On 03/06/11 09:46, Ryan Lane wrote:
In case anyone's wondering what I think of this, I was pretty blunt last time around:
http://lists.wikimedia.org/pipermail/wikitech-l/2011-April/052893.html
To be any more blunt than that, I'd have to press the caps lock key ;)
This post seems to be a much more optimistic view of the 1.17 deployment than I remember. Wasn't it rolled back twice? I also remember getting reports of random issues for weeks after the deployment as well. People were asking for us to do a rollback a couple weeks after the deployment even.
The first deployment caused disruption because we tried to deploy it to all wikis at once, hopefully we're not going to try that again.
The second deployment caused disruption because it was done while I was asleep, and the people involved decided to spend a long time trying to debug the problem while the site mostly down, instead of immediately reverting and isolating it offline. Hopefully we won't try that again either.
Despite these "learning experiences", I think it went a lot better than previous deployments. The number of bugs that needed fixing was relatively small.
1.17 still doesn't have a tarball release. It is *very* late. It doesn't seem to me that this process is working.
We have a beta release, do you mean a stable release? That's mostly due to priorities, not process. Also the Berlin meeting stopped pretty much all work on it for a week, and there were a couple of security issues reported that we'd like to get fixed before 1.17.0 is released.
I agree with the testing process, but I disagree with the deployment process. We can apply the testing process to a continuous integration deployment model. From an ops perspective, this is much nicer; if something is going wrong, we know fairly quickly what caused it to go wrong.
I'm not sure what you're proposing here, in terms of how branches would be used and what sort of testing would be done. Can you elaborate?
-- Tim Starling