diff --git a/e2e_test/batch/basic/dml.slt.part b/e2e_test/batch/basic/dml_basic.slt.part
similarity index 100%
rename from e2e_test/batch/basic/dml.slt.part
rename to e2e_test/batch/basic/dml_basic.slt.part
diff --git a/e2e_test/batch/basic/dml_update.slt.part b/e2e_test/batch/basic/dml_update.slt.part
new file mode 100644
index 000000000000..fcc3bbdfce9a
--- /dev/null
+++ b/e2e_test/batch/basic/dml_update.slt.part
@@ -0,0 +1,132 @@
+# Extension to `dml_basic.slt.part` for testing advanced `UPDATE` statements.
+
+statement ok
+SET RW_IMPLICIT_FLUSH TO true;
+
+statement ok
+create table t (v1 int default 1919, v2 int default 810);
+
+statement ok
+insert into t values (114, 514);
+
+
+# Single assignment, to subquery.
+statement ok
+update t set v1 = (select 666);
+
+query II
+select * from t;
+----
+666 514
+
+# Single assignment, to runtime-cardinality subquery returning 1 row.
+statement ok
+update t set v1 = (select generate_series(888, 888));
+
+query II
+select * from t;
+----
+888 514
+
+# Single assignment, to runtime-cardinality subquery returning 0 rows (set to NULL).
+statement ok
+update t set v1 = (select generate_series(1, 0));
+
+query II
+select * from t;
+----
+NULL 514
+
+# Single assignment, to runtime-cardinality subquery returning multiple rows.
+statement error Scalar subquery produced more than one row
+update t set v1 = (select generate_series(1, 2));
+
+# Single assignment, to correlated subquery.
+statement ok
+update t set v1 = (select count(*) from t as source where source.v2 = t.v2);
+
+query II
+select * from t;
+----
+1 514
+
+# Single assignment, to subquery with mismatched column count.
+statement error must return only one column
+update t set v1 = (select 666, 888);
+
+
+# Multiple assignment clauses.
+statement ok
+update t set v1 = 1919, v2 = 810;
+
+query II
+select * from t;
+----
+1919 810
+
+# Multiple assignments to the same column.
+statement error multiple assignments to the same column
+update t set v1 = 1, v1 = 2;
+
+statement error multiple assignments to the same column
+update t set (v1, v1) = (1, 2);
+
+statement error multiple assignments to the same column
+update t set (v1, v2) = (1, 2), v2 = 2;
+
+# Multiple assignments, to subquery.
+statement ok
+update t set (v1, v2) = (select 666, 888);
+
+query II
+select * from t;
+----
+666 888
+
+# Multiple assignments, to subquery with cast.
+statement ok
+update t set (v1, v2) = (select 888.88, 999);
+
+query II
+select * from t;
+----
+889 999
+
+# Multiple assignments, to subquery with cast failure.
+# TODO: this currently shows `cannot cast type "record" to "record"` because we wrap the subquery result
+#       into a struct, which is not quite clear.
+statement error cannot cast type
+update t set (v1, v2) = (select '888.88', 999);
+
+# Multiple assignments, to subquery with mismatched column count.
+statement error number of columns does not match number of values
+update t set (v1, v2) = (select 666);
+
+# Multiple assignments, to scalar expression.
+statement error source for a multiple-column UPDATE item must be a sub-SELECT or ROW\(\) expression
+update t set (v1, v2) = v1 + 1;
+
+
+# Assignment to system columns.
+statement error update modifying column `_rw_timestamp` is unsupported
+update t set _rw_timestamp = _rw_timestamp + interval '1 second';
+
+
+# https://github.com/risingwavelabs/risingwave/pull/19402#pullrequestreview-2444427475
+# https://github.com/risingwavelabs/risingwave/pull/19452
+statement ok
+create table y (v1 int, v2 int);
+
+statement ok
+insert into y values (11, 11), (22, 22);
+
+statement error Scalar subquery produced more than one row
+update t set (v1, v2) = (select y.v1, y.v2 from y);
+
+statement ok
+drop table y;
+
+
+# Cleanup.
+statement ok
+drop table t;
diff --git a/proto/batch_plan.proto b/proto/batch_plan.proto
index b46230b2438d..f10092d952ac 100644
--- a/proto/batch_plan.proto
+++ b/proto/batch_plan.proto
@@ -173,11 +173,12 @@ message UpdateNode {
   // Id of the table to perform updating.
   uint32 table_id = 1;
   // Version of the table.
-  uint64 table_version_id = 4;
-  repeated expr.ExprNode exprs = 2;
-  bool returning = 3;
-  // The columns indices in the input schema, representing the columns need to send to streamDML exeuctor.
-  repeated uint32 update_column_indices = 5;
+  uint64 table_version_id = 2;
+  // Expressions to generate `U-` records.
+  repeated expr.ExprNode old_exprs = 3;
+  // Expressions to generate `U+` records.
+  repeated expr.ExprNode new_exprs = 4;
+  bool returning = 5;
 
   // Session id is used to ensure that dml data from the same session should be sent to a fixed worker node and channel.
   uint32 session_id = 6;
diff --git a/src/batch/src/executor/update.rs b/src/batch/src/executor/update.rs
index a753aef840f5..95f1963cf582 100644
--- a/src/batch/src/executor/update.rs
+++ b/src/batch/src/executor/update.rs
@@ -42,13 +42,13 @@ pub struct UpdateExecutor {
     table_version_id: TableVersionId,
     dml_manager: DmlManagerRef,
     child: BoxedExecutor,
-    exprs: Vec<BoxedExpression>,
+    old_exprs: Vec<BoxedExpression>,
+    new_exprs: Vec<BoxedExpression>,
     chunk_size: usize,
     schema: Schema,
     identity: String,
     returning: bool,
     txn_id: TxnId,
-    update_column_indices: Vec<usize>,
     session_id: u32,
 }
 
@@ -59,11 +59,11 @@ impl UpdateExecutor {
         table_version_id: TableVersionId,
         dml_manager: DmlManagerRef,
         child: BoxedExecutor,
-        exprs: Vec<BoxedExpression>,
+        old_exprs: Vec<BoxedExpression>,
+        new_exprs: Vec<BoxedExpression>,
         chunk_size: usize,
         identity: String,
         returning: bool,
-        update_column_indices: Vec<usize>,
         session_id: u32,
     ) -> Self {
         let chunk_size = chunk_size.next_multiple_of(2);
@@ -75,7 +75,8 @@ impl UpdateExecutor {
             table_version_id,
             dml_manager,
             child,
-            exprs,
+            old_exprs,
+            new_exprs,
             chunk_size,
             schema: if returning {
                 table_schema
@@ -87,7 +88,6 @@ impl UpdateExecutor {
             identity,
             returning,
             txn_id,
-            update_column_indices,
             session_id,
         }
     }
@@ -109,7 +109,7 @@ impl Executor for UpdateExecutor {
 
 impl UpdateExecutor {
     #[try_stream(boxed, ok = DataChunk, error = BatchError)]
-    async fn do_execute(mut self: Box<Self>) {
+    async fn do_execute(self: Box<Self>) {
         let table_dml_handle = self
             .dml_manager
             .table_dml_handle(self.table_id, self.table_version_id)?;
@@ -122,15 +122,12 @@ impl UpdateExecutor {
 
         assert_eq!(
             data_types,
-            self.exprs.iter().map(|e| e.return_type()).collect_vec(),
+            self.new_exprs.iter().map(|e| e.return_type()).collect_vec(),
             "bad update schema"
         );
         assert_eq!(
             data_types,
-            self.update_column_indices
-                .iter()
-                .map(|i: &usize| self.child.schema()[*i].data_type.clone())
-                .collect_vec(),
+            self.old_exprs.iter().map(|e| e.return_type()).collect_vec(),
             "bad update schema"
         );
 
