Re: generic plans and "initial" pruning

On Sun, Feb 23, 2025 at 2:03 AM Tender Wang <tndrwang(at)gmail(dot)com> wrote:
> Alexander Lakhin <exclusion(at)gmail(dot)com> 于2025年2月22日周六 23:00写道：
>> 21.02.2025 05:40, Amit Langote wrote:
>>
>> I pushed the final piece yesterday.
>>
>>
>> Please look at new error, produced by the following script,
>> starting from 525392d57:
>> CREATE TABLE t(id int) PARTITION BY RANGE (id);
>> CREATE INDEX idx on t(id);
>> CREATE TABLE tp_1 PARTITION OF t FOR VALUES FROM (10) TO (20);
>> CREATE TABLE tp_2 PARTITION OF t FOR VALUES FROM (20) TO (30) PARTITION BY RANGE(id);
>> CREATE TABLE tp_2_1 PARTITION OF tp_2 FOR VALUES FROM (21) to (22);
>> CREATE TABLE tp_2_2 PARTITION OF tp_2 FOR VALUES FROM (22) to (23);
>> CREATE FUNCTION stable_one() RETURNS INT AS $$ BEGIN RETURN 1; END; $$ LANGUAGE plpgsql STABLE;
>>
>> SELECT min(id) OVER (PARTITION BY id ORDER BY id) FROM t WHERE id >= stable_one();
>>
>> ERROR: XX000: trying to open a pruned relation
>> LOCATION: ExecGetRangeTableRelation, execUtils.c:830
>>
>> This issue was discovered with SQLsmith.

Thanks for the report.

> The error message was added in commit 525392d57. In this case, the estate->es_unpruned_relids only includes 1, which is the offset of table t.
> In register_partpruneinfo(), we collect glob->prunableRelids; in this case, it contains 2,3,4,5. Then we will do:
> result->unprunableRelids = bms_difference(glob->allRelids,
> glob->prunableRelids);
> so the result->unprunableRelids only contains 1.
>
> But tp_2 is also partition table, and its partpruneinfo created by create_append_plan() is put into the head of global list.
> So we first process it in ExecDoInitialPruning(). Then error reports because we only contain 1 in estate->es_unpruned_relids.

Thanks for checking.

The RT index of tp_2 should appear in PlannedStmt.unprunableRelids,
because it needs to be opened in CreatePartitionPruneState() for
setting up its PartitionPruneInfo. We use ExecGetRangeTableRelation()
to open, which expects the relation to be locked, so the error.

To ensure tp_2 appears in PlannedStmt.unprunableRelids, we should
prevent make_partitionedrel_pruneinfo() from placing the RT index into
leafpart_rti_map[], as the current condition for inclusion doesn’t
account for whether the partition is itself partitioned.

I've come up with the attached.

--
Thanks, Amit Langote