fuel-dev team mailing list archive

[Dmitry Borodaenko <dborodaenko@xxxxxxxxxxxx>]

All,

We still have a few pain points left in our development process that I
think are easy to fix with a bunch of simple rules. I think releasing
4.0 will be less painful if we try to address these.

1. Branch management for maintenance releases

We already had this discussion during 3.2.1 release cycle, and agreed
to follow the approach that is in line with what OpenStack and most
other free software projects are following. Still, I think we should
do better at actually following the process we agreed to.

To see how good we were at following it for 3.2.1, open two terminal
windows and run:

git whatchanged 3.2..3.2-fixes
git whatchanged 3.2..master

and for each commit in 3.2-fixes, try to find a matching fix in
master. Last time I checked there were still many cases where bugfixes
were merged to 3.2-fixes before (or even without) merging them to
master. Did anyone actually check that we're not missing any important
fixes from 3.2.1 in 4.0?

We should create a new stable/4.0 branch as soon as 4.0 code freeze is
announced (ideally, the announcement itself should direct committers
to the new branch). Reviewers should REJECT all commits to stable/4.0
that have not been merged into master, unless a justification is
provided in the COMMIT MESSAGE.

2. Management and code review of feature development branches

Yet another thing that everyone seems to agree on is that huge
long-lived feature branches with many commits and thousands of lines
worth of changes are evil and dangerous. Luckily, the move to Gerrit
will make it hard enough to maintain and merge multi-commit branches,
and will push people towards committing and merging changes in smaller
self-sufficient chunks.

A recent negative example is the fuel-library pull request #911 that
has merged 104 duplicate commits from ancient alternative history into
master, instead of simply rebasing a single commit. The only way to
prevent something like this from happening is to summarily reject
changes that are too large and/or contain messy revision history.

The other side of the same problem is holding back small reasonable
changes for too long, placing unnecessary burden on authors to keep
rebasing their change on top of other changes that got merged earlier.

For example, my own fuel-docs pull request #67 sat unreviewed for a
week only to be obsoleted by the move of the repo to StackForge (after
being obsoleted couple more times by changes that were merged ahead of
it). I suspect most other developers had similar experiences. On top
of obvious frustration, holding a change back tempts the author to
keep piling changes onto the same request instead of creating a new
review request on top of updated master for their next set of changes.
To use the same example, most of the third commit on #67 should really
have been a separate pull request.

The fix is once again rather obvious: when going through reviews,
start with fixes for critical bugs, then go through remaining reviews
starting with the least recently updated ones. Don't merge a review
request if there's an older review request that can also be merged.

I'm using this link to see all our outstanding review requests:
https://review.openstack.org/#/q/status:open+project:^stackforge/fuel-.*,n,z

Right now I see that there are review requests that have +1 from CI
and from reviewers (meaning they can be merged) sitting unchanged
since Nov 25, and a few unreviewed requests going as far back as Nov
3. We shouldn't have a request sit untouched by an approver for more
than a week, let alone a month. If there's a any reason you don't want
to merge it, give it -1 and explain. Otherwise, there's no reason not
to give it +2. If you have time to review and merge a newer request,
you have time for that older one, too.

3. Bugs triage

Moving our bug tracking to public launchpad was an important step
towards opening up our development process, now we should improve
visibility of our bugs triage and release management processes. In
addition to announcing target release dates, we should also have well
defined release criteria (for example, no critical bugs affecting the
upcoming release, no more than 5 bugs with high importantce, etc.),
and documented rules on how to set importance of a bug. We don't have
to be rigid and beaurocratic about it, but having documented criteria
will help all participants of the process prioritize their own work
and understand how it fits into the state of the whole project. It
will also help avoid situations like missing an important bugfix in a
release, by forcing us to review priorities of all open bugs before
announcing a release.

-- 
Dmitry Borodaenko