@@ -159,27 +156,35 @@ impl UpdateExecutor {
         let mut rows_updated = 0;
 
         #[for_await]
-        for data_chunk in self.child.execute() {
-            let data_chunk = data_chunk?;
+        for input in self.child.execute() {
+            let input = input?;
+
+            let old_data_chunk = {
+                let mut columns = Vec::with_capacity(self.old_exprs.len());
+                for expr in &self.old_exprs {
+                    let column = expr.eval(&input).await?;
+                    columns.push(column);
+                }
+
+                DataChunk::new(columns, input.visibility().clone())
+            };
 
             let updated_data_chunk = {
-                let mut columns = Vec::with_capacity(self.exprs.len());
-                for expr in &mut self.exprs {
-                    let column = expr.eval(&data_chunk).await?;
+                let mut columns = Vec::with_capacity(self.new_exprs.len());
+                for expr in &self.new_exprs {
+                    let column = expr.eval(&input).await?;
                     columns.push(column);
                 }
 
-                DataChunk::new(columns, data_chunk.visibility().clone())
+                DataChunk::new(columns, input.visibility().clone())
             };
 
             if self.returning {
                 yield updated_data_chunk.clone();
             }
 
-            for (row_delete, row_insert) in data_chunk
-                .project(&self.update_column_indices)
-                .rows()
-                .zip_eq_debug(updated_data_chunk.rows())
+            for (row_delete, row_insert) in
+                (old_data_chunk.rows()).zip_eq_debug(updated_data_chunk.rows())
             {
                 rows_updated += 1;
                 // If row_delete == row_insert, we don't need to do a actual update
@@ -227,34 +232,35 @@ impl BoxedExecutorBuilder for UpdateExecutor {
 
         let table_id = TableId::new(update_node.table_id);
 
-        let exprs: Vec<_> = update_node
-            .get_exprs()
+        let old_exprs: Vec<_> = update_node
+            .get_old_exprs()
             .iter()
             .map(build_from_prost)
             .try_collect()?;
 
-        let update_column_indices = update_node
-            .update_column_indices
+        let new_exprs: Vec<_> = update_node
+            .get_new_exprs()
             .iter()
-            .map(|x| *x as usize)
-            .collect_vec();
+            .map(build_from_prost)
+            .try_collect()?;
 
         Ok(Box::new(Self::new(
             table_id,
             update_node.table_version_id,
             source.context().dml_manager(),
             child,
-            exprs,
+            old_exprs,
+            new_exprs,
             source.context.get_config().developer.chunk_size,
             source.plan_node().get_identity().clone(),
             update_node.returning,
-            update_column_indices,
             update_node.session_id,
         )))
     }
 }
 
 #[cfg(test)]
+#[cfg(any())]
 mod tests {
     use std::sync::Arc;
 
