Gate new ScalarSubqueryExec node behind session property (#22530)

## Which issue does this PR close?

Related to discussion on #21240 and
#21080 (comment).


PR #21240 introduced `ScalarSubqueryExec` / `ScalarSubqueryExpr` to
execute uncorrelated scalar subqueries during physical execution. The
two communicate via shared in process state (a `slot` in
`ExecutionProps`), which breaks distributed execution that may split
execution across a network boundary between the producer
(`ScalarSubqueryExec`) and the consumer expression
(`ScalarSubqueryExpr`). See more details on this explanation in
[datafusion-contrib/datafusion-distributed#460](datafusion-contrib/datafusion-distributed#460)

## What changes are included in this PR?

Adds a new optimizer config option
`datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery`
(default true, preserving the current behavior). When true (default),
behavior is unchanged from current main; when false, all scalar
subqueries are rewritten to left joins by `ScalarSubqueryToJoin` and
`ScalarSubqueryExec` is never constructed (which was the previous
behavior).

## Are these changes tested?

Yes all tests pass and added
`uncorrelated_scalar_subquery_rewritten_when_flag_off` to test the
negative case.

## Are there any user-facing changes?

Yes, a new config option
`datafusion.optimizer.physical_uncorrelated_scalar_subquery` (this just
changes the way the query is executed but not the results)

Author: LiaCastaneda

datafusion/common/src/config.rs

@@ -1124,6 +1124,22 @@ config_namespace! {
         /// into the file scan phase.
         pub enable_topk_dynamic_filter_pushdown: bool, default = true
 
+        /// When set to true, uncorrelated scalar subqueries are
+        /// left in the logical plan and executed by `ScalarSubqueryExec` during
+        /// physical execution. When set to false, all scalar subqueries
+        /// (including uncorrelated ones) are rewritten to left joins by the
+        /// `ScalarSubqueryToJoin` optimizer rule.
+        ///
+        /// Note disabling this option is not recommended. It restores
+        /// pre <https://github.com/apache/datafusion/pull/21240>
+        /// behavior, which silently produces incorrect results for
+        /// multi-row subqueries and does not support scalar subqueries in
+        /// ORDER BY / JOIN ON / aggregate-function arguments. This option is
+        /// intended as a temporary escape hatch for distributed execution
+        /// frameworks and is planned to be removed in a future DataFusion
+        /// release.
+        pub enable_physical_uncorrelated_scalar_subquery: bool, default = true
+
         /// When set to true, the optimizer will attempt to push down Join dynamic filters
         /// into the file scan phase.
         pub enable_join_dynamic_filter_pushdown: bool, default = true

datafusion/core/src/physical_planner.rs

@@ -437,7 +437,20 @@ impl DefaultPhysicalPlanner {
         session_state: &'a SessionState,
     ) -> futures::future::BoxFuture<'a, Result<Arc<dyn ExecutionPlan>>> {
         Box::pin(async move {
-            let all_subqueries = Self::collect_scalar_subqueries(logical_plan);
+            // When `enable_physical_uncorrelated_scalar_subquery` is disabled, the
+            // `ScalarSubqueryToJoin` optimizer rule rewrites all uncorrelated
+            // scalar subqueries to joins, so none should reach this point.
+            // Skip collection in that case to avoid creating a no-op
+            // `ScalarSubqueryExec` wrapper.
+            let all_subqueries = if session_state
+                .config_options()
+                .optimizer
+                .enable_physical_uncorrelated_scalar_subquery
+            {
+                Self::collect_scalar_subqueries(logical_plan)
+            } else {
+                Vec::new()
+            };
             let (links, index_map) = self
                 .plan_scalar_subqueries(all_subqueries, session_state)
                 .await?;

datafusion/optimizer/src/scalar_subquery_to_join.rs

@@ -15,7 +15,7 @@
 // specific language governing permissions and limitations
 // under the License.
 
