Using logging and monitoring for system components

This document shows how to configure logging and monitoring for system components in GKE on VMware.

By default, Cloud Logging, Cloud Monitoring, and Google Cloud Managed Service for Prometheus are enabled.

For more information about the options, see Logging and monitoring overview.

Monitored resources

Monitored resources are how Google represents resources such as clusters, nodes, Pods, and containers. To learn more, refer to Cloud Monitoring's Monitored resource types documentation.

To query for logs and metrics, you'll need to know at least these resource labels:

  • project_id: Project ID of the cluster's logging-monitoring project. You provided this value in the stackdriver.projectID field of your cluster configuration file.

  • location: A Google Cloud region where you want to store Cloud Logging logs and Cloud Monitoring metrics. It's a good idea to choose a region that is near your on-premises data center. You provided this value during installation in the stackdriver.clusterLocation field of your cluster configuration file.

  • cluster_name: Cluster name that you chose when you created the cluster.

    You can retrieve the cluster_name value for either the admin or the user cluster by inspecting the Stackdriver custom resource:

    kubectl get stackdriver stackdriver --namespace kube-system \
    --kubeconfig CLUSTER_KUBECONFIG --output yaml | grep 'clusterName:'
    

    where

    • CLUSTER_KUBECONFIG is the path to the admin cluster's or user cluster's kubeconfig file for which the cluster name is required.

Using Cloud Logging

You don't have to take any action to enable Cloud Logging for a cluster. However, you must specify the Google Cloud project where you want to view logs. In the cluster configuration file, you specify the Google Cloud project in the stackdriver section.

You can access logs using the Logs Explorer in the Google Cloud console. For example, to access a container's logs:

  1. Open the Logs Explorer in Google Cloud console for your project.
  2. Find logs for a container by:
    1. Clicking on the top-left log catalog drop-down box and selecting Kubernetes Container.
    2. Selecting the cluster name, then the namespace, and then a container from the hierarchy.

Viewing logs for controllers in bootstrap cluster

  1. Find onprem-admin-cluster-controller / clusterapi-controllers pod name

    By default, the kind cluster name is gkectl-bootstrap-cluster.

    "ADMIN_CLUSTER_NAME"
    resource.type="k8s_container"
    resource.labels.cluster_name="gkectl-bootstrap-cluster"
    
  2. Modify the query using the pod name you find, and get the log

    resource.type="k8s_container"
    resource.labels.cluster_name="gkectl-bootstrap-cluster"
    resource.labels.pod_name="POD_NAME"
    

Using Cloud Monitoring

You don't have to take any action to enable Cloud Monitoring for a cluster. However, you must specify the Google Cloud project where you want to view metrics. In the cluster configuration file, you specify the Google Cloud project in the stackdriver section.

You can choose from over 1,500 metrics by using Metrics Explorer. To access Metrics Explorer, do the following:

  1. In the Google Cloud console, select Monitoring, or use the following button:

    Go to Monitoring

  2. Select Resources > Metrics Explorer.

You can also view metrics in dashboards in the Google Cloud console. For information about creating dashboards and viewing metrics, see Creating dashboards.

Viewing fleet-level monitoring data

For an overall view of your fleet's resource utilization using Cloud Monitoring data, including your GKE on VMware, you can use the GKE Enterprise overview in the Google Cloud console. See Use the GKE Enterprise overview to find out more.

Default Cloud Monitoring quota limits

GKE on VMware monitoring has a default limit of 6000 API calls per minute for each project. If you exceed this limit, your metrics may not be displayed. If you need a higher monitoring limit, request one through the Google Cloud console.

Using Managed Service for Prometheus

Google Cloud Managed Service for Prometheus is part of Cloud Monitoring and is available by default. The benefits of Managed Service for Prometheus include the following:

  • You can continue to use your existing Prometheus based monitoring without altering your alerts and Grafana dashboards.

  • If you use both GKE and GKE on VMware, you can use the same PromQL for metrics on all your clusters. You can also use the PROMQL tab in Metrics Explorer in the Google Cloud console.