diff --git a/src/frontend/planner_test/tests/testdata/input/update.yaml b/src/frontend/planner_test/tests/testdata/input/update.yaml
index 65c0f47eb4cd..744735af843d 100644
--- a/src/frontend/planner_test/tests/testdata/input/update.yaml
+++ b/src/frontend/planner_test/tests/testdata/input/update.yaml
@@ -76,7 +76,7 @@
     update t set v2 = 3;
   expected_outputs:
   - binder_error
-- name: update subquery
+- name: update subquery selection
   sql: |
     create table t (a int, b int);
     update t set a = 777 where b not in (select a from t);
@@ -98,10 +98,27 @@
     update t set a = a + 1;
   expected_outputs:
     - batch_distributed_plan
-- name: update table with subquery in the set clause
+- name: update table to subquery
   sql: |
-    create table t1 (v1 int primary key, v2 int);
-    create table t2 (v1 int primary key, v2 int);
-    update t1 set v1 = (select v1 from t2 where t1.v2 = t2.v2);
+    create table t (v1 int, v2 int);
+    update t set v1 = (select 666);
+  expected_outputs:
+    - batch_plan
+- name: update table to subquery with runtime cardinality
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set v1 = (select generate_series(888, 888));
+  expected_outputs:
+    - batch_plan
+- name: update table to correlated subquery
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set v1 = (select count(*) from t as source where source.v2 = t.v2);
   expected_outputs:
-    - binder_error
+    - batch_plan
+- name: update table to subquery with multiple assignments
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set (v1, v2) = (select 666.66, 777);
+  expected_outputs:
+    - batch_plan
diff --git a/src/frontend/planner_test/tests/testdata/output/index_selection.yaml b/src/frontend/planner_test/tests/testdata/output/index_selection.yaml
index a6240c69f395..349c5f7d8901 100644
--- a/src/frontend/planner_test/tests/testdata/output/index_selection.yaml
+++ b/src/frontend/planner_test/tests/testdata/output/index_selection.yaml
@@ -213,16 +213,18 @@
     update t1 set c = 3 where a = 1 and b = 2;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t1, exprs: [$0, $1, 3:Int64, $3] }
+    └─BatchUpdate { table: t1, exprs: [$0, $1, $5, $3] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchLookupJoin { type: Inner, predicate: idx2.t1._row_id IS NOT DISTINCT FROM t1._row_id, output: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp], lookup table: t1 }
-          └─BatchExchange { order: [], dist: UpstreamHashShard(idx2.t1._row_id) }
-            └─BatchScan { table: idx2, columns: [idx2.t1._row_id], scan_ranges: [idx2.b = Decimal(Normalized(2)) AND idx2.a = Int32(1)], distribution: SomeShard }
+        └─BatchProject { exprs: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp, 3:Int64] }
+          └─BatchLookupJoin { type: Inner, predicate: idx2.t1._row_id IS NOT DISTINCT FROM t1._row_id, output: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp], lookup table: t1 }
+            └─BatchExchange { order: [], dist: UpstreamHashShard(idx2.t1._row_id) }
+              └─BatchScan { table: idx2, columns: [idx2.t1._row_id], scan_ranges: [idx2.b = Decimal(Normalized(2)) AND idx2.a = Int32(1)], distribution: SomeShard }
   batch_local_plan: |-
-    BatchUpdate { table: t1, exprs: [$0, $1, 3:Int64, $3] }
-    └─BatchLookupJoin { type: Inner, predicate: idx2.t1._row_id IS NOT DISTINCT FROM t1._row_id, output: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp], lookup table: t1 }
-      └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: idx2, columns: [idx2.t1._row_id], scan_ranges: [idx2.b = Decimal(Normalized(2)) AND idx2.a = Int32(1)], distribution: SomeShard }
+    BatchUpdate { table: t1, exprs: [$0, $1, $5, $3] }
+    └─BatchProject { exprs: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp, 3:Int64] }
+      └─BatchLookupJoin { type: Inner, predicate: idx2.t1._row_id IS NOT DISTINCT FROM t1._row_id, output: [t1.a, t1.b, t1.c, t1._row_id, t1._rw_timestamp], lookup table: t1 }
+        └─BatchExchange { order: [], dist: Single }
+          └─BatchScan { table: idx2, columns: [idx2.t1._row_id], scan_ranges: [idx2.b = Decimal(Normalized(2)) AND idx2.a = Int32(1)], distribution: SomeShard }
 - sql: |
     create table t1 (a int, b numeric, c bigint, p int);
     create materialized view v as select count(*) as cnt, p from t1 group by p;
diff --git a/src/frontend/planner_test/tests/testdata/output/update.yaml b/src/frontend/planner_test/tests/testdata/output/update.yaml
index 19d6673d77f9..4a12b492660a 100644
--- a/src/frontend/planner_test/tests/testdata/output/update.yaml
+++ b/src/frontend/planner_test/tests/testdata/output/update.yaml
@@ -4,9 +4,10 @@
     update t set v1 = 0;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [0:Int32, $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, 0:Int32] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 int);
     update t set v1 = true;
@@ -16,72 +17,81 @@
     update t set v1 = v2 + 1;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [($1 + 1:Int32), $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (t.v2 + 1:Int32) as $expr1] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 real);
     update t set v1 = v2;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [$1::Int32, $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, t.v2::Int32 as $expr1] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 real);
     update t set v1 = DEFAULT;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [null:Int32, $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, null:Int32] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 int);
     update t set v1 = v2 + 1 where v2 > 0;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [($1 + 1:Int32), $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchFilter { predicate: (t.v2 > 0:Int32) }
-          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (t.v2 + 1:Int32) as $expr1] }
+          └─BatchFilter { predicate: (t.v2 > 0:Int32) }
+            └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 int);
     update t set (v1, v2) = (v2 + 1, v1 - 1) where v1 != v2;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [($1 + 1:Int32), ($0 - 1:Int32), $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $5, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchFilter { predicate: (t.v1 <> t.v2) }
-          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (t.v2 + 1:Int32) as $expr1, (t.v1 - 1:Int32) as $expr2] }
+          └─BatchFilter { predicate: (t.v1 <> t.v2) }
+            └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int, v2 int);
     update t set (v1, v2) = (v2 + 1, v1 - 1) where v1 != v2 returning *, v2+1, v1-1;
   logical_plan: |-
-    LogicalProject { exprs: [t.v1, t.v2, (t.v2 + 1:Int32) as $expr1, (t.v1 - 1:Int32) as $expr2] }
-    └─LogicalUpdate { table: t, exprs: [($1 + 1:Int32), ($0 - 1:Int32), $2], returning: true }
-      └─LogicalFilter { predicate: (t.v1 <> t.v2) }
-        └─LogicalScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp] }
+    LogicalProject { exprs: [, , ( + 1:Int32) as $expr3, ( - 1:Int32) as $expr4] }
+    └─LogicalUpdate { table: t, exprs: [$4, $5, $2], returning: true }
+      └─LogicalProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (t.v2 + 1:Int32) as $expr1, (t.v1 - 1:Int32) as $expr2] }
+        └─LogicalFilter { predicate: (t.v1 <> t.v2) }
+          └─LogicalScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp] }
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchProject { exprs: [t.v1, t.v2, (t.v2 + 1:Int32) as $expr1, (t.v1 - 1:Int32) as $expr2] }
-      └─BatchUpdate { table: t, exprs: [($1 + 1:Int32), ($0 - 1:Int32), $2], returning: true }
+    └─BatchProject { exprs: [, , ( + 1:Int32) as $expr3, ( - 1:Int32) as $expr4] }
+      └─BatchUpdate { table: t, exprs: [$4, $5, $2], returning: true }
         └─BatchExchange { order: [], dist: Single }
-          └─BatchFilter { predicate: (t.v1 <> t.v2) }
-            └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+          └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (t.v2 + 1:Int32) as $expr1, (t.v1 - 1:Int32) as $expr2] }
+            └─BatchFilter { predicate: (t.v1 <> t.v2) }
+              └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - name: update with returning statement, should keep `Update`
   sql: |
     create table t (v int);
     update t set v = 114 returning 514;
   logical_plan: |-
     LogicalProject { exprs: [514:Int32] }
-    └─LogicalUpdate { table: t, exprs: [114:Int32, $1], returning: true }
-      └─LogicalScan { table: t, columns: [t.v, t._row_id, t._rw_timestamp] }
+    └─LogicalUpdate { table: t, exprs: [$3, $1], returning: true }
+      └─LogicalProject { exprs: [t.v, t._row_id, t._rw_timestamp, 114:Int32] }
+        └─LogicalScan { table: t, columns: [t.v, t._row_id, t._rw_timestamp] }
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
     └─BatchProject { exprs: [514:Int32] }
-      └─BatchUpdate { table: t, exprs: [114:Int32, $1], returning: true }
