Solution
Managed Proxmox Cluster
Three nodes with quorum, HA Manager and Ceph, operated around the clock.
Solution · Storage
In brief
Ceph is distributed storage that keeps multiple copies of data on different nodes and restores redundancy on its own after failures. In a managed Ceph cluster from managed-cluster.net, every data block exists three times as standard; we build Ceph into your Proxmox VE cluster and run it 24/7, from €449.99 net per month with the Essentials package. Replication does not replace backups.
How it works
Ceph spreads your virtual machines' data across the disks of every node; there is no central storage system. Each data block is stored three times on different nodes, and a write is only acknowledged once the copies on the other nodes are in place. That is why any virtual machine can start on any node.
In a managed Proxmox cluster, virtual machines and Ceph storage run on the same nodes; Proxmox calls this a hyper-converged setup. There is no separate storage system such as a SAN (storage area network), and with it goes a central component whose failure would hit every node at once.
Each disk is managed by its own Ceph service, the OSD (object storage daemon). Ceph computes where a block lives with the CRUSH algorithm, without a central lookup table. By default, Proxmox VE creates three copies (size 3); at least two must be reachable for reads and writes to continue (min_size 2).
Replication runs over a dedicated storage network, separate from management and public traffic. Proxmox highly recommends this separation because recovery traffic could otherwise interfere with cluster communication over Corosync. If a node fails, the Proxmox HA Manager restarts its virtual machines on another node; see how high availability in the cluster fits together.
Sources: Proxmox VE Administration Guide, “Deploy Hyper-Converged Ceph Cluster” (version 9.2); Ceph documentation, “Architecture” (Tentacle release); retrieved October 9, 2026 · As of October 2026
Failure behavior
If a disk or an entire node fails, no data is lost: Ceph keeps every block in three copies on different nodes and restores the target redundancy on its own as soon as enough disks and nodes are available for it. Recovery increases I/O load; with several simultaneous failures, some data may be temporarily unavailable until recovery completes.
| Scenario | Your data | What Ceph and the cluster do | What we do |
|---|---|---|---|
| One disk fails | Available No data loss; reads and writes continue. | Ceph rewrites the missing copies to the remaining disks (self-healing); I/O load rises until it finishes. | On-call duty is alerted, we replace the disk, you get a post-incident report. |
| One node fails | Available No data loss; every block still exists twice. | The HA Manager restarts the node's virtual machines on another node. Because each copy must sit on a different node, a three-node cluster only regains its third copy once the node is repaired. | On-call duty is alerted, we analyze the fault and replace the hardware, you get a post-incident report. |
| Two disks fail at the same time | Partial No data loss; some data temporarily unavailable. | Same node: every block remains reachable twice. Different nodes: some data has only one copy left, and I/O to it pauses until Ceph has restored a second copy. | On-call duty is alerted, we replace both disks, you get a post-incident report. |
Assumption: Proxmox default of size 3 and min_size 2. According to the Ceph documentation, a replicated pool can lose one OSD fewer than it has copies, all at the same time, without losing data. Sources: Proxmox VE Administration Guide; Ceph documentation, “Pools” (Tentacle release), retrieved October 9, 2026 · As of October 2026
At the compute layer, the virtual machines of a failed node restart on another node. Services running on several nodes behind the load balancer remain available. Our high availability cluster page walks through the step-by-step failover.
Decision guide
The Essentials and Business packages use Ceph. We use ZFS replication in individually designed, more compact setups. The key difference is timing: Ceph writes synchronously to three nodes, while ZFS replicates at intervals. That interval determines how much data can be lost when hardware fails.
| Criterion | Ceph | ZFS replication |
|---|---|---|
| Minimum size | three nodes | more compact setups, including two nodes with an external quorum device |
| Replication | synchronous, three copies on different nodes | at intervals, configurable from every minute to once a week according to Proxmox |
| Impact on the RPO (recovery point objective) after a hardware failure | Acknowledged writes already exist on several nodes; losing a disk or a node loses no data. | Changes since the last sync can be lost; the interval sets the RPO. |
| VM restart after a node failure | on any remaining node | on a node holding a replica, with the data of the last sync |
| In the packages | Yes Essentials and Business | No individually designed setups |
| Replaces backups | No replicates errors and deletions too | No replicates errors and deletions too |
Source on ZFS replication: Proxmox VE Administration Guide, “Storage Replication” (version 9.2), retrieved October 9, 2026. More on RPO and replication methods · As of October 2026
Which method fits depends on data volume, I/O profile and RPO target; we work this out during design and sizing.
