CHANGES.txt

CHANGES
*******

0.13 (unreleased)
=================

- Drop support for Python below 3.6

- Use GitHub Actions for CI


0.12 (2020-01-29)
=================

- **Breaking change**

  reg.arginfo(callable) returns now a FullArgSpec tuple
  instead of the deprecated ArgSpec tuple. FullArgSpec has
  full support for annotations and keyword-only parameters.

  Fixes `#55`_.

  .. _#55: https://github.com/morepath/reg/issues/55

- **Removed**: Removed support for Python 2.

  You have to upgrade to Python 3 if you want to use this version.

- Dropped support for Python 3.3.

- Added support for Python 3.5, 3.6, 3.7, 3.8 and PyPy 3.6.

- Make Python 3.7 the default testing environment.

- Add integration for the Black code formatter.


0.11 (2016-12-23)
=================

- **Breaking change**

  The ``key_predicate`` function is gone.  You can now use
  ``Predicate(..., index=KeyIndex)`` or ``match_key`` instead.

- **Breaking change**

  The ``class_predicate`` function is gone.  You can now use
  ``Predicate(..., index=ClassIndex)``, ``match_instance`` or
  ``match_class`` instead.

- **Breaking change**

  The undocumented ``Sentinel`` class and ``NOT_FOUND`` object are
  gone.

- **Breaking change**

  The class ``PredicateRegistry`` is not longer part of the API.
  Internally, the classes ``MultiPredicate``, ``MultiIndex``,
  ``SingleValueRegistry`` have all been merged into
  ``PredicateRegistry``, which should now considered an implementation
  detail.

- The second argument for ``match_key`` is now optional; if you
  don't supply it ``match_key`` will generate a predicate function
  that extracts that name by default.

- The documentation now includes a section describing the internals of
  Reg.

- Upload universal wheels to pypi during release.


0.10 (2016-10-04)
=================

- **Breaking change**

  Reg has undergone another API breaking change. The goals of this
  change were:

  * Make everything explicit.

  * A simpler implementation structure -- dispatch functions maintain
    their own registries, which allows for less interacting objects.

  * Make the advanced context-dependent dispatch more Pythonic by
    using classes with special dispatch methods.

  Detailed changes:

  * ``reg.Registry`` is gone. Instead you register directly on the
    dispatch function::

      @reg.dispatch('a')
      def foo(a):
        ...

      def foo_implementation(a):
        ...

      foo.register(foo_implementation, a=Document)

  * Caching is now per dispatch function, not globally per lookup. You
    can pass a ``get_key_lookup`` function that wraps
    ``reg.PredicateRegistry`` instance inside a
    ``reg.DictCachingKeyLookup`` cache. You can also use a
    ``reg.LruCachingKeyLookup`` if you expect a dispatch to be called
    with a large amount of possible predicate combinations, to
    preserve memory.

  * The whole concept of a "lookup" is gone:

    * ``reg.implicit`` is gone: everything is explicit. There is no more
      implicit lookup.

    * ``reg.Lookup`` itself is gone -- its now implemented directly in the
      dispatch object, but was already how you accessed it.

    * The special ``lookup`` argument to pass through the current
      ``Lookup`` is gone. If you need context-dependent dispatch, you
      use dispatch methods.

    * If you need context dependent dispatch, where the functions
      being dispatched to depend on application context (such as
      Morepath's application mounting), you use
      ``reg.dispatch_method`` to create a dispatch method. A dispatch
      method maintains an entirely separate dispatch registry for each
      subclass. You use ``reg.methodify`` to register a dispatch
      function that takes an optional context first argument.

  If you do not use the context-dependent dispatch feature, then to
  upgrade your code:

  * remove any ``reg.set_implicit`` from your code, setup of
    ``Lookup`` and the like.

  * If you use an explicit ``lookup`` argument you can just remove them.

  * You also need to change your registration code: no more
    ``reg.Registry`` setup.

  * Change your registrations to be on the dispatch objects itself
    using ``Dispatch.register``.

