Editor's note: today's post is by Janet Kuo and Kenneth Owens, Software Engineers at Google.
Google
This post talks about recent updates to the DaemonSet and StatefulSet API objects for Kubernetes. We explore these features using Apache ZooKeeper and Apache KafkaStatefulSets and a Prometheus node exporterDaemonSet.
In Kubernetes 1.6, we added the RollingUpdate update strategy to the DaemonSet API Object. Configuring your DaemonSets with the RollingUpdate strategy causes the DaemonSet controller to perform automated rolling updates to the Pods in your DaemonSets when their spec.template are updated.
In Kubernetes 1.7, we enhanced the DaemonSet controller to track a history of revisions to the PodTemplateSpecs of DaemonSets. This allows the DaemonSet controller to roll back an update. We also added the RollingUpdate strategy to the StatefulSet API Object, and implemented revision history tracking for the StatefulSet controller. Additionally, we added the Parallel pod management policy to support stateful applications that require Pods with unique identities but not ordered Pod creation and termination.
StatefulSet rolling update and Pod management policy
First, we’re going to demonstrate how to use StatefulSet rolling updates and Pod management policies by deploying a ZooKeeper ensemble and a Kafka cluster.
Prerequisites
To follow along, you’ll need to set up a Kubernetes 1.7 cluster with at least 3 schedulable nodes. Each node needs 1 CPU and 2 GiB of memory available. You will also need either a dynamic provisioner to allow the StatefulSet controller to provision 6 persistent volumes (PVs) with 10 GiB each, or you will need to manually provision the PVs prior to deploying the ZooKeeper ensemble or deploying the Kafka cluster.
Deploying a ZooKeeper ensemble
Apache ZooKeeper is a strongly consistent, distributed system used by other distributed systems for cluster coordination and configuration management.
Note: You can create a ZooKeeper ensemble using this zookeeper_mini.yaml manifest. You can learn more about running a ZooKeeper ensemble on Kubernetes here, as well as a more in-depth explanation of the manifest and its contents.
When you apply the manifest, you will see output like the following.
$ kubectl apply -f zookeeper_mini.yaml service "zk-hs" created service "zk-cs" created poddisruptionbudget "zk-pdb" created statefulset "zk" created |
The manifest creates an ensemble of three ZooKeeper servers using a StatefulSet, zk; a Headless Service, zk-hs, to control the domain of the ensemble; a Service, zk-cs, that clients can use to connect to the ready ZooKeeper instances; and a PodDisruptionBugdet, zk-pdb, that allows for one planned disruption. (Note that while this ensemble is suitable for demonstration purposes, it isn’t sized correctly for production use.)
If you use kubectl get to watch Pod creation in another terminal you will see that, in contrast to the OrderedReady strategy (the default policy that implements the full version of the StatefulSet guarantees), all of the Pods in the zk StatefulSet are created in parallel.
$ kubectl get po -lapp=zk -w NAME READY STATUS RESTARTS AGE zk-0 0/1 Pending 0 0s zk-0 0/1 Pending 0 0s zk-1 0/1 Pending 0 0s zk-1 0/1 Pending 0 0s zk-0 0/1 ContainerCreating 0 0s zk-2 0/1 Pending 0 0s zk-1 0/1 ContainerCreating 0 0s zk-2 0/1 Pending 0 0s zk-2 0/1 ContainerCreating 0 0s zk-0 0/1 Running 0 10s zk-2 0/1 Running 0 11s zk-1 0/1 Running 0 19s zk-0 1/1 Running 0 20s zk-1 1/1 Running 0 30s zk-2 1/1 Running 0 30s |
This is because the zookeeper_mini.yaml manifest sets the podManagementPolicy of the StatefulSet to Parallel.
apiVersion: apps/v1beta1 kind: StatefulSet metadata: name: zk spec: serviceName: zk-hs replicas: 3 updateStrategy: type: RollingUpdate podManagementPolicy: Parallel ... |
Many distributed systems, like ZooKeeper, do not require ordered creation and termination for their processes. You can use the Parallel Pod management policy to accelerate the creation and deletion of StatefulSets that manage these systems. Note that, when Parallel Pod management is used, the StatefulSet controller will not block when it fails to create a Pod. Ordered, sequential Pod creation and termination is performed when a StatefulSet’s podManagementPolicy is set to OrderedReady.
Deploying a Kafka Cluster
Apache Kafka is a popular distributed streaming platform. Kafka producers write data to partitioned topics which are stored, with a configurable replication factor, on a cluster of brokers. Consumers consume the produced data from the partitions stored on the brokers.
