Node Roles

Node	Planned role	Workload intent	Storage / accelerator
NucBox G5	RKE2 server, control plane, etcd	Cluster management only; may also physically host external Forgejo	Intel N97, ~12 GB RAM
Sager laptop	RKE2 worker	General workloads, secondary game-server worker, Jellyfin candidate, GTX 1070 workloads	GTX 1070; local Jellyfin cache/transcode; shared TrueNAS game storage planned
OptiPlex 5070	RKE2 worker	Primary game-server worker, shared with Sager where practical	Existing local PVs initially; shared TrueNAS-backed game storage planned
Media server	RKE2 GPU worker	ComfyUI and other approved GPU workloads	RTX 3080 Ti; ~786 GB of node-local ComfyUI assets
TrueNAS SCALE	External storage appliance	Shared and Kubernetes-managed persistent storage	Existing ZFS datasets under /mnt/homelab/apps/ and /mnt/homelab/media

Scheduling policy

NucBox control plane

The NucBox carries a control-plane taint that normal application workloads do not tolerate. Cluster components that legitimately require the server remain allowed. This protects the relatively small control-plane node from application resource pressure.

Forgejo remains outside Kubernetes initially but may still run on the same physical NucBox. It is therefore outside the Kubernetes failure boundary, not necessarily the NucBox hardware failure boundary.

Sager and OptiPlex game-server placement

The OptiPlex is the initial primary game-server worker and may continue using existing local PVs during early migration. The target is to move suitable game-server persistent data to shared TrueNAS-backed storage so workloads can schedule across the OptiPlex and Sager where practical.

Scheduling policy should distinguish between:

• Game servers still tied to OptiPlex local data
• Game servers using shared TrueNAS storage and eligible for either worker
• Workloads that should not consume game-server capacity

Labels, taints, tolerations, and required affinity must reflect the selected storage placement. A toleration alone is not a placement guarantee.

Sager workload intent

The Sager is planned as:

• A secondary game-server worker
• A Jellyfin candidate
• A GTX 1070 workload host where useful
• The owner of local Jellyfin cache/transcode data if Jellyfin moves there

Jellyfin media remains on the existing TrueNAS media dataset.

Media-server GPU placement

The media server advertises its RTX 3080 Ti as a schedulable resource. ComfyUI remains tied to this node because both the GPU and approximately 786 GB of model assets are local. Those assets use a node-affined local PV.

Sager GPU

The GTX 1070 is a possible secondary GPU resource. It should be enabled only if driver compatibility and useful workload demand justify the added operational surface.

Failure considerations

Node-local data does not gain high availability merely because it is represented by a Kubernetes PV. Local-PV workloads require documented backup, node-loss recovery, and restoration procedures. Moving suitable game-server state to TrueNAS improves placement flexibility but does not remove the need for application-aware backup and recovery.