From 5a572d3ce39c91f677d0a391be371257ee78e9ae Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Tue, 18 Aug 2026 02:07:55 +0800 Subject: [PATCH 1/6] BUG: accept the seed type a Monte Carlo worker is handed A parallel run spawns a SeedSequence per worker and passes it to environment, rocket and flight. _sampler_seed then fed it to SeedSequence(entropy=...), which takes an int or a sequence of ints, so the first worker raised TypeError before drawing anything. The call was reached only from the custom sampler reset until #1117 added the list-choice generator, which every model goes through. A real two-worker run passes at d21abde6^ in 2.32s and does not finish on develop: the worker's own error path raises UnboundLocalError on inputs_json, so the parent never learns it died and the run hangs. The children of one root share their entropy and differ by spawn_key, so the value is folded through generate_state rather than read off entropy, which would put every worker on one sampler stream. Nothing is consumed, and an int or None seed keeps the stream it had. The fold lives in rocketpy.tools, since the component streams and the per-index seeding both need the same one and three copies would drift on width and word order. _sampler_seed does its own final fold through it as well rather than repeating the four lines. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- rocketpy/stochastic/stochastic_model.py | 20 ++++- rocketpy/tools.py | 11 +++ .../test_monte_carlo_parallel_runs.py | 32 ++++++++ tests/unit/stochastic/test_seed_types.py | 81 +++++++++++++++++++ 4 files changed, 140 insertions(+), 4 deletions(-) create mode 100644 tests/unit/simulation/test_monte_carlo_parallel_runs.py create mode 100644 tests/unit/stochastic/test_seed_types.py diff --git a/rocketpy/stochastic/stochastic_model.py b/rocketpy/stochastic/stochastic_model.py index d42fb76c5..1dadb2f01 100644 --- a/rocketpy/stochastic/stochastic_model.py +++ b/rocketpy/stochastic/stochastic_model.py @@ -8,7 +8,7 @@ from rocketpy.mathutils.function import Function from rocketpy.stochastic.custom_sampler import CustomSampler -from ..tools import get_distribution +from ..tools import _seed_sequence_to_int, get_distribution def _names_as_spawn_key(input_names): @@ -41,6 +41,18 @@ def _format_number(value): return f"array of shape {np.shape(value)}" +def _seed_as_entropy(seed): + """A seed as something ``SeedSequence`` will take as entropy. + + A parallel run is handed a ``SeedSequence``, which it will not take. Any + other seed goes through untouched, so the stream an int reaches stays where + it was. + """ + if not isinstance(seed, np.random.SeedSequence): + return seed + return _seed_sequence_to_int(seed) + + def _sampler_seed(seed, input_names): """Derive a seed for one sampler, or for one group that shares a generator. @@ -54,10 +66,10 @@ def _sampler_seed(seed, input_names): # Sorted here rather than trusting the caller, so a future call site cannot # give one group two different seeds by listing its members another way. root = np.random.SeedSequence( - entropy=seed, spawn_key=_names_as_spawn_key(tuple(sorted(input_names))) + entropy=_seed_as_entropy(seed), + spawn_key=_names_as_spawn_key(tuple(sorted(input_names))), ) - words = root.generate_state(4, dtype=np.uint32) - return sum(int(word) << (32 * position) for position, word in enumerate(words)) + return _seed_sequence_to_int(root) # TODO: Stop using assert in production code. Use exceptions instead. diff --git a/rocketpy/tools.py b/rocketpy/tools.py index 0d7f1a74e..7f31f3e19 100644 --- a/rocketpy/tools.py +++ b/rocketpy/tools.py @@ -1377,6 +1377,17 @@ def euler313_to_quaternions(phi, theta, psi): return e0, e1, e2, e3 +def _seed_sequence_to_int(seed_sequence): + """Returns a ``SeedSequence`` as the 128-bit ``int`` it can be rebuilt from. + + Folded through ``generate_state`` rather than read off ``entropy``, since + the children of one root differ only by ``spawn_key``, and combined by + value so it does not depend on byte order. + """ + words = seed_sequence.generate_state(4, dtype=np.uint32) + return sum(int(word) << (32 * position) for