Comment by coldnebo
11 years ago
There is a lot of confusion between typing systems, metatyping systems (that can implement arbitrary type systems) and the transport representation of such systems.
I agree with your counterpoints, but the cool kids are still having issues with transport representation of arbitrary types. Sure (eg Ruby) can use Kernel dump and load to marshall arbitrary types, but what happens when the other end doesn't have the type loaded or available?
Ouch, maybe we should invite Java Enterprise Beans to the party to comment on the semantics of distributed types?
JSON is currently deceptively simple precisely because its wire representation (with simple types) is equivalent to its type definition which can be directly evaled in js to produce a memory instantiation. But none of that holds in the general case... Try for example marshalling a JSON object with a tree structure.
Maybe we end up going in Alan Kay's direction, where we stop trying to send representations across the wire and just send code... But that too was tried in the past. It has other limitations.
It's complicated.
>JSON is currently deceptively simple precisely because its wire representation (with simple types) is equivalent to its type definition which can be directly evaled in js to produce a memory instantiation.
It's not deceptively simple. It's just simple. The fact that it can be evaluated in JS is incidental to its use as a general purpose format for serializing data structures.
>Try for example marshalling a JSON object with a tree structure.
I've done this lots of times. I don't see any issue with it.
The core problem is definitely the support of custom types. I agree, if you refuse custom types everything gets a lot simpler.
Here's a very simple example of marshaling comparing Marshal dump and load in Ruby using YAML, vs. custom JSON marshalers: http://www.skorks.com/2010/04/serializing-and-deserializing-...
Note that the post shows a tree structure in YAML because Marshall gives it to you for free (synonymous with Java serialize behavior). But the post punts on implementing the same tree in JSON, probably because it's messy and complicated.
Nothing about that looks simple to me. For example, the JSON class label "A" has to be interpreted by something to actually instantiate an A Object. YAML is a bit better-- it at least defines that there is a Ruby class instance being marshaled -- but it doesn't help you if that class doesn't exist the same on client and server.
Pretty soon this leads to madness in production systems where the server evolves "A" (v2) among a sea of separately evolving clients "A" (v1, v1.1, v1.2). Then versioned class annotations get introduced, followed by attempts to differentiate between serializable and non-serializable fields, deep and shallow serialization, etc. etc. Pretty soon, your JSON marshaling isn't so simple anymore.
>The core problem is definitely the support of custom types. I agree, if you refuse custom types everything gets a lot simpler.
Which makes perfect sense. If you cut everything down to bool, null, number, string, list and map you can represent anything and you get to remain language agnostic.
Dates can be encoded as strings and so can most of the other more 'awkward' types. This is additional work, but it's not that complicated and not a lot is gained anyhow by putting this stuff in the spec.
You really can't get more complicated than this anyhow, without introducing nasty security vulnerabilities.
1 reply →
As an example of the deceptive simplicity, could you please describe how many bytes are required to deserialize numbers represented in JSON? Note: answer is not 2,4, or 8.
The JSON spec has no opinion on that and that's a good thing. It is up to you, your language and your parser to decide how large the largest JSON number should be.
Would it be simpler to enforce a global 32 bit or 64 bit limit on those numbers in the spec? I don't think so. Why should the limitations of embedded microcontrollers exchanging data apply to those of astronomical systems and vice versa?
I should have been more specific... it's not really about trees per se, it's more about custom types. (I was thinking of trees of custom types).
I'm a fan of the techniques as described in : http://smile.amazon.com/Expert-Business-Objects-Books-Profes...
Basically, it uses the underlying runtime to serialize the entire object to the other machine. The theory being, if object oriented programming ties data to functions, why not ship both over the wire?
It works really well with the micro-services paradigm.
Not exactly a 100% solution, but it does solve the issues at least among the languages the runtime supports.