Update the interval division to throw same type exceptions as Spark

[HaoYang670]

Signed-off-by: remzi [email protected]
This is a subtask of #5196.

Rationale of this PR

We should throw the same type exceptions as Spark when doing interval division.

Changes in this PR

1. Check the divByZero before we do interval division, which includes:

• column / column
• column / scalar
• scalar / column.

3. Use the overflowInIntegralDivideError in Spark instead of implementing our own one.
4. Split the test cases of test_day_time_interval_division_overflow based on the kind of errors they throw.
5. Add more tests to cover interval division with scalars
6. Update some other exceptions to use the RapidsErrorUtils as much as possible.

[HaoYang670]

build

[res-life]

If we use a new error SparkArithmeticException, it's better to remove the original error ArithmeticException.
Can we just update the original check?

Original check

    if (IntervalUtils.hasZero(q)) {
      throw new ArithmeticException("overflow: interval / zero")
    }

Original error:

java.lang.ArithmeticException: overflow: interval / zero
	at org.apache.spark.sql.rapids.shims.IntervalUtils$.divWithHalfUpModeWithOverflowCheck(intervalExpressions.scala:246)

[res-life]

Should also check another path, the numScalar can be zero.

override def doColumnar(interval: GpuColumnVector, numScalar: GpuScalar)

[HaoYang670]

Should also check another path, the numScalar can be zero.

override def doColumnar(interval: GpuColumnVector, numScalar: GpuScalar)

Updated!

[HaoYang670]

If we use a new error SparkArithmeticException, it's better to remove the original error ArithmeticException. Can we just update the original check?

Original check

    if (IntervalUtils.hasZero(q)) {
      throw new ArithmeticException("overflow: interval / zero")
    }

Original error:

java.lang.ArithmeticException: overflow: interval / zero
	at org.apache.spark.sql.rapids.shims.IntervalUtils$.divWithHalfUpModeWithOverflowCheck(intervalExpressions.scala:246)

Updated!.

BTW, I don't update this error:

  def divWithHalfUpModeWithOverflowCheck(p: BinaryOperable, q: BinaryOperable): ColumnVector = {
    // 1. overflow check q is 0
    if (IntervalUtils.hasZero(q)) {
      throw new ArithmeticException("overflow: interval / zero")
    }
    ...

because we don't have the value origin: Origin in this context

[HaoYang670]

I guess we don't need to check the divide by zero error in the 2 scalars case:

  override def doColumnar(numRows: Int, intervalScalar: GpuScalar,
      numScalar: GpuScalar): ColumnVector = {
    withResource(GpuColumnVector.from(intervalScalar, numRows, interval.dataType)) { expandedLhs =>
      doColumnar(expandedLhs, numScalar)
    }
  }

Because the expression is checked by Spark:
(branch 22.08 spark 330)

scala> df.selectExpr("INTERVAL 1 SECOND / 0.0f").collect
org.apache.spark.SparkArithmeticException: Division by zero. Use `try_divide` to tolerate divisor being 0 and return NULL instead. If necessary set "spark.sql.ansi.enabled" to "false" (except for ANSI interval type) to bypass this error.
== SQL(line 1, position 1) ==
INTERVAL 1 SECOND / 0.0f
^^^^^^^^^^^^^^^^^^^^^^^^

  at org.apache.spark.sql.errors.QueryExecutionErrors$.divideByZeroError(QueryExecutionErrors.scala:184)
  at org.apache.spark.sql.catalyst.expressions.IntervalDivide.divideByZeroCheck(intervalExpressions.scala:617)
  at org.apache.spark.sql.catalyst.expressions.IntervalDivide.divideByZeroCheck$(intervalExpressions.scala:614)
  at org.apache.spark.sql.catalyst.expressions.DivideDTInterval.divideByZeroCheck(intervalExpressions.scala:713)
  at org.apache.spark.sql.catalyst.expressions.DivideDTInterval.nullSafeEval(intervalExpressions.scala:737)
  at org.apache.spark.sql.catalyst.expressions.BinaryExpression.eval(Expression.scala:642)
  at org.apache.spark.sql.catalyst.expressions.Alias.eval(namedExpressions.scala:158)
  ...

[HaoYang670]

build

[HaoYang670]

build

[HaoYang670]

build

[HaoYang670]

build

[revans2]

build

[revans2]

Looks like there was a network glitch in the previous build

[res-life]

LGTM, but it's better to test all the interval cases related to this change on all the Spark versions before merging this PR.

[firestarman]

LGTM