The pieces

• RADOS — the reliable object store underneath everything.
• MON (monitors) — hold the cluster maps and form quorum. Run 3 (or 5).
• MGR (managers) — metrics, dashboard, modules. Run at least 2.
• OSD (object storage daemons) — one per disk; store data, replicate, recover.
• MDS — only for CephFS (file system).

Clients — Proxmox VMs using RBD block devices — talk directly to OSDs. There is no central controller in the data path, which is why Ceph scales.

CRUSH

Ceph does not keep a lookup table of where every object lives. The CRUSH algorithm calculates placement from the cluster map and the failure domain rules (host, rack, room). With size=3 and failure domain host, the three copies always land on three different servers.

Placement groups (PGs)

Objects are grouped into PGs; PGs are placed on OSDs. The PG autoscaler sets sensible counts — leave it on unless you have a reason.