Description
We encountered a issue that join conditions not pushed down when we are running spark app on spark2.3, after carefully looking into the code and debugging, we found that it's because there is a bug in the rule `PushPredicateThroughJoin`:
It will try to push parent filter down though the join, however, when the parent filter is wholly pushed down through the join, the join will become the top node, and then the `transform` method will skip the join to apply the rule.
Suppose we have two tables: table1 and table2:
table1: (a: string, b: string, c: string)
table2: (d: string)
sql as:
select * from table1 left join (select d, 'w1' as r from table2) on a = d and r = 'w2' where b = 2
let's focus on the following optimizer rules:
PushPredicateThroughJoin
FodablePropagation
BooleanSimplification
PruneFilters
In the above case, on the first iteration of these rules:
PushPredicateThroughJoin ->
select * from table1 where b=2 left join (select d, 'w1' as r from table2) on a = d and r = 'w2'
FodablePropagation ->
select * from table1 where b=2 left join (select d, 'w1' as r from table2) on a = d and 'w1' = 'w2'
BooleanSimplification ->
select * from table1 where b=2 left join (select d, 'w1' as r from table2) on false
PruneFilters -> No effective
After several iteration of these rules, the join condition will still never be pushed to the
right hand of the left join. thus, in some case(e.g. Large right table), the `BroadcastNestedLoopJoin` may be slow or oom.