[occm] Multi region openstack cluster

[sergelogvinov]

What this PR does / why we need it:

Openstack CCM multi region support, if it has one Identity provider.

Which issue this PR fixes(if applicable):
fixes #1924

Special notes for reviewers:

CCM config changes:

[Global]
auth-url=https://auth.openstack.example.com/v3/
region=REGION1
# new param 'regions' can be specified multiple times
regions=REGION1
regions=REGION2
regions=REGION3

Optionally can be set in cloud.conf

clouds:
  kubernetes:
    auth:
      auth_url: https://auth.openstack.example.com/v3
    region_name: "REGION1"
    regions:
      - REGION1
      - REGION2
      - REGION3

During the initialization process, OCCM checks for the existence of providerID. If providerID does not exist, it defaults to using node.name, as it did previously. Additionally, if the node has the label topology.kubernetes.io/region, OCCM will prioritize using this region as the first one to check. This approach ensures that in the event of a region outage, OCCM can continue to function.

In addition, we can assist CCM in locating the node by providing kubelet parameters:

• --provider-id=openstack:///$InstanceID - InstanceID exists in metadata
• --provider-id=openstack://$REGION/$InstanceID - if you can define the region (by default meta server does not have this information)
• --node-labels=topology.kubernetes.io/region=$REGION set preferred REGION in label, OCCM will then prioritize searching for the node in this specified region

Release note:

NONE

[k8s-ci-robot]

Hi @sergelogvinov. Thanks for your PR.

I'm waiting for a kubernetes member to verify that this patch is reasonable to test. If it is, they should reply with /ok-to-test on its own line. Until that is done, I will not automatically test new commits in this PR, but the usual testing commands by org members will still work. Regular contributors should join the org to skip this step.

Once the patch is verified, the new status will be reflected by the ok-to-test label.

I understand the commands that are listed here.

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[sergelogvinov]

@mdbooth can you take a look on this PR.
Probably I need to add more configuration checks.

Thanks.

[mdbooth]

I haven't taken a deep look into this, but I much prefer this in principal: it only tells cloud-provider things we know to be true. This makes me much more confident that this will continue to work correctly as cloud-provider evolves.

Would you still run multiple CCMs, or switch to a single active CCM?

[mdbooth]

@pierreprinetti how much work is performed when initialising a new service client? Is it local-only, or do we have to go back to keystone?

I might be inclined to intialise this lazily anyway, tbh.

[jichenjc]

similar thought, maybe init them until real usage?

[sergelogvinov]

I've had one issue in proxmox with lazy initialization. The regions cannot exist, and during the rollout of OCCM, it starts without errors. The kubernetes administrator will think that all configuration is correct.

So we can check all regions here and crush if needed. What do you think?

[pierreprinetti]

Apologies for the late response. Building a ProviderClient requires a Keystone roundtrip; building ServiceClients is cheap.

[mdbooth]

I'd probably be a bit more explicit with where we're looking for stuff here. IIUC there are 2 possible places we can get a specific region from:

• providerID
• LabelTopologyRegion

Both may be unset because the node has not yet been adopted by the node-controller.

providerID may not contain a region either because it was set before we became multi-region, or because it was set by kubelet without a region and it's immutable.

But the end result is that either we know the region or we don't. If we know the region we should look only in that region. If we don't know the region we should look everywhere.

How about logic something like:

instanceID, instanceRegion, err := instanceIDFromProviderID(node.Spec.ProviderID)
..err omitted...

if instanceRegion == "" {
  instanceRegion = node.Labels[v1.LabelTopologyRegion]
}

var searchRegions []string
if instanceRegion != "" {
  if !slices.Contains(i.regions, instanceRegion) {
    return ...bad region error...
  }
  searchRegions = []string{instanceRegion}
} else {
  searchRegions = ..all the regions, preferred first...
}

for region := range searchRegions {
  mc := ...
  if instanceID != "" {
    getInstanceByID()
  } else {
    getInstanceByName()
  }
  mc.ObserveRequest()
}

[sergelogvinov]

Thank you, this very very good idea.
I've changed the implementation. But one thought - i cannot trust LabelTopologyRegion, something can change it and node-lifecycle will remove node (for instance on reboot/upgrade event)...

So i can use LabelTopologyRegion only as prefered-region. And check this region first.
Thanks.

[MatthieuFin]

Hi @sergelogvinov
I propose an implementation of multi cloud support for cinder-csi-plugin, which offer multiple openstack clusters support, not only multiple regions, I haven't take look of occm implementation yet, but is it possible to adapt it to support multiple cloud definitions instead of only multiple regions ?

