> On Mar 23, 2016, at 10:25 AM, Chris Nauroth <cnaur...@hortonworks.com> wrote:
>
> 2. Apache feature branches: Sign-off may come from designated branch
> committers in addition to full committers. It's OK to break the branch
> for work in progress, but it must be fixed before a merge. It's still
> review then commit though.
...
> Allen, are there particular reasons that you favor #2 (but with
> commit-then-review) instead of #3 for your work on HADOOP-12857 and
> HADOOP-12930? Maybe you want the patches on Apache infrastructure so you
> can get the pre-commit sign-offs quickly? Is it something else?
HADOOP-12930 is sort of a weird case. I’ve only written some test code,
but... ignoring unit tests and some variable name cleanup, there isn’t much
code being truly added or modified. However there is significant code
_movement_ due to some refactoring that needs to take place. As a result, this
has some implications on reviewers, the testing infrastructure, and how it is
committed:
* A combined patch file will be huge and overwhelming since a lot of effort
will need to be spent actually finding the changes rather than having them as
logically separated units of work.
* As all of you know, precommit doesn't work properly if any 2 or more of
hadoop-common, hadoop-hdfs, hadoop-mapreduce-project, and a handful of others
are touched due to the time it takes to do the unit test runs. Guess how many
of these are being touched? :)
* Committing piece meal means that feature-set wise, it could/would be
incomplete until the last block is pushed into place. This definitely needs to
get committed as a logical set (either as many smaller commits at once or a
rebased/merged single commit) in trunk.
* Being in a branch means that reviewers can checkout and test rather than
having to apply a series of patches [despite how easy smart-apply-patch makes
it. ;) ]
* One of the big criticisms during the 9902 review was that it would have been
better to do it as a feature branch due to the size. This is me trying to
acquiesce to that request. (It’s worthwhile pointing out that 9902 couldn’t
have been easily broken up, however, since hadoop-config.sh was touched. That
was a major ‘beak the world’ moment.)
A big problem with the PMC’s ruleset around feature branches is that
they are pretty much built around the idea that teams of people are going to be
working on a feature. Realistically, I don’t see that being the case here
(despite some mild interest from other committers in the feature) esp given I
can probably finish this feature off in a few days. That’s why I’m asking if
CTR into the branch followed by RTM into trunk is a ‘legal' strategy here,
since branch merges require reviews anyway.
Yes, I could make a branch in my own repo on gitlab, but I doubt that
qualifies as something that is actually reviewable by the ASF’s rules…. which
brings us back here anyway if I turn this into a giant patch file.
