diff --git a/app/vmselect/promql/rollup.go b/app/vmselect/promql/rollup.go
index 247743ddc..c711f04ad 100644
--- a/app/vmselect/promql/rollup.go
+++ b/app/vmselect/promql/rollup.go
@@ -338,9 +338,12 @@ type rollupFuncArg struct {
 	// Timestamps for values.
 	timestamps []int64
 
-	// Actual value preceeding values without restrictions on staleness interval.
+	// Real value preceeding values without restrictions on staleness interval.
 	realPrevValue float64
 
+	// Real value which goes after values.
+	realNextValue float64
+
 	// Current timestamp for rollup evaluation.
 	currTimestamp int64
 
@@ -558,6 +561,11 @@ func (rc *rollupConfig) doInternal(dstValues []float64, tsm *timeseriesMap, valu
 		} else {
 			rfa.realPrevValue = nan
 		}
+		if j < len(values) {
+			rfa.realNextValue = values[j]
+		} else {
+			rfa.realNextValue = nan
+		}
 		rfa.currTimestamp = tEnd
 		value := rc.Func(rfa)
 		rfa.idx++
@@ -1282,6 +1290,8 @@ func rollupDelta(rfa *rollupFuncArg) float64 {
 		d := float64(10)
 		if len(values) > 1 {
 			d = values[1] - values[0]
+		} else if !math.IsNaN(rfa.realNextValue) {
+			d = rfa.realNextValue - values[0]
 		}
 		if math.Abs(values[0]) < 10*(math.Abs(d)+1) {
 			prevValue = 0
diff --git a/app/vmselect/promql/rollup_test.go b/app/vmselect/promql/rollup_test.go
index e0eaacb6c..dff89e2c8 100644
--- a/app/vmselect/promql/rollup_test.go
+++ b/app/vmselect/promql/rollup_test.go
@@ -1103,12 +1103,13 @@ func testRowsEqual(t *testing.T, values []float64, timestamps []int64, valuesExp
 }
 
 func TestRollupDelta(t *testing.T) {
-	f := func(prevValue, realPrevValue float64, values []float64, resultExpected float64) {
+	f := func(prevValue, realPrevValue, realNextValue float64, values []float64, resultExpected float64) {
 		t.Helper()
 		rfa := &rollupFuncArg{
 			prevValue:     prevValue,
 			values:        values,
 			realPrevValue: realPrevValue,
+			realNextValue: realNextValue,
 		}
 		result := rollupDelta(rfa)
 		if math.IsNaN(result) {
@@ -1121,26 +1122,36 @@ func TestRollupDelta(t *testing.T) {
 			t.Fatalf("unexpected result; got %v; want %v", result, resultExpected)
 		}
 	}
-	f(nan, nan, nil, nan)
+	f(nan, nan, nan, nil, nan)
 
 	// Small initial value
-	f(nan, nan, []float64{1}, 1)
-	f(nan, nan, []float64{10}, 10)
-	f(nan, nan, []float64{100}, 100)
-	f(nan, nan, []float64{1, 2, 3}, 3)
-	f(1, nan, []float64{1, 2, 3}, 2)
-	f(nan, nan, []float64{5, 6, 8}, 8)
-	f(2, nan, []float64{5, 6, 8}, 6)
+	f(nan, nan, nan, []float64{1}, 1)
+	f(nan, nan, nan, []float64{10}, 10)
+	f(nan, nan, nan, []float64{100}, 100)
+	f(nan, nan, nan, []float64{1, 2, 3}, 3)
+	f(1, nan, nan, []float64{1, 2, 3}, 2)
+	f(nan, nan, nan, []float64{5, 6, 8}, 8)
+	f(2, nan, nan, []float64{5, 6, 8}, 6)
 
 	// Too big initial value must be skipped.
-	f(nan, nan, []float64{1000}, 0)
-	f(nan, nan, []float64{1000, 1001, 1002}, 2)
+	f(nan, nan, nan, []float64{1000}, 0)
+	f(nan, nan, nan, []float64{1000, 1001, 1002}, 2)
 
-	// Delta calculations against non-nan realPrevValue
-	f(nan, 900, []float64{1000}, 100)
-	f(nan, 900, []float64{1000, 1001, 1002}, 102)
+	// Non-nan realPrevValue
+	f(nan, 900, nan, []float64{1000}, 100)
+	f(nan, 1000, nan, []float64{1000}, 0)
+	f(nan, 1100, nan, []float64{1000}, -100)
+	f(nan, 900, nan, []float64{1000, 1001, 1002}, 102)
+
+	// Small delta between realNextValue and values
+	f(nan, nan, 990, []float64{1000}, 0)
+	f(nan, nan, 1005, []float64{1000}, 0)
+
+	// Big delta between relaNextValue and values
+	f(nan, nan, 800, []float64{1000}, 1000)
+	f(nan, nan, 1300, []float64{1000}, 1000)
 
 	// Empty values
-	f(1, nan, nil, 0)
-	f(100, nan, nil, 0)
+	f(1, nan, nan, nil, 0)
+	f(100, nan, nan, nil, 0)
 }