GitHub user maropu opened a pull request:
https://github.com/apache/spark/pull/18882
[SPARK-21652][SQL] Filter out meaningless constraints inferred in
inferAdditionalConstraints
## What changes were proposed in this pull request?
This pr added code to filter out meaningless constraints inferred in
`inferAdditionalConstraints` (e.g., given constraint `a = 1`, `b = 1`, `a = c`,
and `b = c`, we inferred `a = b` and this predicate was trivially true). These
constraints possibly cause some `Optimizer` overhead and, for example;
```
scala> Seq((1, 2)).toDF("col1", "col2").write.saveAsTable("t1")
scala> Seq(1, 2).toDF("col").write.saveAsTable("t2")
scala> spark.sql("SELECT * FROM t1, t2 WHERE t1.col1 = 1 AND 1 = t1.col2
AND t1.col1 = t2.col AND t1.col2 = t2.col").explain(true)
```
In this query, `InferFiltersFromConstraints` infers a new constraint
'(col2#33 = col1#32)' that is appended to the join condition, then
`PushPredicateThroughJoin` pushes it down, `ConstantPropagation` replaces
'(col2#33 = col1#32)' with '1 = 1' based on other propagated constraints,
`ConstantFolding` replaces '1 = 1' with 'true and `BooleanSimplification`
finally removes this predicate. However, `InferFiltersFromConstraints` will
again infer '(col2#33 = col1#32)' on the next iteration and the process will
continue until the limit of iterations is reached.
See below for more details
```
=== Applying Rule
org.apache.spark.sql.catalyst.optimizer.InferFiltersFromConstraints ===
!Join Inner, ((col1#32 = col#34) && (col2#33 = col#34))
Join Inner, ((col2#33 = col1#32) && ((col1#32 = col#34) &&
(col2#33 = col#34)))
:- Filter ((isnotnull(col1#32) && isnotnull(col2#33)) && ((col1#32 = 1) &&
(1 = col2#33))) :- Filter ((isnotnull(col1#32) && isnotnull(col2#33)) &&
((col1#32 = 1) && (1 = col2#33)))
: +- Relation[col1#32,col2#33] parquet
: +- Relation[col1#32,col2#33] parquet
+- Filter ((1 = col#34) && isnotnull(col#34))
+- Filter ((1 = col#34) && isnotnull(col#34))
+- Relation[col#34] parquet
+- Relation[col#34] parquet
=== Applying Rule
org.apache.spark.sql.catalyst.optimizer.PushPredicateThroughJoin ===
!Join Inner, ((col2#33 = col1#32) && ((col1#32 = col#34) && (col2#33 =
col#34))) Join Inner, ((col1#32 = col#34) && (col2#33 = col#34))
!:- Filter ((isnotnull(col1#32) && isnotnull(col2#33)) && ((col1#32 = 1) &&
(1 = col2#33))) :- Filter (col2#33 = col1#32)
!: +- Relation[col1#32,col2#33] parquet
: +- Filter ((isnotnull(col1#32) && isnotnull(col2#33)) &&
((col1#32 = 1) && (1 = col2#33)))
!+- Filter ((1 = col#34) && isnotnull(col#34))
: +- Relation[col1#32,col2#33] parquet
! +- Relation[col#34] parquet
+- Filter ((1 = col#34) && isnotnull(col#34))
!
+- Relation[col#34] parquet
=== Applying Rule org.apache.spark.sql.catalyst.optimizer.CombineFilters ===
Join Inner, ((col1#32 = col#34) && (col2#33 = col#34))
Join Inner, ((col1#32 = col#34) && (col2#33 = col#34))
!:- Filter (col2#33 = col1#32)
:- Filter (((isnotnull(col1#32) && isnotnull(col2#33)) &&
((col1#32 = 1) && (1 = col2#33))) && (col2#33 = col1#32))
!: +- Filter ((isnotnull(col1#32) && isnotnull(col2#33)) && ((col1#32 = 1)
&& (1 = col2#33))) : +- Relation[col1#32,col2#33] parquet
!: +- Relation[col1#32,col2#33] parquet
+- Filter ((1 = col#34) && isnotnull(col#34))
!+- Filter ((1 = col#34) && isnotnull(col#34))
+- Relation[col#34] parquet
! +- Relation[col#34] parquet
=== Applying Rule
org.apache.spark.sql.catalyst.optimizer.ConstantPropagation ===
Join Inner, ((col1#32 = col#34) && (col2#33 = col#34))
Join Inner, ((col1#32 = col#34) &&
(col2#33 = col#34))
!:- Filter (((isnotnull(col1#32) && isnotnull(col2#33)) && ((col1#32 = 1)
&& (1 = col2#33))) && (col2#33 = col1#32)) :- Filter (((isnotnull(col1#32) &&
isnotnull(col2#33)) && ((col1#32 = 1) && (1 = col2#33))) && (1 = 1))
: +- Relation[col1#32,col2#33] parquet
