Hello everyone!
I've been examining a floating bug in pg_visibility. To reproduce it, you have to:
1) apply a patch to pg_visibility's tests (you can find it attached)
2) run test-icx.sh (attached) from the root of the postgresql's source tree
pg_vis_clean.patch adds one additional test to pg_visibility test suite and normally it is expected to pass. However, when (I assume so) it happens to run in parallel with autovacuum, the test is failed. test-icx.sh runs several parallel installchecks on contrib/pg_visibility tests and there is a chance that some of them will catch this error.
This is the set-up for a bug reproduction:
create table vacuum_test as select 42 i;
vacuum (disable_page_skipping) vacuum_test;
This is the actual test result:
diff -U3 /tmp/icx/ic01-1/contrib/pg_visibility/expected/pg_visibility.out /tmp/icx/ic01-1/contrib/pg_visibility/results/pg_visibility.out
--- /tmp/icx/ic01-1/contrib/pg_visibility/expected/pg_visibility.out 2022-02-17 23:46:26.005500630 +0300
+++ /tmp/icx/ic01-1/contrib/pg_visibility/results/pg_visibility.out 2022-02-17 23:47:36.066838872 +0300
@@ -247,7 +247,7 @@
select count(*) > 0 from pg_check_visible('vacuum_test');
?column?
----------
- f
+ t
(1 row)
-- cleanup
The thing is that the underlying file is one page long which should contain the only all-visible tuple and pg_check_visible should return nothing. Is this a real bug or do I miss something? Thank you in advance for the response
P. S.: SELECT version(); result: PostgreSQL 14.2 on x86_64-pc-linux-gnu, compiled by gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0, 64-bit
