diff --git a/deps/oblib/src/common/object/ob_obj_compare.h b/deps/oblib/src/common/object/ob_obj_compare.h
index d93287bf6..f7c080eed 100644
--- a/deps/oblib/src/common/object/ob_obj_compare.h
+++ b/deps/oblib/src/common/object/ob_obj_compare.h
@@ -259,7 +259,15 @@ public:
     }
     const double l = obj1.get_double();
     const double r = obj2.get_double();
-    if (l == r || fabs(l - r) < p) {
+    if (isnan(l) || isnan(r)) {
+      if (isnan(l) && isnan(r)) {
+        ret = 0;
+      } else if (isnan(l)) {
+        ret = 1;
+      } else {
+        ret = -1;
+      }
+    } else if (l == r || fabs(l - r) < p) {
       ret = 0;
     } else {
       ret = (l < r ? -1 : 1);
diff --git a/src/share/datum/ob_datum_cmp_func_def.h b/src/share/datum/ob_datum_cmp_func_def.h
index bee82eedb..2d9d3e870 100644
--- a/src/share/datum/ob_datum_cmp_func_def.h
+++ b/src/share/datum/ob_datum_cmp_func_def.h
@@ -147,7 +147,17 @@ struct ObFixedDoubleCmp: public ObDefined<>
     cmp_ret = 0;
     const double l = l_datum.get_double();
     const double r = r_datum.get_double();
-    if (l == r || fabs(l - r) < P) {
+    if (isnan(l) || isnan(r)) {
+      if (isnan(l) && isnan(r)) {
+        cmp_ret = 0;
+      } else if (isnan(l)) {
+        // l is nan, r is not nan:left always bigger than right
+        cmp_ret = 1;
+      } else {
+        // l is not nan, r is nan, left always less than right
+        cmp_ret = -1;
+      }
+    } else if (l == r || fabs(l - r) < P) {
       cmp_ret = 0;
     } else {
       cmp_ret = (l < r ? -1 : 1);
diff --git a/src/share/vector/vector_basic_op.h b/src/share/vector/vector_basic_op.h
index 283530703..56b753c1f 100644
--- a/src/share/vector/vector_basic_op.h
+++ b/src/share/vector/vector_basic_op.h
@@ -632,7 +632,17 @@ struct VecTCCmpCalc<VEC_TC_FIXED_DOUBLE, VEC_TC_FIXED_DOUBLE>
     const double l = *reinterpret_cast<const double *>(l_v);
     const double r = *reinterpret_cast<const double *>(r_v);
     const double P = 5 / LOG_10[l_meta.get_scale() + 1];
-    if (l == r || fabs(l - r) < P) {
+    if (isnan(l) || isnan(r)) {
+      if (isnan(l) && isnan(r)) {
+        cmp_ret = 0;
+      } else if (isnan(l)) {
+        // l is nan, r is not nan:left always bigger than right
+        cmp_ret = 1;
+      } else {
+        // l is not nan, r is nan, left always less than right
+        cmp_ret = -1;
+      }
+    } else if (l == r || fabs(l - r) < P) {
       cmp_ret = 0;
     } else {
       cmp_ret = (l < r ? -1: 1);
diff --git a/src/sql/engine/expr/ob_expr_func_round.cpp b/src/sql/engine/expr/ob_expr_func_round.cpp
index 501cce07d..1235789c6 100644
--- a/src/sql/engine/expr/ob_expr_func_round.cpp
+++ b/src/sql/engine/expr/ob_expr_func_round.cpp
@@ -305,7 +305,7 @@ static int do_round_by_type(
     }
     case ObDecimalIntType: {
       if (OB_FAIL(ObExprFuncRound::calc_round_decimalint(
-                  in_meta, out_meta, round_scale, x_datum, res_datum))) {
+                  in_meta, out_meta, GET_SCALE_FOR_CALC(round_scale), x_datum, res_datum))) {
         LOG_WARN("calc_round_decimalint failed", K(ret), K(in_meta), K(out_meta), K(round_scale));
       }
       break;
@@ -434,7 +434,7 @@ static int do_round_by_type_batch_with_check(const int64_t scale, const ObExpr &
           if (x_datum.is_null()) {
             results[i].set_null();
           } else if (OB_FAIL(ObExprFuncRound::calc_round_decimalint(
-                             in_meta, out_meta, scale, x_datum, results[i]))) {
+                             in_meta, out_meta, GET_SCALE_FOR_CALC(scale), x_datum, results[i]))) {
             LOG_WARN("calc_round_decimalint failed", K(ret), K(in_meta), K(out_meta), K(scale));
           }
         }
@@ -567,7 +567,7 @@ static int do_round_by_type_vector(const int64_t scale, const ObExpr &expr,
