Thread: 8.1 vs 8.2.1 view optimization

8.1 vs 8.2.1 view optimization

From
Nathan Bell
Date:
Firing up 8.2.1 I notice that sub-items in a view are optimized out if
they aren't being selected.

For example, "select item1, item2 from a_view" would take just as long
as "select item1, item2, item3, item4 from a_view"

This isn't usually a problem, but if item3 or item4 are significantly
more complex (and slow) than item1 and item2 this is a big problem. In
8.1 and previous versions of postgresql this wouldn't happen. Is there
some setting that can be set to re-enable this feature?

Attached is a small sql script that shows the problem. When run on 8.2
or 8.2.1 it takes twice as long as when run on 8.1 because of this
un-feature. This can be run on a completely fresh, blank database and it
will create all of the languages, tables, etc that it needs. It will
also clean up everything afterwards. The funciton "slow_function" is for
illistration purposes only, and is made only to take time.

Thanks in advance for any help,

Nathan Bell
IT Engineer
Action Target, Inc.
create trusted language 'plpgsql'
handler plpgsql_call_handler lancompiler 'PL/pgSQL';

create table small ( only_item int4 );

create or replace function slow_function(int4, int4) returns int4 as $$
declare
    x int4;
    y int4;
    ret int4 := 2;
begin
    for x in 1..$1 loop
        for y in 1..$2 loop
            ret := ret+(x/y);
        end loop;
    end loop;
    return ret;
end;
$$ language plpgsql;

create or replace view small_v as

    select i.only_item as item1,
    slow_function(i.only_item,i.only_item) as item2
    from small i
  ;


insert into small values (1);
insert into small values (3);
insert into small values (10);
insert into small values (25);
insert into small values (100);
insert into small values (250);
insert into small values (1000);
insert into small values (2500);

select item1 from small_v;
select item2 from small_v;


drop view small_v;
drop function slow_function(int4,int4);
drop table small;
drop language 'plpgsql';

Re: 8.1 vs 8.2.1 view optimization

From
Tom Lane
Date:
Nathan Bell <nathanb@actarg.com> writes:
> Firing up 8.2.1 I notice that sub-items in a view are optimized out if
> they aren't being selected.

You mean "not optimized out", I suppose.  Declare your function as
non-volatile if you want the optimizer to assume it's OK to discard.

            regards, tom lane

Re: 8.1 vs 8.2.1 view optimization

From
Nathan Bell
Date:
Yeah, I saw the "not optimized out" typo as soon as I hit send.

What if the item that is taking a long time isn't a function, but rather
a sub-select?
Can I set the sub-select to stable, or perhaps set the entire view to
non-volatile to achieve the same result?
If not, can I set the sub-select to a different non-volatile view or do
I need to create a non-volatile function that returns the result?

Thanks for the help.

Tom Lane wrote:

>Nathan Bell <nathanb@actarg.com> writes:
>
>
>>Firing up 8.2.1 I notice that sub-items in a view are optimized out if
>>they aren't being selected.
>>
>>
>
>You mean "not optimized out", I suppose.  Declare your function as
>non-volatile if you want the optimizer to assume it's OK to discard.
>
>            regards, tom lane
>
>

Re: 8.1 vs 8.2.1 view optimization

From
Tom Lane
Date:
Nathan Bell <nathanb@actarg.com> writes:
> What if the item that is taking a long time isn't a function, but rather
> a sub-select?

The point is that the view won't be flattened if there are nonvolatile
functions in its SELECT list.

            regards, tom lane

Re: 8.1 vs 8.2.1 view optimization

From
Tom Lane
Date:
I wrote:
> The point is that the view won't be flattened if there are nonvolatile
> functions in its SELECT list.

Sheesh ... s/nonvolatile/volatile/ of course ... this thread seems
afflicted with getting-it-backward disease :-(

            regards, tom lane

Re: 8.1 vs 8.2.1 view optimization

From
Martijn van Oosterhout
Date:
On Tue, Jan 16, 2007 at 02:55:08PM -0700, Nathan Bell wrote:
> Yeah, I saw the "not optimized out" typo as soon as I hit send.
>
> What if the item that is taking a long time isn't a function, but rather
> a sub-select?

The planner should be able to see that the item is non-volatile itself.
It only needs to be told for functions because it can't see into them.

> Can I set the sub-select to stable, or perhaps set the entire view to
> non-volatile to achieve the same result?

No.

> If not, can I set the sub-select to a different non-volatile view or do
> I need to create a non-volatile function that returns the result?

You could create a function that does the job, but that's generally not
necessary.

Have a nice day,
--
Martijn van Oosterhout   <kleptog@svana.org>   http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.

Attachment