Log suggestion scores in 10 discrete buckets.

components/ntp_snippets/content_suggestions_metrics.cc

 // Copyright 2016 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "components/ntp_snippets/content_suggestions_metrics.h"
 
+#include <cmath>
 #include <string>
 #include <type_traits>
 
 #include "base/metrics/histogram.h"
 #include "base/metrics/histogram_macros.h"
 #include "base/metrics/user_metrics.h"
 #include "base/strings/stringprintf.h"
 #include "base/template_util.h"
 
 namespace ntp_snippets {
@@ skipping 144 matching lines @@
 void UmaHistogramAge(const std::string& name, const base::TimeDelta& value) {
   base::Histogram::FactoryTimeGet(
       name, base::TimeDelta::FromSeconds(1), base::TimeDelta::FromDays(7), 100,
       base::HistogramBase::kUmaTargetedHistogramFlag)
       ->AddTime(value);
 }
 
 // This corresponds to UMA_HISTOGRAM_CUSTOM_COUNTS (with min/max appropriate
 // for the score of suggestions) for use with dynamic histogram names.
 void UmaHistogramScore(const std::string& name, float value) {
-  base::Histogram::FactoryGet(name, 1, 100000, 50,
+  // Scores are typically reported in a range of [0,1]. As UMA does not support
+  // floats, we put them on a discrete scale of [1,10]. We keep the extra bucket

[jkrcal, 2017/01/10 19:31:47]

nit: Could you expand the comment, e.g. "For instance, discrete value 1
represents values (0.0, 0.1] and value 10 represents values (0.9, 1.0]."

Could you also put the same explanation in the definition in histograms.xml?

[tschumann, 2017/01/12 12:25:31]

Done.

+  // for unexpected over-flows.
+  base::Histogram::FactoryGet(name, 1, 11, 12,

[jkrcal, 2017/01/10 19:31:47]

I think you should use LinearHistogram, instead. Histogram uses exponential
scale for buckets. This may be the same as the linear one when 
  number of buckets = cardinality of the range
but I think using Linear directly is clearer.

I got puzzled by having max_value = 11. Is not (name, 1, 10, 11, ..) enough?
Does the unit-test fail?

Additionally, I do not think you can change type of a histogram without renaming
it. Did you try to run Chrome with your patch?

[tschumann, 2017/01/12 12:25:31]

Apparently, we need to create a new histogram for this :-/
(https://groups.google.com/a/google.com/forum/#!msg/uma-users/L-aaSsAMMaM/CNC1...)

Done and also changed to LinearHistogram.

I added an explicit overflow bucket so that we can distinguish a score of 0.91
from something weird like a score of 42.

[jkrcal, 2017/01/12 15:13:10]

Acknowledged.

                               base::HistogramBase::kUmaTargetedHistogramFlag)
-      ->Add(value);
+      ->Add(ceil(value*10));
 }
 
 void LogCategoryHistogramEnumeration(const char* base_name,
                                      Category category,
                                      int value,
                                      int boundary_value) {
   std::string name = GetCategoryHistogramName(base_name, category);
   // Since the histogram name is dynamic, we can't use the regular macro.
   UmaHistogramEnumeration(name, value, boundary_value);
 }
@@ skipping 179 matching lines @@
 }
 
 void OnCategoryDismissed(Category category) {
   UMA_HISTOGRAM_ENUMERATION(kHistogramCategoryDismissed,
                             static_cast<int>(GetHistogramCategory(category)),
                             static_cast<int>(HistogramCategories::COUNT));
 }
 
 }  // namespace metrics
 }  // namespace ntp_snippets