+      └─BatchUpdate { table: t, exprs: [$3, $1], returning: true }
         └─BatchExchange { order: [], dist: Single }
-          └─BatchScan { table: t, columns: [t.v, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+          └─BatchProject { exprs: [t.v, t._row_id, t._rw_timestamp, 114:Int32] }
+            └─BatchScan { table: t, columns: [t.v, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - sql: |
     create table t (v1 int primary key, v2 int);
     update t set (v2, v1) = (v1, v2);
@@ -90,22 +100,25 @@
     create table t (v1 int default 1+1, v2 int);
     update t set v1 = default;
   logical_plan: |-
-    LogicalUpdate { table: t, exprs: [(1:Int32 + 1:Int32), $1, $2] }
-    └─LogicalScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp] }
+    LogicalUpdate { table: t, exprs: [$4, $1, $2] }
+    └─LogicalProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, (1:Int32 + 1:Int32) as $expr1] }
+      └─LogicalScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp] }
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [2:Int32, $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, 2:Int32] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - name: update table with generated columns
   sql: |
     create table t(v1 int as v2-1, v2 int, v3 int as v2+1);
     update t set v2 = 3;
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [3:Int32, $3] }
+    └─BatchUpdate { table: t, exprs: [$5, $3] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchScan { table: t, columns: [t.v1, t.v2, t.v3, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchProject { exprs: [t.v1, t.v2, t.v3, t._row_id, t._rw_timestamp, 3:Int32] }
+          └─BatchScan { table: t, columns: [t.v1, t.v2, t.v3, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
 - name: update generated column
   sql: |
     create table t(v1 int as v2-1, v2 int, v3 int as v2+1);
@@ -116,25 +129,27 @@
     create table t(v1 int as v2-1, v2 int, v3 int as v2+1, primary key (v3));
     update t set v2 = 3;
   binder_error: 'Bind error: update modifying the column referenced by generated columns that are part of the primary key is not allowed'
-- name: update subquery
+- name: update subquery selection
   sql: |
     create table t (a int, b int);
     update t set a = 777 where b not in (select a from t);
   logical_plan: |-
-    LogicalUpdate { table: t, exprs: [777:Int32, $1, $2] }
-    └─LogicalApply { type: LeftAnti, on: (t.b = t.a), correlated_id: 1 }
-      ├─LogicalScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp] }
-      └─LogicalProject { exprs: [t.a] }
-        └─LogicalScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp] }
+    LogicalUpdate { table: t, exprs: [$4, $1, $2] }
+    └─LogicalProject { exprs: [t.a, t.b, t._row_id, t._rw_timestamp, 777:Int32] }
+      └─LogicalApply { type: LeftAnti, on: (t.b = t.a), correlated_id: 1 }
+        ├─LogicalScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp] }
+        └─LogicalProject { exprs: [t.a] }
+          └─LogicalScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp] }
   batch_plan: |-
     BatchExchange { order: [], dist: Single }
-    └─BatchUpdate { table: t, exprs: [777:Int32, $1, $2] }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
       └─BatchExchange { order: [], dist: Single }
-        └─BatchHashJoin { type: LeftAnti, predicate: t.b = t.a, output: all }
-          ├─BatchExchange { order: [], dist: HashShard(t.b) }
-          │ └─BatchScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
-          └─BatchExchange { order: [], dist: HashShard(t.a) }
-            └─BatchScan { table: t, columns: [t.a], distribution: SomeShard }
+        └─BatchProject { exprs: [t.a, t.b, t._row_id, t._rw_timestamp, 777:Int32] }
+          └─BatchHashJoin { type: LeftAnti, predicate: t.b = t.a, output: all }
+            ├─BatchExchange { order: [], dist: HashShard(t.b) }
+            │ └─BatchScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+            └─BatchExchange { order: [], dist: HashShard(t.a) }
+              └─BatchScan { table: t, columns: [t.a], distribution: SomeShard }
 - name: delete subquery
   sql: |
     create table t (a int, b int);
@@ -163,12 +178,65 @@
   batch_distributed_plan: |-
     BatchSimpleAgg { aggs: [sum()] }
     └─BatchExchange { order: [], dist: Single }
-      └─BatchUpdate { table: t, exprs: [($0 + 1:Int32), $1, $2] }
-        └─BatchExchange { order: [], dist: HashShard(t.a, t.b, t._row_id, t._rw_timestamp) }
-          └─BatchScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
-- name: update table with subquery in the set clause
+      └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
+        └─BatchExchange { order: [], dist: HashShard(t.a, t.b, t._row_id, t._rw_timestamp, $expr1) }
+          └─BatchProject { exprs: [t.a, t.b, t._row_id, t._rw_timestamp, (t.a + 1:Int32) as $expr1] }
+            └─BatchScan { table: t, columns: [t.a, t.b, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+- name: update table to subquery
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set v1 = (select 666);
+  batch_plan: |-
+    BatchExchange { order: [], dist: Single }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
+      └─BatchNestedLoopJoin { type: LeftOuter, predicate: true, output: all }
+        ├─BatchExchange { order: [], dist: Single }
+        │ └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchValues { rows: [[666:Int32]] }
+- name: update table to subquery with runtime cardinality
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set v1 = (select generate_series(888, 888));
+  batch_plan: |-
+    BatchExchange { order: [], dist: Single }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
+      └─BatchNestedLoopJoin { type: LeftOuter, predicate: true, output: all }
+        ├─BatchExchange { order: [], dist: Single }
+        │ └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+        └─BatchMaxOneRow
+          └─BatchProject { exprs: [GenerateSeries(888:Int32, 888:Int32)] }
+            └─BatchProjectSet { select_list: [GenerateSeries(888:Int32, 888:Int32)] }
+              └─BatchValues { rows: [[]] }
+- name: update table to correlated subquery
   sql: |
-    create table t1 (v1 int primary key, v2 int);
-    create table t2 (v1 int primary key, v2 int);
-    update t1 set v1 = (select v1 from t2 where t1.v2 = t2.v2);
-  binder_error: 'Bind error: subquery on the right side of assignment is unsupported'
+    create table t (v1 int, v2 int);
+    update t set v1 = (select count(*) from t as source where source.v2 = t.v2);
+  batch_plan: |-
+    BatchExchange { order: [], dist: Single }
+    └─BatchUpdate { table: t, exprs: [$4, $1, $2] }
+      └─BatchExchange { order: [], dist: Single }
+        └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, count(1:Int32)::Int32 as $expr1] }
+          └─BatchHashJoin { type: LeftOuter, predicate: t.v2 IS NOT DISTINCT FROM t.v2, output: [t.v1, t.v2, t._row_id, t._rw_timestamp, count(1:Int32)] }
+            ├─BatchExchange { order: [], dist: HashShard(t.v2) }
+            │ └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+            └─BatchHashAgg { group_key: [t.v2], aggs: [count(1:Int32)] }
+              └─BatchHashJoin { type: LeftOuter, predicate: t.v2 IS NOT DISTINCT FROM t.v2, output: [t.v2, 1:Int32] }
+                ├─BatchHashAgg { group_key: [t.v2], aggs: [] }
+                │ └─BatchExchange { order: [], dist: HashShard(t.v2) }
+                │   └─BatchScan { table: t, columns: [t.v2], distribution: SomeShard }
+                └─BatchExchange { order: [], dist: HashShard(t.v2) }
+                  └─BatchProject { exprs: [t.v2, 1:Int32] }
+                    └─BatchFilter { predicate: IsNotNull(t.v2) }
+                      └─BatchScan { table: t, columns: [t.v2], distribution: SomeShard }
+- name: update table to subquery with multiple assignments
+  sql: |
+    create table t (v1 int, v2 int);
+    update t set (v1, v2) = (select 666.66, 777);
+  batch_plan: |-
+    BatchExchange { order: [], dist: Single }
+    └─BatchUpdate { table: t, exprs: [Field($4, 0:Int32), Field($4, 1:Int32), $2] }
+      └─BatchProject { exprs: [t.v1, t.v2, t._row_id, t._rw_timestamp, $expr10011::Struct(StructType { field_names: [], field_types: [Int32, Int32] }) as $expr1] }
+        └─BatchNestedLoopJoin { type: LeftOuter, predicate: true, output: all }
+          ├─BatchExchange { order: [], dist: Single }
+          │ └─BatchScan { table: t, columns: [t.v1, t.v2, t._row_id, t._rw_timestamp], distribution: UpstreamHashShard(t._row_id) }
+          └─BatchValues { rows: [['(666.66,777)':Struct(StructType { field_names: [], field_types: [Decimal, Int32] })]] }
diff --git a/src/frontend/src/binder/expr/subquery.rs b/src/frontend/src/binder/expr/subquery.rs
index 51819116771f..c31a5d653aeb 100644
--- a/src/frontend/src/binder/expr/subquery.rs
+++ b/src/frontend/src/binder/expr/subquery.rs
@@ -15,20 +15,16 @@
 use risingwave_sqlparser::ast::Query;
 
