[openstack-dev] [stable][neutron] backports vs. vendor decomposition

Ihar Hrachyshka ihrachys at redhat.com
Wed Jan 21 21:42:13 UTC 2015


Hi all,

as per: 
https://github.com/openstack/neutron-specs/blob/master/specs/kilo/core-vendor-decomposition.rst, 
neutron is going to spin off vendor plugins into separate trees outside 
of neutron core team control. This raises several questions on how we 
are going to handle stable branches that will still include plugin code 
for several cycles.

1) If a plugin is already spinned off and a patch is applicable for 
stable branches, there are two cases:
- patch is not merged into vendor repo;
- patch is merged into the vendor repo.

My take is:
- if it's merged in the vendor repo, then we just cherry-pick from there 
(it should just work if vendor repo was created with the whole master 
history saved).
- if it's not merged into the repo, I would recommend the author to 
propose and merge it there first. If there are any justifiable issues 
with proposing it for vendor repo inclusion, then we can consider 
stable-only merge.

2) If a plugin is in the middle of spinning off and a patch is 
applicable for stable branch, then there are two options:
- require plugin to spin off first and then apply the patch to vendor 
repo, or
- allow some types of patches to be merged into master while vendors are 
working on spinning off the code.

Examples of those patches are:
- https://review.openstack.org/#/c/147976/
- https://review.openstack.org/#/c/148369/

Currently the patches above are blocked for master inclusion assuming 
the spin off must take place first, and then bugs should be fixed in 
vendor repo. At the same time, we usually avoid backports unless the 
code is not in master anymore, but that's not the case here. So the 
current approach effectively blocks any bug fixes for plugins in stable 
branches.

If we would be sure that a plugin is out of the tree till Kilo, then it 
would indeed be a waste of time to review the code for neutron/master 
since it would be guaranteed to be released as a separate packagr e 
anyway. In that case, it would be ok to forbid any patches for the  
plugin code till its spin off from master, and the patch would go 
directly to stable branches.

That said, it would potentially introduce regressions if we consider 
upgrades from Juno to Kilo + vendor repo. We may say that since the 
regression would be on vendor plugin side, and neutron team does not 
have anything to do with spinned off plugins, that would be fine for us.

But: we cannot guarantee that a plugin wil leave the neutron tree this 
cycle. The spec explicitly gives permission to stay in the tree till 
L-cycle, and in that case it will be our responsibility to handle 
regressions in Kilo that we may introduce by blocking master fixes.

I think we should try to set procedure that would avoid potential 
regressions even if they will come from vendor repos.

We allow fixes that are not applicable for final releases for master if 
it's to be backported in stable branches. F.e. see 
https://review.openstack.org/#/c/127633/ that was merged into master 
while pecan migration should make it useless for Kilo.

It's my belief plugin code bug fixes in stable branches should be 
treated the same way.

That said, we should expect vendors to run third party CI for stable 
branches if they want to see backports merged in.

***
I think the correct approach here is:
- once a plugin is spinned off, consider it is a 'master' for the code, 
and backport to stable branches directly from there;
- before a plugin is spinned off, assume that it's not going to be 
spinned off in Kilo, and hence allow bug fixes in neutron/master (but 
not new features);
- once we get to L release that requires all vendor plugin to go out, 
forbid any fixes for the code, assuming they will either spin off or 
will be dropped anyway.
***

The approach is pretty similar to how oslo project handles new library 
spin-offs from oslo-incubator. Yes, there is a difference here: in 
neutron, we loose any control on spinned off repos. Though I don't feel 
it justifies stable-only fixes while we can easily add value to vendor 
code by asking people to consider fixing the bug there first. More 
importantly, nothing should justify blocking bug fixing for stable branches.

Thoughts?

/Ihar



More information about the OpenStack-dev mailing list