Ceph Expertise: Distributed Storage for Private Cloud

Ceph is the open-source distributed storage system behind a large share of private clouds. LaMeDuSe Cloud designs, deploys and runs Ceph clusters on its own infrastructure: volumes for virtual machines and Kubernetes, a shared file system and S3-compatible object storage.

The problem you're facing

Traditional storage arrays are expensive, lock you into a vendor and sooner or later hit their capacity limits. Ceph solves all three, but a poorly designed or poorly run cluster can become slow, or even lose resilience just when a disk fails. Storage is the layer where there's the least room for error.

Our answer

We design Ceph clusters suited to your workloads: hardware and network choices, replication or erasure coding level, failure domains, NVMe and capacity storage classes. Then we run them: detailed monitoring, disk replacement, rebalancing, upgrades and regular resilience tests.

How Ceph works

Ceph is built on RADOS, a distributed, self-managing object store. Each disk in the cluster is handled by an OSD (Object Storage Daemon) that stores data, replicates it to other OSDs and takes part in recovery after a failure. Monitors (MON) maintain the cluster map and guarantee its consistency; managers (MGR) collect metrics and provide administration modules.

What sets Ceph apart is the CRUSH algorithm: instead of looking up a central index to find where data lives, each client computes the location itself from the cluster map. There is no single bottleneck, and the cluster can grow from a few servers to several hundred simply by adding nodes.

  • OSD: one daemon per disk, storing and replicating data
  • MON: the cluster map and consensus
  • MGR: metrics, dashboard and orchestration modules
  • MDS: CephFS file system metadata
  • RGW: S3-compatible object gateway

RBD: block storage

RBD (RADOS Block Device) provides virtual disks split into objects spread across the whole cluster. It's the natural storage for OpenStack and Proxmox virtual machines, and for Kubernetes persistent volumes through the CSI driver. Each volume benefits from the combined performance of many disks and the resilience of replication.

RBD supports snapshots, instant clones (a virtual machine boots in seconds from an image, without copying), online resizing and asynchronous replication to a second cluster for disaster recovery.

CephFS: the shared file system

CephFS is a POSIX-compatible file system mounted simultaneously by many servers or pods. It suits directories shared between several instances of an application, media repositories, compute data and Kubernetes volumes with multiple read-write access (ReadWriteMany).

Metadata (directory tree, permissions, file names) is handled by dedicated MDS servers, which we deploy in active and standby mode to avoid any interruption. Per-directory quotas and snapshots allow fine-grained space management and quick restoration of a file deleted by mistake.

S3-compatible object storage

The RADOS Gateway exposes the cluster through the S3 API, the same one used by most modern applications, backup tools and data analytics frameworks. You can store backups, user files, logs or datasets there, with buckets, access policies, versioning and object lifecycle management.

It's a sovereign alternative to hyperscaler object storage services: the same API, data hosted in France, Germany or Switzerland, and no per-request fees to budget case by case.

Replication, erasure coding and failure domains

By default we replicate every piece of data three times on three different servers: the cluster keeps running normally even if an entire server stops. For large, less frequently accessed data (archives, backups, object storage), erasure coding splits data into fragments and adds parity fragments: resilience is comparable for far less space.

CRUSH rules define the failure domains: disk, server, rack or room. We configure them so that no foreseeable hardware failure (a server, a power supply, a switch) can make data unavailable. The choice between replication and erasure coding is made pool by pool, based on the performance and cost you're after.

  • 3x replication for production volumes
  • Erasure coding for archive and object data
  • Failure domains per server, rack or room
  • Separate NVMe and capacity pools

Resilience: what happens when a disk fails

In a healthy Ceph cluster, a disk failure has no impact on your applications. The cluster detects the failed OSD, marks it out, then automatically rebuilds the missing copies from the other replicas, spreading that work across all disks. Our teams then replace the disk and the cluster rebalances.

The real question is controlling that recovery: it must neither take too long nor degrade performance enough to disrupt production. We size the storage network, tune recovery priority and keep enough spare capacity to absorb the loss of an entire server. We test these scenarios before go-live and regularly afterwards.

Architecture and sizing

A Ceph cluster's performance depends as much on its architecture as on its disks. We separate the public network (client access) from the internal replication network, size network links so recovery doesn't saturate production, balance the number of disks per server, and reserve enough CPU and memory for each OSD.

We start with at least three storage servers, then grow the cluster as you need. Deployment uses cephadm, the official orchestration tool, or Rook when Ceph is integrated into a Kubernetes cluster.

Day-to-day operations

We monitor cluster health, latency and throughput for each pool, fill rate, disk condition (SMART indicators) and the progress of the scrubbing operations that continuously verify data integrity. Alerts are handled by our 24/7 on-call team.

We roll out Ceph upgrades progressively, daemon by daemon, without service interruption, and plan capacity additions early enough never to exceed the recommended fill thresholds. A full cluster is a fragile cluster: it's one of the most common mistakes we fix during our audits.

What LaMeDuSe Cloud takes care of

  • Ceph architecture: hardware, network, failure domains
  • RBD block storage for OpenStack, Proxmox and Kubernetes
  • CephFS shared file system
  • S3-compatible object storage (RADOS Gateway)
  • Replication or erasure coding depending on use
  • Monitoring, disk replacement and upgrades
  • Backups and replication to a second site
  • Audit and takeover of existing Ceph clusters

Our methodology

  1. 1

    Context audit

    Volumes, criticality, hardware constraints and goals: context dictates the architecture.

  2. 2

    Tailored architecture

    Argued recommendations, sized for your real loads and peaks.

  3. 3

    Implementation

    Deployment by teams who run the same stack on their own production.

  4. 4

    Handover

    Documentation, training and knowledge transfer for your autonomy.

Technologies

CephRBDCephFSRADOS Gateway (S3)cephadmRookPrometheus & Grafana

Security

Hardened baseline configuration, secrets managed off the servers, encrypted backups and regular restore tests.

Hosting & infrastructure

Our own production runs on these technologies in our European datacenters: expertise is proven by operations.

Maintenance

Version tracking, security watch and support by the teams who administer the same stack daily.

Why LaMeDuSe Cloud

Ceph is at the heart of our own infrastructure: we run it daily in our Paris, Frankfurt and Zürich datacenters for our storage, private cloud and managed Kubernetes offers. Our engineers intervene on hardware within 2 to 4 hours, which limits how long a cluster runs in degraded mode.

Frequently asked questions

LaMeDuSe Cloud runs Ceph daily on its own infrastructure, and can deploy, audit or take over the operation of Ceph clusters for its clients, with 24/7 monitoring.

At least three storage servers for three-way replication across three separate machines. In practice we often recommend more nodes to ease recovery and growth.

Yes, with an NVMe pool and a properly sized network. For the most latency-sensitive databases, we can also combine Ceph with local NVMe disks, depending on the application's architecture.

RBD provides virtual disks attached to one machine or pod at a time, ideal for virtual machines and databases. CephFS is a shared file system mounted by several servers at once.

No. Replication protects against hardware failures, not accidental deletion or ransomware. We set up snapshots and off-site backups, to another cluster or another datacenter.

Yes. The RADOS Gateway exposes an S3-compatible API: most applications and backup tools work unchanged, with data hosted in Europe.