Ceph storage for Proxmox VE clusters
Ceph turns the local disks of your Proxmox VE servers into shared, redundant storage. inett sizes the hardware and network, builds the cluster and supports it in production.
The problem with a central storage array
Live migration and high availability need storage that every node in a virtualisation cluster can reach. Traditionally that means a SAN or NAS. The array is expensive, ties you to one vendor and is often a single point of failure itself. When it fills up, you buy a larger model or an expansion shelf at the vendor’s price.
Many organisations leaving VMware face this question at the same time: keep the old array, buy a new one, or move storage into the servers.
What inett does
Ceph is open source software-defined storage. It spreads data across many disks on several servers and keeps multiple copies of everything. When a disk fails, Ceph rebuilds the missing copies on other disks automatically. When a whole server fails, virtual machines keep running on the remaining copies.
Proxmox VE ships with Ceph. You install and manage it from the web interface or with the pveceph tool. The same servers then provide compute and storage, which is what hyper-converged means. VM disks live in Ceph RBD as block devices. CephFS holds ISO images, templates and shared files. inett is a Proxmox Gold Partner and builds Ceph clusters in this integrated form as well as standalone storage clusters.
Ceph is unforgiving of poor sizing. Too few nodes, a slow network or a RAID controller in the wrong place lead to poor performance, or to data loss when something fails. So we start with the hardware:
- Nodes. At least three servers, ideally identical. With three copies per object, the cluster survives the loss of one server and stays writable.
- Network. A dedicated Ceph network of at least 10 Gbit/s, and 25 Gbit/s or more with NVMe drives. Proxmox VE cluster traffic (Corosync) runs on a separate link. For three to five nodes without fast switches, a full-mesh network is an option.
- Disks. Enterprise SSDs, which shorten recovery time after a failure, especially in small clusters. Disks connect through an HBA, not a RAID controller.
- Memory. Proxmox recommends at least 8 GiB of RAM per OSD (one OSD per disk) on top of what the VMs need.
We then build the cluster, configure pools and placement rules, and test a node failure with you before go-live. In operation we monitor cluster health, apply updates and add disks or servers as you grow.
We also run Ceph without Proxmox VE, as a standalone storage cluster built with cephadm. It provides:
- block storage (RBD) for other virtualisation platforms or Kubernetes
- a file system (CephFS), also over NFS if required
- object storage with an S3 interface via the RADOS Gateway, for example as a backup target
Planning, deployment, operation and expansion work the same way as for a hyper-converged cluster.
Technical details
- Ceph releases in Proxmox VE 9.2
- Ceph Tentacle 20.2 and Ceph Squid 19.2 (as of October 2026). The Ceph project lists 31 October 2026 as the estimated end of life for Squid (Ceph releases)
- Minimum size
- three nodes, ideally identical; Proxmox recommends at least 12 OSDs, evenly distributed
- Replication
- default pool with 3 replicas (size 3) and min_size 2; Proxmox explicitly advises against min_size 1
- Network
- dedicated Ceph network of at least 10 Gbit/s; 25 Gbit/s or more with NVMe; Corosync on a separate link, e.g. 1 Gbit/s
- Disks
- SSDs recommended, especially in small clusters; HBA rather than RAID controller
- Memory
- at least 8 GiB per OSD recommended (default 4 GiB), plus headroom for recovery
- Interfaces
- RBD for VM and container disks, CephFS as a file system
- Source
- Proxmox VE administration guide, Ceph chapter
What is included
- Review of capacity, performance and availability requirements
- Hardware and network design for the hyper-converged cluster
- Ceph installation through Proxmox VE
- Set-up of monitors, managers, OSDs, pools and CephFS
- Disk and node failure tests before go-live
- Documentation and a handover session for your team
- Ongoing operation, updates and expansion through our managed services on request
Who it is for
Ceph suits you if you plan a Proxmox VE cluster of three or more servers and do not want to buy or replace a SAN. It grows with you: add disks or servers, and Ceph rebalances the data.
It is the wrong choice for one or two servers. There, local ZFS with Proxmox VE’s asynchronous storage replication works better. If your network cannot provide 10 Gbit/s for storage, plan the network first.
Frequently asked questions
Why at least three servers?
By default Ceph keeps three copies of each object on different servers. The Ceph monitors also need a majority to make decisions. With three servers, the cluster survives the loss of one.
How much of the raw capacity can I use?
With three replicas, roughly a third. Keep additional space free so Ceph can redistribute copies after a failure.