[Python-ideas] Json object-level serializer

[Nick Coghlan]

On Thu, Jul 29, 2010 at 11:11 PM, Antoine Pitrou <solipsis at pitrou.net> wrote:
> On Thu, 29 Jul 2010 22:51:11 +1000
> Nick Coghlan <ncoghlan at gmail.com> wrote:
>>
>> Each individual time this question comes up people tend to react with
>> "oh, that's too complicated and overkill, but magic methods are
>> simple, so let's just define another magic method". The sum total of
>> all those magic methods starts to accumulate into a lot of complexity
>> of its own though :P
>
> I don't agree. __json__ only matters to people who do JSON
> encoding/decoding. Other people can safely ignore it.

Which is exactly the attitude I was talking about: for each individual
case, people go "oh, I understand magic methods, those are easy". It's
the overall process of identifying the need for and gathering
consensus on magic methods that is unwieldy (and ultimately fails to
scale, leading to non-extensible interfaces by default, with pretty
printing being the classic example, and JSON serialisation the
latest).

> And I don't see how generic functions bring less cognitive overhead.
> (they actually bring more of it, since most implementations are more
> complicated to begin with)

Mostly because the fully fledged generic implementations like
PEAK-rules tend to get brought into discussions when they aren't
needed. Single-type generic dispatch is actually so common they gave
it a name: object-oriented programming. All single-type generic
dispatch is about is having a registry for a particular operation that
says "to perform this operation, with objects of this type, use this
function". Instead of having a protocol that says "look up this magic
method in the object's own namespace" (which requires a) agreement on
the magic name to use and b) that the original author of the type in
question both knew and cared about the operation the application
developer is interested in) you instead have a protocol that says
"here is a standard mechanism for declaring a type registry for a
function, so you only have to learn how to register a function once".

Is it really harder for people to learn how to write things like:

    json.dumps.overload(mytype, mytype.to_json)
    json.dumps.overload(third_party_type, my_third_party_type_serialiser)

than it is for them to figure out that implementing a __json__ method
will allow them to change how their object is serialised? (Not to
mention that a __json__ method can only be used via monkey-patching if
the type you want to serialise differently came from a library module
rather than your own code).

The generic function registration approach is incidentally
discoverable via dir(json.dumps) to see that a function provides the
relevant generic function registration methods. Magic method protocols
can *only* be discovered by reading documentation.

Function registration is a solved problem, with much better solutions
than the ad hoc YAMM (yet-another-magic-method) approach we currently
use. We just keep getting scared away from the right answer by the
crazily complex overloading schemes that libraries like PEAK-rules
allow.

Cheers,
Nick.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia