Advertise the readiness of the driver on each node. (#72)

* Advertise the readiness of the driver on each node.

Set the label gcs.csi.ofek.dev/driver-ready={true,false}
on each node running a controller

* upgrade driver-ready set failure to warns

* Document how gcs.csi.ofek.dev/driver-ready can be used.

* documentation reword

* According to JSON patch RFC, add should be used here, not replace even if k8s implementation seems forgiveful.

* Honnor driver-name flag when setting the driver-ready label on nodes.

deploy/base/rbac.yaml

@@ -16,7 +16,7 @@ rules:
   verbs: ["delete"]
 - apiGroups: [""]
   resources: ["nodes"]
-  verbs: ["get", "list", "update"]
+  verbs: ["get", "list", "update", "patch"]
 - apiGroups: [""]
   resources: ["namespaces"]
   verbs: ["get", "list"]

docs/troubleshooting.md

@@ -1,5 +1,47 @@
 # Troubleshooting
+## Warnings events from pods scheduled on newly started nodes when mounting csi-gcs volumes
 
------
+Warnings, like the one below, can be seen from pods scheduled on newly started nodes.
 
-## TODO
+```
+MountVolume.MountDevice failed for volume "xxxx" : kubernetes.io/csi: attacher.MountDevice failed to create newCsiDriverClient: driver name gcs.csi.ofek.dev not found in the list of registered CSI drivers
+```
+
+Those warnings are temporary and reflect the csi-gcs driver is still starting. Kubernetes will retry until the csi-gcs driver is ready.
+
+It's possible to avoid those warnings by adding a node selector or affinity using the node label `gcs.csi.ofek.dev/driver-ready=true`.
+
+> Adding such node selector or affinity will trade the time spend waiting for volume mounting retries with time waiting for scheduling.
+
+> The exact label added is `<driver name>/driver-ready`, by default `<driver name>` is `gcs.csi.ofek.dev`
+
+```
+apiVersion: v1
+kind: Pod
+metadata:
+    name: pod-mount-csi-gcs-volume
+spec:
+  // ...
+  nodeSelector:
+    gcs.csi.ofek.dev/driver-ready: "true"
+```
+
+```
+apiVersion: v1
+kind: Pod
+metadata:
+    name: pod-mount-csi-gcs-volume
+spec:
+  // ...
+  affinity:
+    nodeAffinity:
+      requiredDuringSchedulingIgnoredDuringExecution:
+        nodeSelectorTerms:
+        - matchExpressions:
+          - key: gcs.csi.ofek.dev/driver-ready
+            operator: In
+            values:
+            - "true"
+```
+
+You can also add an admission mutating webhook to automatically inject such node selector or affinity in all pods mounting csi-gcs volumes.

pkg/driver/driver.go

@@ -38,6 +38,11 @@ func NewGCSDriver(name, node, endpoint string, version string, deleteOrphanedPod
 }
 
 func (d *GCSDriver) Run() error {
+	// set the driver-ready label to false at the beginning to handle edge-case where the controller didn't terminated gracefully
+	if err := util.SetDriverReadyLabel(d.name, d.nodeName, false); err != nil {
+		klog.Warningf("Unable to set driver-ready=false label on the node, error: %v", err)
+	}
+
 	if len(d.mountPoint) == 0 {
 		return errors.New("--bucket-mount-path is required")
 	}
@@ -75,11 +80,17 @@ func (d *GCSDriver) Run() error {
 	csi.RegisterIdentityServer(d.server, d)
 	csi.RegisterNodeServer(d.server, d)
 	csi.RegisterControllerServer(d.server, d)
+	if err = util.SetDriverReadyLabel(d.name, d.nodeName, true); err != nil {
+		klog.Warningf("unable to set driver-ready=true label on the node, error: %v", err)
+	}
 	return d.server.Serve(listener)
 }
 
 func (d *GCSDriver) stop() {
 	d.server.Stop()
+	if err := util.SetDriverReadyLabel(d.name, d.nodeName, false); err != nil {
+		klog.Warningf("Unable to set driver-ready=false label on the node, error: %v", err)
+	}
 	klog.V(1).Info("CSI driver stopped")
 }
 
@@ -93,7 +104,7 @@ func (d *GCSDriver) RunPodCleanup() (err error) {
 		// Killing Pod because its Volume is no longer mounted
 		err = util.DeletePod(publishedVolume.Spec.Pod.Namespace, publishedVolume.Spec.Pod.Name)
 		if err == nil {
-			klog.V(4).Infof("Deleted Pod %s/%s bacause its volume was no longer mounted", publishedVolume.Spec.Pod.Namespace, publishedVolume.Spec.Pod.Name)
+			klog.V(4).Infof("Deleted Pod %s/%s because its volume was no longer mounted", publishedVolume.Spec.Pod.Namespace, publishedVolume.Spec.Pod.Name)
 		} else {
 			klog.Errorf("Could not delete pod %s/%s because it was no longer mounted because of error: %v", publishedVolume.Spec.Pod.Namespace, publishedVolume.Spec.Pod.Name, err)
 		}

pkg/util/common.go

@@ -2,6 +2,7 @@ package util
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"hash/crc32"
 	"io/ioutil"
@@ -20,6 +21,7 @@ import (
 	"google.golang.org/grpc/status"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/labels"
+	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/kubernetes"
 	"k8s.io/client-go/rest"
 	"k8s.io/klog"
@@ -185,6 +187,48 @@ func GetPvcAnnotations(pvcName string, pvcNamespace string) (annotations map[str
 	return pvc.ObjectMeta.Annotations, nil
 }
 
+// DriverReadyLabel returns the driver-ready label according to the driver name.
+func DriverReadyLabel(driverName string) string {
+	return driverName + "/driver-ready"
+}
+
+// DriverReadyLabelJSONPatchEscaped returns the driver-ready label according to the driver name but espcaed to be used in a JSONPatch path.
+func DriverReadyLabelJSONPatchEscaped(driverName string) string {
+	return strings.ReplaceAll(DriverReadyLabel(driverName), "/", "~1")
+}
+
+// SetDriverReadyLabel set the label <driver name>/driver-ready=<isReady> on the given node.
+func SetDriverReadyLabel(driverName string, nodeName string, isReady bool) (err error) {
+	config, err := rest.InClusterConfig()
+	if err != nil {
+		return err
+	}
+	clientset, err := kubernetes.NewForConfig(config)
+	if err != nil {
+		return err
+	}
+
+	patch := []struct {
+		Op    string `json:"op"`
+		Path  string `json:"path"`
+		Value string `json:"value"`
+	}{{
+		Op:    "add",
+		Path:  "/metadata/labels/" + DriverReadyLabelJSONPatchEscaped(driverName),
+		Value: strconv.FormatBool(isReady),
+	}}
+	patchBytes, err := json.Marshal(patch)
+	if err != nil {
+		return err
+	}
+
+	_, err = clientset.CoreV1().Nodes().Patch(nodeName, types.JSONPatchType, patchBytes)
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
 func DeletePod(namespace string, name string) (err error) {
 	config, err := rest.InClusterConfig()
 	if err != nil {