diff --git a/src/query/sql/src/planner/optimizer/optimizers/rule/scalar_rules/rule_merge_eval_scalar.rs b/src/query/sql/src/planner/optimizer/optimizers/rule/scalar_rules/rule_merge_eval_scalar.rs
index ce3589089b7e2..08abba6261dcc 100644
--- a/src/query/sql/src/planner/optimizer/optimizers/rule/scalar_rules/rule_merge_eval_scalar.rs
+++ b/src/query/sql/src/planner/optimizer/optimizers/rule/scalar_rules/rule_merge_eval_scalar.rs
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::BTreeSet;
+use std::collections::HashMap;
 use std::sync::Arc;
 
 use databend_common_exception::Result;
@@ -22,9 +24,14 @@ use crate::optimizer::ir::SExpr;
 use crate::optimizer::optimizers::rule::Rule;
 use crate::optimizer::optimizers::rule::RuleID;
 use crate::optimizer::optimizers::rule::TransformResult;
+use crate::plans::walk_expr_mut;
 use crate::plans::EvalScalar;
 use crate::plans::RelOp;
+use crate::plans::ScalarItem;
+use crate::plans::VisitorMut;
 use crate::ColumnSet;
+use crate::IndexType;
+use crate::ScalarExpr;
 
 // Merge two adjacent `EvalScalar`s into one
 pub struct RuleMergeEvalScalar {
@@ -60,25 +67,19 @@ impl Rule for RuleMergeEvalScalar {
     fn apply(&self, s_expr: &SExpr, state: &mut TransformResult) -> Result<()> {
         let up_eval_scalar: EvalScalar = s_expr.plan().clone().try_into()?;
         let down_eval_scalar: EvalScalar = s_expr.child(0)?.plan().clone().try_into()?;
+
         let mut used_columns = ColumnSet::new();
-        for item in up_eval_scalar.items.iter() {
-            used_columns = used_columns
-                .union(&item.scalar.used_columns())
-                .cloned()
-                .collect();
-        }
+        let merged_items = Self::merge_items(up_eval_scalar, down_eval_scalar, &mut used_columns)?;
 
         let rel_expr = RelExpr::with_s_expr(s_expr.child(0)?);
         let input_prop = rel_expr.derive_relational_prop_child(0)?;
-        // Check if the up EvalScalar depends on the down EvalScalar
+
+        // Check that all used columns are available
         if used_columns.is_subset(&input_prop.output_columns) {
             // TODO(leiysky): eliminate duplicated scalars
-            let items = up_eval_scalar
-                .items
-                .into_iter()
-                .chain(down_eval_scalar.items)
-                .collect();
-            let merged = EvalScalar { items };
+            let merged = EvalScalar {
+                items: merged_items,
+            };
 
             let new_expr = SExpr::create_unary(
                 Arc::new(merged.into()),
@@ -100,3 +101,60 @@ impl Default for RuleMergeEvalScalar {
         Self::new()
     }
 }
+
+impl RuleMergeEvalScalar {
+    fn merge_items(
+        up_eval_scalar: EvalScalar,
+        down_eval_scalar: EvalScalar,
+        used_columns: &mut BTreeSet<IndexType>,
+    ) -> Result<Vec<ScalarItem>> {
+        let mut replace_set = HashMap::with_capacity(down_eval_scalar.items.len());
+
+        for item in &down_eval_scalar.items {
+            replace_set.insert(item.index, item.scalar.clone());
+        }
+
+        struct ReplaceColumnVisitor {
+            replace_set: HashMap<IndexType, ScalarExpr>,
+        }
+
+        impl VisitorMut<'_> for ReplaceColumnVisitor {
+            fn visit(&mut self, expr: &'_ mut ScalarExpr) -> Result<()> {
+                if let ScalarExpr::BoundColumnRef(column_ref) = expr {
+                    if let Some(v) = self.replace_set.get(&column_ref.column.index) {
+                        *expr = v.clone();
+                    }
+
+                    return Ok(());
+                }
+
+                walk_expr_mut(self, expr)
+            }
+        }
+
+        let mut visitor = ReplaceColumnVisitor { replace_set };
+
+        let mut new_items = down_eval_scalar.items;
+        for mut item in up_eval_scalar.items {
+            // Skip #X AS #X
+            if let ScalarExpr::BoundColumnRef(column_ref) = &item.scalar {
+                if column_ref.column.index == item.index
+                    && visitor.replace_set.contains_key(&item.index)
+                {
+                    continue;
+                }
+            }
+
+            visitor.visit(&mut item.scalar)?;
+
+            *used_columns = used_columns
+                .union(&item.scalar.used_columns())
+                .cloned()
+                .collect();
+
+            new_items.push(item);
+        }
+
+        Ok(new_items)
+    }
+}