In T396495#10941826, we ran a PoC ML training Airflow pipeline end-to-end using dummy data.
As a next step, I have been working on adapting this pattern to run an ML training Airflow pipeline with real-world data. Specifically, I worked on a simple add_a_link Airflow pipeline with the following components:
- Job logic: ml-pipelines/add_a_link
- Scheduling logic: airflow-dags/ml/dags/add_a_link_dag.py
However, when running this pipeline in the airflow-devenv, the DAG fails due to a Kerberos authentication error when attempting to connect to the Hive metastore. This prevents the Spark job from accessing the required Hive tables.
For reference, here are the detailed YARN logs from the failed runs: https://phabricator.wikimedia.org/P78736