Major Description
Currently, for most of the cases, the project https://issues.apache.org/jira/browse/SPARK-38615 is able to show where the runtime errors happen within the original query.
However, after trying on production, I found that the following queries won't show where the divide by 0 error happens
create table aggTest(i int, j int, k int, d date) using parquet insert into aggTest values(1, 2, 0, date'2022-01-01') select sum(j)/sum(k),percentile(i, 0.9) from aggTest group by d
With `percentile` function in the query, the plan can't execute with whole stage codegen. Thus the child plan of `Project` is serialized to executors for execution, from ProjectExec:
protected override def doExecute(): RDD[InternalRow] = {
child.execute().mapPartitionsWithIndexInternal { (index, iter) =>
val project = UnsafeProjection.create(projectList, child.output)
project.initialize(index)
iter.map(project)
}
}
Note that the `TreeNode.origin` is not serialized to executors since `TreeNode` doesn't extend the trait `Serializable`, which results in an empty query context on errors. For more details, please read https://issues.apache.org/jira/browse/SPARK-39140
A dummy fix is to make `TreeNode` extend the trait `Serializable`. However, it can be performance regression if the query text is long (every `TreeNode` carries it for serialization).
A better fix is to introduce a new trait `SupportQueryContext` and materialize the truncated query context for special expressions. This jira targets on binary arithmetic expressions only. I will create follow-ups for the remaining expressions which support runtime error query context.
