Branch: master
Find file Copy path
2e82759 Nov 20, 2018
5 contributors

Users who have contributed to this file

@DirectXMan12 @tmatias @zioproto @natefox @aschepis
372 lines (286 sloc) 12.7 KB


This walkthrough will go over the basics of setting up the Prometheus adapter on your cluster and configuring an autoscaler to use application metrics sourced from the adapter.


Cluster Configuration

Before getting started, ensure that the main components of your cluster are configured for autoscaling on custom metrics. As of Kubernetes 1.7, this requires enabling the aggregation layer on the API server and configuring the controller manager to use the metrics APIs via their REST clients.

Detailed instructions can be found in the Kubernetes documentation under Horizontal Pod Autoscaling.

Make sure that you've properly configured metrics-server (as default in Kubernetes 1.9+), or enabling custom metrics autoscaling support will disable CPU autoscaling support.

Note that most of the API versions in this walkthrough target Kubernetes 1.9+. Note that current versions of the adapter only work with Kubernetes 1.8+. Version 0.1.0 works with Kubernetes 1.7, but is significantly different.

Binaries and Images

In order to follow this walkthrough, you'll need container images for Prometheus and the custom metrics adapter.

The Prometheus Operator, makes it easy to get up and running with Prometheus. This walkthrough will assume you're planning on doing that -- if you've deployed it by hand instead, you'll need to make a few adjustments to the way you expose metrics to Prometheus.

The adapter has different images for each arch, which can be found at directxman12/k8s-prometheus-adapter-${ARCH}. For instance, if you're on an x86_64 machine, use the directxman12/k8s-prometheus-adapter-amd64 image.

If you're feeling adventurous, you can build the latest version of the custom metrics adapter by running make docker-build or make build-local-image.

Special thanks to @luxas for providing the demo application for this walkthrough.

The Scenario

Suppose that you've written some new web application, and you know it's the next best thing since sliced bread. It's ready to unveil to the world... except you're not sure that just one instance will handle all the traffic once it goes viral. Thankfully, you've got Kubernetes.

Deploy your app into your cluster, exposed via a service so that you can send traffic to it and fetch metrics from it:

apiVersion: apps/v1
kind: Deployment
  name: sample-app
    app: sample-app
  replicas: 1
      app: sample-app
        app: sample-app
      - image: luxas/autoscale-demo:v0.1.2
        name: metrics-provider
      - name: http
        port: 8080
$ kubectl create -f sample-app.deploy.yaml
$ kubectl create service clusterip sample-app --tcp=80:8080

Now, check your app, which exposes metrics and counts the number of accesses to the metrics page via the http_requests_total metric:

$ curl http://$(kubectl get service sample-app -o jsonpath='{ .spec.clusterIP }')/metrics

Notice that each time you access the page, the counter goes up.

Now, you'll want to make sure you can autoscale your application on that metric, so that you're ready for your launch. You can use a HorizontalPodAutoscaler like this to accomplish the autoscaling:

kind: HorizontalPodAutoscaler
apiVersion: autoscaling/v2beta1
  name: sample-app
    # point the HPA at the sample application
    # you created above
    apiVersion: apps/v1
    kind: Deployment
    name: sample-app
  # autoscale between 1 and 10 replicas
  minReplicas: 1
  maxReplicas: 10
  # use a "Pods" metric, which takes the average of the
  # given metric across all pods controlled by the autoscaling target
  - type: Pods
      # use the metric that you used above: pods/http_requests
      metricName: http_requests
      # target 500 milli-requests per second,
      # which is 1 request every two seconds
      targetAverageValue: 500m

If you try creating that now (and take a look at your controller-manager logs), you'll see that the that the HorizontalPodAutoscaler controller is attempting to fetch metrics from /apis/*/http_requests?selector=app%3Dsample-app, but right now, nothing's serving that API.

Before you can autoscale your application, you'll need to make sure that Kubernetes can read the metrics that your application exposes.

Launching Prometheus and the Adapter

In order to expose metrics beyond CPU and memory to Kubernetes for autoscaling, you'll need an "adapter" that serves the custom metrics API. Since you've got Prometheus metrics, it makes sense to use the Prometheus adapter to serve metrics out of Prometheus.

Launching Prometheus

First, you'll need to deploy the Prometheus Operator. Check out the getting started guide for the Operator to deploy a copy of Prometheus.

This walkthrough assumes that Prometheus is deployed in the prom namespace. Most of the sample commands and files are namespace-agnostic, but there are a few commands or pieces of configuration that rely on that namespace. If you're using a different namespace, simply substitute that in for prom when it appears.

