I have taken one of the problems I used to ask in interviews a few years ago and applied it to the world of AI, since it seems like AI is everywhere these days.
Someone who is Senior/Principal Engineer should be able to easily answer this (I assume...)
Now - please be fair! I’m hoping you won’t rely too much on LLMs (though I won’t stop you if you do). The goal is to help understanding the subject how it can manifest in the real world...
Here is the problem:
Context:
- Kubernetes Cluster
- A single GPU node runs ~200 inference pods (can it?)
- Everything healthy. High QPS. GPU fully utilized.
Now:
- containerd suddenly gets SIGKILL’ed.
- Nothing crashes.
- Traffic continues.
- Pods remain Running.
- metrics still scraped.
- GPU utilization unchanged.
- Inference latency looks normal.
And then: You restart containerd.
Things start getting weird:
- k8s logs work
- However, container not found
- Deleting pods hangs in Terminating, never gets deleted.
- New pods schedule successfully — but fail GPU allocation intermittently.
- Old pods still serve traffic and still hold GPU memory.
- Only a node reboot clears the state.
Questions:
1. Why are workloads:
- alive at the kernel level
- visible to k8s cluster
- still consuming GPU
- and yet partially invisible to the runtime?
2. And why does GPU scheduling degrade only after containerd restarts?
Bonne Chance! 🙂
@RealAnkush @sunnykgupta@arpit_bhayani - please share among your acquaintances
@arpit_bhayani Good things to draw from but I feel folks should have originality and ask things what they feel like as well. Except for office food or amenities ofc.
Also @arpit_bhayani have you gotten any good answers from the first q? Surprised if well oiled machines also have issues like G