 use crate::binder::Binder;
-use crate::error::{ErrorCode, Result};
+use crate::error::{bail_bind_error, Result};
 use crate::expr::{ExprImpl, Subquery, SubqueryKind};
 
 impl Binder {
-    pub(super) fn bind_subquery_expr(
-        &mut self,
-        query: Query,
-        kind: SubqueryKind,
-    ) -> Result<ExprImpl> {
+    pub fn bind_subquery_expr(&mut self, query: Query, kind: SubqueryKind) -> Result<ExprImpl> {
         let query = self.bind_query(query)?;
-        if !matches!(kind, SubqueryKind::Existential) && query.data_types().len() != 1 {
-            return Err(
-                ErrorCode::BindError("Subquery must return only one column".to_string()).into(),
-            );
+        if !matches!(kind, SubqueryKind::Existential | SubqueryKind::UpdateSet)
+            && query.data_types().len() != 1
+        {
+            bail_bind_error!("Subquery must return only one column");
         }
         Ok(Subquery::new(query, kind).into())
     }
diff --git a/src/frontend/src/binder/mod.rs b/src/frontend/src/binder/mod.rs
index b346dc45ca2d..4560e51bd656 100644
--- a/src/frontend/src/binder/mod.rs
+++ b/src/frontend/src/binder/mod.rs
@@ -58,7 +58,7 @@ pub use relation::{
 pub use select::{BoundDistinct, BoundSelect};
 pub use set_expr::*;
 pub use statement::BoundStatement;
-pub use update::BoundUpdate;
+pub use update::{BoundUpdate, UpdateProject};
 pub use values::BoundValues;
 
 use crate::catalog::catalog_service::CatalogReadGuard;
diff --git a/src/frontend/src/binder/update.rs b/src/frontend/src/binder/update.rs
index 9cc80dbde447..f57ad1d19798 100644
--- a/src/frontend/src/binder/update.rs
+++ b/src/frontend/src/binder/update.rs
@@ -12,23 +12,42 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::hash_map::Entry;
 use std::collections::{BTreeMap, HashMap};
 
 use fixedbitset::FixedBitSet;
 use itertools::Itertools;
 use risingwave_common::catalog::{Schema, TableVersionId};
+use risingwave_common::types::DataType;
 use risingwave_common::util::iter_util::ZipEqFast;
 use risingwave_sqlparser::ast::{Assignment, AssignmentValue, Expr, ObjectName, SelectItem};
 
 use super::statement::RewriteExprsRecursive;
 use super::{Binder, BoundBaseTable};
 use crate::catalog::TableId;
-use crate::error::{ErrorCode, Result, RwError};
-use crate::expr::{Expr as _, ExprImpl, InputRef};
+use crate::error::{bail_bind_error, bind_error, ErrorCode, Result, RwError};
+use crate::expr::{Expr as _, ExprImpl, SubqueryKind};
 use crate::user::UserId;
 use crate::TableCatalog;
 
+/// Project into `exprs` in `BoundUpdate` to get the new values for updating.
+#[derive(Debug, Clone, Copy)]
+pub enum UpdateProject {
+    /// Use the expression at the given index in `exprs`.
+    Simple(usize),
+    /// Use the `i`-th field of the expression (returning a struct) at the given index in `exprs`.
+    Composite(usize, usize),
+}
+
+impl UpdateProject {
+    /// Offset the index by `i`.
+    pub fn offset(self, i: usize) -> Self {
+        match self {
+            UpdateProject::Simple(index) => UpdateProject::Simple(index + i),
+            UpdateProject::Composite(index, j) => UpdateProject::Composite(index + i, j),
+        }
+    }
+}
+
 #[derive(Debug, Clone)]
 pub struct BoundUpdate {
     /// Id of the table to perform updating.
@@ -48,10 +67,14 @@ pub struct BoundUpdate {
 
     pub selection: Option<ExprImpl>,
 
-    /// Expression used to project to the updated row. The assigned columns will use the new
-    /// expression, and the other columns will be simply `InputRef`.
+    /// Expression used to evaluate the new values for the columns.
     pub exprs: Vec<ExprImpl>,
 
+    /// Mapping from the index of the column to be updated, to the index of the expression in `exprs`.
+    ///
+    /// By constructing two `Project` nodes with `exprs` and `projects`, we can get the new values.
+    pub projects: HashMap<usize, UpdateProject>,
+
     // used for the 'RETURNING" keyword to indicate the returning items and schema
     // if the list is empty and the schema is None, the output schema will be a INT64 as the
     // affected row cnt
@@ -124,107 +147,112 @@ impl Binder {
 
         let selection = selection.map(|expr| self.bind_expr(expr)).transpose()?;
 
-        let mut assignment_exprs = HashMap::new();
-        for Assignment { id, value } in assignments {
-            // FIXME: Parsing of `id` is not strict. It will even treat `a.b` as `(a, b)`.
-            let assignments = match (id.as_slice(), value) {
-                // _ = (subquery)
-                (_ids, AssignmentValue::Expr(Expr::Subquery(_))) => {
-                    return Err(ErrorCode::BindError(
-                        "subquery on the right side of assignment is unsupported".to_owned(),
-                    )
-                    .into())
-                }
-                // col = expr
-                ([id], value) => {
-                    vec![(id.clone(), value)]
-                }
-                // (col1, col2) = (expr1, expr2)
-                // TODO: support `DEFAULT` in multiple assignments
-                (ids, AssignmentValue::Expr(Expr::Row(values))) if ids.len() == values.len() => id
-                    .into_iter()
-                    .zip_eq_fast(values.into_iter().map(AssignmentValue::Expr))
-                    .collect(),
-                // (col1, col2) = <other expr>
-                _ => {
-                    return Err(ErrorCode::BindError(
-                        "number of columns does not match number of values".to_owned(),
-                    )
-                    .into())
-                }
+        let mut exprs = Vec::new();
+        let mut projects = HashMap::new();
+
+        macro_rules! record {
+            ($id:expr, $project:expr) => {
+                let id_index = $id.as_input_ref().unwrap().index;
+                projects
+                    .try_insert(id_index, $project)
+                    .map_err(|_e| bind_error!("multiple assignments to the same column"))?;
             };
+        }
 
