


Contents (13 sections)
- Why back up, when you host for others
- High availability is not a backup
- Sort your data before protecting it
- Our setup in 2026, copy by copy
- Copy 1: a backup server outside the cluster
- Copy 2: a mirror in France, outside the hosting provider, out of reach
- On top of machines: databases, code, witnesses
- 3-2-1-1-0: where we really stand
- Which protection against which risk?
- The limits of our strategy
- A backup is only worth something if it restores
- Your turn: a six-point action plan
- Sources
On 12 November, Fx published an article on Cachem, a French NAS blog, that I recommend: "Sauvegarde 2026: comment je protège réellement mes données" (Backups in 2026: how I really protect my data). He describes his personal strategy, that of a NAS enthusiast with a family and photos that must not be lost: a Synology NAS, an external drive unplugged after each backup and stored in a fireproof case, an encrypted copy in OneDrive. And, rarer still, he writes down the limits of his own setup in black and white.
I wanted to do the same exercise, with the same outline, for Sanfytech. Our data is not holiday photos: it is the databases of our clients' applications, their files, their websites, our code. You cannot put a server cluster in a case. But the questions are exactly the same: what is not a backup, how many copies, where, out of whose reach, and how do you know the restore will work on the day you need it.
Why back up, when you host for others #
Fx tells how, when their first child was born, his partner was very clear: not a single photo may disappear. We have the same rule, simply written into contracts. An association that loses its member database, a school that loses its report cards, a small business that loses its invoices: none of them will care whether a disk, an operator or an attacker was to blame.
And the nature of the risks has changed. Hardware still fails without warning: we wrote about an NVMe drive failure that slowed one of our servers down for weeks without ever bringing it down. But the threat that shapes everything is ransomware. A competent attacker does not encrypt your servers the day they get in: they look for the backups first, and destroy them. Fx's conclusion holds at every scale: never depend on a single device, a single service, or a single set of credentials.
High availability is not a backup #
Fx insists that RAID is not a backup. Our equivalent is even more misleading, because it is more impressive. Our production runs on three dedicated servers, in three French data centres (Strasbourg, Roubaix, Gravelines), joined in a Proxmox cluster. Machines enrolled in failover are copied to another server every fifteen minutes; if a server becomes unreachable, the cluster restarts them elsewhere, with no one touching anything. We wrote about what the tutorials leave out of this high availability.
It is excellent protection against losing a disk, a server, even a whole data centre. It is no protection at all against a table deleted by mistake, an update that corrupts a database, or ransomware: replication faithfully copies the mistake to the neighbouring server a quarter of an hour later. Service continuity and backup answer two different questions. The first asks "is the service still running?", the second "can I go back to yesterday?".
Sort your data before protecting it #
Fx sorts his files into three categories: important, useful, of little value. It is the most underrated advice in his article, and it applies as is to a business. Here is how we sort ours:
- Irreplaceable: the databases and the files uploaded by the users of the applications we host, our code forge with its full history, our secret stores. Nobody can rebuild them.
- Rebuildable: the machines themselves, their configuration, our application images. All of it is described in versioned repositories and can be rebuilt from code. Rebuilding takes time, which is why we back it up anyway, but nothing is lost without it.
- Disposable: the working folders on our workstations and of our development agents, caches, build output. Our rule is simple: the forge is the source of truth, everything else is a local copy. What has not been pushed to the forge does not exist.
Sorting also says what you agree to lose. One of the projects we host has a 200 GB video volume deliberately left out of backups. It is a decision, not an oversight; it still has to be written down, so that nobody discovers it on the day of a restore.
Our setup in 2026, copy by copy #
Fx has a NAS, an external drive and a cloud. We have production, a backup server and a mirror. The principle is the same; what changes is that everything has to run with no manual step, every night, for some forty machines.
Copy 1: a backup server outside the cluster #
- Every night at 10 pm (Paris time), every machine in the cluster is backed up by Proxmox Backup Server, whose deduplication and retention we have covered in detail. The three machines with a special profile have their own weekly job.
- On a dedicated server, in a fourth data centre, in Germany. Outside the cluster, by design: a backup server that belongs to the cluster it backs up goes down with it.
- Encrypted at the source. Data is encrypted on the originating machine before it leaves: the backup server only receives blocks it cannot read. It is the equivalent of the encryption Fx applies before sending his files to OneDrive.
- Write-only for production. A backed-up machine can drop off its backups, not delete them: pruning is driven by the backup server, with other credentials. Each client has its own sealed space.
- Re-read every day. A verification job re-reads every night the backups that have never been checked, and a full monthly re-read covers everything older than 30 days. It is the "check backup integrity" option of Hyper Backup that Fx recommends, automated.
Copy 2: a mirror in France, outside the hosting provider, out of reach #
Every night, a second copy goes to a mirror located in France, with a different provider from the one running production. Two details make all the difference. First, it is the mirror that pulls the backups: the backup server holds no access to it and does not even know it exists. Second, the mirror is on the "never" list of our SSH bastion: none of our usual administration tools can connect to it, neither us nor the AI agents that assist us. If an attacker took over production, then the backup server, then our administration access, they would still find no path to this copy.
On top of machines: databases, code, witnesses #
- Daily database dumps for several applications, on top of machine backups: restoring a single database to yesterday, without bringing back the whole machine, takes a few minutes.
- The code lives on a forge we host, backed up like any machine in the fleet. Every clone on a workstation is also a full copy of the history, and the main branch of each repository is protected against direct pushes: a slip of the hand does not rewrite history.
- One witness per backup in our monitoring, which runs on a separate server outside the cluster. Fx notes that a manual backup depends on your discipline; an automatic backup depends on someone noticing it has stopped. The witness worries about silence: if a backup does not check in, the alert goes out.
