default root tester.

Extended closure

The hook binder.py enables binding to a class definition in the making. The namespace of a class being defined is not available inside functions being defined in the class body. See this example:

def function_a():
    a = 42
    def function_b():
        assert a = 43
    function_b()            # runs fine
function_a()

class class_a:
    a = 42
    def function_b():
        assert a = 43
    function_b()            # NameError: name 'a' is not defined

In order to be able to embed tests in class definitions, the binder hook extends the binding of freevars in test functions to include those of the enclosing class. Sorry for the magic.

This hook is enabled by default, but only performs binding when the option bind=True is present.

class class_a:
    a = 42
    @test(bind=True)
    def function_b():
        assert a = 43

3. Running Tests

Development

Running tests during development can be done by just running or importing your module:

$ python <mymodule.py>
$ python -c "import mymodule"

When you only want to develop a submodule, just cd down into that directory and do the same. Only the tests of that submodule (and everything in imports) will be tested.

The methode above just prints crude messages and has no way to use options. For that use the main that comes with selftest:

$ selftest --help
Usage: selftest [options] module

Options:
  -h, --help            show this help message and exit
  -f FILTER, --filter=FILTER
                      only run tests whose qualified name contains FILTER
  -t , --threshold=THRESHOLD
                      only run tests whose level is >= THRESHOLD

For example to run your tests but not the imported ones from other packages:

$ selftest --filter mymodule mymodule

If you want to run the tests for selftest itself, go to the selftest project directory and use:

$ python -c "import selftest"  selftest.selftest

The argument selftest.selftest lets selftest run its own tests, which are normally skipped. Also, this avoids using the selftest main program because it is not guaranteed that an installed old version of selftest can run its own test from the future.

Production

During production, all tests are automatically run during startup when all needed modules are imported. If an the application configures a specific root, for example by calling logging.basicConfig(), the tests will automatically log there. Alternatively, you can setup a separate Logger for running tests. See Core API.

You can als filter tests or run tests for a specific level only. Or suppress them all. See the source code of __main__.py for ideas.

4. Misc

TODO

• also implement timeout for synchrounous code
• log stats without the loglevel