scalar

The Kubernetes HorizontalPodAutoscaler controller you've been missing.

When do I use scalar?

Simple, if you use Deployments or ReplicationControllers, have heapster configured, and don't want to manually scale pods (or manually create the artifacts to automatically do so), you use scalar. Kubernetes has a built in autscaling feature in the HorizontalPodAutoscaler. This native resource monitors a Deployment's or ReplicationController's CPU usage, and when the average among all related Pods reaches a pre-configured threshold, more Pods are created to lower it. Find more info in the k8s docs.

So, rather than having to manually monitor usage of a Deployment's Pods and use a command like kubectl scale deployment nginx --replicas=5, use a HorizontalPodAutoscaler to automate it. And even better than manually creating one of these for every deployment a coworker or tenant makes like so kubectl autoscale deployment nginx --min=2 --max=10, let scalar do it for you!

What does scalar actually do?

Again, it's simple. scalar doesn't do anything special. It just lives its monotonous life, watching for Deployment and ReplicationController creation and deletion events in your cluster. When one of these is created, scalar springs into action, creating a HorizontalPodAutoscaler for the newly created artifact and adds a reference of it to its cache. While scalar is humming along, it will give periodic status updates of your cluster's scaling activities by logging the status of each active HorizontalPodAutoscaler (even those it didn't create). The frequency of said updates is configurable, and "togglable". Finally, when a Deployment or ReplicationController is deleted, if scalar had created a HorizontalPodAutoscaler for it, this will be deleted as well, and removed from the cache.

How does it do the stuff?

Keeping the trend of simplicity, scalar's implementation intends to be just that. Using the Kubernetes watch interface (docs here), scalar follows all CRUD events for Deployments and ReplicationControllers, filtering out the events it doesn't care about, and reacting to the ones it does. The implementation uses the kubernetes/client-go package to generate an in-cluster configuration used to communicate with the k8s apiserver to do all the things and stuff (docs here).

Are there any cluster dependencies?

In order for scalar to be at all effective, you need a few things. First, and most important, you need a functioning set-up of heapster in your cluster. If you don't have this, follow the instructions layed out here. Secondly, scalar needs to run in a namespace where the default ServiceAccount has authorization to watch resource events and create/delete resources everywhere necessary. This is an out of the box configuration item, if you and your cluster have a different authz pattern, do your thang!

Let's deploy this thing!

This part is easy. Copy the manifest in master, configure it as you see fit, deploy:

> kubectl create -f scalar.yaml

Verify that it worked:

> kubectl get pods
NAME                     READY     STATUS    RESTARTS   AGE
scalar-700111452-0vwas   1/1       Running   0          10s

Check that scalar is happy:

> kubectl logs scalar-700111452-0vwas
2016/10/17 03:56:22 Scalar is configured and ready to scale!

Test that it works:

> kubectl run nginx --image=nginx --port=80
deployment "nginx" created
> kubectl logs scalar-700111452-0vwas
2016/10/17 03:56:22 Scalar is configured and ready to scale!
2016/10/17 03:57:12 Creating horizontal pod autoscaler for deployment nginx default

You're good to go! scalar is scaling away for you!

What do the status updates look like?

It's ok to want to see what you are getting, turn it off if you think it is annoying. Here is what a status update might look like:

2016/10/17 03:57:12 Status for HPA "nginx" in default: | ObservedGeneration: 0 | LastScaleTime: <nil> | CurrentReplicas: 2 | DesiredReplicas: 2 | CurrentCPUUtilizationPercentage: 10 |

What knobs can I turn?

Glad you asked! Here is a list of available configuration values.

Env Var	Default	Description
`SCALAR_SELECTOR`	`""`	A label selector used to filter the watcher events to specifically labeled objects
`SCALAR_PRINT_STATUS`	`"true"`	Flag that toggles printing of autocaling statuses
`SCALAR_STATUS_TIMER`	`"1800"`	Frequency in seconds of autoscaling status print
`SCALAR_MIN_REPLICAS`	`"2"`	Minimum number of active replicas in a Deployment/ReplicationController
`SCALAR_MAX_REPLICAS`	`"8"`	Maximum number of active replicas in a Deployment/ReplicationController
`SCALAR_TARGET_CPU`	`"75"`	Target CPU Utilization Percentage for the autoscaling threshold

If you have any questions or want to contribute, open an issue or PR in the repo! Thank you and have fun with scalar!

noahdietz / scalar