Office Hours for March 2021

Hear ye, hear ye! It’s the third Wednesday of the month on the 17th, you know what that means, it’s time for office hours, our monthly livestream where we answer your user questions with a panel of experts, so if you’re stuck with something, then come on by and let us take a stab at it:

For more info check out:

1 Like

Alright, here are the raw show notes!

Office Hours March 2021

EU Edition


  • David McKay - Equinix Metal
  • Pierre Humberdroz, Spectrm
  • Chris Carty, Google Cloud
  • Chauncey Thorn, Phase2
  • Dave Strebel, Microsoft
  • Jorge Castro, Unusual.VC
  • Puja Abbassi, Giant Swarm
  • Mario Loria, Carta
  • Ayrat Khayretdinov, Google Cloud
  • Rachel Leekin, VMware


my question(s) is/are; how does everyone stay up to date on kubernetes or other work topics, how much time do you spend a week, during office hours? (pun intended) and how do you combine it with your family

Haroune Mohammedi:
Hello, we are having a hard time running k8s jobs that use a 30 GB image. The image takes up to 50 minutes to be pulled on a newly created node. Reducing the image size is not an option and the pull time is bound by decompression, not download, so local registries won’t help. We use autoscaling to allocate resources on demand but this is irritating because running a job that takes 5 minutes will take almost one hour if the node is newly created. Also, having a node up with a prepulled image waiting for jobs is very expensive for us because the nodes has GPUs. Any way to deal with that? Thanks in advance

General concensus seems to be to put the data in a volume so you mount that when the nodes come up instead of copying it over to every new node.

  • Person: igstan

  • Question: My question relates to one of the things that I noticed in the setup of our clusters, namely that each application has its own namespace, named after the application itself.

    There’s no way to get all pods across several namespaces in a single command
    For each app I have in mind I need to type out its particular namespace
    I can’t use kubectl config set-context --current --namespace to alleviate some of the typing

  • Answer: ? (k get pods --all-namespaces or -A)

  • Links:

:13 in the video

GitHub - derailed/k9s: 🐶 Kubernetes CLI To Manage Your Clusters In Style! (for a more interactive workflow)

Rachel: I’ve seen customers namespace by team, this helps them organize.
Walid: OpenShift uses projects, kubectx is a must
Mario: I start with default namespaces and then grow from there, on small, single purpose clusters this might (or might not, depends) remove complexity for the end user consuming the process. It’s not wrong to keep it simple. David: yeah, depends on who is consuming your cluster.
Archy: I like using gitops for this, helps keep everything seperated.
Puja: As kelsey said, kubectl is the new ssh, we’re trying to move towards automated processes
Chauncey: I recommend Octant if you want your developers to see a picture of what the cluster looks like:
Chris: Argo for me!
Puja: We like GitHub - derailed/k9s: 🐶 Kubernetes CLI To Manage Your Clusters In Style! and Mirantis has lens: Lens | Mirantis and also check out: GitHub - sbstp/kubie: A more powerful alternative to kubectx and kubens
Yogi: VSCode + K8s

Person: iG
Q: Does Kubernetes update process from 1.13 to 1.20 have to go through all versions? Any good documentation for that?

P: andeep
I’m at google at a big client and I’m constantly getting the question “how will GKE scale at their limits” what’s the bottleneck? etcd? masters? scheduler?

:35 (I’ll like each of these to the vid)

  • Person: Ravikanth
  • Question: My kubernetes cluster is deploying the pod in the node which is not in my cluster. Warning FailedScheduling default-scheduler node “ip-10-0-.ec2.internal” not found in cache the node is not in the cluster but still it is picking the node which is not in the cluster for deployment
  • Answer: ?
  • Links: