Spendlik/Claude_Homelab
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Document 2026-08-11 accidental host apt upgrade incident and resolution (kernel pin reaffirmed + apt-mark hold), add NVIDIA-on-host expected-behavior gotcha, add Recently Resolved section

Browse Source
This commit is contained in:
Spendlik 2026-08-11 09:03:06 +00:00
parent 2d1247fda2
commit e3a3117c36
1 changed files with 11 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
00_index.md
View File

@ -15,7 +15,7 @@
| **WebSupport service ID** | `15056760` |
| **Gitea owner** | `spendlik` |
| **MCP server** | `mcp.spendlik.sk` (CT 112, port 3500) |
| **Proxmox kernel** | `6.14.11-5-pve` PINNED — 6.17.x and 7.0.x installed but won't boot |
| **Proxmox kernel** | `6.14.11-5-pve` PINNED + `apt-mark hold` on all other installed kernel packages (as of 2026-08-11) — see gotchas |
---
@ -117,7 +117,7 @@ Script location: `/usr/local/bin/ddns-update.sh` in CT 108. Full credentials map
| **WebSupport API — PUT silently resets `name`** | A `PUT` to `/v2/service/{id}/dns/record/{recordId}` that omits the `name` field does NOT preserve the existing name — it silently renames the record to the bare zone apex. Always include `name` explicitly in PUT payloads, or avoid partial-field updates. |
| **WebSupport API — POST `name` is relative** | Creating a record via POST: `name` is relative to the zone, not a full FQDN. Passing `"email.spendlik.sk"` creates `email.spendlik.sk.spendlik.sk`. Use the bare subdomain only (`"email"`, `"*"`, `"git"`). |
| **Gitea writes** | Always read file content before writing — writes replace entire file |
| **Proxmox kernel** | PINNED to `6.14.11-5-pve`. Do NOT unpin or upgrade without verifying NVIDIA 550 DKMS support first |
| **Proxmox kernel** | PINNED to `6.14.11-5-pve` (`proxmox-boot-tool kernel pin`). **2026-08-11 incident**: an accidental `apt upgrade -y` run directly on the Proxmox host (meant for a container) pulled in `proxmox-kernel-6.14.11-9-pve` and the entire `proxmox-kernel-7.0.x` branch. No impact occurred — the boot pin held, running kernel never changed, NVIDIA DKMS module remained built against the pinned `6.14.11-5-pve` only. Resolved by re-affirming the pin and running `apt-mark hold proxmox-kernel-6.14 proxmox-kernel-7.0 proxmox-kernel-6.14.11-9-pve proxmox-kernel-7.0.14-11-pve-signed proxmox-kernel-7.0.6-2-pve`, which now blocks apt from proposing these kernel packages at all. **Verify hold status with `apt-mark showhold` before assuming any future `apt upgrade` on the host is kernel-safe.** |
| **LXC template versions** | Verify exact template filename in `obsidian-vault/05 Resources/Proxmox LXC Templates.md` (or live `pveam list local`) before every `pct create` — versions bump periodically and guides can reference stale strings. Once a container is created, template version is irrelevant — `apt upgrade` keeps it current, no need to recreate existing containers on a newer template. |
| **OpenRGB** | Uninstalled — SMBus interaction with ADATA XPG GAMMIX D35 poses hardware risk. Do not reinstall. |
| **Hairpin NAT** | Slovak Telekom router blocks hairpin — test external domains from mobile data, not LAN |
@ -131,6 +131,7 @@ Script location: `/usr/local/bin/ddns-update.sh` in CT 108. Full credentials map
| **Koillection upgrades** | Always back up PostgreSQL DB before upgrading — data migrations can have edge cases |
| **Radicale + Authelia** | CalDAV clients use Basic Auth on every request, not browser SSO — bypass Authelia for `cal.spendlik.sk`, rely on Radicale htpasswd + HTTPS instead |
| **pveproxy crash** | `Can't call method "process" on an undefined value at pveproxy.pm line 304` — seen 2026-07-28, likely NFS-cascade freeze (host + SSH fully unresponsive). Fix: physical power cycle. Confirmed **all CTs (100–115) + VM 110 now set to "Start at boot"** as of 2026-07-28, so recovery should be automatic next time. |
| **NVIDIA driver on host shows as unavailable** | `nvidia-smi` failing on the Proxmox host with "couldn't communicate with the NVIDIA driver" is **expected**, not a fault, whenever VM 110 is running with GPU passthrough active — the GPU is owned by `vfio-pci` for the VM, not the host. Only investigate if VM 110 is stopped and the host still can't see the GPU. |
---
@ -153,6 +154,14 @@ Script location: `/usr/local/bin/ddns-update.sh` in CT 108. Full credentials map
---
## ✅ Recently Resolved
| Item | Resolution |
|---|---|
| Accidental `apt upgrade -y` on Proxmox host (2026-08-11) | Ran against host instead of a container; pulled in `6.14.11-9-pve` + `7.0.x` kernel packages. Boot pin held throughout, no actual impact on running kernel or NVIDIA passthrough. Sealed with `proxmox-boot-tool kernel pin 6.14.11-5-pve` (reaffirmed) + `apt-mark hold` on all 5 non-pinned kernel packages. Verified via `apt-mark showhold`. |
---
## 🧠 Life OS — Obsidian Vault Structure
| Folder | Purpose |