Thread

1. BUG #19105: Parallel Hash Join crash when work_mem is small

   PG Bug reporting form <noreply@postgresql.org> — 2025-11-08T22:01:45Z

   The following bug has been logged on the website:
   
   Bug reference:      19105
   Logged by:          Nagesh Ka
   Email address:      nageshdbaa@gmail.com
   PostgreSQL version: 16.2
   Operating system:   Ubuntu 22.04 x86_64
   Description:        
   
   PostgreSQL version: 16.2 (Ubuntu 22.04 x86_64)
   Compiled by gcc (Ubuntu 11.4.0)
   work_mem = 1MB
   max_parallel_workers_per_gather = 4
   
   CREATE TABLE t1 AS SELECT i, md5(i::text) AS txt FROM
   generate_series(1,1000000) i;
   CREATE TABLE t2 AS SELECT i, md5(i::text) AS txt FROM
   generate_series(1,1000000) i;
   
   SET work_mem = '1MB';
   SET max_parallel_workers_per_gather = 4;
   
   SELECT t1.i, count(*)
   FROM t1 JOIN t2 USING(i)
   GROUP BY t1.i;
   
   -->Expected Result
   
   Query should complete successfully and return a count for each key.
   
   -->Observed Result
   
   Server terminates with:
   
   ERROR:  could not attach to dynamic shared area
   DETAIL:  invalid dsa_handle
   
   

2. Re: BUG #19105: Parallel Hash Join crash when work_mem is small

   Tom Lane <tgl@sss.pgh.pa.us> — 2025-11-09T15:47:25Z

   PG Bug reporting form <noreply@postgresql.org> writes:
   > PostgreSQL version: 16.2
   
   > Server terminates with:
   
   > ERROR:  could not attach to dynamic shared area
   > DETAIL:  invalid dsa_handle
   
   I could not reproduce this problem with current sources
   (either master or v16 branch tip).  16.2 is pretty old,
   so maybe an update will fix it for you.
   
   			regards, tom lane
   
   
   
   

3. Re: BUG #19105: Parallel Hash Join crash when work_mem is small

   Tomas Vondra <tomas@vondra.me> — 2025-11-09T17:49:13Z

   On 11/9/25 16:47, Tom Lane wrote:
   > PG Bug reporting form <noreply@postgresql.org> writes:
   >> PostgreSQL version: 16.2
   > 
   >> Server terminates with:
   > 
   >> ERROR:  could not attach to dynamic shared area
   >> DETAIL:  invalid dsa_handle
   > 
   > I could not reproduce this problem with current sources
   > (either master or v16 branch tip).  16.2 is pretty old,
   > so maybe an update will fix it for you.
   > 
   
   FWIW I couldn't reproduce it on 16.2 either, even after forcing it to
   actually do parallel hash join (the posted instructions were not enough
   to make that happen on my system).
   
   Presumably there's some additional detail making it fail, but there are
   too many options to guess. Maybe it's some kernel limit, something in
   ubuntu 22.04, or something else.
   
   It's interesting the report claims this happens only "when work_mem is
   small". I wonder if that happens simply because the plan changes.
   
   
   regards
   
   -- 
   Tomas Vondra
   
   
   
   
   

4. Re: BUG #19105: Parallel Hash Join crash when work_mem is small

   Thomas Munro <thomas.munro@gmail.com> — 2025-11-09T22:35:10Z

   On Mon, Nov 10, 2025 at 4:23 AM PG Bug reporting form
   <noreply@postgresql.org> wrote:
   > ERROR:  could not attach to dynamic shared area
   > DETAIL:  invalid dsa_handle
   
           ereport(ERROR,
                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                    errmsg("could not attach to dynamic shared area")));
   
   Where do we emit that DETAIL message?
   
   
   
   

5. Re: BUG #19105: Parallel Hash Join crash when work_mem is small

   Tom Lane <tgl@sss.pgh.pa.us> — 2025-11-09T23:40:55Z

   Thomas Munro <thomas.munro@gmail.com> writes:
   > On Mon, Nov 10, 2025 at 4:23 AM PG Bug reporting form
   > <noreply@postgresql.org> wrote:
   >> ERROR:  could not attach to dynamic shared area
   >> DETAIL:  invalid dsa_handle
   
   >         ereport(ERROR,
   >                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
   >                  errmsg("could not attach to dynamic shared area")));
   
   > Where do we emit that DETAIL message?
   
   I'd not picked up on that point, but you're right: we don't emit any
   such DETAIL anywhere, not in current and not in 16.2 either.  Must
   be a locally-modified tree, or maybe just a copy-paste error in the
   report?
   
   			regards, tom lane