Re: extending relations more efficiently

On Wed, May 2, 2012 at 12:26 PM, Alvaro Herrera
<alvherre(at)commandprompt(dot)com> wrote:
> Excerpts from Robert Haas's message of mié may 02 08:14:36 -0400 2012:
>> On Wed, May 2, 2012 at 7:16 AM, Jeroen Vermeulen <jtv(at)xs4all(dot)nl> wrote:
>> > On 2012-05-01 22:06, Robert Haas wrote:
>> >> It might also be interesting to provide a mechanism to pre-extend a
>> >> relation to a certain number of blocks, though if we did that we'd
>> >> have to make sure that autovac got the memo not to truncate those
>> >> pages away again.
>> >
>> > Good point.  And just to check before skipping over it, do we know that
>> > autovacuum not leaving enough slack space is not a significant cause of the
>> > bottlenecks in the first place?
>>
>> I'm not sure exactly what you mean by this: autovacuum doesn't need
>> any slack space.  Regular DML operations can certainly benefit from
>> slack space, both within each page and overall within the relation.
>> But there's no evidence that vacuum is doing too good a job cleaning
>> up the mess, forcing the relation to be re-extended.  Rather, the
>> usual (and frequent) complaint is that vacuum is leaving way too much
>> slack space - i.e. bloat.
>
> Hm.  I see those two things as different -- to me, bloat is unremoved
> dead tuples, whereas slack space would be free space that can be reused
> by new tuples.  Slack space is useful as it avoids relation extension;
> bloat is not.

I guess I think of bloat as including both unremoved dead tuples and
unwanted internal free space. If you create a giant table, delete 9
out of every 10 tuples, and vacuum, the table is still "bloated", IMV.

> I wonder, though, if we should set a less-than-100 fillfactor for heap
> relations.  Just like default_statistic_target, it seems that the
> default value should be a conservative tradeoff between two extremes.
> This doesn't help extension for bulk insertions a lot, of course, but
> it'd be useful for tables where HOT updates happen with some regularity.

Perhaps, but in theory that should be self-correcting: the data should
spread itself onto the number of pages where HOT pruning is able to
prevent further relation extension.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company