  * To enable caching you need to set up ``get_key_lookup`` on the
    dispatch functions. You can create a partially applied version of
    ``dispatch`` to make this less verbose::

       import reg
       from functools import partial

       def get_caching_key_lookup(r):
           return reg.CachingKeyLookup(r, 5000, 5000, 5000)

       dispatch = partial(reg.dispatch, get_key_lookup=get_caching_key_lookup)

  * ``dispatch_external_predicates`` is gone. Just use ``dispatch``
    directly.  You can add predicates to an existing Dispatch object
    using the ``add_predicates`` method.

  If you do use the context-dependent dispatch feature, then you also
  need to:

  * identify the context class in your application (or create one).

  * move the dispatch functions to this class, marking them with
    ``@reg.dispatch_method`` instead of ``@reg.dispatch``.

  * Registration is now using
    ``<context_class>.<method>.register``. Functions you register this
    way behave as methods to ``context_class``, so get an instance of
    this class as the first argument.

  * You can also use ``reg.methodify`` to register implementation
    functions that do not take the context as the first argument --
    this is useful when upgrading existing code.

  * Call your context-dependent methods as methods on the context
    instance. This way you can indicate what context you are calling
    your dispatch methods in, instead of using the `lookup`` argument.

  In some cases you want a context-dependent method that actually does
  not dispatch on any of its arguments. To support this use case you
  can simply set function (that takes an app argument) as a the method
  on the context class directly::

     Context.my_method = some_function

  If you want to set up a function that doesn't take a reference to a
  ``Context`` instance as its first argument, you can use
  ``reg.methodify`` to turn it into a method that ignores its first
  argument::

     Context.my_method = reg.methodify(some_function)

  If you want to register a function that might or might not have a
  reference to a ``Context`` instance as its first argument, called,
  e.g., ``app``, you can use the following::

     Context.my_method = reg.methodify(some_function, selfname='app')

- **Breaking change**

  Removed the helper function ``mapply`` from the API.

- **Breaking change**

  Removed the exception class ``KeyExtractorError`` from the API.
  When passing the wrong number of arguments to a dispatch function,
  or when using the wrong argument names, you will now get a
  TypeError, in conformity with standard Python behaviour.

- **Breaking change**

  Removed the ``KeyExtractor`` class from the API. Callables used in
  predicate construction now expect the same arguments as the dispatch
  function.

- **Breaking change**

  Removed the ``argnames`` attribute from ``Predicate`` and its
  descendant.

- **Breaking change**

  Remove the ``match_argname`` predicate.  You can now use
  ``match_instance`` with no callable instead.

- The second argument for ``match_class`` is now optional; if you
  don't supply it ``match_class`` will generate a predicate function
  that extracts that name by default.

- The second argument for ``match_instance`` is now optional; if you
  don't supply it ``match_instance`` will generate a predicate function
  that extracts that name by default.

- Include doctests in Tox and Travis.

- We now use virtualenv and pip instead of buildout to set up the
  development environment. The development documentation has been
  updated accordingly.

- As we reached 100% code coverage for pytest, coveralls integration
  was replaced by the ``--fail-under=100`` argument of ``coverage
  report`` in the tox coverage test.

0.9.3 (2016-07-18)
==================

- Minor fixes to documentation.

- Add tox test environments for Python 3.4 and 3.5, PyPy 3 and PEP 8.

- Make Python 3.5 the default Python environment.

- Changed location ``NoImplicitLookupError`` was imported from in
  ``__init__.py``.

0.9.2 (2014-11-13)
==================

- Reg was a bit too strict; when you had multiple (but not single)
  predicates, Reg would raise KeyError when you put in an unknown
  key. Now they're just being silently ignored, as they don't do any
  harm.

- Eliminated a check in ``ArgExtractor`` that could never take place.

- Bring test coverage back up to 100%.

- Add converage configuration to ignore test files in coverage
  reporting.

0.9.1 (2014-11-11)
==================

- A bugfix in the behavior of the fallback logic. In situations with
  multiple predicates of which one is a class predicate it was
  possible for a fallback not to be found even though a fallback was
  available.

0.9 (2014-11-11)
================

Total rewrite of Reg! This includes a range of changes that can break
code. The primary motivations for this rewrite:

* unify predicate system with class-based lookup system.

