> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.fabricgateway.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.fabricgateway.ai/_mcp/server.

# Maintain the data plane and its data

> Learn what state the data plane keeps and where, what you need to back up, how long stored data lasts, and the routine tasks that keep a data plane healthy.

This page applies to the Enterprise plan, where you run the data plane in your own VPC. The gateway and embedder pods keep no durable state of their own. Everything that has to survive a restart, an upgrade, or a reinstall lives outside the Helm release, in systems you run. This page lists that state, what to back up, how long stored data lasts, and the routine tasks to put on your calendar.

## What state the data plane keeps

| State                                                                            | Where it lives                                                                                      | Survives pod replacement        | Who backs it up                           |
| -------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | ------------------------------- | ----------------------------------------- |
| The credentials users connect, client registrations, and short-lived login state | Your MySQL database, or the distributed store's per-pod snapshots                                   | Yes                             | You                                       |
| The encryption key, `FABRIC_GW_SECRET_KEY`, that protects stored credentials     | The app Secret                                                                                      | Yes                             | You                                       |
| Traces, metrics, and logs                                                        | Your telemetry backend, such as ClickHouse                                                          | Yes                             | You                                       |
| Each pod's certificate, private key, and configuration cache                     | `/data/cache` in the pod. An `emptyDir` by default, or a per-pod volume with the distributed store. | Only with the distributed store | Nobody. A pod without them enrolls again. |
| The embedder's model                                                             | An `emptyDir`, or a PVC with `embedder.persistence.enabled`                                         | Only with a PVC                 | Nobody. The model downloads again.        |
| Routes, services, plugins, and the rest of your configuration                    | The control plane, run by Postman                                                                   | Yes                             | Postman                                   |

Your configuration isn't stored in the data plane. Each pod receives it from the control plane and caches it locally, so a data plane you rebuild from scratch gets the same configuration back when it enrolls.

## Back up the secret store

Backups of the secret store are your responsibility. The chart doesn't back anything up, and doesn't rely on Helm's release history to restore anything.

* **With MySQL,** back up the data plane's database on your usual schedule.
* **With the distributed store,** each pod's snapshot volume holds an encrypted copy of the store. Back up those volumes with your volume snapshot tooling.

Back up `FABRIC_GW_SECRET_KEY` with the same care, and keep it with your database backups. Stored credentials are encrypted with this key, so a database restored without its key is unreadable. Every replica must use the same key. To change it safely, see [Manage data plane secrets and keys](/run/secrets-and-keys).

If you lose the secret store without a backup, users have to connect their accounts again, and MCP clients have to register again.

## How long stored data lasts

With the MySQL store, each gateway pod removes expired rows every 60 seconds:

| Data                                            | Removed after                                                                           |
| ----------------------------------------------- | --------------------------------------------------------------------------------------- |
| Login state and one-time codes from OAuth flows | 10 minutes                                                                              |
| MCP client registrations                        | 90 days without use. Each use resets the clock, so registrations in use aren't removed. |
| Credentials users connect                       | Never. They stay until the user or an admin removes them.                               |

Telemetry retention is set by your backend. The chart's collector writes to ClickHouse but doesn't create its tables, so retention follows the tables you create.

## Schema migrations

With MySQL, every install and upgrade runs a schema migration Job before the gateway pods change, and the Job applies only migrations the database hasn't recorded yet. Day-to-day operation needs nothing else. Two rules keep it working:

* Never edit a migration that has already been applied. The next upgrade stops with a checksum mismatch.
* Keep `migrations.baselineVersion` at `"0"` if you adopt a database that already has the schema.

To run migrations from your own pipeline instead, set `migrations.enabled: false`. For least privilege, give the Job its own DDL-capable user with `migrations.dbUser` and `migrations.existingSecret`.

Because the release holds no state, you can uninstall and reinstall the chart against the same MySQL database. The migration Job adopts the existing schema and applies only what's missing. For the full model, see [Database migrations](/run/database-migrations).

## Routine maintenance

* **Keep a way to enroll new pods,** if you enroll with a bootstrap token. You mint the token in the console, and it lasts 10 minutes by default and 30 minutes at most, so pods created after it expires can't enroll. Before you add or replace pods, mint a fresh token and put it in the app Secret. Workload attestation needs no token. See [Connect the data plane to the control plane](/run/connect-a-data-plane).

* **Rotate other secrets** in the app Secret, or in the secret manager that syncs it. The gateway reads new values when its pods restart:

  ```bash
  kubectl -n pm-fabric-dp rollout restart deploy/fabric-gateway-chart
  ```

  With the distributed store, the gateway runs as a StatefulSet, so use `statefulset/fabric-gateway-chart` instead.

* **Update the image** on every planned upgrade, by setting `image.tag` or pinning `image.digest`. See [Upgrade the data plane](/run/upgrade).

* **Watch telemetry export.** An increase in `otelcol_exporter_send_failed_*` means telemetry isn't reaching your backend, before it turns into a gap in your data. See [Monitor the data plane](/run/monitor#metrics).

* **Back up the secret store and its key,** and test a restore.

Certificate renewal needs no maintenance. Each pod renews its own certificate before it expires.

## Next steps

* To keep the data plane current, see [Upgrade the data plane](/run/upgrade).
* To remove a data plane, see [Uninstall the data plane](/run/uninstall).
* To fix a problem, see [Troubleshoot the data plane](/run/troubleshooting).