BUG #19530: Crash (SIGSEGV/SIGBUS) in parallel B-tree index vacuum during plain VACUUM

Поиск
Список
Период
Сортировка
Искать

BUG #19530: Crash (SIGSEGV/SIGBUS) in parallel B-tree index vacuum during plain VACUUM

От:
PG Bug reporting form <noreply@postgresql.org>
Дата:
The following bug has been logged on the website:

Bug reference:      19530
Logged by:          Maciej
Email address:      webmaster@algedi.pl
PostgreSQL version: 17.7
Operating system:   FreeBSD 14.4-RELEASE, x86-64
Description:        

PostgreSQL 17.7 intermittently crashes a parallel maintenance worker during
routine VACUUM, taking the server through automatic crash recovery each
  time.

  Summary:
  Recurring backend / parallel-worker terminations with signal 11 (SIGSEGV)
and signal 10 (SIGBUS). Each crash is followed by "terminating any other
  active server processes" and crash recovery (~5 s). It is intermittent:
about 16 process terminations over 5 days, with fully clean days in between;
  the same nightly maintenance succeeds on most days. Hardware was checked
and is clean (no MCE/ECC/I/O errors).

  Environment:
  PostgreSQL 17.7, FreeBSD 14.4-RELEASE, x86-64, physical server, 256 GB
RAM. dynamic_shared_memory_type=posix, huge_pages=try.
  max_parallel_workers=8, max_parallel_maintenance_workers=4,
max_parallel_workers_per_gather=4. Extensions: pg_stat_statements, pg_trgm,
unaccent,
  plpgsql only — no custom C extensions, no PL/Perl/Python.

  What the crashing process was doing:
  The crashing process is a parallel maintenance worker. A preserved core
dump shows the process title "parallel worker for PID ", with memory
  contexts "BTree Vacuum State" and "AutoVacuum Data" present and no
executor/query state — i.e. parallel index vacuum, not parallel query. The
  trigger is a scheduled VACUUM (SKIP_DATABASE_STATS, VERBOSE, ANALYZE) run
via vacuumdb over several large tables, each carrying many (~10–30) btree
  indexes, so parallel index vacuuming is engaged. The SIGBUS variant is
consistent with a DSM (POSIX shared memory) problem in the parallel path.

  Expected vs actual:
  Expected: VACUUM completes normally. Actual: a parallel index-vacuum
worker crashes (SIGSEGV/SIGBUS) and the server goes through crash recovery.

  Possibly related:
  This appears to reach the same btbulkdelete / btvacuumscan code as the
report "Segmentation fault in PostgreSQL 17.7 during REINDEX TABLE
  CONCURRENTLY"
(https://www.postgresql.org/message-id/VI0PR07MB10718A4DC292E068C51404AB2974EA@VI0PR07MB10718.eurprd07.prod.outlook.com),
but reached
  via the plain-VACUUM index-vacuum path rather than validate_index /
ReindexRelationConcurrently. That suggests the fault is in the (parallel)
B-tree
  index-vacuum code, reachable from both paths.

  Workaround:
  Setting max_parallel_maintenance_workers = 0 (sequential index vacuuming)
appears to stop the crashes (observation ongoing).

Re: BUG #19530: Crash (SIGSEGV/SIGBUS) in parallel B-tree index vacuum during plain VACUUM

От:
John Naylor <johncnaylorls@gmail.com>
Дата:
On Mon, Jun 22, 2026 at 3:19 PM PG Bug reporting form
 wrote:
> PostgreSQL 17.7 intermittently crashes a parallel maintenance worker during
> routine VACUUM, taking the server through automatic crash recovery each
>   time.

>   What the crashing process was doing:
>   The crashing process is a parallel maintenance worker. A preserved core
> dump shows the process title "parallel worker for PID ", with memory
>   contexts "BTree Vacuum State" and "AutoVacuum Data" present and no

This doesn't really give us any actionable info. Are you able to get a
backtrace from the core dump?

(Autovacuum in v17 cannot use parallel workers, so I'm not sure where
that came from)

>   Possibly related:
>   This appears to reach the same btbulkdelete / btvacuumscan code as the
> report "Segmentation fault in PostgreSQL 17.7 during REINDEX TABLE
>   CONCURRENTLY"
> (https://www.postgresql.org/message-id/VI0PR07MB10718A4DC292E068C51404AB2974EA@VI0PR07MB10718.eurprd07.prod.outlook.com),
> but reached
>   via the plain-VACUUM index-vacuum path rather than validate_index /
> ReindexRelationConcurrently. That suggests the fault is in the (parallel)
> B-tree
>   index-vacuum code, reachable from both paths.

That operation was writing TIDs to a file, which is not what vacuum
does, and it's not clear from that thread whether there's a bug in the
first place.

--
John Naylor
Amazon Web Services
FAQ