swiss-ai-hub-backup

The centralized backup, restore, and PostgreSQL-maintenance service for Swiss AI Hub — a self-contained Dagster instance that snapshots every stateful service to S3.

---

What is Swiss AI Hub?

Swiss AI Hub is an open-source, self-hosted AI platform for enterprises. One docker compose up starts ~30 integrated containers across several stateful stores — PostgreSQL, FerretDB, Milvus, Neo4j, ClickHouse, Valkey, and NATS JetStream. This package keeps that data safe.

What is this package?

swiss-ai-hub-backup is the platform's backup/restore and database-maintenance plane. It runs as its own independent Dagster instance (separate from the data pipelines, with its own SQLite storage) and:

• Backs up PostgreSQL (×2), Milvus, Neo4j, ClickHouse, Valkey, and NATS JetStream to S3 (SeaweedFS) on a schedule — gracefully stopping and restarting the managed containers around each run for consistent snapshots.
• Restores any service from a chosen backup timestamp.
• Maintains the platform PostgreSQL online: prunes verbose Dagster event_logs, tunes autovacuum, and runs pg_repack — so deployments stay bounded over time without downtime.

Each stateful service has a BackupHandler (postgres, milvus, neo4j, clickhouse, valkey, nats); the whole thing is wired into a Dagster asset graph by backup_definitions(). Because it operates on the storage layer and needs to stop containers, it requires read access to the Docker socket (/var/run/docker.sock), which it uses to discover platform containers via their com.docker.compose.project label.

Unlike the other Swiss AI Hub packages, this is an operational service, not a library you build agents/APIs on. It is licensed AGPL-3.0-or-later (the rest of the SDK is Apache-2.0).

Should you use this package?

Most operators don't install it directly — it ships with the platform as the backup-* containers (a gRPC code server, a daemon, and a webserver UI on :3004). You'd reach for this PyPI package to run the backup plane standalone, embed its logic, or extend it — for example, adding a BackupHandler for a stateful service of your own.

What it does

Job	Schedule	Stops containers?
Full backup (all services → S3)	daily	Yes (consistent snapshots)
Restore (service ← chosen timestamp)	on demand	Yes
event_logs cleanup + autovacuum tuning	weekly	No (online-safe)
pg_repack (reclaim disk)	monthly	No (online-safe)

Installation

pip install swiss-ai-hub-backup
# or
uv add swiss-ai-hub-backup

Requires Python 3.13.

---

Quick start

The backup plane is a Dagster code location built by backup_definitions():

# my_backup/__init__.py
from swiss_ai_hub.backup.dagster.definitions import backup_definitions

defs = backup_definitions()   # 26 assets, 4 jobs: backup, restore, cleanup, repack

Inspect and run it with the Dagster UI (it keeps its own state in DAGSTER_HOME):

export DAGSTER_HOME=/tmp/backup-dagster && mkdir -p "$DAGSTER_HOME"
set -a && source .env && set +a          # S3 + DB credentials, BACKUP_* settings
dagster dev -m my_backup                 # http://localhost:3000

From the UI you can materialize the online-safe maintenance jobs (cleanup, pg_repack) against a running stack without disruption. The full backup/restore jobs stop and restart containers, so run those deliberately — and note they need access to the Docker socket and to all the stateful services. dagster definitions validate -m my_backup loads the whole code location without running anything (a fast CI/sanity check).

Settings are not auto-loaded from the environment. Connection and BACKUP_* settings are read only when constructed, so export them in the process that runs Dagster (set -a && source .env && set +a).

---

How it's deployed

In production the backup plane runs as three containers from one image, forming a self-contained Dagster instance:

Container	Role	Notes
backup-code	Dagster gRPC code server (dagster api grpc … :4266)	mounts /var/run/docker.sock:ro to stop/start containers; on data + storage
backup-daemon	Dagster daemon	runs the schedules and sensors; on data
backup-webserver	Dagster UI (:3004)	inspect runs, trigger restores; on proxy + data

Because it needs the Docker socket and the platform's stateful services, the canonical deployment is the platform's own backup compose. See infra/deployment and the documentation for the full container setup, retention config, and the BACKUP_* environment variables. If you run your own variant, mirror that three-container shape and grant the code server read access to the Docker socket.

Extending — add a service to back up

Implement the BackupHandler ABC for your service in services/, then register it in HANDLER_FACTORIES — the Dagster asset wiring picks it up automatically (handlers are synchronous by design):

from swiss_ai_hub.backup.services.base import BackupHandler

class MyServiceHandler(BackupHandler):
    def backup(self, context) -> ...:
        ...   # dump your service's state to S3
    def restore(self, context) -> ...:
        ...   # restore it from a backup

If the handler needs Docker access, type-hint a DockerManager parameter in __init__ and the factory injects it. The maintenance subsystem follows the same pattern (MaintenanceHandler + CLEANUP_HANDLER_NAMES). See the documentation for the full handler contract.

---

Links

• Source & issues: https://github.com/bbvch-ai/aihub-core
• Documentation: https://bbvch-ai.github.io/aihub-core/

License

AGPL-3.0-or-later — see packages/backup/LICENSE. Note this differs from the Apache-2.0 SDK packages; for the full per-package license matrix, see LICENSES.md.

---

Part of Swiss AI Hub. Built in Switzerland by bbv Software Services.