How to nest an Airflow DAG dynamically?

I have a simple DAG of three operators. The first one is PythonOperator with our own functionality, the other two are standard operators from airflow.contrib (FileToGoogleCloudStorageOperator and GoogleCloudStorageToBigQueryOperator to be precise). They work in sequence. Our custom task produces a number of files, typically between 2 and 5, depending on the parameters. All of these files have to be processed by subsequent tasks separately. That means I want several downstream branches, but it's unknowable how many exactly before the DAG is run.

How would you approach this problem?

UPDATE:

Using BranchPythonOperator that jhnclvr mentioned in his another reply as a point of departure, I created an operator that would skip or continue executing a branch, depending on condition. This approach is feasible only because highest possible number of branches is known and sufficiently small.

The operator:

class SkipOperator(PythonOperator):
    def execute(self, context):
        boolean = super(SkipOperator, self).execute(context)
        session = settings.Session()
        for task in context['task'].downstream_list:
            if boolean is False:
                ti = TaskInstance(
                    task, execution_date=context['ti'].execution_date)
                ti.state = State.SKIPPED
                ti.start_date = datetime.now()
                ti.end_date = datetime.now()
                session.merge(ti)
        session.commit()
        session.close()

Usage:

def check(i, templates_dict=None, **kwargs):
    return len(templates_dict["data_list"].split(",")) > i

dag = DAG(
    dag_name,
    default_args=default_args,
    schedule_interval=None
)

load = CustomOperator(
    task_id="load_op",
    bash_command=' '.join([
        './command.sh'
        '--data-list {{ dag_run.conf["data_list"]|join(",") }}'
    ]),
    dag=dag
)

for i in range(0, 5):
    condition = SkipOperator(
        task_id=f"{dag_name}_condition_{i}",
        python_callable=partial(check, i),
        provide_context=True,
        templates_dict={"data_list": '{{ dag_run.conf["data_list"]|join(",") }}'},
        dag=dag
    )
    gs_filename = 'prefix_{{ dag_run.conf["data_list"][%d] }}.json' % i

    load_to_gcs = CustomFileToGoogleCloudStorageOperator(
        task_id=f"{dag_name}_to_gs_{i}",
        src='/tmp/{{ run_id }}_%d.{{ dag_run.conf["file_extension"] }}' % i,
        bucket=gs_bucket,
        dst=gs_filename,
        mime_type='application/json',
        google_cloud_storage_conn_id=connection_id,
        dag=dag
    )
    load_to_bq = GoogleCloudStorageToBigQueryOperator(
        task_id=f"{dag_name}_to_bq_{i}",
        bucket=gs_bucket,
        source_objects=[gs_filename, ],
        source_format='NEWLINE_DELIMITED_JSON',
        destination_project_dataset_table='myproject.temp_{{ dag_run.conf["data_list"][%d] }}' % i,
        bigquery_conn_id=connection_id,
        schema_fields={},
        google_cloud_storage_conn_id=connection_id,
        write_disposition='WRITE_TRUNCATE',
        dag=dag
    )

    condition.set_upstream(load)
    load_to_gcs.set_upstream(condition)
    load_to_bq.set_upstream(load_to_gcs)

Basically, you can't add tasks to a DAG when it's running. You would need to know ahead of time how many tasks you wanted to add.

You could process N files using a single operator.

Or, if you have another separate dag that processes a file you could trigger that DAG N times, passing the name of the file in the conf.

See here for an example of the TriggerDagRunOperator.

See here for the DAG that would be triggered.

And lastly see this post from which the above examples are from.

来源：https://stackoverflow.com/questions/44636774/how-to-nest-an-airflow-dag-dynamically

标签

airflow