Enabling and disabling Managed Service for Prometheus

Managed Service for Prometheus is enabled by default in GKE on VMware.

To disable Managed Service for Prometheus in a cluster:

  1. Open the Stackdriver object named stackdriver for editing:

    kubectl --kubeconfig CLUSTER_KUBECONFIG --namespace kube-system \
        edit stackdriver stackdriver
    
  2. Add the enableGMPForSystemMetrics feature gate, and set it to false:

    apiVersion: addons.gke.io/v1alpha1
    kind: Stackdriver
    metadata:
      name: stackdriver
      namespace: kube-system
    spec:
      featureGates:
        enableGMPForSystemMetrics: false
    
  3. Close your editing session.

Viewing metric data

When Managed Service for Prometheus is enabled, metrics for the following components have a different format for how they are stored and queried in Cloud Monitoring:

  • kube-apiserver
  • kube-scheduler
  • kube-controller-manager
  • kubelet and cadvisor
  • kube-state-metrics
  • node-exporter

In the new format, you can query the preceding metrics by using either PromQL or Monitoring Query Language (MQL).

PromQL example:

histogram_quantile(0.95, sum(rate(apiserver_request_duration_seconds_bucket[5m])) by (le))

To use MQL, set the monitored resource to prometheus_target, and add the Prometheus type as a suffix to the metric.

MQL example:

fetch prometheus_target
| metric 'kubernetes.io/anthos/apiserver_request_duration_seconds/histogram'
| align delta(5m)
| every 5m
| group_by [], [value_histogram_percentile: percentile(value.histogram, 95)]

Configuring Grafana dashboards with Managed Service for Prometheus

To use Grafana with metrics data from Managed Service for Prometheus, follow the steps in Query using Grafana to authenticate and configure a Grafana data source to query data from Managed Service for Prometheus.

A set of sample Grafana dashboards are provided in the anthos-samples repository on GitHub. To install the sample dashboards, do the following:

  1. Download the sample .json files:

    git clone https://github.com/GoogleCloudPlatform/anthos-samples.git
    cd anthos-samples/gmp-grafana-dashboards
    
  2. If your Grafana data source was created with a name different with Managed Service for Prometheus, change the datasource field in all the .json files:

    sed -i "s/Managed Service for Prometheus/[DATASOURCE_NAME]/g" ./*.json
    

    Replace [DATASOURCE_NAME] with the name of the data source in your Grafana that was pointed to the Prometheus frontend service.

  3. Access Grafana UI from your browser, and select + Import under the Dashboards menu.

    Navigating to dashboard import in Grafana.

  4. Either upload the .json file, or copy and paste the file content and select Load. Once the file content is successfully loaded, select Import. Optionally you can also change the dashboard name and UID before importing.

    Importing dashboard in Grafana.

  5. The imported dashboard should load successfully if your GKE on VMware and the data source are configured correctly. For example, the following screenshot shows the dashboard configured by cluster-capacity.json.

    Cluster capacity dashboard in Grafana.

Additional resources

For more information about Managed Service for Prometheus, see the following:

Using Prometheus and Grafana

Starting in version 1.16, Prometheus and Grafana are not available in newly created clusters. We recommend that you use Managed Service for Prometheus as a replacement for in-cluster monitoring.

If you upgrade a 1.15 cluster that has Prometheus and Grafana enabled to 1.16, Prometheus and Grafana will continue to work as is, but they will not be updated or given security patches.

If you want to delete all the Prometheus and Grafana resources after upgrading to 1.16, run the following command:

kubectl --kubeconfig KUBECONFIG delete -n kube-system \
    statefulsets,services,configmaps,secrets,serviceaccounts,clusterroles,clusterrolebindings,certificates,deployments \
    -l addons.gke.io/legacy-pg=true

As an alternative to using the Prometheus and Grafana components included in earlier versions of GKE on VMware, you can switch to an open source community version of Prometheus and Grafana.

Known issue

In user clusters, Prometheus and Grafana get automatically disabled during upgrade. However, the configuration and metrics data are not lost.

