BUG #9135: PostgreSQL doesn't want use index scan instead of (index scan+sort+limit) - Mailing list pgsql-bugs

The following bug has been logged on the website:

Bug reference:      9135
Logged by:          Maxim Boguk
Email address:      maxim.boguk@gmail.com
PostgreSQL version: 9.3.2
Operating system:   Linux
Description:

Hi,

One of my customers have very curious situation with simple query and index
usage. I tried different ideas but it doesn't work anyway and now I out of
ideas. It's looks like a bug if I not missing something.

Detail:
4GB liexWebmasterProducts table with interesting fields:

 lwpid                         | integer                     | not null
default nextval('liexwebmasterproducts_lwpid_seq'::regclass)
 lwpname                       | text                        |
...
 lwpwebsiteid                  | integer                     |
...
 lwpnotforsale                 | boolean                     | not null
...
 lwpcreatedate                 | timestamp without time zone | not null
default now()
...

Index on the last three fields defined as:
    "i_liexwebmasterproducts_2" btree (lwpwebsiteid, lwpnotforsale,
lwpcreatedate)

Target query and plan:
select *
from liexWebmasterProducts this_
where
this_.lwpWebsiteId=5935
and this_.lwpnotForSale=FALSE
order by this_.lwpCreateDate desc limit 1;

 Limit  (cost=122.18..122.19 rows=1 width=902) (actual time=13.505..13.506
rows=1 loops=1)
   ->  Sort  (cost=122.18..124.57 rows=953 width=902) (actual
time=13.503..13.503 rows=1 loops=1)
         Sort Key: lwpcreatedate
         Sort Method: top-N heapsort  Memory: 27kB
         ->  Index Scan using i_liexwebmasterproducts_2 on
liexwebmasterproducts this_  (cost=0.43..117.42 rows=953 width=902) (actual
time=0.171..10.429 rows=1674 loops=1)
               Index Cond: ((lwpwebsiteid = 5935) AND (lwpnotforsale =
false))
               Filter: (NOT lwpnotforsale)
 Total runtime: 13.626 ms


I have no idea why Postgresql doesn't want use simple index scan over 3
fields...
set enable_sort to 0;
have no effect:

 Limit  (cost=10000000119.90..10000000119.90 rows=1 width=902) (actual
time=6.591..6.592 rows=1 loops=1)
   ->  Sort  (cost=10000000119.90..10000000122.24 rows=935 width=902)
(actual time=6.588..6.588 rows=1 loops=1)
         Sort Key: lwpcreatedate
         Sort Method: top-N heapsort  Memory: 27kB
         ->  Index Scan using i_liexwebmasterproducts_2 on
liexwebmasterproducts this_  (cost=0.43..115.22 rows=935 width=902) (actual
time=0.050..3.733 rows=1673 loops=1)
               Index Cond: ((lwpwebsiteid = 5935) AND (lwpnotforsale =
false))
               Filter: (NOT lwpnotforsale)
 Total runtime: 6.670 ms


It seems somehow related to the: "Filter: (NOT lwpnotforsale)" part of the
query which look like redundant, but it's my pure guessing.

Reindexing the index, vacuum analyze table - provide zero effect on the
plan.

Generating whole new subset and table via:
shop=# create table test as select lwpid,lwpwebsiteid, lwpnotforsale,
lwpcreatedate from liexwebmasterproducts;
SELECT 6799176
shop=#  create index CONCURRENTLY test_index_1 on test(lwpwebsiteid,
lwpnotforsale, lwpcreatedate);
CREATE INDEX
shop=# vacuum analyze test;
VACUUM

Have no effect as well (plan over test table stay the same).

Changing order of the two first fields in index via:
create index CONCURRENTLY test_index_2 on test(lwpnotforsale, lwpwebsiteid,
wpcreatedate);
have no effect on the plan too.

Kindly Regards,
Maksym