Re: PG17 optimizations to vacuum

On 02/09/2024 23:35, Pavel Luzanov wrote:
> On 02.09.2024 22:23, Melanie Plageman wrote:
>> For some reason I stopped being able to reproduce Pavel's case.

I also cannot reproduce this.

> I repeated the test on another computer, but compared master with v15.
> The results are the same. The test can be simplified as follows:
>
> CREATE TABLE t(id integer) WITH (autovacuum_enabled = off);
> INSERT INTO t SELECT gen.id FROM generate_series(1,3_500_000) gen(id);
> CREATE INDEX t_id ON t(id);
> VACUUM FREEZE t;
> UPDATE t SET id = id + 1;
> VACUUM FREEZE VERBOSE t;
>
> My results (only line with WAL info from the last VACUUM command).
>
> master:
> WAL usage: 119583 records, 37231 full page images, 272631468 bytes
>
> v15:
> WAL usage: 96565 records, 47647 full page images, 217144602 bytes

Can you dump the stats with pg_waldump please. Something like:

CREATE TABLE t(id integer) WITH (autovacuum_enabled = off);
INSERT INTO t SELECT gen.id FROM generate_series(1,3_500_000) gen(id);
CREATE INDEX t_id ON t(id);
SET maintenance_work_mem = '1MB';
UPDATE t SET id = id + 1;
select pg_current_wal_insert_lsn(); -- <<< PRINT WAL POS BEFORE VACUUM
VACUUM FREEZE VERBOSE t;

And then:

pg_waldump -p data/pg_wal/ -s 1/F4474498 --stats=record

where "1/F4474498" is the position printed by the "SELECT
pg_current_wal_insert_lsn()" above.

Do you have any non-default settings? "select name,
current_setting(name), source from pg_settings where setting <>
boot_val;" would show that.

--
Heikki Linnakangas
Neon (https://neon.tech)