Apache Airflow version: 2.0.0-b3

Kubernetes version: v1.18.3+47c0e71, OpenShift v4.5

What happened:

When starting a worker pod with the KubernetesOperator, the python script fails to import airflow

What you expected to happen:

Running a docker container in OpenShift is usually done with the "Restricted" Security Context Constraint, meaning that the container will be run with an arbitrary UID, and GID 0. The airflow docker image supports this (#9545), which relies on the following lines from the container entrypoint:

However, the entrypoint is currently overridden:

command changes the container ENTRYPOINT (not the CMD as one might think 🤪 ), which means that the production image entrypoint is not used by the worker pod, preventing the airflow command from executing properly. To execute a custom command while keeping the original entrypoint, one should use the args parameter instead of command

Looking at this with @pbastia. Sounds like a solution would be to:

• add an args parameter to PodGenerator.construct_pod
• in the kubernetes_executor, set args to command[1:], since the container entrypoint expects an airflow command by default (see below)

Happy to open a PR for it if that sounds like an acceptable change.
cc @dimberman

How to reproduce it:

Run any dag in OpenShift, using the restricted SCC