Measure quality of data against firefox location service telemetry

[graydon]

Current location service in firefox is recording a telemetry statistic showing distribution of accuracy in location queries. Histogram or CDF of that is the plot-of-accuracy to compare against.

Accuracy plots to make and compare against base plot:

• Response accuracy of live queries arriving against MLS
• Simulated response accuracy for queries at each "ground truth" measurement site reported from the stumbler (stored in the measurements DB)
  • Bonus round: look at accuracy plots for subsets of data, city-by-city or region-by-region.

[graydon]

Update: the telemetry statistic was added in bug https://bugzilla.mozilla.org/show_bug.cgi?id=989682 and is live and collecting here: http://telemetry.mozilla.org/#nightly/31/GEOLOCATION_ACCURACY and it shows a pretty crude bimodal distribution: 100k responses at "over 18km" accuracy, 22k responses at "less than 100m" accuracy, a few hundred responses in the 100-200m range and a scattering of responses across the remainder. Basically there's a "wifi" clump and a "geoip" clump at opposite ends of the histogram and nothing much between.

Moveover we are currently only returning fixed accuracy values ourselves, from MLS: either 500m (for wifi), 35km (for cell) and 40km (for geoip). At this point there's no real use taking a "histogram" of these since they are just going to take on those three values; so in the meantime (until we have any improvement in accuracy estimation) we'll just collect counters for each type, this is bug #172 .

I may open a fresh bug / reopen the existing telemetry bug to request a second finer-grained metric for the wifi clump, so we can see if there's any meaningful distribution in there.

[hannosch]

I think we are good here. We have telemetry for Firefox desktop and we have since added fine-grained stats for tracking accuracy from our own service.