Ceph: A Viable Alternative Object Store
March 1, 2026
Why Ceph’s RADOS architecture makes it a compelling alternative to MinIO.
I recently found out that MinIO is no longer actively maintained. That sent me back to an object store I'd shelved a while ago because it looked like too much work: Ceph.
At first glance, Ceph is intimidating. It hands you three different storage interfaces and lets you figure out the rest:
- CephFS — a POSIX-compliant distributed filesystem
- Ceph RBD — a block storage device
- Ceph RADOSGW — object storage with an S3-compatible API
Three products, three mental models to learn? Not quite. Once you dig in, the trick reveals itself:
All three interfaces are just different front-ends to the same underlying system.
That system is RADOS — the Reliable Autonomic Distributed Object Store. Everything Ceph does eventually flows through it, which is what makes the whole thing click once it clicks.
The core model
Under the hood, Ceph is a distributed object store built on a primary-copy replication model.
Data is bucketed into Placement Groups (PGs), and each PG is assigned a primary OSD — the leader responsible for sequencing every write to that group. This "leaderful" design isn't an accident. Because Ceph also serves block and filesystem workloads, it has to handle partial writes: change 4 KB in the middle of a 1 GB volume without rewriting the whole thing.
Partial writes mean read-modify-write, and read-modify-write means you need someone in charge to keep updates consistent. That someone is the primary OSD.
How MinIO differs
MinIO makes the opposite bet. It's a leaderless, quorum-based system built purely for object storage, and it sidesteps the coordination problem by refusing to do partial writes at all:
- Objects are never modified in place.
- An "update" is just a brand-new version of the object.
With immutability as the rule, there's nothing partial to coordinate. Any node can accept a write, and as long as a strict quorum acknowledges it, the object is safely stored. No leader required.
The trade-off
| | Ceph | MinIO | |---|---|---| | Coordination | Leaderful (primary-copy) | Leaderless (quorum) | | Workloads | Object + block + filesystem | Object only | | Wins at | Versatility | Raw object throughput | | Pays for it with | Coordination overhead | Less flexibility |
Ceph pays what I think of as a "management tax": placement-group peering, cluster-map propagation, primary coordination. In exchange, it can be an object store, a block device, and a filesystem — all from one cluster.
MinIO skips the tax entirely and pours everything into a lean, fast data path. But that speed comes from doing exactly one thing.
Takeaway
Ceph looks like three storage systems bolted together. It isn't. It's one distributed object store with three interfaces layered on top, and once you internalize that everything is RADOS underneath, the intimidation mostly evaporates.
If you want to go deeper, read the original RADOS paper by Sage Weil, Ceph's creator. It's short, and it's where all of this starts.