Stealing from the busiest list

Most designs keep a small local list per worker so they do not all contend on one shared queue. That reintroduces the original problem in miniature, because an item already on a local list is claimed and a free worker elsewhere cannot reach it.

Work stealing is the escape hatch. An idle worker may take an item from the tail of a busy worker's local list, so a long item on one list does not leave another worker waiting while the queue is empty.

Keep the local lists short. Large lists defeat both halves of the pattern: they reduce the chance the shared queue has anything left to take, and they leave less on the tail for a stealing worker, so the pool drifts back towards fixed assignment.

STEAL_LIMIT = 3

def steal(victim, limit=STEAL_LIMIT):
    if len(victim) <= limit:
        return None
    return victim.pop()

The limit stops a pool from shuffling a short list back and forth instead of doing work.