Обсуждение: somewhat slow query with subselect

Hi list,

I have a table, not that many rows (400k) but with about 55 columns.
There are some maintenance selects I want to do (every now and then)
that I don't want to add indexes for.

select
    ...
from
    quite_big_table qbt
where
    qbt.owner = 123 AND
    exists (
        select null
        from quite_big_table qbt2
        where
            qbt2.owner = qbt.owner AND
            qbt2.id <> qbt.id AND
            qbt2.filelength = qbt.filelength
    )

In a case with some 5000 rows belonging to owner 123, this select really
takes a long time. Way longer than without the subselect and order by
filelength. I agree that with the subselect it would take some extra
juice, but in my mind it would do some hash in memory which wouldn't be
too slow to lookup in.

shared_buffers = 16GB
temp_buffers = 1GB
work_mem = 32MB
maintenance_work_mem = 1GB
#server has 64GB (64bit)

I guess there is some penalty because of the size (# columns) of the
table since it has to go thru more blocks on the disk. Would it have
been beneficial if filelength was in a joined child table instead?

How would you do this? Create a temporary table with owner+filelenght?

Thanks!
Marcus

Marcus Engene <mengpg2@engene.se> writes:
> ... In a case with some 5000 rows belonging to owner 123, this select really
> takes a long time. Way longer than without the subselect and order by
> filelength. I agree that with the subselect it would take some extra
> juice, but in my mind it would do some hash in memory which wouldn't be
> too slow to lookup in.

8.4 can turn EXISTS subqueries into hash joins, but previous releases
won't...

            regards, tom lane

Tom Lane wrote:
> Marcus Engene <mengpg2@engene.se> writes:
>
>> ... In a case with some 5000 rows belonging to owner 123, this select really
>> takes a long time. Way longer than without the subselect and order by
>> filelength. I agree that with the subselect it would take some extra
>> juice, but in my mind it would do some hash in memory which wouldn't be
>> too slow to lookup in.
>>
>
> 8.4 can turn EXISTS subqueries into hash joins, but previous releases
> won't...
>
>             regards, tom lane
>
>
Thank you very much for your answer, Tom. I tried to join the table
instead and it was way faster.

  Sort  (cost=46769.87..46770.51 rows=258 width=48)
   Sort Key: pic.filesize
   ->  Nested Loop  (cost=34.30..46759.54 rows=258 width=48)
         Join Filter: ((picsame.objectid <> pic.objectid) AND
(pic.filesize = picsame.filesize))
         ->  Nested Loop  (cost=8.27..3099.28 rows=16 width=56)
               ->  HashAggregate  (cost=8.27..8.28 rows=1 width=4)
                     ->  Index Scan using user_c2 on user pu2
(cost=0.00..8.27 rows=1 width=4)
                           Index Cond: ((username_locase)::text =
'prolificarts'::text)
               ->  Index Scan using item_common_x1 on item_common pic
(cost=0.00..3081.41 rows=767 width=52)
                     Index Cond: (pic.user = pu2.objectid)
         ->  Bitmap Heap Scan on item_common picsame
(cost=26.03..2715.34 rows=767 width=16)
               Recheck Cond: (picsame.user = pic.user)
               ->  Bitmap Index Scan on item_common_x1
(cost=0.00..25.84 rows=767 width=0)
                     Index Cond: (picsame.user = pic.user)

Best regards,
Marcus