fix: exclude precision-losing integer-to-float conversions from CastExpr::check_bigger_cast (#23808)

datafusion/core/tests/data/int_to_float_cast_precision.csv

@@ -0,0 +1,3 @@
+k,v
+1,16777217
+2,16777216

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

@@ -179,11 +179,8 @@ impl CastExpr {
                 | (UInt8, UInt16 | UInt32 | UInt64)
                 | (UInt16, UInt32 | UInt64)
                 | (UInt32, UInt64)
-                | (
-                    Int8 | Int16 | Int32 | UInt8 | UInt16 | UInt32,
-                    Float32 | Float64
-                )
-                | (Int64 | UInt64, Float64)
+                | (Int8 | Int16 | UInt8 | UInt16, Float32)
+                | (Int8 | Int16 | Int32 | UInt8 | UInt16 | UInt32, Float64)
                 | (Utf8, LargeUtf8)
         )
     }
@@ -1208,6 +1205,31 @@ mod tests {
 
         Ok(())
     }
+
+    #[test]
+    fn test_check_bigger_cast_precision_loss() {
+        use DataType::*;
+
+        // Exact conversions without precision loss
+        assert!(CastExpr::check_bigger_cast(&Int16, &Int8));
+        assert!(CastExpr::check_bigger_cast(&Int64, &Int32));
+        assert!(CastExpr::check_bigger_cast(&Float32, &Int16));
+        assert!(CastExpr::check_bigger_cast(&Float32, &UInt16));
+        assert!(CastExpr::check_bigger_cast(&Float64, &Int32));
+        assert!(CastExpr::check_bigger_cast(&Float64, &UInt32));
+        assert!(CastExpr::check_bigger_cast(&LargeUtf8, &Utf8));
+
+        // Precision-losing int-to-float conversions should return false
+        assert!(!CastExpr::check_bigger_cast(&Float32, &Int32));
+        assert!(!CastExpr::check_bigger_cast(&Float32, &UInt32));
+        assert!(!CastExpr::check_bigger_cast(&Float64, &Int64));
+        assert!(!CastExpr::check_bigger_cast(&Float64, &UInt64));
+
+        // Signed <-> Unsigned conversions should return false (not order-preserving due to negative values)
+        assert!(!CastExpr::check_bigger_cast(&UInt16, &Int8));
+        assert!(!CastExpr::check_bigger_cast(&UInt32, &Int16));
+        assert!(!CastExpr::check_bigger_cast(&Int16, &UInt8));
+    }
 }
 
 /// Tests for the `try_to_proto` / `try_from_proto` hooks.

datafusion/sqllogictest/test_files/monotonic_projection_test.slt

@@ -252,3 +252,66 @@ ORDER BY a, b;
 ----
 a 1
 a0 1
+
+# Test that precision-losing int-to-float casts do not invalidate suffix sort keys.
+#
+# When CAST(Int32 AS Float32) collapses distinct integer values (e.g., 16777216 and
+# 16777217 both become 16777216.0), the suffix sort key (k) must still be sorted
+# correctly. Before the fix, the optimizer incorrectly reused the pre-existing sort
+# order and dropped the SortExec, producing wrong results.
+#
+# t1 is declared with a sort order, t2 is not — their results should be identical
+# since CAST(v AS FLOAT) is not injective for 32-bit integers.
+statement ok
+CREATE EXTERNAL TABLE t1_int_float (k int, v int)
+STORED AS CSV
+WITH ORDER (v DESC, k DESC)
+LOCATION '../core/tests/data/int_to_float_cast_precision.csv'
+OPTIONS ('format.has_header' 'true');
+
+statement ok
+CREATE EXTERNAL TABLE t2_int_float (k int, v int)
+STORED AS CSV
+LOCATION '../core/tests/data/int_to_float_cast_precision.csv'
+OPTIONS ('format.has_header' 'true');
+
+# Both queries must return the same result: k=2 before k=1.
+# (v_=16777216.0 for both rows; when tied on v_, DESC on k means k=2 comes first)
+query IR
+SELECT k, cast(v as float) v_ FROM t1_int_float ORDER BY v_ DESC, k DESC;
+----
+2 16777216
+1 16777216
+
+query IR
+SELECT k, cast(v as float) v_ FROM t2_int_float ORDER BY v_ DESC, k DESC;
+----
+2 16777216
+1 16777216
+
+# Widening cast (Int32 -> Int64) is strictly 1-to-1, so the optimizer CAN
+# legally reuse the pre-existing sort order and omit a SortExec.
+statement ok
+CREATE EXTERNAL TABLE t3_int_bigint (k int, v int)
+STORED AS CSV
+WITH ORDER (v DESC, k DESC)
+LOCATION '../core/tests/data/int_to_float_cast_precision.csv'
+OPTIONS ('format.has_header' 'true');
+
+# CAST(Int32 AS BIGINT) is injective, so suffix key ordering is preserved.
+query II
+SELECT k, cast(v as bigint) v_ FROM t3_int_bigint ORDER BY v_ DESC, k DESC;
+----
+1 16777217
+2 16777216
+
+# Cleanup
+statement ok
+DROP TABLE t1_int_float;
+
+statement ok
+DROP TABLE t2_int_float;
+
+statement ok
+DROP TABLE t3_int_bigint;
+