-//! [`ScalarSubqueryToJoin`] rewriting correlated scalar subquery filters to `JOIN`s
+//! [`ScalarSubqueryToJoin`] rewriting scalar subquery filters to `JOIN`s
 
 use std::collections::{BTreeSet, HashMap};
 use std::sync::Arc;
@@ -36,9 +36,14 @@ use datafusion_expr::logical_plan::{JoinType, Subquery};
 use datafusion_expr::utils::conjunction;
 use datafusion_expr::{Expr, LogicalPlan, LogicalPlanBuilder, lit, not, when};
 
-/// Optimizer rule that rewrites correlated scalar subquery filters to joins and
-/// places an additional projection on top of the filter, to preserve the
-/// original schema.
+/// Optimizer rule that rewrites scalar subquery filters to joins and places an
+/// additional projection on top of the filter to preserve the original schema.
+///
+/// When [`datafusion_common::config::OptimizerOptions::enable_physical_uncorrelated_scalar_subquery`] is
+/// true (the default), only *correlated* scalar subqueries are rewritten here;
+/// uncorrelated ones are left for physical execution via `ScalarSubqueryExec`.
+/// When the option is false, all scalar subqueries — correlated and
+/// uncorrelated — are rewritten to left joins by this rule.
 #[derive(Default, Debug)]
 pub struct ScalarSubqueryToJoin {}
 
@@ -63,10 +68,12 @@ impl ScalarSubqueryToJoin {
         &self,
         predicate: &Expr,
         alias_gen: &Arc<AliasGenerator>,
+        physical_uncorrelated: bool,
     ) -> Result<(Vec<(Subquery, String)>, Expr)> {
         let mut extract = ExtractScalarSubQuery {
             sub_query_info: vec![],
             alias_gen,
+            physical_uncorrelated,
         };
         predicate
             .clone()
@@ -88,15 +95,23 @@ impl OptimizerRule for ScalarSubqueryToJoin {
     ) -> Result<Transformed<LogicalPlan>> {
         match plan {
             LogicalPlan::Filter(filter) => {
+                let physical_uncorrelated = config
+                    .options()
+                    .optimizer
+                    .enable_physical_uncorrelated_scalar_subquery;
                 // Optimization: skip the rest of the rule and its copies if
-                // there are no scalar subqueries
-                if !contains_correlated_scalar_subquery(&filter.predicate) {
+                // there are no scalar subqueries this rule should rewrite
+                if !contains_scalar_subquery_to_rewrite(
+                    &filter.predicate,
+                    physical_uncorrelated,
+                ) {
                     return Ok(Transformed::no(LogicalPlan::Filter(filter)));
                 }
 
                 let (subqueries, mut rewrite_expr) = self.extract_subquery_exprs(
                     &filter.predicate,
                     config.alias_generator(),
+                    physical_uncorrelated,
                 )?;
 
                 assert_or_internal_err!(
@@ -141,13 +156,15 @@ impl OptimizerRule for ScalarSubqueryToJoin {
                 Ok(Transformed::yes(new_plan))
             }
             LogicalPlan::Projection(projection) => {
+                let physical_uncorrelated = config
+                    .options()
+                    .optimizer
+                    .enable_physical_uncorrelated_scalar_subquery;
                 // Optimization: skip the rest of the rule and its copies if there
-                // are no correlated scalar subqueries
-                if !projection
-                    .expr
-                    .iter()
-                    .any(contains_correlated_scalar_subquery)
-                {
+                // are no scalar subqueries this rule should rewrite
+                if !projection.expr.iter().any(|expr| {
+                    contains_scalar_subquery_to_rewrite(expr, physical_uncorrelated)
+                }) {
                     return Ok(Transformed::no(LogicalPlan::Projection(projection)));
                 }
 
@@ -156,8 +173,11 @@ impl OptimizerRule for ScalarSubqueryToJoin {
                 let mut rewrite_exprs: Vec<Expr> =
                     Vec::with_capacity(projection.expr.len());
                 for (idx, expr) in projection.expr.iter().enumerate() {
-                    let (subqueries, rewrite_expr) =
-                        self.extract_subquery_exprs(expr, config.alias_generator())?;
+                    let (subqueries, rewrite_expr) = self.extract_subquery_exprs(
+                        expr,
+                        config.alias_generator(),
+                        physical_uncorrelated,
+                    )?;
                     for (_, alias) in &subqueries {
                         alias_to_index.insert(alias.clone(), idx);
                     }
@@ -228,29 +248,42 @@ impl OptimizerRule for ScalarSubqueryToJoin {
     }
 }
 
