I'm running a simple query with 2 inner joins (say A, B and C). Each of the join columns has indexes. If I run queries that join just A and B, or just B and C, postgres uses indexes. But if I run "A join B join C" then the "B join C" part starts using a sequential scan and I can't figure out why.
Here's the query, which basically retrieves all meta-data for all messages in a given forum. The relationship is pretty simple. Forums contain threads, which contain messages, which each have associated meta-data:
SELECT message.message_id, message_meta_data.value FROM thread JOIN message USING (thread_id) JOIN message_meta_data ON (message.message_id=message_meta_data.message_id) WHERE thread.forum_id=123;
Explaining: Hash Join (cost=337.93..1267.54 rows=180 width=35) Hash Cond: ("outer".message_id = "inner".message_id) -> Seq Scan on message_meta_data (cost=0.00..739.19 rows=37719 width=30) -> Hash (cost=337.79..337.79 rows=57 width=13) -> Nested Loop (cost=0.00..337.79 rows=57 width=13) -> Index Scan using thread_forum_id_idx on thread (cost=0.00..41.61 rows=13 width=4) Index Cond: (forum_id = 6) -> Index Scan using message_thread_id_idx on message (cost=0.00..22.72 rows=5 width=17) Index Cond: ("outer".thread_id = message.thread_id)
As you can see, the message and message_meta_data tables use a Seq Scan. The only way I can think of forcing it to use the Index Scan in all cases would be to use two separate nested queries: The outer query would retrieve the list of messages in the forum, and the inner query would retrieve the list of metadata for an individual message. Obviously I want to avoid having to do that if possible.