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

This is a maintained fork of the original deislabs' Osiris (which has been archived and is not maintained anymore).

Note that it was forked before the HTTPS and HTTP/2 support PR, because we observed failed requests with the proxy following this change.

Introduction

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

Various types of Kubernetes resources can be Osiris-enabled using an annotation.

Osiris-enabled pods are automatically instrumented with a metrics-collecting proxy deployed as a sidecar container.

Osiris-enabled deployments or statefulSets (if already scaled to a configurable minimum number of replicas-- one by default) automatically have metrics from their pods continuously scraped and analyzed by the zeroscaler component. 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:

Setup

Prerequisites:

Helm (v2.11.0+, or v3+)
A running Kubernetes cluster.

Installation

First, add the Osiris charts repository:

helm repo add osiris https://dailymotion-oss.github.io/osiris/charts

And then install it:

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

Installation Options

Osiris global configuration is minimal - because most of it will be done by the users with annotations on the Kubernetes resources.

The following table lists the configurable parameters of the Helm chart and their default values.

Parameter	Description	Default
`zeroscaler.metricsCheckInterval`	The interval in which the zeroScaler would repeatedly track the pod http request metrics. The value is the number of seconds of the interval. Note that this can also be set on a per-deployment basis, with an annotation.	`150`

Example of installation with Helm and a custom configuration:

helm install osiris/osiris \
  --name osiris \
  --namespace osiris-system \
  --set zeroscaler.metricsCheckInterval=600

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.dm.gg/enableScaling: "true"
spec:
  replicas: 1
  selector:
    matchLabels:
      app: my-app
  template:
    metadata:
      labels:
        app: nginx
      annotations:
        osiris.dm.gg/collectMetrics: "true"
    # ...
  # ...

Note that the template for the pod also uses an annotation to enable Osiris-- in this case, it enables the metrics-collecting proxy sidecar container on all of the deployment's pods.

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.dm.gg/manageEndpoints: "true"
    osiris.dm.gg/deployment: my-app
spec:
  selector:
    app: my-app
  # ...

Configuration

Most of Osiris configuration is done with Kubernetes annotations - as seen in the Usage section.

Deployment & StatefulSet Annotations

The following table lists the supported annotations for Kubernetes Deployments and StatefulSets, and their default values.

Annotation	Description	Default
`osiris.dm.gg/enableScaling`	Enable the zeroscaler component to scrape and analyze metrics from the deployment's or statefulSet's pods and scale the deployment/statefulSet to zero when idle. Allowed values: `y`, `yes`, `true`, `on`, `1`.	no value (= disabled)
`osiris.dm.gg/minReplicas`	The minimum number of replicas to set on the deployment/statefulSet when Osiris will scale up. If you set `2`, Osiris will scale the deployment/statefulSet from `0` to `2` replicas directly. Osiris won't collect metrics from deployments/statefulSets which have more than `minReplicas` replicas - to avoid useless collections of metrics.	`1`
`osiris.dm.gg/metricsCheckInterval`	The interval in which Osiris would repeatedly track the pod http request metrics. The value is the number of seconds of the interval. Note that this value override the global value defined by the `zeroscaler.metricsCheckInterval` Helm value.	value of the `zeroscaler.metricsCheckInterval` Helm value
`osiris.dm.gg/metricsCollector`	Configure the collection of metrics for a pod. The value is a JSON object with at least a `type` string, and an optional `implementation` object. See the Metrics Scraping section for more.	`{ "type": "osiris" }`
`osiris.dm.gg/dependencies`	A list of (comma-separated) dependent deployments/statefulsets to scale down/up with this one. Format: `kind:namespace/name`. Example: `deployment:my-ns/my-deployment,statefulset:my-ns/my-statefulset`.	no value

Pod Annotations

The following table lists the supported annotations for Kubernetes Pods and their default values.

Annotation	Description	Default
`osiris.dm.gg/collectMetrics`	Enable the metrics collecting proxy to be injected as a sidecar container into this pod. This is required for metrics collection. Allowed values: `y`, `yes`, `true`, `on`, `1`.	no value (= disabled)
`osiris.dm.gg/ignoredPaths`	The list of (url) paths that should be "ignored" by Osiris. Requests to such paths won't be "counted" by the proxy. Format: comma-separated string.	no value

Service Annotations

The following table lists the supported annotations for Kubernetes Services and their default values.

Annotation	Description	Default
`osiris.dm.gg/manageEndpoints`	Enable this service's endpoints to be managed by the Osiris endpoints controller. Allowed values: `y`, `yes`, `true`, `on`, `1`.	no value (= disabled)
`osiris.dm.gg/deployment`	Name of the deployment which is behind this service. This is required to map the service with its deployment.	no value
`osiris.dm.gg/statefulset`	Name of the statefulSet which is behind this service. This is required to map the service with its statefulSet.	no value
`osiris.dm.gg/loadBalancerHostname`	Map requests coming from a specific hostname to this service. Note that if you have multiple hostnames, you can set them with different annotations, using `osiris.dm.gg/loadBalancerHostname-1`, `osiris.dm.gg/loadBalancerHostname-2`, ...	no value
`osiris.dm.gg/ingressHostname`	Map requests coming from a specific hostname to this service. If you use an ingress in front of your service, this is required to create a link between the ingress and the service. Note that if you have multiple hostnames, you can set them with different annotations, using `osiris.dm.gg/ingressHostname-1`, `osiris.dm.gg/ingressHostname-2`, ...	no value
`osiris.dm.gg/ingressDefaultPort`	Custom service port when the request comes from an ingress. Default behaviour if there are more than 1 port on the service, is to look for a port named `http`, and fallback to the port `80`. Set this if you have multiple ports and using a non-standard port with a non-standard name.	no value

Note that you might see an osiris.dm.gg/selector annotation - this is for internal use only, and you shouldn't try to set/update or delete it.

Metrics Scraping Configuration

Scraping the metrics from the pods is done automatically using Osiris provided sidecar container by default. But if you don't want to use the auto-injected sidecar container, you can also configure a custom metrics scraper, using the osiris.dm.gg/metricsCollector annotation on your deployment/statefulset.

The following scrapers are supported:

osiris

This is the default scraper, which doesn't need any configuration.

prometheus

The prometheus scraper retrieves metrics about the request count from your own prometheus endpoint. To use it, your application need to expose an endpoint with metrics in the prometheus format. You can then set the following annotation:

annotations:
  osiris.dm.gg/metricsCollector: |
    {
      "type": "prometheus",
      "implementation": {
        "port": 8080,
        "path": "/metrics",
        "requestCountMetricName": "requests"
      }
    }

The schema of the prometheus implementation configuration is:

a mandatory port integer
an optional path string - default to /metrics if not set
a mandatory requestCountMetricName string, for the name of the metric that expose the number of requests
an optional requestCountMetricLabels object, for all labels that should match the metric for request count

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.

Credit

Deislabs for their original work on Osiris.

Oriflame/osiris