There is a naive implementation of parallel processing that kicks off all available pending jobs as soon as possible. Obviously it's not 100% bulletproof and requires user to keep this in mind.
Instead it should be possible to limit the number of jobs that could be running within single worker in parallel.