openstack team mailing list archive

[Chris Wright <chrisw@xxxxxxxxxxxx>]

Overall, I think you've described problem and solutions well.  A few
thoughts below.

* Mark McLoughlin (markmc@xxxxxxxxxx) wrote:
> Firstly, problem definition:
> 
>   - Nova is big, complex and has a fairly massive rate of churn. While 
>     the nova-core team is big, there isn't enough careful review going 
>     on by experts in particular areas and there's a consistently large
>     backlog of reviews.
> 
>   - Developers working on features are very keen to have their work 
>     land somewhere and this leads to half-finished features being 
>     merged onto master rather than developers collaborating to get a 
>     feature to a level of completeness and polish before merging into 
>     master.

I'm not sure having subsystem trees and feature branches (as defined
below) actually address the desire to land somewhere anymore than
developers' github forks do.  A feature needs to be just as complete
when it lands in a subsystem tree as it does when it lands in master.
I don't think it's a shortcoming to the model here, but should be made
nice and clear.

> Some assumptions about the solution:
> 
>   - There should be a small number of domain experts who can approve 
>     changes to each of major subsystems. This will encourage 
>     specialization and give more clear lines of responsibility.
> 
>   - There should be a small number of project dictators who have final 
>     approval on merge proposals, but who are not expected to review 
>     every patch in great detail. This is good because we need someone 
>     with an overall view of the project who can make sure efforts in 
>     the various subsystems are coordinated, without that someone being 
>     massively overloaded.
> 
>   - New features should be developed on a branch and brought to a level 
>     of completeness before being merged into master. This is good 
>     because we don't want half-baked stuff in master but also because 
>     it encourages developers to break their features into stages where 
>     each stage of the work can be brought to completion and merged 
>     before moving on to the next stage.
> 
>   - In essence, we're assuming some variation of the kernel distributed 
>     development model.
> 
>     (FWIW, my instinct is to avoid the kernel model on projects. Mostly 
>     because it's extremely complex and massive overkill for most 
>     projects. Looking at the kernel history with gitk is enough to send 
>     anyone screaming for the hills. However, Nova seems to be big 
>     enough that we're experiencing the same pressures that drove the 
>     kernel to adopt their model)

I'm certainly biased, but I think this is a great model ;)

> Ok, what are "subsystem branches" and how would they work?
> 
>   - Subsystem branches would have a small number of maintainers who can 
>     approve a change. These would be domain experts providing strong 
>     oversight over a particular area.
> 
>     (In gerrit, this is a branch with a small team or single person who 
>     can +1 approve a review)
> 
>   - Project dictators don't need to do detailed reviews of merge 
>     proposals from subsystem maintainers. The dictator's role is mostly 
>     just to sign off on the merge proposal. However, the dictator can 
>     comment in the proposal on things which could have been done better 
>     and the subsystem maintainer should take note of these comments and 
>     perhaps retroactively fix them up. Ultimately, though, the dictator 
>     can have exercise a veto if the merge proposal is unacceptable or 
>     if the subsystem maintainer is consistently making the same 
>     mistakes.
> 
>   - It would be up to the project dictators to help drive patches 
>     through the right subsystem branches - e.g. they might object if 
>     one subsystem maintainer merged a patch that inappropriately cut 
>     into another subsystem or they might refuse to merge a given patch
>     into the main branch unless it went through the appropriate 
>     subsystem branch.
> 
>     (In gerrit, this would mean a small team or single person who can 
>     +1 approve merge proposals on master. They would -1 proposals
>     submitted against master which should have been submitted against a 
>     subsystem branch.)
> 
>   - Subsystem branches might not necessarily be blessed centrally. It 
>     might be a case that anyone can create such a branch and, over 
>     time, establish trust with the project dictators. Subsystem 
>     branches would come and go. This is the mechanism by which 
>     subsystem maintainership is transferred between people over time.
> 
>     (In gerrit, this means people need to easily be able to create 
>     their own branches)
> 
>     (What's more difficult to imagine in gerrit is how a new, potential 
>     subsystem maintainer comes along, starts hoovering up patches into 
>     her branch and submitting them in batches. Where does she hoover 
>     them up from and how does she say "I've merged this into my branch, 
>     don't merge it via another branch")
> 
>   - Bisectability remains important. Subsystem maintainers don't merge 
>     broken commits into their subsystem branch and the project 
>     dictators can enforce this using their veto. It is not good enough 
>     for subsystem maintainers to consistently merge broken commits into 
>     their branch, fix it up with a later commit and include both 
>     commits their merge proposals.

Yes!

