Job queue with worker pool

Design a job/task queue with a pool of worker processes for a background-processing platform: producers enqueue ~20k jobs/sec, jobs take 50ms–5min, and the system must guarantee at-least-once execution even if a worker dies mid-job. Requirements: no job is silently lost, a stuck job is retried, and a single poison job can't block its queue forever. Describe enqueue/dequeue, how in-flight jobs are tracked, and how concurrent workers avoid running the same job twice in the common case.