Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
3.4.0
Description
When calculate slope and intercept using regr_slope & regr_intercept aggregate:
(using Java api)
spark.sql("drop table if exists tab"); spark.sql("CREATE TABLE tab(y int, x int) using parquet"); spark.sql("INSERT INTO tab VALUES (1, 1)"); spark.sql("INSERT INTO tab VALUES (2, 3)"); spark.sql("INSERT INTO tab VALUES (3, 5)"); spark.sql("INSERT INTO tab VALUES (NULL, 3)"); spark.sql("INSERT INTO tab VALUES (3, NULL)"); spark.sql("SELECT " + "regr_slope(x, y), " + "regr_intercept(x, y)" + "FROM tab").show();
Spark result:
+------------------+--------------------+ | regr_slope(x, y)|regr_intercept(x, y)| +------------------+--------------------+ |1.4545454545454546| 0.09090909090909083| +------------------+--------------------+
The correct answer should be 2.0 and -1.0 obviously.
Reason:
In sql/catalyst/expressions/aggregate/linearRegression.scala,
case class RegrSlope(left: Expression, right: Expression) extends DeclarativeAggregate with ImplicitCastInputTypes with BinaryLike[Expression] { private val covarPop = new CovPopulation(right, left) private val varPop = new VariancePop(right) ......
CovPopulation will filter tuples which right OR left is NULL
But VariancePop will only filter null right expression.
This will cause wrong result when some of the tuples' left is null (and right is not null).
Same reason with RegrIntercept.
A possible fix:
case class RegrSlope(left: Expression, right: Expression) extends DeclarativeAggregate with ImplicitCastInputTypes with BinaryLike[Expression] { private val covarPop = new CovPopulation(right, left) private val varPop = new VariancePop(If(And(IsNotNull(left), IsNotNull(right)), right, Literal.create(null, right.dataType))) .....
same fix to RegrIntercept