Skip to content

Fleet inventory — October 21, 2026

Purpose: list every machine in the lab, with what runs on it and how loaded it is. Status: dated, read-only snapshot. It replaces the October 15 inventory as the current host list. Disks are maintained in the disk map.

Collection: about 13:10–13:20 UTC, from Prometheus (physical-node exporters and in-cluster node-exporter), the Kubernetes API, and Longhorn objects. Queries ran separately, so this is not an atomic snapshot. CPU % is a 10-minute average. No benchmarks or changes were made.

What changed since October 15

  • Dell OptiPlex 7060 (.16) retired (#3067). Its VM, machine class, Proxmox provider and Longhorn node are gone, and the host is powered off.
  • Two bare-metal Talos workers joined, PXE-booted through booter on the Pi (#3069, #3074, #3077). They are the first nodes not running as Proxmox VMs.
  • Omni v1.12.4 and Talos v1.14.4 on every node (#3068).
  • Paseo turned off (#3072). Hermes is the cluster agent.
  • Two single-replica volumes faulted after the upgrade reboots. On 2026-10-21 tubesync/config-pvc was restored from kopiur and paperless-ngx/paperless-export-pvc was recreated (#3071).

Physical machines

Machine Address CPU RAM Platform Runs CPU used RAM free Uptime
Threadripper (X399) .14 Ryzen TR 2950X · 32 threads 125.7 GiB Proxmox VE 9 (kernel 7.0.14-25-pve) GPU worker VM (30 vCPU, 100 GiB, 2× RTX 3090) 49% 20.2 GiB 0.9 d
HP ProDesk 600 G4 SFF .21 i5-8500 · 6 threads 62.6 GiB (4× 16 GB, full) Proxmox VE 9 (7.0.14-25-pve) control plane VM (4 vCPU, 12 GiB) + worker VM (6 vCPU, 40 GiB) 37% 11.3 GiB 0.9 d
HP Elite Mini 600 G9 .22 i5-13500T · 20 threads 30.6 GiB (2× 16 GB DDR5) Proxmox VE 9 (7.0.14-20-pve) worker VM (16 vCPU, 26 GiB) + Proxmox Datacenter Manager LXC 16% 2.1 GiB 4.4 d
HP ProDesk 600 G4 DM (shed) .20 i5-8500T · 6 threads 31.1 GiB Proxmox VE 9 (7.0.14-25-pve) worker VM (4 vCPU, 12 GiB), USB radios, Wi-Fi bridge 5% 24.1 GiB 0.9 d
Dell Pro Micro QCM1250 .181 Core Ultra 5 245T · 14 cores, iGPU + NPU 16 GiB (1 DIMM, 1 slot free) Talos v1.14.4 bare metal worker mini-pc-dell-ultra-worker 4% see nodes 0.4 d
HP Elite Mini 800 G9 .82 i5-12500T · 12 threads 16 GiB (2× 8 GB, both slots used) Talos v1.14.4 bare metal worker hp-elite-gen12 6% see nodes 0.8 d
HPE ProLiant DL360 Gen9 (TrueNAS) .133 Xeon E5-2680 v4 · 28 threads 377.6 GiB TrueNAS 28.0 (kernel 6.18.55) NFS/SMB storage, kopiur repository 3% 25.3 GiB (ARC uses the rest) 7.8 d
Raspberry Pi 5 .15 Cortex-A76 · 4 threads 7.7 GiB Ubuntu 26.04 Omni v1.12.4, 4 Proxmox providers, booter v0.4.0 (PXE), Technitium DNS, ProxCenter 1% 5.8 GiB 34.4 d
Dell OptiPlex 7060 .16 — — retired, powered off — — — —

The Proxmox hosts .14, .21 and .20 were apt-upgraded and rebooted during the Talos rollout on 2026-10-20/21. That explains their 0.9-day uptime.

Kubernetes nodes

Cluster talos-prod-cluster-v2: 7 nodes, all Ready, Talos v1.14.4, kernel 6.18.57, Kubernetes v1.37.1, containerd 2.3.7.

