Improve performance on SQL query with Nested Loop - PostgreSQL

Question

I am using PostgreSQL and I have a weird problem with my SQL query. Depending on wich date paramter I'm using. My request doesn't do the same operation.

This is my working query :

SELECT DISTINCT app.id_application 
FROM stat sj
LEFT OUTER JOIN groupe gp ON gp.id_groupe = sj.id_groupe 
LEFT OUTER JOIN application app ON app.id_application = gp.id_application 
WHERE date_stat >= '2016/3/01' 
AND date_stat <= '2016/3/31' 
AND ( date_stat = date_gen-1 or (date_gen = '2016/04/01' AND date_stat = '2016/3/31')) 
AND app.id_application IS NOT NULL 

This query takes around 2 secondes (which is OKAY for me because I have a lots of rows). When I run EXPLAIN ANALYSE for this query I have this:

HashAggregate  (cost=375486.95..375493.62 rows=667 width=4) (actual time=2320.541..2320.656 rows=442 loops=1)
    ->  Hash Join  (cost=254.02..375478.99 rows=3186 width=4) (actual time=6.144..2271.984 rows=263274 loops=1)
    Hash Cond: (gp.id_application = app.id_application)
    ->  Hash Join  (cost=234.01..375415.17 rows=3186 width=4) (actual time=5.926..2200.671 rows=263274 loops=1)
          Hash Cond: (sj.id_groupe = gp.id_groupe)
          ->  Seq Scan on stat sj  (cost=0.00..375109.47 rows=3186 width=8) (actual time=3.196..2068.357 rows=263274 loops=1)
                Filter: ((date_stat >= '2016-03-01'::date) AND (date_stat <= '2016-03-31'::date) AND ((date_stat = (date_gen - 1)) OR ((date_gen = '2016-04-01'::date) AND (date_stat = '2016-03-31'::date))))
                Rows Removed by Filter: 7199514
          ->  Hash  (cost=133.45..133.45 rows=8045 width=12) (actual time=2.677..2.677 rows=8019 loops=1)
                Buckets: 1024  Batches: 1  Memory Usage: 345kB
                ->  Seq Scan on groupe gp  (cost=0.00..133.45 rows=8045 width=12) (actual time=0.007..1.284 rows=8019 loops=1)
    ->  Hash  (cost=11.67..11.67 rows=667 width=4) (actual time=0.206..0.206 rows=692 loops=1)
          Buckets: 1024  Batches: 1  Memory Usage: 25kB
          ->  Seq Scan on application app  (cost=0.00..11.67 rows=667 width=4) (actual time=0.007..0.101 rows=692 loops=1)
                Filter: (id_application IS NOT NULL)
    Total runtime: 2320.855 ms

Now, When I'm trying the same query for the current month (we are the 6th of April, so I'm trying to get all the application_id of April) with the same query

SELECT DISTINCT app.id_application 
FROM stat sj
LEFT OUTER JOIN groupe gp ON gp.id_groupe = sj.id_groupe 
LEFT OUTER JOIN application app ON app.id_application = gp.id_application 
WHERE date_stat >= '2016/04/01' 
AND date_stat <= '2016/04/30' 
AND ( date_stat = date_gen-1 or ( date_gen = '2016/05/01' AND date_job = '2016/04/30')) 
AND app.id_application IS NOT NULL 

This query takes now 120 seconds. So I also ran EXPLAIN ANALYZE on this query and now it doesn't have the same operations:

HashAggregate  (cost=375363.50..375363.51 rows=1 width=4) (actual time=186716.468..186716.532 rows=490 loops=1)
->  Nested Loop  (cost=0.00..375363.49 rows=1 width=4) (actual time=1.945..186619.404 rows=118990 loops=1)
    Join Filter: (gp.id_application = app.id_application)
    Rows Removed by Join Filter: 82222090
    ->  Nested Loop  (cost=0.00..375343.49 rows=1 width=4) (actual time=1.821..171458.237 rows=118990 loops=1)
          Join Filter: (sj.id_groupe = gp.id_groupe)
          Rows Removed by Join Filter: 954061820
          ->  Seq Scan on stat sj  (cost=0.00..375109.47 rows=1 width=8) (actual time=0.235..1964.423 rows=118990 loops=1)
                Filter: ((date_stat >= '2016-04-01'::date) AND (date_stat <= '2016-04-30'::date) AND ((date_stat = (date_gen - 1)) OR ((date_gen = '2016-05-01'::date) AND (date_stat = '2016-04-30'::date))))
                Rows Removed by Filter: 7343798
          ->  Seq Scan on groupe gp  (cost=0.00..133.45 rows=8045 width=12) (actual time=0.002..0.736 rows=8019 loops=118990)
    ->  Seq Scan on application app  (cost=0.00..11.67 rows=667 width=4) (actual time=0.003..0.073 rows=692 loops=118990)
          Filter: (id_application IS NOT NULL)
  Total runtime: 186716.635 ms

So I decided to search where the problem came from by reducing the number of conditions from my query until the performances is acceptable again.

So with only this parameter

WHERE date_stat >= '2016/04/01'

It takes only 1.9secondes (like the first working query) and it's also working with 2 parameters :

WHERE date_stat >= '2016/04/01' 
AND app.id_application IS NOT NULL 

BUT when I try to add one of those line I have the Nested loop in the Explain

AND date_stat <= '2016/04/30' 
AND ( date_stat = date_gen-1 or ( date_gen = '2016/05/01' AND date_stat = '2016/04/30')) 

Does someone have any idea where it could come from?

Solution

Ok, it looks like there's problem with optimizer estimations. He thiks that for april there will be only 1 row so he choose NESTED LOOP which is very inefficient for big number of rows (118,990 in that case).

1. Perform VACUUM ANALYZE for every table. This will clean up dead tuples and refresh statistics.
2. consider adding index based on dates like CREATE INDEX date_stat_idx ON <table with date_stat> USING btree (date_stat);

Rerun the query,

Comments

[Gabriel's Messanger]

Perform EXPLAIN ANALYZE on both queries and add both outputs to your question. Also do you have any indexes on that tables?

[a_horse_with_no_name]

The problem is that in the second query Postgres completely underestimates the rows that are returned by that condition (estimated: 1 row, actual: 118990 rows). So this looks like stale statistics (in the first query the number of rows is also underestimated, but that doesn't cause a bad plan). Check if running analyze stat; changes anything. It doesn't look like as if you have an index on stat (date_stat). Creating one should help as well.