Tuning random_page_cost

Markus Wollny

Hi!

I've got a query that has a where clause on a timestamp field:

select t.board_id
, t.thread_id
from public.board_thread t
where t.last_reply <= now()-'6 months'::interval
limit 1

I've got random_page_cost set to 1.4 which is fine for most queries; yet
here the planner prefers a (slower) sequential scan:

QUERY PLAN
Limit (cost=0.00..0.14 rows=1 width=8) (actual time=2.598..2.600 rows=1
loops=1)
-> Seq Scan on board_thread t (cost=0.00..4613.87 rows=33674
width=8) (actual time=2.592..2.592 rows=1 loops=1)
Filter: (last_reply <= (now() - '6 mons'::interval))
Total runtime: 2.711 ms

When I set random_page_cost extremely low, the planner makes a better
decision; the "breaking point" ist at set random_page_cost = 0.16938,
which seems much to unrealistic to use as a general setting:

QUERY PLAN
Limit (cost=0.00..0.14 rows=1 width=8) (actual time=0.142..0.143 rows=1
loops=1)
-> Index Scan using idx_bt_last_reply on board_thread t
(cost=0.00..4613.72 rows=33674 width=8) (actual time=0.137..0.137 rows=1
loops=1)
Index Cond: (last_reply <= (now() - '6 mons'::interval))
Total runtime: 0.281 ms

When I change my query slightly so as to retrieve only rows with an
"equals" match, the planner prefers the index-scan alright, even with
the 1.4-setting:

select t.board_id
, t.thread_id
from public.board_thread t
where t.last_reply <= now()-'6 months'::interval
limit 1

QUERY PLAN
Limit (cost=0.00..1.83 rows=1 width=8) (actual time=0.321..0.321 rows=0
loops=1)
-> Index Scan using idx_bt_last_reply on board_thread t
(cost=0.00..3.66 rows=2 width=8) (actual time=0.315..0.315 rows=0
loops=1)
Index Cond: (last_reply = (now() - '6 mons'::interval))
Total runtime: 0.457 ms

Now here's my question: Is there some other screw to turn in order to
tune general planner behaviour so it better matches this specific kind
of query, too, or is it preferrable to leave general settings as they
are and just use a set enable_seqscan = off; before this type of query?

Kind regards

Markus

Nov 23 '05 #1

Bruno Wolff III

On Tue, Jul 13, 2004 at 13:30:29 +0200,
Markus Wollny <Ma***********@computec.de> wrote:

Hi!

I've got a query that has a where clause on a timestamp field:

select t.board_id
, t.thread_id
from public.board_thread t
where t.last_reply <= now()-'6 months'::interval
limit 1

I've got random_page_cost set to 1.4 which is fine for most queries; yet
here the planner prefers a (slower) sequential scan:

If you know that an index scan is better you can fudge the query to
change the planner's estimate of the number of rows that will be returned.
The normal fudge is to add a >= check for some timestamp that is earlier
than any in the column, so that you will have a range condition, but not
any change in the rows returned.

---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to ma*******@postgresql.org

Nov 23 '05 #2

Scott Marlowe

On Tue, 2004-07-13 at 05:30, Markus Wollny wrote:

Hi!

I've got a query that has a where clause on a timestamp field:

select t.board_id
, t.thread_id
from public.board_thread t
where t.last_reply <= now()-'6 months'::interval
limit 1

I've got random_page_cost set to 1.4 which is fine for most queries;
yet here the planner prefers a (slower) sequential scan:

QUERY PLAN
Limit (cost=0.00..0.14 rows=1 width=8) (actual time=2.598..2.600
rows=1 loops=1)
-> Seq Scan on board_thread t (cost=0.00..4613.87 rows=33674
width=8) (actual time=2.592..2.592 rows=1 loops=1)
Filter: (last_reply <= (now() - '6 mons'::interval))
Total runtime: 2.711 ms
Did you happen to notice that the estimated number of rows is 33674 and
the actual number is only 1?

Are you analyzing this table, and if so, have you tried upping your
target statistics on the column here?
When I set random_page_cost extremely low, the planner makes a better
decision; the "breaking point" ist at set random_page_cost = 0.16938,
which seems much to unrealistic to use as a general setting:

Yes, but this is the wrong way. First, the planner needs the right
numbers. Bonking it on the head with a sledge hammer is not how to get
it to behave. Analyzing with a high enough stats target is.

Note that these types of questions are generally better handled on the
performance mailing list.
---------------------------(end of broadcast)---------------------------
TIP 8: explain analyze is your friend

Nov 23 '05 #3