fix some orphan pods cleaner bugs

cmd/controller-manager/main.go

@@ -141,8 +141,23 @@ func main() {
 	bsController := backupschedule.NewController(kubeCli, cli, informerFactory, kubeInformerFactory)
 	controllerCtx, cancel := context.WithCancel(context.Background())
 	defer cancel()
-	go informerFactory.Start(controllerCtx.Done())
-	go kubeInformerFactory.Start(controllerCtx.Done())
+
+	// Start informer factories after all controller are initialized.
+	informerFactory.Start(controllerCtx.Done())
+	kubeInformerFactory.Start(controllerCtx.Done())
+
+	// Wait for all started informers' cache were synced.
+	for v, synced := range informerFactory.WaitForCacheSync(wait.NeverStop) {
+		if !synced {
+			glog.Fatalf("error syncing informer for %v", v)
+		}
+	}
+	for v, synced := range kubeInformerFactory.WaitForCacheSync(wait.NeverStop) {
+		if !synced {
+			glog.Fatalf("error syncing informer for %v", v)
+		}
+	}
+	glog.Infof("cache of informer factories sync successfully")
 
 	onStarted := func(ctx context.Context) {
 		go wait.Forever(func() { backupController.Run(workers, ctx.Done()) }, waitDuration)

pkg/controller/backup/backup_controller.go

@@ -118,10 +118,6 @@ func (bkc *Controller) Run(workers int, stopCh <-chan struct{}) {
 	glog.Info("Starting backup controller")
 	defer glog.Info("Shutting down backup controller")
 
-	if !cache.WaitForCacheSync(stopCh, bkc.backupListerSynced) {
-		return
-	}
-
 	for i := 0; i < workers; i++ {
 		go wait.Until(bkc.worker, time.Second, stopCh)
 	}

pkg/controller/backupschedule/backup_schedule_controller.go

@@ -114,10 +114,6 @@ func (bsc *Controller) Run(workers int, stopCh <-chan struct{}) {
 	glog.Info("Starting backup schedule controller")
 	defer glog.Info("Shutting down backup schedule controller")
 
-	if !cache.WaitForCacheSync(stopCh, bsc.bsListerSynced) {
-		return
-	}
-
 	for i := 0; i < workers; i++ {
 		go wait.Until(bsc.worker, time.Second, stopCh)
 	}

pkg/controller/pod_control.go

@@ -23,6 +23,7 @@ import (
 	"github.com/pingcap/tidb-operator/pkg/label"
 	"github.com/pingcap/tidb-operator/pkg/pdapi"
 	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	utilruntime "k8s.io/apimachinery/pkg/util/runtime"
 	coreinformers "k8s.io/client-go/informers/core/v1"
 	"k8s.io/client-go/kubernetes"
@@ -192,7 +193,9 @@ func (rpc *realPodControl) DeletePod(tc *v1alpha1.TidbCluster, pod *corev1.Pod)
 	ns := tc.GetNamespace()
 	tcName := tc.GetName()
 	podName := pod.GetName()
-	err := rpc.kubeCli.CoreV1().Pods(ns).Delete(podName, nil)
+	preconditions := metav1.Preconditions{UID: &pod.UID}
+	deleteOptions := metav1.DeleteOptions{Preconditions: &preconditions}
+	err := rpc.kubeCli.CoreV1().Pods(ns).Delete(podName, &deleteOptions)
 	if err != nil {
 		glog.Errorf("failed to delete Pod: [%s/%s], TidbCluster: %s, %v", ns, podName, tcName, err)
 	} else {

pkg/controller/restore/restore_controller.go

@@ -117,10 +117,6 @@ func (rsc *Controller) Run(workers int, stopCh <-chan struct{}) {
 	glog.Info("Starting restore controller")
 	defer glog.Info("Shutting down restore controller")
 
-	if !cache.WaitForCacheSync(stopCh, rsc.restoreListerSynced) {
-		return
-	}
-
 	for i := 0; i < workers; i++ {
 		go wait.Until(rsc.worker, time.Second, stopCh)
 	}

pkg/controller/tidbcluster/tidb_cluster_controller.go

@@ -170,6 +170,7 @@ func NewController(
 				podInformer.Lister(),
 				podControl,
 				pvcInformer.Lister(),
+				kubeCli,
 			),
 			recorder,
 		),
@@ -210,10 +211,6 @@ func (tcc *Controller) Run(workers int, stopCh <-chan struct{}) {
 	glog.Info("Starting tidbcluster controller")
 	defer glog.Info("Shutting down tidbcluster controller")
 
-	if !cache.WaitForCacheSync(stopCh, tcc.tcListerSynced, tcc.setListerSynced) {
-		return
-	}
-
 	for i := 0; i < workers; i++ {
 		go wait.Until(tcc.worker, time.Second, stopCh)
 	}

pkg/manager/member/orphan_pods_cleaner.go

@@ -18,17 +18,35 @@ import (
 	"github.com/pingcap/tidb-operator/pkg/apis/pingcap.com/v1alpha1"
 	"github.com/pingcap/tidb-operator/pkg/controller"
 	"github.com/pingcap/tidb-operator/pkg/label"
+	v1 "k8s.io/api/core/v1"
 	"k8s.io/apimachinery/pkg/api/errors"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
 	corelisters "k8s.io/client-go/listers/core/v1"
 )
 