position, word in enumerate(words)) + + def get_matplotlib_supported_file_endings(): """Gets the file endings supported by matplotlib. diff --git a/tests/unit/simulation/test_monte_carlo_parallel_runs.py b/tests/unit/simulation/test_monte_carlo_parallel_runs.py new file mode 100644 index 000000000..4ab0be440 --- /dev/null +++ b/tests/unit/simulation/test_monte_carlo_parallel_runs.py @@ -0,0 +1,32 @@ +import pytest + +from rocketpy.simulation.monte_carlo import MonteCarlo + + +@pytest.mark.parametrize("parallel", [False, True]) +def test_a_monte_carlo_run_finishes( + stochastic_environment, stochastic_calisto, stochastic_flight, tmp_path, parallel +): + # The parallel path hands each worker a SeedSequence rather than an int, and + # nothing else in the suite exercises that. A worker that dies on it is not + # reported, so this reads as a hang rather than as a failure. + # + # Built here rather than taken from the monte_carlo_calisto fixture, whose + # own filename is fixed, since `filename` is a plain attribute and the three + # working paths are settled when the object is constructed. + analysis = MonteCarlo( + filename=str(tmp_path / "study"), + environment=stochastic_environment, + rocket=stochastic_calisto, + flight=stochastic_flight, + ) + + analysis.simulate( + number_of_simulations=2, + append=False, + parallel=parallel, + n_workers=2 if parallel else None, + ) + + assert analysis.num_of_loaded_sims == 2 + assert str(tmp_path) in str(analysis.output_file) diff --git a/tests/unit/stochastic/test_seed_types.py b/tests/unit/stochastic/test_seed_types.py new file mode 100644 index 000000000..ba3d42583 --- /dev/null +++ b/tests/unit/stochastic/test_seed_types.py @@ -0,0 +1,81 @@ +import numpy as np +import pytest + +from rocketpy.stochastic.stochastic_model import ( + _names_as_spawn_key, + _sampler_seed, +) +from rocketpy.tools import _seed_sequence_to_int + + +def _a_worker_seed(index=0, workers=2): + # What MonteCarlo.__run_in_parallel spawns and hands to each worker, which + # passes it straight to environment/rocket/flight._set_stochastic. + return np.random.SeedSequence().spawn(workers)[index] + + +def test_the_seed_type_a_worker_is_handed_is_accepted(stochastic_calisto): + stochastic_calisto._set_stochastic(_a_worker_seed()) + + stochastic_calisto.create_object() + + +def test_a_parachute_derives_its_noise_seed_from_a_worker_seed( + stochastic_main_parachute, +): + stochastic_main_parachute._set_stochastic(_a_worker_seed()) + + assert stochastic_main_parachute.create_object().noise[2] is not None + + +def test_two_workers_do_not_share_a_sampler_stream(): + first, second = np.random.SeedSequence(7).spawn(2) + # They come off one root, so they carry the same entropy and differ only in + # spawn_key. Reading the entropy alone would put both on one stream. + assert first.entropy == second.entropy + + assert _sampler_seed(first, ("__list_choice__",)) != _sampler_seed( + second, ("__list_choice__",) + ) + + +def test_a_caller_seed_sequence_is_not_consumed(): + root = np.random.SeedSequence(42) + + _sampler_seed(root, ("__list_choice__",)) + + assert root.n_children_spawned == 0 + assert root.spawn(1)[0].spawn_key == (0,) + + +def test_the_same_seed_sequence_twice_gives_the_same_sampler_seed(): + root = np.random.SeedSequence(42) + + first = _sampler_seed(root, ("pressure_noise", "main")) + second = _sampler_seed(root, ("pressure_noise", "main")) + + assert first == second + + +@pytest.mark.parametrize("seed", [42, 7, [1, 2, 3]]) +@pytest.mark.parametrize("names", [("__list_choice__",), ("pressure_noise", "main")]) +def test_a_seed_that_is_not_a_sequence_reaches_numpy_untouched(seed, names): + # The control. Every fixed-seed baseline in the suite was recorded through + # this path, so anything but a SeedSequence has to arrive as it always did. + # Compared with the expression rather than with a recorded number, which + # would go red on a NumPy release instead of on a change of ours. + unchanged = np.random.SeedSequence( + entropy=seed, spawn_key=_names_as_spawn_key(tuple(sorted(names))) + ) + + assert _sampler_seed(seed, names) == _seed_sequence_to_int(unchanged) + + +def