# Self-hosting (Enterprise plan)

> Run DataxLand in your own Kubernetes infrastructure with a Helm chart, an offline license and an air-gapped mode; when to choose it, what it includes, how it differs from the cloud and its known limits.

On the Enterprise plan you can run DataxLand **in your own infrastructure**: data, warehouse and backups live in your cloud
account or data centre, the AI model can be one you host yourself, and the installation can run without any outbound Internet
access.

## When to self-host

- Internal policy or regulation requires that data never leaves your company network.
- You need an in-house AI model instead of an external model provider.
- Your source databases are only reachable from the internal network and are never exposed outside it.
- You have an infrastructure team that runs Kubernetes, Postgres and object storage.

Without those constraints, [DataxLand Cloud](/en/pricing) means far less to operate: backups, upgrades and monitoring are
handled by DataxLand.

## The model: one installation per customer

- One installation serves **one organisation**. It can hold many workspaces and many users, within the limits of your license.
- Self-hosted DataxLand runs **the same code and the same images** as DataxLand Cloud. Only the configuration and the license
  file differ, so features and fixes arrive at the same time.
- It installs with a **Helm chart** on any Kubernetes (1.27 or later). **Reference Terraform modules** for AWS, GCP and Azure
  are included as well.

## Components

| Component | Role | Default replicas |
| --- | --- | --- |
| `web` | User interface | 2 |
| `api` | API, dashboards, asking Si | 2 |
| `mcp` | MCP server for external agents (can be disabled) | 1 |
| `worker` | Ingest and builds, refresh schedules, backups, warehouse maintenance | always 1 |
| `migrate` Job | Upgrades the Postgres schema before every upgrade | runs once |
| Backup CronJob | Optional, when you want Kubernetes to own the backup schedule | off |

Outside the cluster, you provide:

- **Managed Postgres 16**: application metadata and the data warehouse catalog.
- **S3-compatible object storage**: the data warehouse (Parquet) and backups.
- **A shared volume** (PersistentVolumeClaim): uploaded files, workspace configuration, encrypted secrets and the build working
  directory.

Details in [Requirements and sizing](/en/docs/self-host/requirements).

## How it differs from the cloud

| | DataxLand Cloud | Self-hosted |
| --- | --- | --- |
| Infrastructure, backups, upgrades | Run by DataxLand | Your infrastructure team |
| Sign-up | Through the sign-up page | Invitation only by default; installation-wide admins are declared by email |
| Plan | As subscribed | Enterprise plan, every feature |
| Licensing | Billed per plan | [Offline license file](/en/docs/self-host/license), never calls home to DataxLand |
| AI model | Models chosen by DataxLand | The default catalog (calls the model provider's API), or your own in-house model |
| Outbound Internet | Yes | Your choice; an [air-gapped mode](/en/docs/self-host/airgapped) is available |
| Data sent to DataxLand | Not applicable | Nothing is sent to DataxLand |

## Known limits

Read these carefully before you plan:

- **The shared volume must be ReadWriteMany when you run several nodes.** `api`, `mcp` and `worker` read and write the same
  directory. In the cloud that means EFS (AWS), Filestore (GCP) or Azure Files (Azure); on-prem, NFS or CephFS. ReadWriteOnce
  is only enough when the whole installation runs on a single node. NFS performance on the build directory is the biggest
  operational dependency.
- **`worker` runs as a single replica.** The chart pins one replica and updates it with the Recreate strategy: during an
  upgrade, background work (builds, refresh schedules, backups) pauses for a few minutes. Dashboards and asking Si keep working.
- **One object storage key for the whole bucket in the cloud.** On S3, GCS or s3proxy every workspace shares one key; data is
  separated by path prefix. Per-workspace keys are only available with MinIO.
- **Azure Blob goes through s3proxy and has not been tested on a real account.** If you need certainty, run your own MinIO on
  Azure Disk.
- **Terraform is a reference template.** It passes `terraform fmt` and `terraform validate` but has not been applied to a real
  cloud account. See [Reference Terraform](/en/docs/self-host/terraform).
- **Only one installation per namespace**, because the `api` Service has the fixed name `api`.

## Steps

1. Prepare the infrastructure following [Requirements and sizing](/en/docs/self-host/requirements), or build it with the
   [reference Terraform](/en/docs/self-host/terraform).
2. Contact DataxLand to get your license file and image pull access.
3. [Install with Helm](/en/docs/self-host/helm).
4. If the network has no Internet access: [Air-gapped mode](/en/docs/self-host/airgapped).
5. Schedule backups and rehearse a restore: [Upgrades and backups](/en/docs/self-host/upgrade-backup).

## Related

- [Architecture](/en/docs/architecture)
- [Security and privacy](/en/docs/security-privacy)
- [Enterprise plan](/en/enterprise)
