charts

Ollama Helm Chart

Version: 0.1.0 AppVersion: latest Type: application Artifact Hub

What is Ollama?

Ollama runs large language models locally — it pulls, manages and serves models (Llama, Mistral, Gemma, Qwen, …) behind a simple HTTP API on port 11434, with an OpenAI-compatible /v1 surface.

This chart deploys the ollama/ollama server on Kubernetes with a persistent model store and optional NVIDIA GPU acceleration. Behaviour is driven by the bjw-s-labs common library, so almost every knob lives in values.yaml.

Key features:

TL;DR

helm repo add obeone https://charts.obeone.cloud
helm install ollama obeone/ollama

The exporter / transparent proxy

This is the headline feature of the chart, so it gets its own section.

The exporter sidecar is a transparent proxy that sits in front of Ollama: it listens on :8000, forwards every API request to the Ollama server in the same pod (http://localhost:11434), and exposes Prometheus metrics on /metrics for the traffic it sees. It has instrumented routes for /api/chat and /api/generate and transparently passes through every other endpoint (including /v1/*), so nothing breaks by routing the API through it.

The chart defaults to a multi-arch image published on GHCR (ghcr.io/obeone/ollama-exporter), built from frcooper/ollama-exporter (Unlicense) and signed with cosign (keyless).

Because the proxy is in the API path, the chart wires the http Service port (11434) to the sidecar’s :8000 rather than straight to Ollama. All client traffic therefore flows through the proxy and gets accounted for.

A single switch toggles the whole path

You do not juggle several coupled values to turn the sidecar on or off. The chart keys everything off the exporter container’s own enabled flag and rewrites the Service wiring for you (in templates/common.yaml, before the bjw-s loader renders):

controllers:
  main:
    containers:
      exporter:
        enabled: false   # <-- the only line you touch
exporter.enabled http Service port → targetPort metrics port (:8000) Sidecar container
true (default) 11434 → 8000 (through proxy) present present
false 11434 → 11434 (direct to Ollama) dropped dropped

So when you disable it, the API keeps working — traffic just goes straight to Ollama and there is no dangling :8000 Service port left pointing at a container that no longer exists. No other value needs changing.

ℹ️ The default ships the proxy on. If you have no metrics/proxy image to run, set exporter.enabled: false and Ollama is served directly.

Prerequisites

Installing the chart

helm install ollama obeone/ollama

The chart will:

  1. Create a Deployment with one replica (single RWO PVC → Recreate strategy).
  2. Provision a 100 GiB PVC mounted at /root/.ollama for models and blobs.
  3. Expose a ClusterIP service on port 11434 (through the proxy by default).

GPU + direct-serving example:

helm install ollama obeone/ollama \
  --set defaultPodOptions.runtimeClassName=nvidia \
  --set controllers.main.containers.exporter.enabled=false \
  --set persistence.data.size=200Gi

Uninstalling the chart

helm uninstall ollama

The PVC is retained by default — delete it explicitly for a clean slate:

kubectl delete pvc -l app.kubernetes.io/name=ollama

Configuration

Example values.yaml

defaultPodOptions:
  runtimeClassName: nvidia          # GPU; leave "" for CPU-only

controllers:
  main:
    containers:
      ollama:
        image:
          repository: ollama/ollama
          tag: "0.6.0"              # pin for reproducibility
        env:
          OLLAMA_KEEP_ALIVE: 30m
          OLLAMA_CONTEXT_LENGTH: "64000"
        resources:
          requests:
            memory: 10Gi            # bump to fit your models
      exporter:
        enabled: true               # transparent proxy + /metrics (see above)
        image:
          repository: ghcr.io/obeone/ollama-exporter

persistence:
  data:
    size: 200Gi

serviceMonitor:
  metrics:
    enabled: true                   # needs a Prometheus Operator stack

Parameters

Pod / GPU

Key Type Default Description
defaultPodOptions.runtimeClassName string "" NVIDIA RuntimeClass for GPU; empty = CPU-only
defaultPodOptions.automountServiceAccountToken bool false Ollama does not call the Kubernetes API
defaultPodOptions.nodeSelector object {} Pin the workload to specific node(s), typically a GPU host

