Getting started with kube-prometheus-stack

I recently added monitoring and alerting to my Kubernetes stack to discover and respond to failures faster. It took me a lot longer I than expected to get my head around kube-prometheus-stack and get it doing what I wanted it to, so here's what I found out.


I've spent the last year building and running infrastructure to host virtual conferences (Climate:Red, MozFest '21 & '22 & Planet:Red). Running these large deployments by myself has led to a lot of automation. I start off automating testing and integration, then building and releasing containers and then deployment, configuration and scaling of those containers.

I'd previously used kube-prometheus-stack to see what was going on inside my Kubernetes clusters, but always knew there was a lot more to it than the fancy Grafana graphs. For MozFest 2022 I decided it was time to look into this properly.

The conference infrastructure relies on several background tasks, running as Kubernetes CronJob resources, which can fail very quietly. If one of the jobs failed, the schedule might become stale or new users might not be able to sign in. I needed something to quickly surface those background errors and report them.


kube-promethus-stack is pretty easy to get installed, DigitalOcean even offers a one-click install after deploying a Kubernetes cluster! That internally uses Helm to deploy prometheus-community/kube-prometheus-stack. So you can:

helm repo add prometheus-community
helm repo update
helm install kube-prometheus-stack prometheus-community/kube-prometheus-stack

This deploys prometheus-operator/kube-prometheus, which is an operator that uses CRDs to deploy and configure instances of Prometheus and Alertmanager. It uses them to deploy a standard instance of Prometheus and Alertmanager and adds a tonne of rules for monitoring standard Kubernetes resources.

One confusion point was that internally kube-prometheus-stack is deployed using Jsonnet templates and the helm offering is community-operated, so there are some differences between both sets of documentation.

How it works

Prometheus is a time series database which periodically scrapes different servers to fetch metrics. It offers a query language, PromQL, to query those metrics over time and a rule system that triggers alerts based on queries. Alerts have a name and key-value labels of information about them, like the namespace the resource was in or name of the related container.

The stack comes setup to monitor the resources within Kubernetes using kube-state-metrics and has lots of rules already setup to monitor things like Pods crashing.

Alertmanager takes Prometheus' alerts and lets you group and filter them based on the associated labels. It provides routing to send different alerts to different receivers, like an email or Slack message.

Configuring Alertmanager

The default Prometheus instance is already configured to collect metrics on Kubernetes resources and alert on failures. I was interested in KubeJobFailed and KubePodCrashLooping rules to detect failed jobs or bad deployments. Alertmanager needs to be configured with a receiver to send them somewhere. prometheus-operator provides the AlertmanagerConfig custom resource to do just that. The default Alertmanager instance automatically regenerates it's internal configuration whenever any AlertmanagerConfig is changed in any namespace. So you can create a custom-alertmanagerconfig.yml and kubectl apply -f -it:

kind: AlertmanagerConfig
name: custom-config
namespace: production
groupBy: ['job']
groupWait: 30s
groupInterval: 5m
repeatInterval: 12h
receiver: 'sendgrid-smtp'
- key: severity
value: warning

- name: 'sendgrid-smtp'
- to:
authUsername: apikey
name: sendgrid-alerts

It's worth pointing out AlertmanagerConfig is a v1alpha1 CRD so it will most likely change in the future.

This configuration is a camel-case version of regular Alertmanager config and the operator merges them together into one configuration. This creates a route that groups jobs by their job label and sends them to the sendgrid-smtp receiver. You can use matchers to filter the alerts the route receives based on their labels.

There is a bit of hidden logic in how prometheus-operator merges these configurations back together. It will ignore any namespace matchers you set and instead match the namespace the AlertmanagerConfig itself is in. In this case it would be namespace=production, so make sure to put the config in the same namespace you want alerts from.

Receivers configure how to actually send alerts, in this example via smtp. This one will email alerts to via SendGrid's smtp server. It references a secret, which needs to be in the same namespace as the AlertmanagerConfig, to specify the password. So you don't have to put credentials in your CRD.


