Templated Variables being treated as String by KubernetesPodOperator in Airflow 2.0

I’m encountering an issue that I believe is similar to #13348, but am opening a new issue to expand on what is happening in more detail.

Apache Airflow version: 2.0.0

Kubernetes version (if you are using kubernetes) (use kubectl version): 1.19.4

Environment:

• Cloud provider or hardware configuration: GCP, using GKE

What happened:

My DAG in Airflow 1.10.13 was using a couple of built-in variables and Jinja templating to pass information between KubernetesPodOperator tasks (e.g. file paths, execution dates); it looked similar to this:

from airflow import DAG
from airflow.contrib.operators.kubernetes_pod_operator import KubernetesPodOperator

default_args = {'owner': 'airflow'}

example_workflow = DAG('my_dag',
                       default_args=default_args,
                       default_view='graph',
                       schedule_interval='@daily')

with example_workflow:
    example_k8s_task = KubernetesPodOperator(namespace='my_namespace',
                                             image="ubuntu:latest",
                                             name="pod1",
                                             task_id='my_task_id',
                                             env_vars={'TASK_ID': '{{ task.task_id }}'},
                                             cmds=["echo $TASK_ID"]
                                             )

    example_k8s_task

which successfully echoed 'my_task_id' in the logs when the DAG was triggered and ran successfully.

With the changes to the KubernetesPodOperator in Airflow 2.0, I followed the guidelines here and refactored the above code to look like this:

from airflow import DAG
from airflow.providers.cncf.kubernetes.operators.kubernetes_pod import KubernetesPodOperator
from kubernetes.client import models as k8s

default_args = {'owner': 'airflow'}

my_env_vars = [
    k8s.V1EnvVar(
        name='TASK_ID',
        value='{{ task.task_id }}'
    )]

example_workflow = DAG(dag_id='my_dag',
                       default_args=default_args,
                       default_view='graph',
                       schedule_interval='@daily')

with example_workflow:
    example_k8s_task = KubernetesPodOperator(namespace='my_namespace',
                                             image="ubuntu:latest",
                                             name="pod1",
                                             task_id='my_task_id',
                                             env_vars=my_env_vars,
                                             cmds=["echo $TASK_ID"]
                                             )

    example_k8s_task

However, this echoed '{{ task.task_id }}' in the logs, meaning the templated variable did not render properly.

What I expected to happen:

I expected to have the templated variable rendered properly and return the variable value, e.g. 'my_task_id'.

What I think actually happened:

While KubernetesPodOperator env_var parameter is marked as “(templated)” in the source code, when a string is passed in the value parameter of kubernetes.client V1EnvVar object (source code), it does not render the template, but rather treats it as a literal string.

How to reproduce it:

Set up Airflow on minikube or use your cloud provider’s Airflow instance (so easy to set up, I know /s) and run the second Python code snippet above (using Airflow 2.0).

Thanks!

[dimberman]

@phlope yes when we release the next backport provider it will also have this fix.

[dimberman]

Hi @shihabcsedu09

I’ve created and linked a PR. Because the KPO lives in a provider once this is merged we can release and you can upgrade for this fix without needing to upgrade your airflow version (just pull the newest provider 😃 )