I like the idea of lieutenants, but another option would be a "multi-lieutenant" model.
The model used at google is that each directory has a file called "OWNERS" which lists several usernames, one per line. For any given patch, you are expected to get a review such that, for each modified file, one of the OWNERS listed in that directory (or any parent thereof) has +1ed. So, for example, imagine that hbase/OWNERS has only Stack, and hbase/foo/component1/OWNERS has "jxiang,larsh". If I make a patch which touches something in foo/component1/bar/, I'd need a review from at least one of Jimmy, Lars, or Stack. The assumption is that you try to get review from the most specific owner, but if those people are MIA, you get review from someone higher up the stack. The multi-person-per-dir model also ensures that, if someone's on vacation or otherwise busy, we don't get blocked. And it formalizes in the actual source tree who you should probably email if you have questions about an area. It also means that wide-ranging patches that touch multiple components need a lot of reviewers (or someone higher up the chain of command who has "permission" on the whole tree). So if I had a mondo patch that touched the region server, the master, and the IPC layer, I'd probably need at least three separate people to sign off. Whatever we do, rather than making it a strict policy, let's start out with a soft touch. Perhaps declare the component maintainers and try to pick reviewers based on the criteria. But if people are busy and work needs to get done, we don't need to be anal about it :) -Todd On Fri, Sep 14, 2012 at 12:17 PM, Stack <st...@duboce.net> wrote: > At the contributor's pow wow a few days ago [1], during a discussion > about whether or not commits should have more friction applied -- i.e. > have more review before they go in -- it was thought that we might > benefit if we had "lieutenants" over-seeing individual HBase > components. A lieutenant would be someone who has an interest and an > understanding of how a particular component works (or should work). A > lieutenant does not need to be a committer. Before committing a patch > that touched on a particular component, the patch would have to have > been +1'd by the component lieutenant before it could go in (or if the > lieutenant is MIA, it was suggested by the Mighty Jon Hsieh that two > +1s by other contributors/committers would do instead; this latter > rule would probably also apply when a patch spanned components). > > We already have a few folks signed up, knowingly or otherwise, as > component owners [1]. > > What do folks think? > > Should we go ahead w/ this project? If so, any volunteers (I signed > up a few of the obvious component leads)? I can add you as component > lieutenant into JIRA. We can add more components if you don't see > your interest listed. > > St.Ack > > 1. http://www.meetup.com/hbaseusergroup/events/80621872/ -- Todd Lipcon Software Engineer, Cloudera
