I think you meant the changes introduced by bugfix 6708183 (disabled by default), right?
Some details are in note 6708183.8 including the references to other enhancements in the same area

Thanks,
Mauro

seems few of the keywords got missed relational operator and value after level and rowum, not sure if you left them out intentionally.

Pleasure reading your posts.

1000 Rows – getting selected for removal (First collection)
– Index Range scan 500 Rows (Second Collection)
As you said 1000 Fetches 500 (worst case 1000 * 500) else average of (500*500) – as said across in description

Why can’t we prode 1000 rows go for Hash Anti join with needed 500 rows, that will lead to max and min (500) verifications
instead of average

some thing like

Only delaing across with two collections separately
select x1, x2, x3 from table(cast(:b2 as jpl_table3)) v2
select x1 from table(cast(:b1 as jpl_table)) v1

Result to join across with source table t1

Hash Join (Anti) or Hash Join
HasH Join
Collection 2 (Probing first collection) that make my CR less and CPU inturn
Collection 1
T1

I was trying for that..

it looks as if the optimizer has some new ways (including dynamic sampling) of handling collections and subquery manipulation of collections in 11.2

True – it’s cardinality feedback. Another thing is Oracle uses collection size at run-time with the help of bind peeking.