Railway

Railway runs each compose service as a service of its own, so there is no server to look after. In return, less of the security model holds, and the warehouse is Postgres.

Caution

Planned, and not built. The guide will be written and tested by deploying from it once before a release.

What changes on Railway

Railway attaches a volume to one service only, puts every service of a project on one private network, and has no depends_on (Railway’s compose guide). So:

• The warehouse is postgres. DuckDB’s hand-off needs a volume shared by three containers, and Railway cannot share one. Switching is one setting.
• anfra’s data source and Cerbos’s policies go into their images at mb build, instead of onto a shared volume.
• Secrets come through Railway’s variables, with the env provider.
• Network separation is gone: every service can reach every other. The database roles, the per-container secrets, Cerbos, the api’s sign-in and the guarded fetcher still hold. The network rules, read-only root filesystems and dropped capabilities do not.

Choose Railway when you want no server more than you want the full security model.