# Self-host storage

There is no object store and no managed filesystem here, and neither is
coming. An S3-compatible endpoint or a shared filesystem is something you run
on a VM, on volumes.

## The shape

One VM, its root disk for the OS and the service, one or more volumes for the
data ([which disk holds what](../vm.md#two-kinds-of-disk)).

```
POST /v1/vms       {"name":"store1","ram_mb":4096,"disk_gb":40, ...}
POST /v1/volumes   {"name":"bulk1","size_gb":2000}
POST /v1/volumes/{id}/actions {"type":"attach","vm_id":"vm_..."}
POST /v1/vms/{id}/actions     {"type":"start"}
```

Then mount the volume by UUID and point the service at it:
[Volumes](../vm.md#volumes).

## S3-compatible: MinIO

Serves the S3 API over one directory, which is your mounted volume. Any SDK
that talks to S3 talks to it once you set the endpoint.

- Give it its own data directory on the volume, not the mount point itself.
- Put a DNS name in front of it from the start, so the machine underneath can
  be recreated without touching any client.
- Terminate TLS on the machine (a certificate for that DNS name), because the
  traffic crosses the public Internet even when both ends are your VMs.
- Open only its port inbound: [Harden a VM](harden-your-vm.md).

## Shared filesystem: NFS

A volume attaches to **one VM at a time** — these are block devices, not a
shared filesystem. So "several machines, same files" means one machine
exporting and the others mounting, with the export scoped to your own
addresses in the server's firewall.

Every VM has a public address and there is no private network between them, so
either scope the export tightly or put the mesh in first:
[Build a private mesh](wireguard-mesh.md).

## What it costs

**Egress is metered at your VM's own interface**, so serving another of your
VMs costs the same as serving the Internet. Inbound is free. Reading a dataset
repeatedly across the network costs more than putting it on the machine that
reads it.

Volumes bill from creation to deletion, attached or not, and growing is
one-way.

## Back it up somewhere that is not here

A snapshot is stored alongside the disk it covers, so it covers your own
deletes and nothing else. For anything you cannot lose, push copies off the
platform — `restic` or `borg` on a schedule to a target you hold elsewhere —
and test a restore.