Node Host Role / pool CPU (alloc) Memory (alloc) CPU used Memory used (working set) Memory requested Pods Accelerators
talos-prod-cluster-v2-control-planes-7mxjxv .21 VM control plane (NoSchedule) 3.95 11.1 GiB 0.32 5.8 GiB 4.2 GiB 20 —
talos-prod-cluster-v2-gpu-workers-jxfkdr .14 VM gpu 29.95 97.7 GiB 12.0 39.8 GiB 66.8 GiB (68%) 62 nvidia.com/gpu=2 (RTX 3090)
talos-prod-cluster-v2-hp-sff-workers-h76ncc .21 VM general 5.95 38.7 GiB 0.65 12.5 GiB 23.6 GiB (61%) 44 gpu.intel.com/i915=4
talos-prod-cluster-v2-hp-elite-workers-l5zlgr .22 VM general 15.95 24.9 GiB 1.87 9.0 GiB 15.7 GiB (63%) 69 gpu.intel.com/i915=4
talos-prod-cluster-v2-hp-micro-workers-pstbbd .20 VM edge (taint link=wifi) 3.95 11.2 GiB 0.06 2.1 GiB 2.6 GiB 13 —
mini-pc-dell-ultra-worker .181 bare metal general 13.25 13.5 GiB 0.27 6.3 GiB 8.0 GiB (59%) 48 gpu.intel.com/i915=4, NPU extension (no device plugin yet)
hp-elite-gen12 .82 bare metal general 11.25 13.7 GiB 0.30 3.6 GiB 7.2 GiB (53%) 45 gpu.intel.com/i915=4
Total 84.25 210.8 GiB 15.5 79.1 GiB 128.1 GiB 301

The bare-metal minis give Kubernetes 13.5–13.7 GiB of their 16 GiB, after the kubelet reservations (system 1 GiB, kube 512 MiB). The 16 GiB shed VM, by comparison, gives 11.2 GiB.

Storage

Longhorn: 66 volumes, 59 attached and healthy, 7 detached (normal). All 78 PVCs are Bound.

Node Longhorn disk Scheduling Size Free Scheduled (requested)
GPU worker gpu-bulk on 799.6 GiB 756.4 GiB 478.1 GiB
GPU worker ssd-flash (PLP, clone-ok) on 439.9 GiB 222.5 GiB 502.0 GiB (over-provisioned)
hp-sff worker hp-sff-ssd on 689.7 GiB 611.7 GiB 316.0 GiB
hp-elite worker hp-elite-nvme (Intel 660p, 76% worn) on 439.8 GiB 409.8 GiB 176.0 GiB
hp-micro worker hp-micro-ssd off (Wi-Fi node) 849.6 GiB 833.2 GiB 0
both minis — compute-only, no Longhorn disk — — —

Each node also has a talos-ephemeral disk with scheduling off. NAS pools and dataset usage were not re-collected; see the October 15 inventory.

GitOps health

98 Argo CD applications. 97 are Synced/Healthy. my-apps-posthog is OutOfSync (Healthy) while it waits on its clickhouse-migrations-init hook.

Findings

  1. HP Elite 600 G9 (.22) host has 2.1 GiB free. Its 26 GiB VM plus the PDM LXC leave Proxmox little headroom. This is the tightest host, and #3071 step 8 plans to convert it to bare metal (hp-elite-gen13).
  2. The shed host (.20) has 24 GiB free. Its VM uses 12 GiB of 31.1 GiB. The machine-class comment still describes a 16 GiB host. Growing the VM is possible, but the node is tainted for Wi-Fi-only workloads, so extra RAM only helps pods that tolerate that taint.
  3. GPU worker memory requests are at 68%, the highest in the cluster. The descheduler now sheds nodes above 65% (#3079).
  4. Threadripper reports 87–89 °C on k10temp. The 2950X reports Tctl with a +27 °C offset, so the die is about 60–62 °C. That is normal under 49% load. Other hosts peak at 57–68 °C, and the minis at 44–50 °C.
  5. Control plane is still one VM. The 1 → 3 move with both minis is the next planned change (#3071 step 4, draft #3014).

Limits

  • Physical CPU and RAM come from node-exporter on each host. VM sizes come from the Omni machine classes and the October 15 VM table. Proxmox was not queried.
  • "Memory requested" counts running pods only. Percentages are of allocatable memory.
  • Disk serials stay out of this page. See the disk map for drive models and roles.