Skip to main content

04 · Storage data paths

Two data paths, two disk owners: Longhorn owns the local SSDs and serves VM root disks; the external Rook/Ceph cluster owns the passthrough NVMe and serves the application cluster's durable data. · interactive original ↗

A · Longhorn — hyperconverged
DATA PATH A · LONGHORN (HYPERCONVERGED)VM root disks · images · fast local-tier PVCsVM / podvirtio-blkLonghorn engineon host1 replica, localstrict data-localitylocal SSD tiersame host as VM
B · External Rook/Ceph
DATA PATH B · EXTERNAL ROOK / CEPHbulk PVCs · shared FS · object storageApp pod PVCceph-csiCeph publicVLAN 210MON / OSDstorage RKE2 VMsNVMe OSDspassthrough · repl. VLAN 211

Data path A — Longhorn (hyperconverged)​

Scope​

VM root disks, images, fast local-tier PVCs.

Behaviour​

Root disks: replica 1 + strict data-locality — reads/writes stay on the host, zero storage-network dependency. Trade-off: host loss = rebuild VM from image/backup; durable data lives on Ceph. VLAN 200 still reserved for rebuild/migration traffic.

Data path B — External Rook/Ceph​

Scope​

Bulk PVCs, shared FS, object storage for the application cluster.

Placement rules​

6 OSD VMs, 2 per zone, each alone on its host with all its NVMe (OSD count = disks, not VMs) · masters on separate hosts · CRUSH failure domain = zone — replica 3 puts one copy per rack · Ceph never stores on Longhorn.