>     (I don't think we'd use Jenkins to enforce this, but subsystem 
>     maintainers might use it as a tool to help them catch issues. So, 
>     the full set of gating tests would only gate merges into master but 
>     subsystem branches might choose to gate merges into their branch on 
>     the unit tests. Subsystem maintainers might also use Smokestack to 
>     pre-gate merge proposals to the subsystem branch)
> 
>   - Subsystem branches would not rebase unless the project dictator 
>     outright rejects a merge request from the subsystem branch (i.e.
>     "I'm not merging commit abcdef0! Fix it and rebase!"). This means 
>     the subsystem maintainer will need to regularly (probably only when 
>     there are conflicts to be dealt with) merge master back into the
>     subsystem branch.

Any branch that has a downstream user (someone who's cloned that branch)
should not rebase.  If the subsystem->master merge prop is rejected, the
subsystem should generate a new branch, re-apply the relevant changes
(dropping, rearranging, etc as necessary), and start a new merge prop.
I'm unclear on where gerrit fits on that rebase step?

Regular merging master back is one culprit in your "screaming for the
hills" review of kernel git history.  In fact, it's discouraged there.
The workflow is helped by the merge windows which makes for discrete
merge points.

>   - Plausible subsystem branches are e.g.:
> 
>       - OpenStack APIs
>       - EC2 API
>       - virt
>          - libvirt driver
>          - xenapi driver
>          - vmware driver
>       - networking
>       - volumes
>       - scheduler
>       - RPC
> 
>     Deciding which areas make sense as a subsystem branch is 
>     non-trivial.
> 
>     Should there be a "DB" subsystem? Probably not, because that would 
>     mean every new feature needs to come through this branch or, 
>     alternatively, the DB maintainer would need to accept DB schema 
>     additions without the context of how it's being used higher up the 
>     stack.

This does touch on the reality that some changes will span subsystems
which will require coordination between subsystems.  Having a mechanism
for this coordination ahead of time is helpful.  In Linux it's one job
of the linux-next tree, but that doesn't address jenkins side.

>     Ok, so why does it make sense to have an "OpenStack APIs" 
>     subsystem? Don't all features affect that branch too? Well, maybe, 
>     but the APIs really do need strong oversight. Perhaps we can be 
>     confident that we can add e.g. a new scheduler feature through the
>     scheduler branch and then later merge any API additions through the 
>     APIs branch.
> 
> And how about feature branches?
> 
>   - Feature branches are relatively short-lived (i.e. weeks or months
>     rather than years) branches for a specific feature. They are a
>     mechanism for developers to work on a patch series in the open until
>     the feature is complete enough to be merged into a subsystem branch
>     or master.
> 
>     (I'm not sure gerrit is right for this. Why not just do it in 
>     folk's github forks? I think all people are looking for is for 
>     people to be more aware of feature branches. How about if you put 
>     details of your feature branch in the blueprint for the feature?)
> 
>     (If not using gerrit, can developers configure Jenkins to CI their 
>     branch? Or is Smokestack the right tool?)
> 
>   - Feature branches rebase, do not contain merge commits and each 
>     commit on the branch is functional, bisectable and self-contained.
> 
>   - When a feature branch is ready to be merged into a subsystem 
>     branch, the patch series is submitted for review. The subsystem 
>     maintainer will likely require changes to individual patches and 
>     these changes would be made on the feature branch and squashed back 
>     into the appropriate individual patch.
>
>     (Ideally gerrit's "topic review" feature will get upstream and 
>     we'll use that. This would mean that a patch series could be 
>     proposed for review as a single logical unit while still keeping 
>     individual patches as separate commits)

One topic that came up at the design summit that I think you mean by the
above is the "to squash or not to squash?" question.

The series may need rework (rebase, fix+internal squash, etc) while
it's a feature branch, but it gets merge prop'd as a real stand alone,
bisectable patchset (minus the true work-in-progress fiuxps/changes,
of course).  Put another way... 'git rebase -i' is your friend ;)

>   - Because feature branches rebase, active day-to-day collaboration
>     with others is difficult. You certainly can't have multiple people
>     rebasing the same branch, that way lies madness.
> 
>     There are ways to have multiple people work actively on the same 
>     rebasing branch e.g.
> 
>       http://blogs.gnome.org/markmc/2011/02/26/git-rebasing-cont/
> 
>     but, ultimately, feature branches are going to be owned by a single 
>     person who might incorporate patches from others.
> 
>     (Incorporating the work of others, rebasing and squashing means a 
>     patch might have multiple contributors but only one author listed in
>     git. That makes CLA enforcement impossible, but we should drop the 
>     CLA in favour of the kernel-like Signed-off-by: tag. See this 
>     discussion: https://lists.launchpad.net/openstack/msg06544.html )
> 
>   - One option for longer-lived, active collaboration is for a subsystem
>     maintainer to create a feature branch and review the work as it is 
>     ongoing. The idea being that the subsystem maintainer commits to 
>     not requiring the feature branch to be rebased before it is merged 
>     into the subsystem branch.