[Python-Dev] proto-pep: plugin proposal (for unittest)
Michael Foord
fuzzyman at voidspace.org.uk
Fri Jul 30 00:55:50 CEST 2010
Hello all,
My apologies in advance if email mangles whitespace in the code
examples. I can reformulate as a PEP if that is deemed useful and this
document can be found online at:
http://hg.python.org/unittest2/file/tip/description.txt
(Please excuse errors and omissions - but do feel free to point them out.)
This is a description, and request for feedback, of the unittest plugin
system that I am currently prototyping in the plugins branch of
unittest2_. My goal is to merge the plugin system back into unittest
itself in Python 3.2.
.. _unittest2: http://hg.python.org/unittest2
As part of the prototype I have been implementing some example plugins
(in unittest2/plugins/) so I can develop the mechanism against real
rather than imagined use cases. Jason Pellerin, creator of nose, has
been providing me with feedback and has been trying it out by porting
some of the nose plugins to unittest [#]_. He comments on the system "it
looks very flexible and clean". ;-)
Example plugins available and included:
* a pep8 and pyflakes checker
* a debugger plugin that drops you into pdb on test fail / error
* a doctest loader (looks for doctests in all text files in the
project)
* use a regex for matching files in test discovery instead of a glob
* growl notifications on test run start and stop
* filter individual *test methods* using a regex
* load test functions from modules as well as TestCases
* integration with the coverage module for coverage reporting
In addition I intend to create a plugin that outputs junit compatible
xml from a test run (for integration with tools like the hudson
continuous integration server) and a test runner that runs tests in
parallel using multiprocessing.
Not all of these will be included in the merge to unittest. Which ones
will is still an open question.
I'd like feedback on the proposal, and hopefully approval to port it
into unittest after discussion / amendment / completion. In particular
I'd like feedback on the basic system, plus which events should be
available and what information should be available in them. Note that
the system is *not* complete in the prototype. Enough is implemented to
get "the general idea" and to formalise the full system. It still needs
extensive tests and the extra work in TestProgram makes it abundantly
clear that refactoring there is well overdue...
In the details below open questions and todos are noted. I *really*
value feedback (but will ignore bikeshedding ;-)
.. note::
Throughout this document I refer to the prototype implementation
using names like ``unittest2.events.hooks``. Should this proposal be
accepted then the names will live in the unittest package instead of
unittest2.
The core classes for the event system live in the current
implementation in the ``unittest2.events`` namespace.
Abstract
========
unittest lacks a standard way of extending it to provide commonly
requested functionality, other than subclassing and overriding (and
reimplementing) parts of its behaviour. This document describes a plugin
system already partially prototyped in unittest2.
Aspects of the plugin system include:
* an events mechanism where handlers can be registered and called during
a test run
* a Plugin class built over the top of this for easy creation of plugins
* a configuration system for specifying which plugins should be loaded
and for configuring individual plugins
* command line integration
* the specific set of events and the information / actions available to them
As the plugin system essentially just adds event calls to key places it
has few backwards compatibility issues. Unfortunately existing
extensions that override the parts of unittest that call these events
will not be compatible with plugins that use them. Framework authors who
re-implement parts of unittest, for example custom test runners, may
want to add calling these events in appropriate places.
Rationale
=========
Why a plugin system for unittest?
unittest is the standard library test framework for Python but in recent
years has been eclipsed in functionality by frameworks like nose and
py.test. Among the reasons for this is that these frameworks are easier
to extend with plugins than unittest. unittest makes itself particularly
difficult to extend by using subclassing as its basic extension
mechanism. You subclass and override behaviour in its core classes like
the loader, runner and result classes.
This means that where you have more than one "extension" working in the
same area it is very hard for them to work together. Whilst various
extensions to unittest do exist (e.g. testtools, zope.testrunner etc)
they don't tend to work well together. In contrast the plugin system
makes creating extensions to unittest much simpler and less likely that
extensions will clash with each other.
nose itself exists as a large system built over the top of unittest.
Extending unittest in this way was very painful for the creators of
nose, and every release of Python breaks nose in some way due to changes
in unittest. One of the goals of the extension mechanism is to allow
nose2 to be a much thinner set of plugins over unittest(2) that is much
simpler to maintain [#]_. The early indications are that the proposed
system is a good fit for this goal.
