Update test_add_overflow_with_ansi_enabled and test_subtraction_overflow_with_ansi_enabled to check the exception type for Integral case.

[HaoYang670]

Signed-off-by: remzi 13716567376yh@gmail.com
This a subtask of #5196.

Changes in this PR.

Update test_add_overflow_with_ansi_enabled and test_subtraction_overflow_with_ansi_enabled to check the type of exception thrown.
Please note that we only update the IntegralType case and don't modify the DecimalType case, because it is difficult to get a Spark's Decimal from a cuDF's ColumnView, which makes it hard to use the RapidsErrorUtil.cannotChangeDecimalPrecisionError: https://github.com/NVIDIA/spark-rapids/blob/branch-22.08/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuCast.scala#L1501-L1516

[HaoYang670]

build

[jlowe]

Please note that we only update the IntegralType case and don't modify the DecimalType case, because it is difficult to get a Spark's Decimal from a cuDF's ColumnView, which makes it hard to use the RapidsErrorUtil.cannotChangeDecimalPrecisionError:

Why are we trying to get the Spark type from the cudf type? We know both the Spark type of the input and the Spark type of the output from the query plan. We may not know it in fixDecimalBounds given its limited parameters, but the code that eventually calls that method knows the input and output decimal types.

[jlowe]

build

[jlowe]

build

[jlowe]

build

[jlowe]

Hiccup on the artifact server, kicking the build again

[jlowe]

build

[HaoYang670]

We may not know it in fixDecimalBounds given its limited parameters, but the code that eventually calls that method knows the input and output decimal types.

The fixDecimalBounds method is called from here: https://github.com/NVIDIA/spark-rapids/blob/branch-22.08/sql-plugin/src/main/scala/com/nvidia/spark/rapids/decimalExpressions.scala#L45-L60.
If we want to get the decimal value, we should also pass input to the GpuCast.checkNFixDecimalBounds. Otherwise, we have to rebuild the GpuColumnVector from ColumnView in fixDecimalBounds, which is difficult.

As this would cause API changes or maybe more memory allocation, I'm not sure whether this is worth to do when our target is just to match the error message with Spark's.

[tgravescs]

build

[jlowe]

If we want to get the decimal value, we should also pass input to the GpuCast.checkNFixDecimalBounds.

Yep, that or separately pass the input DataType. Either works.

As this would cause API changes or maybe more memory allocation

We own the APIs in question (they are internal, not public), so this would not be a major change. There's no additional memory allocation involved to pass an existing object instance to a method, not sure what you're getting at there?

Determining the input datatype takes no additional memory, but besides the datatype we would need to extract a specific, offending value from the input to call cannotChangeDecimalPrecisionError which does require additional GPU processing and memory. As such I'm fine with this just matching the base exception type for now. We can put in the effort to extract a value in a followup PR if deemed necessary.