Hello hackers,
I happened to observe an interesting failure of `make check` being
executed concurrently against multiple databases:
diff -U3 --minimal .../regress_input_27/expected/domain.out .../regress_output_27/results/domain.out
--- .../expected/domain.out 2026-10-02 08:03:59.474274001 +0000
+++ .../results/domain.out 2026-10-02 08:26:49.225597020 +0000
@@ -1337,11 +1337,7 @@
FROM information_schema.domain_constraints
WHERE domain_name IN ('con', 'dom', 'pos_int', 'things'))
ORDER BY constraint_name;
- constraint_catalog | constraint_schema | constraint_name | check_clause
---------------------+-------------------+------------------+-------------------
- regress027 | public | con_check | (VALUE > 0)
- regress027 | public | meow | (VALUE < 11)
- regress027 | public | pos_int_check | (VALUE > 0)
- regress027 | public | pos_int_not_null | VALUE IS NOT NULL
-(4 rows)
-
+ERROR: deadlock detected
+DETAIL: Process 322216 waits for AccessShareLock on relation 1197300 of database 16411; blocked by process 322217.
+Process 322217 waits for AccessExclusiveLock on relation 1197314 of database 16411; blocked by process 322216.
+HINT: See server log for query details.
I've managed to reproduce this with the attached patch:
# diff -U3 .../src/test/regress/expected/alter_table.out .../src/test/regress/results/alter_table.out
# --- .../src/test/regress/expected/alter_table.out 2026-10-07 12:53:24.843538437 -0400
# +++ .../src/test/regress/results/alter_table.out 2026-10-07 14:58:08.602062642 -0400
# @@ -4808,7 +4808,10 @@
# (1 row)
#
# alter table attbl alter column a type numeric; -- someday this should work
# -ERROR: cannot alter table "attbl" because column "atref.b" uses its row type
# +ERROR: deadlock detected
# +DETAIL: Process 2884690 waits for AccessExclusiveLock on relation 45440 of database 16384; blocked by process
2884685.
# +Process 2884685 waits for AccessShareLock on relation 45437 of database 16384; blocked by process 2884690.
# +HINT: See server log for query details.
# alter table atref drop constraint atref_b_check;
# create statistics atref_stat on ((b).a is not null) from atref;
# alter table attbl alter column a type numeric; -- someday this should work
That is, the deadlock occurs between
"SELECT * FROM information_schema.check_constraint" / pg_get_constraintdef()
and
"alter table attbl alter column a type numeric;" / ATPostAlterTypeCleanup().
Reproduced starting from 8319e5cb5.
Best regards,
Alexander