To work around this issue, after the upgrade, open monitoring-sample for editing and set enablePrometheus to true.

Accessing monitoring metrics from Grafana dashboards

Grafana displays metrics gathered from your clusters. To view these metrics, you need to access Grafana's dashboards:

  1. Get the name of the Grafana Pod running in a user cluster's kube-system namespace:

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] -n kube-system get pods

    where [USER_CLUSTER_KUBECONFIG] is the user cluster's kubeconfig file.

  2. The Grafana Pod has an HTTP server listening on TCP localhost port 3000. Forward a local port to port 3000 in the Pod, so that you can view Grafana's dashboards from a web browser.

    For example, suppose the name of the Pod is grafana-0. To forward port 50000 to port 3000 in the Pod, enter this command::

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] -n kube-system port-forward grafana-0 50000:3000
  3. From a web browser, navigate to http://localhost:50000.

  4. On the login page, enter admin for username and password.

  5. If login is successful, you will see a prompt to change the password. After you have changed the default password, the user cluster's Grafana Home Dashboard should load.

  6. To access other dashboards, click the Home drop-down menu in the top-left corner of the page.

For an example of using Grafana, see Create a Grafana dashboard.

Accessing alerts

Prometheus Alertmanager collects alerts from the Prometheus server. You can view these alerts in a Grafana dashboard. To view the alerts, you need to access the dashboard:

  1. The container in the alertmanager-0 Pod listens on TCP port 9093. Forward a local port to port 9093 in the Pod:

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] port-forward \
       -n kube-system alertmanager-0 50001:9093
  2. From a web browser, navigate to http://localhost:50001.

Changing Prometheus Alertmanager configuration

You can change Prometheus Alertmanager's default configuration by editing your user cluster's monitoring.yaml file. You should do this if you want to direct alerts to a specific destination, rather than keep them in the dashboard. You can learn how to configure Alertmanager in Prometheus' Configuration documentation.

To change the Alertmanager configuration, perform the following steps:

  1. Make a copy of the user cluster's monitoring.yaml manifest file:

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] -n kube-system \
       get monitoring monitoring-sample -o yaml > monitoring.yaml
  2. To configure Alertmanager, make changes to the fields under spec.alertmanager.yml. When you're finished, save the changed manifest.

  3. Apply the manifest to your cluster:

    kubectl apply --kubeconfig [USER_CLUSTER_KUBECONIFG] -f monitoring.yaml

Create a Grafana dashboard

You've deployed an application that exposes a metric, verified that the metric is exposed, and verified that Prometheus scrapes the metric. Now you can add the application-level metric to a custom Grafana dashboard.

To create a Grafana dashboard, perform the following steps:

  1. If necessary, gain access to Grafana.
  2. From the Home Dashboard, click the Home drop-down menu in the top-left corner of the page.
  3. From the right-side menu, click New dashboard.
  4. From the New panel section, click Graph. An empty graph dashboard appears.
  5. Click Panel title, then click Edit. The bottom Graph panel opens to the Metrics tab.
  6. From the Data Source drop-down menu, select user. Click Add query, and enter foo in the search field.
  7. Click the Back to dashboard button in the top-right corner of the screen. Your dashboard is displayed.
  8. To save the dashboard, click Save dashboard in the top-right corner of the screen. Choose a name for the dashboard, then click Save.

Disabling Prometheus and Grafana

Starting from version 1.16, Prometheus and Grafana are no longer controlled by the enablePrometheus field in the monitoring-sample object. See Using Prometheus and Grafana for details.

Example: Adding application-level metrics to a Grafana dashboard

The following sections walk you through adding metrics for an application. In this section, you complete the following tasks:

  • Deploy an example application that exposes a metric called foo.
  • Verify that Prometheus exposes and scrapes the metric.
  • Create a custom Grafana dashboard.

Deploy the example application

The example application runs in a single Pod. The Pod's container exposes a metric, foo, with a constant value of 40.

