Обсуждение: Vacuum Full does not release the disk size space after delete from table
Hi hackers,
We meet a corner case that related to the behavior of Vacuum Full.
There are two SQL scripts
SQL1:
-------------------------------------------------------------
-- Test vacuum fullcreate schema s2;
set search_path to s2;
create table a (i int);
create table b (i int);
insert into a select generate_series(1,50000);
delete from a where i > 10;
-- Disk space should release after vacuum full
vacuum full a;
drop table a, b;
reset search_path;
drop schema s2;
-----------------------------------------------------------------------------------------
SQL2:
------------------------------------------------------------------------------------------
create schema s1;set search_path to s1;
create table c(i int);
insert into c select generate_series(1,50000);
select pg_sleep(5);
reset search_path;
drop schema s1;
-----------------------------------------------------------------------------------------------
If we run both sql scripts on same database in parallel, the "VACUUM FULL a;" will not release the disk space.
But if we run then on same database in sequence, "VACUUM FULL a" can release the disk space.
We use the PG master branch and the autovacuum has set to off.
Could someone please help us to figure out the root cause of this issue?
Thanks a lot!
Regards,
Haozhou
Haozhou Wang <hawang@pivotal.io> writes: > We meet a corner case that related to the behavior of Vacuum Full. > ... > If we run both sql scripts on same database in parallel, the "VACUUM FULL > a;" will not release the disk space. I think what's happening is that the delete in script 1 happens after the "pg_sleep" in script 2 starts. Then the pg_sleep has an open snapshot that could potentially see the deleted rows, so they can't be removed yet. You could check this theory by changing the vacuum to use VERBOSE, and seeing what it says about rows that can't be removed yet. regards, tom lane
Thank Tom!
We will check it.
On Fri, Nov 2, 2018 at 10:35 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Haozhou Wang <hawang@pivotal.io> writes:
> We meet a corner case that related to the behavior of Vacuum Full.
> ...
> If we run both sql scripts on same database in parallel, the "VACUUM FULL
> a;" will not release the disk space.
I think what's happening is that the delete in script 1 happens after the
"pg_sleep" in script 2 starts. Then the pg_sleep has an open snapshot
that could potentially see the deleted rows, so they can't be removed yet.
You could check this theory by changing the vacuum to use VERBOSE, and
seeing what it says about rows that can't be removed yet.
regards, tom lane
Regards,
Haozhou