Collect MKE cluster metrics with Prometheus

Collect MKE cluster metrics with Prometheus

Prometheus is an open-source systems monitoring and alerting toolkit. You can configure Docker as a Prometheus target. This topic shows you how to configure Docker, set up Prometheus to run as a Docker container, and monitor your Docker instance using Prometheus.

In UCP 3.0, Prometheus servers were standard containers. In MKE 3.1, Prometheus runs as a Kubernetes deployment. By default, this will be a DaemonSet that runs on every manager node. One benefit of this change is you can set the DaemonSet to not schedule on any nodes, which effectively disables Prometheus if you don’t use the MKE web interface.

The data is stored locally on disk for each Prometheus server, so data is not replicated on new managers or if you schedule Prometheus to run on a new node. Metrics are not kept longer than 24 hours.

Events, logs, and metrics are sources of data that provide observability of your cluster. Metrics monitors numerical data values that have a time-series component. There are several sources from which metrics can be derived, each providing different kinds of meaning for a business and its applications.

MKE provides a base set of metrics that gets you running and into production without having to rely on external or third-party tools. Docker strongly encourages the use of additional monitoring to provide more comprehensive visibility into your specific Docker environment, but recognizes the need for a basic set of metrics built into the product. The following are examples of these metrics:

Business metrics

These are high-level aggregate metrics that typically combine technical, financial, and organizational data to create metrics for business leaders of the IT infrastructure. Some examples of business metrics might be:

  • Company or division-level application downtime

  • Aggregate resource utilization

  • Application resource demand growth

Application metrics

These are metrics about domain of APM tools like AppDynamics or DynaTrace and provide metrics about the state or performance of the application itself.

  • Service state metrics

  • Container platform metrics

  • Host infrastructure metrics

Service state metrics

These are metrics about the state of services running on the container platform. These types of metrics have very low cardinality, meaning the values are typically from a small fixed set of possibilities, commonly binary.

  • Application health

  • Convergence of K8s deployments and Swarm services

  • Cluster load by number of services or containers or pods

Web UI disk usage metrics, including free space, only reflect the Docker managed portion of the filesystem: /var/lib/docker. To monitor the total space available on each filesystem of a MKE worker or manager, you must deploy a third party monitoring solution to monitor the operating system.

Deploy Prometheus on worker nodes

MKE deploys Prometheus by default on the manager nodes to provide a built-in metrics backend. For cluster sizes over 100 nodes or for use cases where scraping metrics from the Prometheus instances are needed, we recommend that you deploy Prometheus on dedicated worker nodes in the cluster.

To deploy Prometheus on worker nodes in a cluster:

  1. Begin by sourcing an admin bundle.

  2. Verify that ucp-metrics pods are running on all managers.

    $ kubectl -n kube-system get pods -l k8s-app=ucp-metrics -o wide NAME
    READY     STATUS    RESTARTS   AGE       IP              NODE
    ucp-metrics-hvkr7   3/3       Running   0          4h   3a724a-0
  3. Add a Kubernetes node label to one or more workers. Here we add a label with key “ucp-metrics” and value “” to a node with name “3a724a-1”.

    $ kubectl label node 3a724a-1 ucp-metrics=
    node "test-3a724a-1" labeled

    SELinux Prometheus Deployment for MKE 3.1.0, 3.1.1, and 3.1.2

    If you are using SELinux, you must label your ucp-node-certs directories properly on your worker nodes before you move the ucp-metrics workload to them. To run ucp-metrics on a worker node, update the ucp-node-certs label by running sudo chcon -R system_u:object_r:container_file_t:s0 /var/lib/docker/volumes/ucp-node-certs/_data.

  4. Patch the ucp-metrics DaemonSet’s nodeSelector using the same key and value used for the node label. This example shows the key “ucp-metrics” and the value “”.

    $ kubectl -n kube-system patch daemonset ucp-metrics --type json -p
    '[{"op": "replace", "path": "/spec/template/spec/nodeSelector", "value":
    {"ucp-metrics": ""}}]' daemonset "ucp-metrics" patched
  5. Observe that ucp-metrics pods are running only on the labeled workers.

    $ kubectl -n kube-system get pods -l k8s-app=ucp-metrics -o wide NAME
    READY     STATUS        RESTARTS   AGE       IP              NODE
    ucp-metrics-88lzx   3/3       Running       0          12s    3a724a-1 ucp-metrics-hvkr7   3/3       Terminating   0
    4h   3a724a-0

Configure external Prometheus to scrape metrics from MKE

To configure your external Prometheus server to scrape metrics from Prometheus in MKE:

  1. Begin by sourcing an admin bundle.

  2. Create a Kubernetes secret containing your bundle’s TLS material.

    (cd $DOCKER_CERT_PATH && kubectl create secret generic prometheus --from-file=ca.pem --from-file=cert.pem --from-file=key.pem)
  3. Create a Prometheus deployment and ClusterIP service using YAML as follows.

    On AWS with Kube’s cloud provider configured, you can replace ClusterIP with LoadBalancer in the service YAML then access the service through the load balancer. If running Prometheus external to MKE, change the following domain for the inventory container in the Prometheus deployment from ucp-controller.kube-system.svc.cluster.local to an external domain to access MKE from the Prometheus node.

    kubectl apply -f - <<EOF
    apiVersion: v1
    kind: ConfigMap
      name: prometheus
      prometheus.yaml: |
          scrape_interval: 10s
        - job_name: 'ucp'
            ca_file: /bundle/ca.pem
            cert_file: /bundle/cert.pem
            key_file: /bundle/key.pem
            server_name: proxy.local
          scheme: https
          - files:
            - /inventory/inventory.json
    apiVersion: apps/v1
    kind: Deployment
      name: prometheus
      replicas: 2
          app: prometheus
            app: prometheus
          - name: inventory
            image: alpine
            command: ["sh", "-c"]
            - apk add --no-cache curl &&
              while :; do
                curl -Ss --cacert /bundle/ca.pem --cert /bundle/cert.pem --key /bundle/key.pem --output /inventory/inventory.json https://ucp-controller.kube-system.svc.cluster.local/metricsdiscovery;
                sleep 15;
            - name: bundle
              mountPath: /bundle
            - name: inventory
              mountPath: /inventory
          - name: prometheus
            image: prom/prometheus
            command: ["/bin/prometheus"]
            - --config.file=/config/prometheus.yaml
            - --storage.tsdb.path=/prometheus
            - --web.console.libraries=/etc/prometheus/console_libraries
            - --web.console.templates=/etc/prometheus/consoles
            - name: bundle
              mountPath: /bundle
            - name: config
              mountPath: /config
            - name: inventory
              mountPath: /inventory
          - name: bundle
              secretName: prometheus
          - name: config
              name: prometheus
          - name: inventory
              medium: Memory
    apiVersion: v1
    kind: Service
      name: prometheus
      - port: 9090
        targetPort: 9090
        app: prometheus
      sessionAffinity: ClientIP
  4. Determine the service ClusterIP.

    $ kubectl get service prometheus
    NAME         TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)    AGE
    prometheus   ClusterIP   <none>        9090/TCP   1h
  5. Forward port 9090 on the local host to the ClusterIP. The tunnel created does not need to be kept alive and is only intended to expose the Prometheus UI.

    ssh -L 9090: ANY_NODE
  6. Visit to explore the MKE metrics being collected by Prometheus.