A question came up in Fundraising yesterday that I now have results for -- basically it's how many page requests view the site via IPv6; and how much can we trust ip geolocation on those IPs.
tldr + a little extra: We can trust IPv6 geolocation for rough pinpointing of countries. Based on complaints we get from donors, that for those countries with not great connectivity we should take the GeoIP resolution with a large grain of salt. City resolution for the US seems to work OK with the exception of mobile -- I have no anecdotal evidence for anywhere else.
--
[snip]
Basically it comes down like this: Last month we served 6,166,994,300 banner requests (served to every page in a content namespace -- so not Special pages but pretty much everything else.) - of which 28,739,500 were IPv6 (0.47%) - of which 39,800 failed to resolve to a country, even with IPv4 fallback (0.0006%) -- (in the noise floor we had 300 people resolve to anonymous proxy or 'European Union' IPs; there were no satellite provider or 'Asia Pacific' resolutions recorded)
I had Ottomata do a filter on geoiplookup.wikimedia.org (our IPv4 only fallback host). In the period between 22:58 and 23:17 we served 344 requests. In that time period we had roughly 210,000 IPv6 requests. Meaning that our fallback is important for ~0.16% of IPv6 users.
I still think we should, in fundraising, do an interstitial page because that's a friendly thing to do. But for sitewide operations; I think the success rate is fairly good. Most of the 'crummyness' that Josh routinely reports is fuzzyness around borders -- e.g. people in India resolving to being in Pakistan.
[snip]
~Matt Walker Wikimedia Foundation Fundraising Technology Team