Re: Bugs in guile-json builder

[Aleix Conchillo Flaqué]

Hi,

On Wed, Jan 2, 2019 at 7:14 AM John Cowan <address@hidden> wrote:
>
> I'm not sure where best to report these, so sending this email.
>

Same as before, github issues should work.

> JSON numbers have to be, in Scheme terms, either exact integers or inexact 
> reals excluding +inf.0, -inf.0, and +nan.0.  Guile-json will blithely output 
> any of these three as well as 1/2, 3.4+5.6i, etc., which will not work with 
> other JSON parsers.
>

Good catch!

> Symbols are currently output as JSON strings.  This will be problematic if 
> anyone depends on it and you decide to switch to the symbol null for JSON 
> null, as I suggested in my previous email.  I think symbols should be invalid 
> as JSON values.
>

Not sure about this one. I think it's nice to use symbols not sure if
it's worth to remove support of symbols for just this case. There must
be something else we can do. I'll investigate.

> Symbols and numbers are allowed as object keys, converting them to strings, 
> which is reasonable but should be documented.
>

Makes sense.

> If you pass anything but a string, a symbol, or a number as an object key, 
> you get a low-level error from string->list instead of a proper json-invalid 
> error.
>

Thanks!

> The builder doesn't validate the native object first.  This requires an extra 
> tree walk, but for example if you call (json-scm (vector 1 2 3 #u8(1 2 3))) 
> it will output "[1,2,3" before throwing an exception.  This is not the Right 
> Thing: a procedure that does output should work correctly or fail cleanly 
> without outputting anything.
>

Agreed. There's should be some validation before printing anything.

Again, if you could file these issues separately on github it would be
fantastic! We can follow discussions, patches, etc. there.

Thanks,

Aleix