On Wed, May 7, 2014 at 1:39 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> The indexscan is incorrectly returning rows where the queried key exists
> but isn't at top-level.
>
> We could fix this either by giving up on no-recheck for existence queries,
> or by changing the way that non-top-level keys get indexed. However
> I suspect the latter would break containment queries, or at least make
> their lives a lot more difficult.
Thanks for looking into this.
I don't know that it would make life all that much harder for
containment. jsonb_ops is already not compelling for testing
containment where there is a lot of nesting. ISTM that
make_scalar_key() could be taught to respect that distinction. ISTM
that containment will indifferently treat the top-level keys as top
level, and the nested keys as nested, because if you're looking at
something nested you have to put a top-level key in the rhs value to
do so, and it will be marked as top level, or not top level in a
bijective fashion.
--
Peter Geoghegan