-/// Returns true if the expression contains a correlated scalar subquery, false
-/// otherwise.  Uncorrelated scalar subqueries are handled by the physical
-/// planner via `ScalarSubqueryExec` and do not need to be converted to joins.
-fn contains_correlated_scalar_subquery(expr: &Expr) -> bool {
+/// Returns true if the expression contains a scalar subquery that this rule
+/// should rewrite to a join.
+///
+/// When `enable_physical_uncorrelated_scalar_subquery` is true (the default) only
+/// correlated scalar subqueries are rewritten — uncorrelated ones are handled
+/// by the physical planner via `ScalarSubqueryExec`. When it is false, all
+/// scalar subqueries (correlated and uncorrelated) are rewritten.
+fn contains_scalar_subquery_to_rewrite(expr: &Expr, physical_uncorrelated: bool) -> bool {
     expr.exists(|expr| {
-        Ok(matches!(expr, Expr::ScalarSubquery(sq) if !sq.outer_ref_columns.is_empty()))
+        Ok(matches!(
+            expr,
+            Expr::ScalarSubquery(sq)
+                if !physical_uncorrelated || !sq.outer_ref_columns.is_empty()
+        ))
     })
     .expect("Inner is always Ok")
 }
 
 struct ExtractScalarSubQuery<'a> {
     sub_query_info: Vec<(Subquery, String)>,
     alias_gen: &'a Arc<AliasGenerator>,
+    physical_uncorrelated: bool,
 }
 
 impl TreeNodeRewriter for ExtractScalarSubQuery<'_> {
     type Node = Expr;
 
     fn f_down(&mut self, expr: Expr) -> Result<Transformed<Expr>> {
         match expr {
-            // Skip uncorrelated scalar subqueries
+            // Match scalar subqueries this rule should rewrite to a join. When
+            // `physical_uncorrelated` is true, only correlated subqueries are
+            // rewritten — uncorrelated ones are handled later by the physical
+            // planner. When false, both are rewritten.
             Expr::ScalarSubquery(ref subquery)
-                if !subquery.outer_ref_columns.is_empty() =>
+                if !self.physical_uncorrelated
+                    || !subquery.outer_ref_columns.is_empty() =>
             {
                 let subquery = subquery.clone();
                 let scalar_expr = subquery
@@ -288,9 +321,15 @@ impl TreeNodeRewriter for ExtractScalarSubQuery<'_> {
 /// where c.balance > o."avg(total)"
 /// ```
 ///
+/// When [`datafusion_common::config::OptimizerOptions::enable_physical_uncorrelated_scalar_subquery`] is
+/// false, this function also handles uncorrelated scalar subqueries, rewriting
+/// them as a `Left Join: Filter: Boolean(true)` instead of leaving them for
+/// `ScalarSubqueryExec`.
+///
 /// # Arguments
 ///
-/// * `subquery` - The correlated scalar subquery to decorrelate.
+/// * `subquery` - The scalar subquery to rewrite (correlated, or uncorrelated
+///   when `enable_physical_uncorrelated_scalar_subquery` is false).
 /// * `outer_input` - The outer plan that the decorrelated subquery is
 ///   left-joined onto — the input of the `Filter` or `Projection` node
 ///   that contained the subquery.
@@ -308,10 +347,9 @@ fn build_join(
     outer_input: &LogicalPlan,
     subquery_alias: &str,
 ) -> Result<Option<(LogicalPlan, HashMap<Column, Expr>)>> {
-    assert_or_internal_err!(
-        !subquery.outer_ref_columns.is_empty(),
-        "build_join should only be called for correlated subqueries"
-    );
+    // `build_join` also handles uncorrelated scalar subqueries (as a left
+    // join with `Boolean(true)`) when the
+    // `enable_physical_uncorrelated_scalar_subquery` option is disabled.
     let subquery_plan = subquery.subquery.as_ref();
     let mut pull_up = PullUpCorrelatedExpr::new().with_need_handle_count_bug(true);
     let decorrelated_subquery = subquery_plan.clone().rewrite(&mut pull_up).data()?;
@@ -1159,4 +1197,52 @@ mod tests {
         "
         )
     }
+
+    #[test]
+    fn uncorrelated_scalar_subquery_rewritten_when_flag_off() -> Result<()> {
+        use datafusion_common::config::ConfigOptions;
+
+        let sq = Arc::new(
+            LogicalPlanBuilder::from(scan_tpch_table("orders"))
+                .aggregate(Vec::<Expr>::new(), vec![max(col("orders.o_custkey"))])?
+                .project(vec![max(col("orders.o_custkey"))])?
+                .build()?,
+        );
+
+        let plan = LogicalPlanBuilder::from(scan_tpch_table("customer"))
+            .filter(col("customer.c_custkey").eq(scalar_subquery(sq)))?
+            .project(vec![col("customer.c_custkey")])?
+            .build()?;
+
+        let mut options = ConfigOptions::default();
+        options
+            .optimizer
+            .enable_physical_uncorrelated_scalar_subquery = false;
+        let context = crate::OptimizerContext::new_with_config_options(Arc::new(options));
+
+        let rule: Arc<dyn OptimizerRule + Send + Sync> =
+            Arc::new(ScalarSubqueryToJoin::new());
+        let optimizer = crate::Optimizer::with_rules(vec![rule]);
+        let optimized_plan = optimizer
+            .optimize(plan, &context, |_, _| {})
+            .expect("failed to optimize plan");
+        let formatted_plan = optimized_plan.display_indent_schema();
+
+        insta::assert_snapshot!(
+            formatted_plan,
+            @r"
+        Projection: customer.c_custkey [c_custkey:Int64]
+          Projection: customer.c_custkey, customer.c_name [c_custkey:Int64, c_name:Utf8]
+            Filter: customer.c_custkey = __scalar_sq_1.max(orders.o_custkey) [c_custkey:Int64, c_name:Utf8, max(orders.o_custkey):Int64;N]
+              Left Join:  Filter: Boolean(true) [c_custkey:Int64, c_name:Utf8, max(orders.o_custkey):Int64;N]
+                TableScan: customer [c_custkey:Int64, c_name:Utf8]
+                SubqueryAlias: __scalar_sq_1 [max(orders.o_custkey):Int64;N]
+                  Projection: max(orders.o_custkey) [max(orders.o_custkey):Int64;N]
+                    Aggregate: groupBy=[[]], aggr=[[max(orders.o_custkey)]] [max(orders.o_custkey):Int64;N]
+                      TableScan: orders [o_orderkey:Int64, o_custkey:Int64, o_orderstatus:Utf8, o_totalprice:Float64;N]
+        "
+        );
+
+        Ok(())
+    }
 }