Low Level Mechanism
====================
The basic mechanism is having events fired at various points during a
test run. Plugins can register event handler functions that will be
called with an event object. Multiple functions may be registered to
handle an event and event handlers can also be removed.
Over the top of this is a ``Plugin`` class that simplifies building
plugins on top of this mechanism. This is described in a separate section.
The events live on the ``unittest2.events.hooks`` class. Handlers are
added using ``+=`` and removed using ``-=``, a syntax borrowed from the
.NET system.
For example adding a handler for the ``startTestRun`` event::
from unittest2.events import hooks
def startTestRun(event):
print 'test run started at %s' % event.startTime
hooks.startTestRun += startTestRun
Handlers are called with an Event object specific to the event. Each
event provides different information on its event objects as attributes.
For example the attributes available on ``StartTestRunEvent`` objects are:
* ``suite`` - the test suite for the full test run
* ``runner`` - the test runner
* ``result`` - the test result
* ``startTime``
The name of events, whether any should be added or removed, and what
information is available on the event objects are all valid topics for
discussion. Specific events and the information available to them is
covered in a section below.
An example plugin using events directly is the ``doctestloader`` plugin.
Framework authors who re-implement parts of unittest, for example custom
test runners, may want to add calling these events in appropriate
places. This is very simple. For example the ``pluginsLoaded`` event is
fired with a ``PluginsLoadedEvent`` object that is instantiated without
parameters::
from unittest2.events import hooks, PluginsLoadedEvent
hooks.pluginsLoaded(PluginsLoadedEvent())
Why use event objects and not function parameters?
--------------------------------------------------
There are several reasons to use event objects instead of function
parameters. The *disadvantage* of this is that the information available
to an event is not obvious from the signature of a handler. There are
several compelling advantages however:
* the signature of all handler functions is identical and therefore easy
to remember
* backwards compatibility - new attributes can be added to event objects
(and parameters deprecated) without breaking existing plugins. Changing
the way a function is called (unless all handlers have a ``**kw``
signature) is much harder.
* several of the events have a lot of information available. This would
make the signature of handlers huge. With an event object handlers only
need to be aware of attributes they are interested in and ignore
information they aren't interested in ("only pay for what you eat").
* some of the attributes are mutable - the event object is shared
between all handlers, this would be less obvious if function parameters
were used
* calling multiple handlers and still returning a value (see the handled
pattern below)
The handled pattern
--------------------
Several events can be used to *override* the default behaviour. For
example the 'matchregexp' plugin uses the ``matchPath`` event to replace
the default way of matching files for loading as tests during test
discovery. The handler signals that it is handling this event, and the
default implementation should not be run, by setting ``event.handled =
True``::
def matchRegexp(event):
pattern = event.pattern
name = event.name
event.handled = True
path = event.path
if matchFullPath:
return re.match(pattern, path)
return re.match(pattern, name)
Where the default implementation returns a value, for example creating a
test suite, or in the case of ``matchPath`` deciding if a path matches a
file that should be loaded as a test, the handler can return a result.
If an event sets handled on an event then no more handlers will be
called for that event. Which events can be handled, and which not, is
discussed in the events section.
The Plugin Class
================
A sometimes-more-convenient way of creating plugins is to subclass the
``unittest2.events.Plugin`` class. By default subclassing ``Plugin``
will auto-instantiate the plugin and store the instance in a list of
loaded plugins.
Each plugin has a ``register()`` method that auto-hooks up all methods
whose names correspond to events. Plugin classes may also provide
``configSection`` and ``commandLineSwitch`` class attributes which
simplifies enabling the plugin through the command line and making
available a section from the configuration file(s).
