For some time we’ve wanted to phase, gradually roll out, updates to expanding subsets of Ubuntu users so that we may monitor for regressions and stop the update process if there are any. The support for phased updates has existed in update-manager for a while, but we did not have the server side part implemented. Thanks to the work of Colin Watson, Evan Dandrea, and myself this is now done.
Who is participating?
Users of Ubuntu 13.04 who install updates with update-manager are automatically participating in this process. For every package, in -updates, update-manager will generate a random number and if that number is less than the Phased-Update-Percentage the package will be installed. One can opt out of the Phased Update process by adding ‘Update-Manager::Never-Include-Phased-Updates “True”;’ to the configuration file “/etc/apt/apt.conf”.
How does the Phased Update process work?
When a Stable Release Update is released to 13.04 it will have its phased update percentage initially set to 10%. A job is run, every 6 hours, in the data center that checks to see if there are any regression about the package, and if there are none then the phased update percentage will be incremented by 10%. The phased update percentage for a binary package is available at the publishing history page for it. Here is an example with apport. If there is no value for “Phased updates” then the update is fully phased at 100%.
What are the regression checks?
The Ubuntu Error Tracker (errors.ubuntu.com) has been modified to help us determine if there are any regressions about the package. We do this by checking to see if there are any crashes reported about the new version of the package that were not reported about the previous version of the package. (You can actually check this yourself using a query like: https://errors.ubuntu.com/api/1.0/package-version-new-buckets/?format=json&package=unattended-upgrades&previous_version=0.76&new_version=0.76ubuntu1) Additionally, we check the error tracker to see if there is an increased rate of crashes about the package. This is done by examining the quantity of errors reported today and comparing it to the average number of crashes per day for the past two weeks multiplied by the portion of the day that has passed.
If either of these types of regressions are detected then the phasing of the update is stopped by setting it to 0. This will prevent other users from receiving the updated version of the package. There is also a report of packages currently undergoing phasing that displays the phased update percentage for the package and any detected regressions. Additionally, an email is sent to the signer of the package (uploader) and its creator (uploader or sponsee). The email notifies them of the problem and that phasing of the update has been stopped.
There is support in the phased-updater for overriding specific problems, for example if we determine that a regression was not introduced in a specific version of a package. It also keeps track of emails sent so that we do not send an email about the same problem more than once.
If you encounter any issues as a user installing updates or as a developer who has uploaded packages please let me know.