tools/metrics/histograms/histograms.xml

Index: tools/metrics/histograms/histograms.xml
diff --git a/tools/metrics/histograms/histograms.xml b/tools/metrics/histograms/histograms.xml
index 8ddf95ea8c71cfbf083369e042291f04ff0d513d..eb8b6bdb48c787dbeb5bb76f60ed8dd6b1cb0058 100644
--- a/tools/metrics/histograms/histograms.xml
+++ b/tools/metrics/histograms/histograms.xml
@@ -25325,6 +25325,17 @@ http://cs/file:chrome/histograms.xml - but prefer this file for new entries.
   </summary>
 </histogram>
 
+<histogram name="Media.Video.KeyFrameDistanceAverage" units="ms">
+  <owner>avayvod@chromium.org</owner>
+  <owner>dalecurtis@chromium.org</owner>
+  <summary>
+    Records the moving average of time distance between consequent keyframes in
+    a video based on the last N frames (see N in decoder_stream_traits.cc). The
+    new average is calculated for each video key frame decoded and is recorded
+    every time the average increases from the previous one.

[Ilya Sherman, 2016/12/05 23:48:04]

Recording every time the average increases, but not when it decreases, seems
like an odd choice; I'd expect it to make the data hard to interpret.  If you
just want to record the max seen over the lifetime of the video, could you
record it after the video finishes playing?

[whywhat, 2016/12/06 05:16:28]

I think the main reason to report it continuously is because I suspect the
majority of videos are not watched till the end (renderer gets killed, the user
doesn't watch until the titles finish and so on).

We care about the average frame distance to avoid excluding videos with a few
frames being rare from the optimization (the max could be much bigger than
average) we'd like to enable, but reporting it only when max changes presumably
makes the reporting less frequent - average frame distance may change a bit with
every key frame but max seems to stabilize over time in my local testing.

Dale suggested reporting the actual distance for each keyframe (as an example:
every 5s for YouTube, every 0.5s for Netflix) and have histograms take care of
accumulating. WDYT?

[Ilya Sherman, 2016/12/06 22:03:55]

I think recording once every keyframe makes more sense, in that it'll be easier
to understand the data.  Otherwise, I'd expect that you'd have a very odd bias:
videos where the max distance changes more often get more votes; and videos
where the max is quite different from the typical average will report unhelpful
values.  The downside of recording for every keyframe is that it doesn't give
you a great way to get an average per video, since data recorded to histograms
lacks timestamps.

Here's a question then: When you're deciding whether or not to apply the
optimization, how will you make that decision?  Presumably it'll be based on
some heuristic, since you won't have the full video stream available
immediately.  Would it make sense to record that heuristic prediction (or if
you're trying to calibrate that heuristic, the values that go into making the
prediction)?

+  </summary>
+</histogram>
+
 <histogram name="Media.VideoCapture.AspectRatio" units="%">
   <owner>mcasas@chromium.org</owner>
   <summary>