Note: Details of the manifests contents can be found here. You can learn more about running a Kafka cluster on Kubernetes here.
To create a cluster, you only need to download and apply the kafka_mini.yaml manifest. When you apply the manifest, you will see output like the following:
$ kubectl apply -f kafka_mini.yaml service "kafka-hs" created poddisruptionbudget "kafka-pdb" created statefulset "kafka" created |
The manifest creates a three broker cluster using the kafka StatefulSet, a Headless Service, kafka-hs, to control the domain of the brokers; and a PodDisruptionBudget, kafka-pdb, that allows for one planned disruption. The brokers are configured to use the ZooKeeper ensemble we created above by connecting through the zk-cs Service. As with the ZooKeeper ensemble deployed above, this Kafka cluster is fine for demonstration purposes, but it’s probably not sized correctly for production use.
If you watch Pod creation, you will notice that, like the ZooKeeper ensemble created above, the Kafka cluster uses the Parallel podManagementPolicy.
$ kubectl get po -lapp=kafka -w NAME READY STATUS RESTARTS AGE kafka-0 0/1 Pending 0 0s kafka-0 0/1 Pending 0 0s kafka-1 0/1 Pending 0 0s kafka-1 0/1 Pending 0 0s kafka-2 0/1 Pending 0 0s kafka-0 0/1 ContainerCreating 0 0s kafka-2 0/1 Pending 0 0s kafka-1 0/1 ContainerCreating 0 0s kafka-1 0/1 Running 0 11s kafka-0 0/1 Running 0 19s kafka-1 1/1 Running 0 23s kafka-0 1/1 Running 0 32s |
Producing and consuming data
You can use kubectl run to execute the kafka-topics.sh script to create a topic named test.
$ kubectl run -ti --image=gcr.io/google_containers/kubernetes-kafka:1.0-10.2.1 createtopic --restart=Never --rm -- kafka-topics.sh --create \ > --topic test \ > --zookeeper zk-cs.default.svc.cluster.local:2181 \ > --partitions 1 \ > --replication-factor 3 |
Now you can use kubectl run to execute the kafka-console-consumer.sh command to listen for messages.
$ kubectl run -ti --image=gcr.io/google_containers/kubnetes-kafka:1.0-10.2.1 consume --restart=Never --rm -- kafka-console-consumer.sh --topic test --bootstrap-server kafka-0.kafka-hs.default.svc.cluster.local:9093 |
In another terminal, you can run the kafka-console-producer.sh command.
$kubectl run -ti --image=gcr.io/google_containers/kubernetes-kafka:1.0-10.2.1 produce --restart=Never --rm \ > -- kafka-console-producer.sh --topic test --broker-list kafka-0.kafka-hs.default.svc.cluster.local:9093,kafka-1.kafka-hs.default.svc.cluster.local:9093,kafka-2.kafka-hs.default.svc.cluster.local:9093 |
Output from the second terminal appears in the first terminal. If you continue to produce and consume messages while updating the cluster, you will notice that no messages are lost. You may see error messages as the leader for the partition changes when individual brokers are updated, but the client retries until the message is committed. This is due to the ordered, sequential nature of StatefulSet rolling updates which we will explore further in the next section.
Updating the Kafka cluster
StatefulSet updates are like DaemonSet updates in that they are both configured by setting the spec.updateStrategy of the corresponding API object. When the update strategy is set to OnDelete, the respective controllers will only create new Pods when a Pod in the StatefulSet or DaemonSet has been deleted. When the update strategy is set to RollingUpdate, the controllers will delete and recreate Pods when a modification is made to the spec.template field of a DaemonSet or StatefulSet. You can use rolling updates to change the configuration (via environment variables or command line parameters), resource requests, resource limits, container images, labels, and/or annotations of the Pods in a StatefulSet or DaemonSet. Note that all updates are destructive, always requiring that each Pod in the DaemonSet or StatefulSet be destroyed and recreated. StatefulSet rolling updates differ from DaemonSet rolling updates in that Pod termination and creation is ordered and sequential.
You can patch the kafka StatefulSet to reduce the CPU resource request to 250m.
$ kubectl patch sts kafka --type='json' -p='[{"op": "replace", "path": "/spec/template/spec/containers/0/resources/requests/cpu", "value":"250m"}]' statefulset "kafka" patched |
If you watch the status of the Pods in the StatefulSet, you will see that each Pod is deleted and recreated in reverse ordinal order (starting with the Pod with the largest ordinal and progressing to the smallest). The controller waits for each updated Pod to be running and ready before updating the subsequent Pod.
