Proxmox NUC — Pacific¶

Last Updated: 2026-07-19 Status: ✅ Active — rebuilt on new disk 2026-07-19, see below Purpose: Runs Home Assistant, Frigate (doorbell camera + AI detection), AdGuard Home (primary DNS), GMUPS-27 (UPS monitoring), and a Tailscale exit node.

Hardware specs: inventory.md → Pacific

⚠️ Temporary state as of 2026-07-19¶

Home Assistant (VM 112) is currently running on Lotus, not Pacific. Pacific's original boot disk failed (see Hardware Notes below) and was replaced; as part of that migration, HA was deliberately failed over to the standby VM on Lotus for the duration of the rebuild, per HA Recovery → Scenario B. VM 112 exists on Pacific again (restored from backup) but is intentionally left powered off until a single controlled failback once Pacific has been soak-tested. Check qm status 112 on Pacific and Lotus's VM manager before assuming which one is authoritative. This section will be updated once the failback happens.

All other guests (AdGuard, Frigate, GMUPS-27, Tailscale) are running natively on the rebuilt Pacific as normal.

Access¶

Method	Detail
Proxmox Web UI	https://192.168.1.10:8006
SSH	`ssh root@192.168.1.10` — key-based (`brabham-claude` key, same as Lotus/Cooper). Re-added 2026-07-19 after the fresh install (host SSH keys don't survive a reinstall — old entries in `known_hosts` need `ssh-keygen -R 192.168.1.10` before reconnecting)

Storage¶

Replaced 2026-07-19 — original 238.5GB SanDisk SDSSDHP256G failed (SMART showed climbing uncorrectable-sector counts; see Lessons Learned). Replaced with a Samsung 860 PRO 512GB (repurposed from Brabham).


Primary SSD	512GB Samsung 860 PRO (was 238.5GB SanDisk, failed)
Proxmox root	94GB LVM (was 59.3GB)
VM/LXC pool	~365GB LVM (was 152.6GB)
Onboard eMMC	29.1GB (unused, not part of the OS install)
Proxmox version	VE 9.2.2, kernel 7.0.2-6-pve (Debian 13 Trixie) — was 8.4.17

Virtual Machines¶

VMID	Name	RAM	Disk	Status	Notes
112	HomeAssistant	8GB	42GB	⏸️ Stopped (intentional)	Home Assistant OS — currently running on Lotus instead, see warning above

LXC Containers¶

CTID	Name	Status	Purpose
101	AdGuard	✅ Running	DNS filtering — active primary DNS for the network (see DNS Filtering). Root disk still only 3.9GB (unchanged since the 2026-07-19 disk-full incident) but healthy at ~31% used after the query-log retention fix
103	TailscaleExitNode	✅ Running	Tailscale exit node for the home network
105	kuma	❌ Stopped	Uptime Kuma monitoring — intentionally not backed up/restored during the 2026-07-19 rebuild, not needed
108	frigate	✅ Running	Frigate NVR — processes Reolink doorbell via Coral TPU. Required manual driver work after the disk rebuild, see Hardware Notes
130	GMUPS-27	✅ Running	UPS monitoring — NUT netclient, see NUT section below
200	dockertemplate	❌ Stopped	Template container — intentionally not backed up/restored, not needed

UPS / NUT (power-loss shutdown)¶

Not previously documented. The APC Back-UPS ES 650G2 connects via USB directly to Pacific — this is host-OS-level config, not captured by any VM/CT backup, and must be manually recreated after any fresh Proxmox install.

Packages: nut nut-client nut-server
Driver: usbhid-ups, auto-detected (port = auto in /etc/nut/ups.conf)
Mode: netserver — Pacific's upsd listens on 0.0.0.0:3493 and serves UPS status to network clients
upsmon on Pacific itself runs as master, monitoring apc@localhost, and triggers shutdown -h on critical battery — this is what protects Pacific from power-loss corruption
3 accounts in /etc/nut/upsd.users share one password (not in Bitwarden as of 2026-07-19 — added there after the 2026-07-19 rebuild, check if still current): upsadmin (master), upsuser (slave — used by GMUPS-27), homeassistant (used by HA's NUT integration to show battery/UPS sensors)
GMUPS-27 (LXC 130) is a NUT netclient — MONITOR apc@192.168.1.10 1 upsuser <pw> slave — it doesn't own the UPS itself, just monitors Pacific's server over the network
Home Assistant's UPS sensors also come from polling Pacific over the network (the homeassistant account), not from any USB passthrough to the HA VM

Config file contents (captured 2026-07-19 — for recreation after a fresh install)¶

Server config only, no secrets. The shared password for upsd.users/upsmon.conf is in Bitwarden as NUT UPS - Pacific.

