[sevensor]

My complaint about JSON is that it’s not minimal enough. The receiver always has to validate anyway, so what has syntax typing done for us? Different implementations of JSON disagree about what constitutes a valid value. For instance, is

    {“x”: NaN}

valid JSON? How about 9007199254740993? Or -.053? If so, will that text round trip through your JSON library without loss of precision? Is that desirable if it does?

Basically I think formats with syntax typed primitives always run into this problem: even if the encoder and decoder are consistent with each other about what the values are, the receiver still has to decide whether it can use the result. This after all is the main benefit of a library like Pydantic. But if we’re doing all this work to make sure the object is correct, we know what the value types are supposed to be on the receiving end, so why are we making a needlessly complex decoder guess for us?

[aidenn0]

NaN is not a valid value in JSON. Neither are 0123 or .123 (there must always be at least one digit before the decimal marker, but extraneous leading zeroes are disallowed).

JSON was originally parsed in javascript with eval() which allowed many things that aren't JSON through, but that doesn't make JSON more complex.

[sevensor]

That’s my point, though! I’ve run into popular JSON libraries that will emit all of those! 9007199254740993 is problematic because it’s not representable as a 64 bit float. Python’s JSON library is happy to write it, even though you need an int to represent it, and JSON doesn’t have ints.

Edit: I didn’t see my thought all the way through here. Syntax typing invites this kind of nonconformity, because different programming languages mean different things by “number,” “string,” “date,” or even “null.” They will bend the format to match their own semantics, resulting in incompatibility.

[dragonwriter]

> 9007199254740993 is problematic because it’s not representable as a 64 bit float. Python’s JSON library is happy to write it, even though you need an int to represent it

JSON numbers have unlimited range in terms of the format standard, but implementations are explicitly permitted to set limits on the range and precision they generate and handle, and users are warned that:

   [...] Since software that implements IEEE 754 binary64 (double precision)
   numbers is generally available and widely used, good interoperability can be 
   achieved by implementations that expect no more precision or range than these
   provide, in the sense that implementations will approximate JSON
   numbers within the expected precision.

Also, you don't need an int to represent it (a wide enough int will represent it, so will unlimited precision decimals, wide enough binary floats -- of standard formats, IEEE 754 binary128 works -- etc.).

[sevensor]

RFC 8259 is a good read and I wish more people would make the effort. I really don’t mean to bash JSON here. It was a great idea and it continues to be a great idea, especially if you are using javascript. However, the passage you quote illustrates the same shortcoming I’m complaining about: RFC 8259 basically says “valid primitive types in json are the valid primitive types in your programming language,” but this results in implementations like Python’s json library emitting invalid tokens like bare NaN, which can cause decoders to choke.

I think what JSON gets right is that it gives us a universal way of expressing structure: arrays and objects map onto basic notions of sequence and association that are useful in many contexts and can be represented in a variety of ways by programming languages. My ideal data interchange format would stop there and let the user decide what to do with the value text after the structure has been decoded.

[aidenn0]

Before your edit, I was going to object to your premise because it seems like a format could get worse just by more implementations being made.

After your edit, I see that it's rather that syntax-typed formats are prone to this form of implementation divergence.

I don't think this is limited to syntax-typed formats though. For example, TNetstrings[1] have type tags, but "#" is an integer. The specification requires that integers fit into 63 bits (since the reference encoder will refuse to encode a python long), but implementations in C tend to allow 64 bits and in other languages allow bignums. It does explicitly allow "nan", "inf", and "-inf" FWIW.

1: https://tnetstrings.info/

[sevensor]

Agreed; I think there’s a problem with self-describing data as a concept. It just begs for implementation defined weirdness.

[conartist6]

Yeah I would emit NaN and just hope the receiver handles it.

What's the point of lying about the data?

The format offers you no data type that would not be an outright lie when applied to this data, so you may as well not lie and break the format

[sevensor]

Your other option is to comply with the spec, emit a string, and expect the receiver to deal with that at validation time rather than parse time.