How to run Apache Beam Python pipelines on Flink cluster in Kubernetes?
By default (as of this writing), Beam stages runtime dependencies ("artifacts") to a certain directory (/tmp/staged by default) that needs to be accessible to both the job server (in your case, the client) and the Beam worker.
You can get around this by setting the --flink_submit_uber_jar pipeline option. When --flink_submit_uber_jar is set, Beam wraps all your dependencies in a jar that is submitted to Flink.