Kube-vip depends on `nodename == hostname` for service election, which easily breaks

[d-uzlov]

Kube-vip depends on `nodename == hostname` for service election, which easily breaks

[d-uzlov]

Currently kube-vip attempts to handle FQDN hostname cases. For example, here:

kube-vip/pkg/manager/watch_endpoints.go

Lines 96 to 120 in 0cedf6a

	// 1. Compare the hostname on the endpoint to the hostname
	// 2. Compare the nodename on the endpoint to the hostname
	// 3. Drop the FQDN to a shortname and compare to the nodename on the endpoint
	// 1. Compare the Hostname first (should be FQDN)
	log.Debugf("[%s] processing endpoint [%s]", ep.label, ep.endpoints.Subsets[subset].Addresses[address].IP)
	if id == ep.endpoints.Subsets[subset].Addresses[address].Hostname {
	log.Debugf("[%s] found local endpoint - address: %s, hostname: %s",
	ep.label, ep.endpoints.Subsets[subset].Addresses[address].IP, ep.endpoints.Subsets[subset].Addresses[address].Hostname)
	localEndpoints = append(localEndpoints, ep.endpoints.Subsets[subset].Addresses[address].IP)
	} else {
	// 2. Compare the Nodename (from testing could be FQDN or short)
	if ep.endpoints.Subsets[subset].Addresses[address].NodeName != nil {
	if id == *ep.endpoints.Subsets[subset].Addresses[address].NodeName {
	log.Debugf("[%s] found local endpoint - address: %s, hostname: %s, node: %s",
	ep.label, ep.endpoints.Subsets[subset].Addresses[address].IP, ep.endpoints.Subsets[subset].Addresses[address].Hostname,
	*ep.endpoints.Subsets[subset].Addresses[address].NodeName)
	localEndpoints = append(localEndpoints, ep.endpoints.Subsets[subset].Addresses[address].IP)
	} else if shortnameErr == nil && shortname == *ep.endpoints.Subsets[subset].Addresses[address].NodeName {
	log.Debugf("[%s] found local endpoint - address: %s, shortname: %s, node: %s",
	ep.label, ep.endpoints.Subsets[subset].Addresses[address].IP, shortname, *ep.endpoints.Subsets[subset].Addresses[address].NodeName)
	localEndpoints = append(localEndpoints, ep.endpoints.Subsets[subset].Addresses[address].IP)
	}
	}
	}

kube-vip/pkg/manager/watch_endpointslices.go

Lines 76 to 103 in 0cedf6a

	// 1. Compare the hostname on the endpoint to the hostname
	// 2. Compare the nodename on the endpoint to the hostname
	// 3. Drop the FQDN to a shortname and compare to the nodename on the endpoint
	// 1. Compare the Hostname first (should be FQDN)
	log.Debugf("[%s] processing endpoint [%s]", ep.label, ep.endpoints.Endpoints[i].Addresses[j])
	if ep.endpoints.Endpoints[i].Hostname != nil && id == *ep.endpoints.Endpoints[i].Hostname {
	if *ep.endpoints.Endpoints[i].Conditions.Serving {
	log.Debugf("[%s] found endpoint - address: %s, hostname: %s", ep.label, ep.endpoints.Endpoints[i].Addresses[j], *ep.endpoints.Endpoints[i].Hostname)
	localEndpoints = append(localEndpoints, ep.endpoints.Endpoints[i].Addresses[j])
	}
	} else {
	// 2. Compare the Nodename (from testing could be FQDN or short)
	if ep.endpoints.Endpoints[i].NodeName != nil {
	if id == *ep.endpoints.Endpoints[i].NodeName && *ep.endpoints.Endpoints[i].Conditions.Serving {
	if ep.endpoints.Endpoints[i].Hostname != nil {
	log.Debugf("[%s] found endpoint - address: %s, hostname: %s, node: %s", ep.label, ep.endpoints.Endpoints[i].Addresses[j], *ep.endpoints.Endpoints[i].Hostname, *ep.endpoints.Endpoints[i].NodeName)
	} else {
	log.Debugf("[%s] found endpoint - address: %s, node: %s", ep.label, ep.endpoints.Endpoints[i].Addresses[j], *ep.endpoints.Endpoints[i].NodeName)
	}
	localEndpoints = append(localEndpoints, ep.endpoints.Endpoints[i].Addresses[j])
	// 3. Compare to shortname
	} else if shortnameErr != nil && shortname == *ep.endpoints.Endpoints[i].NodeName && *ep.endpoints.Endpoints[i].Conditions.Serving {
	log.Debugf("[%s] found endpoint - address: %s, shortname: %s, node: %s", ep.label, ep.endpoints.Endpoints[i].Addresses[j], shortname, *ep.endpoints.Endpoints[i].NodeName)
	localEndpoints = append(localEndpoints, ep.endpoints.Endpoints[i].Addresses[j])
	}
	}
	}

But it still doesn't work because in other places it doesn't:

1. Here id goes straight into watchEndpoint, which could handle FQDN names but it doesn't because id = os.Hostname() is already a short name)
   

   kube-vip/pkg/manager/watch_services.go

   Line 52 in 0cedf6a

   id, err := os.Hostname()

   
   

   kube-vip/pkg/manager/watch_services.go

   Line 177 in 0cedf6a

   if err = sm.watchEndpoint(activeServiceLoadBalancer[string(svc.UID)], id, svc, &wg, provider); err != nil {

2. Here hostname := os.Hostname() is a short name, but kubernetes.io/hostname can be set to either FQDN or an arbitrary custom string:
   

   kube-vip/pkg/manager/watch_annotations.go

   Line 31 in 0cedf6a

   hostname, err := os.Hostname()

   
   

   kube-vip/pkg/manager/watch_annotations.go

   Line 36 in 0cedf6a

   labelSelector := metav1.LabelSelector{MatchLabels: map[string]string{"kubernetes.io/hostname": hostname}}

[d-uzlov]

Now that v0.8.0 is released I think we can close this.