perf: Unregister the topology domain when failing NodeClaim creation (#…

…1819)

pkg/controllers/provisioning/scheduling/nodeclaim.go

@@ -39,6 +39,7 @@ type NodeClaim struct {
 	topology        *Topology
 	hostPortUsage   *scheduling.HostPortUsage
 	daemonResources v1.ResourceList
+	hostname        string
 }
 
 var nodeID int64
@@ -59,6 +60,7 @@ func NewNodeClaim(nodeClaimTemplate *NodeClaimTemplate, topology *Topology, daem
 		hostPortUsage:     scheduling.NewHostPortUsage(),
 		topology:          topology,
 		daemonResources:   daemonResources,
+		hostname:          hostname,
 	}
 }
 
@@ -119,6 +121,10 @@ func (n *NodeClaim) Add(pod *v1.Pod) error {
 	return nil
 }
 
+func (n *NodeClaim) Destroy() {
+	n.topology.Unregister(v1.LabelHostname, n.hostname)
+}
+
 // FinalizeScheduling is called once all scheduling has completed and allows the node to perform any cleanup
 // necessary before its requirements are used for instance launching
 func (n *NodeClaim) FinalizeScheduling() {

pkg/controllers/provisioning/scheduling/scheduler.go

@@ -289,6 +289,7 @@ func (s *Scheduler) add(ctx context.Context, pod *corev1.Pod) error {
 		}
 		nodeClaim := NewNodeClaim(nodeClaimTemplate, s.topology, s.daemonOverhead[nodeClaimTemplate], instanceTypes)
 		if err := nodeClaim.Add(pod); err != nil {
+			nodeClaim.Destroy() // Ensure we cleanup any changes that we made while mocking out a NodeClaim
 			errs = multierr.Append(errs, fmt.Errorf("incompatible with nodepool %q, daemonset overhead=%s, %w",
 				nodeClaimTemplate.NodePoolName,
 				resources.String(s.daemonOverhead[nodeClaimTemplate]),

pkg/controllers/provisioning/scheduling/topology.go

@@ -201,6 +201,20 @@ func (t *Topology) Register(topologyKey string, domain string) {
 	}
 }
 
+// Unregister is used to unregister a domain as available across topologies for the given topology key.
+func (t *Topology) Unregister(topologyKey string, domain string) {
+	for _, topology := range t.topologies {
+		if topology.Key == topologyKey {
+			topology.Unregister(domain)
+		}
+	}
+	for _, topology := range t.inverseTopologies {
+		if topology.Key == topologyKey {
+			topology.Unregister(domain)
+		}
+	}
+}
+
 // updateInverseAffinities is used to identify pods with anti-affinity terms so we can track those topologies.  We
 // have to look at every pod in the cluster as there is no way to query for a pod with anti-affinity terms.
 func (t *Topology) updateInverseAffinities(ctx context.Context) error {

pkg/controllers/provisioning/scheduling/topologygroup.go

@@ -128,6 +128,14 @@ func (t *TopologyGroup) Register(domains ...string) {
 	}
 }
 
+// Unregister removes the topology group from being aware of the domain
+func (t *TopologyGroup) Unregister(domains ...string) {
+	for _, domain := range domains {
+		delete(t.domains, domain)
+		t.emptyDomains.Delete(domain)
+	}
+}
+
 func (t *TopologyGroup) AddOwner(key types.UID) {
 	t.owners[key] = struct{}{}
 }