update affinity assistant creation implementation

Before this commit, the affinity assistant was created in the beginning of
the pipleineRun. And the same affinity assistant was relied on for the
entire lifecycle of a PR. Now, there could be a case when the node
on which affinity assistant pod is created goes down. In this case,
the rest of the pipelineRun is stuck and cannot run to the completition
since the affinity assistant (StatefulSet) tries to schedule rest
of the pods (taskRuns) on the same node but that node is cordoned or not
scheduling anything new.

This commit always makes an attempt to create Affinity Assistant (StatefulSet)
in case it does not exist. If it exist, the controller checks if the node on
which Affinity Assistant pod is created is healthy to schedule subsequent
pods. If not, the controller deletes Affinity Assistant pod so that
StatefulSet can upscale the replicas (set to 1) on other node in the cluster.

config/200-clusterrole.yaml

@@ -25,6 +25,10 @@ rules:
     # Controller needs to watch Pods created by TaskRuns to see them progress.
     resources: ["pods"]
     verbs: ["list", "watch"]
+  - apiGroups: [ "" ]
+    # Controller needs to watch nodes for their health over the course of a single run
+    resources: [ "nodes" ]
+    verbs: [ "get", "list", "watch" ]
     # Controller needs cluster access to all of the CRDs that it is responsible for
     # managing.
   - apiGroups: ["tekton.dev"]

pkg/reconciler/pipelinerun/affinity_assistant.go

@@ -20,7 +20,6 @@ import (
 	"context"
 	"crypto/sha256"
 	"fmt"
-
 	"github.com/tektoncd/pipeline/pkg/apis/config"
 	"github.com/tektoncd/pipeline/pkg/apis/pipeline"
 	"github.com/tektoncd/pipeline/pkg/apis/pipeline/pod"
@@ -56,7 +55,7 @@ func (c *Reconciler) createAffinityAssistants(ctx context.Context, wb []v1beta1.
 	for _, w := range wb {
 		if w.PersistentVolumeClaim != nil || w.VolumeClaimTemplate != nil {
 			affinityAssistantName := getAffinityAssistantName(w.Name, pr.Name)
-			_, err := c.KubeClientSet.AppsV1().StatefulSets(namespace).Get(ctx, affinityAssistantName, metav1.GetOptions{})
+			a, err := c.KubeClientSet.AppsV1().StatefulSets(namespace).Get(ctx, affinityAssistantName, metav1.GetOptions{})
 			claimName := getClaimName(w, *kmeta.NewControllerRef(pr))
 			switch {
 			case apierrors.IsNotFound(err):
@@ -68,6 +67,37 @@ func (c *Reconciler) createAffinityAssistants(ctx context.Context, wb []v1beta1.
 				if err == nil {
 					logger.Infof("Created StatefulSet %s in namespace %s", affinityAssistantName, namespace)
 				}
+			case err == nil && a != nil && a.Status.ReadyReplicas == 1:
+				// the node which hosts the affinity assistant pod is unschedulable or cordoned for some reason,
+				// delete the affinity assistant pod such that a StatefulSet can recreate the same pod on a different node
+				// get the list of pods associated with a given StatefulSet based on the following label:
+				// app.kubernetes.io/instance=affinity-assistant-641c8a0038
+				pods, err := c.KubeClientSet.CoreV1().Pods("").List(ctx, metav1.ListOptions{
+					LabelSelector: workspace.LabelInstance + "=" + a.Name,
+				})
+				if err != nil {
+					return fmt.Errorf("could not list affinity assistant pods for StatefulSet %s: %s", a.Name, err)
+				}
+				// there should be only one pod created for a given StatefulSet since we have set the replicas to 1
+				if len(pods.Items) > 1 {
+					return fmt.Errorf("the affinity assistant only creates one replica but we found \"%d\" pods", len(pods.Items))
+				}
+				// iterate over the list of pods
+				// at most there can be only one pod but could be zero due to some network delay
+				for _, pod := range pods.Items {
+					// get the node based on the affinity assistant pod - pod.Spec.nodeName
+					n, err := c.KubeClientSet.CoreV1().Nodes().Get(ctx, pod.Spec.NodeName, metav1.GetOptions{})
+					if err != nil {
+						return fmt.Errorf("could not get the node \"%s\" on which affinity assistant pod is created, err: %s", pod.Spec.NodeName, err)
+					}
+					// delete the affinity assistant pod if the node is cordoned or unschedulable anymore
+					if n.Spec.Unschedulable {
+						err = c.KubeClientSet.CoreV1().Pods(pod.Namespace).Delete(ctx, pod.Name, metav1.DeleteOptions{})
+						if err != nil {
+							return fmt.Errorf("error deleting affinity assistant pod %s in ns %s: %s", pod.Name, pod.Namespace, err)
+						}
+					}
+				}
 			case err != nil:
 				errs = append(errs, fmt.Errorf("failed to retrieve StatefulSet %s: %w", affinityAssistantName, err))
 			}

pkg/reconciler/pipelinerun/pipelinerun.go

@@ -607,16 +607,18 @@ func (c *Reconciler) reconcile(ctx context.Context, pr *v1beta1.PipelineRun, get
 				return controller.NewPermanentError(err)
 			}
 		}
+	}
 
-		if !c.isAffinityAssistantDisabled(ctx) {
-			// create Affinity Assistant (StatefulSet) so that taskRun pods that share workspace PVC achieve Node Affinity
-			if err = c.createAffinityAssistants(ctx, pr.Spec.Workspaces, pr, pr.Namespace); err != nil {
-				logger.Errorf("Failed to create affinity assistant StatefulSet for PipelineRun %s: %v", pr.Name, err)
-				pr.Status.MarkFailed(ReasonCouldntCreateAffinityAssistantStatefulSet,
-					"Failed to create StatefulSet for PipelineRun %s/%s correctly: %s",
-					pr.Namespace, pr.Name, err)
-				return controller.NewPermanentError(err)
-			}
+	// Make an attempt to create Affinity Assistant if it does not exist
+	// if the Affinity Assistant already exists, handle the possibility of assigned node going down by deleting the pod
+	if !c.isAffinityAssistantDisabled(ctx) {
+		// create Affinity Assistant (StatefulSet) so that taskRun pods that share workspace PVC achieve Node Affinity
+		if err = c.createAffinityAssistants(ctx, pr.Spec.Workspaces, pr, pr.Namespace); err != nil {
+			logger.Errorf("Failed to create affinity assistant StatefulSet for PipelineRun %s: %v", pr.Name, err)
+			pr.Status.MarkFailed(ReasonCouldntCreateAffinityAssistantStatefulSet,
+				"Failed to create StatefulSet for PipelineRun %s/%s correctly: %s",
+				pr.Namespace, pr.Name, err)
+			return controller.NewPermanentError(err)
 		}
 	}