Re: question on visibility map

On Fri, Apr 24, 2026 at 10:13 AM Fabrice Chapuis
<commeilmeplait(at)gmail(dot)com> wrote:
>
> Hi,
>
> In this test I found that not all pages are not marked as all_visible after vacuuming, any explanation?
>
> CREATE TABLE test (
> id SERIAL PRIMARY KEY,
> name TEXT,
> created_at TIMESTAMP DEFAULT now()
> );
>
> INSERT INTO test (name)
> SELECT
> 'name_' || g
> FROM generate_series(1, 100000) AS g;
> CREATE TABLE
> INSERT 0 100000
> cpr [2720498]=#
> cpr [2720498]=#
> cpr [2720498]=# table test limit 10;
> +----+---------+----------------------------+
> | id | name | created_at |
> +----+---------+----------------------------+
> | 1 | name_1 | 2026-04-24 09:35:46.561014 |
> | 2 | name_2 | 2026-04-24 09:35:46.561014 |
> | 3 | name_3 | 2026-04-24 09:35:46.561014 |
> | 4 | name_4 | 2026-04-24 09:35:46.561014 |
> | 5 | name_5 | 2026-04-24 09:35:46.561014 |
> | 6 | name_6 | 2026-04-24 09:35:46.561014 |
> | 7 | name_7 | 2026-04-24 09:35:46.561014 |
> | 8 | name_8 | 2026-04-24 09:35:46.561014 |
> | 9 | name_9 | 2026-04-24 09:35:46.561014 |
> | 10 | name_10 | 2026-04-24 09:35:46.561014 |
> +----+---------+----------------------------+
> (10 rows)
>
> # ALTER TABLE test SET (autovacuum_enabled = false);
> ALTER TABLE
> # SELECT reloptions
> FROM pg_class
> WHERE relname = 'test';
> +----------------------------+
> | reloptions |
> +----------------------------+
> | {autovacuum_enabled=false} |
> +----------------------------+
> (1 row)
>
> # update test set name = 'name_x' where id = 1;
> UPDATE 1
>
> # create extension pg_visibility;
> CREATE EXTENSION
> # SELECT
> c.relpages AS total_pages,
> s.all_visible,
> s.all_frozen,
> round(100.0 * s.all_visible / NULLIF(c.relpages, 0), 1) AS pct_visible,
> round(100.0 * s.all_frozen / NULLIF(c.relpages, 0), 1) AS pct_frozen
> FROM pg_class c
> CROSS JOIN LATERAL pg_visibility_map_summary(c.oid) s
> WHERE c.relname = 'test';
> +-------------+-------------+------------+-------------+------------+
> | total_pages | all_visible | all_frozen | pct_visible | pct_frozen |
> +-------------+-------------+------------+-------------+------------+
> | 637 | 635 | 0 | 99.7 | 0.0 |
> +-------------+-------------+------------+-------------+------------+
> (1 row)
>
> vacuum test;
> VACUUM
> # SELECT
> c.relpages AS total_pages,
> s.all_visible,
> s.all_frozen,
> round(100.0 * s.all_visible / NULLIF(c.relpages, 0), 1) AS pct_visible,
> round(100.0 * s.all_frozen / NULLIF(c.relpages, 0), 1) AS pct_frozen
> FROM pg_class c
> CROSS JOIN LATERAL pg_visibility_map_summary(c.oid) s
> WHERE c.relname = 'test';
> +-------------+-------------+------------+-------------+------------+
> | total_pages | all_visible | all_frozen | pct_visible | pct_frozen |
> +-------------+-------------+------------+-------------+------------+
> | 637 | 636 | 0 | 99.8 | 0.0 |
> +-------------+-------------+------------+-------------+------------+
> (1 row)
>