A simple plugin using this is the 'debugger' plugin that starts ``pdb``
when the ``onTestFail`` event fires::
from unittest2.events import Plugin
import pdb
import sys
class Debugger(Plugin):
configSection = 'debugger'
commandLineSwitch = ('D', 'debugger', 'Enter pdb on test fail
or error')
def __init__(self):
self.errorsOnly = self.config.as_bool('errors-only',
default=False)
def onTestFail(self, event):
value, tb = event.exc_info[1:]
test = event.test
if self.errorsOnly and isinstance(value,
test.failureException):
return
original = sys.stdout
sys.stdout = sys.__stdout__
try:
pdb.post_mortem(tb)
finally:
sys.stdout = original
A plugin that doesn't want to be auto-instantiated (for example a base
class used for several plugins) can set ``autoCreate = False`` as a
class attribute. (This attribute is only looked for on the class
directly and so isn't inherited by subclasses.) If a plugin is
auto-instantiated then the instance will be set as the ``instance``
attribute on the plugin class.
``configSection`` and ``commandLineSwitch`` are described in the
`configuration system`_ and `command line integration`_ sections.
Plugin instances also have an ``unregister`` method that unhooks all
events. It doesn't exactly correspond to the ``register`` method (it
undoes some of the work done when a plugin is instantiated) and so can
only be called once.
Plugins to be loaded are specified in configuration files. For
frameworks not using the unittest test runner and configuration system
APIs for loading plugins are available in the form of the
``loadPlugins`` function (which uses the configuration system to load
plugins) and ``loadPlugin`` which loads an individual plugin by module
name. Loading plugins just means importing the module containing the plugin.
Configuration system
====================
By default the unittest2 test runner (triggered by the unit2 script or
for unittest ``python -m unittest``) loads two configuration files to
determine which plugins to load.
A user configuration file, ~/unittest.cfg (alternative name and location
would be possible), can specify plugins that will always be loaded. A
per-project configuration file, unittest.cfg which should be located in
the current directory when unit2 is launched, can specify plugins for
individual projects.
To support this system several command line options have been added to
the test runner::
--config=CONFIGLOCATIONS
Specify local config file location
--no-user-config Don't use user config file
--no-plugins Disable all plugins
Several config files can be specified using ``--config``. If the user
config is being loaded then it will be loaded first (if it exists),
followed by the project config (if it exists) *or* any config files
specified by ``--config``. ``--config`` can point to specific files, or
to a directory containing a ``unittest.cfg``.
Config files loaded later are merged into already loaded ones. Where a
*key* is in both the later key overrides the earlier one. Where a
section is in both but with different keys they are merged. (The
exception to keys overriding is the 'plugins' key in the unittest
section - these are combined to create a full list of plugins. Perhaps
multiline values in config files could also be merged?)
plugins to be loaded are specified in the ``plugins`` key of the
``unittest`` section::
[unittest]
plugins =
unittest2.plugins.checker
unittest2.plugins.doctestloader
unittest2.plugins.matchregexp
unittest2.plugins.moduleloading
unittest2.plugins.debugger
unittest2.plugins.testcoverage
unittest2.plugins.growl
unittest2.plugins.filtertests
The plugins are simply module names. They either hook themselves up
manually on import or are created by virtue of subclassing ``Plugin``. A
list of all loaded plugins is available as
``unittest2.events.loadedPlugins`` (a list of strings).
For accessing config values there is a ``getConfig(sectionName=None)``
function. By default it returns the whole config data-structure but it
an also return individual sections by name. If the section doesn't exist
an empty section will be returned. The config data-structure is not
read-only but there is no mechanism for persisting changes.
The config is a dictionary of ``Section`` objects, where a section is a
dictionary subclass with some convenience methods for accessing values::
section = getConfig(sectionName)
integer = section.as_int('foo', default=3)
number = section.as_float('bar', default=0.0)
# as_list returns a list with empty lines and comment lines removed
items = section.as_list('items', default=[])
# as_bool allows 'true', '1', 'on', 'yes' for True (matched
case-insensitively) and
# 'false', 'off', '0', 'no', '' for False
value = section.as_bool('value', default=True)
If a plugin specifies a ``configSection`` as a class attribute then that
section will be fetched and set as the ``config`` attribute on instances.
By convention plugins should use the 'always-on' key in their config
section to specify that the plugin should be switched on by default. If
'always-on' exists and is set to 'True' then the ``register()`` method
will be called on the plugin to hook up all events. If you don't want a
plugin to be auto-registered you should fetch the config section
yourself rather than using ``configSection``.