 const (
-	skipReasonOrphanPodsCleanerIsNotPDOrTiKV  = "orphan pods cleaner: member type is not pd or tikv"
-	skipReasonOrphanPodsCleanerPVCNameIsEmpty = "orphan pods cleaner: pvcName is empty"
-	skipReasonOrphanPodsCleanerPVCIsFound     = "orphan pods cleaner: pvc is found"
+	skipReasonOrphanPodsCleanerIsNotPDOrTiKV   = "orphan pods cleaner: member type is not pd or tikv"
+	skipReasonOrphanPodsCleanerPVCNameIsEmpty  = "orphan pods cleaner: pvcName is empty"
+	skipReasonOrphanPodsCleanerPVCIsFound      = "orphan pods cleaner: pvc is found"
+	skipReasonOrphanPodsCleanerPodIsNotPending = "orphan pods cleaner: pod is not pending"
+	skipReasonOrphanPodsCleanerPodIsNotFound   = "orphan pods cleaner: pod does not exist anymore"
+	skipReasonOrphanPodsCleanerPodChanged      = "orphan pods cleaner: pod changed before deletion"
 )
 
 // OrphanPodsCleaner implements the logic for cleaning the orphan pods(has no pvc)
+//
+// In scaling out and failover, we will try to delete the old PVC to prevent it
+// from being used by the new pod. However, the PVC might not be deleted
+// immediately in the apiserver because of finalizers (e.g.
+// kubernetes.io/pvc-protection) and the statefulset controller may not have
+// received PVC delete event when it tries to create the new replica and the
+// new pod will be pending forever because no PVC to use. We need to clean
+// these orphan pods and let the statefulset controller to create PVC(s) for
+// them.
+//
+// https://github.com/kubernetes/kubernetes/blob/84fe3db5cf58bf0fc8ff792b885465ceaf70a435/pkg/controller/statefulset/stateful_pod_control.go#L175-L199
+//
 type OrphanPodsCleaner interface {
 	Clean(*v1alpha1.TidbCluster) (map[string]string, error)
 }
@@ -37,13 +55,15 @@ type orphanPodsCleaner struct {
 	podLister  corelisters.PodLister
 	podControl controller.PodControlInterface
 	pvcLister  corelisters.PersistentVolumeClaimLister
+	kubeCli    kubernetes.Interface
 }
 
 // NewOrphanPodsCleaner returns a OrphanPodsCleaner
 func NewOrphanPodsCleaner(podLister corelisters.PodLister,
 	podControl controller.PodControlInterface,
-	pvcLister corelisters.PersistentVolumeClaimLister) OrphanPodsCleaner {
-	return &orphanPodsCleaner{podLister, podControl, pvcLister}
+	pvcLister corelisters.PersistentVolumeClaimLister,
+	kubeCli kubernetes.Interface) OrphanPodsCleaner {
+	return &orphanPodsCleaner{podLister, podControl, pvcLister, kubeCli}
 }
 
 func (opc *orphanPodsCleaner) Clean(tc *v1alpha1.TidbCluster) (map[string]string, error) {
@@ -68,6 +88,12 @@ func (opc *orphanPodsCleaner) Clean(tc *v1alpha1.TidbCluster) (map[string]string
 			continue
 		}
 
+		if pod.Status.Phase != v1.PodPending {
+			skipReason[podName] = skipReasonOrphanPodsCleanerPodIsNotPending
+			continue
+		}
+
+		// TODO support multiple pvcs case?
 		var pvcName string
 		for _, vol := range pod.Spec.Volumes {
 			if vol.PersistentVolumeClaim != nil {
@@ -80,7 +106,9 @@ func (opc *orphanPodsCleaner) Clean(tc *v1alpha1.TidbCluster) (map[string]string
 			continue
 		}
 
-		_, err := opc.pvcLister.PersistentVolumeClaims(ns).Get(pvcName)
+		var err error
+		// check informer cache
+		_, err = opc.pvcLister.PersistentVolumeClaims(ns).Get(pvcName)
 		if err == nil {
 			skipReason[podName] = skipReasonOrphanPodsCleanerPVCIsFound
 			continue
@@ -89,6 +117,33 @@ func (opc *orphanPodsCleaner) Clean(tc *v1alpha1.TidbCluster) (map[string]string
 			return skipReason, err
 		}
 
+		// if PVC not found in cache, re-check from apiserver directly to make sure the PVC really not exist
+		_, err = opc.kubeCli.CoreV1().PersistentVolumeClaims(ns).Get(pvcName, metav1.GetOptions{})
+		if err == nil {
+			skipReason[podName] = skipReasonOrphanPodsCleanerPVCIsFound
+			continue
+		}
+		if !errors.IsNotFound(err) {
+			return skipReason, err
+		}
+
+		// if the PVC is not found in apiserver (also informer cache) and the
+		// phase of the Pod is Pending, delete it and let the stateful
+		// controller to create the pod and its PVC(s) again
+		apiPod, err := opc.kubeCli.CoreV1().Pods(ns).Get(podName, metav1.GetOptions{})
+		if errors.IsNotFound(err) {
+			skipReason[podName] = skipReasonOrphanPodsCleanerPodIsNotFound
+			continue
+		}
+		if err != nil {
+			return skipReason, err
+		}
+		// try our best to avoid deleting wrong object in apiserver
+		// TODO upgrade to use deleteOption.Preconditions.ResourceVersion in client-go 1.14+
+		if apiPod.UID != pod.UID || apiPod.ResourceVersion != pod.ResourceVersion {
+			skipReason[podName] = skipReasonOrphanPodsCleanerPodChanged
+			continue
+		}
 		err = opc.podControl.DeletePod(tc, pod)
 		if err != nil {
 			glog.Errorf("orphan pods cleaner: failed to clean orphan pod: %s/%s, %v", ns, podName, err)