Hi Fabrice,

with both recent 12 and 13.23 I'm was getting as the result:

total_pages | all_visible | all_frozen | pct_visible | pct_frozen
-------------+-------------+------------+-------------+------------
637 | 637 | 0 | 100.0 | 0.0

but starting with 14+, I've started getting those 1-2 not visible
blocks too at the begging:

total_pages | all_visible | all_frozen | pct_visible | pct_frozen
-------------+-------------+------------+-------------+------------
637 | 636 | 636 | 99.8 | 99.8

and with select * from pg_visibility_map('test'::regclass) where blkno
<= 2 and all_visible='f';

blkno | all_visible | all_frozen
-------+-------------+------------
0 | f | f

so it is always block #0 where Your update landed. After VACUUM FULL +
VACUUM however it back to 100% again. Now, re-trying Your's test case
(to get it again to 636/637) to inspect it further we apparently have
this situation on 14..master:

CREATE EXTENSION pageinspect;
SELECT lp, lp_off, lp_flags, lp_len, t_xmin, t_xmax, t_ctid,
t_infomask, t_infomask2 FROM heap_page_items(get_raw_page('test', 0))

lp | lp_off | lp_flags | lp_len | t_xmin | t_xmax | t_ctid |
t_infomask | t_infomask2
-----+--------+----------+--------+--------+--------+---------+------------+-------------
1 | 0 | 3 | 0 | | | |
|
2 | 8144 | 1 | 48 | 753 | 0 | (0,2) |
2306 | 3
3 | 8096 | 1 | 48 | 753 | 0 | (0,3) |
2306 | 3
4 | 8048 | 1 | 48 | 753 | 0 | (0,4) |
2306 | 3
5 | 8000 | 1 | 48 | 753 | 0 | (0,5) |
2306 | 3
6 | 7952 | 1 | 48 | 753 | 0 | (0,6) |
2306 | 3

Note lp_flags==3, so with 14+ we started zero-ing out the block
and putting it at the end of relation when looking just at the
result of the testcase:

SELECT blkno, count(*) AS item_count FROM generate_series(0, (SELECT
pg_relation_size('test') / 8192) - 1) AS blkno, LATERAL
heap_page_items(get_raw_page('test', blkno::int)) GROUP BY 1 ORDER BY
1;
[..]
634 | 157
635 | 157
636 | 149 // this has 148 -> 149 LPs now, due to that single UPDATE
appending it there

So with blkno=636 being larger by one LP, you can new row @ 149 LP:

lp | lp_off | lp_flags | lp_len | t_xmin | t_xmax | t_ctid |
t_infomask | t_infomask2
-----+--------+----------+--------+--------+--------+-----------+------------+-------------
[..]
146 | 1184 | 1 | 48 | 799 | 0 | (636,146) |
2306 | 3
147 | 1136 | 1 | 48 | 799 | 0 | (636,147) |
2306 | 3
148 | 1088 | 1 | 48 | 799 | 0 | (636,148) |
2306 | 3
149 | 1040 | 1 | 48 | 801 | 0 | (636,149) |
10498 | 3

So on 14+, I think UPDATE just zeroed out blkno#0 lp_1 (and later it had set
lp_flags == 3 == LP_DEAD (!) flag was set there by VACUUM), and that UPDATE
put new record at the end of table, but VACUUM somewhat is not able to remove
that LP_DEAD unless indexes are guranteed to be cleaned up from stale entries.

If you take a look on 14+ then can force it do so
`VACUUM (INDEX_CLEANUP true) test;` and it will have 100% pct_visible set
afterwards. See more for default toggle 'INDEX_CLEANUP AUTO' which says
`the default is AUTO, which allows VACUUM to skip index vacuuming when
appropriate` and this was introduced in 14.0 as part of`Allow vacuum to
skip index vacuuming when the number of removable index entries is
insignificant`. So the LP_DEAD is just sign that VACUUM in Your's
testcase did not clean indexes (unless forced to do so). That feature
references two commits [0][1] and first one even says in the commitmsg:

"We only skip index vacuuming when 2% or less of the table's pages have
one or more LP_DEAD items -- bypassing index vacuuming as an
optimization must not noticeably impede setting bits in the visibility
map."

Hope that helps.

-J.

[0] - https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=5100010ee
[1] - https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=3499df0de