Index "misbehavior" in PostgreSQL 8.2.2?

Salut,

I have four tables, two of which are relevant for this:

searchengine=# \d websites
Table "public.websites"
Column | Type | Modifiers
------------+--------------------------+-------------------------------------------------------
id | bigint | not null default nextval('websites_id_seq'::regclass)
url | text | not null
title | text |
abstract | text |
lastindex | timestamp with time zone |
authority | bigint | not null default 0
failed | boolean | not null default false
spamminess | smallint | not null default 0
Indexes:
"websites_pkey" PRIMARY KEY, btree (id)
"websites_url_key" UNIQUE, btree (url)
"websites_authority_key" btree (authority)
"websites_failed_key" btree (failed)
"websites_lastindex_key" btree (lastindex)
"websites_spamminess_key" btree (spamminess)

searchengine=# SELECT pg_size_pretty (pg_relation_size ('websites'));
pg_size_pretty
----------------
293 MB
(1 row)

searchengine=# SELECT COUNT(*) FROM websites;
count
--------
828778
(1 row)

searchengine=# \d sitewords
Table "public.sitewords"
Column | Type | Modifiers
------------+--------+-----------
id_site | bigint | not null
id_keyword | bigint | not null
count | bigint | not null
ratio | bigint |
Indexes:
"sitewords_key" UNIQUE, btree (id_site, id_keyword)
"sitewords_id_keyword_key" btree (id_keyword)
"sitewords_id_site_key" btree (id_site)

searchengine=# SELECT COUNT(*) FROM sitewords;
count
----------
46696260
(1 row)

searchengine=# SELECT pg_size_pretty (pg_relation_size ('sitewords'));
pg_size_pretty
----------------
3041 MB
(1 row)

searchengine=#

My effective_cache_size is set to 768MB.

I am trying to run a query:

searchengine=# EXPLAIN ANALYZE SELECT COUNT(me.id_keyword) * SUM(ratio) AS relevance, id_site FROM sitewords me JOIN websites ws ON me.id_site = ws.id WHERE me.id_keyword IN (4241,28303) GROUP BY id_site ORDER BY relevance DESC LIMIT 10;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=150319.90..150319.91 rows=3 width=24) (actual time=9211.903..9212.013 rows=10 loops=1)
-> Sort (cost=150319.90..150319.91 rows=3 width=24) (actual time=9211.895..9211.931 rows=10 loops=1)
Sort Key: ((count(me.id_keyword))::numeric * sum(me.ratio))
-> HashAggregate (cost=150319.81..150319.87 rows=3 width=24) (actual time=9180.744..9197.080 rows=3320 loops=1)
-> Hash Join (cost=60470.11..150116.47 rows=27112 width=24) (actual time=7816.860..9151.601 rows=5747 loops=1)
Hash Cond: (me.id_site = ws.id)
-> Bitmap Heap Scan on sitewords me (cost=609.31..85356.43 rows=27112 width=24) (actual time=3.245..137.026 rows=5748 loops=1)
Recheck Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Bitmap Index Scan on sitewords_id_keyword_key (cost=0.00..602.53 rows=27112 width=0) (actual time=2.169..2.169 rows=5748 loops=1)
Index Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Hash (cost=45674.80..45674.80 rows=816080 width=8) (actual time=7786.940..7786.940 rows=829078 loops=1)
-> Seq Scan on websites ws (cost=0.00..45674.80 rows=816080 width=8) (actual time=0.026..4126.069 rows=829078 loops=1)
Total runtime: 9212.679 ms
(13 rows)

searchengine=#

The only way to get PostgreSQL to use the index appears to be to set the
random_page_cost to 1.0 (1.1 also doesn't work). In that case, I get:

searchengine=# EXPLAIN ANALYZE SELECT COUNT(me.id_keyword) * SUM(ratio) AS relevance, id_site FROM sitewords me JOIN websites ws ON me.id_site = ws.id WHERE me.id_keyword IN (4241,28303) GROUP BY id_site ORDER BY relevance DESC LIMIT 10;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=78647.65..78647.66 rows=3 width=24) (actual time=418.368..418.478 rows=10 loops=1)
-> Sort (cost=78647.65..78647.66 rows=3 width=24) (actual time=418.359..418.395 rows=10 loops=1)
Sort Key: ((count(me.id_keyword))::numeric * sum(me.ratio))
-> HashAggregate (cost=78647.57..78647.63 rows=3 width=24) (actual time=386.734..403.241 rows=3320 loops=1)
-> Nested Loop (cost=310.29..78444.23 rows=27112 width=24) (actual time=2.878..354.418 rows=5747 loops=1)
-> Bitmap Heap Scan on sitewords me (cost=310.29..26849.19 rows=27112 width=24) (actual time=2.789..135.915 rows=5748 loops=1)
Recheck Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Bitmap Index Scan on sitewords_id_keyword_key (cost=0.00..303.51 rows=27112 width=0) (actual time=1.762..1.762 rows=5748 loops=1)
Index Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Index Scan using websites_pkey on websites ws (cost=0.00..1.89 rows=1 width=8) (actual time=0.017..0.022 rows=1 loops=5748)
Index Cond: (me.id_site = ws.id)
Total runtime: 418.791 ms
(12 rows)

searchengine=#

which is fine. However, if I leave random_page_cost at 4 and do the following
query:

searchengine=# EXPLAIN ANALYZE select x.ratio * x.count, x.id_site from (SELECT me.id_site, count(me.id_site), sum(me.ratio) as ratio FROM sitewords me WHERE me.id_keyword IN (4241, 28303) group by id_site) x join websites ws on (ws.id = x.id_site);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------
Nested Loop (cost=85559.77..85585.13 rows=3 width=48) (actual time=169.707..337.717 rows=3320 loops=1)
-> HashAggregate (cost=85559.77..85559.82 rows=3 width=16) (actual time=169.567..184.975 rows=3321 loops=1)
-> Bitmap Heap Scan on sitewords me (cost=609.31..85356.43 rows=27112 width=16) (actual time=2.897..137.590 rows=5748 loops=1)
Recheck Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Bitmap Index Scan on sitewords_id_keyword_key (cost=0.00..602.53 rows=27112 width=0) (actual time=1.820..1.820 rows=5748 loops=1)
Index Cond: (id_keyword = ANY ('{4241,28303}'::bigint[]))
-> Index Scan using websites_pkey on websites ws (cost=0.00..8.41 rows=1 width=8) (actual time=0.025..0.029 rows=1 loops=3321)
Index Cond: (ws.id = x.id_site)
Total runtime: 351.053 ms
(9 rows)

searchengine=#

the result is even more acceptable than with the above JOIN statement.

The tables are analyzed, and I have vacuumed regularly.

Any idea why PostgreSQL attempts a full sequential scan here?

Tonnerre