DEV Community

Nerav Doshi
Nerav Doshi

Posted on

Deployed Ollama to Local Kubernetes With No Resource Limits

Before testing whether last entry's real Podman number (1.663GB) holds up as an actual Kubernetes resource limit, I wanted a baseline: what happens with nothing set at all. kind runs a full local Kubernetes cluster inside containers acting as nodes — Kubernetes on top of the same containers everything else in this series has been using.

Spun up a fresh cluster and deployed Ollama with no resources section in the pod spec whatsoever:

kind create cluster --name today-i-ran
kubectl apply -f ollama-pod-naive.yaml
kubectl exec -it ollama-naive -- ollama pull llama3.2:1b
kubectl exec -it ollama-naive -- ollama run llama3.2:1b
Enter fullscreen mode Exit fullscreen mode

Asked it a real question — "What does BGP do in OpenShift Networking" — to force a full response, then checked the actual resource configuration:

kubectl describe pod ollama-naive | grep -A5 "Limits\|Requests"
Enter fullscreen mode Exit fullscreen mode

Pulled and loaded with zero constraints. QoS Class: BestEffort, which is Kubernetes plainly telling you no requests or limits were set. One real annoyance worth flagging: the image pull took 4 minutes 26 seconds on this fresh node, against maybe 30 seconds under Podman in the last entry. A brand new kind cluster has nothing cached — the first pull anywhere is always going to be the slow one.

Also worth being honest about: the model's actual answer to the OpenShift networking question wasn't good. It invented a fake expansion for "OpenShift Networking" and made up BGP/BGPsec behavior that doesn't match how OpenShift actually works. A 1B local model sounding confident on something specialized like this is a different risk than the more mechanical oc questions from earlier entries — worth remembering not to trust this stuff just because it sounds sure of itself.

With no limit, Kubernetes will let a pod take whatever it wants — which is exactly why this baseline matters. Without it, there'd be no way to tell whether the Podman number from last entry was actually a safe ceiling or not. Spoiler for the next one: it wasn't.

Top comments (0)