Backup and Restore

Backups should cover structured data, object storage, and configuration. Take them together: documents, knowledge, and the entity filesystem span PostgreSQL and MinIO and Redis-held JuiceFS metadata, and a restore is only consistent if the pieces are from the same point in time.

What to Back Up

What	Where it lives	Why
PostgreSQL	pg_data volume	All relational data: users, workspaces, tasks, goals, runs, messages.
MinIO	minio_data volume	Uploaded files, generated artifacts, and JuiceFS data blocks.
Redis DB /1	redis_data volume	JuiceFS metadata. Without it the JuiceFS data blocks in MinIO are unreadable.
.env	host	Secrets, provider credentials, and deployment configuration.
OAuth app registrations	external providers	Client ids/secrets and callback URLs needed to reconnect integrations.

Built-in Export API

The API ships a tenant export surface for logical backups:

GET  /api/v1/backup/summary          # what an export would contain
GET  /api/v1/backup/export           # export the tenant's data
POST /api/v1/backup/export/download  # download an export archive

This is a data-level export (suitable for migrations and offsite copies), not a substitute for volume-level backups of a whole deployment.

PostgreSQL

Dump:

docker compose exec postgres pg_dump -U manor manor > manor.sql

Restore into a compatible PostgreSQL version:

cat manor.sql | docker compose exec -T postgres psql -U manor manor

For large databases prefer the custom format (pg_dump -Fc) with pg_restore, and run dumps on a schedule (cron on the host is fine for a single-server deployment).

MinIO and JuiceFS

Use MinIO client tooling (mc mirror) or volume snapshots of minio_data. Two rules:

• Keep object backups coordinated with database backups — a document row whose file blob is missing (or vice versa) is a broken restore.
• JuiceFS data is only usable together with its metadata. Snapshot redis_data (or at minimum Redis database /1) at the same time as minio_data. Restoring MinIO without the matching Redis metadata leaves the entity filesystem empty even though the blocks exist.

Stop or quiesce the workers before snapshotting if you need a strictly consistent filesystem image; the FUSE mount flushes its cache on clean stop.

Restore Practice

Do not wait for an incident to test restores. Periodically restore into a separate environment and confirm:

1. Login works and users/workspaces are present.
2. Documents open and knowledge search returns results (checks PostgreSQL, MinIO, and JuiceFS metadata alignment).
3. An agent run completes.
4. Integrations reconnect (or are cleanly re-authorized).

Keep at least one backup from immediately before every upgrade — see Upgrades and Releases.