How to optimize this Postgres count query - postgresql

Question

EXPLAIN ANALYZE
SELECT count(*)
FROM "businesses"
WHERE (
source = 'facebook'
OR EXISTS(
SELECT *
FROM provider_business_map pbm
WHERE
pbm.hotstepper_business_id=businesses.id
AND pbm.provider_name='facebook'
)
);
PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Aggregate (cost=233538965.74..233538965.75 rows=1 width=0) (actual time=116169.720..116169.721 rows=1 loops=1)
-> Seq Scan on businesses (cost=0.00..233521096.48 rows=7147706 width=0) (actual time=11.284..116165.646 rows=3693 loops=1)
Filter: (((source)::text = 'facebook'::text) OR (alternatives: SubPlan 1 or hashed SubPlan 2))
SubPlan 1
-> Index Scan using idx_provider_hotstepper_business on provider_business_map pbm (cost=0.00..16.29 rows=1 width=0) (never executed)
Index Cond: (((provider_name)::text = 'facebook'::text) AND (hotstepper_business_id = businesses.id))
SubPlan 2
-> Index Scan using idx_provider_hotstepper_business on provider_business_map pbm (cost=0.00..16.28 rows=1 width=4) (actual time=0.045..5.685 rows=3858 loops=1)
Index Cond: ((provider_name)::text = 'facebook'::text)
Total runtime: 116169.820 ms
(10 rows)
This query takes over a minute and it's doing a count that results in ~3000. It seems the bottleneck is the sequential scan but I'm not sure what index I would need on the database to optimize this. It's also worth noting that I haven't tuned postgres so if there's any tuning that would help it may be worth considering. Although my DB is 15GB and I don't plan on trying to fit all of that in memory anytime soon so I'm not sure changing RAM related values would help a lot.

Answer 1

OR is notorious for lousy performance. Try splitting it into a union of two completely separate queries on the two tables:
SELECT COUNT(*) FROM (
SELECT id
FROM businesses
WHERE source = 'facebook'
UNION -- union makes the ids unique in the result
SELECT hotstepper_business_id
FROM provider_business_map
WHERE provider_name = 'facebook'
AND hotstepper_business_id IS NOT NULL
) x
If hotstepper_business_id can not be null, you may remove the line
AND hotstepper_business_id IS NOT NULL
If you want the whole business row, you'd could simply wrap the above query with an IN (...):
SELECT * FROM businesses
WHERE ID IN (
-- above inner query
)
But a much better performing query would be to modify the above query use use a join:
SELECT *
FROM businesses
WHERE source = 'facebook'
UNION
SELECT b.*
FROM provider_business_map m
JOIN businesses b
ON b.id = m.hotstepper_business_id
WHERE provider_name = 'facebook'

Answer 2

I'd at least try rewriting the dependent subquery as;
SELECT COUNT(DISTINCT b.*)
FROM businesses b
LEFT JOIN provider_business_map pbm
ON b.id=pbm.hotstepper_business_id
WHERE b.source = 'facebook'
OR pbm.provider_name = 'facebook';
Unless I'm mis-reading something, an index on businesses.id exists, but make sure there are also indexes on provider_business_map.hotstepper_business_id, businesses.source and provider_business_map.provider_name for best performance.

Answer 3

create index index_name on businesses(source);
Since there 3,693 rows matches in more than 7 million rows it will probably use the index. Do not forget to
analyse businesses;