test_no_seed_still_means_no_seed(): + # None is left out above on purpose: it asks NumPy for fresh entropy, so + # two calls must not agree, and comparing one against another would be + # asserting the opposite of what an unseeded run promises. + first = _sampler_seed(None, ("__list_choice__",)) + second = _sampler_seed(None, ("__list_choice__",)) + + assert first != second From 73773f22c819edfac7c8d7f76f80d5ba9d56c215 Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Tue, 18 Aug 2026 05:16:07 +0800 Subject: [PATCH 2/6] BUG: number serial simulations the way parallel numbers them The two run paths named the same simulation differently. Three of them wrote 1, 2, 3 through the serial path and 0, 1, 2 through the parallel one, so a row could not be compared with its counterpart and an index meant nothing on its own. Serial counts from zero now, which is what the parallel path already did and what append already assumed: num_of_loaded_sims counts rows, so a two-row checkpoint resumes at 2, an index the serial path never used. Existing serial results are numbered one higher than the same run would be numbered now. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- rocketpy/simulation/monte_carlo.py | 12 ++- .../test_monte_carlo_simulation_index.py | 76 +++++++++++++++++++ 2 files changed, 84 insertions(+), 4 deletions(-) create mode 100644 tests/unit/simulation/test_monte_carlo_simulation_index.py diff --git a/rocketpy/simulation/monte_carlo.py b/rocketpy/simulation/monte_carlo.py index c2dcd4030..cec94fbc3 100644 --- a/rocketpy/simulation/monte_carlo.py +++ b/rocketpy/simulation/monte_carlo.py @@ -413,14 +413,18 @@ def __run_in_serial(self): n_simulations=self.number_of_simulations, start_time=time(), ) + sim_idx = sim_monitor.count try: while sim_monitor.keep_simulating(): - sim_monitor.increment() + # Counted from zero, as the parallel path already does. The two + # named the same simulation differently: three of them wrote + # 1, 2, 3 here and 0, 1, 2 there. + sim_idx = sim_monitor.increment() - 1 inputs_json, outputs_json = "", "" flight = self.__run_single_simulation() - inputs_json = self.__evaluate_flight_inputs(sim_monitor.count) - outputs_json = self.__evaluate_flight_outputs(flight, sim_monitor.count) + inputs_json = self.__evaluate_flight_inputs(sim_idx) + outputs_json = self.__evaluate_flight_outputs(flight, sim_idx) self._append_simulation_record(inputs_json, outputs_json) @@ -434,7 +438,7 @@ def __run_in_serial(self): f.write(inputs_json) except Exception as error: - print(f"Error on iteration {sim_monitor.count}: {error}") + print(f"Error on iteration {sim_idx}: {error}") with open(self._error_file, "a", encoding="utf-8") as f: f.write(inputs_json) raise error diff --git a/tests/unit/simulation/test_monte_carlo_simulation_index.py b/tests/unit/simulation/test_monte_carlo_simulation_index.py new file mode 100644 index 000000000..7d2d6a52a --- /dev/null +++ b/tests/unit/simulation/test_monte_carlo_simulation_index.py @@ -0,0 +1,76 @@ +import json + +import pytest + +from rocketpy.simulation.monte_carlo import MonteCarlo + + +def _indices(analysis): + with open(analysis.output_file, "r", encoding="utf-8") as written: + return sorted(json.loads(line)["index"] for line in written if line.strip()) + + +def _a_study(tmp_path, stem, environment, rocket, flight): + return MonteCarlo( + filename=str(tmp_path / stem), + environment=environment, + rocket=rocket, + flight=flight, + ) + + +@pytest.mark.parametrize("parallel", [False, True]) +def test_a_run_numbers_its_simulations_from_zero( + stochastic_environment, stochastic_calisto, stochastic_flight, tmp_path, parallel +): + # Serial used to write 1, 2, 3 while parallel wrote 0, 1, 2, so the same + # simulation had two names depending on how the run was started. + analysis = _a_study( + tmp_path, + f"study-{parallel}", + stochastic_environment, + stochastic_calisto, + stochastic_flight, + ) + + analysis.simulate( + number_of_simulations=3, + append=False, + parallel=parallel, + n_workers=2 if parallel else None, + ) + + assert _indices(analysis) == [0, 1, 2] + + +def