Monitoring Your Application

In order to monitor your application, you'll need to set up a ServiceMonitor pointing at the application. Assuming you've set up your Prometheus instance to use ServiceMonitors with the app: sample-app label, create a ServiceMonitor to monitor the app's metrics via the service:

kind: ServiceMonitor
  name: sample-app
    app: sample-app
      app: sample-app
  - port: http
$ kubectl create -f service-monitor.yaml

Now, you should see your metrics appear in your Prometheus instance. Look them up via the dashboard, and make sure they have the namespace and pod labels.

Launching the Adapter

Now that you've got a running copy of Prometheus that's monitoring your application, you'll need to deploy the adapter, which knows how to communicate with both Kubernetes and Prometheus, acting as a translator between the two.

The deploy/manifests directory contains the appropriate files for creating the Kubernetes objects to deploy the adapter.

See the deployment README for more information about the steps to deploy the adapter. Note that if you're deploying on a non-x86_64 (amd64) platform, you'll need to change the image field in the Deployment to be the appropriate image for your platform.

The default adapter configuration should work for this walkthrough and a standard Prometheus Operator configuration, but if you've got custom relabelling rules, or your labels above weren't exactly namespace and pod, you may need to edit the configuration in the ConfigMap. The configuration walkthrough provides an overview of how configuration works.

The Registered API

As part of the creation of the adapter Deployment and associated objects (performed above), we registered the API with the API aggregator (part of the main Kubernetes API server).

The API is registered as, and you can find more information about aggregation at Concepts: Aggregation.

Double-Checking Your Work

With that all set, your custom metrics API should show up in discovery.

Try fetching the discovery information for it:

$ kubectl get --raw /apis/

Since you've set up Prometheus to collect your app's metrics, you should see a pods/http_request resource show up. This represents the http_requests_total metric, converted into a rate, aggregated to have one datapoint per pod. Notice that this translates to the same API that our HorizontalPodAutoscaler was trying to use above.

You can check the value of the metric using kubectl get --raw, which sends a raw GET request to the Kubernetes API server, automatically injecting auth information:

$ kubectl get --raw "/apis/*/http_requests?selector=app%3Dsample-app"

Because of the adapter's configuration, the cumulative metric http_requests_total has been converted into a rate metric, pods/http_requests, which measures requests per second over a 2 minute interval. The value should currently be close to zero, since there's no traffic to your app, except for the regular metrics collection from Prometheus.

Try generating some traffic using cURL a few times, like before:

$ curl http://$(kubectl get service sample-app -o jsonpath='{ .spec.clusterIP }')/metrics

Now, if you fetch the metrics again, you should see an increase in the value. If you leave it alone for a bit, the value should go back down again.

Quantity Values

Notice that the API uses Kubernetes-style quantities to describe metric values. These quantities use SI suffixes instead of decimal points. The most common to see in the metrics API is the m suffix, which means milli-units, or 1000ths of a unit. If your metric is exactly a whole number of units on the nose, you might not see a suffix. Otherwise, you'll probably see an m suffix to represent fractions of a unit.

For example, here, 500m would be half a request per second, 10 would be 10 requests per second, and 10500m would be 10.5 requests per second.

Troubleshooting Missing Metrics

If the metric does not appear, or is not registered with the right resources, you might need to modify your adapter configuration, as mentioned above. Check your labels via the Prometheus dashboard, and then modify the configuration appropriately.

As noted in the main README, you'll need to also make sure your metrics relist interval is at least your Prometheus scrape interval. If it's less that that, you'll see metrics periodically appear and disappear from the adapter.


Now that you finally have the metrics API set up, your HorizontalPodAutoscaler should be able to fetch the appropriate metric, and make decisions based on it.

If you didn't create the HorizontalPodAutoscaler above, create it now:

$ kubectl create -f sample-app-hpa.yaml

Wait a little bit, and then examine the HPA:

$ kubectl describe hpa sample-app

You should see that it succesfully fetched the metric, but it hasn't tried to scale, since there's not traffic.

Since your app is going to need to scale in response to traffic, generate some via cURL like above:

$ curl http://$(kubectl get service sample-app -o jsonpath='{ .spec.clusterIP }')/metrics

Recall from the configuration at the start that you configured your HPA to have each replica handle 500 milli-requests, or 1 request every two seconds (ok, so maybe you still have some performance issues to work out before your beta period ends). Thus, if you generate a few requests, you should see the HPA scale up your app relatively quickly.

If you describe the HPA again, you should see that the last observed metric value roughly corresponds to your rate of requests, and that the HPA has recently scaled your app.

Now that you've got your app autoscaling on the HTTP requests, you're all ready to launch! If you leave the app alone for a while, the HPA should scale it back down, so you can save precious budget for the launch party.

Next Steps

For more information on how the HPA controller consumes different kinds of metrics, take a look at the HPA walkthrough.

Also try exposing a non-cumulative metric from your own application, or scaling on application on a metric provided by another application by setting different labels or using the Object metric source type.

For more information on how metrics are exposed by the Prometheus adapter, see config documentation, and check the default configuration.