Scaling and Self-Healing

While Services provide routing and load balancing for Pods, which may go in and out of existence, ReplicaSet (RS) and ReplicationController (RC) are used to specify and then ensure the desired number of Pods (replicas) are in existence. For example, if you always want your application server to be scaled to 3 Pods (instances), a ReplicaSet is needed. Without an RS, any Pods that are killed or somehow die/exit are not automatically restarted. ReplicaSets and ReplicationController are how OpenShift "self heals" and while Deployments control ReplicaSets, ReplicationController here are controlled by DeploymentConfigs.

From the deployments documentation:

In Kubernetes, a Deployment (D) defines how something should be deployed. In almost all cases, you will end up using the Pod, Service, ReplicaSet and Deployment resources together. And, in almost all of those cases, OpenShift will create all of them for you.

There are some edge cases where you might want some Pods and an RS without a D or a Service, and others, so feel free to ask us about them after the labs.

Now that we know the background of what a ReplicaSet and Deployment are, we can explore how they work and are related. Take a look at the Deployment (D) that was created for you when you told OpenShift to stand up the parksmap image:

To get more details, we can look into the ReplicaSet (RS).

Take a look at the ReplicaSet (RS) that was created for you when you told OpenShift to stand up the parksmap image:

This lets us know that, right now, we expect one Pod to be deployed (Desired), and we have one Pod actually deployed (Current). By changing the desired number, we can tell OpenShift that we want more or less Pods.

OpenShift’s HorizontalPodAutoscaler effectively monitors the CPU usage of a set of instances and then manipulates the RCs accordingly.

You can learn more about the CPU-based Horizontal Pod Autoscaler here

Let’s scale our parksmap "application" up to 2 instances. We can do this with the scale command. You could also do this by incrementing the Desired Count in the OpenShift web console. Pick one of these methods; it’s your choice.

You can also scale up to two pods in the Developer Perspective. From the Topology view, first click the parksmap deployment config and select the Details tab:

Next, click the ^ icon next to the Pod visualization to scale up to 2 pods.

To verify that we changed the number of replicas, issue the following command:

You can see that we now have 2 replicas. Let’s verify the number of pods with the oc get pods command:

And lastly, let’s verify that the Service that we learned about in the previous lab accurately reflects two endpoints:

You will see something like the following output:

Another way to look at a Service's endpoints is with the following:

And you will see something like the following:

Your IP addresses will likely be different, as each pod receives a unique IP within the OpenShift environment. The endpoint list is a quick way to see how many pods are behind a service.

You can also see that both Pods are running in the Developer Perspective:

Overall, that’s how simple it is to scale an application (Pods in a Service). Application scaling can happen extremely quickly because OpenShift is just launching new instances of an existing image, especially if that image is already cached on the node.

Because OpenShift’s RSs are constantly monitoring to see that the desired number of Pods actually are running, you might also expect that OpenShift will "fix" the situation if it is ever not right. You would be correct!

Since we have two Pods running right now, let’s see what happens if we "accidentally" kill one. Run the oc get pods command again, and choose a Pod name. Then, do the following:

Did you notice anything? One container has been deleted, and there’s a new container already being created.

Also, the names of the Pods are slightly changed. That’s because OpenShift almost immediately detected that the current state (1 Pod) didn’t match the desired state (2 Pods), and it fixed it by scheduling another Pod.

Additionally, OpenShift provides rudimentary capabilities around checking the liveness and/or readiness of application instances. If the basic checks are insufficient, OpenShift also allows you to run a command inside the container in order to perform the check. That command could be a complicated script that uses any installed language.

Based on these health checks, if OpenShift decided that our parksmap application instance wasn’t alive, it would kill the instance and then restart it, always ensuring that the desired number of replicas was in place.

More information on probing applications is available in the Application Health section of the documentation and later in this guide.

Before we continue, go ahead and scale your application down to a single instance. Feel free to do this using whatever method you like.