If the plugin is configured to be 'always-on', and is auto-registered,
then it doesn't need a command line switch to turn it on (although it
may add other command line switches or options) and
``commandLineSwitch`` will be ignored.
Command Line Interface
======================
Plugins may add command line options, either switches with a callback
function or options that take values and will be added to a list. There
are two functions that do this: ``unittest2.events.addOption`` and
``unittest2.events.addDiscoveryOption``. Some of the events are only
applicable to test discovery (``matchPath`` is the only one currently I
think), options that use these events should use ``addDiscoveryOption``
which will only be used if test discovery is invoked.
Both functions have the same signature::
addDiscoveryOption(callback, opt=None, longOpt=None, help=None)
addOption(plugin.method, 'X', '--extreme', 'Run tests in extreme mode')
* ``callback`` is a callback function (taking no arguments) to be
invoked if the option is on *or* a list indicating that this is an
option that takes arguments, values passed in at the command line will
be added to the list
* ``opt`` is a short option for the command (or None) not including the
leading '-'
* ``longopt`` a long option for the command (or None) not including the
leading '--'
* ``help`` is optional help text for the option, to be displayed by
``unit2 -h``
Lowercase short options are reserved for use by unittest2 internally.
Plugins may only add uppercase short options.
If a plugin needs a simple command line switch (on/off) then it can set
the ``commandLineSwitch`` class attribute to a tuple of ``(opt, longOpt,
help)``. The ``register()`` method will be used as the callback
function, automatically hooking the plugin up to events if it is
switched on.
The Events
==========
This section details the events implemented so far, the order they are
called in, what attributes are available on the event objects, whether
the event is 'handleable' (and what that means for the event), plus the
intended use case for the event.
Events in rough order are:
* ``pluginsLoaded``
* ``handleFile``
* ``matchPath``
* ``loadTestsFromNames``
* ``loadTestsFromName``
* ``loadTestsFromModule``
* ``loadTestsFromTestCase``
* ``getTestCaseNames``
* ``runnerCreated``
* ``startTestRun``
* ``startTest``
* ``onTestFail``
* ``stopTest``
* ``stopTestRun``
pluginsLoaded
-------------
This event is useful for plugin initialisation. It is fired after all
plugins have been loaded, the config file has been read and command line
options processed.
The ``PluginsLoadedEvent`` has one attribute: ``loadedPlugins`` which
is a list of strings referring to all plugin modules that have been loaded.
handleFile
----------
This event is fired when a file is looked at in test discovery or a
*filename* is passed at the command line. It can be used for loading
tests from non-Python files, like doctests from text files, or adding
tests for a file like pep8 and pyflakes checks.
A ``HandleFileEvent`` object has the following attributes:
* ``extraTests`` - a list, extend this with tests to *add* tests that
will be loaded from this file without preventing the default test loading
* ``name`` - the name of the file
* ``path`` - the full path of the file being looked at
* ``loader`` - the ``TestLoader`` in use
* ``pattern`` - the pattern being used to match files, or None if not
called during test discovery
* ``top_level_directory`` - the top level directory of the project tests
are being loaded from, or the current working directory if not called
during test discovery
This event *can* be handled. If it is handled then the handler should
return a test suite or None. Returning None means no tests will be
loaded from this file. If any plugin has created any ``extraTests`` then
these will be used even if a handler handles the event and returns None.
If this event is not handled then it will be matched against the pattern
(test discovery only) and either be rejected or go through for standard
test loading.
matchPath
---------
``matchPath`` is called to determine if a file should be loaded as a
test module. This event only fires during test discovery.
``matchPath`` is only fired if the filename can be converted to a valid
python module name, this is because tests are loaded by importing. If
you want to load tests from files whose paths don't translate to valid
python identifiers then you should use ``handleFile`` instead.
A ``MatchPathEvent`` has the following attributes:
* ``path`` - full path to the file
* ``name`` - filename only
* ``pattern`` - pattern being used for discovery
This event *can* be handled. If it is handled then the handler should
return True or False to indicate whether or not test loading should be
attempted from this file. If this event is not handled then the pattern
supplied to test discovery will be used as a glob pattern to match the
filename.
loadTestsFromNames
------------------
This event is fired when ``TestLoader.loadTestsFromNames`` is called.
