On 05.12.2010 18:26, Tom Lane wrote:
> Andres Freund<andres@anarazel.de> writes:
>> On Sunday 05 December 2010 17:42:59 Tom Lane wrote:
>>> I think the reason the given example fails is just that it's all being
>>> done in one transaction. If the null-containing row were known dead
>>> it wouldn't get indexed. So: commit.
>
>> Um I doubt it.
>
> [ gets out gdb... ] Oh: the reason GIN is complaining is that it's just
> looking at ARR_HASNULL(), and the array's has-nulls flag is still set
> because we don't bother to try to clear it after replacing one element
> of the array. (Which in general would be an expensive thing to try to
> do...)
>
> If we were intending to leave GIN in its current nulls-hating state,
> the thing to do would be to replace the stupid ARR_HASNULL check with
> something more intelligent. But really it needs to be fixed to handle
> nulls properly, so I'm thinking that might be a dead-end patch.
Sounds like we'd still want to just replace ARR_HASNULL() with something
more intelligent in back-branches though.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com