Detecting errors

First, you want to make sure the operator is accepting accepting your AlertmanagerConfig. It will log any errors from the config or let you know that it successfully reloaded. You can watch those logs:

kubectl logs -f -n kube-prometheus-stack deploy/kube-prometheus-stack-operator

Inspecting generated config

If the configuration is loading correctly, you can inspect the generated secret that it is stored in:

kubectl get secret \
-n kube-prometheus-stack \
-o=jsonpath="{.data.alertmanager\.yaml}" \
alertmanager-kube-prometheus-stack-alertmanager-generated \
| base64 -d

Configuration options

If you're unsure about what you can put into an AlertmanagerConfig, is a great website that describes CRDs based on their schema. →

Inspecting prometheus & alertmanager

You can port-forward Prometheus and Alertmanager to see what is going on in a browser. Looking at Prometheus is useful to explore the default alerts and their generated labels and also inspect the raw metrics. Looking at Alertmanager lets you see the grouping that is applied and the alerts that have been triggered.

# Inspect prometheus
kubectl -n kube-prometheus-stack port-forward svc/prometheus-operated 9090:9090
open http://localhost:9090

# Inspect alertmanager
kubectl -n kube-prometheus-stack port-forward svc/alertmanager-operated 9093:9093
open http://localhost:9093


To make sure alerting is setup, create an obviously bad workload and see if it triggers an alert. e.g. this deployment.yml:

apiVersion: apps/v1
kind: Deployment
name: bad-app
namespace: production
matchLabels: bad-app
labels: bad-app
- name: bad-app
image: busybox
command: ['exit', '1']

After you kubectl apply -f it, you should first see a "Pending" alert in Prometheus After the for duration has passed, it should fire the alert. Once fired, you will see the new alert in Alertmanager and receive it in whatever receiver(s) you have setup.

Next steps

Choosing rules is the next logical step. This setup will alert on all of kube-prometheus-stack's default alerts, which are a good first step. But if that's too much info, you need to find the alerts that are useful to your specific stack and setup matchers. Looking through prometheus's default alerts should be a good first step, i.e. localhost:9090/alerts.

Meta alerting is useful to ensure confidence that monitoring and alerting is all going as expected. Prometheus constantly fires a Watchdog alert so you could use a dead-man's-switch type service that alerts you when the alert stops firing.

Different receivers let you send different notifications to different places, perhaps based on severity or the team responsible. You could have seperate alerts for frontend or backend teams for example.

Setup persistence, the default kube-prometheus-stack doesn't persist any metrics beyond a reboot. For alerting purposes this is fine but if you need to keep those metrics around you'll want to look into configuring some PersistentVolumeClaims.

Custom metrics is a very interesting next step. You'll need to tell your Prometheus instance to start scraping your services or pods directly using a ServiceMonitor or PodMonitor and define custom /metrics endpoints on your containers that expose prometheus metrics specific to your stack. For MozFest there is a 'site-visitors' widget and it could be interesting to track that over time and have a custom alert when it is over a certain value.


These were helpful links I used while getting this setup:

  • provides a relatively good overview of whats going on, but does miss out on some of the nuances that go on behind the scenes. The design page is a very good starting point explaining what each CRD is for. The user-guides section in their GitHub repo has some good guides for getting started and configuration.
  • was really helpful debugging what needs to go inside the AlertmanagerConfig resource and the descriptions helped to describe what different fields do and how they operate together.
  • Alertmanager Configuration is useful to know what should and shouldn't be going into those configs and debugging what the generated configurations do. The visual editor was kind of useful, but not very.
  • The Artifact Hub page has some useful information on it too.


These were the steps I ended up with to get notifications when things start going wrong inside my Kubernetes cluster. It helpfully surfaces things that are going wrong and lets you know about them. If you found this helpful or have questions, let me know on Twitter!