Developing on a native mobile platform such as iOS or Android is very different from developing for the web. In addition to the standardized development environments necessary to develop code, there is also the issue of posting updates to an App Store, waiting for it to be approved (if the app store requires approval) and then waiting for users to apply updates. You can't push early and often like you can with web applications.

The difference in the update methodology possible on the Web vs. native mobile is actually a really important issue. Let's say that you want to roll out a release that uses a new backend infrastructure that you've developed that offers more features and costs less money to operate. How soon after the update can you start scaling back the old infrastructure and turn it off? Will it ever be possible? What about if you want to release an update that fixes bugs and crashes? How many users are going to bother to update instead of writing off your app as buggy? These are questions that we've been asking ourselves as we've improved our iOS application over the past few months.

Apple provides reporting on downloads. The information is admittedly very sparse -- they only have info on the number of new app downloads per day and the number of updates per day. Nevertheless, there is some information to be gleaned from these simple metrics.

Here's a chart showing the update distribution after a recent update we made and a curve fit of the Pareto distribution over it:

The chance of a user updating after N days seems to follow a power law distribution, this specific fit being a Pareto distribution. Power law distributions are good models for many common phenomena and statistics, such as the populations of cities, the frequencies of words in a book, song popularities, earth quake intensities, file sizes on a disk, etc. In our case, it seems like we get 90% of users updating after 10 days, an example of the 80-20 law at work:

So far, we only have a hypothesis that app updates follow a power law curve. The test of our hypothesis will be to see if the update curves look the same over upcoming updates (ie. our hypothesis being useful as a predictive measure). Stay tuned!