doc: Document caveats of swap_inputs() interface in join executors (#17373)

* doc caveats of `swap_inputs()` interface in join executors

* fix ci

* Update datafusion/physical-plan/src/joins/hash_join/exec.rs

Co-authored-by: Jonathan Chen <[email protected]>

---------

Co-authored-by: Jonathan Chen <[email protected]>

Author: 2010YOUY01

datafusion/physical-plan/src/joins/cross_join.rs

@@ -175,6 +175,12 @@ impl CrossJoinExec {
     /// Returns a new `ExecutionPlan` that computes the same join as this one,
     /// with the left and right inputs swapped using the  specified
     /// `partition_mode`.
+    ///
+    /// # Notes:
+    ///
+    /// This function should be called BEFORE inserting any repartitioning
+    /// operators on the join's children. Check [`super::HashJoinExec::swap_inputs`]
+    /// for more details.
     pub fn swap_inputs(&self) -> Result<Arc<dyn ExecutionPlan>> {
         let new_join =
             CrossJoinExec::new(Arc::clone(&self.right), Arc::clone(&self.left));

datafusion/physical-plan/src/joins/hash_join/exec.rs

@@ -638,6 +638,21 @@ impl HashJoinExec {
     ///
     /// This function is public so other downstream projects can use it to
     /// construct `HashJoinExec` with right side as the build side.
+    ///
+    /// For using this interface directly, please refer to below:
+    ///
+    /// Hash join execution may require specific input partitioning (for example,
+    /// the left child may have a single partition while the right child has multiple).
+    ///
+    /// Calling this function on join nodes whose children have already been repartitioned
+    /// (e.g., after a `RepartitionExec` has been inserted) may break the partitioning
+    /// requirements of the hash join. Therefore, ensure you call this function
+    /// before inserting any repartitioning operators on the join's children.
+    ///
+    /// In DataFusion's default SQL interface, this function is used by the `JoinSelection`
+    /// physical optimizer rule to determine a good join order, which is
+    /// executed before the `EnforceDistribution` rule (the rule that may
+    /// insert `RepartitionExec` operators).
     pub fn swap_inputs(
         &self,
         partition_mode: PartitionMode,

datafusion/physical-plan/src/joins/nested_loop_join.rs

@@ -348,6 +348,12 @@ impl NestedLoopJoinExec {
 
     /// Returns a new `ExecutionPlan` that runs NestedLoopsJoins with the left
     /// and right inputs swapped.
+    ///
+    /// # Notes:
+    ///
+    /// This function should be called BEFORE inserting any repartitioning
+    /// operators on the join's children. Check [`super::HashJoinExec::swap_inputs`]
+    /// for more details.
     pub fn swap_inputs(&self) -> Result<Arc<dyn ExecutionPlan>> {
         let left = self.left();
         let right = self.right();

datafusion/physical-plan/src/joins/sort_merge_join/exec.rs

@@ -304,6 +304,11 @@ impl SortMergeJoinExec {
         ))
     }
 
+    /// # Notes:
+    ///
+    /// This function should be called BEFORE inserting any repartitioning
+    /// operators on the join's children. Check [`super::super::HashJoinExec::swap_inputs`]
+    /// for more details.
     pub fn swap_inputs(&self) -> Result<Arc<dyn ExecutionPlan>> {
         let left = self.left();
         let right = self.right();