Attributes on the ``LoadFromNamesEvent`` object are:
* ``loader`` - the test loader
* ``names`` - a list of the names being loaded
* ``module`` - the module passed to ``loader.loadTestsFromNames(...)``
* ``extraTests`` - a list of extra tests to be added to the suites
loaded from the names
This event can be handled. If it is handled then the handler should
return a list of suites or None. Returning None means no tests will be
loaded from these names. If any plugin has created any ``extraTests``
then these will be used even if a handler handles the event and returns
None.
If this event is not handled then ``loader.loadTestFromName`` will be
called for each name to build up the list of suites.
loadTestsFromName
-----------------
This event is fired when ``TestLoader.loadTestsFromName`` is called.
Attributes on the ``LoadFromNameEvent`` object are:
* ``loader`` - the test loader
* ``name`` - the name being loaded
* ``module`` - the module passed to ``loader.loadTestsFromName(...)``
* ``extraTests`` - a suite of extra tests to be added to the suite
loaded from the name
This event can be handled. If it is handled then the handler should
return a TestSuite or None. Returning None means no tests will be loaded
from this name. If any plugin has created any ``extraTests`` then these
will be used even if a handler handles the event and returns None.
If the event is not handled then each name will be resolved and tests
loaded from it, which may mean calling ``loader.loadTestsFromModule`` or
``loader.loadTestsFromTestCase``.
loadTestsFromModule
-------------------
This event is fired when ``TestLoader.loadTestsFromModule`` is called.
It can be used to customise the loading of tests from a module, for
example loading tests from functions as well as from TestCase classes.
Attributes on the ``LoadFromModuleEvent`` object are:
* ``loader`` - the test loader
* ``module`` - the module object tests
* ``extraTests`` - a suite of extra tests to be added to the suite
loaded from the module
This event can be handled. If it is handled then the handler should
return a TestSuite or None. Returning None means no tests will be loaded
from this module. If any plugin has created any ``extraTests`` then
these will be used even if a handler handles the event and returns None.
If the event is not handled then ``loader.loadTestsFromTestCase`` will
be called for every TestCase in the module.
Event if the event is handled, if the module defines a ``load_tests``
function then it *will* be called for the module. This removes the
responsibility for implementing the ``load_tests`` protocol from plugin
authors.
loadTestsFromTestCase
---------------------
This event is fired when ``TestLoader.loadTestsFromTestCase`` is called.
It could be used to customise the loading of tests from a TestCase, for
example loading tests with an alternative prefix or created generative /
parameterized tests.
Attributes on the ``LoadFromTestCaseEvent`` object are:
* ``loader`` - the test loader
* ``testCase`` - the test case class being loaded
* ``extraTests`` - a suite of extra tests to be added to the suite
loaded from the TestCase
This event can be handled. If it is handled then the handler should
return a TestSuite or None. Returning None means no tests will be loaded
from this module. If any plugin has created any ``extraTests`` then
these will be used even if a handler handles the event and returns None
If the event is not handled then ``loader.getTestCaseNames`` will be
called to get method names from the test case and a suite will be
created by instantiating the TestCase class with each name it returns.
getTestCaseNames
----------------
This event is fired when ``TestLoader.getTestCaseNames`` is called. It
could be used to customise the method names used to load tests from a
TestCase, for example loading tests with an alternative prefix from the
default or filtering for specific names.
Attributes on the ``GetTestCaseNamesEvent`` object are:
* ``loader`` - the test loader
* ``testCase`` - the test case class that tests are being loaded from
* ``testMethodPrefix`` - set to None, modify this attribute to *change*
the prefix being used for this class
* ``extraNames`` - a list of extra names to use for this test case as
well as the default ones
* ``excludedNames`` - a list of names to exclude from loading from this
class
This event can be handled. If it is handled it should return a list of
strings. Note that if this event returns an empty list (or None which
will be replaced with an empty list then ``loadTestsFromTestCase`` will
still check to see if the TestCase has a ``runTest`` method.
