v4.0.2
======
    * The unpickler is now more resilient to malformed "py/id" and "py/repr" data.
      (+546)
    * The unpickler is now more resilient to invalid "py/b85" and "py/b64" data.
      (+547)
    * The unpickler's support for read-only str attributes was improved.
      (+548) (#478)

v4.0.1
======
    * The unpickler is now more resilient to malformed "py/reduce", "py/set",
      "py/tuple", "py/b64", "py/b85", and "py/iterator" input data. (+544) (+545)
    * The test suite was updated to leverage more pytest features.
    * The ``jsonpickle.compat`` module is no longer used. It is still provided
      for backwards compatibility but it may be removed in a future version.

v4.0.0
======
    * **Breaking Change**: Python 3.7 is no longer supported.
    * **Breaking Change**: Support for pre-0.7.0 ``repr``-serialized objects is no
      longer enabled by default. The ``safe`` option to ``decode()`` was changed from
      ``False`` to ``True``. Users can still pass ``safe=False`` to ``decode()`` in order
      to enable this feature for the purposes of loading older files, but beware that
      this feature relies on unsafe behavior through its use of ``eval()``. Users are
      encouraged to re-pickle old data in order to migrate away from the the unsafe loading
      feature. (+514)
    * The pickler no longer produces ``py/repr`` tags when pickling modules.
      ``py/mod`` is used instead, as it is clearer and uses one less byte. (+514)
    * The test suite no longer uses the deprecated ``datetime.datetime.utcnow()``
      function. (+539)

v3.4.2
======
    * The breaking changes from v4 were inadvertedly included in v3.4.1, which has
      been yanked. This release remedies this by reverting the v4 changes.

v3.4.1
======
    * Support decoding pandas dataframes encoded with versions 3.3.0 and older. (+536)

v3.4.0
======
    * Officially support Python 3.12 in the GitHub Actions testing matrix, and update
      GHA package versions used. (+524)
    * Improve reproducibility of benchmarking commands on Linux by using taskset and
      adding a "HOWTO" run benchmarks section in ``benchmarking/README.md``. (+526)
    * The ``setup.cfg`` packaging configuration has been replaced by
      ``pyproject.toml``. (+527)
    * ``yaml`` is now supported as a jsonpickle backend. (+528)
    * `OSSFuzz <https://github.com/google/oss-fuzz>`_ scripts are now available in
      the ``fuzzing/`` directory. (+525)
    * Pure-python dtypes are now preserved across ``encode()``/``decode()`` roundtrips
      for the pandas extension. (#407) (+534)
    * Pandas dataframe columns with an ``object`` dtype that contain multiple different
      types within (e.g. a column of type ``list[Union[str, int]]``) now preserve the types
      upon being roundtripped. (#457) (#358) (+534)
    * Fix warnings in the test suite regarding numpy.compat usage. (#533) (+535)

v3.3.0
======
    * The unpickler was updated to avoid using ``eval``, which helps improve its
      security. Users can still pass ``safe=False`` to ``decode`` to use the old
      behavior, though this is not recommended. (+513)
    * Objects can now exclude specific attributes from pickling by providing a
      ``_jsonpickle_exclude`` class or instance attribute. This attribute should contain
      the list of attribute names to exclude when pickling the object.

v3.2.2
======
    * A bug with the incorrect (de)serialization of NoneType objects has been fixed.
      (+507)
    * ``tests/benchmark.py`` was updated to avoid Python 2 syntax. (+508)
    * The unpickler was updated to avoid creating temporary functions. (+508)
    * Some basic scripts have been made to analyze benchmark results. (+511)
    * Fix test suite compatibility with Numpy 2.x (+512)
    * `setup.cfg` was updated to use `license_files` instead of `license_file`.