Our latest published measurements, dated 5 September 2026: 322 out of 322 backups completed in seven days, 7 out of 7 integrity re-reads with no error, 1.4 TB replicated to the mirror. They are listed with their method on the sovereign hosting page: a figure without its date and method is not a measurement.
3-2-1-1-0: where we really stand #
Fx starts from the 3-2-1 rule and adds its two modern requirements: one offline or immutable copy, and zero errors in checks and restores. We have already explained it in detail in our article on immutable backups. Here is our scorecard, digit by digit, without rounding up:
- 3 copies: production, the backup server, the mirror. The high-availability replica does not count: it copies mistakes.
- 2 different media: the hypervisors' disks, the backup server's disks, then another provider's infrastructure. Three systems, three sets of credentials.
- 1 off-site copy: both backup copies are outside the production data centres, and the mirror is even outside the hosting provider. A suspended account, a billing error or a disaster at our main provider would not take everything with it.
- 1 immutable or offline copy: halfway. Our copies are protected by architecture (write-only, pull-mode mirror, unreachable by our tools), not by being physically unplugged. More on this under the limits.
- 0 errors: integrity re-reads are daily and error-free over the published period, and a full restore has been carried out and documented. What is missing is a regular drill.
Which protection against which risk? #
Fx sums up his setup in a risk-by-risk table. It is the most useful exercise in his article, and I redo it with our building blocks:
- A failed disk or server: high availability is enough. Machines restart on another server, with at most fifteen minutes of data to catch up.
- Losing a whole data centre: same answer, the other two take over; the backups are elsewhere anyway.
- Accidental deletion, an update that breaks a database: replication is useless, it has already copied the mistake. Last night's dump or backup is what helps, up to a month back.
- Ransomware on a production machine: the machine cannot delete its backups. We restore the state from before the infection, provided it was detected within the retention window.
- Our administration access compromised: the mirror, which none of that access can reach.
- A major incident at our main hosting provider: the mirror, with another provider, in France.
- Losing the code forge: its backup, and every full clone on the team's workstations.
The limits of our strategy #
This is the part of Fx's article I liked best, so I hold myself to it too. Our internal audit of August 2026 raised these points; they are known, and not all of them are solved.
- Short retention on production. For hosted machines, we keep 7 daily, 2 weekly and 1 monthly backups. An intrusion discovered after six weeks could no longer be restored from the backup server. Both levers are known: detect earlier, hence every machine's logs centralised in our SIEM, and lengthen retention, which deduplication makes cheap.
- No strictly offline copy. We have no equivalent of Fx's drive in its case. Our "1" rests on architecture: it protects against an attacker who holds our servers or our access, not against a flaw in the backup software itself, which would affect every copy it manages. ANSSI, the French cybersecurity agency, recommends an offline copy; we honestly describe ours as "out of reach", not "offline".
- The "0" is not yet a habit. A full restore has been carried out and documented, which is already rare. But a one-off restore is not a yearly drill with a dated report, and only the latter proves the procedure still holds after a year of changes.
Fx points out a trap worth turning on yourself: without the encryption key, an encrypted backup is unreadable. Since our backups are encrypted at the source, the key must not live only inside the infrastructure it protects, and the restore procedure must be something a person other than its author can follow. On a fleet as on a family NAS, backups must depend neither on a single machine nor on a single person.
A backup is only worth something if it restores #
Fx distinguishes three levels of proof, and I take them as they are because they are right. Checking the integrity of the copies: for us, automatic and daily. Restoring a sample to another location and opening the files: that is what restoring a database to yesterday on a test machine gives you. Once a year, going all the way to a full restore, stopwatch in hand: that is where the real recovery time is measured, the one our article on disaster recovery plans describes.
One last word on a reader's comment under the Cachem article: OneDrive as an off-site copy would suit a business or a public body poorly. That is right, and it is not only about reliability. For client data, the remote copy must stay under European jurisdiction, with a provider that is not subject to the Cloud Act, and encrypted with a key you hold. That is why both our copies stay in the European Union, in Germany and in France, and why the server that receives them cannot read them.
Your turn: a six-point action plan #
Fx's plan, adapted to a small business that runs on a few servers or a hosting contract:
- Sort your data into irreplaceable, rebuildable and disposable, and write down what you agree to lose.
- Separate backup from production: another machine, another site, other credentials. The office NAS, on the same network as the servers, is not off-site.
- Put one copy out of reach, immutable or unplugged. The test: from a compromised production server, with all its passwords, can someone delete your backups? If the answer is not a demonstrable "no", it is not done.
- Encrypt at the source everything that leaves your premises, with a provider under European law.
- Keep keys, access and procedure outside the infrastructure, known to at least two people.
- Watch for silence and actually restore: an alert when a backup does not check in, and a full restore once a year, in the calendar and done.
Fx ends with a question I put to you again: when did you last restore your data? If the answer is "never" or "I don't know", you have copies, not yet a backup. This setup is the one behind our Proxmox managed services and our sovereign hosting; the infrastructure that carries it is described in our case study. If you want to put yours through the six points, let's talk.
Sources #
- Cachem, Fx, Sauvegarde 2026 : comment je protège réellement mes données (12 November 2026, in French), readers' comments included.
- ANSSI, Backing up information systems: the fundamentals (25 October 2023, in French).
- Proxmox, Proxmox Backup Server documentation: remotes and sync jobs, pruning, garbage collection and verification.
- Sanfytech, operating measurements of 5 September 2026 (backup frequency, success and re-reads, retention, copy locations), published with their method on the sovereign hosting page.






