SnowPro Advanced: MLOps Engineer free question 7 / 10 · Domain 4: Pipeline Orchestration and Automation (CI/CD)
A company runs about 200 Airflow DAGs for its data platform. It is adding a training pipeline that uses Snowflake ML, and training reads large tables. Which approach does the Snowflake ML guide recommend?
- Keep Airflow and send the heavy training steps to ML Jobs.
- Replace Airflow with Dagster, which Snowflake ML favors.
- Pull the tables onto Airflow workers and train them there.
- Move all the DAGs to Task Graphs before training starts.
Show the answer and explanation
Answer: A
Why this is correct
The section "Pipeline orchestration" says Snowflake ML is compatible with most orchestration frameworks, including Airflow, Dagster and Prefect. If you already have an existing workflow or DAG setup, it recommends simply integrating your existing workflows with Snowflake ML features and offloading compute- or data-intensive steps to ML Jobs or UDFs. If you do not have an existing DAG setup, you can use Snowflake Task Graphs as a Snowflake-native solution. The company keeps its Airflow DAGs and sends the training step, which reads large tables, to ML Jobs.
Why the other options are wrong
"Move all the DAGs to Task Graphs before training starts." … The guide offers Snowflake Task Graphs for teams that do not have an existing DAG setup. This company has one, and the guide's advice for that case is to integrate with it rather than rebuild it.
"Pull the tables onto Airflow workers and train them there." … Moving large tables out to the workers is the opposite of what the guide recommends. It says to offload compute- or data-intensive steps to ML Jobs or UDFs, so the work runs where the data is.
"Replace Airflow with Dagster, which Snowflake ML favors." … The guide says Snowflake ML is compatible with most orchestration frameworks, including Airflow, Dagster and Prefect. It does not rank one above the others, so a switch gains nothing here.
Orchestrator stays, heavy work moves
The orchestrator decides when and in what order steps run; it does not have to be where the compute happens. Keeping the scheduler you already operate and pushing the data-heavy steps into Snowflake gives you retries, history and alerts from the tool you know, and the training runs next to the data.