Repository navigation
Test full parallel MonteCarlo determinism under spawn and forkserver, not only seed derivation #1076
Description
Activity
- added a commit that references this issue
on Jul 20, 2026 Half of this is covered now, recording which half so the issue stays accurate.
#1054 adds
test_an_index_keeps_its_inputs_when_the_workers_are_spawned, which runs a realsimulate()on two and four workers with the start method forced tospawn, and compares the sampled inputs against a serial run of the same seed. No stubbedFlight, and no skip on the start method, so the end-to-end half of this holds forspawn.Getting there took fixing the comparison rather than the code. The first version compared serialized records directly and failed, which looked like the guarantee breaking under
spawn. It was not:RocketPyEncoderwritessignature["hash"] = hash(obj), which is process-local, and a callablesourceas a dill blob whose bytes differ between processes for a value nothing varies. The test now strips those two fields before comparing.Still open:
forkserveris exercised only through the derivation helpers, the same as before. Leaving this issue open for that.
Context
#1054 added
test_seed_derivation_is_start_method_invariant, which checks that the per-index seed a worker derives matches the main process under fork, spawn and forkserver. It runs in ordinary CI on all three start methods, which covers the seeding math, but it deliberately exercises only the pure derivation helpers, not a realsimulate().The gap
The end-to-end
serial == parallel(n)test still stubs the module-levelFlightand skips unless the start method isfork, because the stub reaches workers only through fork inheritance. So the full production path (the realmultiprocesspool,__sim_producer, the manager, bound-method pickling, the stochastic object graph, the file writes) is only verified under fork. Python 3.14 madeforkserverthe POSIX default, and Windows and macOS usespawn, so fork is no longer the default anywhere. The untested modes are the ones most users will actually hit.A possible direction
Give the worker-invariance test a picklable, top-level flight factory (or a very small real
Flight) instead of a module-level monkeypatch, so it can run a shortsimulate(parallel=True)under spawn and forkserver and assert that serial and parallel produce identical inputs per index, with no gaps or duplicates and an empty error file.Follow-up to the review discussion on #1054.