Use correct numeric types for the ArrayHistogram calculations

`float` only has 23-bit long mantissa, so the maximum integer that can
be safely represented is 8 388 608, which is too low for correct
calculation of MFU thresholds (we can easily have more than 8 million
items in vBuckets).

`std::aggregate` uses the type of the last parameter, which was `int` in
this case. Not sufficient to hold the sum of all bucket counts given a
generic `CountType`.

Change-Id: I93f9af90ecc420af5355c6f82eda9a8c837e5d90
Reviewed-on: https://review.couchbase.org/c/kv_engine/+/190045
Tested-by: Vesko Karaganev <[email protected]>
Reviewed-by: Dave Rigby <[email protected]>

Author: veselink1

engines/ep/src/array_histogram.cc

@@ -111,7 +111,7 @@ size_t ArrayHistogram<CountType, ArraySize, StoredType>::getValueAtPercentile(
     // With the bias:
     //                      (50.0       / 100.0) * 3 + 0.5
     //  would give 2.0 samples, and would find the 50th percentile to be 20.
-    auto threshold = CountType(((percentile / 100.0f) * totalCount) + 0.5f);
+    auto threshold = CountType(((percentile / 100.0) * totalCount) + 0.5);
 
     if (!threshold) {
         // the histogram is not empty, but the requested percentile
@@ -139,7 +139,7 @@ size_t ArrayHistogram<CountType, ArraySize, StoredType>::getValueAtPercentile(
 template <class CountType, size_t ArraySize, class StoredType>
 CountType ArrayHistogram<CountType, ArraySize, StoredType>::getNumberOfSamples()
         const {
-    return std::accumulate(valueArray.begin(), valueArray.end(), 0);
+    return std::accumulate(valueArray.begin(), valueArray.end(), CountType(0));
 }
 
 template <class CountType, size_t ArraySize, class StoredType>

engines/ep/tests/module_tests/array_histogram_test.cc

@@ -132,6 +132,18 @@ TEST(ArrayHistogramTest, Percentile) {
         SCOPED_TRACE("50%ile and 100%ile bumped to 120");
         expect(100, 120, 120);
     }
+
+    // Test with a large number of samples. Previously, we used floats for the
+    // threshold calculation, and because of the 24-bit mantissa in floats,
+    // these values would be off. We also used to use 32-bit ints for the total
+    // number of samples. Compare implementations above those limits.
+    hdr.add(110, std::numeric_limits<int32_t>::max());
+    hist.add(110, std::numeric_limits<int32_t>::max());
+
+    {
+        SCOPED_TRACE("50%ile bumped to 110");
+        expect(100, 110, 120);
+    }
 }
 
 TEST(ArrayHistogramTest, ThreadSafety) {
@@ -160,4 +172,4 @@ TEST(ArrayHistogramTest, ThreadSafety) {
 
     thread1.join();
     thread2.join();
-}
+}