v3.2.1
======
    * The ``ignorereserved`` parameter to the private ``_restore_from_dict()``
      function has been restored for backwards compatibility. (+501)

v3.2.0
======
    * Nested dictionaries in `py/state` are now correctly restored when
      tracking object references. (+501) (#500)

v3.1.0
======
    * `jsonpickle.ext.numpy.register_handlers` now provides options that are forwarded
      to the `NumpyNDArrayHandler` constructor. (+489)
    * Fix bug of not handling ``classes`` argument to `jsonpickle.decode`
      being a dict. Previously, the keys were ignored and only values were
      used. (+494)
    * Allow the ``classes`` argument to `jsonpickle.pickle` to have class
      objects as keys. This extends the current functionality of only having
      class name strings as keys. (+494)
    * The ``garden setup/dev`` action and ``requirements-dev.txt`` requirements file
      now include test dependencies for use during development.
    * Added support for Python 3.13. (+505) (#504)

v3.0.4
======
    * Fixed an issue with django.SafeString and other classes inheriting from
      str having read-only attribute errors (#478) (+481)
    * The test suite was made compatible with `pytest-ruff>=0.3.0`. (+482)
    * A `garden.yaml` file was added for use with the
      `garden <https://crates.io/crates/garden-tools>_` command runner. (+486)
    * The test suite was updated to avoid deprecated SQLALchemy APIs.
    * The `jaraco.packaging.sphinx` documentation dependency was removed.

v3.0.3
======
    * Compatibilty with Pandas and Cython 3.0 was added. (#460) (+477)
    * Fixed a bug where pickling some built-in classes (e.g. zoneinfo) 
      could return a ``None`` module. (#447)
    * Fixed a bug where unpickling a missing class would return a different object
      instead of ``None``. (+471)
    * Fixed the handling of missing classes when setting ``on_missing`` to ``warn``
      or ``error``. (+471)
    * The test suite was made compatible with Python 3.12.
    * The tox configuration was updated to generate code coverage reports.
    * The suite now uses ``ruff`` to validate python code.
    * The documentation can now be built offline when ``rst.linker`` and
      ``jaraco.packaging.sphinx`` are not available.

v3.0.2
======
    * Properly raise warning if a custom pickling handler returns None. (#433)
    * Fix issue with serialization of certain sklearn objects breaking when
      the numpy handler was enabled. (#431) (+434)
    * Allow custom backends to not implement _encoder_options (#436) (+446)
    * Implement compatibility with pandas 2 (+446)
    * Fix encoding/decoding of dictionary subclasses with referencing (+455)
    * Fix depth tracking for list/dict referencing (+456)

v3.0.1
======
    * Remove accidental pin of setuptools to versions below 59. This allows
      jsonpickle to build with CPython 3.11 and 3.12 alphas. (#424)
    * Remove accidental dependency on pytz in pandas tests. (+421)
    * Fix issue with decoding bson.bson.Int64 objects (#422)

v3.0.0
======
    * Drop support for CPython<3.7. CPython 3.6 and below have reached EOL
      and no longer receive security updates. (#375)
    * Add support for CPython 3.11. (#395) (+396)
    * Remove jsonlib and yajl backends (py2 only)
    * Add ``include_properties`` option to the pickler. This should only
      be used if analyzing generated json outside of Python. (#297) (+387)
    * Allow the ``classes`` argument to ``jsonpickle.decode`` to be a dict
      of class name to class object. This lets you decode arbitrary dumps
      into different classes. (#148) (+392)
    * Fix bug with deserializing `numpy.poly1d`. (#391)
    * Allow frozen dataclasses to be deserialized. (#240)
    * Fixed a bug where pickling a function could return a ``None`` module. (#399)
    * Removed old bytes/quopri and ref decoding abaility from the unpickler.
      These were last used in jsonpickle<1. Removing them causes a slight speedup
      in unpickling lists (~5%). (+403)
    * Fixed a bug with namedtuples encoding on CPython 3.11. (#411)
    * When using the ``sort_keys`` option for the ``simplejson`` backend,
      jsonpickle now produces correct object references with py/id tags. (#408)
    * Speed up the internal method ``_restore_tags`` by ~10%. This should speed
      up unpickling of almost every object.