Create the following Pod manifest, pro-pod.yaml:

apiVersion: v1
kind: Pod
metadata:
  name: prometheus-example
  annotations:
    prometheus.io/scrape: 'true'
    prometheus.io/port: '8080'
    prometheus.io/path: '/metrics'
spec:
  containers:
  - image: registry.k8s.io/prometheus-dummy-exporter:v0.1.0
    name: prometheus-example
    command:
    - /bin/sh
    - -c
    - ./prometheus_dummy_exporter --metric-name=foo --metric-value=40 --port=8080

Then apply the Pod manifest to your user cluster:

kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] apply -f pro-pod.yaml

Verify that the metric is exposed and scraped

  1. The container in the prometheus-example pod listens on TCP port 8080. Forward a local port to port 8080 in the Pod:

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] port-forward prometheus-example 50002:8080
  2. To verify that the application exposes the metric, run the following command:

    curl localhost:50002/metrics | grep foo
    

    The command returns the following output:

    # HELP foo Custom metric
    # TYPE foo gauge
    foo 40
  3. The container in the prometheus-0 Pod listens on TCP port 9090. Forward a local port to port 9090 in the Pod:

    kubectl --kubeconfig [USER_CLUSTER_KUBECONFIG] port-forward prometheus-0 50003:9090
  4. To verify that Prometheus is scraping the metric, navigate to http://localhost:50003/targets, which should take you to the prometheus-0 Pod under the prometheus-io-pods target group.

  5. To view metrics in Prometheus, navigate to http://localhost:50003/graph. From the search field, enter foo, then click Execute. The page should display the metric.

Configuring the Stackdriver custom resource

When you create a cluster, GKE on VMware automatically creates a Stackdriver custom resource. You can edit the spec in the custom resource to override the default values for CPU and memory requests and limits for a Stackdriver component, and you can separately override the default storage size and storage class.

Override default values for requests and limits for CPU and memory

To override these defaults, do the following:

  1. Open your Stackdriver custom resource in a command line editor:

    kubectl --kubeconfig=KUBECONFIG -n kube-system edit stackdriver stackdriver

    where KUBECONFIG is the path to your kubeconfig file for the cluster. This can be either an admin cluster or user cluster.

  2. In the Stackdriver custom resource, add the resourceAttrOverride field under the spec section:

    resourceAttrOverride:
          POD_NAME_WITHOUT_RANDOM_SUFFIX/CONTAINER_NAME:
            LIMITS_OR_REQUESTS:
              RESOURCE: RESOURCE_QUANTITY

    Note that the resourceAttrOverride field overrides all existing default limits and requests for the component you specify. The following components are supported by resourceAttrOverride:

    • gke-metrics-agent/gke-metrics-agent
    • stackdriver-log-forwarder/stackdriver-log-forwarder
    • stackdriver-metadata-agent-cluster-level/metadata-agent
    • node-exporter/node-exporter
    • kube-state-metrics/kube-state-metrics

An example file looks like the following

apiVersion: addons.gke.io/v1alpha1
    kind: Stackdriver
    metadata:
      name: stackdriver
      namespace: kube-system
    spec:
      projectID: my-project
      clusterName: my-cluster
      clusterLocation: us-west-1a
      resourceAttrOverride:
        gke-metrics-agent/gke-metrics-agent:
          requests:
            cpu: 110m
            memory: 240Mi
          limits:
            cpu: 200m
            memory: 4.5Gi

  1. Save changes and quit your command line editor.

  2. Check the health of your Pods:

    kubectl --kubeconfig=KUBECONFIG -n kube-system get pods | grep gke-metrics-agent

    For example, a healthy Pod looks like the following:

    gke-metrics-agent-4th8r                                1/1     Running   0          5d19h
  3. Check the Pod spec of the component to make sure the resources are set correctly.

    kubectl --kubeconfig=KUBECONFIG -n kube-system describe pod POD_NAME

    where POD_NAME is the name of the Pod you just changed. For example, stackdriver-prometheus-k8s-0

    The response looks like the following:

      Name:         gke-metrics-agent-4th8r
      Namespace:    kube-system
      ...
      Containers:
        gke-metrics-agent:
          Limits:
            cpu: 200m
            memory: 4.5Gi
          Requests:
            cpu: 110m
            memory: 240Mi
          ...
          

