Fix float64 comparison test failure on archs using FMA

Architectures using FMA optimization yield slightly different results so
we cannot assume floating point values will be precisely the same across
different architectures.

The solution in this change is to check "abs(a-b) < tolerance" instead
of comparing the exact values. This will give us confidence that the
histogram buckets are near identical.

Signed-off-by: Seth Bunce <[email protected]>

Author: sbunce

prometheus/histogram_test.go

@@ -355,13 +355,12 @@ func TestBuckets(t *testing.T) {
 
 	got = ExponentialBucketsRange(1, 100, 10)
 	want = []float64{
-		1.0, 1.6681005372000588, 2.782559402207125,
-		4.641588833612779, 7.742636826811273, 12.915496650148842,
-		21.544346900318846, 35.93813663804629, 59.94842503189414,
-		100.00000000000007,
+		1.0, 1.6681, 2.7825, 4.6415, 7.7426, 12.9154, 21.5443,
+		35.9381, 59.9484, 100.0000,
 	}
-	if !reflect.DeepEqual(got, want) {
-		t.Errorf("exponential buckets range: got %v, want %v", got, want)
+	const tolerance = 0.0001
+	if !equalFloat64s(got, want, 0.0001) {
+		t.Errorf("exponential buckets range: got %v, want %v (tolerance %f)", got, want, tolerance)
 	}
 }
 
@@ -467,3 +466,23 @@ func TestHistogramExemplar(t *testing.T) {
 		}
 	}
 }
+
+// equalFloat64 returns true if a and b are within the tolerance. We have this
+// because float comparison varies on different architectures. For example,
+// architectures which do FMA yield slightly different results.
+// https://github.com/prometheus/client_golang/pull/899#issuecomment-1244506390
+func equalFloat64(a, b, tolerance float64) bool {
+	return math.Abs(a-b) < tolerance
+}
+
+func equalFloat64s(a, b []float64, tolerance float64) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if !equalFloat64(a[i], b[i], tolerance) {
+			return false
+		}
+	}
+	return true
+}