Ceph is standard in the cluster packages: Ceph HA from the Essentials package at €449.99 net per month, with three nodes and full operations.
All prices are net, plus VAT.
Capacity
With three-way replication, roughly a third of the raw capacity is available for data. You should not fill all of it: Ceph needs free space to rewrite missing copies after a disk failure, and by default it warns once a disk is 85 percent full.
| Step | Calculation | Result |
|---|---|---|
| Raw capacity | All disks of the three nodes combined | 100% |
| Usable with three copies | 100% ÷ 3 | about 33% |
| Up to the “nearfull” warning | 33% × 0.85 (Ceph default) | about 28% |
| With headroom for one disk failure per node | A failed disk's data spreads across the node's three remaining disks: 0.85 × ¾ ≈ 64% utilization beforehand, so 33% × 0.64 | about 21% |
Simplified, assuming even distribution and excluding metadata. Ceph defaults: “nearfull” 0.85, “full” 0.95. Source: Ceph documentation, “Monitor Config Reference” (Tentacle release), retrieved October 9, 2026 · As of October 2026
The last row follows from placement: because each copy must sit on a different node, a failed disk's data in a three-node cluster lands on the remaining disks of the same node. If a disk reaches the “full” threshold, Ceph blocks I/O to prevent data loss, according to its documentation.
Larger disks increase storage density, but after a failure Ceph then has to recover more data at once (per Proxmox). We discuss the hardware of the cluster nodes (CPU, RAM, storage, network) individually and define it in the design proposal, together with headroom for disk failures, load peaks and growth.
Operations
We monitor your Ceph storage around the clock as part of your cluster: disks, storage network and Ceph services. Critical alerts reach on-call duty immediately and are handled right away, at any hour. We replace failed disks, and every incident ends with a post-incident report on cause and remediation.
In the FRA01 data center in the Rhine-Main region, failed components are replaced around the clock via remote hands.
Scope
Ceph protects against hardware failure, not against errors in your data. It writes every change synchronously to all copies, including an accidental deletion or encryption by ransomware. To recover, you need an earlier, undamaged copy of your data outside the cluster.
A high availability cluster is not a backup.
Backups run on infrastructure separate from the cluster, with defined retention periods and regular restore tests. In the Business package they are stored, and restore-tested, at the second site NBG01. Background and recovery targets: why replication is not a backup.
FAQ
From three nodes, usually Ceph; for more compact setups, ZFS replication. Ceph replicates synchronously and lets any virtual machine start on any node; the Essentials and Business packages use it. ZFS syncs data at intervals. We make the call during the design phase, based on data volume, I/O profile and RPO target.
At least three. Ceph places the three copies of a block on three different nodes, and Proxmox also requires at least three servers for a hyper-converged Ceph cluster. Smaller environments can use ZFS replication; see Ceph vs ZFS compared.
No data is lost. Ceph rewrites the missing copies to the remaining disks on its own, which temporarily increases I/O load. Our on-call duty is alerted, we replace the disk and document the incident in a post-incident report.
About a third of the raw capacity, before headroom. Because Ceph needs free space for recovery and warns at 85 percent utilization, plan with about 21 percent of the raw capacity in the example of three nodes with four disks each. The capacity example shows the full calculation.
No. Replication also carries deletions and errors to every copy. Our backup and disaster recovery page explains how separate backups with restore tests cover this.
Next step
Tell us your data volume, I/O profile and RPO target. You will receive a written design with an architecture diagram, a component list and a fixed price for setup and operations.
Discuss storage requirementsRelated