etcd operator

Project status: alpha Not all planned feature are completed. The API, spec, status and other user facing objects are subject to change. We do not support backward-compability for the alpha releases.

etcd operator manages etcd clusters atop Kubernetes, automating their creation and administration:

• Create and destroy
• Resize
• Recover a member
• Backup and restore a cluster
• Rolling upgrade

Requirements

• Kubernetes 1.4+
• etcd 3.0+

Deploy etcd operator

$ kubectl create -f example/deployment.yaml
deployment "etcd-operator" created

etcd operator will create a Kubernetes Third-Party Resource (TPR) called "etcd-cluster" automatically.

$ kubectl get thirdpartyresources
NAME                      DESCRIPTION             VERSION(S)
etcd-cluster.coreos.com   Managed etcd clusters   v1

Create and destroy an etcd cluster

$ kubectl create -f example/example-etcd-cluster.yaml

A 3 member etcd cluster will be created.

$ kubectl get pods
NAME                             READY     STATUS    RESTARTS   AGE
etcd-cluster-0000                1/1       Running   0          23s
etcd-cluster-0001                1/1       Running   0          16s
etcd-cluster-0002                1/1       Running   0          8s

$ kubectl get services
NAME                       CLUSTER-IP     EXTERNAL-IP   PORT(S)             AGE
etcd-cluster-0000          10.0.84.34     <none>        2380/TCP,2379/TCP   37s
etcd-cluster-0001          10.0.51.78     <none>        2380/TCP,2379/TCP   30s
etcd-cluster-0002          10.0.140.141   <none>        2380/TCP,2379/TCP   22s

If you are working with minikube locally create a nodePort service and test out that etcd is responding:

kubectl create -f example/example-etcd-cluster-nodeport-service.json
export ETCDCTL_API=3
export ETCDCTL_ENDPOINTS=$(minikube service etcd-cluster-client-service --url)
etcdctl put foo bar

Destroy etcd cluster:

$ kubectl delete -f example/example-etcd-cluster.yaml

Resize an etcd cluster

kubectl apply doesn't work for TPR at the moment. See kubernetes/#29542. As a workaround, we use cURL to resize the cluster.

Create an etcd cluster:

$ kubectl create -f example/example-etcd-cluster.yaml

Use kubectl to create a reverse proxy:

$ kubectl proxy --port=8080
Starting to serve on 127.0.0.1:8080

Now we can talk to apiserver via "http://127.0.0.1:8080".

Create a json file with the new configuration:

$ cat body.json
{
  "apiVersion": "coreos.com/v1",
  "kind": "EtcdCluster",
  "metadata": {
    "name": "etcd-cluster",
    "namespace": "default"
  },
  "spec": {
    "size": 5
  }
}

In another terminal, use the following command to change the cluster size from 3 to 5.

$ curl -H 'Content-Type: application/json' -X PUT --data @body.json http://127.0.0.1:8080/apis/coreos.com/v1/namespaces/default/etcdclusters/etcd-cluster

{
   "apiVersion":"coreos.com/v1",
   "kind":"EtcdCluster",
   "metadata":{
      "name":"etcd-cluster",
      "namespace":"default",
      "selfLink":"/apis/coreos.com/v1/namespaces/default/etcdclusters/etcd-cluster",
      "uid":"4773679d-86cf-11e6-9086-42010af00002",
      "resourceVersion":"438492",
      "creationTimestamp":"2016-09-30T05:32:29Z"
   },
   "spec":{
      "size":5
   }
}

We should see

$ kubectl get pods
NAME                             READY     STATUS              RESTARTS   AGE
etcd-cluster-0000                1/1       Running             0          3m
etcd-cluster-0001                1/1       Running             0          2m
etcd-cluster-0002                1/1       Running             0          2m
etcd-cluster-0003                1/1       Running             0          9s
etcd-cluster-0004                0/1       ContainerCreating   0          1s

Now we can decrease the size of cluster from 5 back to 3.

Create a json file with cluster size of 3:

$ cat body.json
{
  "apiVersion": "coreos.com/v1",
  "kind": "EtcdCluster",
  "metadata": {
    "name": "etcd-cluster",
    "namespace": "default"
  },
  "spec": {
    "size": 3
  }
}

Apply it to API Server:

$ curl -H 'Content-Type: application/json' -X PUT --data @body.json http://127.0.0.1:8080/apis/coreos.com/v1/namespaces/default/etcdclusters/etcd-cluster

We should see that etcd cluster will eventually reduce to 3 pods:

$ kubectl get pods
NAME                             READY     STATUS    RESTARTS   AGE
etcd-cluster-0002                1/1       Running   0          3m
etcd-cluster-0003                1/1       Running   0          1m
etcd-cluster-0004                1/1       Running   0          1m

Member recovery

If the minority of etcd members crash, the etcd operator will automatically recover the failure. Let's walk through in the following steps.

Create an etcd cluster:

$ kubectl create -f example/example-etcd-cluster.yaml

Wait until all three members are up. Simulate a member failure by deleting a pod:

$ kubectl delete pod etcd-cluster-0000 --now

The etcd operator will recover the failure by creating a new pod etcd-cluster-0003:

$ kubectl get pods
NAME                READY     STATUS    RESTARTS   AGE
etcd-cluster-0001   1/1       Running   0          5s
etcd-cluster-0002   1/1       Running   0          5s
etcd-cluster-0003   1/1       Running   0          5s

Destroy etcd cluster:

$ kubectl delete -f example/example-etcd-cluster.yaml

etcd operator recovery

If the etcd operator restarts, it can recover its previous state. Let's walk through in the following steps.

