Looking into the job manager logs, I observe the following events leading up to the failure:
1. Registering stream consumer
2. Starting consumption from stream
3. Required number of readers: 2, Registered readers: 0
4. Required number of readers: 2, Registered readers: 1
5. Can change the parallelism of the job. Restarting the job.
6. Job switched from state RUNNING to CANCELLING
7. Job switched from state CANCELLING to CANCELED
8. De-registering stream consumer
9. Running initialization on master for JOB_NAME
10. Waiting for stream consumer to be deregistered
11. Job switched from state CREATED to RUNNING.
12. Registering stream consumer
13. Found existing consumer. Proceeding to read from consumer.
14. Starting consumption from stream
15. Stream consumer has been deregistered
16. Required number of readers: 2, Registered readers: 0 (indefinitely, operator stalls on initializing phase)