$kubectl get po -lapp=kafka -w NAME READY STATUS RESTARTS AGE kafka-0 1/1 Running 0 13m kafka-1 1/1 Running 0 13m kafka-2 1/1 Running 0 13m kafka-2 1/1 Terminating 0 14m kafka-2 0/1 Terminating 0 14m kafka-2 0/1 Terminating 0 14m kafka-2 0/1 Terminating 0 14m kafka-2 0/1 Pending 0 0s kafka-2 0/1 Pending 0 0s kafka-2 0/1 ContainerCreating 0 0s kafka-2 0/1 Running 0 10s kafka-2 1/1 Running 0 21s kafka-1 1/1 Terminating 0 14m kafka-1 0/1 Terminating 0 14m kafka-1 0/1 Terminating 0 14m kafka-1 0/1 Terminating 0 14m kafka-1 0/1 Pending 0 0s kafka-1 0/1 Pending 0 0s kafka-1 0/1 ContainerCreating 0 0s kafka-1 0/1 Running 0 11s kafka-1 1/1 Running 0 21s kafka-0 1/1 Terminating 0 14m kafka-0 0/1 Terminating 0 14m kafka-0 0/1 Terminating 0 14m kafka-0 0/1 Terminating 0 14m kafka-0 0/1 Pending 0 0s kafka-0 0/1 Pending 0 0s kafka-0 0/1 ContainerCreating 0 0s kafka-0 0/1 Running 0 10s kafka-0 1/1 Running 0 22s |
Note that unplanned disruptions will not lead to unintentional updates during the update process. That is, the StatefulSet controller will always recreate the Pod at the correct version to ensure the ordering of the update is preserved. If a Pod is deleted, and if it has already been updated, it will be created from the updated version of the StatefulSet’s spec.template. If the Pod has not already been updated, it will be created from the previous version of the StatefulSet’s spec.template. We will explore this further in the following sections.
Staging an update
Depending on how your organization handles deployments and configuration modifications, you may want or need to stage updates to a StatefulSet prior to allowing the roll out to progress. You can accomplish this by setting a partition for the RollingUpdate. When the StatefulSet controller detects a partition in the updateStrategy of a StatefulSet, it will only apply the updated version of the StatefulSet’s spec.template to Pods whose ordinal is greater than or equal to the value of the partition.
You can patch the kafka StatefulSet to add a partition to the RollingUpdate update strategy. If you set the partition to a number greater than or equal to the StatefulSet’s spec.replicas (as below), any subsequent updates you perform to the StatefulSet’s spec.template will be staged for roll out, but the StatefulSet controller will not start a rolling update.
$ kubectl patch sts kafka -p '{"spec":{"updateStrategy":{"type":"RollingUpdate","rollingUpdate":{"partition":3}}}}' statefulset "kafka" patched |
If you patch the StatefulSet to set the requested CPU to 0.3, you will notice that none of the Pods are updated.
$ kubectl patch sts kafka --type='json' -p='[{"op": "replace", "path": "/spec/template/spec/containers/0/resources/requests/cpu", "value":"0.3"}]' statefulset "kafka" patched |
Even if you delete a Pod and wait for the StatefulSet controller to recreate it, you will notice that the Pod is recreated with current CPU request.
$ kubectl delete po kafka-1 pod "kafka-1" deleted $ kubectl get po kafka-1 -w NAME READY STATUS RESTARTS AGE kafka-1 0/1 ContainerCreating 0 10s kafka-1 0/1 Running 0 19s kafka-1 1/1 Running 0 21s $ kubectl get po kafka-1 -o yaml apiVersion: v1 kind: Pod metadata: ... resources: requests: cpu: 250m memory: 1Gi |
Rolling out a canary
Often, we want to verify an image update or configuration change on a single instance of an application before rolling it out globally. If you modify the partition created above to be 2, the StatefulSet controller will roll out a canary that can be used to verify that the update is working as intended.
$ kubectl patch sts kafka -p '{"spec":{"updateStrategy":{"type":"RollingUpdate","rollingUpdate":{"partition":2}}}}' statefulset "kafka" patched |
You can watch the StatefulSet controller update the kafka-2 Pod and pause after the update is complete.
$ kubectl get po -lapp=kafka -w NAME READY STATUS RESTARTS AGE kafka-0 1/1 Running 0 50m kafka-1 1/1 Running 0 10m kafka-2 1/1 Running 0 29s kafka-2 1/1 Terminating 0 34s kafka-2 0/1 Terminating 0 38s kafka-2 0/1 Terminating 0 39s kafka-2 0/1 Terminating 0 39s kafka-2 0/1 Pending 0 0s kafka-2 0/1 Pending 0 0s kafka-2 0/1 Terminating 0 20s kafka-2 0/1 Terminating 0 20s kafka-2 0/1 Pending 0 0s kafka-2 0/1 Pending 0 0s kafka-2 0/1 ContainerCreating 0 0s kafka-2 0/1 Running 0 19s kafka-2 1/1 Running 0 22s |
Phased roll outs
Similar to rolling out a canary, you can roll out updates based on a phased progression (e.g. linear, geometric, or exponential roll outs).
