diff --git a/pkg/expression/builtin_cast.go b/pkg/expression/builtin_cast.go
index b623a525cfc5c..70121f37c0862 100644
--- a/pkg/expression/builtin_cast.go
+++ b/pkg/expression/builtin_cast.go
@@ -23,6 +23,7 @@
 package expression
 
 import (
+	"bytes"
 	"fmt"
 	"math"
 	"strconv"
@@ -1033,6 +1034,39 @@ func (b *builtinCastRealAsStringSig) Clone() builtinFunc {
 	return newSig
 }
 
+func formatFloat(f float64, bitSize int) string {
+	// MySQL makes float to string max 6 significant digits
+	// MySQL makes double to string max 17 significant digits
+
+	// Unified to display behavior of TiDB. TiKV should also follow this rule.
+	const (
+		expFormatBig   = 1e15
+		expFormatSmall = 1e-15
+	)
+
+	absVal := math.Abs(f)
+	isEFormat := false
+
+	if bitSize == 32 {
+		isEFormat = float32(absVal) >= float32(expFormatBig) || (float32(absVal) != 0 && float32(absVal) < float32(expFormatSmall))
+	} else {
+		isEFormat = absVal >= expFormatBig || (absVal != 0 && absVal < expFormatSmall)
+	}
+
+	dst := make([]byte, 0, 24)
+	if isEFormat {
+		dst = strconv.AppendFloat(dst, f, 'e', -1, bitSize)
+		if idx := bytes.IndexByte(dst, '+'); idx != -1 {
+			copy(dst[idx:], dst[idx+1:])
+			dst = dst[:len(dst)-1]
+		}
+	} else {
+		dst = strconv.AppendFloat(dst, f, 'f', -1, bitSize)
+	}
+
+	return string(dst)
+}
+
 func (b *builtinCastRealAsStringSig) evalString(ctx EvalContext, row chunk.Row) (res string, isNull bool, err error) {
 	val, isNull, err := b.args[0].EvalReal(ctx, row)
 	if isNull || err != nil {
@@ -1046,7 +1080,7 @@ func (b *builtinCastRealAsStringSig) evalString(ctx EvalContext, row chunk.Row)
 		// If we strconv.FormatFloat the value with 64bits, the result is incorrect!
 		bits = 32
 	}
-	res, err = types.ProduceStrWithSpecifiedTp(strconv.FormatFloat(val, 'f', -1, bits), b.tp, typeCtx(ctx), false)
+	res, err = types.ProduceStrWithSpecifiedTp(formatFloat(val, bits), b.tp, typeCtx(ctx), false)
 	if err != nil {
 		return res, false, err
 	}
diff --git a/pkg/expression/builtin_cast_test.go b/pkg/expression/builtin_cast_test.go
index 606151c46f737..2bb01914459aa 100644
--- a/pkg/expression/builtin_cast_test.go
+++ b/pkg/expression/builtin_cast_test.go
@@ -1116,6 +1116,61 @@ func TestCastFuncSig(t *testing.T) {
 	require.False(t, iRes.IsNull())
 	require.Equal(t, types.KindInt64, iRes.Kind())
 	require.Equal(t, int64(0), iRes.GetInt64())
+
+	// test cast real to string
+	zero := float32(0.00)
+	negZero := -zero
+	castFloat32ToStringCases := []struct {
+		v      float32
+		expect string
+	}{
+		{negZero, "-0"},
+		{00000.0, "0"},
+		{4474.78125, "4474.7812"},
+		{1e15, "1e15"},
+		{-1e15, "-1e15"},
+		{9.99999e14, "999999000000000"},
+		{-9.99999e14, "-999999000000000"},
+		{1e15 - 1.0, "1e15"},
+		{-3.4028235e38, "-3.4028235e38"},
+		{3.4028235e38, "3.4028235e38"},
+		{1.1754944e-38, "1.1754944e-38"},
+		{1.000, "1"},
+		{-123456789123000.0, "-123456790000000"},
+		{1e-15, "0.000000000000001"},
+		{9.9999e-16, "9.9999e-16"},
+		{1.23456789123000e-9, "0.0000000012345679"},
+	}
+
+	for _, d := range castFloat32ToStringCases {
+		require.Equal(t, formatFloat(float64(d.v), 32), d.expect)
+	}
+
+	castFloat64ToStringCases := []struct {
+		v      float64
+		expect string
+	}{
+		{float64(negZero), "-0"},
+		{00000.0, "0"},
+		{4474.78125, "4474.78125"},
+		{1e15, "1e15"},
+		{-1e15, "-1e15"},
+		{9.99999e14, "999999000000000"},
+		{-9.99999e14, "-999999000000000"},
+		{1e15 - 1.0, "999999999999999"},
+		{-1.7976931348623157e308, "-1.7976931348623157e308"},
+		{1.7976931348623157e308, "1.7976931348623157e308"},
+		{2.2250738585072014e-308, "2.2250738585072014e-308"},
+		{1.000, "1"},
+		{-123456789123000.0, "-123456789123000"},
+		{1e-15, "0.000000000000001"},
+		{9.9999e-16, "9.9999e-16"},
+		{1.23456789123000e-9, "0.00000000123456789123"},
+	}
+
+	for _, d := range castFloat64ToStringCases {
+		require.Equal(t, formatFloat(d.v, 64), d.expect)
+	}
 }
 
 func TestCastJSONAsDecimalSig(t *testing.T) {
diff --git a/pkg/expression/builtin_cast_vec.go b/pkg/expression/builtin_cast_vec.go
index bce2383171338..0ffe672b7f0e1 100644
--- a/pkg/expression/builtin_cast_vec.go
+++ b/pkg/expression/builtin_cast_vec.go
@@ -221,7 +221,7 @@ func (b *builtinCastRealAsStringSig) vecEvalString(ctx EvalContext, input *chunk
 			result.AppendNull()
 			continue
 		}
-		res, err = types.ProduceStrWithSpecifiedTp(strconv.FormatFloat(v, 'f', -1, bits), b.tp, tc, false)
+		res, err = types.ProduceStrWithSpecifiedTp(formatFloat(v, bits), b.tp, tc, false)
 		if err != nil {
 			return err
 		}