Zookeeper Helm Chart

This helm chart provides an implementation of the ZooKeeper StatefulSet found in Kubernetes Contrib Zookeeper StatefulSet.

Prerequisites

• Kubernetes 1.6
• PersistentVolume support on the underlying infrastructure
• A dynamic provisioner for the PersistentVolumes
• A familiarity with Apache ZooKeeper 3.4.x

Chart Components

This chart will do the following:

• Create a fixed size ZooKeeper ensemble using a StatefulSet.
• Create a PodDisruptionBudget so kubectl drain will respect the Quorum size of the ensemble.
• Create a Headless Service to control the domain of the ZooKeeper ensemble.
• Create a Service configured to connect to the available ZooKeeper instance on the configured client port.
• Optionally, apply a Pod Anti-Affinity to spread the ZooKeeper ensemble across nodes.

Installing the Chart

You can install the chart with the release name myzk as below.

$ helm repo add incubator http://storage.googleapis.com/kubernetes-charts-incubator
$ helm install --name myzk incubator/zookeeper

If you do not specify a name, helm will select a name for you.

Installed Components

You can use kubectl get to view all of the installed components.

$ kubectl get all -l app=zookeeper
NAME                   READY     STATUS    RESTARTS   AGE
po/myzk-zookeeper-0   1/1       Running   0          2m
po/myzk-zookeeper-1   1/1       Running   0          1m
po/myzk-zookeeper-2   1/1       Running   0          52s

NAME                           CLUSTER-IP    EXTERNAL-IP   PORT(S)             AGE
svc/myzk-zookeeper            10.3.255.86   <none>        2181/TCP            2m
svc/myzk-zookeeper-headless   None          <none>        2888/TCP,3888/TCP   2m

NAME                           DESIRED   CURRENT   AGE
statefulsets/myzk-zookeeper   3         3         2m

1. statefulsets/myzk-zookeeper is the StatefulSet created by the chart.
2. po/myzk-zookeeper-<0|1|2> are the Pods created by the StatefulSet. Each Pod has a single container running a ZooKeeper server.
3. svc/myzk-zookeeper-headless is the Headless Server used to control the network domain of the ZooKeeper ensemble.
4. svc/myzk-zookeeper is a Service that can be used by clients to connect to an available ZooKeeper server.

Configuration

You can specify each parameter using the --set key=value[,key=value] argument to helm install.

Alternatively, a YAML file that specifies the values for the parameters can be provided while installing the chart. For example,

$ helm install --name my-release -f values.yaml incubator/zookeeper

Tip: You can use the default values.yaml

Resources

The configuration parameters in this section control the resources requested and utilized by the ZooKeeper ensemble.

Network

These parameters control the network ports on which the ensemble communicates.

Time

ZooKeeper uses the Zab protocol to replicate its state machine across the ensemble. The following parameters control the timeouts for the protocol.

Log Retention

ZooKeeper writes its WAL (Write Ahead Log) and periodic snapshots to storage media. These parameters control the retention policy for snapshots and WAL segments. If you do not configure the ensemble to automatically periodically purge snapshots and logs, it is important to implement such a mechanism yourself. Otherwise, you will eventually exhaust all available storage media.

Spreading

Spreading allows you specify an anti-affinity between ZooKeeper servers in the ensemble. This will prevent the Pods from being scheduled on the same node.

Logging

In order to allow for the default installation to work well with the log rolling and retention policy of Kubernetes, all logs are written to stdout. This should also be compatible with logging integrations such as Google Cloud Logging and ELK.

Liveness and Readiness

The servers in the ensemble have both liveness and readiness checks specified. These parameters can be used to tune the sensitivity of the liveness and readiness checks.

ImagePull

This parameter controls when the image is pulled from the repository.

Deep dive

Image Details

The image used for this chart is based on Ubuntu 16.04 LTS. This image is larger than Alpine or BusyBox, but it provides glibc, rather than ulibc or mucl, and a JVM release that is built against it. You can easily convert this chart to run against a smaller image with a JVM that is build against that images libc. However, as far as we know, no Hadoop vendor supports, or has verified, ZooKeeper running on such a JVM.

JVM Details

The Java Virtual Machine used for this chart is the OpenJDK JVM 8u111 JRE (headless).

ZooKeeper Details

The ZooKeeper version is the latest stable version (3.4.9). The distribution is installed into /opt/zookeeper-3.4.9. This directory is symbolically linked to /opt/zookeeper. Symlinks are created to simulate a rpm installation into /usr.

Failover

You can test failover by killing the leader. Insert a key:

$ kubectl exec <RELEASE-NAME>-zookeeper-0 -- /opt/zookeeper/bin/zkCli.sh create /foo bar;
$ kubectl exec <RELEASE-NAME>-zookeeper-2 -- /opt/zookeeper/bin/zkCli.sh get /foo;

Watch existing members:

$ kubectl run --attach bbox --image=busybox --restart=Never -- sh -c 'while true; do for i in 0 1 2; do echo zk-${i} $(echo stats | nc <pod-name>-${i}.<headless-service-name>:2181 | grep Mode); sleep 1; done; done';

zk-2 Mode: follower
zk-0 Mode: follower
zk-1 Mode: leader
zk-2 Mode: follower

Delete Pods and wait for the StatefulSet controller to bring them back up:

$ kubectl delete po -l app=zookeeper
$ kubectl get po --watch-only
NAME                READY     STATUS    RESTARTS   AGE
myzk-zookeeper-0   0/1       Running   0          35s
myzk-zookeeper-0   1/1       Running   0         50s
myzk-zookeeper-1   0/1       Pending   0         0s
myzk-zookeeper-1   0/1       Pending   0         0s
myzk-zookeeper-1   0/1       ContainerCreating   0         0s
myzk-zookeeper-1   0/1       Running   0         19s
myzk-zookeeper-1   1/1       Running   0         40s
myzk-zookeeper-2   0/1       Pending   0         0s
myzk-zookeeper-2   0/1       Pending   0         0s
myzk-zookeeper-2   0/1       ContainerCreating   0         0s
myzk-zookeeper-2   0/1       Running   0         19s
myzk-zookeeper-2   1/1       Running   0         41s

...

zk-0 Mode: follower
zk-1 Mode: leader
zk-2 Mode: follower

Check the previously inserted key:

$ kubectl exec myzk-zookeeper-1 -- /opt/zookeeper/bin/zkCli.sh get /foo
ionid = 0x354887858e80035, negotiated timeout = 30000

WATCHER::

WatchedEvent state:SyncConnected type:None path:null
bar

Scaling

ZooKeeper can not be safely scaled in versions prior to 3.5.x. There are manual procedures for scaling an ensemble, but as noted in the ZooKeeper 3.5.2 documentation these procedures require a rolling restart, are known to be error prone, and often result in a data loss.

While ZooKeeper 3.5.x does allow for dynamic ensemble reconfiguration (including scaling membership), the current status of the release is still alpha, and it is not recommended for production use.

Limitations

• StatefulSet and PodDisruptionBudget are beta resources.
• Only supports storage options that have backends for persistent volume claims.