If you patch the kafka StatefulSet to set the partition to 1, the StatefulSet controller updates one more broker.
$ kubectl patch sts kafka -p '{"spec":{"updateStrategy":{"type":"RollingUpdate","rollingUpdate":{"partition":1}}}}' statefulset "kafka" patched |
If you set it to 0, the StatefulSet controller updates the final broker and completes the update.
$ kubectl patch sts kafka -p '{"spec":{"updateStrategy":{"type":"RollingUpdate","rollingUpdate":{"partition":0}}}}' statefulset "kafka" patched |
Note that you don’t have to decrement the partition by one. For a larger StatefulSet--for example, one with 100 replicas--you might use a progression more like 100, 99, 90, 50, 0. In this case, you would stage your update, deploy a canary, roll out to 10 instances, update fifty percent of the Pods, and then complete the update.
Cleaning up
To delete the API Objects created above, you can use kubectl delete on the two manifests you used to create the ZooKeeper ensemble and the Kafka cluster.
$ kubectl delete -f kafka_mini.yaml service "kafka-hs" deleted poddisruptionbudget "kafka-pdb" deleted Statefulset “kafka” deleted $ kubectl delete -f zookeeper_mini.yaml service "zk-hs" deleted service "zk-cs" deleted poddisruptionbudget "zk-pdb" deleted statefulset "zk" deleted |
By design, the StatefulSet controller does not delete any persistent volume claims (PVCs): the PVCs created for the ZooKeeper ensemble and the Kafka cluster must be manually deleted. Depending on the storage reclamation policy of your cluster, you many also need to manually delete the backing PVs.
DaemonSet rolling update, history, and rollback
In this section, we’re going to show you how to perform a rolling update on a DaemonSet, look at its history, and then perform a rollback after a bad rollout. We will use a DaemonSet to deploy a Prometheus node exporter on each Kubernetes node in the cluster. These node exporters export node metrics to the Prometheus monitoring system. For the sake of simplicity, we’ve omitted the installation of the Prometheus server and the service for communication with DaemonSet pods from this blogpost.
Prerequisites
To follow along with this section of the blog, you need a working Kubernetes 1.7 cluster and kubectl version 1.7 or later. If you followed along with the first section, you can use the same cluster.
DaemonSet rolling update: Prometheus node exporters
First, prepare the node exporter DaemonSet manifest to run a v0.13 Prometheus node exporter on every node in the cluster:
$ cat >> node-exporter-v0.13.yaml <<EOF apiVersion: extensions/v1beta1 kind: DaemonSet metadata: name: node-exporter spec: updateStrategy: type: RollingUpdate template: metadata: labels: app: node-exporter name: node-exporter spec: containers: - image: prom/node-exporter:v0.13.0 name: node-exporter ports: - containerPort: 9100 hostPort: 9100 name: scrape hostNetwork: true hostPID: true EOF |
Note that you need to enable the DaemonSet rolling update feature by explicitly setting DaemonSet.spec.updateStrategy.type to RollingUpdate.
Apply the manifest to create the node exporter DaemonSet:
$ kubectl apply -f node-exporter-v0.13.yaml --record daemonset "node-exporter" created |
Wait for the first DaemonSet rollout to complete:
$ kubectl rollout status ds node-exporter daemon set "node-exporter" successfully rolled out |
You should see each of your node runs one copy of the node exporter pod:
$ kubectl get pods -l app=node-exporter -o wide |
To perform a rolling update on the node exporter DaemonSet, prepare a manifest that includes the v0.14 Prometheus node exporter:
$ cat node-exporter-v0.13.yaml | sed "s/v0.13.0/v0.14.0/g"> node-exporter-v0.14.yaml |
Then apply the v0.14 node exporter DaemonSet:
$ kubectl apply -f node-exporter-v0.14.yaml --record daemonset "node-exporter" configured |
Wait for the DaemonSet rolling update to complete:
$ kubectl rollout status ds node-exporter ... Waiting for rollout to finish: 3 out of 4 new pods have been updated... Waiting for rollout to finish: 3 of 4 updated pods are available... daemon set "node-exporter" successfully rolled out |
We just triggered a DaemonSet rolling update by updating the DaemonSet template. By default, one old DaemonSet pod will be killed and one new DaemonSet pod will be created at a time.
