Re: JSON constructors and window functions

On 4/3/22 22:46, Andrew Dunstan wrote:
> On 4/3/22 20:11, Andres Freund wrote:
>> Hi,
>>
>> On 2022-04-03 18:56:39 -0400, Andrew Dunstan wrote:
>>> Haven't found the issue yet :-( It happens on the second call for the
>>> partition to  json_check_unique_key().
>>>
>>> Here's a more idiomatic and self-contained query that triggers the problem.
>>>
>>>
>>> select json_objectagg('10' : ref_0.level2 with unique keys)
>>>     over (partition by ref_0.parent_no order by ref_0.level2)
>>> from (values (1::int,1::int),(1,2),(2,1),(2,2)) as ref_0(parent_no,level2);
>> The hash was created in a context that's already freed.
>>
> [...]
>>
>> I don't think you're allowed to free stuff in a finalfunc - we might reuse the
>> transition state for further calls to the aggregate.
>>
>
> Doh! Of course! I'll fix it in the morning. Thanks.
>
>

I've committed a fix for this. I didn't find something to clean out the
hash table, so I just removed the 'hash_destroy' and left it at that.
All the test I did came back with expected results.

Maybe a hash_reset() is something worth having assuming it's possible? I
note that simplehash has a reset function.

cheers

andrew

--
Andrew Dunstan
EDB: https://www.enterprisedb.com