Override storage size defaults

To override these defaults, do the following:

  1. Open your Stackdriver custom resource in a command line editor:

    kubectl --kubeconfig=KUBECONFIG -n kube-system edit stackdriver stackdriver
  2. Add the storageSizeOverride field under the spec section. You can use the component stackdriver-prometheus-k8s or stackdriver-prometheus-app. The section takes this format:

    storageSizeOverride:
    STATEFULSET_NAME: SIZE
    

    This example uses the statefulset stackdriver-prometheus-k8s and size 120Gi.

    apiVersion: addons.gke.io/v1alpha1
    kind: Stackdriver
    metadata:
      name: stackdriver
      namespace: kube-system
    spec:
      projectID: my-project
      clusterName: my-cluster
      clusterLocation: us-west-1a
      storageSizeOverride:
        stackdriver-prometheus-k8s: 120Gi
      
  3. Save, and quit your command line editor.

  4. Check the health of your Pods:

    kubectl --kubeconfig=KUBECONFIG -n kube-system get pods | grep stackdriver
    For example, a healthy Pod looks like the following:
    stackdriver-prometheus-k8s-0                                2/2     Running   0          5d19h
  5. Check the Pod spec of the component to make sure the storage size is correctly overridden.

    kubectl --kubeconfig=KUBECONFIG -n kube-system describe statefulset STATEFULSET_NAME

    The response looks like the following:

    Volume Claims:
     Name:          my-statefulset-persistent-volume-claim
     StorageClass:  my-storage-class
     Labels:
     Annotations:
     Capacity:      120Gi
     Access Modes:  [ReadWriteOnce]          

Override storage class defaults

Prerequisite

You must first create a StorageClass you want to use.

To override the default storage class for persistent volumes claimed by logging and monitoring components:

  1. Open your Stackdriver custom resource in a command line editor:

    kubectl --kubeconfig=KUBECONFIG -n kube-system edit stackdriver stackdriver

    where KUBECONFIG is the path to your kubeconfig file for the cluster. This can be either an admin cluster or user cluster.

  2. Add the storageClassName field under the spec section:

    storageClassName: STORAGECLASS_NAME

    Note that the storageClassName field overrides the existing default storage class, and applies to all logging and monitoring components with persistent volumes claimed. An example file looks like the following:

    apiVersion: addons.gke.io/v1alpha1
    kind: Stackdriver
    metadata:
      name: stackdriver
      namespace: kube-system
    spec:
      projectID: my-project
      clusterName: my-cluster
      clusterLocation: us-west-1a
      proxyConfigSecretName: my-secret-name
      enableVPC: 
      optimizedMetrics: true
      storageClassName: my-storage-class
  3. Save changes.

  4. Check the health of your Pods:

    kubectl --kubeconfig=KUBECONFIG -n kube-system get pods | grep stackdriver

    For example, a healthy Pod looks like the following:

    stackdriver-prometheus-k8s-0                                1/1     Running   0          5d19h
  5. Check the Pod spec of a component to make sure the storage class is set correctly.

    kubectl --kubeconfig=KUBECONFIG -n kube-system describe statefulset STATEFULSET_NAME

    For example, using the stateful set stackdriver-prometheus-k8s, the response looks like the following:

    Volume Claims:
     Name:          stackdriver-prometheus-data
     StorageClass:  my-storage-class
     Labels:
     Annotations:
     Capacity:      120Gi
     Access Modes:  [ReadWriteOnce]          

Disable optimized metrics

By default, the metrics agents running in the cluster collect and report an optimized set of container, kubelet and kube-state-metrics metrics to Stackdriver. If you require additional metrics, we recommend that you find a replacement from the list of GKE Enterprise metrics.