-            for (id, value) in assignments {
-                let id_expr = self.bind_expr(Expr::Identifier(id.clone()))?;
-                let id_index = if let Some(id_input_ref) = id_expr.clone().as_input_ref() {
-                    let id_index = id_input_ref.index;
-                    if table
-                        .table_catalog
-                        .pk()
-                        .iter()
-                        .any(|k| k.column_index == id_index)
-                    {
-                        return Err(ErrorCode::BindError(
-                            "update modifying the PK column is unsupported".to_owned(),
-                        )
-                        .into());
-                    }
-                    if table
-                        .table_catalog
-                        .generated_col_idxes()
-                        .contains(&id_index)
-                    {
-                        return Err(ErrorCode::BindError(
-                            "update modifying the generated column is unsupported".to_owned(),
-                        )
-                        .into());
+        for Assignment { id, value } in assignments {
+            let ids: Vec<_> = id
+                .into_iter()
+                .map(|id| self.bind_expr(Expr::Identifier(id)))
+                .try_collect()?;
+
+            match (ids.as_slice(), value) {
+                // `SET col1 = DEFAULT`, `SET (col1, col2, ...) = DEFAULT`
+                (ids, AssignmentValue::Default) => {
+                    for id in ids {
+                        let id_index = id.as_input_ref().unwrap().index;
+                        let expr = default_columns_from_catalog
+                            .get(&id_index)
+                            .cloned()
+                            .unwrap_or_else(|| ExprImpl::literal_null(id.return_type()));
+
+                        exprs.push(expr);
+                        record!(id, UpdateProject::Simple(exprs.len() - 1));
                     }
-                    if cols_refed_by_generated_pk.contains(id_index) {
-                        return Err(ErrorCode::BindError(
-                            "update modifying the column referenced by generated columns that are part of the primary key is not allowed".to_owned(),
-                        )
-                        .into());
+                }
+
+                // `SET col1 = expr`
+                ([id], AssignmentValue::Expr(expr)) => {
+                    let expr = self.bind_expr(expr)?.cast_assign(id.return_type())?;
+                    exprs.push(expr);
+                    record!(id, UpdateProject::Simple(exprs.len() - 1));
+                }
+                // `SET (col1, col2, ...) = (val1, val2, ...)`
+                (ids, AssignmentValue::Expr(Expr::Row(values))) => {
+                    if ids.len() != values.len() {
+                        bail_bind_error!("number of columns does not match number of values");
                     }
-                    id_index
-                } else {
-                    unreachable!()
-                };
-
-                let value_expr = match value {
-                    AssignmentValue::Expr(expr) => {
-                        self.bind_expr(expr)?.cast_assign(id_expr.return_type())?
+
+                    for (id, value) in ids.iter().zip_eq_fast(values) {
+                        let expr = self.bind_expr(value)?.cast_assign(id.return_type())?;
+                        exprs.push(expr);
+                        record!(id, UpdateProject::Simple(exprs.len() - 1));
                     }
-                    AssignmentValue::Default => default_columns_from_catalog
-                        .get(&id_index)
-                        .cloned()
-                        .unwrap_or_else(|| ExprImpl::literal_null(id_expr.return_type())),
-                };
-
-                match assignment_exprs.entry(id_expr) {
-                    Entry::Occupied(_) => {
-                        return Err(ErrorCode::BindError(
-                            "multiple assignments to same column".to_owned(),
-                        )
-                        .into())
+                }
+                // `SET (col1, col2, ...) = (SELECT ...)`
+                (ids, AssignmentValue::Expr(Expr::Subquery(subquery))) => {
+                    let expr = self.bind_subquery_expr(*subquery, SubqueryKind::UpdateSet)?;
+
+                    if expr.return_type().as_struct().len() != ids.len() {
+                        bail_bind_error!("number of columns does not match number of values");
                     }
-                    Entry::Vacant(v) => {
-                        v.insert(value_expr);
+
+                    let target_type = DataType::new_unnamed_struct(
+                        ids.iter().map(|id| id.return_type()).collect(),
+                    );
+                    let expr = expr.cast_assign(target_type)?;
+
+                    exprs.push(expr);
+
+                    for (i, id) in ids.iter().enumerate() {
+                        record!(id, UpdateProject::Composite(exprs.len() - 1, i));
                     }
                 }
+
+                (_ids, AssignmentValue::Expr(_expr)) => {
+                    bail_bind_error!("source for a multiple-column UPDATE item must be a sub-SELECT or ROW() expression");
+                }
             }
         }
 
-        let exprs = table
-            .table_catalog
-            .columns()
-            .iter()
-            .enumerate()
-            .filter_map(|(i, c)| {
-                c.can_dml()
-                    .then_some(InputRef::new(i, c.data_type().clone()).into())
-            })
-            .map(|c| assignment_exprs.remove(&c).unwrap_or(c))
-            .collect_vec();
+        // Check whether updating these columns is allowed.
+        for &id_index in projects.keys() {
+            if (table.table_catalog.pk())
+                .iter()
+                .any(|k| k.column_index == id_index)
+            {
+                return Err(ErrorCode::BindError(
+                    "update modifying the PK column is unsupported".to_owned(),
+                )
+                .into());
+            }
+            if (table.table_catalog.generated_col_idxes()).contains(&id_index) {
+                return Err(ErrorCode::BindError(
+                    "update modifying the generated column is unsupported".to_owned(),
+                )
+                .into());
+            }
+            if cols_refed_by_generated_pk.contains(id_index) {
+                return Err(ErrorCode::BindError(
+                    "update modifying the column referenced by generated columns that are part of the primary key is not allowed".to_owned(),
+                )
+                .into());
+            }
+
+            let col = &table.table_catalog.columns()[id_index];
+            if !col.can_dml() {
+                bail_bind_error!("update modifying column `{}` is unsupported", col.name());
+            }
+        }
 
         let (returning_list, fields) = self.bind_returning_list(returning_items)?;
         let returning = !returning_list.is_empty();
@@ -236,6 +264,7 @@ impl Binder {
             owner,
             table,
             selection,
+            projects,
             exprs,
             returning_list,
             returning_schema: if returning {
diff --git a/src/frontend/src/error.rs b/src/frontend/src/error.rs
index 3092c9bee91a..f0cf35e85966 100644
--- a/src/frontend/src/error.rs
+++ b/src/frontend/src/error.rs
@@ -33,8 +33,8 @@ use tokio::task::JoinError;
 // - Some variants are never constructed.
 // - Some variants store a type-erased `BoxedError` to resolve the reverse dependency.
 //   It's not necessary anymore as the error type is now defined at the top-level.
-#[derive(Error, thiserror_ext::ReportDebug, thiserror_ext::Box)]
-#[thiserror_ext(newtype(name = RwError, backtrace))]
+#[derive(Error, thiserror_ext::ReportDebug, thiserror_ext::Box, thiserror_ext::Macro)]
+#[thiserror_ext(newtype(name = RwError, backtrace), macro(path = "crate::error"))]
 pub enum ErrorCode {
     #[error("internal error: {0}")]
     InternalError(String),
@@ -105,7 +105,7 @@ pub enum ErrorCode {
     // TODO: use a new type for bind error
     // TODO(error-handling): should prefer use error types than strings.
     #[error("Bind error: {0}")]
-    BindError(String),
+    BindError(#[message] String),
     // TODO: only keep this one
     #[error("Failed to bind expression: {expr}: {error}")]
     BindErrorRoot {
diff --git a/src/frontend/src/expr/subquery.rs b/src/frontend/src/expr/subquery.rs
index 62f59c934dd6..8460f73d5fbb 100644
--- a/src/frontend/src/expr/subquery.rs
+++ b/src/frontend/src/expr/subquery.rs
@@ -24,6 +24,9 @@ use crate::expr::{CorrelatedId, Depth};
 pub enum SubqueryKind {
     /// Returns a scalar value (single column single row).
     Scalar,
+    /// Returns a scalar struct value composed of multiple columns.
+    /// Used in `UPDATE SET (col1, col2) = (SELECT ...)`.
+    UpdateSet,
     /// `EXISTS` | `NOT EXISTS` subquery (semi/anti-semi join). Returns a boolean.
     Existential,
     /// `IN` subquery.
@@ -88,6 +91,7 @@ impl Expr for Subquery {
                 assert_eq!(types.len(), 1, "Subquery with more than one column");
                 types[0].clone()
             }
+            SubqueryKind::UpdateSet => DataType::new_unnamed_struct(self.query.data_types()),
             SubqueryKind::Array => {
                 let types = self.query.data_types();
                 assert_eq!(types.len(), 1, "Subquery with more than one column");
diff --git a/src/frontend/src/optimizer/plan_node/batch_update.rs b/src/frontend/src/optimizer/plan_node/batch_update.rs
index d0351e6fdec2..28dfa79916cc 100644
--- a/src/frontend/src/optimizer/plan_node/batch_update.rs
+++ b/src/frontend/src/optimizer/plan_node/batch_update.rs
@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use itertools::Itertools;
 use risingwave_common::catalog::Schema;
 use risingwave_pb::batch_plan::plan_node::NodeBody;
 use risingwave_pb::batch_plan::UpdateNode;
@@ -84,20 +83,21 @@ impl ToDistributedBatch for BatchUpdate {
 
