## What changes were proposed in this pull request?
This PR contains 3 optimizations:
1) it improves significantly the operation `--` on `AttributeSet`. As a benchmark for the `--` operation, the following code has been run
```
test("AttributeSet -- benchmark") {
val attrSetA = AttributeSet((1 to 100).map { i => AttributeReference(s"c$i", IntegerType)() })
val attrSetB = AttributeSet(attrSetA.take(80).toSeq)
val attrSetC = AttributeSet((1 to 100).map { i => AttributeReference(s"c2_$i", IntegerType)() })
val attrSetD = AttributeSet((attrSetA.take(50) ++ attrSetC.take(50)).toSeq)
val attrSetE = AttributeSet((attrSetC.take(50) ++ attrSetA.take(50)).toSeq)
val n_iter = 1000000
val t0 = System.nanoTime()
(1 to n_iter) foreach { _ =>
val r1 = attrSetA -- attrSetB
val r2 = attrSetA -- attrSetC
val r3 = attrSetA -- attrSetD
val r4 = attrSetA -- attrSetE
}
val t1 = System.nanoTime()
val totalTime = t1 - t0
println(s"Average time: ${totalTime / n_iter} us")
}
```
The results are:
```
Before PR - Average time: 67674 us (100 %)
After PR - Average time: 28827 us (42.6 %)
```
2) In `ColumnPruning`, it replaces the occurrences of `(attributeSet1 -- attributeSet2).nonEmpty` with `attributeSet1.subsetOf(attributeSet2)` which is order of magnitudes more efficient (especially where there are many attributes). Running the previous benchmark replacing `--` with `subsetOf` returns:
```
Average time: 67 us (0.1 %)
```
3) Provides a more efficient way of building `AttributeSet`s, which can greatly improve the performance of the methods `references` and `outputSet` of `Expression` and `QueryPlan`. This basically avoids unneeded operations (eg. creating many `AttributeEqual` wrapper classes which could be avoided)
The overall effect of those optimizations has been tested on `ColumnPruning` with the following benchmark:
```
test("ColumnPruning benchmark") {
val attrSetA = (1 to 100).map { i => AttributeReference(s"c$i", IntegerType)() }
val attrSetB = attrSetA.take(80)
val attrSetC = attrSetA.take(20).map(a => Alias(Add(a, Literal(1)), s"${a.name}_1")())
val input = LocalRelation(attrSetA)
val query1 = Project(attrSetB, Project(attrSetA, input)).analyze
val query2 = Project(attrSetC, Project(attrSetA, input)).analyze
val query3 = Project(attrSetA, Project(attrSetA, input)).analyze
val nIter = 100000
val t0 = System.nanoTime()
(1 to nIter).foreach { _ =>
ColumnPruning(query1)
ColumnPruning(query2)
ColumnPruning(query3)
}
val t1 = System.nanoTime()
val totalTime = t1 - t0
println(s"Average time: ${totalTime / nIter} us")
}
```
The output of the test is:
```
Before PR - Average time: 733471 us (100 %)
After PR - Average time: 362455 us (49.4 %)
```
The performance improvement has been evaluated also on the `SQLQueryTestSuite`'s queries:
```
(before) org.apache.spark.sql.catalyst.optimizer.ColumnPruning 518413198 / 1377707172 2756 / 15717
(after) org.apache.spark.sql.catalyst.optimizer.ColumnPruning 415432579 / 1121147950 2756 / 15717
% Running time 80.1% / 81.3%
```
Also other rules benefit especially from (3), despite the impact is lower, eg:
```
(before) org.apache.spark.sql.catalyst.analysis.Analyzer$ResolveReferences 307341442 / 623436806 2154 / 16480
(after) org.apache.spark.sql.catalyst.analysis.Analyzer$ResolveReferences 290511312 / 560962495 2154 / 16480
% Running time 94.5% / 90.0%
```
The reason why the impact on the `SQLQueryTestSuite`'s queries is lower compared to the other benchmark is that the optimizations are more significant when the number of attributes involved is higher. Since in the tests we often have very few attributes, the effect there is lower.
## How was this patch tested?
run benchmarks + existing UTs
Closes#22364 from mgaido91/SPARK-25379.
Authored-by: Marco Gaido <marcogaido91@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
44a71741d5