[SPARK-17986][ML] SQLTransformer should remove temporary tables

## What changes were proposed in this pull request?

A call to the method `SQLTransformer.transform` previously would create a temporary table and never delete it. This change adds a call to `dropTempView()` that deletes this temporary table before returning the result so that the table will not remain in spark's table catalog. Because `tableName` is randomized and not exposed, there should be no expected use of this table outside of the `transform` method.

## How was this patch tested?

A single new assertion was added to the existing test of the `SQLTransformer.transform` method that all temporary tables are removed. Without the corresponding code change, this new assertion fails. I am not aware of any circumstances in which removing this temporary view would be bad for performance or correctness in other ways, but some expertise here would be helpful.

Author: Drew Robb <[email protected]>

Closes #15526 from drewrobb/SPARK-17986.

(cherry picked from commit ab3363e)
Signed-off-by: Yanbo Liang <[email protected]>

Author: drewrobb

mllib/src/main/scala/org/apache/spark/ml/feature/SQLTransformer.scala

@@ -67,7 +67,9 @@ class SQLTransformer @Since("1.6.0") (@Since("1.6.0") override val uid: String)
     val tableName = Identifiable.randomUID(uid)
     dataset.createOrReplaceTempView(tableName)
     val realStatement = $(statement).replace(tableIdentifier, tableName)
-    dataset.sparkSession.sql(realStatement)
+    val result = dataset.sparkSession.sql(realStatement)
+    dataset.sparkSession.catalog.dropTempView(tableName)
+    result
   }
 
   @Since("1.6.0")

mllib/src/test/scala/org/apache/spark/ml/feature/SQLTransformerSuite.scala

@@ -43,6 +43,7 @@ class SQLTransformerSuite
     assert(result.schema.toString == resultSchema.toString)
     assert(resultSchema == expected.schema)
     assert(result.collect().toSeq == expected.collect().toSeq)
+    assert(original.sparkSession.catalog.listTables().count() == 0)
   }
 
   test("read/write") {