Re: Optimization inner join

Gustavo Rezende Montesino <gustavo(dot)montesino(at)trtsp(dot)jus(dot)br> writes:
> Being the client in question, I would like to make a little remark: What
> we thought could be optimized here at first is on the row estimate of
> the index scan; which could take null_frac into account. To put things
> into perspective, our similar case in production has a table with 6
> million lines where only 9.5k aren´t null for the join field, an the
> over-estimation is throwing away good plans (like ~150ms execution time)
> in favor of pretty bad ones (~80s execution time).

Please provide a concrete test case for that. AFAIK the null fraction
should be accounted for in join size estimates. Here's a little test
case showing that it is:

regression=# create table t1 as select generate_series(1,1000000) as f1;
SELECT 1000000
regression=# analyze t1;
ANALYZE
regression=# create table t2 as select generate_series(1,1000000) as f1;
SELECT 1000000
regression=# analyze t2;
ANALYZE
regression=# explain select * from t1,t2 where t1.f1=t2.f1;
QUERY PLAN
------------------------------------------------------------------------
Hash Join (cost=30832.00..70728.00 rows=1000000 width=8)
Hash Cond: (t1.f1 = t2.f1)
-> Seq Scan on t1 (cost=0.00..14425.00 rows=1000000 width=4)
-> Hash (cost=14425.00..14425.00 rows=1000000 width=4)
-> Seq Scan on t2 (cost=0.00..14425.00 rows=1000000 width=4)
(5 rows)

regression=# insert into t2 select null from generate_series(1,1000000);
INSERT 0 1000000
regression=# analyze t2;
ANALYZE
regression=# explain select * from t1,t2 where t1.f1=t2.f1;
QUERY PLAN
------------------------------------------------------------------------
Hash Join (cost=30832.00..95727.00 rows=1000000 width=8)
Hash Cond: (t2.f1 = t1.f1)
-> Seq Scan on t2 (cost=0.00..27862.00 rows=2000000 width=4)
-> Hash (cost=14425.00..14425.00 rows=1000000 width=4)
-> Seq Scan on t1 (cost=0.00..14425.00 rows=1000000 width=4)
(5 rows)

The join size estimate is still correct even though it knows there are
many more rows in t2.

As for inserting a not-null test at the scan level, I'm not exactly
convinced that it's a win:

regression=# \timing
Timing is on.
regression=# select count(*) from t1,t2 where t1.f1=t2.f1;
count
---------
1000000
(1 row)

Time: 562.914 ms
regression=# select count(*) from t1,t2 where t1.f1=t2.f1 and t2.f1 is not null;
count
---------
1000000
(1 row)

Time: 564.896 ms

[ ftr, these times are best-of-three-trials ]

It's possible that in the case where an explicit sort has to be inserted,
reducing the amount of data passing through the sort would be worth doing;
but in the general case that's unproven.

regards, tom lane