datafusion/sqllogictest/test_files/information_schema.slt

@@ -303,6 +303,7 @@ datafusion.optimizer.enable_distinct_aggregation_soft_limit true
 datafusion.optimizer.enable_dynamic_filter_pushdown true
 datafusion.optimizer.enable_join_dynamic_filter_pushdown true
 datafusion.optimizer.enable_leaf_expression_pushdown true
+datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery true
 datafusion.optimizer.enable_piecewise_merge_join false
 datafusion.optimizer.enable_round_robin_repartition true
 datafusion.optimizer.enable_sort_pushdown true
@@ -453,6 +454,7 @@ datafusion.optimizer.enable_distinct_aggregation_soft_limit true When set to tru
 datafusion.optimizer.enable_dynamic_filter_pushdown true When set to true attempts to push down dynamic filters generated by operators (TopK, Join & Aggregate) into the file scan phase. For example, for a query such as `SELECT * FROM t ORDER BY timestamp DESC LIMIT 10`, the optimizer will attempt to push down the current top 10 timestamps that the TopK operator references into the file scans. This means that if we already have 10 timestamps in the year 2025 any files that only have timestamps in the year 2024 can be skipped / pruned at various stages in the scan. The config will suppress `enable_join_dynamic_filter_pushdown`, `enable_topk_dynamic_filter_pushdown` & `enable_aggregate_dynamic_filter_pushdown` So if you disable `enable_topk_dynamic_filter_pushdown`, then enable `enable_dynamic_filter_pushdown`, the `enable_topk_dynamic_filter_pushdown` will be overridden.
 datafusion.optimizer.enable_join_dynamic_filter_pushdown true When set to true, the optimizer will attempt to push down Join dynamic filters into the file scan phase.
 datafusion.optimizer.enable_leaf_expression_pushdown true When set to true, the optimizer will extract leaf expressions (such as `get_field`) from filter/sort/join nodes into projections closer to the leaf table scans, and push those projections down towards the leaf nodes.
+datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery true When set to true, uncorrelated scalar subqueries are left in the logical plan and executed by `ScalarSubqueryExec` during physical execution. When set to false, all scalar subqueries (including uncorrelated ones) are rewritten to left joins by the `ScalarSubqueryToJoin` optimizer rule. Note disabling this option is not recommended. It restores pre <https://github.com/apache/datafusion/pull/21240> behavior, which silently produces incorrect results for multi-row subqueries and does not support scalar subqueries in ORDER BY / JOIN ON / aggregate-function arguments. This option is intended as a temporary escape hatch for distributed execution frameworks and is planned to be removed in a future DataFusion release.
 datafusion.optimizer.enable_piecewise_merge_join false When set to true, piecewise merge join is enabled. PiecewiseMergeJoin is currently experimental. Physical planner will opt for PiecewiseMergeJoin when there is only one range filter.
 datafusion.optimizer.enable_round_robin_repartition true When set to true, the physical plan optimizer will try to add round robin repartitioning to increase parallelism to leverage more CPU cores
 datafusion.optimizer.enable_sort_pushdown true Enable sort pushdown optimization. When enabled, attempts to push sort requirements down to data sources that can natively handle them (e.g., by reversing file/row group read order). Returns **inexact ordering**: Sort operator is kept for correctness, but optimized input enables early termination for TopK queries (ORDER BY ... LIMIT N), providing significant speedup. Memory: No additional overhead (only changes read order). Future: Will add option to detect perfectly sorted data and eliminate Sort completely. Default: true

