CODEWITHSUNDEEP
pythonpython-3.xairflow
aaron
aaron

Reputation: 6489

How to dynamically iterate over the output of an upstream task to create parallel tasks in airflow?

Consider the following example of a DAG where the first task, get_id_creds, extracts a list of credentials from a database. This operation tells me what users in my database I am able to run further data preprocessing on and it writes those ids to the file /tmp/ids.txt. I then scan those ids into my DAG and use them to generate a list of upload_transaction tasks that can be run in parallel.

My question is: Is there a more idiomatically correct, dynamic way to do this using airflow? What I have here feels clumsy and brittle. How can I directly pass a list of valid IDs from one process to that defines the subsequent downstream processes? 

from datetime import datetime, timedelta
import os
import sys

from airflow.models import DAG
from airflow.operators.python_operator import PythonOperator

import ds_dependencies

SCRIPT_PATH = os.getenv('DASH_PREPROC_PATH')
if SCRIPT_PATH:
    sys.path.insert(0, SCRIPT_PATH)
    import dash_workers
else:
    print('Define DASH_PREPROC_PATH value in environmental variables')
    sys.exit(1)

default_args = {
  'start_date': datetime.now(),
  'schedule_interval': None
}

DAG = DAG(
  dag_id='dash_preproc',
  default_args=default_args
)

get_id_creds = PythonOperator(
    task_id='get_id_creds',
    python_callable=dash_workers.get_id_creds, 
    provide_context=True,
    dag=DAG)

with open('/tmp/ids.txt', 'r') as infile:
    ids = infile.read().splitlines()

for uid in uids:
    upload_transactions = PythonOperator(
        task_id=uid,
        python_callable=dash_workers.upload_transactions,
        op_args=[uid],
        dag=DAG)
    upload_transactions.set_downstream(get_id_creds)

Upvotes: 7

Views: 7891

Answers (2)

aaron
aaron

Reputation: 6489

Per @Juan Riza's suggestion I checked out this link: Proper way to create dynamic workflows in Airflow. This was pretty much the answer, although I was able to simplify the solution enough that I thought I would offer my own modified version of the implementation here: 

from datetime import datetime
import os
import sys

from airflow.models import DAG
from airflow.operators.python_operator import PythonOperator

import ds_dependencies

SCRIPT_PATH = os.getenv('DASH_PREPROC_PATH')
if SCRIPT_PATH:
    sys.path.insert(0, SCRIPT_PATH)
    import dash_workers
else:
    print('Define DASH_PREPROC_PATH value in environmental variables')
    sys.exit(1)

ENV = os.environ

default_args = {
  # 'start_date': datetime.now(),
  'start_date': datetime(2017, 7, 18)
}

DAG = DAG(
  dag_id='dash_preproc',
  default_args=default_args
)

clear_tables = PythonOperator(
  task_id='clear_tables',
  python_callable=dash_workers.clear_db,
  dag=DAG)

def id_worker(uid):
    return PythonOperator(
        task_id=uid,
        python_callable=dash_workers.main_preprocess,
        op_args=[uid],
        dag=DAG)

for uid in capone_dash_workers.get_id_creds():
    clear_tables >> id_worker(uid)

clear_tables cleans the database that will be re-built as a result of the process. id_worker is a function that dynamically generates new preprocessing tasks, based on the array of ID values returned from get_if_creds. The task ID is just the corresponding user ID, though it could easily have been an index, i, as in the example mentioned above.

NOTE That bitshift operator (<<) looks backwards to me, as the clear_tables task should come first, but it's what seems to be working in this case.

Upvotes: 6

sdikby
sdikby

Reputation: 1471

Considering that Apache Airflow is a workflow management tool, ie. it determines the dependencies between task that the user defines in comparison (as an example) with apache Nifi which is a dataflow management tool, ie. the dependencies here are data which are transferd through the tasks.

That said, i think that your approach is quit right (my comment is based on the code posted) but Airflow offers a concept called XCom. It allows tasks to "cross-communicate" between them by passing some data. How big should the passed data be ? it is up to you to test! But generally it should be not so big. I think it is in the form of key,value pairs and it get stored in the airflow meta-database,ie you can't pass files for example but a list with ids could work.

Like i said you should test it your self. I would be very happy to know your experience. Here is an example dag which demonstrates the use of XCom and here is the necessary documentation. Cheers!

Upvotes: 3

Related Questions