$ kubectl create -f example/example-etcd-cluster.yaml

Wait until all three members are up. Then

$ kubectl delete -f example/deployment.yaml
deployment "etcd-operator" deleted

$ kubectl delete pod etcd-cluster-0000 --now
pod "etcd-cluster-0000" deleted

Then restart the etcd operator. It should recover itself and the etcd clusters it manages.

$ kubectl create -f example/deployment.yaml
deployment "etcd-operator" created

$ kubectl get pods
NAME                READY     STATUS    RESTARTS   AGE
etcd-cluster-0001   1/1       Running   0          5s
etcd-cluster-0002   1/1       Running   0          5s
etcd-cluster-0003   1/1       Running   0          5s

Disaster recovery

If the majority of etcd members crash, but at least one backup exists for the cluster, the etcd operator can restore the entire cluster from the backup.

By default, the etcd operator creates a storage class on initialization:

$ kubectl get storageclass
NAME                            TYPE
etcd-operator-backup-gce-pd   kubernetes.io/gce-pd

This is used to request the persistent volume to store the backup data. (AWS EBS supported using --pv-provisioner=kubernetes.io/aws-ebs. See AWS deployment)

To enable backup, create an etcd cluster with backup enabled spec.

$ kubectl create -f example/example-etcd-cluster-with-backup.yaml

A persistent volume claim is created for the backup pod:

$ kubectl get pvc
NAME               STATUS    VOLUME                                     CAPACITY   ACCESSMODES   AGE
pvc-etcd-cluster   Bound     pvc-164d18fe-8797-11e6-a8b4-42010af00002   1Gi        RWO           14m

Let's try to write some data into etcd:

$ kubectl run --rm -i --tty fun --image quay.io/coreos/etcd --restart=Never -- /bin/sh
/ # ETCDCTL_API=3 etcdctl --endpoints http://etcd-cluster-with-backup-0002:2379 put foo bar
OK
(ctrl-D to exit)

Now let's kill two pods to simulate a disaster failure:

$ kubectl delete pod etcd-cluster-with-backup-0000 etcd-cluster-with-backup-0001 --now
pod "etcd-cluster-with-backup-0000" deleted
pod "etcd-cluster-with-backup-0001" deleted

Now quorum is lost. The etcd operator will start to recover the cluster by:

• Creating a new seed member to recover from the backup
• Add more members until the size reaches to specified number

$ kubectl get pods
NAME                                         READY     STATUS     RESTARTS   AGE
etcd-cluster-with-backup-0003                0/1       Init:0/2   0          11s
etcd-cluster-with-backup-backup-tool-e9gkv   1/1       Running    0          18m
...
$ kubectl get pods
NAME                                         READY     STATUS    RESTARTS   AGE
etcd-cluster-with-backup-0003                1/1       Running   0          3m
etcd-cluster-with-backup-0004                1/1       Running   0          3m
etcd-cluster-with-backup-0005                1/1       Running   0          3m
etcd-cluster-with-backup-backup-tool-e9gkv   1/1       Running   0          22m

Note: There could be a race that it will fall to single member recovery if a pod is recovered before another is deleted.

Upgrade an etcd cluster

Have the following yaml file ready:

$ cat 3.0-etcd-cluster.yaml
apiVersion: "coreos.com/v1"
kind: "EtcdCluster"
metadata:
  name: "etcd-cluster"
spec:
  size: 3
  version: "v3.0.13"

Create an etcd cluster with the version specified (3.0.13) in the yaml file:

$ kubectl create -f 3.0-etcd-cluster.yaml
$ kubectl get pods
NAME                   READY     STATUS    RESTARTS   AGE
etcd-cluster-0000      1/1       Running   0          37s
etcd-cluster-0001      1/1       Running   0          25s
etcd-cluster-0002      1/1       Running   0          14s

The container image version should be 3.0.13:

$ kubectl get pod etcd-cluster-0000 -o yaml | grep "image:" | uniq
    image: quay.io/coreos/etcd:v3.0.13

kubectl apply doesn't work for TPR at the moment. See kubernetes/#29542. We use cURL to update the cluster as a workaround.

Use kubectl to create a reverse proxy:

$ kubectl proxy --port=8080
Starting to serve on 127.0.0.1:8080

Have following json file ready: (Note that the version field is changed from v3.0.13 to v3.1.0-alpha.1)

$ cat body.json
{
  "apiVersion": "coreos.com/v1",
  "kind": "EtcdCluster",
  "metadata": {
    "name": "etcd-cluster"
  },
  "spec": {
    "size": 3,
    "version": "v3.1.0-alpha.1"
  }
}

Then we update the version in spec.

$ curl -H 'Content-Type: application/json' -X PUT --data @body.json \
    http://127.0.0.1:8080/apis/coreos.com/v1/namespaces/default/etcdclusters/etcd-cluster

Wait ~30 seconds. The container image version should be updated to v3.1.0-alpha.1:

$ kubectl get pod etcd-cluster-0000 -o yaml | grep "image:" | uniq
    image: quay.io/coreos/etcd:v3.1.0-alpha.1

Check the other two pods and you should see the same result.

Limitations

• Backup works only for data in etcd3 storage, not for data in etcd2 storage.
• Backup requires PV to work, and it only works on GCE(kubernetes.io/gce-pd) and AWS(kubernetes.io/aws-ebs) for now.
• Migration, the process of allowing the etcd operator to manage existing etcd3 clusters, only supports a single-member cluster, with all nodes running in the same Kubernetes cluster.