On Wed, Mar 3, 2021 at 7:35 AM Slawek Kaplonski <skaplons@redhat.com> wrote:
Hi,
Dnia środa, 3 marca 2021 05:01:24 CET Sean Mooney pisze:
On Wed, 2021-03-03 at 15:59 +1300, Lingxian Kong wrote:
Hi,
Thanks for all your hard work on this.
I'm wondering is there any doc proposed for devstack to tell people who are not interested in OVN to keep the current devstack behaviour? I have a feeling that using OVN as default Neutron driver would break the CI jobs for some projects like Octavia, Trove, etc. which rely on ovs port for the set up.
You can look on how we set some of our jobs to use ML2/OVS: https:// review.opendev.org/c/openstack/neutron-tempest-plugin/+/749503/26/zuul.d/ master_jobs.yaml#121
Thanks Lingxian and Slaweq for the suggestion/replies. Apart from what Slaweq pointed out, the DevStack documentation have a sample config file (https://docs.openstack.org/devstack/latest/#create-a-local-conf) as part of the steps to deploy it, I was thinking I could propose another sample file that would enable ML2/OVS for the deployment and add a note in that doc; if you think it's worth it. Also what gives us more confidence regarding projects like Octavia is that TripleO already uses ML2/OVN as the default driver for their overcloud, so many of these projects are already being tested/used with OVN.
well ovn is just an alternivie contoler for ovs. so ovn replace the neutron l2 agent it does not replace ovs. project like octavia or trove that deploy loadblances or dbs in vms should not be able to observe a difference. they may still want to deploy ml2/ovs but unless they are doing something directly on the host like adding port directly to ovs because they are not using vms they should not be aware of this change.
my reticense to cahnge at this point in the cyle for nova is motiveated maily by gate stablity. the active contibutes to nova dont really have experince with ovn and how to debug it in the gate. we also are getting close to FF when we tend to get a lot of patches and the gate stablity become even more imporant so adding a new vaiabrly to that mix but swaping out the networkbacked between now and the wallaby release seams problematic.
in any cases swapping back should ideallly be as simple as setting Q_AGENT=openvswitch.
i have not looked at the patches but to swap betwween ovs and linuxbidge you just define Q_AGENT=linuxbridge for the most part so im expecting that we would just enable Q_AGENT=ovn or somthing simialr for ovn.
i know ovn used to have its own devstack plugin but if we are makeing it the default that means it need to be support nativly in devstack not as a plugin so useing Q_AGENT=ovn to enable it and make that the new default would seam to be the simplest way to manage that.
but yes documenting how to enabel the old behavior is still important the example nova patch shows how to hardcode the old behavior https://review.opendev.org/c/openstack/nova/+/776944/5/.zuul.yaml#234 it seams to be doing this a little more explictly then i would like but its not that hard. i would suggest adding a second sample loca.conf in devstack for standard ovs deployments.
--- Lingxian Kong Senior Cloud Engineer (Catalyst Cloud) Trove PTL (OpenStack) OpenStack Cloud Provider Co-Lead (Kubernetes)
On Wed, Mar 3, 2021 at 2:03 PM Goutham Pacha Ravi <gouthampravi@gmail.com>
wrote:
On Mon, Mar 1, 2021 at 10:29 AM Sean Mooney <smooney@redhat.com> wrote:
On Mon, 2021-03-01 at 16:07 +0000, Lucas Alvares Gomes wrote:
Hi all,
As part of the Victoria PTG [0] the Neutron community agreed upon switching the default backend in Devstack to OVN. A lot of work has been done since, from porting the OVN devstack module to the DevStack tree, refactoring the DevStack module to install OVN from distro packages, implementing features to close the parity gap with ML2/OVS, fixing issues with tests and distros, etc...
We are now very close to being able to make the switch and we've thought about sending this email to the broader community to raise awareness about this change as well as bring more attention to the patches that are current on review.
Note that moving DevStack to ML2/OVN does not mean that ML2/OVS is discontinued and/or not supported anymore. The ML2/OVS driver is still going to be developed and maintained by the upstream Neutron community.
can we ensure that this does not happen until the xena release. in generall i think its ok to change the default but not this late in
the cycle.
i would also like to ensure we keep at least one non ovn based multi
node job in
nova until https://review.opendev.org/c/openstack/nova/+/602432 is
merged and possible after.
right now the event/neutorn interaction is not the same during move
operations.
Below is a e per project explanation with relevant links and issues of where we stand with this work right now:
* Keystone:
Everything should be good for Keystone, the gate is happy with the changes. Here is the test patch: https://review.opendev.org/c/openstack/keystone/+/777963
* Glance:
Everything should be good for Glace, the gate is happy with the changes. Here is the test patch: https://review.opendev.org/c/openstack/glance/+/748390
* Swift:
Everything should be good for Swift, the gate is happy with the changes. Here is the test patch: https://review.opendev.org/c/openstack/swift/+/748403
* Ironic:
Since chainloading iPXE by the OVN built-in DHCP server is work in progress, we've changed most of the Ironic jobs to explicitly enable ML2/OVS and everything is merged, so we should be good for Ironic too. Here is the test patch: https://review.opendev.org/c/openstack/ironic/+/748405
* Cinder:
Cinder is almost complete. There's one test failure in the "tempest-slow-py3" job run on the "test_port_security_macspoofing_port" test.
