1

We just moved to kubernetes, but the engineer that helped launch it is on paternity leave before we had hoped (never trust a baby not to be eager!).

Now we're trying to do maintenance tasks and one off work and having nodes get killed in the middle of things.

I've looked at using kubernetes jobs for this, but that's overkill. We don't want to write manifest files for everything.

We just need long lived shell access to do this and that.

What's the pattern for this so your maintenance task doesn't get killed?

Improve this question
3
  • Could you be a bit more explicit as to what you mean by "maintenance"? Are you talking about working on the node itself, or Kubernetes objects ?
    – Bruce Becker
    May 27, 2022 at 15:35
  • Like, I want bash access to run an ETL script, say. Or to do some random maintenance task that might take a few hours or a few days to run. I'm not at the point where I could do maintenance on the cluster itself. Or, like, if I need to do a database migration script, where do I run that without having to make a complex job manifest?
    – mlissner
    May 27, 2022 at 21:22
  • My 2c, so take it for what it's worth and consider it a comment, not an answer, but dealing with "complex job manifest" is going to be unavoidable. My opinion is that if you have made the decision to adopt K8s, you should also accept that you're going to have to write a lot of YAML. The tasks you mention sound like they should be implemented as sidecars or batch jobs.
    – Bruce Becker
    May 28, 2022 at 6:02

1 Answer 1

Reset to default
1

We were eventually able to answer this by following the rules for when a node is terminated. According to the FAQ, there are a number of ways that pods can prevent the cluster autoscaler from removing a node. One type of pod is:

Pods that are not backed by a controller object (so not created by deployment, replica set, job, stateful set etc).

So our solution is to create a pod in that way via a manifest file. This lets us have a pod named maintenance that sticks around and isn't killed by the cluster autoscaler:

---
apiVersion: v1
kind: Pod
metadata:
  name: maintenance
  namespace: blah
  labels:
    type: maintenance
spec:
  containers:
    - name: web
      image: whatever
      imagePullPolicy: IfNotPresent
      command: [bash]
      stdin: true
      tty: true
Improve this answer

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.