       for (int64_t j = bound.start(); OB_SUCC(ret) && j < bound.end(); ++j) {
         CHECK_ROUND_VECTOR();
         if (OB_FAIL((ObExprFuncRound::calc_round_decimalint<LeftVec, ResVec>)(
-                            in_meta, out_meta, scale, left_vec, res_vec, j))) {
+                            in_meta, out_meta, GET_SCALE_FOR_CALC(scale), left_vec, res_vec, j))) {
           LOG_WARN("calc_round_decimalint failed", K(ret), K(in_meta), K(out_meta), K(scale));
         }
         eval_flags.set(j);
@@ -656,7 +656,7 @@ static int do_round_by_type_batch_without_check(const int64_t scale, const ObExp
       const ObDatumMeta &out_meta = expr.datum_meta_;
       for (int64_t i = 0; OB_SUCC(ret) && i < batch_size; ++i) {
         if (OB_FAIL(ObExprFuncRound::calc_round_decimalint(
-                    in_meta, out_meta, scale, x_datums[i], results[i]))) {
+                    in_meta, out_meta, GET_SCALE_FOR_CALC(scale), x_datums[i], results[i]))) {
           LOG_WARN("calc_round_decimalint failed", K(ret), K(in_meta), K(out_meta), K(scale));
         }
       }
diff --git a/src/sql/engine/expr/ob_expr_util.cpp b/src/sql/engine/expr/ob_expr_util.cpp
index 7b06d87bb..9e4ddeef2 100644
--- a/src/sql/engine/expr/ob_expr_util.cpp
+++ b/src/sql/engine/expr/ob_expr_util.cpp
@@ -459,13 +459,13 @@ int ObExprUtil::get_mb_str_info(const ObString &str,
 
 double ObExprUtil::round_double(double val, int64_t dec)
 {
-  const double pow_val = std::pow(10, static_cast<double>(std::abs(dec)));
-  volatile double val_div_tmp = val / pow_val;
-  volatile double val_mul_tmp = val * pow_val;
-  volatile double res = 0.0;
+  const double pow_val = std::pow(10.0, static_cast<double>(std::abs(dec)));
+  double val_div_tmp = val / pow_val;
+  double val_mul_tmp = val * pow_val;
+  double res = 0.0;
   if (dec < 0 && std::isinf(pow_val)) {
     res = 0.0;
-  } else if (dec >= 0 && std::isinf(val_mul_tmp)) {
+  } else if (dec >= 0 && !std::isfinite(val_mul_tmp)) {
     res = val;
   } else {
     res = dec < 0 ? rint(val_div_tmp) * pow_val : rint(val_mul_tmp) / pow_val;
@@ -474,23 +474,6 @@ double ObExprUtil::round_double(double val, int64_t dec)
   return res;
 }
 
-double ObExprUtil::round_double_nearest(double val, int64_t dec)
-{
-  const double pow_val = std::pow(10, static_cast<double>(std::abs(dec)));
-  volatile double val_div_tmp = val / pow_val;
-  volatile double val_mul_tmp = val * pow_val;
-  volatile double res = 0.0;
-  if (dec < 0 && std::isinf(pow_val)) {
-    res = 0.0;
-  } else if (dec >= 0 && std::isinf(val_mul_tmp)) {
-    res = val;
-  } else {
-    res = dec < 0 ? std::round(val_div_tmp) * pow_val : std::round(val_mul_tmp) / pow_val;
-  }
-  LOG_DEBUG("round double done", K(val), K(dec), K(res));
-  return res;
-}
-
 uint64_t ObExprUtil::round_uint64(uint64_t val, int64_t dec)
 {
   uint64_t res = 0;
@@ -524,7 +507,7 @@ double ObExprUtil::trunc_double(double val, int64_t dec)
   volatile double res = 0.0;
   if (dec < 0 && std::isinf(pow_val)) {
     res = 0.0;
-  } else if (dec >= 0 && std::isinf(val_mul_tmp)) {
+  } else if (dec >= 0 && !std::isfinite(val_mul_tmp)) {
     res = val;
   } else {
     if (val >= 0) {
diff --git a/src/sql/engine/expr/ob_expr_util.h b/src/sql/engine/expr/ob_expr_util.h
index 9e3db848b..e9dfa98e9 100644
--- a/src/sql/engine/expr/ob_expr_util.h
+++ b/src/sql/engine/expr/ob_expr_util.h
@@ -80,8 +80,6 @@ public:
                              common::ObIArray<size_t> &byte_offset);
   // 将double round到小数点后或者小数点前指定位置
   static double round_double(double val, int64_t dec);
-  // 将double round到小数点后指定位置，.5 -> 1 而不是0
-  static double round_double_nearest(double val, int64_t dec);
   static uint64_t round_uint64(uint64_t val, int64_t dec);
   // 将double trunc到小数点后或者小数点前指定位置
   static double trunc_double(double val, int64_t dec);