Support binary mathematical operators work with NULL literals

datafusion/core/tests/sql/expr.rs

@@ -1254,3 +1254,51 @@ async fn comparisons_with_null_lt() {
         assert!(batch.columns()[0].is_null(1));
     }
 }
+
+#[tokio::test]
+async fn binary_mathematical_operator_with_null_lt() {
+    let ctx = SessionContext::new();
+
+    let cases = vec![
+        // 1. Integer and NULL
+        "select column1 + NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column1 - NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column1 * NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column1 / NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column1 % NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        // 2. Float and NULL
+        "select column2 + NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column2 - NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column2 * NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column2 / NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select column2 % NULL from (VALUES (1, 2.3), (2, 5.4)) as t",
+        // ----
+        // ---- same queries, reversed argument order
+        // ----
+        // 3. NULL and Integer
+        "select NULL + column1 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL - column1 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL * column1 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL / column1 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL % column1 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        // 4. NULL and Float
+        "select NULL + column2 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL - column2 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL * column2 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL / column2 from (VALUES (1, 2.3), (2, 5.4)) as t",
+        "select NULL % column2 from (VALUES (1, 2.3), (2, 5.4)) as t",
+    ];
+
+    for sql in cases {
+        println!("Computing: {}", sql);
+
+        let mut actual = execute_to_batches(&ctx, sql).await;
+        assert_eq!(actual.len(), 1);
+
+        let batch = actual.pop().unwrap();
+        assert_eq!(batch.num_rows(), 2);
+        assert_eq!(batch.num_columns(), 1);
+        assert!(batch.columns()[0].is_null(0));
+        assert!(batch.columns()[0].is_null(1));
+    }
+}

datafusion/expr/src/binary_rule.rs

@@ -282,7 +282,7 @@ fn mathematics_numerical_coercion(
     use arrow::datatypes::DataType::*;
 
     // error on any non-numeric type
-    if !is_numeric(lhs_type) || !is_numeric(rhs_type) {
+    if !both_numeric_or_null_and_numeric(lhs_type, rhs_type) {
         return None;
     };
 
@@ -412,6 +412,15 @@ pub fn is_numeric(dt: &DataType) -> bool {
         }
 }
 
+/// Determine if at least of one of lhs and rhs is numeric, and the other must be NULL or numeric
+fn both_numeric_or_null_and_numeric(lhs_type: &DataType, rhs_type: &DataType) -> bool {
+    match (lhs_type, rhs_type) {
+        (_, DataType::Null) => is_numeric(lhs_type),
+        (DataType::Null, _) => is_numeric(rhs_type),
+        _ => is_numeric(lhs_type) && is_numeric(rhs_type),
+    }
+}
+
 /// Coercion rules for dictionary values (aka the type of the  dictionary itself)
 fn dictionary_value_coercion(
     lhs_type: &DataType,

datafusion/physical-expr/src/expressions/binary.rs

@@ -701,16 +701,18 @@ macro_rules! compute_bool_op {
 /// LEFT is array, RIGHT is scalar value
 macro_rules! compute_op_scalar {
     ($LEFT:expr, $RIGHT:expr, $OP:ident, $DT:ident) => {{
-        let ll = $LEFT
-            .as_any()
-            .downcast_ref::<$DT>()
-            .expect("compute_op failed to downcast array");
-        // generate the scalar function name, such as lt_scalar, from the $OP parameter
-        // (which could have a value of lt) and the suffix _scalar
-        Ok(Arc::new(paste::expr! {[<$OP _scalar>]}(
-            &ll,
-            $RIGHT.try_into()?,
-        )?))
+        if $RIGHT.is_null() {
+            Ok(Arc::new(new_null_array($LEFT.data_type(), $LEFT.len())))
+        } else {
+            let ll = $LEFT
+                .as_any()
+                .downcast_ref::<$DT>()
+                .expect("compute_op failed to downcast array");
+            Ok(Arc::new(paste::expr! {[<$OP _scalar>]}(
+                &ll,
+                $RIGHT.try_into()?,
+            )?))
+        }
     }};
 }