 impl ToBatchPb for BatchUpdate {
     fn to_batch_prost_body(&self) -> NodeBody {
-        let exprs = self.core.exprs.iter().map(|x| x.to_expr_proto()).collect();
-
-        let update_column_indices = self
-            .core
-            .update_column_indices
+        let old_exprs = (self.core.old_exprs)
+            .iter()
+            .map(|x| x.to_expr_proto())
+            .collect();
+        let new_exprs = (self.core.new_exprs)
             .iter()
-            .map(|i| *i as _)
-            .collect_vec();
+            .map(|x| x.to_expr_proto())
+            .collect();
+
         NodeBody::Update(UpdateNode {
-            exprs,
             table_id: self.core.table_id.table_id(),
             table_version_id: self.core.table_version_id,
             returning: self.core.returning,
-            update_column_indices,
+            old_exprs,
+            new_exprs,
             session_id: self.base.ctx().session_ctx().session_id().0 as u32,
         })
     }
@@ -125,6 +125,6 @@ impl ExprRewritable for BatchUpdate {
 
 impl ExprVisitable for BatchUpdate {
     fn visit_exprs(&self, v: &mut dyn ExprVisitor) {
-        self.core.exprs.iter().for_each(|e| v.visit_expr(e));
+        self.core.visit_exprs(v);
     }
 }
diff --git a/src/frontend/src/optimizer/plan_node/generic/update.rs b/src/frontend/src/optimizer/plan_node/generic/update.rs
index 61d044f53c99..d68af1a01ae3 100644
--- a/src/frontend/src/optimizer/plan_node/generic/update.rs
+++ b/src/frontend/src/optimizer/plan_node/generic/update.rs
@@ -21,7 +21,7 @@ use risingwave_common::types::DataType;
 
 use super::{DistillUnit, GenericPlanNode, GenericPlanRef};
 use crate::catalog::TableId;
-use crate::expr::{ExprImpl, ExprRewriter};
+use crate::expr::{Expr, ExprImpl, ExprRewriter, ExprVisitor};
 use crate::optimizer::plan_node::utils::childless_record;
 use crate::optimizer::property::FunctionalDependencySet;
 use crate::OptimizerContextRef;
@@ -35,15 +35,15 @@ pub struct Update<PlanRef: Eq + Hash> {
     pub table_id: TableId,
     pub table_version_id: TableVersionId,
     pub input: PlanRef,
-    pub exprs: Vec<ExprImpl>,
+    pub old_exprs: Vec<ExprImpl>,
+    pub new_exprs: Vec<ExprImpl>,
     pub returning: bool,
-    pub update_column_indices: Vec<usize>,
 }
 
 impl<PlanRef: GenericPlanRef> Update<PlanRef> {
     pub fn output_len(&self) -> usize {
         if self.returning {
-            self.input.schema().len()
+            self.new_exprs.len()
         } else {
             1
         }
@@ -56,18 +56,19 @@ impl<PlanRef: GenericPlanRef> GenericPlanNode for Update<PlanRef> {
 
     fn schema(&self) -> Schema {
         if self.returning {
-            self.input.schema().clone()
+            Schema::new(
+                self.new_exprs
+                    .iter()
+                    .map(|e| Field::unnamed(e.return_type()))
+                    .collect(),
+            )
         } else {
             Schema::new(vec![Field::unnamed(DataType::Int64)])
         }
     }
 
     fn stream_key(&self) -> Option<Vec<usize>> {
-        if self.returning {
-            Some(self.input.stream_key()?.to_vec())
-        } else {
-            Some(vec![])
-        }
+        None
     }
 
     fn ctx(&self) -> OptimizerContextRef {
@@ -81,27 +82,31 @@ impl<PlanRef: Eq + Hash> Update<PlanRef> {
         table_name: String,
         table_id: TableId,
         table_version_id: TableVersionId,
-        exprs: Vec<ExprImpl>,
+        old_exprs: Vec<ExprImpl>,
+        new_exprs: Vec<ExprImpl>,
         returning: bool,
-        update_column_indices: Vec<usize>,
     ) -> Self {
         Self {
             table_name,
             table_id,
             table_version_id,
             input,
-            exprs,
+            old_exprs,
+            new_exprs,
             returning,
-            update_column_indices,
         }
     }
 
     pub(crate) fn rewrite_exprs(&mut self, r: &mut dyn ExprRewriter) {
-        self.exprs = self
-            .exprs
-            .iter()
-            .map(|e| r.rewrite_expr(e.clone()))
-            .collect();
+        for exprs in [&mut self.old_exprs, &mut self.new_exprs] {
+            *exprs = exprs.iter().map(|e| r.rewrite_expr(e.clone())).collect();
+        }
+    }
+
+    pub(crate) fn visit_exprs(&self, v: &mut dyn ExprVisitor) {
+        for exprs in [&self.old_exprs, &self.new_exprs] {
+            exprs.iter().for_each(|e| v.visit_expr(e));
+        }
     }
 }
 
@@ -109,7 +114,7 @@ impl<PlanRef: Eq + Hash> DistillUnit for Update<PlanRef> {
     fn distill_with_name<'a>(&self, name: impl Into<Str<'a>>) -> XmlNode<'a> {
         let mut vec = Vec::with_capacity(if self.returning { 3 } else { 2 });
         vec.push(("table", Pretty::from(self.table_name.clone())));
-        vec.push(("exprs", Pretty::debug(&self.exprs)));
+        vec.push(("exprs", Pretty::debug(&self.new_exprs)));
         if self.returning {
             vec.push(("returning", Pretty::display(&true)));
         }
diff --git a/src/frontend/src/optimizer/plan_node/logical_update.rs b/src/frontend/src/optimizer/plan_node/logical_update.rs
index 127b6ed8b317..a5590501715b 100644
--- a/src/frontend/src/optimizer/plan_node/logical_update.rs
+++ b/src/frontend/src/optimizer/plan_node/logical_update.rs
@@ -12,17 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use risingwave_common::catalog::TableVersionId;
-
 use super::generic::GenericPlanRef;
 use super::utils::impl_distill_by_unit;
 use super::{
     gen_filter_and_pushdown, generic, BatchUpdate, ColPrunable, ExprRewritable, Logical,
     LogicalProject, PlanBase, PlanRef, PlanTreeNodeUnary, PredicatePushdown, ToBatch, ToStream,
 };
-use crate::catalog::TableId;
 use crate::error::Result;
-use crate::expr::{ExprImpl, ExprRewriter, ExprVisitor};
+use crate::expr::{ExprRewriter, ExprVisitor};
 use crate::optimizer::plan_node::expr_visitable::ExprVisitable;
 use crate::optimizer::plan_node::{
     ColumnPruningContext, PredicatePushdownContext, RewriteStreamContext, ToStreamContext,
@@ -46,25 +43,6 @@ impl From<generic::Update<PlanRef>> for LogicalUpdate {
     }
 }
 