Ollama container

Key Type Default Description
controllers.main.containers.ollama.image.repository string ollama/ollama Ollama server image
controllers.main.containers.ollama.image.tag string latest Image tag — pin to a release for reproducibility
controllers.main.containers.ollama.env.OLLAMA_KEEP_ALIVE string 30m How long a model stays loaded after the last call
controllers.main.containers.ollama.env.OLLAMA_NUM_PARALLEL string "2" Parallel requests per model
controllers.main.containers.ollama.env.OLLAMA_MAX_LOADED_MODELS string "4" Max models kept loaded simultaneously
controllers.main.containers.ollama.env.OLLAMA_CONTEXT_LENGTH string "64000" Default context window size
controllers.main.containers.ollama.env.OLLAMA_FLASH_ATTENTION string "true" Enable Flash Attention

Exporter / proxy sidecar

Key Type Default Description
controllers.main.containers.exporter.enabled bool true Single switch for the proxy + metrics path (see section above)
controllers.main.containers.exporter.image.repository string ghcr.io/obeone/ollama-exporter Proxy/exporter image (cosign-signed multi-arch build)
controllers.main.containers.exporter.image.tag string latest Image tag
controllers.main.containers.exporter.env.OLLAMA_HOST string http://localhost:11434 Upstream Ollama URL the proxy forwards to

Service

Key Type Default Description
service.main.type string ClusterIP Service type
service.main.ipFamilyPolicy string PreferDualStack Dual-stack, degrades to single-stack
service.main.ports.http.port int 11434 API port; targetPort is managed by exporter.enabled
service.main.ports.metrics.port int 8000 Exporter /metrics port (dropped when the exporter is off)

Persistence

Key Type Default Description
persistence.data.enabled bool true Model/blob store at /root/.ollama
persistence.data.size string 100Gi PVC size — size it for the models you intend to pull
persistence.data.accessMode string ReadWriteOnce PVC access mode
persistence.tmp.enabled bool true emptyDir scratch for OLLAMA_TMPDIR at /tmp/ollama

ServiceMonitor

Key Type Default Description
serviceMonitor.metrics.enabled bool false Prometheus Operator ServiceMonitor scraping the exporter

Ingress

Key Type Default Description
ingress.main.enabled bool false Enable ingress (routes to the proxied API port, not /metrics)
ingress.main.hosts list … Ingress hosts
ingress.main.tls list [] Ingress TLS

Metrics

When the exporter is enabled (default), Prometheus metrics are served on :8000/metrics. With a Prometheus Operator stack present, scrape them by enabling the bundled ServiceMonitor:

helm upgrade ollama obeone/ollama --reuse-values \
  --set serviceMonitor.metrics.enabled=true

Disabling the exporter (exporter.enabled=false) removes both the sidecar and the metrics port, so leave the ServiceMonitor off in that case.

GPU notes

GPU exposure is effective only when defaultPodOptions.runtimeClassName points at an NVIDIA RuntimeClass. The ollama container already carries NVIDIA_VISIBLE_DEVICES=all and NVIDIA_DRIVER_CAPABILITIES=compute,utility, plus OLLAMA_SCHED_SPREAD=1 to spread work across all GPUs. For Sablier scale-to-zero on a GPU group, set pod labels under controllers.main.pod.labels.

Upgrading

helm upgrade ollama obeone/ollama --reuse-values

Troubleshooting

Checking logs

kubectl logs -l app.kubernetes.io/name=ollama -c ollama
kubectl logs -l app.kubernetes.io/name=ollama -c exporter   # if enabled

Pulling and querying a model

kubectl exec -it deploy/ollama -c ollama -- ollama pull llama3.2
kubectl run -it --rm ollama-curl --image=curlimages/curl --restart=Never -- \
  -s http://ollama:11434/api/tags

The ollama Service name assumes a release name of ollama.

Source Code

Requirements

Repository Name Version
https://bjw-s-labs.github.io/helm-charts common 5.0.1

Maintainers

Name Email Url
obeone obeone@obeone.org