v2.2.0
======

    * Classes with a custom ``__getitem__()`` and ``append()``
      now pickle properly. (#362) (+379)
    * Remove the demjson backend, as demjson hasn't been maintained
      for 5 years. (+379)
    * Added new handler for numpy objects when using unpickleable=False.
      (#381) (+382)
    * Added exception handling for class attributes that can't be accessed.
      (#301) (+383)
    * Added a long-requested on_missing attribute to the Unpickler class.
      This lets you choose behavior for when jsonpickle can't find a class
      to deserialize to. (#190) (#193) (+384)
    * Private members of ``__slots__`` are no longer skipped when encoding.
      Any objects encoded with versions prior to 2.2.0 should still decode
      properly. (#318) (+385)

v2.1.0
======

    * Python 3.10 is now officially supported. (+376)
    * Benchmarks were added to aid in optimization efforts.  (#350) (+352)
    * ``is_reducible()`` was sped up by ~80%.  (+353) (+354)
    * ``_restore_tags()`` was sped up by ~100%. Unpickling items
      with a lot of tuples and sets will benefit most. Python 2 users
      and users deserializing pickles from jsonpickle <= 0.9.6 may see
      a slight performance decrease if using a lot of bytes, ref,
      and/or repr objects. (+354)
    * ``is_iterator()`` was sped up by ~20% by removing an unnecessary
      variable assignment. (+354)
    * ``jsonpickle.decode`` has a new option, ``v1_decode`` to assist in
      decoding objects created in jsonpickle version 1. (#364)
    * The ``encode()`` documentation has been updated to help sklearn users.
    * ``demjson`` has been removed from the test suite. (+374)
    * ``SQLALchemy<1.2`` is no longer being tested by jsonpickle.
      Users of sqlalchemy + jsonpickle can always use 1.2 or 1.3.
      When jsonpickle v3 is released we will add SQLAlchemy 1.4 to
      the test suite alongside removal of support for Python 3.5 and earlier.

v2.0.0
======
    * Major release: the serialized JSON format now preserves dictionary
      identity, which is a subtle change in the serialized format.  (#351)
    * Dictionary identity is now preserved.  For example, if the same
      dictionary appears twice in a list, the reconstituted list
      will now contain two references to the same dictionary.  (#255) (+332)

v1.5.2
======
    * Patch release to avoid the change in behavior from the preservation
      of dict identity.  The next release will be v2.0.0.  (#351)
    * This release does *not* include the performance improvements
      from v1.5.1.
    * Pandas DataFrame objects with multilevel columns are now supported.
      (#346) (+347)
    * Numpy 1.20 is now officially supported.  (#336)
    * Python 3.9 is now officially supported.  (+348)
    * Achieved a small speedup for _get_flattener by merging type checks. (+349)

v1.5.1
======
    * The performance of the unpickler was drastically improved by
      avoiding tag checks for basic Python types.  (+340)
    * ``decode()`` documentation improvements.  (+341)
    * Serialization of Pandas DataFrame objects that contain
      timedelta64[ns] dtypes are now supported.  (+330) (#331)
    * Dictionary identity is now preserved.  For example, if the same
      dictionary appears twice in a list, the reconstituted list
      will now contain two references to the same dictionary.  (#255) (+332)
    * Unit tests were added to ensure that sklearn.tree.DecisionTreeClassifier
      objects are properly serialized.  (#155) (+344)
    * The ``is_reducible()`` utility function used by ``encode()`` is now
      4x faster!  Objects that provide ``__getstate__()``, ``__setstate__()``,
      and ``__slots__`` benefit most from these improvements.  (+343)
    * Improved pickler ``flatten()/encode()`` performance.  (+345)