* extract dispatch information from specific arguments instead of all
  arguments.

Some specific changes:

* Replaced ``@reg.generic`` decorator with ``@reg.dispatch()``
  decorator. This decorator can be configured with predicates that
  extract information from the arguments. Rewrite this::

    @reg.generic
    def foo(obj):
       pass

  to this::

    @reg.dispatch('obj')
    def foo(obj):
       pass

  And this::

    @reg.generic
    def bar(a, b):
        pass

  To this::

    @reg.dispatch('a', 'b')
    def bar(a, b):
        pass

  This is to get dispatch on the classes of these instance
  arguments. If you want to match on the class of an attribute of
  an argument (for instance) you can use ``match_instance``
  with a function::

    @reg.dispatch(match_instance('a', lambda a: a.attr))

  The first argument to ``match_instance`` is the name of the
  predicate by which you refer to it in ``register_function``.

  You can also use ``match_class`` to have direct dispatch on classes
  (useful for replicating classmethods), and ``match_key`` to have
  dispatch on the (immutable) value of the argument (useful for a view
  predicate system). Like for ``match_instance``, you supply functions
  to these match functions that extract the exact information to
  dispatch on from the argument.

* The ``register_function`` API replaces the ``register`` API to
  register a function. Replace this::

    r.register(foo, (SomeClass,), dispatched_to)

  with::

    r.register_function(foo, dispatched_to, obj=SomeClass)

  You now use keyword parameters to indicate exactly those arguments
  specified by ``reg.dispatch()`` are actually predicate
  arguments. You don't need to worry about the order of predicates
  anymore when you register a function for it.

* The new ``classgeneric`` functionality is part of the predicate
  system now; you can use ``reg.match_class`` instead. Replace::

    @reg.classgeneric
    def foo(cls):
       pass

  with::

    @reg.dispatch(reg.match_class('cls', lambda cls: cls))
    def foo(cls):
        pass

  You can do this with any argument now, not just the first one.

* pep443 support is gone. Reg is focused on its own dispatch system.

* Compose functionality is gone -- it turns out Morepath doesn't use
  lookup composition to support App inheritance. The cached lookup
  functionality has moved into ``registry.py`` and now also supports
  caching of predicate-based lookups.

* Dependency on the future module is gone in favor of a small amount
  of compatibility code.

0.8 (2014-08-28)
================

- Added a ``@reg.classgeneric``. This is like ``@reg.generic``, but
  the first argument is treated as a class, not as an instance. This
  makes it possible to replace ``@classmethod`` with a generic
  function too.

- Fix documentation on running documentation tests. For some reason
  this did not work properly anymore without running sphinxpython
  explicitly.

- Optimization: improve performance of generic function calls by
  employing ``lookup_mapply`` instead of general ``mapply``, as we
  only care about passing in the lookup argument when it's defined,
  and any other arguments should work as before. Also added a
  ``perf.py`` which is a simple generic function timing script.

0.7 (2014-06-17)
================

- Python 2.6 compatibility. (Ivo van der Wijk)

- Class maps (and thus generic function lookup) now works with old
  style classes as well.

- Marked as production/stable now in ``setup.py``.

0.6 (2014-04-08)
================

- Removed unused code from mapply.py.

- Typo fix in API docs.

0.5 (2014-01-21)
================

- Make ``reg.ANY`` public. Used for predicates that match any value.

0.4 (2014-01-14)
================

- arginfo has been totally rewritten and is now part of the public API of reg.

0.3 (2014-01-06)
================

- Experimental Python 3.3 support thanks to the future module.

0.2 (2013-12-19)
================

- If a generic function implementation defines a ``lookup`` argument
  that argument will be the lookup used to call it.

- Added ``reg.mapply()``. This allows you to call things with more
  keyword arguments than it accepts, ignoring those extra keyword
  args.

- A function that returns ``None`` is not assumed to fail, so no fallback
  to the original generic function is triggered anymore.

- An optional ``precalc`` facility is made available on ``Matcher`` to
  avoid some recalculation.

- Implement a specific ``PredicateMatcher`` that matches a value on
  predicate.

0.1 (2013-10-28)
================

- Initial public release.