Major Description
Since SPARK-28379, Spark has supported non-aggregated single-row correlated subqueries. SPARK-40800 handles the majority of the cases where projects can be collapsed. But Spark can throw exceptions for single-row subqueries with non-deterministic expressions. For example:
CREATE TEMP VIEW t1 AS SELECT ARRAY('a', 'b') a SELECT ( SELECT array_sort(a, (i, j) -> rank[i] - rank[j])[0] + r + r AS sorted FROM (SELECT MAP('a', 1, 'b', 2) rank, rand() as r) ) FROM t1
This throws an exception:
Unexpected operator Join Inner
:- Aggregate [[a,b]], [[a,b] AS a#253]
: +- OneRowRelation
+- Project [map(keys: [a,b], values: [1,2]) AS rank#241, rand(86882494013664043) AS r#242]
+- OneRowRelation
in correlated subquery
This is because when Spark rewrites correlated subqueries, it checks whether a scalar subquery is subject to the COUNT bug. It splits the query into parts above the aggregate, the aggregate, and the parts below the aggregate (see `splitSubquery` in the `RewriteCorrelatedScalarSubquery` rule).
This pattern is very restrictive and does not work well with non-aggregated single-row subqueries. We should fix this issue.