Here are some examples of replacements you might use:

Disabled metric Replacements
kube_pod_start_time container/uptime
kube_pod_container_resource_requests container/cpu/request_cores
container/memory/request_bytes
kube_pod_container_resource_limits container/cpu/limit_cores
container/memory/limit_bytes

To disable the optimized kube-state-metrics metrics default setting (not recommended), do the following:

  1. Open your Stackdriver custom resource in a command line editor:

    kubectl --kubeconfig=KUBECONFIG -n kube-system edit stackdriver stackdriver

    where KUBECONFIG is the path to your kubeconfig file for the cluster. This can be either an admin cluster or user cluster.

  2. Set the optimizedMetrics field to false:

    apiVersion: addons.gke.io/v1alpha1
    kind: Stackdriver
    metadata:
      name: stackdriver
      namespace: kube-system
    spec:
      projectID: my-project
      clusterName: my-cluster
      clusterLocation: us-west-1a
      proxyConfigSecretName: my-secret-name
      enableVPC: 
      optimizedMetrics: false
      storageClassName: my-storage-class
  3. Save changes, and quit your command line editor.

Known issue: Cloud Monitoring error condition

(Issue ID 159761921)

Under certain conditions, the default Cloud Monitoring pod, deployed by default in each new cluster, can become unresponsive. When clusters are upgraded, for example, storage data can become corrupted when pods in statefulset/prometheus-stackdriver-k8s are restarted.

Specifically, monitoring pod stackdriver-prometheus-k8s-0 can be caught in a loop when corrupted data prevents prometheus-stackdriver-sidecar writing to the cluster storage PersistentVolume.

You can manually diagnose and recover the error following the steps below.

Diagnosing the Cloud Monitoring failure

When the monitoring pod has failed, the logs will report the following:

{"log":"level=warn ts=2020-04-08T22:15:44.557Z caller=queue_manager.go:534 component=queue_manager msg=\"Unrecoverable error sending samples to remote storage\" err=\"rpc error: code = InvalidArgument desc = One or more TimeSeries could not be written: One or more points were written more frequently than the maximum sampling period configured for the metric.: timeSeries[0-114]; Unknown metric: kubernetes.io/anthos/scheduler_pending_pods: timeSeries[196-198]\"\n","stream":"stderr","time":"2020-04-08T22:15:44.558246866Z"}

{"log":"level=info ts=2020-04-08T22:15:44.656Z caller=queue_manager.go:229 component=queue_manager msg=\"Remote storage stopped.\"\n","stream":"stderr","time":"2020-04-08T22:15:44.656798666Z"}

{"log":"level=error ts=2020-04-08T22:15:44.663Z caller=main.go:603 err=\"corruption after 29032448 bytes: unexpected non-zero byte in padded page\"\n","stream":"stderr","time":"2020-04-08T22:15:44.663707748Z"}

{"log":"level=info ts=2020-04-08T22:15:44.663Z caller=main.go:605 msg=\"See you next time!\"\n","stream":"stderr","time":"2020-04-08T22:15:44.664000941Z"}

Recovering from the Cloud Monitoring error

To recover Cloud Monitoring manually:

  1. Stop cluster monitoring. Scale down the stackdriver operator to prevent monitoring reconciliation:

    kubectl --kubeconfig /ADMIN_CLUSTER_KUBCONFIG --namespace kube-system scale deployment stackdriver-operator --replicas 0

  2. Delete the monitoring pipeline workloads:

    kubectl --kubeconfig /ADMIN_CLUSTER_KUBCONFIG --namespace kube-system delete statefulset stackdriver-prometheus-k8s

  3. Delete the monitoring pipeline PersistentVolumeClaims (PVCs):

    kubectl --kubeconfig /ADMIN_CLUSTER_KUBCONFIG --namespace kube-system delete pvc -l app=stackdriver-prometheus-k8s

  4. Restart cluster monitoring. Scale up the stackdriver operator to reinstall a new monitoring pipeline and resume reconciliation:

    kubectl --kubeconfig /ADMIN_CLUSTER_KUBCONFIG --namespace kube-system scale deployment stackdriver-operator --replicas=1