Обсуждение: Vacuum Question

Поиск
Список
Период
Сортировка

Vacuum Question

От
Murthy Nunna
Дата:

Version 14.13

 

I have a large database 22 TB, and it has lot of tables. Most of the tables do not change (static). But the age(relfrozenxid) of those tables keep increasing because there are some other tables in the database that are updated. The size of these large static tables are about 200 GB on an average. And to prevent transaction ID wrap around, I have been doing manual vacuum table by table (couple of tables a day due to limited WAL disk space). Each table generates WAL size of 90% of the tablesize approx.

e.g

Tablesize = 200 GB. Time takes to run vacuum = 1 hour 45 minutes. WAL generated 182 GB

 

I tried VACUUM FREEZE also, but the WAL generated and time it takes is no significantly different.

 

Following is an example output of a table vacuum:

 

vacuumdb: vacuuming database "large_db"

INFO:  aggressively vacuuming "public.tab_111"

INFO:  launched 1 parallel vacuum worker for index cleanup (planned: 1)

INFO:  table "tab_111": found 0 removable, 527846215 nonremovable row versions in 15396753 out of 15396753 pages

DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 954951860

Skipped 0 pages due to buffer pins, 0 frozen pages.

CPU: user: 131.12 s, system: 174.14 s, elapsed: 4111.88 s.

INFO:  aggressively vacuuming "pg_toast.pg_toast_17386"

INFO:  table "pg_toast_17386": found 0 removable, 32180684 nonremovable row versions in 7981550 out of 7981550 pages

DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 955034530

Skipped 0 pages due to buffer pins, 0 frozen pages.

CPU: user: 52.96 s, system: 87.86 s, elapsed: 2104.04 s.

 

Is there a way I can minimize WAL generation? My issue is amount of WAL rather than time it takes to run. Since it is not locking the table I do not mind long run time.

I know one way is to pgdump/restore but it takes a long time and further to that I have to rebuild replicas.

 

Please note, I have autovacuum turned on and it is doing what it is supposed to do on tables that change. So, there is no issue there.

 

I very much appreciate any help/advice you can provide.

Re: Vacuum Question

От
Ron Johnson
Дата:
I think you asked the same question 11 years ago. 😀
  • Seriously, though, 200GB is less than 1% of 22TB.  There are bigger problems if you're running that razor-thin on disk space.
  • Is the transaction rate on the active tables sooo high that there's a real chance of wrap-around?
  • According to my interpretation of the docs, if you VACUUM FREEZE the big static tables, then you won't need to vacuum them again, nor worry about wrap-around problems.
PostgreSQL reserves a special XID, FrozenTransactionId, which does not follow the normal XID comparison rules and is always considered older than every normal XID. Frozen row versions are treated as if the inserting XID were FrozenTransactionId, so that they will appear to be “in the past” to all normal transactions regardless of wraparound issues, and so such row versions will be valid until deleted, no matter how long that is.

On Mon, Sep 22, 2025 at 11:06 AM Murthy Nunna <mnunna@fnal.gov> wrote:

Version 14.13

 

I have a large database 22 TB, and it has lot of tables. Most of the tables do not change (static). But the age(relfrozenxid) of those tables keep increasing because there are some other tables in the database that are updated. The size of these large static tables are about 200 GB on an average. And to prevent transaction ID wrap around, I have been doing manual vacuum table by table (couple of tables a day due to limited WAL disk space). Each table generates WAL size of 90% of the tablesize approx.

e.g

Tablesize = 200 GB. Time takes to run vacuum = 1 hour 45 minutes. WAL generated 182 GB

 

I tried VACUUM FREEZE also, but the WAL generated and time it takes is no significantly different.

 

Following is an example output of a table vacuum:

 

vacuumdb: vacuuming database "large_db"

INFO:  aggressively vacuuming "public.tab_111"

INFO:  launched 1 parallel vacuum worker for index cleanup (planned: 1)

INFO:  table "tab_111": found 0 removable, 527846215 nonremovable row versions in 15396753 out of 15396753 pages

DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 954951860

Skipped 0 pages due to buffer pins, 0 frozen pages.

CPU: user: 131.12 s, system: 174.14 s, elapsed: 4111.88 s.

INFO:  aggressively vacuuming "pg_toast.pg_toast_17386"

INFO:  table "pg_toast_17386": found 0 removable, 32180684 nonremovable row versions in 7981550 out of 7981550 pages

DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 955034530

Skipped 0 pages due to buffer pins, 0 frozen pages.