This failure is due to a bug in core OVN [1]. This bug has already been fixed upstream [2] and the fix has been backported down to the branch-20.03 [3] of the OVN project. However, since we install OVN from packages we are currently waiting for this fix to be included in the packages for Ubuntu Focal (it's based on OVN 20.03). I already contacted the package maintainer which has been very supportive of this work and will work on the package update, but he maintain a handful of backports in that package which is not yet included in OVN 20.03 upstream and he's now working with the core OVN community [4] to include it first in the branch and then create a new package for it. Hopefully this will happen soon.
But for now we have a few options moving on with this issue:
1- Wait for the new package version 2- Mark the test as unstable until we get the new package version 3- Compile OVN from source instead of installing it from packages (OVN_BUILD_FROM_SOURCE=True in local.conf)
i dont think we should default to ovn untill a souce build is not
required.
compiling form souce while not supper expensice still adds time to the
job
execution and im not sure we should be paying that cost on every
devstack job run.
we could maybe compile it once and bake the package into the image or
host it on a mirror
but i think we should avoid this option if we have alternitives.
What do you think about it ?
Here is the test patch for Cinder: https://review.opendev.org/c/openstack/cinder/+/748227
* Nova:
There are a few patches waiting for review for Nova, which are:
1- Adapting the live migration scripts to work with ML2/OVN: Basically the scripts were trying to stop the Neutron agent (q-agt) process which is not part of an ML2/OVN deployment. The patch changes the code to check if that system unit exists before trying to stop it.
Patch: https://review.opendev.org/c/openstack/nova/+/776419
2- Explicitly set grenade job to ML2/OVS: This is a temporary change which can be removed one release cycle after we switch DevStack to ML2/OVN. Grenade will test updating from the release version to the master branch but, since the default of the released version is not ML2/OVN, upgrading from ML2/OVS to ML2/OVN as part of the grenade job is not supported.
Patch: https://review.opendev.org/c/openstack/nova/+/776934
3- Explicitly set nova-next job to ML2/OVS: This job uses the QoS minimum bandwidth feature which is not yet supported by ML2/OVN [5][6] therefore we are temporarily enabling ML2/OVS for this job until that feature lands in core OVN.
Patch: https://review.opendev.org/c/openstack/nova/+/776944
I also spoke briefly with Sean Mooney (irc: sean-k-mooney) about these changes and he suggested keeping all the Nova jobs on ML2/OVS for now because he feels like a change in the default network driver a few weeks prior to the upstream code freeze can be concerning. We do not know yet precisely when we are changing the default due to the current patches we need to get merged but, if this is a shared feeling among the Nova community I can work on enabling ML2/OVS on all jobs in Nova until we get a new release in OpenStack.
yep this is still my view. i would suggest we do the work required in the repos but not merge it
until the xena release
is open. thats technically at RC1 so march 25th i think we can safely do the swich after that but i would not change the
defualt in any project
before then.
Here's the test patch for Nova: https://review.opendev.org/c/openstack/nova/+/776945
* DevStack:
And this is the final patch that will make this all happen: https://review.opendev.org/c/openstack/devstack/+/735097
It changes the default in DevStack from ML2/OVS to ML2/OVN. It's been a long and bumpy road to get to this point and I would like to say thanks to everyone involved so far and everyone that read the whole email, please let me know your thoughts.
thanks for working on this.
[0] https://etherpad.opendev.org/p/neutron-victoria-ptg [1] https://bugs.launchpad.net/tempest/+bug/1728886 [2]
https://patchwork.ozlabs.org/project/openvswitch/patch/20200319122641.47 3776-1-numans@ovn.org/> >
[3]
https://github.com/ovn-org/ovn/commit/0c26bc03064f2c21d208f0f860b48d8ab3 9380cb> >
[4]
https://mail.openvswitch.org/pipermail/ovs-discuss/2021-February/050961. html> >
[5]
https://github.com/openstack/nova/blob/ded25f33c734ebff963f06984707a99fe 76a9ee1/gate/post_test_hook.sh#L129-L143> >
[6] https://docs.openstack.org/neutron/latest/ovn/gaps.html
Cheers, Lucas
++ Thank you indeed for working diligently on this important change.
Please do note that devstack, and the base job that you're modifying is used by many other projects besides the ones that you have enumerated in the subject line. I suggest using [all] as a better subject line indicator to get the attention of folks like me who have filters based on the subject line. Also, the network substrate is important for the project I help maintain: Manila, which provides shared file systems over a network - so I followed your lead and submitted a dependent patch. I hope to reach out to you in case we see some breakages: https://review.opendev.org/c/openstack/manila-tempest-plugin/+/778346
-- Slawek Kaplonski Principal Software Engineer Red Hat