Now we’ll cause a rollout to fail by updating the image to an invalid value:
$ cat node-exporter-v0.13.yaml | sed "s/v0.13.0/bad/g"> node-exporter-bad.yaml $ kubectl apply -f node-exporter-bad.yaml --record daemonset "node-exporter" configured |
Notice that the rollout never finishes:
$ kubectl rollout status ds node-exporter Waiting for rollout to finish: 0 out of 4 new pods have been updated... Waiting for rollout to finish: 1 out of 4 new pods have been updated… # Use ^C to exit |
This behavior is expected. We mentioned earlier that a DaemonSet rolling update kills and creates one pod at a time. Because the new pod never becomes available, the rollout is halted, preventing the invalid specification from propagating to more than one node. StatefulSet rolling updates implement the same behavior with respect to failed deployments. Unsuccessful updates are blocked until it corrected via roll back or by rolling forward with a specification.
$ kubectl get pods -l app=node-exporter NAME READY STATUS RESTARTS AGE node-exporter-f2n14 0/1 ErrImagePull 0 3m ... # N = number of nodes $ kubectl get ds node-exporter NAME DESIRED CURRENT READY UP-TO-DATE AVAILABLE NODE SELECTOR AGE node-exporter N N N-1 1 N <none> 46m |
DaemonSet history, rollbacks, and rolling forward
Next, perform a rollback. Take a look at the node exporter DaemonSet rollout history:
$ kubectl rollout history ds node-exporter daemonsets "node-exporter" REVISION CHANGE-CAUSE 1 kubectl apply --filename=node-exporter-v0.13.yaml --record=true 2 kubectl apply --filename=node-exporter-v0.14.yaml --record=true 3 kubectl apply --filename=node-exporter-bad.yaml --record=true |
Check the details of the revision you want to roll back to:
$ kubectl rollout history ds node-exporter --revision=2 daemonsets "node-exporter" with revision #2 Pod Template: Labels: app=node-exporter Containers: node-exporter: Image: prom/node-exporter:v0.14.0 Port: 9100/TCP Environment: <none> Mounts: <none> Volumes: <none> |
You can quickly roll back to any DaemonSet revision you found through kubectl rollout history:
# Roll back to the last revision $ kubectl rollout undo ds node-exporter daemonset "node-exporter" rolled back # Or use --to-revision to roll back to a specific revision $ kubectl rollout undo ds node-exporter --to-revision=2 daemonset "node-exporter" rolled back |
A DaemonSet rollback is done by rolling forward. Therefore, after the rollback, DaemonSet revision 2 becomes revision 4 (current revision):
$ kubectl rollout history ds node-exporter daemonsets "node-exporter" REVISION CHANGE-CAUSE 1 kubectl apply --filename=node-exporter-v0.13.yaml --record=true 3 kubectl apply --filename=node-exporter-bad.yaml --record=true 4 kubectl apply --filename=node-exporter-v0.14.yaml --record=true |
The node exporter DaemonSet is now healthy again:
$ kubectl rollout status ds node-exporter daemon set "node-exporter" successfully rolled out # N = number of nodes $ kubectl get ds node-exporter NAME DESIRED CURRENT READY UP-TO-DATE AVAILABLE NODE SELECTOR AGE node-exporter N N N N N <none> 46m |
If current DaemonSet revision is specified while performing a rollback, the rollback is skipped:
$ kubectl rollout undo ds node-exporter --to-revision=4 daemonset "node-exporter" skipped rollback (current template already matches revision 4) |
You will see this complaint from kubectl if the DaemonSet revision is not found:
$ kubectl rollout undo ds node-exporter --to-revision=10 error: unable to find specified revision 10 in history |
Note that kubectl rollout history and kubectl rollout status support StatefulSets, too!
Cleaning up
$ kubectl delete ds node-exporter |
What’s next for DaemonSet and StatefulSet
Rolling updates and roll backs close an important feature gap for DaemonSets and StatefulSets. As we plan for Kubernetes 1.8, we want to continue to focus on advancing the core controllers to GA. This likely means that some advanced feature requests (e.g. automatic roll back, infant mortality detection) will be deferred in favor of ensuring the consistency, usability, and stability of the core controllers. We welcome feedback and contributions, so please feel free to reach out on Slack, to ask questions on Stack Overflow, or open issues or pull requests on GitHub.
- Post questions (or answer questions) on Stack Overflow
- Join the community portal for advocates on K8sPort
- Connect with the community on Slack
- Get involved with the Kubernetes project on GitHub