Thanks to everyone for this great teamwork!

The updated geographical performance dashboards which Gilles and Mark just posted paint a more optimistic picture than before, which is encouraging:
http://multimedia-metrics.wmflabs.org/dashboards/mmv#geographical_network_performance-graphs-tab

However, these extremely fast load times do not match what we are hearing from our users — or even our own experience on slower connections. Many images still take a much longer time to load in practice, as reported by beta users around the world, from Brazil to Hungary. 

Can we really assume that the mean image load time in India is 691 milliseconds? Seems way too fast, based on my experience traveling in Asia a few weeks ago — where images could take a very long time to load, if at all. 

As Gergo pointed out, these early results may be because our first beta testers may have some faster connections than average users. But could there also be some bots or other traffic which could be distorting the results?

I know that we are working next on histograms that will give us a better sense of how outliers are performing against average users. Can’t wait for that.

But I am still concerned that this chart may be painting a much rosier picture than what’s actually going on in the real world.

Any practical recommendations for addressing this concern? We want to know what’s really happening for average users, so we can determine whether or not regions with slow connections like India should consider making this feature opt-in, rather than opt-out.

Thanks again to you all for helping us gain more clarity on this critical issue :)


Fabrice 


On Apr 18, 2014, at 11:16 AM, Gilles Dubuc <gilles@wikimedia.org> wrote:

Mark deployed the change, the mean and standard deviation on the "Overall network performance" and "Geographical network performance" tabs are now geometric:

http://multimedia-metrics.wmflabs.org/dashboards/mmv

These charts and maps now make a lot more sense! Next I'll be working on distribution histograms, so that we can see the outlier values that are now excluded from those graphs.

Thanks again Aaron, thanks to you these visualizations have become truly useful and meaningful, in the way they were meant to be.


On Thu, Apr 17, 2014 at 6:13 PM, Aaron Halfaker <ahalfaker@wikimedia.org> wrote:
Yikes!  Good catch.  


On Thu, Apr 17, 2014 at 11:12 AM, Gilles Dubuc <gilles@wikimedia.org> wrote:
A solution to this problem is to generate a geometric mean[2] instead.

Thanks a lot for the help, it literally instantly solved my problem!

There was a small mistake in the order of functions in your example, for the record it should be:

EXP(AVG(LOG(event_total))) AS geometric_mean

And conveniently the geometric standard deviation can be calculated the same way:

EXP(STDDEV(LOG(event_total))) AS geometric_stddev

I put it to the test on a specific set of data where we had a huge outlier, and for that data it seems equivalent to excluding the lower and upper 10 percentiles, which is exactly what I was after.





On Wed, Apr 16, 2014 at 4:24 PM, Aaron Halfaker <ahalfaker@wikimedia.org> wrote:
Hi Gilles,

I think I know just the thing you're looking for.   

It turns out that much of this performance data is log-normally distributed[1].    Log-normal distributions tend to have a hockey stick shape where most of the values are close to zero, but occasionally very large values appear[3].  Taking the mean of a log-normal distributions tend to be sensitive to outliers like the ones you describe.  

A solution to this problem is to generate a geometric mean[2] instead.  One convenient thing about log-normal data is that if you log() it, it becomes normal[4] -- and not sensitive to outliers in the usual way.  Also convenient, geometric means are super easy to generate.  All you need to do is this: (1) pass all of the data through log() (2) pass the same data through mean() (or avg() -- whatever) (3) pass the result through exp().  The best thing about this is that you can do it in MySQL.

For example:

SELECT
  country,
  mean(timings) AS regular_mean,
  exp(log(mean(timings)) AS geomteric_mean
FROM log.WhateverSchemaYouveGot
GROUP BY country

4. See distribution.log_normal.logged.svg (33K)

-Aaron

On Wed, Apr 16, 2014 at 8:42 AM, Dan Andreescu <dandreescu@wikimedia.org> wrote:
So, my latest idea for a solution is to write a python script that will import the section (last X days) of data from the EventLogging tables that we're interested in into a temporary sqlite database, then proceed with removing the upper and lower percentiles of the data, according to any column grouping that might be necessary. And finally, once the data preprocessing is done in sqlite, run similar queries as before to export the mean, standard deviation, etc. for given metrics to tsvs. I think using sqlite is cleaner than doing the preprocessing on db1047 anyway.

It's quite an undertaking, it basically means rewriting all our current SQL => TSV conversion. The ability to use more steps in the conversion means that we'd be able to have simpler, more readable SQL queries. It would also be a good opportunity to clean up the giant performance query with a bazillion JOINS: https://gitorious.org/analytics/multimedia/source/a949b1c8723c4c41700cedf6e9e48c3866e8b2f4:perf/template.sql which can actually be divided into several data sources all used in the same graph.

Does that sound like a good idea, or is there a simpler solution out there that someone can think of?

Well, I think this sounds like we need to seriously evaluate how people are using EventLogging data and provide this sort of analysis as a feature.  We'd have to hear from more people but I bet it's the right thing to do long term.

Meanwhile, "simple" is highly subjective here.  If it was me, I'd clean up the indentation of that giant SQL query you have, then maybe figure out some ways to make it faster, then be happy as a clam.  So if sql-lite is the tool you feel happy as a clam with, then that sounds like a great solution.  Alternatives would be python, php, etc.  I forgot if pandas was allowed where you're working but that's a great python library that would make what you're talking about fairly easy.

Another thing for us to seriously consider is PostgreSQL.  This has proper f-ing temporary tables and supports actual people doing actual work with databases.  We could dump data, especially really simple schemas like EventLogging, into PostgreSQL for analysis.

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics



_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics



_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics



_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics


_______________________________________________
Multimedia mailing list
Multimedia@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/multimedia

_______________________________

Fabrice Florin
Product Manager
Wikimedia Foundation