9 Replies Latest reply on Jan 11, 2015 9:06 PM by mithomps

    4.13: Graphs for small intervals look very strange.

    pathduck

      Hi,

      Looking at the line charts for several resources, in this instance requests/min, and data over eth0, the graphs look very strange when looking at time intervals less than 24 hrs.

       

      Examples:

       

       

      We use the default samplerate for this which I believe is 20 minutes.

       

      Should not the graphs be "interpolated" (not sure if correct term) for the times when there is no data, so it forms a curved line. I assume that RHQ actually does not have 'zero' data at every second interval?

       

      I think this started with 4.13, which we recently set into production, and it makes it hard to see detailed statistics.

       

      Seems kind of related to this, but not quite:

      Graphs,  lots of "no data" columns...

       

      Stian

        • 1. Re: 4.13: Graphs for small intervals look very strange.
          mithomps

          Good catch Stian, I will look into taking the undefined data points out of the interpolated points. The graph will still look weird because the line will be disjoint, but better than the current confusing lines.

           

          You also might want to shorten the collection intervals from the current default 20 minutes to 10 minutes and/or zoom out to get rid of the undefined data points in the multi-line graph.

           

          -- Mike

          1 of 1 people found this helpful
          • 2. Re: 4.13: Graphs for small intervals look very strange.
            pathduck

            Thanks Mike - I will try to set the interval to 10 mins and see if the situation improves

             

            To clarify if I understand it; this is caused by the "resolution" of the graphs being 10 minutes for these time ranges, and since there is no data at every second interval it just assumes it as zero instead of interpolating it?

             

            While the resolution of the bigger timespan is smaller, i.e. more time between the points being drawn?

             

            I think it should always interpolate, even if it's just drawing a straight line between two points - it looks better and since we cannot assume anything for the missing points, a line is better than zero anyway.

             

            Edit; setting sample interval to 10 mins made it much better.

             

             

            But, we can't go around tweaking sample intervals just to get pretty graphs, for instance JVM Heapsize is no big point to sample more than 20 mins.

             

            Stian

            • 3. Re: 4.13: Graphs for small intervals look very strange.
              mithomps

              To clarify if I understand it; this is caused by the "resolution" of the graphs being 10 minutes for these time ranges, and since there is no data at every second interval it just assumes it as zero instead of interpolating it?

               

              This is the correct understanding of the issue.

               

              I think it should always interpolate, even if it's just drawing a straight line between two points - it looks better and since we cannot assume anything for the missing points, a line is better than zero anyway.

               

              While I agree a line is better than a zero -- We had many discussions on how missing data should be represented and eventually came to the conclusion that missing data should explicitly be represented by the no-data bars. This because we want to highlight the missing data as many times it is an availability issue. We don't want to imply that there is data in the interval just to make pretty graphs. In this case of the multi-resource graph, the no-data bars might fall over other data points so just leaving the individual resource line empty is the solution I'm recommending.

               

              -- Mike

              • 4. Re: 4.13: Graphs for small intervals look very strange.
                mithomps

                 

                But, we can't go around tweaking sample intervals just to get pretty graphs, for instance JVM Heapsize is no big point to sample more than 20 mins.

                 

                You probably already know this, but for other's benefit -- The collection intervals can be set more or less globally via the Admin Metric Collection templates (coregui/#Administration/Configuration/MetricTemplates). This is usually the first thing I do on a new install is to adjust the collection intervals down to a finer grain sampling.

                • 5. Re: 4.13: Graphs for small intervals look very strange.
                  pilhuhn

                  >> Thanks Mike - I will try to set the interval to 10 mins and see if the situation improves

                   

                  (I may be wrong here)

                  In the past the graphs always had 60 bars which correspond to 60 data points in that line graph.

                  With an interval of 12h that you have chosen, this means that each bar is 12 mins wide. So a collection interval of 12mins

                  (or an even divider of it - 1,2,3,4,6 mins) would make sure that all the bars are "filled". Using an interval of 10mins would

                  still miss one bar (minute 0 and minute 10 would be in the first bar, same repeats every hour).

                   

                  For the default display interval of 8 hours, a bar is 8 minutes wide.

                  • 6. Re: 4.13: Graphs for small intervals look very strange.
                    pathduck

                    Morning Mike (and Heiko) - yes I know about setting it globally for all resources of a type. But there's still a lot of intervals to tune and there are some metrics where setting very low intervals are not really useful (like JVM Heap). And the UI for this is not really user-friendly

                     

                    As for the drawing of missing data values - it absolutely makes sense for bar-graphs, since these can easily be represented as an empty space. But for line-graphs not so much I think, as doing this they are basically (overlapping) bar-graphs instead of line-graphs.

                     

                    I am not sure how you mean a 'disjoint' graphs would look - If every second data point is 'missing' it would just be a series of dots.

                     

                    Anyone used to looking at line-graphs know that the any point in time between two data samples is something we can't know about. Of course they could be lower or even higher at that specific time but we just don't have the data. And that's perfectly acceptable as long as you know where the data points are.

                     

                    Maybe to make things clearer, one could mark each point with a small circle to mark that "here we have sampled data".

                     

                    Anyway I would propose a graph should look like this even if there are points 'missing' (the red line).

                     

                     

                    This is similar to the way it was done in the old graphs system and it's a perfectly valid way of doing things and looks much better. I think the old graphs also did leave out a line if there was a large section with no data.

                     

                    This is from our JON 3.1.2:

                     

                    The new graphs are superior in almost every way, but drawing them like they are now just looks very strange to everyone looking at them. And I am the guy having to explain to others again when I get asked "why do the graphs look so strange?"

                     

                    Stian

                    • 7. Re: 4.13: Graphs for small intervals look very strange.
                      mithomps

                      Hi Stian,

                       

                      Heiko's comment above is correct [I neglected to mention the 60 bucketing intervals].

                       

                      Ok, I'm persuaded to agree with you on these multi-resource charts (with multiple lines instead of bars) interpolating values instead of the no-data sections. I just need to run this past our usability team first.

                       

                      Thanks for your feedback.

                       

                      -- Mike

                      • 8. Re: 4.13: Graphs for small intervals look very strange.
                        pathduck

                        Thanks Mike - looking forward to testing once you have something

                         

                        I guess it won't be in 4.13.1 but maybe in 4.14 dependent on how much there is to do?

                         

                        By the way Mike, have you had a chance to look at the ER for improvements on the charts?

                        https://bugzilla.redhat.com/show_bug.cgi?id=966444

                         

                        Especially the ability to delect resources and hover to find the parentage of a resource would be awesome to have.

                         

                        Stian

                        • 9. Re: 4.13: Graphs for small intervals look very strange.
                          mithomps

                          Stian,

                           

                          Not sure if it will be in 4.13.1 but 4.14 for sure.

                           

                          As for the other improvements, I have been focused on the RHQ.next stuff and have not spent time on the RHQ charting improvements(but instead, on the new more powerful next generation charting system). Although the two enhancements you mention will not take that much time.  I will have to discuss priorities with the higher ups.

                           

                          -- Mike

                          1 of 1 people found this helpful