Re: TRAP: failed Assert("offsets[i] > offsets[i - 1]"), File: "tidstore.c" - Mailing list pgsql-bugs

On 25/04/2026 01:23, Masahiko Sawada wrote:
> On Wed, Apr 22, 2026 at 10:23 AM Andrei Lepikhov <lepihov@gmail.com> wrote:
>> Both tools are experimental and not meant for core use; they are only used to
>> trigger potential issues. In this case, I think the query picked a costly sorted
>> path, which led to the crash.
> 
> Does this imply that array_agg() could return unsorted results
> depending on the plan the optimizer chooses? Or is such a path
> currently never selected by the optimizer?

The array_agg() function does not sort its output. In theory, this means the
join could return results in any order, but in practice, I have not seen this
happen.

> 
> I’m asking because if this scenario never occurs with the current
> optimizer, it might make sense to apply the patch only to HEAD (i.e.,
> for PG20). On the other hand, backpatching to PG17 might be justified,
> given that DISTINCT does not guarantee sorted results in principle,
> and the fix could benefit extension development on stable branches.

In stable versions, the planner's logic remains unchanged. So, it seems
reliable. However, backpatching could help extension developers a little bit.
Since this code fixes a real issue and does not break anything complex, I would
backpatch it. Still, I am fine with just committing it to master if you prefer.

P.S.
    
I looked into the issue further. The problem happens when the join sides are
shuffled. Here is what I found:

EXPLAIN of the successful execution (unnecessary details stripped):

Insert on pg_temp.hideblocks  (cost=1.21..1.66 rows=0 width=0)
   ->  Subquery Scan on unnamed_subquery  (cost=1.21..1.66 rows=5 width=8)
         Output: unnamed_subquery.do_set_block_offsets
         ->  GroupAggregate  (cost=1.21..1.61 rows=5 width=16)
               Output: do_set_block_offsets("*VALUES*".column1,
            (array_agg("*VALUES*_1".column1))::smallint[]), ...
               Group Key: "*VALUES*".column1
               ->  Sort  (cost=1.21..1.27 rows=25 width=12)
                     Output: "*VALUES*".column1, "*VALUES*_1".column1
                     Sort Key: "*VALUES*".column1
                     ->  Nested Loop  (cost=0.00..0.62 rows=25 width=12)
                           Output: "*VALUES*".column1, "*VALUES*_1".column1
                           ->  Values Scan on "*VALUES*"
                (cost=0.00..0.06 rows=5 width=8)
                                 Output: "*VALUES*".column1
                           ->  Values Scan on "*VALUES*_1"
                (cost=0.00..0.06 rows=5 width=4)
                                 Output: "*VALUES*_1".column1

EXPLAIN that causes assertion:

 Insert on pg_temp.hideblocks  (cost=1.03..1.48 rows=0 width=0)
   ->  Subquery Scan on unnamed_subquery  (cost=1.03..1.48 rows=5 width=8)
         Output: unnamed_subquery.do_set_block_offsets
         ->  GroupAggregate  (cost=1.03..1.43 rows=5 width=16)
               Output: do_set_block_offsets("*VALUES*".column1,
            (array_agg("*VALUES*_1".column1))::smallint[]),...
               Group Key: "*VALUES*".column1
               ->  Sort  (cost=1.03..1.09 rows=25 width=12)
                     Output: "*VALUES*".column1, "*VALUES*_1".column1
                     Sort Key: "*VALUES*".column1
                     ->  Nested Loop  (cost=0.00..0.45 rows=25 width=12)
                           Output: "*VALUES*".column1, "*VALUES*_1".column1
                           ->  Values Scan on "*VALUES*_1"
                (cost=0.00..0.06 rows=5 width=4)
                                 Output: "*VALUES*_1".column1
                           ->  Materialize  (cost=0.00..0.09 rows=5 width=8)
                                 Output: "*VALUES*".column1
                                 ->  Values Scan on "*VALUES*"
                    (cost=0.00..0.06 rows=5 width=8)
                                       Output: "*VALUES*".column1

At the second case offsets have come to the aggregation without order that
highlighted the issue.

-- 
regards, Andrei Lepikhov,
pgEdge