mcp: session-notes — PVE Memory Audit, Static IP Confirmation, and Tailscale DNS Decision

Sources/Homelab/2026-05-25-pve-memory-and-dns-audit.md

@@ -0,0 +1,41 @@
+---
+created: '2026-05-25'
+path: Sources/Homelab
+project: homelab-infrastructure
+tags:
+- proxmox
+- tailscale
+- dns
+type: session-notes
+---
+
+## Outcome
+
+Audited herbys-dev memory allocation, confirmed PVE host already has a static IP, and surfaced (but deferred) a Tailscale DNS resilience question.
+
+## Topics Covered
+
+- Docker memory totals on herbys-dev (~2.56 GiB across 16 containers)
+- htop vs PVE memory accounting discrepancy explained (page cache vs process memory)
+- herbys-dev VM right-sized: 24576 MiB → 14336 MiB (applies on next reboot)
+- traefik VM (2 GiB allocated / ~500 MiB used) and knot LXC (512 MiB / ~29 MiB used) reviewed — both left alone
+- PVE host /etc/network/interfaces confirmed: static IP `10.0.21.188/24` on `vmbr0`, gateway `10.0.21.1`, VLAN-aware bridge with vids 2-4094
+- PVE `/etc/resolv.conf` is managed by Tailscale, pointing at `100.100.100.100`
+- Split-DNS confirmed working (resolves both `pihole.herbylab.dev` → 10.0.11.20 and public domains)
+
+## Key Learnings
+
+- **Linux memory accounting**: `free -h` inside the VM is the source of truth. The `available` column is what matters, not `used`. PVE's outside view conflates process memory with page cache and shows allocations as "in use" when they're really evictable. The fear-bump from 24 → 26 GiB was based on misreading PVE.
+- **VM allocation ≠ host commitment, with a caveat**: LXC limits are caps and cost nothing extra. VM allocations are reserved unless ballooning is configured with `balloon:` set below `memory:`. Default PVE behavior is effectively no ballooning.
+- **`memory:` in qm config is MiB**: 14 GiB = 14336, 12 GiB = 12288.
+- **Tailscale `--accept-dns=true` (default) hijacks `/etc/resolv.conf`**: the daemon rewrites it to point at `100.100.100.100` on every start. `--accept-dns=false` tells Tailscale to leave DNS config alone, persists across reboots, costs `*.ts.net` shell resolution but nothing else. VPN tunnel is unrelated to DNS setting.
+- **PVE depending on Tailscale for DNS is a resilience smell**: a hypervisor that loses DNS when a userland VPN daemon hiccups is a bad pattern. Functional today, worth hardening eventually.
+- **Long uptime increases risk over time**: untested boot paths, config drift between running state and on-disk state, accumulated updates. Periodic planned reboots are hygiene, not paranoia.
+
+## Follow-ons
+
+- [ ] Reboot herbys-dev when convenient to apply the 14 GiB memory change
+- [ ] Pre-reboot smoke check: `docker ps --format "table {{.Names}}\t{{.Status}}\t{{.RestartPolicy}}"` to confirm every container has a restart policy
+- [ ] Detach the ubuntu-24.04.4 live-server ISO from VM 100 (`ide2` line in qm config) — install cruft, not urgent
+- [ ] **Optional hardening**: switch PVE host off Tailscale DNS. Run `tailscale set --accept-dns=false`, add `dns-nameservers 10.0.11.20` to the `vmbr0` stanza in `/etc/network/interfaces`, then `ifreload -a`. Removes Tailscale as a DNS dependency for the hypervisor.
+- [ ] Try `dj0024javia.gitea-vscode-pullrequest` extension for VS Code → Lovebug PR review flow without the browser