datafusion/sqllogictest/test_files/subquery.slt

@@ -2153,6 +2153,95 @@ SELECT (SELECT v FROM (SELECT 1 AS v UNION ALL SELECT 2) AS t ORDER BY v LIMIT 1
 ----
 1
 
+#############
+## End-to-end correctness coverage for the flag-off path.
+## When `datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery` is false,
+## uncorrelated scalar subqueries are rewritten to left joins by
+## `ScalarSubqueryToJoin` instead of executed by `ScalarSubqueryExec`. This
+## restores pre-PR-21240 behavior, which has two known shortcomings the
+## physical-execution path was built to fix: multi-row subqueries silently
+## return wrong results, and uncorrelated scalar subqueries do not work in
+## ORDER BY / JOIN ON / aggregate-function arguments. Those cases are
+## intentionally not covered here; the queries below are the ones where both
+## paths agree.
+#############
+
+statement ok
+set datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery = false;
+
+# Scalar subquery returning exactly one row → success
+query I
+SELECT (SELECT v FROM sq_values LIMIT 1);
+----
+1
+
+# Scalar subquery returning exactly one row in WHERE → success
+query I rowsort
+SELECT x FROM sq_main WHERE x > (SELECT v FROM sq_values LIMIT 1);
+----
+10
+20
+
+# Scalar subquery returning zero rows → NULL
+query I
+SELECT (SELECT v FROM sq_empty);
+----
+NULL
+
+# Scalar subquery returning zero rows in arithmetic → NULL propagation
+query I
+SELECT x + (SELECT v FROM sq_empty) FROM sq_main;
+----
+NULL
+NULL
+
+# Scalar subquery returning zero rows in WHERE comparison → no matching rows
+query I
+SELECT x FROM sq_main WHERE x > (SELECT v FROM sq_empty);
+----
+
+# Aggregated subquery always returns one row, even on empty input → success
+query I
+SELECT (SELECT count(*) FROM sq_empty);
+----
+0
+
+# Aggregated subquery on multi-row table → success
+query I
+SELECT (SELECT max(v) FROM sq_values);
+----
+3
+
+# HAVING clause with uncorrelated scalar subquery
+query II rowsort
+SELECT x, count(*) AS cnt FROM sq_main GROUP BY x
+HAVING count(*) > (SELECT min(v) FROM sq_values);
+----
+
+# CASE WHEN with uncorrelated scalar subquery as condition
+query T rowsort
+SELECT CASE WHEN x > (SELECT min(v) FROM sq_values)
+            THEN 'big' ELSE 'small' END AS label
+FROM sq_main;
+----
+big
+big
+
+# Doubly-nested constant subquery
+query I
+SELECT (SELECT (SELECT 42));
+----
+42
+
+# NULL comparison semantics through subquery boundary
+query B
+SELECT 1 = (SELECT CAST(NULL AS INT));
+----
+NULL
+
+statement ok
+RESET datafusion.optimizer.enable_physical_uncorrelated_scalar_subquery;
+
 statement count 0
 DROP TABLE sq_values;