Making object JSON serializable with regular encoder

As I said in a comment to your question, after looking at the json module’s source code, it does not appear to lend itself to doing what you want. However the goal could be achieved by what is known as monkey-patching
(see question What is a monkey patch?).
This could be done in your package’s __init__.py initialization script and would affect all subsequent json module serialization since modules are generally only loaded once and the result is cached in sys.modules.

The patch changes the default json encoder’s default method—the default default().

Here’s an example implemented as a standalone module for simplicity’s sake:

Module: make_json_serializable.py

""" Module that monkey-patches json module when it's imported so
JSONEncoder.default() automatically checks for a special "to_json()"
method and uses it to encode the object if found.
"""
from json import JSONEncoder

def _default(self, obj):
    return getattr(obj.__class__, "to_json", _default.default)(obj)

_default.default = JSONEncoder.default  # Save unmodified default.
JSONEncoder.default = _default # Replace it.

Using it is trivial since the patch is applied by simply importing the module.

Sample client script:

import json
import make_json_serializable  # apply monkey-patch

class Foo(object):
    def __init__(self, name):
        self.name = name
    def to_json(self):  # New special method.
        """ Convert to JSON format string representation. """
        return '{"name": "%s"}' % self.name

foo = Foo('sazpaz')
print(json.dumps(foo))  # -> "{\"name\": \"sazpaz\"}"

To retain the object type information, the special method can also include it in the string returned:

        return ('{"type": "%s", "name": "%s"}' %
                 (self.__class__.__name__, self.name))

Which produces the following JSON that now includes the class name:

"{\"type\": \"Foo\", \"name\": \"sazpaz\"}"

Magick Lies Here

Even better than having the replacement default() look for a specially named method, would be for it to be able to serialize most Python objects automatically, including user-defined class instances, without needing to add a special method. After researching a number of alternatives, the following — based on an answer by @Raymond Hettinger to another question — which uses the pickle module, seemed closest to that ideal to me:

Module: make_json_serializable2.py

""" Module that imports the json module and monkey-patches it so
JSONEncoder.default() automatically pickles any Python objects
encountered that aren't standard JSON data types.
"""
from json import JSONEncoder
import pickle

def _default(self, obj):
    return {'_python_object': pickle.dumps(obj)}

JSONEncoder.default = _default  # Replace with the above.

Of course everything can’t be pickled—extension types for example. However there are ways defined to handle them via the pickle protocol by writing special methods—similar to what you suggested and I described earlier—but doing that would likely be necessary for a far fewer number of cases.

Deserializing

Regardless, using the pickle protocol also means it would be fairly easy to reconstruct the original Python object by providing a custom object_hook function argument on any json.loads() calls that used any '_python_object' key in the dictionary passed in, whenever it has one. Something like:

def as_python_object(dct):
    try:
        return pickle.loads(str(dct['_python_object']))
    except KeyError:
        return dct

pyobj = json.loads(json_str, object_hook=as_python_object)

If this has to be done in many places, it might be worthwhile to define a wrapper function that automatically supplied the extra keyword argument:

json_pkloads = functools.partial(json.loads, object_hook=as_python_object)

pyobj = json_pkloads(json_str)

Naturally, this could be monkey-patched it into the json module as well, making the function the default object_hook (instead of None).

I got the idea for using pickle from an answer by Raymond Hettinger to another JSON serialization question, whom I consider exceptionally credible as well as an official source (as in Python core developer).

Portability to Python 3

The code above does not work as shown in Python 3 because json.dumps() returns a bytes object which the JSONEncoder can’t handle. However the approach is still valid. A simple way to workaround the issue is to latin1 “decode” the value returned from pickle.dumps() and then “encode” it from latin1 before passing it on to pickle.loads() in the as_python_object() function. This works because arbitrary binary strings are valid latin1 which can always be decoded to Unicode and then encoded back to the original string again (as pointed out in this answer by Sven Marnach).

^{(Although the following works fine in Python 2, the latin1 decoding and encoding it does is superfluous.)}

from decimal import Decimal

class PythonObjectEncoder(json.JSONEncoder):
    def default(self, obj):
        return {'_python_object': pickle.dumps(obj).decode('latin1')}


def as_python_object(dct):
    try:
        return pickle.loads(dct['_python_object'].encode('latin1'))
    except KeyError:
        return dct


class Foo(object):  # Some user-defined class.
    def __init__(self, name):
        self.name = name

    def __eq__(self, other):
        if type(other) is type(self):  # Instances of same class?
            return self.name == other.name
        return NotImplemented

    __hash__ = None


data = [1,2,3, set(['knights', 'who', 'say', 'ni']), {'key':'value'},
        Foo('Bar'), Decimal('3.141592653589793238462643383279502884197169')]
j = json.dumps(data, cls=PythonObjectEncoder, indent=4)
data2 = json.loads(j, object_hook=as_python_object)
assert data == data2  # both should be same