v1.5.0
======
    * Previous versions of jsonpickle with `make_refs=False` would emit
      ``null`` when encountering an object it had already seen when
      traversing objects.  All instances of the object are now serialized.
      While this is arguably an improvement in the vast majority of
      scenarios, it is a change in behavior and is thus considered a
      minor-level change.  (#333) (#334) (#337) (+338)
    * Multiple enums are now serialized correctly with `make_refs=False`.  (#235)

v1.4.2
======
    * Use importlib.metadata from the stdlib on Python 3.8.  (+305) (#303)
    * Micro-optimize type checks to use a `set` for lookups. (+327)
    * Documentation improvements.

v1.4.1
======
    * Patch release for Python 3.8 `importlib_metadata` support.
      (#300)

v1.4
====
    * Python 3.8 support.  (#292)
    * ``jsonpickle.encode`` now supports the standard ``indent``
      and ``separators`` arguments, and passes them through to the
      active JSON backend library.  (#183)
    * We now include a custom handler for `array.array` objects.  (#199)
    * Dict key order is preserved when pickling dictionaries on Python3.  (#193)
    * Improved serialization of dictionaries with non-string keys.
      Previously, using an enum that was both the key and a value in
      a dictionary could end up with incorrect references to other
      objects.  The references are now properly maintained for dicts
      with object keys that are also referenced in the dict's values.  (#286)
    * Improved serialization of pandas.Series objects.  (#287)

v1.3
====
    * Improved round tripping of default dicts.  (+283) (#282)

    * Better support for cyclical references when encoding with
      ``unpicklable=False``.  (+264)

v1.2
====
    * Simplified JSON representation for `__reduce__` values.  (+261)

    * Improved Pandas support with new handlers for more Pandas data types.
      (+256)

    * Prevent stack overflows caused by bugs in user-defined `__getstate__`
      functions which cause infinite recursion.  (+260)
      (#259)

    * Improved support for objects that contain dicts with Integer keys.
      Previously, jsonpickle could not restore objects that contained
      dicts with integer keys and provided getstate only.
      These objects are now handled robustly.  (#247).

    * Support for encoding binary data in `base85`_ instead of base64 has been
      added on Python 3. Base85 produces payloads about 10% smaller than base64,
      albeit at the cost of lower throughput.  For performance and backwards
      compatibility with Python 2 the pickler uses base64 by default, but it can
      be configured to use ``base85`` with the new ``use_base85`` argument.
      (#251).

    * Dynamic SQLAlchemy tables in SQLAlchemy >= 1.3 are now supported.
      (#254).

.. _base85: https://en.wikipedia.org/wiki/Ascii85


v1.1
====
    * Python 3.7 `collections.Iterator` deprecation warnings have been fixed.
      (#229).

    * Improved Pandas support for datetime and complex numbers.  (+245)

v1.0
====
    * *NOTE* jsonpickle no longer supports Python2.6, or Python3 < 3.4.
      The officially supported Python versions are now 2.7 and 3.4+.

    * Improved Pandas and Numpy support.  (+227)

    * Improved support for pickling iterators.  (+216)

    * Better support for the stdlib `json` module when `simplejson`
      is not installed.  (+217)

    * jsonpickle will now output python3-style module names when
      pickling builtins methods or functions.  (+223)

    * jsonpickle will always flatten primitives, even when ``max_depth``
      is reached, which avoids encoding unicode strings into their
      ``u'string'`` representation.  (+207) (#180) (#198).

    * Nested classes are now supported on Python 3.  (+206) (#176).

    * Better support for older (pre-1.9) versions of numpy (+195).

v0.9.6
======
    * Better support for SQLAlchemy (#180).

    * Better support for NumPy and SciKit-Learn.  (#184).

