dbt-qod

A dbt adapter for the Quack-on-Demand FlightSQL edge.

qod is a multi-tenant Arrow Flight SQL gateway in front of DuckDB Quack + DuckLake. dbt-qod lets dbt build models, seeds, snapshots, and tests against it over the Arrow ADBC FlightSQL driver, with the qod tenant/pool routing and RBAC applied.

Verified end-to-end (views, table, incremental, seed, snapshot, tests) against a live edge, and the full dbt-tests-adapter conformance suite passes (12/12) - see examples/jaffle_tpch/RUN_REPORT.md.

Install

pip install dbt-core dbt-qod      # dbt-qod from this checkout: pip install -e .

Requires Python 3.10-3.13 (dbt-core does not yet support 3.14).

Profile

# ~/.dbt/profiles.yml
my_project:
  target: dev
  outputs:
    dev:
      type: qod
      host: localhost
      port: 31338
      database: acme_tpch     # qod tenant-db / DuckLake catalog
      schema: dbt_jaffle      # where dbt writes models
      tenant: acme            # gRPC routing header
      pool: bi                # gRPC routing header
      username: admin
      password: admin
      superuser: true         # optional; system-realm login
      use_tls: true           # grpc+tls (default)
      tls_skip_verify: true   # self-signed dev cert
      threads: 4

dbt init is supported too (a profile_template.yml ships with the adapter).

Supported

dbt feature	status
view / table materializations	yes
incremental (append, delete+insert)	yes
seed	yes
snapshot (timestamp / check)	yes
generic tests, sources, refs	yes
dbt docs generate (catalog)	yes

How it works

The backend SQL dialect is DuckDB, so the adapter reuses dbt's default macros and overrides only the DuckDB/Flight specifics. Three design decisions matter:

1. database = tenant-db, schema = DuckLake schema. dbt's three-part relations render as <database>.<schema>.<identifier> (e.g. acme_tpch.dbt_jaffle.stg_customer). That both satisfies the edge's ACL qualification and targets the shared DuckLake catalog, so objects are visible across every node of a pool.
2. No TEMP intermediates. dbt's table/incremental/snapshot flows build intermediate relations. DuckDB TEMPORARY tables are node-local and would break across a multi-node pool, so the adapter materializes intermediates as real, uniquely-suffixed DuckLake tables (visible everywhere) and drops them after.
3. Eager execution in the connection manager. The qod edge executes a statement on do_get (fetch), not at prepare, so the ADBC cursor is wrapped to pull the result and force execution of DDL/DML. Update-count statements (DROP/INSERT/...) return an empty stream against a Count schema, which ADBC flags as inconsistent schema; that one case is treated as benign. ACL denials surface earlier (at prepare) and propagate as dbt errors.

Other adapter notes:

• The edge does not implement parameter-bound statements, so dbt seed inlines literal values instead of using %s bindings.
• DuckDB exposes a single unqualified information_schema spanning all catalogs; metadata macros filter by catalog_name / table_catalog rather than qualifying the path.

Example project

examples/jaffle_tpch is a runnable project (sources over tpch1, staging views, a mart table, an incremental model, a seed, a snapshot, and tests). With a qod stack up and TPC-H loaded:

cd examples/jaffle_tpch
dbt build --full-refresh --profiles-dir .

Layout

dbt/adapters/qod/    credentials, connections (ADBC), relation, column, impl, Plugin
dbt/include/qod/     dbt_project.yml, profile_template.yml, macros/ (DuckDB dialect)
examples/jaffle_tpch/  runnable demo project + RUN_REPORT.md

Development & publishing

python -m venv .venv && source .venv/bin/activate
pip install -e '.[test]'          # adapter + dbt-tests-adapter conformance suite

# conformance suite (needs a live qod edge + TPC-H demo; override via QOD_* env)
pytest tests/functional

# build + validate distributions
python -m build && twine check dist/*

Publishing is automated: pushing a vX.Y.Z tag triggers .github/workflows/publish.yml, which builds and uploads to PyPI via Trusted Publishing (OIDC, no stored token). One-time setup: on PyPI, add a GitHub trusted publisher for this repo pointing at publish.yml and the pypi environment. The version is single-sourced from dbt/adapters/qod/__version__.py.

To get listed in dbt's adapter docs, run the full dbt-tests-adapter suite green and open a PR to dbt-labs/docs.getdbt.com.

License

Apache-2.0