arkitekt-datalayer

The datalayer Django app shared by arkitekt services: S3-backed store rows (media, big file, zarr, parquet, sparse, fabriks, konnektion, kulisse), scoped upload and read grants minted as STS sessions, per-organization/user/upload quotas, and an orphan purge.

Install it as arkitekt_datalayer; its app label is datalayer, so tables are datalayer_* and models are referenced as "datalayer.MediaStore".

INSTALLED_APPS = [..., "polymorphic", "authentikate", "arkitekt_datalayer"]
DATALAYER = conf.datalayer.model_dump(exclude_none=True) if conf.datalayer else {}

Embed arkitekt_datalayer.configuration.DatalayerSettings in the service's config model (narrow the buckets it needs in a subclass). Without a DATALAYER setting the app still installs; grants and writes then raise, naming the missing bucket.

What a service wires itself: the GraphQL mutations it exposes (arkitekt_datalayer.mutations, gated by load_config().upload_roles), scalars.SCALAR_MAP in its schema config, and the store types it declares in its hub contract. Deleting data rows should straddle the delete with storage.stores_orphaned_by / storage.flag_orphaned; manage.py purge_orphaned_stores collects them after store_grace_days.

A KulisseStore holds a device or scene model, one self-contained GLB in the wire format the kulisse package specifies; arkitekt_datalayer.kulisse is the server's own reader of it (range reads, never the whole file, except the streamed SHA-256 that is the model's reference).

Tests run against a real postgres and RustFS via dokker: uv run pytest.