    * Better support for dict sub-classes (#156).

v0.9.5
======
    * Better support for objects that implement the reduce protocol.  (+170)
      This backward-incompatible change removes the SimpleReduceHandler.
      Any projects registering that handler for a particular type should
      instead remove references to the handler and jsonpickle will now
      handle those types directly.

v0.9.4
======
    * Arbitrary byte streams are now better supported.  (#143)

    * Better support for NumPy data types.  The Python3 NumPy support
      is especially robust.

    * Fortran-ordered based NumPy arrays are now properly serialized.

v0.9.3
======
    * UUID objects can now be serialized (#130)

    * Added `set_decoder_options` method to allow decoder specific options
      equal to `set_encoder_options`.

    * Int keys can be encoded directly by e.g. demjson by passing
      `numeric_keys=True` and setting its backend options via
      `jsonpickle.set_encoder_options('demjson', strict=False)`.

    * Newer Numpy versions (v1.10+) are now supported.

v0.9.2
======
    * Fixes for serializing objects with custom handlers.

    * We now properly serialize deque objects constructed with a `maxlen` parameter.

    * Test suite fixes

v0.9.1
======

    * Support datetime objects with FixedOffsets.

v0.9.0
======
    * Support for Pickle Protocol v4.

    * We now support serializing defaultdict subclasses that use `self`
      as their default factory.

    * We now have a decorator syntax for registering custom handlers,
      and allow custom handlers to register themselves for all subclasses.
      (+104)

    * We now support serializing types with metaclasses and their
      instances (e.g., Python 3 `enum`).

    * We now support serializing bytestrings in both Python 2 and Python 3.
      In Python 2, the `str` type is decoded to UTF-8 whenever possible and
      serialized as a true bytestring elsewise; in Python 3, bytestrings
      are explicitly encoded/decoded as bytestrings. Unicode strings are
      always encoded as is in both Python 2 and Python 3.

    * Added support for serializing numpy arrays, dtypes and scalars
      (see `jsonpickle.ext.numpy` module).

v0.8.0
======

    * We now support serializing objects that contain references to
      module-level functions.  (#77)

    * Better Pickle Protocol v2 support.  (#78)

    * Support for string ``__slots__`` and iterable ``__slots__``. (#67) (#68)

    * `encode()` now has a `warn` option that makes jsonpickle emit warnings
      when encountering objects that cannot be pickled.

    * A Javascript implementation of jsonpickle is now included
      in the jsonpickleJS directory.

v0.7.2
======

    * We now properly serialize classes that inherit from classes
      that use `__slots__` and add additional slots in the derived class.
    * jsonpickle can now serialize objects that implement `__getstate__()` but
      not `__setstate__()`.  The result of `__getstate__()` is returned as-is
      when doing a round-trip from Python objects to jsonpickle and back.
    * Better support for collections.defaultdict with custom factories.
    * Added support for `queue.Queue` objects.

v0.7.1
======

    * Added support for Python 3.4.
    * Added support for `posix.stat_result`.

v0.7.0
======

    * Added ``handles`` decorator to :class:`jsonpickle.handlers.BaseHandler`,
      enabling simple declaration of a handler for a class.
    * `__getstate__()` and `__setstate__()` are now honored
      when pickling objects that subclass :class:`dict`.
    * jsonpickle can now serialize :class:`collections.Counter` objects.
    * Object references are properly handled when using integer keys.
    * Object references are now supported when using custom handlers.
    * Decimal objects are supported in Python 3.
    * jsonpickle's "fallthrough-on-error" behavior can now be disabled.
    * Simpler API for registering custom handlers.
    * A new "safe-mode" is provided which avoids eval().
      Backwards-compatible deserialization of repr-serialized objects
      is disabled in this mode.  e.g. `decode(string, safe=True)`

v0.6.1
======

    * Python 3.2 support, and additional fixes for Python 3.

v0.6.0
======

    * Python 3 support!
    * :class:`time.struct_time` is now serialized using the built-in
      `jsonpickle.handlers.SimpleReduceHandler`.

v0.5.0
======

    * Non-string dictionary keys (e.g. ints, objects) are now supported
      by passing `keys=True` to :func:`jsonpickle.encode` and
      :func:`jsonpickle.decode`.
    * We now support namedtuple, deque, and defaultdict.
    * Datetimes with timezones are now fully supported.
    * Better support for complicated structures e.g.
      datetime inside dicts.
    * jsonpickle added support for references and cyclical data structures
      in 0.4.0.  This can be disabled by passing `make_refs=False` to
      :func:`jsonpickle.encode`.