Even if the event is handled ``extraNames`` will still be added to the
list, however *excludedNames`` won't be removed as they are filtered out
by the default implementation which looks for all attributes that are
methods (or callable) whose name begins with ``loader.testMethodPrefix``
(or ``event.testMethodPrefix`` if that is set) and aren't in the list of
excluded names (converted to a set first for efficient lookup).
The list of names will also be sorted using ``loader.sortTestMethodsUsing``.
runnerCreated
-------------
This event is fired when the ``TextTestRunner`` is instantiated. It can
be used to customize the test runner, for example replace the stream and
result class, without needing to write a custom test harness. This
should allow the default test runner script (``unit2`` or ``python -m
untitest``) to be suitable for a greater range of projects. Projects
that want to use custom test reporting should be able to do it through a
plugin rather than having to rebuild the runner and result machinery,
which also requires writing custom test collection too.
The ``RunnerCreatedEvent`` object only has one attribute; ``runner``
which is the runner instance.
startTestRun
------------
This event is fired when the test run is started. This is used, for
example, by the growl notifier that displays a growl notification when a
test run begins. It can also be used for filtering tests after they have
all been loaded or for taking over the test run machinery altogether,
for distributed testing for example.
The ``StartTestRunEvent`` object has the following attributes:
* ``test`` - the full suite of all tests to be run (may be modified in
place)
* ``result`` - the result object
* ``startTime`` - the time the test run started
Currently this event can be handled. This prevents the normal test run
from executing, allowing an alternative implementation, but the return
value is unused. Handling this event (as with handling any event)
prevents other plugins from executing. This means that the it wouldn't
be possible to safely combine a distributed test runner with a plugin
that filters the suite. Fixing this issue is one of the open issues with
the plugin system.
startTest
---------
This event is fired immediately before a test is executed (inside
``TestCase.run(...)``).
The ``StartTestEvent`` object has the following attributes:
* ``test`` - the test to be run
* ``result`` - the result object
* ``startTime`` - the time the test starts execution
This event cannot be handled.
onTestFail
----------
This event is fired when a test setUp, a test, a tearDown or a cleanUp
fails or errors. It is currently used by the debugger plugin. It is
*not* currently called for 'internal' unittest exceptions like
``SkipTest`` or expected failures and unexpected successes.
Attributes on the ``TestFailEvent`` are:
* ``test`` - the test
* ``result`` - the result
* ``exc_info`` - the result of ``sys.exc_info()`` after the error / fail
* ``when`` - one of 'setUp', 'call', 'tearDown', or 'cleanUp'
This event cannot be handled. Should this event be able to suppress
raised exceptions? It should also be able to modify the traceback so
that bare asserts could be used but still provide useful diagnostic
information. Should this event be fired for test skips?
stopTest
--------
This event is fired when a test execution is completed. It includes a
great deal of information about the test and could be used to completely
replace test reporting, making the test result potentially obsolete. It
will be used by the junit-xml plugin to generate the xml reports
describing the test run.
If there are errors during a tearDown or clean up functions then this
event may be fired several times for a test. For each call the ``stage``
will be different, although there could be several errors during clean
up functions.
Attributes on the ``StopTestEvent`` are:
* ``test`` - the test
* ``result`` - the result
* ``exc_info`` - the result of ``sys.exc_info()`` after an error / fail
or None for success
* ``stopTime``- time the test stopped, including tear down and clean up
functions
* ``timeTaken`` - total time for test execution from setUp to clean up
functions
* ``stage`` - one of setUp, call, tearDown, cleanUp, or None for success
* ``outcome`` - one of passed, failed, error, skipped,
unexpectedSuccess, expectedFailure
The outcomes all correspond to an attribute that will be set to True or
False depending on outcome:
* ``passed``
* ``failed``
* ``error``
* ``skipped``
* ``unexpectedSuccess``
* ``expectedFailure``
In addition there is a ``skipReason`` that will be None unless the test
was skipped, in which case it will be a string containing the reason.
This event cannot be handled.
stopTestRun
-----------
This event is fired when the test run completes. It is useful for
reporting tools.
The ``StopTestRunEvent`` event objects have the following attributes:
* ``runner`` - the test runner
* ``result`` - the test result
* ``stopTime`` - the time the test run completes
* ``timeTaken`` - total time taken by the test run
More information about the Python-Dev
mailing list