Integrate an Application with Prometheus Operator and Package with a Helm Chart

In the previous posts, we saw:

how to implement metrics in applications
how to run the monitoring stack locally
how to test and debug metrics generated by a simple Python Flask application

In this post we will:

use Kubernetes Custom Resources to …


This content originally appeared on DEV Community and was authored by Julien Acroute

In the previous posts, we saw:

  • how to implement metrics in applications
  • how to run the monitoring stack locally
  • how to test and debug metrics generated by a simple Python Flask application

In this post we will:

  • use Kubernetes Custom Resources to integrate our application with the Prometheus Operator
  • define some alerts based on the metrics generated by the application
  • deploy a custom dashboard in Grafana
  • package everything in a Helm chart, including a Grafana dashboard

Connect Prometheus to your Application

Prometheus will retrieve metrics from Pods with a /metrics HTTP endpoint. If the Prometheus Operator is deployed in your Kubernetes Cluster, the discovery of the Pods is done by deploying one of the following custom Kubernetes objects:

Using ServiceMonitor

When a ServiceMonitor object is deployed, Prometheus will create one target per address defined in the Endpoints object linked to the Service. This means every Pod is in a ready status used by the Service.

For example, if you have the following Service and Deployment in your cluster:

apiVersion: v1
kind: Service
metadata:
  name: webapp
  labels:
    component: backend
    instance: app
    name: containers-my-app
  namespace: my-app
spec:
  type: ClusterIP
  ports:
  - name: http
    port: 80
    protocol: TCP
    targetPort: webapp
  selector:
    component: backend
    instance: app
    name: containers-my-app

apiVersion: apps/v1
kind: Deployment
metadata:
  name: webapp
  labels:
    component: backend
    instance: app
    name: containers-my-app
  namespace: my-app
spec:
  selector:
    matchLabels:
      component: backend
      instance: app
      name: containers-my-app
  template:
    metadata:
      labels:
        component: backend
        instance: app
        name: containers-my-app
    spec:
      containers:
      - name: app
        image: ghcr.io/camptocamp/course_docker_backend:python
        ports:
        - containerPort: 8080
          name: webapp

As defined in the Deployment’s spec.template.metadata.labels field, Pods will have the following labels:

  • component: backend
  • instance: app
  • name: containers-my-app

The Service has a selector that matches labels of the Pod. Therefore the Service will load balance traffic to Pods deployed by the Deployment.

ServiceMonitor objects also use a selector to discover which Services need to be monitored. Prometheus will scrape metrics from every Pods behind selected Services.

For example, to retrieve metrics from our Pods, we can deploy the following ServiceMonitor object:

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: webapp
  labels:
    component: backend
    instance: app
    name: containers-my-app
    release: prom
  namespace: my-app
spec:
  namespaceSelector:
    matchNames:
    - my-app
  selector:
    matchLabels:
      component: backend
      instance: app
      name: containers-my-app
  endpoints:
  - port: http

Prometheus Operator will search for Services:

  • in the my-app namespace,
  • with the following labels:
    • component: backend
    • instance: app
    • name: containers-my-app
  • with a port named: http

It will then use the Service selector to find Pods. As a result, one target per Pod will be created in the Prometheus configuration.

Using PodMonitor

PodMonitor objects use a selector to find Pods directly. No Service needs to be deployed.

For our Pods, we can use the following PodMonitor:

apiVersion: monitoring.coreos.com/v1
kind: PodMonitor
metadata:
  name: webapp
  labels:
    acomponent: backend
    instance: app
    name: containers-my-app
    release: prom
  namespace: my-app
spec:
  namespaceSelector:
    matchNames:
    - my-app
  selector:
    matchLabels:
      component: backend
      instance: app
      name: containers-my-app
  podMetricsEndpoints:
  - port: webapp

The Prometheus operator will search for Pods:

  • in the my-app namespace,
  • with the following labels:
    • component: backend
    • instance: app
    • name: containers-my-app
  • with a port named: webapp

For each Pod, a new target will be added to the Prometheus configuration.

Configure Alerts

Why use Alerts?

Gathering and storing metrics is very useful to investigate when something goes wrong. But there are often some modifications of one or a combination of metrics before a service becomes completely unusable.

A common example of this is remaining free disk space. Fixing hard thresholds with arbitrary values on disk space is usually inefficient (you might actually end up with 95%, 100% and 101% thresholds). What needs to be monitored is actually the estimated time left until the disk is full, which can be obtained by running a time regression on the metric.

Some other examples:

  • For an online store, having no purchases during the afternoon is unusual, maybe there is an issue that blocks users in the payment process.
  • If the ratio of HTTP responses with code 500 suddenly increases, investigation is needed.

