On 08/29/2015 12:02 PM, Pavel Stehule wrote:
>
>
> 2015-08-29 15:43 GMT+02:00 Shulgin, Oleksandr
> <oleksandr.shulgin@zalando.de <mailto:oleksandr.shulgin@zalando.de>>:
>
> On Sat, Aug 29, 2015 at 3:39 PM, Tom Lane <tgl@sss.pgh.pa.us
> <mailto:tgl@sss.pgh.pa.us>> wrote:
>
> Andrew Dunstan <andrew@dunslane.net
> <mailto:andrew@dunslane.net>> writes:
> > On 08/29/2015 08:47 AM, Shulgin, Oleksandr wrote:
> >> Given there were no loud complaints about this, the current behavior
> >> is appropriate for most users, the rest can still work
> around using
> >> coalesce(to_json(...), json 'null').
>
> > I don't think it's necessarily more correct. But I do agree
> that it's
> > not a good idea to change the behaviour unless there is major
> > unhappiness with it.
>
> I'm not entirely convinced that JSON NULL and SQL NULL should
> be treated
> as the same concept, so I would say that the current behavior
> is fine ---
> at least when you think about it in isolation. However,
> haven't we
> already bought into that equivalence in these examples?
>
> regression=# select row_to_json(row(1,null,2));
> row_to_json
> ---------------------------
> {"f1":1,"f2":null,"f3":2}
> (1 row)
>
> regression=# select array_to_json(array[1,null,2]);
> array_to_json
> ---------------
> [1,null,2]
> (1 row)
>
> or even in to_json itself:
>
> regression=# select to_json(array[1,null,2]);
> to_json
> ------------
> [1,null,2]
> (1 row)
>
> The scalar case is definitely failing to be consistent with these.
>
>
> Yes, that's my argument for correctness also: to_json() on a
> composite object should behave like distribution of to_json()
> calls over object/array elements.
>
> Is consistency a sufficient reason to change it?
>
>
> Not for me.
>
>
> It is bug - and it should be fixed. I agree, so this change is too
> strong for fixing in minor version - but we can change it in
> unreleased major versions - 9.5 and master.
No, frankly that's being far too free with the word bug. It's not even
unambiguously incorrect.
Note that all the to_json functions are strict. In this sense it's quite
consistent. If we change it to being called on null input, what should
we return if a null non-scalar is passed in?
cheers
andrew