Sorry for being a few days late on this.
I totally agree with the direction this conversation is going, but as
someone who has done build engineering in various incarnations over the
last couple of years, I would suggest two additional things. Build per
branch and gitflow. We have had good success in quality control by
requiring developers to create a new branch for all activities, which
automatically creates a new build in the build system. This allows people
to have full builds of their branch, that in theory must be passing, before
they merge back into master/dev/trunk. This is a bit different from the
gerrit solution where you are still committing to master regularly. We
also do a review/pull request system to regulate that merge process. I
know it was said that limiting the input by having a review process would
be bad, I'm a firm believer in having two sets of eyes look at everything
makes things generally better.
This does come with an increased build infrastructure cost since running
builds for all of the branch changes can be costly. We implement this
internally, and have a build server setup in the office which is pulling
several interesting branches from ACS and running builds daily so we can
keep up better on current status and test additional items that are
specific to our environment. This is mirrored into our internal git repo
where we can create branches for testing random fixes and such (we aren't
committers yet, but we can dream), but still have full build and test
support. My goal is to eventually do a full build, deploy, configure, and
automated tests against a scratch built local cloud with every commit from
selected branches or daily from master (Hooray for nested virtualization
and excess hardware). But I'm also a bit of a build nerd.
Even without reviews, something of this nature could help improve quality
as well.
On Sat, Feb 8, 2014 at 12:19 AM, Rohit Yadav <bhais...@apache.org> wrote:
> On Fri, Feb 7, 2014 at 2:16 PM, Hugo Trippaers <h...@trippaers.nl> wrote:
> > Hey David,
> >
> > I would make a distinction between code issues and functional issues.
> Occasionally somebody just plainly breaks the build, i'm guilty of that
> myself actually, and thats just plain stupid. Luckily we have Jenkins to
> catch these errors quickly. I'm in a continuous struggle with Jenkins to
> get the build time to less than 5 minutes. I feel that is an acceptable
> time to get feedback on a commit, any longer and you have moved on to the
> next thing or gone home.
>
> Why not do incremental builds, since last build on a git SHA to speed
> up by cheating?
>
> May I share a hack I use based on an old method once shared by Edison
> to do a fast builds locally:
>
> 1. If you just clone cloudstack, build once this will probably take a
> lot of time, also to get deps and whatnot
>
> 2. Start making changes and build only projects/modules that got
> changed using: (you make create a shell function that wraps this in
> your bashrc or zshrc etc)
>
> mvn -pl `git status --porcelain |sed -n '/\/src/p'| awk '{print $2}'
> |sed 's/\/src/$/'|cut -d $ -f 1|uniq |tr "\n" "," |sed
> 's/,$/,client/'` clean install
>
> 3. Commit often on your branch after build works and when you're
> implemented your stuff do a squash merge on master (or target branch)
> which results in a single (reviewable) commit. This probably will save
> you from breaking builds.
>
> > Also this kind of testing isn't really hard, run the build and unit
> tests. By introducing something like gerrit we can actually make this
> happen before committing it to the repo. Push a patch to gerrit, gerrit
> tells jenkins to test the patch, if +1 from jerkins commit, for non
> committers the step would be to invite somebody for review as well. Second
> nice thing about jenkins is the post-review test, if a contributor submits
> a patch its build by jenkins, if a reviewes approves the patch, jerkins
> will again run a build to ensure that the patch will still apply and
> doesn't break the build. Very handy if there is some time between patch
> submission and patch review.
>
> +1
>
> I think it's a culture issue, while we may think that introducing and
> forcing everyone to go through code reviewing process will slow
> everyone down IMHO over time this will inculcate the habit in everyone
> to do code reviews for others (their patches/branches etc.) so others
> would do for them. It can fail in case there are not enough reviewers
> available for a code review or they lack interest/time (our state with
> reviewboard).
>
> So, the trick could be to have, in addition of the reviewers, few
> assigned maintainers who are responsible for churning out pending
> reviews in parts of the codebase they understand very well and can
> help with reviews process.
>
> >
> > Functional issues are much harder to track. For example yesterday i
> found several issues in the contrail plugin that would not cause any pain
> in a contrail environment, but any other environments creating a network
> would fail. These examples are too common and difficult to catch with unit
> tests. It can be done, but requires some serious effort on the developers
> side and we in general don't seem to be very active at writing unit tests.