This is the purpose of alerts: when such events are detected, the system notifies the right person, allowing you to keep your eyes off dashboards.

After investigating and finding the root cause, you should always ask yourself if you can build an alert to detect such a case. There is also the possibility of increasing the observability if some metrics are missing.

How to define Alerts?

The Prometheus Operator allows the definition of alerts with a custom Kubernetes object: PrometheusRule

In this custom resource, you can define multiple alerts:

apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
metadata:
  name: webapp
  namespace: my-app
  labels:
    component: backend
    instance: app
    name: containers-my-app
    app: kube-prometheus-stack
    release: prom
spec:
  groups:
    - name: app
      rules:
        - alert: NoProductViewSince1h
          expr: (view_total - view_total offset 1h) < 1
          for: 5m
          labels:
            severity: Critical

In the example above, only one alert is defined.

Find below what needs to be defined for usual cases for each alert:

  • alert: the alert name
  • expr: a PromQL expression that triggers an alert if it returns something. This is why most of the time a threshold is used. With the offset function we compute the views from the past hour, if the result is above the threshold 1, then an alert is created and pushed to AlertManager.
  • Optional labels: a set of labels, usually used for alert severity
  • Optional for: delays triggering the alert. The PromQL expression must return some sample during the duration of the field for, before an alert is triggered.

Troubleshooting

There are many selectors involved in this process:

  • selector on ServiceMonitor to find Services
  • selector on Service to find Pods
  • selector on PodMonitor to find Pods

There are also selectors to discover ServiceMonitor, PodMonitor, and PrometheusRule objects. Those selectors are defined in the Prometheus object using the following fields:

  • Discovery of ServiceMonitor:
    • spec.serviceMonitorSelector
    • spec.serviceMonitorNamespaceSelector
  • Discovery of PodMonitor:
    • spec.podMonitorSelector
    • spec.podMonitorNamespaceSelector
  • Dicovery of PrometheusRule:
    • spec.ruleSelector
    • spec.ruleNamespaceSelector

Selectors used by Prometheus Operator

Check Selectors

If the target is not discovered by Prometheus:

  • Check that your ServiceMonitor or PodMonitor is deployed in a Namespace that matches the namespace selector in the Prometheus object.
  • Check that labels on your ServiceMonitor or PodMonitor match the selector in the Prometheus object.
  • Check that the selector on your ServiceMonitor or PodMonitor matches labels defined in the Service or Pod.
  • Check that the Service or Pod are deployed in the Namespace selected by the namespace selector defined in the ServiceMonitor or PodMonitor.

In order to check a label selector, you can use the -l option of kubectl. For example, to check the following selector in a ServiceMonitor:

  selector:
    matchLabels:
      component: backend
      instance: app
      name: containers-my-app

run the following command:

kubectl -l component=backend,instance=app,name=containers-my-app get service

Check Port Name or Number

  • Check that the port number or name matches a port defined in a Service or a Pod.

ServiceMonitor references either an incoming port defined in the Service or a Pod port:

  • ServiceMonitor.spec.endpoints.port references the name of a Service port: Service.spec.ports.name
  • ServiceMonitor.spec.endpoints.targetPort references a Pod port: Pod.spec.containers.ports.containerPort or Pod.spec.containers.ports.name

PodMonitor references port defined on Pod:

  • PodMonitor.spec.podMetricsEndpoints.port reference the name of a Pod port: Pod.spec.containers.ports.name

Note that Prometheus will only use Pods with a Ready state.

Create Grafana Dashboards as ConfigMaps

Grafana includes an auto discovery mechanism for dashboards. Any ConfigMap with a label grafana_dashboard=1 is loaded into Grafana.

The following ConfigMap will create a minimal dashboard in Grafana. Note that this ConfigMap needs to be deployed in the same Namespace as Grafana.

kind: ConfigMap
apiVersion: v1
metadata:
  name: my-dashboard
  labels:
    grafana_dashboard: "1"
data:
  dashboard.json: |
    { "title": "Product Views",
      "time": { "from": "now-6h", "to": "now" },
      "editable": false,
      "panels": [ {
          "gridPos": { "h": 9, "w": 12, "x": 0, "y": 0 },
          "id": 2,
          "targets": [ {
              "exemplar": true,
              "expr": "rate(view_total[2m])",
              "interval": "",
              "legendFormat": "{{product}}",
              "refId": "A"
            } ],
          "title": "Product View",
          "type": "timeseries"
        }
      ]
    }

Package Monitoring Objects with Applications using Helm Charts

Including monitoring objects in Application Helm Charts is a good way to maintain the observability layer of an application. The monitoring components can be versioned with application packaging. Also the deployment of monitoring can follow the same workflow as the application.

