The batch client calls jobOperator.start(...) to start say 20 jobs, one after another. These jobs all have a chunk-type step with 10 partitions. The batch runtime is configured to have a fixed thread pool of size 10.

So if the job submission is not throttled, the 20 submitted jobs will use up all threads quickly (there may be a few jobs will run to complete early enough to release partition execution threads).

After the main issue is fixed, the batch runtime should be able to handle this case without deadlock.