Re: ON ERROR in json_query and the like

> On 17.06.2024, at 08:20, Amit Langote <amitlangote09(at)gmail(dot)com> wrote:
>
> Hi,
>
> (apologies for not replying to this thread sooner)
>
> On Tue, May 28, 2024 at 6:57 PM Pavel Stehule <pavel(dot)stehule(at)gmail(dot)com> wrote:
>> út 28. 5. 2024 v 11:29 odesílatel Markus Winand <markus(dot)winand(at)winand(dot)at> napsal:
>>>
>>> Hi!
>>>
>>> I’ve noticed two “surprising” (to me) behaviors related to
>>> the “ON ERROR” clause of the new JSON query functions in 17beta1.
>>>
>>> 1. JSON parsing errors are not subject to ON ERROR
>>> Apparently, the functions expect JSONB so that a cast is implied
>>> when providing TEXT. However, the errors during that cast are
>>> not subject to the ON ERROR clause.
>>>
>>> 17beta1=# SELECT JSON_QUERY('invalid', '$' NULL ON ERROR);
>>> ERROR: invalid input syntax for type json
>>> DETAIL: Token "invalid" is invalid.
>>> CONTEXT: JSON data, line 1: invalid
>>>
>>> Oracle DB and Db2 (LUW) both return NULL in that case.
>>>
>>> I had a look on the list archive to see if that is intentional but
>>> frankly speaking these functions came a long way. In case it is
>>> intentional it might be worth adding a note to the docs.
>>
>>
>> I remember a talk about this subject years ago. Originally the JSON_QUERY was designed in similar like Oracle, and casting to jsonb was done inside. If I remember this behave depends on the fact, so old SQL/JSON has not json type and it was based just on processing of plain text. But Postgres has JSON, and JSONB and then was more logical to use these types. And because the JSON_QUERY uses these types, and the casting is done before the execution of the function, then the clause ON ERROR cannot be handled. Moreover, until soft errors Postgres didn't allow handling input errors in common functions.
>>
>> I think so this difference should be mentioned in documentation.
>
> Agree that the documentation needs to be clear about this. I'll update
> my patch at [1] to add a note next to table 9.16.3. SQL/JSON Query
> Functions.

Considering another branch of this thread [1] I think the
"Supported Features” appendix of the docs should mention that as well.

The way I see it is that the standards defines two overloaded
JSON_QUERY functions, of which PostgreSQL will support only one.
In case of valid JSON, the implied CAST makes it look as though
the second variant of these function was supported as well but that
illusion totally falls apart once the JSON is not valid anymore.

I think it affects the following feature IDs:

- T821, Basic SQL/JSON query operators
For JSON_VALUE, JSON_TABLE and JSON_EXISTS
- T828, JSON_QUERY

Also, how hard would it be to add the functions that accept
character strings? Is there, besides the effort, any thing else
against it? I’m asking because I believe once released it might
never be changed — for backward compatibility.

[1] https://www.postgresql.org/message-id/CAKFQuwb50BAaj83Np%2B1O6xe3_T6DO8w2mxtFbgSbbUng%2BabrqA%40mail.gmail.com

>
>>> 2. EMPTY [ARRAY|OBJECT] ON ERROR implies ERROR ON EMPTY
>>>
>>> 17beta1=# SELECT JSON_QUERY('[]', '$[*]' EMPTY ARRAY ON ERROR) a;
>>> a
>>> ----
>>> []
>>> (1 row)
>>>
>>> As NULL ON EMPTY is implied, it should give the same result as
>>> explicitly adding NULL ON EMPTY:
>>>
>>> 17beta1=# SELECT JSON_QUERY('[]', '$[*]' NULL ON EMPTY EMPTY ARRAY ON ERROR) a;
>>> a
>>> ---
>>>
>>> (1 row)
>>>
>>> Interestingly, Oracle DB gives the same (wrong) results. Db2 (LUW)
>>> on the other hand returns NULL for both queries.
>>>
>>> I don’t think that PostgreSQL should follow Oracle DB's suit here
>>> but again, in case this is intentional it should be made explicit
>>> in the docs.
>
> This behavior is a bug and result of an unintentional change that I
> made at some point after getting involved with this patch set. So I'm
> going to fix this so that the empty results of jsonpath evaluation use
> NULL ON EMPTY by default, ie, when the ON EMPTY clause is not present.
> Attached a patch to do so.
>

Tested: works.

Thanks :)

-markus