Adam, are you basing these numbers on the WAP detection by the varnish? (Based on the "accepted" http header not containing "HTML" keyword), or is it based on the user agent detection? I suspect (research needed) that many phones may claim they support HTML without actually handling many of our tags, so the question may still be there if we should treat them as WAP and provide simplified HTML.

In other words, I don't feel we should be making this decision simply on the fact that devices no longer request "WAP", but on how the real phones out there behave. Unfortunately, we do NOT have any reasonable number of common phones to test our site on - we mostly test it on android/iphone devices, which is clearly not the whole picture. The online phone testing services don't seem to provide older phone models either.

Proposal: acquire a number of devices with the most common non-smart-phone user agents based on the zero log, and perform adequate testing of all mobile features. Attempt to do the same with WAP mode (URL param?) to see if that makes any significant difference.



On Fri, Feb 28, 2014 at 11:55 PM, Adam Baso <abaso@wikimedia.org> wrote:
(Resending, got stuck in pending queue for size.)

Here's a trendline for about one month of usage.


The figure seems to be holding between 0.2% and 0.3% for that most recent 1-month period of available WML usage data in Wikipedia Zero.

Other content types are not counted in this WML percent; for example, the following hits on <language>.(m|zero).wikipedia.org are excluded: JSON, XML, PNGs, icons, JavaScript and CSS. Furthermore, content served from domains such as upload.wikimedia.org (media, including most article images) andbits.wikimedia.org (frequently accessed core UX elements such as JavaScript, CSS, icons, etc.) is not part of the denominator.

The previous email referring to sub-0.1 percent WML /hit/ ratios on in-scope active carriers - for which I provided some rather basic analysis - holds. However, using the re-worked definition of WML percent described here there's a slightly higher skew.

In addition to the PDF trendline, a quick and dirty look at the 20140214 log for active operators shows the following rough layout.

Less than 0.5% WML
Overall traffic in this zone is 77.03% of the W0 traffic; average of 0.05% WML in-zone

Between 0.5% and 1.35% WML
Overall traffic in this zone is 22.73% of the W0 traffic; average of 1.08% WML in-zone

Between 1.35% and 4.9999999999% WML
Overall traffic in this zone is 0.22% of the W0 traffic, average of 2.07% WML in-zone; 

5% and higher WML
Overall traffic in this bucket is 0.015% of the W0 traffic, average of 34.2% WML in-zone

The percents don't add to 100 perfectly as a consequence of rounding.

Generally, WML seems to be in decline across most Wikipedia Zero operators, but there are some exceptions (e.g., one operator's WML percent as defined here seems to be holding steady around 1% with this definition of WML, although I'm sort of skeptical of whether the user agent is actually WML-only or even a mobile device or is incapable of handling HTML).

We're analyzing whether WML deprecation has a disparate impact measurable enough to defer exclusive concentration on HTML, especially in light of the increasing availability of cheap mobile devices supporting HTML and the tendency for large intermediary search services and appliances to translate HTML to WML on behalf of the actual client device in the rare case of a device only supporting WML.

-Adam


On Fri, Feb 28, 2014 at 12:41 PM, Yuri Astrakhan <yastrakhan@wikimedia.org> wrote:
I just ran some stats on the 1/1000 sample logs for ZERO traffic. Due to nature of Zero - to bring Wikipedia to the developing world - we have to deal with lots of old phones. I would like to reopen WAP discussion before we put the final nail into the coffin or decide if this data warrants further investigation.