A general purpose, scale-to-zero component for Kubernetes
Clone or download

README.md

Osiris - A general purpose, Scale to Zero component for Kubernetes

CircleCI

Osiris enables greater resource efficiency within a Kubernetes cluster by allowing idling workloads to automatically scale-to-zero and allowing scaled-to-zero workloads to be automatically re-activated on-demand by inbound requests.

Osiris, as a concept, is highly experimental and currently remains under heavy development.

How it works

For Osiris-enabled deployments, Osiris automatically instruments application pods with a metrics-collecting proxy deployed as a sidecar container.

For any Osiris-enabled deployment that is already scaled to a configurable minimum number of replicas (one, by default), the zeroscaler component continuously analyzes metrics from each of that deployment's pods. When the aggregated metrics reveal that all of the deployment's pods are idling, the zeroscaler scales the deployment to zero replicas.

Under normal circumstances, scaling a deployment to zero replicas poses a problem: any services that select pods from that deployment (and only that deployment) would lose all of their endpoints and become permanently unavailable. Osiris-enabled services, however, have their endpoints managed by the Osiris endpoints controller (instead of Kubernetes' built-in endpoints controller). The Osiris endpoints controller will automatically add Osiris activator endpoints to any Osiris-enabled service that has lost the rest of its endpoints.

The Osiris activator component receives traffic for Osiris-enabled services that are lacking any application endpoints. The activator initiates a scale-up of a corresponding deployment to a configurable minimum number of replicas (one, by default). When at least one application pod becomes ready, the request will be forwarded to the pod.

After the activator "reactivates" the deployment, the endpoints controller (described above) will naturally observe the availability of application endpoints for any Osiris-enabled services that select those pods and will remove activator endpoints from that service. All subsequent traffic for the service will, once again, flow directly to application pods... until a period of inactivity causes the zeroscaler to take the application offline again.

Scaling to zero and the HPA

Osiris is designed to work alongside the Horizontal Pod Autoscaler and is not meant to replace it-- it will scale your pods from n to 0 and from 0 to n, where n is a configurable minimum number of replicas (one, by default). All other scaling decisions may be delegated to an HPA, if desired.

This diagram better illustrates the different roles of Osiris, the HPA and the Cluster Autoscaler:

diagram

Setup

Prerequisites:

  • Helm (v2.11.0 or greater)
  • A running Kubernetes cluster.

Install Osiris

Osiris' Helm chart is hosted in an Azure Container Registry, which does not yet support anonymous access to charts therein. Until this is resolved, adding the Helm repository from which Osiris can be installed requires use of a shared set of read-only credentials.

Make sure helm is initialized in your running kubernetes cluster.

For more details on initializing helm, Go here

helm repo add osiris https://osiris.azurecr.io/helm/v1/repo \
  --username eae9749a-fccf-4a24-ac0d-6506fe2a6ab3 \
  --password 2fc6a721-85e4-41ca-933d-2ca02e1394c4

Installation requires use of the --devel flag to indicate pre-release versions of the specified chart are eligible for download and installation:

helm install osiris/osiris-edge \
  --name osiris \
  --namespace osiris-system \
  --devel

Usage

Osiris will not affect the normal behavior of any Kubernetes resource without explicitly being directed to do so.

To enabled the zeroscaler to scale a deployment with idling pods to zero replicas, annotate the deployment like so:

apiVersion: apps/v1
kind: Deployment
metadata:
  namespace: my-aoo
  name: my-app
  annotations:
    osiris.deislabs.io/enabled: "true"
spec:
  replicas: 1
  selector:
    matchLabels:
      app: my-app
  template:
    metadata:
      labels:
        app: nginx
    # ...
  # ...

In Kubernetes, there is no direct relationship between deployments and services. Deployments manage pods and services may select pods managed by one or more deployments. Rather than attempt to infer relationships between deployments and services and potentially impact service behavior without explicit consent, Osiris requires services to explicitly opt-in to management by the Osiris endpoints controller. Such services must also utilize an annotation to indicate which deployment should be reactivated when the activator component intercepts a request on their behalf. For example:

kind: Service
apiVersion: v1
metadata:
  namespace: my-namespace
  name: my-app
  annotations:
    osiris.deislabs.io/enabled: "true"
    osiris.deislabs.io/deployment: my-app
spec:
  selector:
    app: my-app
  # ...

Demo

Deploy the example application hello-osiris :

kubectl create -f ./example/hello-osiris.yaml

This will create an Osiris-enabled deployment and service named hello-osiris.

Get the External IP of the hello-osiris service once it appears:

kubectl get service hello-osiris -o jsonpath='{.status.loadBalancer.ingress[*].ip}'

Point your browser to "http://<EXTERNAL-IP>", and verify that hello-osiris is serving traffic.

After about 2.5 minutes, the Osiris-enabled deployment should scale to zero replicas and the one hello-osiris pod should be terminated.

Make a request again, and watch as Osiris scales the deployment back to one replica and your request is handled successfully.

Limitations

It is a specific goal of Osiris to enable greater resource efficiency within Kubernetes clusters, in general, but especially with respect to "nodeless" Kubernetes options such as Virtual Kubelet or Azure Kubernetes Service Virtual Nodes preview, however, due to known issues with those technologies, Osiris remains incompatible with them for the near term.

Contributing

Osiris follows the CNCF Code of Conduct.