I will not explain how to package an application, but I’ll demonstrate how to include the following elements:

  • Dashboard: ConfigMap with dashboard code
  • Alerts: PrometheusRule with alerts definition
  • Metrics endpoint: PodMonitor or ServiceMonitor

In the chart’s values.yaml, add a new section for monitoring:

monitoring:
  alerts: true
  dashboard: true

The monitoring.alerts values controls the deployment of the PrometheusRule object. The deployment of the ConfigMap for the dashboard is controlled by monitoring.dashboard.

For the PodMonitor or ServiceMonitor objects, we can check if the Prometheus Operator is installed using the .Capabilities.APIVersions.Has function:

{{- if .Capabilities.APIVersions.Has "servicemonitor.monitoring.coreos.com/v1" }}
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
...
{{- end }}

Additionally, for alerts and dashboards, we can check the "values" set on the Helm release:

{{- if .Capabilities.APIVersions.Has "prometheusrule.monitoring.coreos.com/v1" }}
{{- if .Values.monitoring.alerts }}
apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
...
{{- end }}
{{- end }}

A common workflow to maintain a dashboard is to:

  • edit dashboards in the Grafana Web UI
  • copy the JSON model from Web UI
  • paste the JSON to a file in the Helm Chart
  • commit and push modifications

⚠ If the JSON representation of the dashboard is stored in the ConfigMap code, you will have to indent the content properly:

kind: ConfigMap
apiVersion: v1
metadata:
  name: my-dashboard
  labels:
    grafana_dashboard: "1"
data:
  dashboard.json: |
    { "title": "Product Views",
      "time": { "from": "now-6h", "to": "now" },
      "editable": false,
      "panels": [ {
          "gridPos": { "h": 9, "w": 12, "x": 0, "y": 0 },
          "id": 2,
          "targets": [ {
              "exemplar": true,
              "expr": "rate(view_total[2m])",
              "interval": "",
              "legendFormat": "{{product}}",
              "refId": "A"
            } ],
          "title": "Product View",
          "type": "timeseries"
        }
      ]
    }

It is generally easier to store the dashboard code in a dedicated file and then load the contents in the ConfigMap with some Helm templating functions:

{{- if .Capabilities.APIVersions.Has "prometheusrule.monitoring.coreos.com/v1" }}
{{- if .Values.monitoring.dashboard }}
apiVersion: v1
kind: ConfigMap
…
data:                                             
  dashboard.json: |                               
    {{- .Files.Get "dashboard.json" | trim | nindent 4 }}
{{- end }}
{{- end }}

Conclusion

Deploying monitoring stacks —Prometheus, Grafana, AlertManager, ElasticSearch, Loki, …— provides much more observability in a project, but at the cost of consuming more resources. Developers not using these features is a waste of resources. The project also has poor observability because only system metrics and maybe http metrics are retrieved. Adding application specific metrics and even business metrics allows you to build beautiful dashboards with colors and graphs that are very fun to report during boring review meetings.

grafana


This content originally appeared on DEV Community and was authored by Julien Acroute


Print Share Comment Cite Upload Translate Updates
APA

Julien Acroute | Sciencx (2021-07-19T08:59:28+00:00) Integrate an Application with Prometheus Operator and Package with a Helm Chart. Retrieved from https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/

MLA
" » Integrate an Application with Prometheus Operator and Package with a Helm Chart." Julien Acroute | Sciencx - Monday July 19, 2021, https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/
HARVARD
Julien Acroute | Sciencx Monday July 19, 2021 » Integrate an Application with Prometheus Operator and Package with a Helm Chart., viewed ,<https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/>
VANCOUVER
Julien Acroute | Sciencx - » Integrate an Application with Prometheus Operator and Package with a Helm Chart. [Internet]. [Accessed ]. Available from: https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/
CHICAGO
" » Integrate an Application with Prometheus Operator and Package with a Helm Chart." Julien Acroute | Sciencx - Accessed . https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/
IEEE
" » Integrate an Application with Prometheus Operator and Package with a Helm Chart." Julien Acroute | Sciencx [Online]. Available: https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/. [Accessed: ]
rf:citation
» Integrate an Application with Prometheus Operator and Package with a Helm Chart | Julien Acroute | Sciencx | https://www.scien.cx/2021/07/19/integrate-an-application-with-prometheus-operator-and-package-with-a-helm-chart/ |

Please log in to upload a file.



There are no updates yet.
Click the Upload button above to add an update.

You must be logged in to translate posts. Please log in or register.

Related Posts