airflow-turbine

Apache Airflow operator for Turbine. Drop TurbineCheckOperator into a DAG, point it at a Turbine server with an Airflow Connection, and your data-quality checks run as native Airflow tasks. The operator is deferrable — it releases the worker slot while Turbine evaluates, then resumes to log every Check Outcome and fail the task on any breach.

Install

uv add airflow-turbine

Python 3.12 or newer. Pulls only apache-airflow and turbine-client — the Turbine engine never enters your Airflow environment.

Minimal example

from airflow_turbine import TurbineCheckOperator

quality = TurbineCheckOperator(
    task_id="quality",
    turbine_conn_id="turbine",
)

The Connection's host / port / schema build the base URL; password (when set) becomes a static bearer token. With no other configuration the operator runs every contract registered with the server.

Contract Selection

TurbineCheckOperator(
    task_id="quality",
    turbine_conn_id="turbine",
    contracts=["orders", "customers"],   # list, single id, or None to fan out across every contract
)

contracts=None (the default) fans out across every registered contract; contracts=[] is a no-op that succeeds without running anything (useful when the list is templated from an upstream task).

Check Window from the DAG's data interval

Opt in to forward data_interval_start / data_interval_end as the Check Window:

TurbineCheckOperator(
    task_id="quality",
    turbine_conn_id="turbine",
    use_data_interval=True,
)

Off by default — every scheduled DAG has a data interval, and forwarding it unconditionally would spray warnings on contracts whose SLAs do not declare a latency element. Setting use_data_interval=True together with incremental=True is rejected at task construction.

Run options

TurbineCheckOperator(
    task_id="quality",
    turbine_conn_id="turbine",
    incremental=True,    # check rows newer than the last watermark; mutually exclusive with use_data_interval
    flag_rows=True,      # persist failing-row PKs to the flag matrix
)

Auth

Three modes, picked from the Connection:

Override the Connection-derived auth explicitly:

from turbine_client import BearerAuth

TurbineCheckOperator(
    task_id="quality",
    turbine_conn_id="turbine",
    auth=BearerAuth(token="explicit"),
)

Any explicit auth=... requires turbine_conn_id to be set — the trigger re-reads the connection on the triggerer process to rebuild the auth without carrying credentials across the deferral boundary.

Behaviour on failure

Each Check Outcome is logged on resume. The task raises RuntimeError when:

• The Check Run finishes with RunStatus.FAILED.
• Any Check Outcome is FAILED or ERRORED.

WARNED outcomes log at WARNING level but do not fail the task — advisory checks (severity: warn in the Quality Spec) surface as a soft signal.

Compatibility

airflow-turbine >= 0.5.12 requires turbine-data >= 0.5.12 on the server.