CPU: user: 52.96 s, system: 87.86 s, elapsed: 2104.04 s.

 

Is there a way I can minimize WAL generation? My issue is amount of WAL rather than time it takes to run. Since it is not locking the table I do not mind long run time.

I know one way is to pgdump/restore but it takes a long time and further to that I have to rebuild replicas.

 

Please note, I have autovacuum turned on and it is doing what it is supposed to do on tables that change. So, there is no issue there.

 

I very much appreciate any help/advice you can provide.



--
Death to <Redacted>, and butter sauce.
Don't boil me, I'm still alive.
<Redacted> lobster!

RE: Vacuum Question

От
Murthy Nunna
Дата:

 

 

From: Ron Johnson <ronljohnsonjr@gmail.com>
Sent: Monday, September 22, 2025 10:39 AM
To: pgsql-admin <pgsql-admin@postgresql.org>; Murthy Nunna <mnunna@fnal.gov>
Subject: Re: Vacuum Question

 

[EXTERNAL] – This message is from an external sender

I think you asked the same question 11 years ago. 😀

  • Seriously, though, 200GB is less than 1% of 22TB.  There are bigger problems if you're running that razor-thin on disk space.
    • I have 10 TB disk for WALs which is separate from /pgdata disk. But with the amount of WALs vacuum is generating, even 10 TB could fill up.
  • Is the transaction rate on the active tables sooo high that there's a real chance of wrap-around?
    • Given the xid numbers, I would say it is high. It went from 786089 to 965376038 in 12 months
  • According to my interpretation of the docs, if you VACUUM FREEZE the big static tables, then you won't need to vacuum them again, nor worry about wrap-around problems.
    • About a year ago, I did pgdump/restore. But I did not perform vacuum though. So, may be I should complete my table by table vacuum at least once then.

PostgreSQL reserves a special XID, FrozenTransactionId, which does not follow the normal XID comparison rules and is always considered older than every normal XID. Frozen row versions are treated as if the inserting XID were FrozenTransactionId, so that they will appear to be “in the past” to all normal transactions regardless of wraparound issues, and so such row versions will be valid until deleted, no matter how long that is.

 

Re: Vacuum Question

От
Laurenz Albe
Дата:
On Mon, 2025-09-22 at 15:06 +0000, Murthy Nunna wrote:
> Version 14.13
>  
> I have a large database 22 TB, and it has lot of tables. Most of the tables do not change (static).
> But the age(relfrozenxid) of those tables keep increasing because there are some other tables in
> the database that are updated. The size of these large static tables are about 200 GB on an
> average. And to prevent transaction ID wrap around, I have been doing manual vacuum table by table
> (couple of tables a day due to limited WAL disk space). Each table generates WAL size of 90% of
> the tablesize approx.
> e.g
> Tablesize = 200 GB. Time takes to run vacuum = 1 hour 45 minutes. WAL generated 182 GB
>  
> I tried VACUUM FREEZE also, but the WAL generated and time it takes is no significantly different.
>  
> Following is an example output of a table vacuum:
>  
> vacuumdb: vacuuming database "large_db"
> INFO:  aggressively vacuuming "public.tab_111"
> INFO:  launched 1 parallel vacuum worker for index cleanup (planned: 1)
> INFO:  table "tab_111": found 0 removable, 527846215 nonremovable row versions in 15396753 out of 15396753 pages
> DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 954951860
> Skipped 0 pages due to buffer pins, 0 frozen pages.
> CPU: user: 131.12 s, system: 174.14 s, elapsed: 4111.88 s.
> INFO:  aggressively vacuuming "pg_toast.pg_toast_17386"
> INFO:  table "pg_toast_17386": found 0 removable, 32180684 nonremovable row versions in 7981550 out of 7981550 pages
> DETAIL:  0 dead row versions cannot be removed yet, oldest xmin: 955034530
> Skipped 0 pages due to buffer pins, 0 frozen pages.
> CPU: user: 52.96 s, system: 87.86 s, elapsed: 2104.04 s.
>  
> Is there a way I can minimize WAL generation?

If the data won't change any more, run a VACUUM (FREEZE) on the table.  That should freeze all
rows, and any subsequent VACUUM will finish very quickly and produce no WAL.

> About a year ago, I did pgdump/restore. But I did not perform vacuum though. So, may be I
> should complete my table by table vacuum at least once then.

If you restore a table from a dump, all the rows will be unfrozen.
It will take another VACUUM (FREEZE) to freeze the rows.

Yours,
Laurenz Albe