[sergelogvinov]

Hi @sergelogvinov I propose an implementation of multi cloud support for cinder-csi-plugin, which offer multiple openstack clusters support, not only multiple regions, I haven't take look of occm implementation yet, but is it possible to adapt it to support multiple cloud definitions instead of only multiple regions ?

Thank you for this PR, it is very interesting. Can we have a call/chat in slack #provider-openstack (Serge Logvinov)?

[jichenjc]

/ok-to-test

[jichenjc]

/ok-to-test

[jichenjc]

similar thought, maybe init them until real usage?

[sergelogvinov]

I've rebased the PR. all tests passed and i've tested manually too

Can you take a look please @jichenjc @mdbooth
It will be great to merge this change into the upcoming release...

Thanks.

[sergelogvinov]

Is anything else we can do here? @jichenjc @mdbooth @kayrus

We had conversation how we need initialize the openstack clients

	for _, region := range os.regions {
		opt := os.epOpts
		opt.Region = region

		compute[region], err = client.NewComputeV2(os.provider, opt)
		if err != nil {
			klog.Errorf("unable to access compute v2 API : %v", err)
			return nil, false
		}

		network[region], err = client.NewNetworkV2(os.provider, opt)
		if err != nil {
			klog.Errorf("unable to access network v2 API : %v", err)
			return nil, false
		}

It seems to be a similar process to the one we followed in cinder-csi-plugin.
I believe @MatthieuFin and I can introduce multi OpenStack authentication support after this PR.

    [Global]
    auth-url="https://auth.cloud.openstackcluster.region-default.local/v3"
    username="region-default-username"
    password="region-default-password"
    region="default"
    tenant-id="region-default-tenant-id"
    tenant-name="region-default-tenant-name"
    domain-name="Default"
    
    [Global "region-one"]
    auth-url="https://auth.cloud.openstackcluster.region-one.local/v3"
    username="region-one-username"
    password="region-one-password"
    region="one"
    tenant-id="region-one-tenant-id"
    tenant-name="region-one-tenant-name"
    domain-name="Default"

Thanks.

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by:
Once this PR has been reviewed and has the lgtm label, please assign zetaab for approval. For more information see the Kubernetes Code Review Process.

The full list of commands accepted by this bot can be found here.

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[sergelogvinov]

/retest

[sergelogvinov]

/test openstack-cloud-controller-manager-e2e-test

[sergelogvinov]

Thank you @kayrus for refactoring of instance.go (instance-v2).
I’ve moved all my changes to instance.go. And it ready to go.

Whenever you have a moment, could you please take a look at this?

[kayrus]

@sergelogvinov hm, I have doubts about this PR so far. Some raw thoughts:

• region vs regions already cause a confusion, I believe there should be a better way to introduce regions. Does your implementation support multiple keystone URLs? Can we extract regions from the config file?
• add more documentation
• add/adjust documentation on how this correlates with cinder-csi and manila-csi plugins.
• multicloud vs multiregion? do we need both, or does one imply another?
• ideally I'd have a per-cloud/per-region struct member/interface that implements only a single cloud/region logic, but I'm afraid that the upstream https://github.com/kubernetes/cloud-provider doesn't support this. Maybe it makes sense to submit a feature request there? Let me know if you have questions on this particular point.

Sorry, I don't have a multicloud/multiregion setup, and the use case is not really clear for me.

[sergelogvinov]

Thank you @kayrus for reviewing my PR.

I completely agree with your point about hybrid/multi-cluster setups (bare metal + OpenStack + AWS + GCP, etc.). This should first be implemented in https://github.com/kubernetes/cloud-provider first. I hope it will be added someday.

Our case is a bit different:

Imagine an OpenStack setup with only one Keystone endpoint but multiple regions for services like Nova, Neutron, Cinder, and Glance. Each region has one available zone called "nova" (the default installation). I think this type of setup is easier to manage and upgrade. I've seen similar setups used by many cloud providers.

So, in this case, it’s not fully separated OpenStack clusters. It looks more like one region with many av-zones, similar to how well-known cloud providers organize their systems.

That’s why this PR supports only one Keystone endpoint. Using this endpoint, we can get Nova/Neutron endpoints for each region.

This feature will be introduced as an alpha feature and can be enabled using a CLI flag. Other components like the load balancer, cinder-csi and manila-csi will also need updates. However, we need to focus on supporting the cloud-node and cloud-node-lifecycle controllers first.

PS. I think I forgot to check backward compatibility with non regional OpenStack setups. I will check this soon.
Thanks.