[openstack-dev] [tripleo][ci][infra] Quickstart Branching

Alex Schultz aschultz at redhat.com
Wed May 23 17:29:47 UTC 2018


On Wed, May 23, 2018 at 10:49 AM, Sagi Shnaidman <sshnaidm at redhat.com> wrote:
> Alex,
>
> the problem is that you're working and focusing mostly on release specific
> code like featuresets and some scripts. But tripleo-quickstart(-extras) and
> tripleo-ci is much *much* more than set of featuresets. Only 10% of the code
> may be related to releases and branches, while other 90% is completely
> independent and not related to releases.
>

It is not necessarily about release specific code, it's about being
able to reduce the impact of a change. From my original reply:

> If there's a high maintenance cost, we haven't properly identified the optimal way to separate functionality between tripleo/quickstart.

IMHO this is a side effect of having a whole bunch of roles in a
single repo.  oooq-extras has a mix of tripleo and non-tripleo related
content. The reproducer IMHO is related to provisioning and could fall
in the oooq repo and not oooq-extras.  This is a structure problem
with quickstart.  If it's not version specific, then don't put it in a
version specific repo. But that doesn't mean don't use version
specific repos at all.

This is one of the reasons why we're opting not to use this pattern of
a bunch of roles in a single repo for tripleo itself[0][1][2].  We
learned with the puppet modules that carrying all this stuff in a
single repo has a huge maintenance cost and if you split them out you
can identify re-usability and establish proper patterns for moving
functionality into a shared place[3].  Yes there is a maintenance cost
of maintaining independent repos, but at the same time there's a
benefit of re-usability by other projects/groups when you expose
important pieces of functionality as a standalone. You can establish
clear ways to interact with each piece, test items, and release
independently.  For example the ansible-role-container-registry is not
tripleo specific and anyone looking to manage a standalone docker
registry can use it & contribute.

> So in 90% code we DO need to backport every change, take for example the
> latest patch to extras: https://review.openstack.org/#/c/570167/, it's
> fixing reproducer. If oooq-extra was branched, we would need to backport
> this fix to every and every branch. And the same for all other 90% of code,
> which is complete nonsense.
> Just because not using "{% if release %}" construct - to block the whole
> work of CI team and make the CI code is absolutely unmaintainable?
>

And you're saying what we currently have is maintainable?  We keep
breaking ourselves, there's big gaps in coverage and it takes
time[4][5] to identify breakages. I don't consider that maintainable
because this is a recurring topic because we clearly haven't fixed it
with the current setup.  It's time to re-evaluate what we have an see
if there's room for improvement.  I know I wasn't proposing to branch
all the repositories, but it might make sense to figure out if there's
a way to reduce our recurring issues with stable branches or
independent modules for some of the functions in CI.

> Some of release related templates we moved recently from tripleo-ci to THT
> repo like scenarios, OC templates, etc. If we discover another things in
> oooq that could be moved to branched THT I'd be only happy for that.
>
> Sometimes it could be hard to maintain one file in extras templates with
> different logic for releases, like we have in tempest configuration for
> example. The solution is to create a few release-related templates and use
> one that match the current branch. It doesn't affect 90% of code and still
> "branch-like" approach. But I didn't see other scripts that are so release
> dependent. If we'll have ones, we could do the same. For now I see "{% if
> release %}" construct working very well.

Considering this is how we broke Queens, I'm not sure I agree.

>
> I didn't see still any advantage of branching CI code, except of a little
> bit nicer jinja templates without "{% if release ", but amount of
> disadvantages is so huge, that it'll literally block all current work in CI.
>

It's about reducing our risk with test coverage. We do not properly
test all jobs and all configurations when we make these changes. This
is a repeated problem and when we have to add version specific logic,
unless we're able to identify what this is actually impacting and
verify with jobs we have a risk of breaking ourselves.  We've seen
that code review is not sufficient for these changes as we merge
things and only find out after they've been merged that we broke
stable branches. Then it takes folks tracking down changes to decipher
what we broke. For example the original patch[4] broke Queens for
about a week.  That's 7 days of nothing being able to be merged,
that's not OK.