-impl LogicalUpdate {
-    #[must_use]
-    pub fn table_id(&self) -> TableId {
-        self.core.table_id
-    }
-
-    pub fn exprs(&self) -> &[ExprImpl] {
-        self.core.exprs.as_ref()
-    }
-
-    pub fn has_returning(&self) -> bool {
-        self.core.returning
-    }
-
-    pub fn table_version_id(&self) -> TableVersionId {
-        self.core.table_version_id
-    }
-}
-
 impl PlanTreeNodeUnary for LogicalUpdate {
     fn input(&self) -> PlanRef {
         self.core.input.clone()
@@ -86,15 +64,15 @@ impl ExprRewritable for LogicalUpdate {
     }
 
     fn rewrite_exprs(&self, r: &mut dyn ExprRewriter) -> PlanRef {
-        let mut new = self.core.clone();
-        new.exprs = new.exprs.into_iter().map(|e| r.rewrite_expr(e)).collect();
-        Self::from(new).into()
+        let mut core = self.core.clone();
+        core.rewrite_exprs(r);
+        Self::from(core).into()
     }
 }
 
 impl ExprVisitable for LogicalUpdate {
     fn visit_exprs(&self, v: &mut dyn ExprVisitor) {
-        self.core.exprs.iter().for_each(|e| v.visit_expr(e));
+        self.core.visit_exprs(v);
     }
 }
 
diff --git a/src/frontend/src/planner/select.rs b/src/frontend/src/planner/select.rs
index a9e7dd3526ed..ebed01351f7d 100644
--- a/src/frontend/src/planner/select.rs
+++ b/src/frontend/src/planner/select.rs
@@ -320,7 +320,7 @@ impl Planner {
     ///
     /// The [`InputRef`]s' indexes start from `root.schema().len()`,
     /// which means they are additional columns beyond the original `root`.
-    fn substitute_subqueries(
+    pub(super) fn substitute_subqueries(
         &mut self,
         mut root: PlanRef,
         mut exprs: Vec<ExprImpl>,
@@ -366,10 +366,27 @@ impl Planner {
             .zip_eq_fast(rewriter.correlated_indices_collection)
             .zip_eq_fast(rewriter.correlated_ids)
         {
+            let return_type = subquery.return_type();
             let subroot = self.plan_query(subquery.query)?;
 
             let right = match subquery.kind {
                 SubqueryKind::Scalar => subroot.into_unordered_subplan(),
+                SubqueryKind::UpdateSet => {
+                    let plan = subroot.into_unordered_subplan();
+
+                    // Compose all input columns into a struct with `ROW` function.
+                    let all_input_refs = plan
+                        .schema()
+                        .data_types()
+                        .into_iter()
+                        .enumerate()
+                        .map(|(i, data_type)| InputRef::new(i, data_type).into())
+                        .collect::<Vec<_>>();
+                    let call =
+                        FunctionCall::new_unchecked(ExprType::Row, all_input_refs, return_type);
+
+                    LogicalProject::create(plan, vec![call.into()])
+                }
                 SubqueryKind::Existential => {
                     self.create_exists(subroot.into_unordered_subplan())?
                 }
diff --git a/src/frontend/src/planner/update.rs b/src/frontend/src/planner/update.rs
index ddf9ab0bdf9a..2db18ac0e292 100644
--- a/src/frontend/src/planner/update.rs
+++ b/src/frontend/src/planner/update.rs
@@ -13,41 +13,92 @@
 // limitations under the License.
 
 use fixedbitset::FixedBitSet;
-use itertools::Itertools;
+use risingwave_common::types::{DataType, Scalar};
+use risingwave_pb::expr::expr_node::Type;
 
 use super::Planner;
-use crate::binder::BoundUpdate;
+use crate::binder::{BoundUpdate, UpdateProject};
 use crate::error::Result;
+use crate::expr::{ExprImpl, FunctionCall, InputRef, Literal};
+use crate::optimizer::plan_node::generic::GenericPlanRef;
 use crate::optimizer::plan_node::{generic, LogicalProject, LogicalUpdate};
 use crate::optimizer::property::{Order, RequiredDist};
 use crate::optimizer::{PlanRef, PlanRoot};
 
 impl Planner {
     pub(super) fn plan_update(&mut self, update: BoundUpdate) -> Result<PlanRoot> {
+        let returning = !update.returning_list.is_empty();
+
         let scan = self.plan_base_table(&update.table)?;
         let input = if let Some(expr) = update.selection {
             self.plan_where(scan, expr)?
         } else {
             scan
         };
-        let returning = !update.returning_list.is_empty();
-        let update_column_indices = update
-            .table
-            .table_catalog
-            .columns()
-            .iter()
-            .enumerate()
-            .filter_map(|(i, c)| c.can_dml().then_some(i))
-            .collect_vec();
+        let old_schema_len = input.schema().len();
+
+        // Extend table scan with updated columns.
+        let with_new: PlanRef = {
+            let mut plan = input;
+
+            let mut exprs: Vec<ExprImpl> = plan
+                .schema()
+                .data_types()
+                .into_iter()
+                .enumerate()
+                .map(|(index, data_type)| InputRef::new(index, data_type).into())
+                .collect();
+
+            exprs.extend(update.exprs);
+
+            // Substitute subqueries into `LogicalApply`s.
+            if exprs.iter().any(|e| e.has_subquery()) {
+                (plan, exprs) = self.substitute_subqueries(plan, exprs)?;
+            }
+
+            LogicalProject::new(plan, exprs).into()
+        };
+
+        let mut olds = Vec::new();
+        let mut news = Vec::new();
+
+        for (i, col) in update.table.table_catalog.columns().iter().enumerate() {
+            // Skip generated columns and system columns.
+            if !col.can_dml() {
+                continue;
+            }
+            let data_type = col.data_type();
+
+            let old: ExprImpl = InputRef::new(i, data_type.clone()).into();
+
+            let new: ExprImpl = match (update.projects.get(&i)).map(|p| p.offset(old_schema_len)) {
+                Some(UpdateProject::Simple(j)) => InputRef::new(j, data_type.clone()).into(),
+                Some(UpdateProject::Composite(j, field)) => FunctionCall::new_unchecked(
+                    Type::Field,
+                    vec![
+                        InputRef::new(j, with_new.schema().data_types()[j].clone()).into(), // struct
+                        Literal::new(Some((field as i32).to_scalar_value()), DataType::Int32)
+                            .into(),
+                    ],
+                    data_type.clone(),
+                )
+                .into(),
+
+                None => old.clone(),
+            };
+
+            olds.push(old);
+            news.push(new);
+        }
 
         let mut plan: PlanRef = LogicalUpdate::from(generic::Update::new(
-            input,
+            with_new,
             update.table_name.clone(),
             update.table_id,
             update.table_version_id,
-            update.exprs,
+            olds,
+            news,
             returning,
-            update_column_indices,
         ))
         .into();