Storage

Solo and Lab (node agents)

Data	Where	Notes
Control plane store	SQLite file in the data directory, or Postgres for larger Labs	einsora backup (planned) takes a consistent copy
Data directory	macOS: ~/Library/Application Support/Einsora. Linux: /var/lib/einsora (service) or ~/.local/share/einsora (user).	Planned
Workload disks (VM images, notebook workspaces)	Local disk on the machine that runs the workload	No replication. A workload is tied to its machine’s disk.
Model cache	One per machine, deduplicated by sha256 and shared between workloads	Validated before first use
Disk caps	Per workload, and a total for guests on shared machines (ADR-0014)	Planned

Durability: a Solo or Lab machine’s disk is as safe as the machine. Back up the store, and keep anything precious off community hosts.

Datacenter (Kubernetes)

Need	Phase 3 start	Later	Referenced as
Block (VM disks, notebook dirs, Postgres, Prometheus)	Longhorn	Ceph RBD	StorageClass einsora-block (cluster-config)
Object (models, datasets, backups)	SeaweedFS or Garage (object-store)	Ceph RGW	S3 endpoint in einsora-models
Shared filesystem	none	CephFS	future einsora-shared
Local scratch	emptyDir on NVMe	TopoLVM or local-path	future

The control plane only knows einsora-block (EINSORA_STORAGE_CLASS), so switching backends changes deploy/, not code.

VM images are CDI DataSources in einsora-images, imported per region. Each VM’s root disk is a clone of one.

Durability and backups

Data	Protection
Postgres (tenants, workloads, events, usage)	WAL archiving plus daily base backups, with monthly restore tests
etcd	k3s etcd-snapshot to object storage
User volumes	Not backed up by default. Snapshots are a later feature.

Regions: every region keeps its own block and object storage, and workload data stays in its region (ADR-0009). Only the control plane’s store is shared across regions.

Encryption: volumes aren’t encrypted at rest yet, and freed blocks can hold earlier users’ data. The plan is encrypted StorageClasses with per-volume keys.