> These kind of issues can only be found by actually running CloudStack and
> executing a series of functional tests. Ideally that is what we have the
> BVT suite for, but i think our current BVT setup is not documented enough
> to give accurate feedback to a developer about which patch broke a certain
> piece of functionality. In jenkins the path from code to BVT is not kept
> yet, so it is almost impossible to see which commits were new in a
> particular run of the bvt suite.
> >
> > Personally i'm trying to get into the habit of running a series of tests
> on devcloud before committing something. Doesn't prove a lot, but does
> guarantee that the bare basic developer functionality is working before
> committing something. After a commit at least i'm sure that anybody will be
> able to spinup devcloud and deploy an instance. I'm trying to get this
> automated as well so we can use this as feedback on a patch. Beers for
> anyone who writes an easy to use script that configures devcloud with a
> zone and tests if a user vm can be instantiated on an isolated sourcenat
> network. If we could include such a script in the tree it might help people
> with testing their patch before committing.
>
> I once discussed an idea that we use something fast to do the testing,
> so instead of vms or nested vms we can use a mocked hypervisor
> (simulator) or LXC container on a VM locally. But again, in every case
> they will be just passing very small set of test cases.
>
> Regards.
>
> >
> > I think we are seeing more and more reverts in the tree. Not necessarily
> a good thing, but at least people know that there is that option if a
> commit really breaks a build. Also please help each other out, everybody
> can make a mistake and commit it. If its a trivial mistake it might not be
> much effort to track it down and fix it, which is way better than a revert
> or a mail that something is broken.
> >
> > In short, we need to make testing more efficient and transparent to
> allow people to easily incorporate it in their personal workflow.
> >
> > Cheers,
> >
> > Hugo
> >
> > On 7 feb. 2014, at 04:50, David Nalley <da...@gnsa.us> wrote:
> >
> >> Hi folks,
> >>
> >> We continue to break things large and small in the codebase, and after
> >> a number of different conversations; I thought I'd bring that
> >> discussion here.
> >>
> >> First - coding quality is only one factor that the PMC considers when
> >> making someone a committer.
> >>
> >> Second - CloudStack is a huge codebase; has a ton of inter-related
> >> pieces, and unintended consequences are easy.
> >>
> >> We also have an pretty heady commit velocity - 20+ commits today alone.
> >>
> >> Some communities have Review-then-commit - which would slow us down,
> >> and presumably help us increase quality. However, I am not personally
> >> convinced that it will do so measurably because even the most
> >> experienced CloudStack developers occasionally break a build or worse.
> >>
> >> We could have an automated pipeline that verifies a number of
> >> different tests pass - before a patch/commit makes it into a mainline
> >> branch. That is difficult with our current tooling; but perhaps
> >> something worth considering.
> >>
> >> At FOSDEM, Hugo and I were discussing his experiences with Gerrit and
> >> OpenDaylight, and he thinks thats a viable option. I think it would
> >> certainly be a step in the right direction.
> >>
> >> Separately, Jake Farrell and I were discussing our git-related
> >> proposal for ApacheCon, and broached the subject of Gerrit. Jake is
> >> the current person bearing most of the load for git at the ASF, and
> >> he's also run Gerrit in other contexts. He points out a number of
> >> difficulties. (And I'd love for him to weigh in on this conversation,
> >> hence the CC) He wants to expand RB significantly, including
> >> pre-commit testing.
> >>
> >> So - thoughts, comments, flames? How do we improve code quality, stop
> >> needless breakage? Much of this is going to be cultural I think, and I
> >> personally think we struggle with that. Many folks have voiced an
> >> opinion about stopping continued commits when the build is broken; but
> >> we haven't been able to do that.
> >>
> >> --David
> >
>
--
*Nate Gordon*Director of Technology | Appcore - the business of cloud
computing®
Office +1.800.735.7104 | Direct +1.515.612.7787
nate.gor...@appcore.com | www.appcore.com
----------------------------------------------------------------------
The information in this message is intended for the named recipients only.
It may contain information that is privileged, confidential or otherwise
protected from disclosure. If you are not the intended recipient, you are
hereby notified that any disclosure, copying, distribution, or the taking
of any action in reliance on the contents of this message is strictly
prohibited. If you have received this e-mail in error, do not print it or
disseminate it or its contents. In such event, please notify the sender by
return e-mail and delete the e-mail file immediately thereafter. Thank you.
