If you subclass json.JSONEncoder to enable serialisation of custom types beyond those supported, you are meant to transform values of said type into a serialisable version within an overridden default method. For example:

    class MyJSONEncoder(json.JSONEncoder):
        def default(self, o):
            if isinstance(o, MyType):
                return str(o)

        # Raise TypeError when we have a type we can't serialise
        super().default(o)

This, however, won't work if your custom type is an instance of one of the supported types. This is because, in Lib/json/encoder.py, the _iterencode function (defined in _make_iterencode) type checks against supported types before it delegates to the default method.

The reason this came up is because I wanted to serialise a named tuple into a JSON object, with keys corresponding to the named tuple's field names. Ignoring the merits (or otherwise) of this desired outcome, this can't work because a named tuple is still a tuple and thus the default method is never called.

In Python 2.7, the _iterencode method was part of the JSONEncoder class, so you could override it in a subclass; even if doing so is somewhat brittle. In Python 3, this method has been moved out into the module namespace (tested in 3.6; looking through the repo, it looks like this change was made in Python 3.1), so it can't easily be monkey-patched without it affecting other things.

I believe this to be a bug. It seems reasonable to subclass at least named tuples, dictionaries and lists in such a way that you'd want a different JSON serialisation to their defaults.

I have written a proof-of-concept implementation and submitted a pull request. See the PR for details: #1558

(My CLA is pending; submitted around 2017-05-12T10:30:00Z+0100)

See also https://bugs.python.org/issue34858

This needs a discussion on python-ideas or -dev.