609985 - apply smoothing and other data points to overall orange factor graph

Status: RESOLVED WONTFIX

(Tree Management Graveyard :: OrangeFactor, defect, P3)

Description

[Joel Maher ( :jmaher ) (UTC -8)]

for the overall orange factor graph we have settled on a per day calculation instead of cumulative.

Does it make sense to have cumulative data?  overall, 7 day?

Also this could be very choppy, it would be nice to either make the default line we display smoother or plot a smooth line in addition to it.

Comment 1

[Jeff Hammel]

I don't know if cumulative information is too useful.  If you fix a bug (or set of bugs) that reduce the instantaneous orange factor, it will only gradually relax to the new average.  For instance, if the orange factor is 4 for 180 days and immediate drops to 2, if it remains that way for another 180 days the orange factor will still read 3.

Shorter term cumulative day (7 day, day, month, etc) could be useful as long as it is highlighted that it is a cumulative average over this period.  

I would prefer more local filtering.  While I would like to see the instantaneous data, which will inherently be noisy, it is more useful to see a smooth line (IMHO).

See https://wiki.mozilla.org/Auto-tools/Projects/WarOnOrange/Statistics#Smoothing for one such simple method.  Fuller Gaussian smoothing could be used instead of just the nearest neighbors with or without weighting.  Low-pass filters could also be used.  There is a full field of study on the subject.  While I'm not advocating we all become experts on statistical smoothing, I do advocate that we give significant flexibility with respect to how we do smoothing so that we can swap out models and apply a wide range of filters depending on our need.

I would guess the simple smoothing algorithm in the above wiki page will suffice with careful choice of parameters (guess+check). 

We should find smoothing appropriate for a simple but accurate implementation and present this to users (preferably, IMHO, with documentation on how we are smoothing) in addition with the very noisy data.  That's what I would want to see.  And again, keep the back-end flexible to allow us to easily play with different filters and smoothing algorithms without having to do extensive code rewrite. (E.g. possibly with query parameters with the defaults chosen to present a good end-user graph.)

Comment 2

[Jeff Hammel]

See also about moving averages for another kind of filter: http://stockcharts.com/school/doku.php?id=chart_school:technical_indicators:moving_averages

Comment 3

[Jeff Hammel]

In general, the filters should appear as a dropdown selection that can be applied to any linegraph on the chart to produce a new line.  It would be nice if (e.g. by using dashes vs lines, or making the original data become points) it was easy to tell what generated line was associated with what raw data.  A dropdown should also be provided that allows selection of what data is displayed.

Comment 4

[Ed Morley [:emorley]]

OrangeFactor is now EOL (bug 1367364), having been replaced by the new Intermittent Failures View (in bug 1367362):
https://treeherder.mozilla.org/intermittent-failures.html

Bugs for the new view can be filed here:
https://bugzilla.mozilla.org/enter_bug.cgi?product=Tree%20Management&component=Intermittent%20Failures%20View