Beam Python on Flink fails when no artifacts staged

When a Beam job with no artifacts staged is deployed on a Beam-on-Flink cluster (without the JobServer), it crashes the Python beam worker pool and does not recover.  This causes that job (and subsequent jobs that would have used that task manager slot) to hang and fail.  Strangely, if a Beam job with artifacts staged is run on that Beam worker pool container instance (i.e. using that task manager slot), subsequent jobs without artifacts staged that end up using that task manager slot run properly and succeed.

When this happens, the error from the worker pool container looks like this:

2020-08-19 01:58:12.287 PDT
2020/08/19 08:58:12 Initializing python harness: /opt/apache/beam/boot --id=1-1 --logging_endpoint=localhost:38839 --artifact_endpoint=localhost:43187 --provision_endpoint=localhost:37259 --control_endpoint=localhost:34931
2020-08-19 01:58:12.305 PDT
2020/08/19 08:58:12 Failed to retrieve staged files: failed to get manifest
2020-08-19 01:58:12.305 PDT
caused by:
2020-08-19 01:58:12.305 PDT
rpc error: code = Unimplemented desc = Method not found: org.apache.beam.model.job_management.v1.LegacyArtifactRetrievalService/GetManifest

To reproduce this, it is sufficient to run the wordcount example modified to set the PipelineOptions flag `save_main_session=False` with Beam 2.23.

python -m apache_beam.examples.wordcount --runner=FlinkRunner --flink_master=$FLINK_MASTER_HOST:8081 --flink_submit_uber_jar --environment_type=EXTERNAL --environment_config=localhost:50000 --input gs://dataflow-samples/shakespeare/kinglear.txt --output gs://my/bucket/location

This was tested in a Beam-on-Flink cluster deployed on Kubernetes (https://github.com/GoogleCloudPlatform/flink-on-k8s-operator) with a modified version of this patch to use Beam 2.23 (https://github.com/GoogleCloudPlatform/flink-on-k8s-operator/pull/301, i.e. with instances of "2.22" replaced with "2.23").