Merge pull request #3990 from chaosi-zju/dev

Response to priority change of ClusterPropagationPolicy

pkg/detector/detector.go

@@ -894,7 +894,7 @@ func (d *ResourceDetector) OnClusterPropagationPolicyAdd(obj interface{}) {
 }
 
 // OnClusterPropagationPolicyUpdate handles object update event and push the object to queue.
-func (d *ResourceDetector) OnClusterPropagationPolicyUpdate(_, newObj interface{}) {
+func (d *ResourceDetector) OnClusterPropagationPolicyUpdate(oldObj, newObj interface{}) {
 	key, err := ClusterWideKeyFunc(newObj)
 	if err != nil {
 		return
@@ -902,6 +902,52 @@ func (d *ResourceDetector) OnClusterPropagationPolicyUpdate(_, newObj interface{
 
 	klog.V(2).Infof("Update ClusterPropagationPolicy(%s)", key)
 	d.clusterPolicyReconcileWorker.Add(key)
+
+	// Temporary solution of corner case: After the priority(.spec.priority) of
+	// ClusterPropagationPolicy changed from high priority (e.g. 5) to low priority(e.g. 3),
+	// we should try to check if there is a ClusterPropagationPolicy(e.g. with priority 4)
+	// could preempt the targeted resources.
+	//
+	// Recognized limitations of the temporary solution are:
+	// - Too much logical processed in an event handler function will slow down
+	//   the overall reconcile speed.
+	// - If there is an error raised during the process, the event will be lost
+	//   and no second chance to retry.
+	//
+	// The idea of the long-term solution, perhaps ClusterPropagationPolicy could have
+	// a status, in that case we can record the observed priority(.status.observedPriority)
+	// which can be used to detect priority changes during reconcile logic.
+	if features.FeatureGate.Enabled(features.PolicyPreemption) {
+		var unstructuredOldObj *unstructured.Unstructured
+		var unstructuredNewObj *unstructured.Unstructured
+
+		unstructuredOldObj, err = helper.ToUnstructured(oldObj)
+		if err != nil {
+			klog.Errorf("Failed to transform oldObj, error: %v", err)
+			return
+		}
+		unstructuredNewObj, err = helper.ToUnstructured(newObj)
+		if err != nil {
+			klog.Errorf("Failed to transform newObj, error: %v", err)
+			return
+		}
+
+		var oldPolicy policyv1alpha1.ClusterPropagationPolicy
+		var newPolicy policyv1alpha1.ClusterPropagationPolicy
+
+		if err = helper.ConvertToTypedObject(unstructuredOldObj, &oldPolicy); err != nil {
+			klog.Errorf("Failed to convert typed ClusterPropagationPolicy(%s/%s): %v", unstructuredOldObj.GetNamespace(), unstructuredOldObj.GetName(), err)
+			return
+		}
+		if err = helper.ConvertToTypedObject(unstructuredNewObj, &newPolicy); err != nil {
+			klog.Errorf("Failed to convert typed ClusterPropagationPolicy(%s/%s): %v", newPolicy.GetNamespace(), newPolicy.GetName(), err)
+			return
+		}
+
+		if newPolicy.ExplicitPriority() < oldPolicy.ExplicitPriority() {
+			d.HandleDeprioritizedClusterPropagationPolicy(oldPolicy, newPolicy)
+		}
+	}
 }
 
 // OnClusterPropagationPolicyDelete handles object delete event and push the object to queue.

pkg/detector/preemption.go

@@ -271,3 +271,49 @@ func (d *ResourceDetector) HandleDeprioritizedPropagationPolicy(oldPolicy policy
 		}
 	}
 }
+
+// HandleDeprioritizedClusterPropagationPolicy responses to priority change of a ClusterPropagationPolicy,
+// if the change is from high priority (e.g. 5) to low priority(e.g. 3), it will
+// check if there is another ClusterPropagationPolicy could preempt the targeted resource,
+// and put the ClusterPropagationPolicy in the queue to trigger preemption.
+func (d *ResourceDetector) HandleDeprioritizedClusterPropagationPolicy(oldPolicy policyv1alpha1.ClusterPropagationPolicy, newPolicy policyv1alpha1.ClusterPropagationPolicy) {
+	klog.Infof("ClusterPropagationPolicy(%s/%s) priority changed from %d to %d",
+		newPolicy.GetNamespace(), newPolicy.GetName(), *oldPolicy.Spec.Priority, *newPolicy.Spec.Priority)
+
+	policies, err := d.clusterPropagationPolicyLister.ByNamespace(newPolicy.GetNamespace()).List(labels.Everything())
+	if err != nil {
+		klog.Errorf("Failed to list ClusterPropagationPolicy from namespace: %s, error: %v", newPolicy.GetNamespace(), err)
+		return
+	}
+
+	// TODO(@RainbowMango): Should sort the listed policies to ensure the
+	// higher priority ClusterPropagationPolicy be process first to avoid possible
+	// multiple preemption.
+
+	for i := range policies {
+		var potentialPolicy policyv1alpha1.ClusterPropagationPolicy
+		if err = helper.ConvertToTypedObject(policies[i], &potentialPolicy); err != nil {
+			klog.Errorf("Failed to convert typed ClusterPropagationPolicy: %v", err)
+			continue
+		}
+		// Re-queue the polies that enables preemption and with the priority
+		// in range (new priority, old priority).
+		// For the polices with higher priority than old priority, it can
+		// perform preempt automatically and don't need to re-queue here.
+		// For the polices with lower priority than new priority, it can't
+		// perform preempt as insufficient priority.
+		if potentialPolicy.Spec.Priority != nil &&
+			potentialPolicy.Spec.Preemption == policyv1alpha1.PreemptAlways &&
+			potentialPolicy.ExplicitPriority() > newPolicy.ExplicitPriority() &&
+			potentialPolicy.ExplicitPriority() < oldPolicy.ExplicitPriority() {
+			var potentialKey util.QueueKey
+			potentialKey, err = ClusterWideKeyFunc(&potentialPolicy)
+			if err != nil {
+				return
+			}
+			klog.Infof("Enqueuing ClusterPropagationPolicy(%s/%s) in case of ClusterPropagationPolicy(%s/%s) priority changes",
+				potentialPolicy.GetNamespace(), potentialPolicy.GetName(), newPolicy.GetNamespace(), newPolicy.GetName())
+			d.clusterPolicyReconcileWorker.Add(potentialKey)
+		}
+	}
+}