Thanks,
-Alex

[0] http://git.openstack.org/cgit/openstack/ansible-role-container-registry/
[1] http://git.openstack.org/cgit/openstack/ansible-role-redhat-subscription/
[2] http://git.openstack.org/cgit/openstack/ansible-role-tripleo-keystone/
[3] http://git.openstack.org/cgit/openstack/puppet-openstacklib/
[4] https://review.openstack.org/#/c/565856/
[5] https://review.openstack.org/#/c/569830

> Thanks
>
>
>
> On Wed, May 23, 2018 at 7:04 PM, Alex Schultz <aschultz at redhat.com> wrote:
>>
>> On Wed, May 23, 2018 at 8:30 AM, Sagi Shnaidman <sshnaidm at redhat.com>
>> wrote:
>> > Hi, Sergii
>> >
>> > thanks for the question. It's not first time that this topic is raised
>> > and
>> > from first view it could seem that branching would help to that sort of
>> > issues.
>> >
>> > Although it's not the case. Tripleo-quickstart(-extras) is part of CI
>> > code,
>> > as well as tripleo-ci repo which have never been branched. The reason
>> > for
>> > that is relative small impact on CI code from product branching. Think
>> > about
>> > backport almost *every* patch to oooq and extras to all supported
>> > branches,
>> > down to newton at least. This will be a really *huge* price and non
>> > reasonable work. Just think about active maintenance of 3-4 versions of
>> > CI
>> > code in each of 3 repositories. It will take all time of CI team with
>> > almost
>> > zero value of this work.
>> >
>>
>> So I'm not sure I completely agree with this assessment as there is a
>> price paid for every {%if release in [...]%} that we have to carry in
>> oooq{,-extras}.  These go away if we branch because we don't have to
>> worry about breaking previous releases or current release (which may
>> or may not actually have CI results).
>>
>> > What regards patch you listed, we would have backport this change to
>> > *every*
>> > branch, and it wouldn't really help to avoid the issue. The source of
>> > problem is not branchless repo here.
>> >
>>
>> No we shouldn't be backporting every change.  The logic in oooq-extras
>> should be version specific and if we're changing an interface in
>> tripleo in a breaking fashion we're doing it wrong in tripleo. If
>> we're backporting things to work around tripleo issues, we're doing it
>> wrong in quickstart.
>>
>> > Regarding catching such issues and Bogdans point, that's right we added
>> > a
>> > few jobs to catch such issues in the future and prevent breakages, and a
>> > few
>> > running jobs is reasonable price to keep configuration working in all
>> > branches. Comparing to maintenance nightmare with branches of CI code,
>> > it's
>> > really a *zero* price.
>> >
>>
>> Nothing is free. If there's a high maintenance cost, we haven't
>> properly identified the optimal way to separate functionality between
>> tripleo/quickstart.  I have repeatedly said that the provisioning
>> parts of quickstart should be separate because those aren't tied to a
>> tripleo version and this along with the scenario configs should be the
>> only unbranched repo we have. Any roles related to how to
>> configure/work with tripleo should be branched and tied to a stable
>> branch of tripleo. This would actually be beneficial for tripleo as
>> well because then we can see when we are introducing backwards
>> incompatible changes.
>>
>> Thanks,
>> -Alex
>>
>> > Thanks
>> >
>> >
>> > On Wed, May 23, 2018 at 3:43 PM, Sergii Golovatiuk <sgolovat at redhat.com>
>> > wrote:
>> >>
>> >> Hi,
>> >>
>> >> Looking at [1], I am thinking about the price we paid for not
>> >> branching tripleo-quickstart. Can we discuss the options to prevent
>> >> the issues such as [1]? Thank you in advance.
>> >>
>> >> [1] https://review.openstack.org/#/c/569830/4
>> >>
>> >> --
>> >> Best Regards,
>> >> Sergii Golovatiuk
>> >>
>> >>
>> >> __________________________________________________________________________
>> >> OpenStack Development Mailing List (not for usage questions)
>> >> Unsubscribe:
>> >> OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>> >> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>> >
>> >
>> >
>> >
>> > --
>> > Best regards
>> > Sagi Shnaidman
>> >
>> >
>> > __________________________________________________________________________
>> > OpenStack Development Mailing List (not for usage questions)
>> > Unsubscribe:
>> > OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>> >
>>
>> __________________________________________________________________________
>> OpenStack Development Mailing List (not for usage questions)
>> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
>
>
> --
> Best regards
> Sagi Shnaidman
>
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>



More information about the OpenStack-dev mailing list