ollama

Metadata

Field	Value
Service	ollama
Purpose	Local LLM inference service with PVC-backed model cache
Criticality	Tier 3
Owner	Application / platform owner
Clusters	homelab
Namespace	ollama
Exposure	internet
Stateful	yes
Backup class	snapshot
RPO / RTO	Model cache can be rebuilt, 1 to 2 hours to restore service
Last reviewed	2026-05-20

1. Service Overview

Ollama serves local model inference APIs for interactive or automation workloads and stores pulled models on a persistent volume.

Summary

If it fails, LLM-backed workflows stop, but model artifacts can usually be re-pulled if the cache is lost.

Dependencies

Dependency	Type	Why it matters
PVC-backed storage	storage	Stores downloaded model artifacts
Traefik	ingress	Exposes the external API
Node scheduling	runtime	Determines where model execution occurs

2. Architecture Diagram

[Client]
  -> [Traefik]
  -> [Ollama API]
  -> [PVC-backed model cache]

3. Deployment Specifications

Item	Value
Source path	ollama/base and ollama/overlays/homelab
Deployment model	Kustomize plus Fleet bundle
Namespace	ollama
Workload kind	Deployment
Chart or image version	See base manifests for the current image tag
Config files	base/kustomization.yaml, overlays/homelab/kustomization.yaml, fleet.yaml

Cluster mapping

Cluster	Overlay path	Notes
homelab	ollama/overlays/homelab	Current deployment target

4. Configuration Guide

Environment variables

Variable	Source	Purpose	Secret?
Ollama runtime settings	overlay manifests and optional secrets	API and model-cache behavior	mixed

ConfigMaps

Resource	Path	Purpose
Kustomize-managed runtime config	ollama/base and ollama/overlays/homelab	Scheduling and service customization

Secrets management

• Secret names: optional runtime secrets in the ollama namespace
• Source of truth: overlay inputs and generated manifests
• Rotation trigger: API or external integration changes
• Recovery note: restore secrets and remount the model cache PVC before restart if needed

5. Access Protocols

Path	URL or endpoint	Audience	Auth	TLS terminates at
Internal	Ollama service in the ollama namespace	Cluster workloads	namespace RBAC	Traefik / Ollama
External	https://ollama.mutana.site	Operators and integrated clients	ingress policy	Traefik

6. Operations and Observability

• Primary health indicators: Deployment healthy, API responsive, and model cache mounted.
• Dashboards or alerts: shared cluster monitoring and pod resource metrics.
• Log locations: Ollama pod logs.
• Known failure modes: large-model disk exhaustion, node scheduling issues, or ingress/API errors.

7. Backup and Recovery Notes

• Backup method: PVC snapshot when local model cache preservation matters.
• Restore prerequisites: enough storage capacity and runtime secrets if used.
• Related runbook: not required for this lower-blast-radius service.

8. Release and Change Notes

• Current deployed app version: see ollama/base image tags.
• Current chart version: N/A.
• Last significant change: homelab overlay documented with the current Traefik exposure path.
• Rollback reference: previous overlay revision in Git.