Skip site navigation (1) Skip section navigation (2)

Re: Vacuumdb - Max_FSM_Pages Problem.

From: Ron <rjpeace(at)earthlink(dot)net>
To: "Peter Childs" <peterachilds(at)gmail(dot)com>
Cc: pgsql-performance(at)postgresql(dot)org
Subject: Re: Vacuumdb - Max_FSM_Pages Problem.
Date: 2007-02-26 20:33:14
Message-ID: (view raw, whole thread or download thread mbox)
Lists: pgsql-performance
At 10:53 AM 2/26/2007, Peter Childs wrote:
>On 26/02/07, Pallav Kalva <pkalva(at)livedatagroup(dot)com> wrote:
>>      I am in the process of cleaning up one of our big table, this table
>>has 187 million records and we need to delete around 100 million of them.
>>      I am deleting around 4-5 million of them daily in order to catchup
>>with vacuum and also with the archive logs space. So far I have deleted
>>around 15million in past few days.
>>      max_fsm_pages value is set to 1200000. Vacuumdb runs once daily,
>>here is the output from last night's vacuum job
>>      INFO:  free space map: 999 relations, 798572 pages stored; 755424
>>total pages needed
>>      DETAIL:  Allocated FSM size: 1000 relations + 1200000 pages = 7096
>>kB shared memory.
>>      VACUUM
>>      From the output it says 755424 total pages needed , this number
>>keeps growing daily even after vacuums are done daily. This was around
>>350K pages before the delete process started.
>>      I am afraid that this number will reach the max_fsm_pages limit
>>soon and vacuums thereafter will never catch up .
>>      Can anyone please explain this behavior ? What should I do to catch
>>up with vacuumdb daily ?
>Vacuum adds to free pages to the fsm so that they can be reused. If
>you don't fill up those free pages the fsm will fill up. Once the fsm
>is full no more pages can be added to the fsm. If you start writing to
>the free pages via inserts when vacuum next runs more free pages will
>be added that did not fit previously in the free space map due to it
>being full.
>If you are really deleting that many records you may be better coping
>those you want to a new table and dropping the old one. To actually
>recover space you need to either run vacuum full or cluster.
>This ought to be in the manual somewhere as this question gets asked
>about once a week.
In fact ,
a= copying data to a new table and dropping the original table
rather than
b= updating the original table
is a "standard best DBA practice" regardless of DB product.

The only thing that changes from DB product to DB product is the 
exact point where the copy is large enough to make "copy, replace" 
better than "update in place".

Rule of Thumb:  No matter what DB product you are using, if it's more 
than 1/2 of any table or more than 1/4 of any table that does not fit 
into memory, it's usually better to copy replace rather then update in place.

...and I completely agree that we should document this sort of 
Industry Best Practice in a way that is easily usable by the pg community.


In response to

pgsql-performance by date

Next:From: Charles SprickmanDate: 2007-02-26 21:24:12
Subject: Re: Writting a "search engine" for a pgsql DB
Previous:From: Jeff DavisDate: 2007-02-26 20:14:05
Subject: Re: Writting a "search engine" for a pgsql DB

Privacy Policy | About PostgreSQL
Copyright © 1996-2017 The PostgreSQL Global Development Group