# /etc/nut/ups.conf
pollinterval = 15
maxretry = 3
offdelay = 120
ondelay = 240
[apc]
driver = usbhid-ups
port = auto
desc = "APC Back-UPS ES 650G2"
vendorid = 051D

# /etc/nut/upsd.conf
LISTEN 0.0.0.0 3493
LISTEN :: 3493

# /etc/nut/nut.conf
MODE=netserver

upsmon.conf on Pacific monitors apc@localhost as master with SHUTDOWNCMD "/sbin/shutdown -h".

Bring it up with systemctl enable --now nut-driver@apc nut-server nut-monitor, then verify with upsc apc@localhost. Reinstall NUT before restoring any guests — GMUPS-27 and HA both just point at 192.168.1.10 and reconnect on their own once the server is back.

Hardware Notes¶

Coral TPU installed and passed through to Frigate LXC for AI object detection (person/vehicle recognition on doorbell camera). After the 2026-07-19 disk rebuild (Proxmox 8→9, kernel 6.8→7.0), the community gasket-dkms driver package failed to build — 4 separate kernel-API incompatibilities (eventfd_signal, no_llseek→noop_llseek, class_create, MODULE_IMPORT_NS). Fixed by building directly from https://github.com/google/gasket-driver (not the packaged version) with 2 manual one-line patches — see Lessons Learned for the full fix. Not DKMS-integrated — a future kernel upgrade will need the same 2 patches reapplied and the module rebuilt manually.
GPU passthrough (/dev/dri, VAAPI hardware video decode for Frigate's doorbell stream) also broke in the same rebuild, for an unrelated reason: a host-side script (/var/lib/lxc/108/mount_hook.sh) that dynamically bind-mounted /dev/apex_0 and /dev/dri/* into the container isn't part of any backup and didn't survive the reinstall. Replaced with static lxc.mount.entry lines in 108.conf instead of recreating the script — more durable since it doesn't depend on an external file.
Pacific is a NUC6CAYS (Celeron J3455, circa 2016) — inventory previously had the model as "NUC6CAYH", corrected 2026-07-19 after checking dmidecode directly (system-product-name = NUC6CAYS). CPU confirmed via lscpu as J3455. Single SATA bay, no spare M.2/second bay — a drive swap is an offline, one-drive-out-one-drive-in operation.
RAM is near capacity (~10GB used of 16GB)
UPS backed — survives brief power cuts without shutting down (see NUT section above)
Home Assistant runs as a VM (not LXC) to support HAOS add-ons and supervisor
Cooling fan noise, discovered 2026-07-19 during the disk swap (unrelated — fan bearing wear on ~decade-old hardware, not caused by the disk work). AVC BAZB0508R5U, "Advanced Hydraulic Bearing" (sealed, not user-lubricatable). Exact-fit replacement part available cheaply on Amazon.co.uk/eBay UK. Not yet replaced as of 2026-07-19 — fan still functions, just noisy; worth doing given Pacific lives in the bedroom.

Backup¶

Pacific VMs and LXC containers are backed up via Proxmox Backup Server, which runs as a Docker container on Lotus and stores backups on the Lotus array.

The PBS-New storage connection uses a dedicated backup@pbs PBS account — separate from and not the same password as any personal PBS login (e.g. admin@pbs). This password isn't recoverable if lost (one-way hash) — only resettable via the PBS web UI (Configuration → Access Control → User Management) if a PBS admin account is available, or CLI (proxmox-backup-manager user update backup@pbs --password ... inside the proxmox-backup-server container on Lotus). Backups themselves use --crypt-mode=none (no client-side encryption), so a password reset never affects the ability to restore existing backups — it's purely an auth credential, unrelated to data-at-rest encryption.

Recovery¶

See Proxmox Recovery for restore procedures — Scenario B (full reinstall) was executed for real on 2026-07-19 and the doc has been updated with what actually happened.