Path to this page:
Subject: CVS commit: pkgsrc/devel/py-ZopeInterface
From: Adam Ciarcinski
Date: 2020-06-08 15:18:06
Message id: 20200608131806.D475FFB27@cvs.NetBSD.org
Log Message:
py-ZopeInterface: updated to 5.1.0
5.1.0:
- Make ``@implementer(*iface)`` and ``classImplements(cls, *iface)``
ignore redundant interfaces. If the class already implements an
interface through inheritance, it is no longer redeclared
specifically for *cls*. This solves many instances of inconsistent
resolution orders, while still allowing the interface to be declared
for readability and maintenance purposes.
- Remove all bare ``except:`` statements. Previously, when accessing
special attributes such as ``__provides__``, ``__providedBy__``,
``__class__`` and ``__conform__``, this package wrapped such access
in a bare ``except:`` statement, meaning that many errors could pass
silently; typically this would result in a fallback path being taken
and sometimes (like with ``providedBy()``) the result would be
non-sensical. This is especially true when those attributes are
implemented with descriptors. Now, only ``AttributeError`` is
caught. This makes errors more obvious.
Obviously, this means that some exceptions will be propagated
differently than before. In particular, ``RuntimeError`` raised by
Acquisition in the case of circular containment will now be
propagated. Previously, when adapting such a broken object, a
``TypeError`` would be the common result, but now it will be a more
informative ``RuntimeError``.
In addition, ZODB errors like ``POSKeyError`` could now be
propagated where previously they would ignored by this package.
- Require that the second argument (*bases*) to ``InterfaceClass`` is
a tuple. This only matters when directly using ``InterfaceClass`` to
create new interfaces dynamically. Previously, an individual
interface was allowed, but did not work correctly. Now it is
consistent with ``type`` and requires a tuple.
- Let interfaces define custom ``__adapt__`` methods. This implements
the other side of the :pep:`246` adaptation protocol: objects being
adapted could already implement ``__conform__`` if they know about
the interface, and now interfaces can implement ``__adapt__`` if
they know about particular objects. There is no performance penalty
for interfaces that do not supply custom ``__adapt__`` methods.
This includes the ability to add new methods, or override existing
interface methods using the new ``@interfacemethod`` decorator.
- Make the internal singleton object returned by APIs like
``implementedBy`` and ``directlyProvidedBy`` for objects that
implement or provide no interfaces more immutable. Previously an
internal cache could be mutated.
5.0.2:
- Ensure that objects that implement no interfaces (such as direct
subclasses of ``object``) still include ``Interface`` itself in
their ``__iro___`` and ``__sro___``. This fixes adapter registry
lookups for such objects when the adapter is registered for
``Interface``.
5.0.1:
- Ensure the resolution order for ``InterfaceClass`` is consistent.
- Ensure the resolution order for ``collections.OrderedDict`` is
consistent on CPython 2. (It was already consistent on Python 3 and PyPy).
- Fix the handling of the ``ZOPE_INTERFACE_STRICT_IRO`` environment
variable. Previously, ``ZOPE_INTERFACE_STRICT_RO`` was read, in
contrast with the documentation.
5.0.0:
- Make an internal singleton object returned by APIs like
``implementedBy`` and ``directlyProvidedBy`` immutable. Previously,
it was fully mutable and allowed changing its ``__bases___``. That
could potentially lead to wrong results in pathological corner
cases.
- Support the ``PURE_PYTHON`` environment variable at runtime instead
of just at wheel build time. A value of 0 forces the C extensions to
be used (even on PyPy) failing if they aren't present. Any other
value forces the Python implementation to be used, ignoring the C
extensions.
- Cache the result of ``__hash__`` method in ``InterfaceClass`` as a
speed optimization. The method is called very often (i.e several
hundred thousand times during Plone 5.2 startup). Because the hash value never
changes it can be cached. This improves test performance from 0.614s
down to 0.575s (1.07x faster). In a real world Plone case a reindex
index came down from 402s to 320s (1.26x faster).
- Change the C classes ``SpecificationBase`` and its subclass
``ClassProvidesBase`` to store implementation attributes in their structures
instead of their instance dictionaries. This eliminates the use of
an undocumented private C API function, and helps make some
instances require less memory.
- Reduce memory usage in other ways based on observations of usage
patterns in Zope (3) and Plone code bases.
- Specifications with no dependents are common (more than 50%) so
avoid allocating a ``WeakKeyDictionary`` unless we need it.
- Likewise, tagged values are relatively rare, so don't allocate a
dictionary to hold them until they are used.
- Use ``__slots___`` or the C equivalent ``tp_members`` in more
common places. Note that this removes the ability to set arbitrary
instance variables on certain objects.
The changes in this release resulted in a 7% memory reduction after
loading about 6,000 modules that define about 2,200 interfaces.
.. caution::
Details of many private attributes have changed, and external use
of those private attributes may break. In particular, the
lifetime and default value of ``_v_attrs`` has changed.
- Remove support for hashing uninitialized interfaces. This could only
be done by subclassing ``InterfaceClass``. This has generated a
warning since it was first added in 2011 (3.6.5). Please call the
``InterfaceClass`` constructor or otherwise set the appropriate
fields in your subclass before attempting to hash or sort it.
- Remove unneeded override of the ``__hash__`` method from
``zope.interface.declarations.Implements``. Watching a reindex index
process in ZCatalog with on a Py-Spy after 10k samples the time for
``.adapter._lookup`` was reduced from 27.5s to 18.8s (~1.5x faster).
Overall reindex index time shrunk from 369s to 293s (1.26x faster).
- Make the Python implementation closer to the C implementation by
ignoring all exceptions, not just ``AttributeError``, during (parts
of) interface adaptation.
