Which Python pool wrappers reseed NumPy's global RNG per worker
Measured with 16 tasks on 4 workers, each returning np.random.random(), from a script that imports pandas (which loads numpy.random in the parent). CPython 3.12.3 (fork default) and 3.14.10 (forkserver default), NumPy 2.4.6/2.5.4. Only 4/16 distinct (the k-th task of every worker repeats) on both: stdlib Pool and ProcessPoolExecutor, joblib backend='multiprocessing', multiprocess 0.70.19, pathos 0.3.5 ProcessPool, mpire 2.10.2 WorkerPool, pebble 5.2.3 ProcessPool, billiard 4.3.2 (it reseeds only stdlib random), tqdm.contrib.concurrent.process_map, and pqdm processes; pandarallel 1.6.5 too on 3.12. Independent 16/16: joblib default loky, loky, dask scheduler='processes' (dask/multiprocessing.py calls np.random.seed() in each worker), distributed LocalCluster, ray, and aiomultiprocess (spawn by default). Stdlib random was 16/16 everywhere. So switching from Pool to mpire or pebble does not fix duplicated draws; pass per-task Generators from SeedSequence(seed).spawn(n), or reseed in an initializer. pathos has an opt-in helper, pathos.helpers.mp_helper.random_state('numpy.random', seed='*').
Reproduced the table on CPython 3.12.16 (fork) and 3.14.10 (forkserver), NumPy 2.5.4: 16 tasks on 4 workers, chunksize=1 where the API has one, 50 ms sleep per task so all workers get work. 4/16 distinct on both versions for stdlib Pool, ProcessPoolExecutor, joblib backend='multiprocessing', multiprocess 0.70.19, pathos 0.3.5 ProcessPool, mpire 2.10.2, pebble 5.2.3, billiard 4.3.2 and tqdm 4.70.1 process_map. 16/16 for joblib 1.6.0 default loky, loky 3.7.0 and dask 2026.8.0 scheduler='processes' (np.random.seed() at dask/multiprocessing.py:260). billiard's popen_fork/forkserver call only random.seed(). I did not test pandarallel, pqdm, distributed, ray or aiomultiprocess.
One correction: the pathos helper does not work for NumPy. random_state('numpy.random', seed='*') computes seed = pid + int(time.time()*1e6), about 1.8e15, and the legacy seed accepts only 0..2**32-1. So it raises "ValueError: Seed must be between 0 and 2**32 - 1" every time, inside a worker or not, with new=True too, and NumPy 1.26.4 rejects the same seed. It only works for module='random'. Even reduced mod 2**32 the scheme is weak by construction: pid+1 at time t-1 µs gives the same seed as pid at t.
A working one-liner for the legacy global: mp.Pool(4, initializer=np.random.seed) gives 16/16 under both fork and forkserver (OS entropy, like dask; not reproducible). For reproducible runs, keep the SeedSequence(seed).spawn(n) per-task Generators you suggest.