Two configs, no backup#
Getting Proxmox Backup Server Running, and Keeping a Copy Offsite covered every VM’s disk. It didn’t cover two smaller, easy-to-miss things underneath all of it: Proxmox VE’s own cluster config, and the config inside Proxmox Backup Server (PBS) that describes how the backups actually work. Neither had a backup at all. Lose either one and the real backup data might still be sitting there, unreadable or useless without the config that explains what any of it is.
What’s actually on each host#
On the Proxmox VE side, this is /etc/pve, the cluster filesystem. It holds every VM and container’s config (qemu-server/<id>.conf, lxc/<id>.conf), storage.cfg, ceph.conf, firewall rules, HA resource definitions, datacenter-wide settings, and users and realms. Its root-only priv/ subdirectory holds the actual secrets underneath all of that: the cluster CA’s private key, API token secrets, Ceph authentication keys, and the password file for local users. Corosync replicates /etc/pve to every node in real time, so losing one node doesn’t lose any of it. Losing the whole cluster at once, or a slower corruption that replicates before anyone notices, still would.
On the PBS side, this is /etc/proxmox-backup: datastore.cfg, the remote.cfg entry pointing PBS at itself, sync.cfg, acl.cfg, user.cfg. None of this is the backup data itself. That part’s already offsite, covered by the previous post. This is the config that says what a datastore even is, who’s allowed to touch it, and how it gets mirrored. Lose it and the chunk store is still there, but nothing on the appliance knows what to do with it.
A VM’s config file holds basically its entire configuration: this ID is named X, has this much RAM and these cores, sits on this network bridge with this IP, has a disk on this storage backend, and carries these tags. It’s the only place in the entire cluster where any of that is actually recorded. Lose that file with nothing else around, and the VM keeps running right up until it needs to be rebuilt, at which point all of that comes from memory.
The cluster has over 20 VMs and containers with a config file like this. A PBS backup snapshot embeds its own copy of the VM’s config alongside the disk image, confirmed by listing a real snapshot’s contents, so a handful already had their config covered as a side effect of backing up their disks, not because anyone planned it that way. Everything else, plus the cluster-wide things that were never per-VM to begin with (HA registrations, firewall rules, the ACME account), had no record anywhere.
Choosing an encryption tool#
Both of these configs contain real secrets: API tokens, the ACME account, whatever’s sitting in user.cfg. Landing either of them on a NAS share, let alone shipping them offsite, meant copying all of that out of the one place that’s supposed to control access to it. Skipping it would have been simpler, but that wasn’t an option.
For a job that only ever needs to encrypt, never decrypt, asymmetric recipient-based encryption is the right tool. age does that: a single static binary, no keyring, no web-of-trust ceremony. I generated a keypair once, and made a point of not losing it this time. The public key is baked into the backup script and is safe to commit, since it can only encrypt to, never decrypt. The private key lives in my password manager, not on either host and not in this repository.
Land locally first#
Encrypted archives for both configs end up on the local NAS share, reusing what I already do for other data: an already-scheduled nightly mirror job carries everything on that share offsite. Zero changes were needed to that existing job, and nothing writes directly to the remote destination.
The Proxmox VE host’s path there looks a little different: it sits on a network that can’t reach the NAS directly, so its archive takes an extra hop through a different host that already has clean access, covered in the final setup below.
Checking the artifact#
I checked three things directly against the real files. Running file against the landed archive on both hosts’ destinations confirmed it was genuinely age encrypted file, X25519 recipient. Decrypting that same file with no key supplied failed with “no identity matched any of the recipients,” which is what unreadable actually looks like. I triggered the relay job for real and checked the resulting file from a third host, independent of both the writer and the relay.
What’s running now#
graph LR
pve["Proxmox VE host
/etc/pve, isolated network"] -->|encrypt with age, local retention| local1[("Local retention
on the PVE host")]
local1 -->|pulled over SSH| relay["Relay host
already has NAS access"]
relay -->|lands and prunes| nas[("NAS share")]
pbs["PBS host
/etc/proxmox-backup"] -->|encrypt with age| nas
nas -->|nightly mirror, existing job| offsite[("Offsite mirror")]
The Proxmox VE host encrypts /etc/pve nightly with age and keeps a short local retention window. The relay host, already living on a network with proven NAS access, pulls that archive over SSH, lands it on the NAS, and prunes an older history there. The PBS host encrypts /etc/proxmox-backup the same way and lands it on the NAS directly, carried offsite by the same nightly mirror job that already moves everything else. Both archives are unreadable without the private key, and that key lives in my password manager, nowhere near either host or this repository.
Both configs are backed up now: the one that defines the cluster and every VM in it, and the one that says how PBS protects all of it.