- Micro-optimization in ``.adapter._lookup`` , ``.adapter._lookupAll``
and ``.adapter._subscriptions``: By loading ``components.get`` into
a local variable before entering the loop a bytcode "LOAD_FAST 0
(components)" in the loop can be eliminated. In Plone, while running
all tests, average speedup of the "owntime" of ``_lookup`` is ~5x.
- Add ``__all__`` declarations to all modules. This helps tools that
do auto-completion and documentation and results in less cluttered
results. Wildcard ("*") are not recommended and may be affected.
- Fix ``verifyClass`` and ``verifyObject`` for builtin types like
``dict`` that have methods taking an optional, unnamed argument with
no default value like ``dict.pop``. On PyPy3, the verification is
strict, but on PyPy2 (as on all versions of CPython) those methods
cannot be verified and are ignored.
- Update the common interfaces ``IEnumerableMapping``,
``IExtendedReadMapping``, ``IExtendedWriteMapping``,
``IReadSequence`` and ``IUniqueMemberWriteSequence`` to no longer
require methods that were removed from Python 3 on Python 3, such as
``__setslice___``. Now, ``dict``, ``list`` and ``tuple`` properly
verify as ``IFullMapping``, ``ISequence`` and ``IReadSequence,``
respectively on all versions of Python.
- Add human-readable ``__str___`` and ``__repr___`` to ``Attribute``
and ``Method``. These contain the name of the defining interface
and the attribute. For methods, it also includes the signature.
- Change the error strings raised by ``verifyObject`` and
``verifyClass``. They now include more human-readable information
and exclude extraneous lines and spaces.
.. caution:: This will break consumers (such as doctests) that
depended on the exact error messages.
- Make ``verifyObject`` and ``verifyClass`` report all errors, if the
candidate object has multiple detectable violations. Previously they
reported only the first error.
Like the above, this will break consumers depending on the exact
output of error messages if more than one error is present.
- Add ``zope.interface.common.collections``,
``zope.interface.common.numbers``, and ``zope.interface.common.io``.
These modules define interfaces based on the ABCs defined in the
standard library ``collections.abc``, ``numbers`` and ``io``
modules, respectively. Importing these modules will make the
standard library concrete classes that are registered with those
ABCs declare the appropriate interface.
- Add ``zope.interface.common.builtins``. This module defines
interfaces of common builtin types, such as ``ITextString`` and
``IByteString``, ``IDict``, etc. These interfaces extend the
appropriate interfaces from ``collections`` and ``numbers``, and the
standard library classes implement them after importing this module.
This is intended as a replacement for third-party packages like
`dolmen.builtins <https://pypi.org/project/dolmen.builtins/>`_.
- Make ``providedBy()`` and ``implementedBy()`` respect ``super``
objects. For instance, if class ``Derived`` implements ``IDerived``
and extends ``Base`` which in turn implements ``IBase``, then
``providedBy(super(Derived, derived))`` will return ``[IBase]``.
Previously it would have returned ``[IDerived]`` (in general, it
would previously have returned whatever would have been returned
without ``super``).
Along with this change, adapter registries will unpack ``super``
objects into their ``__self___`` before passing it to the factory.
Together, this means that ``component.getAdapter(super(Derived,
self), ITarget)`` is now meaningful.
- Fix a potential interpreter crash in the low-level adapter
registry lookup functions.
- Adopt Python's standard `C3 resolution order
<https://www.python.org/download/releases/2.3/mro/>`_ to compute the
``__iro__`` and ``__sro__`` of interfaces, with tweaks to support
additional cases that are common in interfaces but disallowed for
Python classes. Previously, an ad-hoc ordering that made no
particular guarantees was used.
This has many beneficial properties, including the fact that base
interface and base classes tend to appear near the end of the
resolution order instead of the beginning. The resolution order in
general should be more predictable and consistent.
.. caution::
In some cases, especially with complex interface inheritance
trees or when manually providing or implementing interfaces, the
resulting IRO may be quite different. This may affect adapter
lookup.
The C3 order enforces some constraints in order to be able to
guarantee a sensible ordering. Older versions of zope.interface did
not impose similar constraints, so it was possible to create
interfaces and declarations that are inconsistent with the C3
constraints. In that event, zope.interface will still produce a
resolution order equal to the old order, but it won't be guaranteed
to be fully C3 compliant. In the future, strict enforcement of C3
order may be the default.
A set of environment variables and module constants allows
controlling several aspects of this new behaviour. It is possible to
request warnings about inconsistent resolution orders encountered,
and even to forbid them. Differences between the C3 resolution order
and the previous order can be logged, and, in extreme cases, the
previous order can still be used (this ability will be removed in
the future). For details, see the documentation for
``zope.interface.ro``.
- Make inherited tagged values in interfaces respect the resolution
order (``__iro__``), as method and attribute lookup does. Previously
tagged values could give inconsistent results.
- Add ``getDirectTaggedValue`` (and related methods) to interfaces to
allow accessing tagged values irrespective of inheritance.
- Ensure that ``Interface`` is always the last item in the ``__iro__``
and ``__sro__``. This is usually the case, but if classes that do
not implement any interfaces are part of a class inheritance
hierarchy, ``Interface`` could be assigned too high a priority.
- Implement sorting, equality, and hashing in C for ``Interface``
objects. In micro benchmarks, this makes those operations 40% to 80%
faster. This translates to a 20% speed up in querying adapters.
Note that this changes certain implementation details. In
particular, ``InterfaceClass`` now has a non-default metaclass, and
it is enforced that ``__module__`` in instances of
``InterfaceClass`` is read-only.
Files: