On 16.06.2012 8:20, Andrew Garrett wrote:
On Sat, Jun 16, 2012 at 6:53 AM, Rob Lanphierrobla@wikimedia.org wrote:
On Fri, Jun 15, 2012 at 8:48 AM, Sumana Harihareswara sumanah@wikimedia.org wrote:
If you merge into mediawiki/core.git, your change is considered safe for inclusion in a wmf branch. The wmf branch is just branched out of master and then deployed. We don't review it again. Because we're deploying more frequently to WMF sites, the code review for merging into MediaWiki's core.git needs to be more like deployment/shell-level review, and so we gave merge access to people who already had deployment access. We have since added some more people. The current list: https://gerrit.wikimedia.org/r/#/admin/groups/11,members
Let me elaborate on this. As unclear as our process is for giving access, it's even less clear what our policy is for taking it away. If we can settle on a policy for taking access away/suspending access, it'll make it much easier to loosen up about giving access.
Here's the situation we want to avoid: we give access to someone who probably shouldn't have it. They continually introduce deployment blockers into the code, making us need to slow down our frequent deployment process. Two hour deploy windows become six hour deploy windows as we need time to fix up breakage introduced during the window. Even with the group we have, there are times where things that really shouldn't slip through do. It's manageable now, but adding more people is going to multiply this problem as we get back into a situation where poorly conceived changes become core dependencies.
We haven't had a culture of making a big deal about the case when someone introduces a breaking change or does something that brings the db to its knees or introduces a massive security hole or whatever. That means that if the situation were to arise that we needed to revoke someones access, we have to wait until it gets egregious and awful, and even then the person is likely to be shocked that their rights are being revoked (if we even do it then). To be less conservative about giving access, we also need to figure out how to be less conservative about taking it away. We also want to be as reasonably objective about it. It's always going to be somewhat subjective, and we don't want to completely eliminate the role of common sense.
It would also be nice if we didn't have to resort to the nuclear option to get the point across. One low-stakes way we can use to make sure people are more careful is to have some sort of rotating "oops" award. At one former job I had, we had a Ghostbusters Stay Puft doll named "Buster" that was handed out when someone broke the build that they had to prominently display in their office. At another job, it was a pair of Shrek ears that people had to wear when they messed something up in production. In both cases, it was something you had to wear until someone else came along. Perhaps we should institute something similar (maybe as simple as asking people to append "OOPS" to their IRC nicks when they botch something).
In general I think we'd want to start by making sure that the person who broke something actually found out that they broke it. I don't think we need to get into "punishment" unless we actually start having serious problems. Otherwise this is a solution looking for a problem.
In terms of low stakes "punishment" for breaking the build, I've heard of an organisation where the last person who broke it is responsible for some unpleasant task (resolving merge conflicts?). In more homogenous co-located organisations I can see something like "has to buy a drink for the people who cleaned up after it" working, but that doesn't really work in our organisation.
Shrek ears?? Why people have to be humiliated? Why don't just: 1. temporary reduce the salary 2. permanently reduce the salary 3. fire ? Dmitriy