Providing Prometheus Alert Rules
When writing a component that manages a critical piece of infrastructure, you should provide alerts that notify the operator if it fails. Writing good alerts and runbooks is difficult. This document should give you some best practices that worked for us so far.
In nearly all cases you can provide Prometheus alert rules through the PrometheusRule CRD. This definition is then picked up by the responsible monitoring component.
For OpenShift cluster this generally means labeling the namespace with
openshift.io/cluster-monitoring: 'true' and for clusters with rancher monitoring this would mean labeling it with
Alerts need to be actionable
Try to imagine what you would do if you received this alert. If the answer is "I don’t know" or wait and see if it resolves itself, you probably shouldn’t emit this alert.
Label your alert
Label your alerts, so that they can be routed effectively. At the very least add labels
syn_component: 'COMPONENT_NAME'to indicate that this alert is managed by the syn component, and a label
How critical is this alert? We generally differentiate three severity levels.
infofor alerts that don’t need urgent intervention. These are things that someone should look into, but it can usually wait up to a few days. Info alerts could also often just be part of a dashboard.
warningfor alerts that should be looked at as soon as possible, but it can usually wait until regular office hours.
criticalfor alerts that need immediate attention, even outside office hours.
Carefully decide in which category your alert should be and add the appropriate
severitylabel. But keep in mind that if all alerts are critical none of them are.
Make alerts tunable
You most likely won’t be able to write a perfect alert out of the box. It will either be too noisy, not sensitive enough, or in some other way not relevant for the user. With that in mind, give the user a way to tune your alert.
At the very least provide ways to selectively enable or disable individual alerts. It’s considered best practice to let the user overwrite all of the alert specification if they wish. However, it’s a good idea to also provide some more convenient parameters to tune configuration that often need to be adapted such as alert labels or alert specific parameters like a list of relevant namespaces.
Try to imagine what a user might need to change and make tuning it as easy as possible.
Provide a runbook
You should always provide a link to a runbook in an annotation
runbook_url. See the section below on writing good runbooks.
Following these guidelines, you should get a usable alert. There are still some pitfalls when writing Prometheus alerts, but there are also many guides to help you write them. You can look at the official documentation or check out how Cloudscale writes alert rules.
When installing third party software there are often upstream alerts. It’s a good idea to reuse these alerts, but the best practices still apply.
Don’t blindly include all upstream alerts. Check if they’re actionable, add labels, make them tunable, and provide a runbook, even if you didn’t write the alert yourself.
Every alert rule should have a runbook. The runbook is the first place a user looks to get information on the alert and how to debug it.
It should tell the reader:
What does this alert mean?
Tell the reader why they got the alert. What exactly doesn’t work as it should? Maybe also tell the user how the alert was measured and if there might be false positive.
What’s the impact?
Who and what’s effected? How fast should the reader react? The alert labels should already give an impression how critical the alert is, but try to be more explicit in the runbook.
How do I diagnose this?
Provide some input on how to debug this. Where might the reader get the relevant events or logs? How to narrow down the possible root causes?
How may I mitigate the issue?
List some possible mitigation strategies or ways to resolve this alert for good.
Ideally, you shouldn’t alert on issues that could be fixed automatically. If you have one clear way to resolve this alert, check if you could resolve this automatically.
How do I tune the alert?
Maybe this alert wasn’t actionable, or maybe the alert was raised far too late. Give the reader options to tune the alert to make it less noisy or more sensitive.
Whenever possible try to provide code snippets and precise instructions.
If the reader got a critical alert, they don’t have the time or nerves to build the
jq query they need right now or to find out exactly which controller is responsible for this CRD.
It’s considered best practice to put all your runbooks at
docs/modules/ROOT/pages/runbooks/ALERTNAME.adoc, but there might be good reasons to deviate from this.
Just make sure to adjust the runbook links as necessary.
Finally, a runbook doesn’t have to be perfect. Maybe you don’t really know how this might fail or how to debug this, or maybe you simply don’t have the resources right now to write a comprehensive runbook. Add one anyway. Any input can be valuable when debugging an alert and at the very least there is now a basis on which to improve on when we learn more.
Removing or Renaming Alert Rules
Sometimes alerts become obsolete. Maybe the system can now resolve the issue automatically, or the responsible part simply doesn’t exist anymore.
However, you need to make sure that you never break a runbook link. There might be people using older releases of your component and their runbook links should still lead to valid runbooks.
If you follow these three rules, runbook links should always stay relevant.