Обсуждение: plan question - query with order by and limit not choosing index depends on size of limit, table

Thanks for the assistance.  

Here is an explain analyze of the query with the problem limit:

production=# explain analyze select * from landing_page.messages where ((messages.topic = E'x') AND (messages.processed = 'f'))  ORDER BY messages.created_at ASC limit 10;

                                                                                QUERY PLAN                    

                                                            

--------------------------------------------------------------------------------------------------------------

------------------------------------------------------------

 Limit  (cost=0.00..2891.06 rows=10 width=1340) (actual time=207922.586..207922.586 rows=0 loops=1)

   ->  Index Scan using idx_landing_page_messages_created_at on messages  (cost=0.00..449560.48 rows=1555 widt

h=1340) (actual time=207922.581..207922.581 rows=0 loops=1)

         Filter: ((NOT processed) AND ((topic)::text = 'x'::text))

 Total runtime: 207949.413 ms

(4 rows)

and an explain analyze with a higher limit that hits the index:

production=# explain analyze select * from landing_page.messages where ((messages.topic = E'x') AND (messages.processed = 'f'))  ORDER BY messages.created_at ASC limit 25;

                                                                      QUERY PLAN                              

                                         

--------------------------------------------------------------------------------------------------------------

-----------------------------------------

 Limit  (cost=5885.47..5885.54 rows=25 width=1340) (actual time=80.931..80.931 rows=0 loops=1)

   ->  Sort  (cost=5885.47..5889.36 rows=1555 width=1340) (actual time=80.926..80.926 rows=0 loops=1)

         Sort Key: created_at

         Sort Method:  quicksort  Memory: 17kB

         ->  Bitmap Heap Scan on messages  (cost=60.45..5841.59 rows=1555 width=1340) (actual time=64.404..64.

404 rows=0 loops=1)

               Recheck Cond: ((topic)::text = 'x'::text)

               Filter: (NOT processed)

               ->  Bitmap Index Scan on idx_messages_topic_processed  (cost=0.00..60.06 rows=1550 width=0) (ac

tual time=56.207..56.207 rows=0 loops=1)

                     Index Cond: (((topic)::text = 'x'::text) AND (p

rocessed = false))

 Total runtime: 88.051 ms

(10 rows)

overrides in postgresql.conf

shared_buffers = 256MB

work_mem = 8MB

max_fsm_pages = 2000000

max_fsm_relations = 2000

checkpoint_segments = 10

archive_mode = on

random_page_cost = 3.0

effective_cache_size = 6GB

default_statistics_target = 250

logging_collector = on

Forgot to mention this is Postgres 8.3.8 with 6GB memory on the server.

When you ask how big is the active portion of the database I am not sure how to answer.  The whole database server is about 140GB, but there are other applications that use this database, this particular table is about 1.6GB and growing.  Currently there are jobs that query from this table every minute.

Thanks again

Mike

On Wed, Jan 5, 2011 at 5:10 PM, Kevin Grittner <Kevin.Grittner@wicourts.gov> wrote:

Mike Broers <mbroers@gmail.com> wrote:

> Hello performance, I need help explaining the performance of a
> particular query

You provided some of the information needed, but you should review
this page and post a bit more:

http://wiki.postgresql.org/wiki/SlowQueryQuestions

In particular, post the result of EXPLAIN ANALYZE, not just EXPLAIN.
Also, showing all overrides in your postgresql.conf file is
important, and some information about your hardware.  How big is the
active portion of your database (the frequently read portion)?

> Why does the smaller limit cause it to skip the index?

Because the optimizer thinks the query will return rows sooner that
way.

> Is there a way to help the planner choose the better plan?

You might get there by adjusting your memory settings and/or costing
settings, but we need to see more information to know that.

-Kevin

On Thu, Jan 6, 2011 at 4:36 PM, Mike Broers <mbroers@gmail.com> wrote:
> Thanks for the assistance.
> Here is an explain analyze of the query with the problem limit:
> production=# explain analyze select * from landing_page.messages where
> ((messages.topic = E'x') AND (messages.processed = 'f'))  ORDER BY
> messages.created_at ASC limit 10;
>
>    QUERY PLAN
>
> --------------------------------------------------------------------------------------------------------------
> ------------------------------------------------------------
>  Limit  (cost=0.00..2891.06 rows=10 width=1340) (actual
> time=207922.586..207922.586 rows=0 loops=1)
>    ->  Index Scan using idx_landing_page_messages_created_at on messages
>  (cost=0.00..449560.48 rows=1555 widt
> h=1340) (actual time=207922.581..207922.581 rows=0 loops=1)
>          Filter: ((NOT processed) AND ((topic)::text = 'x'::text))
>  Total runtime: 207949.413 ms
> (4 rows)

You're not the first person to have been bitten by this.  The
optimizer thinks that rows WHERE NOT processed and topic = 'x' are
reasonably common, so it figures that it can just index scan until it
finds 10 of them.  But when it turns out that there are none at all,
it ends up having to scan the entire index, which stinks big-time.

The alternative plan is to use a different index to find ALL the
relevant rows, sort them, and then take the top 10.   That would suck
if there actually were tons of rows like this, but there aren't.

So the root of the problem, in some sense, is that the planner's
estimate of the selectivity of "NOT processed and topic = 'x'" is not
very good.  Some things to try:

- increase the statistics target for the "processed" and "topic"
columns even higher
- put the processed rows in one table and the not processed rows in
another table
- do something like SELECT * FROM (SELECT .. LIMIT 200 OFFSET 0) LIMIT
10 to try to fool the planner into planning based on the higher, inner
limit
- create a partial index on messages (topic) WHERE NOT processed and
see if the planner will use it

...Robert