0.4.0
=====

    * Switch build from setuptools to distutils
    * Consistent dictionary key ordering
    * Fix areas with improper support for unpicklable=False
    * Added support for cyclical data structures
      (#16).
    * Experimental support for  `jsonlib <http://pypi.python.org/pypi/jsonlib/>`_
      and `py-yajl <http://github.com/rtyler/py-yajl/>`_ backends.
    * New contributors David K. Hess and Alec Thomas

    .. warning::

        To support cyclical data structures
        (#16),
        the storage format has been modified.  Efforts have been made to
        ensure backwards-compatibility.  jsonpickle 0.4.0 can read data
        encoded by jsonpickle 0.3.1, but earlier versions of jsonpickle may be
        unable to read data encoded by jsonpickle 0.4.0.


0.3.1
=====

    * Include tests and docs directories in sdist for distribution packages.

0.3.0
=====

    * Officially migrated to git from subversion. Project home now at
      `<http://jsonpickle.github.com/>`_. Thanks to Michael Jone's
      `sphinx-to-github <http://github.com/michaeljones/sphinx-to-github>`_.
    * Fortified jsonpickle against common error conditions.
    * Added support for:

     * List and set subclasses.
     * Objects with module references.
     * Newstyle classes with `__slots__`.
     * Objects implementing `__setstate__()` and `__getstate__()`
       (follows the :mod:`pickle` protocol).

    * Improved support for Zope objects via pre-fetch.
    * Support for user-defined serialization handlers via the
      jsonpickle.handlers registry.
    * Removed cjson support per John Millikin's recommendation.
    * General improvements to style, including :pep:`257` compliance and
      refactored project layout.
    * Steps towards Python 2.3 and Python 3 support.
    * New contributors Dan Buch and Ian Schenck.
    * Thanks also to Kieran Darcy, Eoghan Murray, and Antonin Hildebrand
      for their assistance!

0.2.0
=====

    * Support for all major Python JSON backends (including json in Python 2.6,
      simplejson, cjson, and demjson)
    * Handle several datetime objects using the repr() of the objects
      (Thanks to Antonin Hildebrand).
    * Sphinx documentation
    * Added support for recursive data structures
    * Unicode dict-keys support
    * Support for Google App Engine and Django
    * Tons of additional testing and bug reports (Antonin Hildebrand, Sorin,
      Roberto Saccon, Faber Fedor,
      `FirePython <http://github.com/darwin/firepython/tree/master>`_, and
      `Joose <http://code.google.com/p/joose-js/>`_)

0.1.0
=====

    * Added long as basic primitive (thanks Adam Fisk)
    * Prefer python-cjson to simplejson, if available
    * Major API change, use python-cjson's decode/encode instead of
      simplejson's load/loads/dump/dumps
    * Added benchmark.py to compare simplejson and python-cjson

0.0.5
=====

    * Changed prefix of special fields to conform with CouchDB
      requirements (Thanks Dean Landolt). Break backwards compatibility.
    * Moved to Google Code subversion
    * Fixed unit test imports

0.0.3
=====

    * Convert back to setup.py from pavement.py (issue found by spidaman)

0.0.2
=====

    * Handle feedparser's FeedParserDict
    * Converted project to Paver
    * Restructured directories
    * Increase test coverage

0.0.1
=====

    Initial release