On Tue, Apr 6, 2010 at 12:03 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Petr Jelinek <pjmodos@pjmodos.net> writes:
>> Dne 6.4.2010 7:57, Joseph Adams napsal(a):
>>> To me, the most logical approach is to do the obvious thing: make
>>> JSON's 'null' be SQL's NULL. For instance, SELECTing on a table with
>>> NULLs in it and converting the result set to JSON would yield a
>>> structure with 'null's in it. 'null'::JSON would yield NULL. I'm not
>>> sure what startling results would come of this approach, but I'm
>>> guessing this would be most intuitive and useful.
>
>> +1
>
> I think it's a pretty bad idea for 'null'::JSON to yield NULL. AFAIR
> there is no other standard datatype for which the input converter can
> yield NULL from a non-null input string, and I'm not even sure that the
> InputFunctionCall protocol allows it. (In fact a quick look indicates
> that it doesn't...)
Oh. I missed this aspect of the proposal. I agree - that's a bad idea.
> To me, what this throws into question is not so much whether JSON null
> should equate to SQL NULL (it should), but whether it's sane to accept
> atomic values.
With this, I disagree. I see no reason to suppose that a JSON NULL
and an SQL NULL are the same thing.
> If I understood the beginning of this discussion, that's
> not strictly legal. I think it would be better for strict input mode
> to reject this, and permissive mode to convert it to a non-atomic value.
> Thus jsonify('null') wouldn't yield NULL but a structure containing a
> null.
There's no obvious "structure" to convert this into.
...Robert