7

I am currently converting workflows that were implemented in bash scripts before to Airflow DAGs. In the bash scripts, I was just exporting the variables at run time with

export HADOOP_CONF_DIR="/etc/hadoop/conf"

Now I'd like to do the same in Airflow, but haven't found a solution for this yet. The one workaround I found was setting the variables with os.environ[VAR_NAME]='some_text' outside of any method or operator, but that means they get exported the moment the script gets loaded, not at run time.

Now when I try to call os.environ[VAR_NAME] = 'some_text' in a function that gets called by a PythonOperator, it does not work. My code looks like this

def set_env():
    os.environ['HADOOP_CONF_DIR'] = "/etc/hadoop/conf"
    os.environ['PATH'] = "somePath:" + os.environ['PATH']
    os.environ['SPARK_HOME'] = "pathToSparkHome"
    os.environ['PYTHONPATH'] = "somePythonPath"
    os.environ['PYSPARK_PYTHON'] = os.popen('which python').read().strip()
    os.environ['PYSPARK_DRIVER_PYTHON'] = os.popen('which python').read().strip()

set_env_operator = PythonOperator(
    task_id='set_env_vars_NOT_WORKING',
    python_callable=set_env,
    dag=dag)

Now when my SparkSubmitOperator gets executed, I get the exception:

Exception in thread "main" java.lang.Exception: When running with master 'yarn' either HADOOP_CONF_DIR or YARN_CONF_DIR must be set in the environment.

My use case where this is relevant is that I have SparkSubmitOperator, where I submit jobs to YARN, therefore either HADOOP_CONF_DIR or YARN_CONF_DIR must be set in the environment. Setting them in my .bashrc or any other config is sadly not possible for me, which is why I need to set them at runtime.

Preferably I'd like to set them in an Operator before executing the SparkSubmitOperator, but if there was the possibility to pass them as arguments to the SparkSubmitOperator, that would be at least something.

stefanobaghino
  • 11,253
  • 4
  • 35
  • 63
Christopher Beck
  • 735
  • 8
  • 19
  • I just found this: https://github.com/apache/incubator-airflow/pull/3268/commits, which seems to fix this for the SparkSubmitOperator in the next airflow version, but what about if I have another Operator? – Christopher Beck Jul 12 '18 at 23:47
  • 2
    That won't work since afaik every task is being executed as a "separate" Python script process, if you will. Since the first execution sets the variables and the second execution runs the Spark operator, the second execution never sees the variables. – tobi6 Jul 13 '18 at 07:28

1 Answers1

2

From what I can see in the spark submit operator you can pass in environment variables to spark-submit as a dictionary. 

:param env_vars: Environment variables for spark-submit. It
                 supports yarn and k8s mode too.
:type env_vars: dict

Have you tried this?

Simon D
  • 5,730
  • 2
  • 17
  • 31
  • As far as I can see this is not in the released version yet. – Christopher Beck Jul 13 '18 at 00:18
  • 1
    You could always just pick those changes out and dump it into your own plugin. Or just overwrite the copy of the spark submit operator on your own version with the one from the master branch - that's a bit dodgy though :) – Simon D Jul 13 '18 at 00:23
  • I just realized that I've created every part of my workflow with SubDags because I need to create tasks at run time, which by my understanding also means that if I just put the call to os.environ[Var] = Var in there, it should get executed when the main dag gets loaded and when the subdag gets executed. I'm not 100% happy that it will execute when the main dag gets loaded but otherwise that might actually be ok :) – Christopher Beck Jul 13 '18 at 01:27