Merge pull request #1673 from imorph/faster_find_bucket

PERF: faster algorithm to discover bucket of an histogram observation

Author: ArthurSens

prometheus/histogram.go

@@ -858,15 +858,35 @@ func (h *histogram) Write(out *dto.Metric) error {
 // findBucket returns the index of the bucket for the provided value, or
 // len(h.upperBounds) for the +Inf bucket.
 func (h *histogram) findBucket(v float64) int {
-	// TODO(beorn7): For small numbers of buckets (<30), a linear search is
-	// slightly faster than the binary search. If we really care, we could
-	// switch from one search strategy to the other depending on the number
-	// of buckets.
-	//
-	// Microbenchmarks (BenchmarkHistogramNoLabels):
-	// 11 buckets: 38.3 ns/op linear - binary 48.7 ns/op
-	// 100 buckets: 78.1 ns/op linear - binary 54.9 ns/op
-	// 300 buckets: 154 ns/op linear - binary 61.6 ns/op
+	n := len(h.upperBounds)
+	if n == 0 {
+		return 0
+	}
+
+	// Early exit: if v is less than or equal to the first upper bound, return 0
+	if v <= h.upperBounds[0] {
+		return 0
+	}
+
+	// Early exit: if v is greater than the last upper bound, return len(h.upperBounds)
+	if v > h.upperBounds[n-1] {
+		return n
+	}
+
+	// For small arrays, use simple linear search
+	// "magic number" 35 is result of tests on couple different (AWS and baremetal) servers
+	// see more details here: https://github.com/prometheus/client_golang/pull/1662
+	if n < 35 {
+		for i, bound := range h.upperBounds {
+			if v <= bound {
+				return i
+			}
+		}
+		// If v is greater than all upper bounds, return len(h.upperBounds)
+		return n
+	}
+
+	// For larger arrays, use stdlib's binary search
 	return sort.SearchFloat64s(h.upperBounds, v)
 }
 

prometheus/histogram_test.go

@@ -1455,3 +1455,91 @@ func compareNativeExemplarValues(t *testing.T, exps []*dto.Exemplar, values []fl
 		}
 	}
 }
+
+var resultFindBucket int
+
+func benchmarkFindBucket(b *testing.B, l int) {
+	h := &histogram{upperBounds: make([]float64, l)}
+	for i := range h.upperBounds {
+		h.upperBounds[i] = float64(i)
+	}
+	v := float64(l / 2)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		resultFindBucket = h.findBucket(v)
+	}
+}
+
+func BenchmarkFindBucketShort(b *testing.B) {
+	benchmarkFindBucket(b, 20)
+}
+
+func BenchmarkFindBucketMid(b *testing.B) {
+	benchmarkFindBucket(b, 40)
+}
+
+func BenchmarkFindBucketLarge(b *testing.B) {
+	benchmarkFindBucket(b, 100)
+}
+
+func BenchmarkFindBucketHuge(b *testing.B) {
+	benchmarkFindBucket(b, 500)
+}
+
+func BenchmarkFindBucketInf(b *testing.B) {
+	h := &histogram{upperBounds: make([]float64, 500)}
+	for i := range h.upperBounds {
+		h.upperBounds[i] = float64(i)
+	}
+	v := 1000.5
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		resultFindBucket = h.findBucket(v)
+	}
+}
+
+func BenchmarkFindBucketLow(b *testing.B) {
+	h := &histogram{upperBounds: make([]float64, 500)}
+	for i := range h.upperBounds {
+		h.upperBounds[i] = float64(i)
+	}
+	v := -1.1
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		resultFindBucket = h.findBucket(v)
+	}
+}
+
+func TestFindBucket(t *testing.T) {
+	smallHistogram := &histogram{upperBounds: []float64{1, 2, 3, 4, 5}}
+	largeHistogram := &histogram{upperBounds: make([]float64, 50)}
+	for i := range largeHistogram.upperBounds {
+		largeHistogram.upperBounds[i] = float64(i)
+	}
+
+	tests := []struct {
+		h        *histogram
+		v        float64
+		expected int
+	}{
+		{smallHistogram, -1, 0},
+		{smallHistogram, 0.5, 0},
+		{smallHistogram, 2.5, 2},
+		{smallHistogram, 5.5, 5},
+		{largeHistogram, -1, 0},
+		{largeHistogram, 25.5, 26},
+		{largeHistogram, 49.5, 50},
+		{largeHistogram, 50.5, 50},
+		{largeHistogram, 5000.5, 50},
+	}
+
+	for _, tt := range tests {
+		result := tt.h.findBucket(tt.v)
+		if result != tt.expected {
+			t.Errorf("findBucket(%v) = %d; expected %d", tt.v, result, tt.expected)
+		}
+	}
+}