test_both_modes_agree_on_what_a_simulation_is_called( + stochastic_environment, stochastic_calisto, stochastic_flight, tmp_path +): + one = _a_study( + tmp_path, + "serial", + stochastic_environment, + stochastic_calisto, + stochastic_flight, + ) + other = _a_study( + tmp_path, "para", stochastic_environment, stochastic_calisto, stochastic_flight + ) + + one.simulate(number_of_simulations=3, append=False, parallel=False) + other.simulate(number_of_simulations=3, append=False, parallel=True, n_workers=2) + + assert _indices(one) == _indices(other) + + +def test_an_appended_run_carries_on_from_the_last_index( + stochastic_environment, stochastic_calisto, stochastic_flight, tmp_path +): + analysis = _a_study( + tmp_path, "study", stochastic_environment, stochastic_calisto, stochastic_flight + ) + + analysis.simulate(number_of_simulations=2, append=False) + analysis.simulate(number_of_simulations=4, append=True) + + assert _indices(analysis) == [0, 1, 2, 3] From 5fd8408f1eacf63161280a7102231a65e026d433 Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Tue, 18 Aug 2026 05:32:34 +0800 Subject: [PATCH 3/6] ENH: derive each simulation's seed from its own index A run seeded its workers, one seed each, from fresh entropy every time, and the serial path never reseeded at all. So the same study gave different results run to run, different results in the two modes, and different results again when the worker count changed. Addresses #1053. simulate() takes random_seed now, keyword-only, and every simulation takes the child of that root belonging to its index. The child is derived directly rather than by spawning the ones before it: spawn appends n_children_spawned + i to the parent key, so rebuilding that one child reproduces it bit for bit, and a worker reaches any index from four picklable values instead of a list a million long. There is a test comparing it with spawn(n)[i] for every seed type. Measured over real flights, four simulations: serial(42) == serial(42) True serial(42) == parallel(2 workers, 42) True serial(42) == parallel(4 workers, 42) True serial(42) == serial(7) False The per-worker seed is gone rather than kept alongside, since a worker now decides nothing about sampling and how many there are cannot reach it. Appending continues the same stream when the same seed is given, since an index maps to a seed and nothing else. Nothing here checks that the caller did give the same one; persisting the root so it can be checked is #1075. Fixed-seed results change: every study is sampled from a different place. Nothing that was reproducible before stops being so, because nothing was. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- rocketpy/simulation/monte_carlo.py | 92 ++++++++-- .../simulation/test_monte_carlo_seeding.py | 157 ++++++++++++++++++ 2 files changed, 238 insertions(+), 11 deletions(-) create mode 100644 tests/unit/simulation/test_monte_carlo_seeding.py diff --git a/rocketpy/simulation/monte_carlo.py b/rocketpy/simulation/monte_carlo.py index cec94fbc3..7c4b86a34 100644 --- a/rocketpy/simulation/monte_carlo.py +++ b/rocketpy/simulation/monte_carlo.py @@ -18,6 +18,7 @@ import os import traceback import warnings +from copy import deepcopy from numbers import Real from pathlib import Path from time import time @@ -31,6 +32,7 @@ from rocketpy.prints.monte_carlo_prints import _MonteCarloPrints from rocketpy.simulation.flight import Flight from rocketpy.tools import ( + _seed_sequence_to_int, generate_monte_carlo_ellipses, generate_monte_carlo_ellipses_coordinates, import_optional_dependency, @@ -44,6 +46,57 @@ _SIMULATION_LOG_SUFFIX = ".txt" +def _root_seed_sequence(random_seed): + """The immutable root a run derives every simulation's seed from. + + A ``SeedSequence`` is rebuilt from its full state rather than used as + given, since ``spawn`` advances a counter the caller still holds. A + ``Generator`` is refused rather than read, because using a consume-on-use + object as an immutable seed cannot mean what it says. + """ + if isinstance(random_seed, np.random.SeedSequence): + return np.random.SeedSequence(**random_seed.state) + if isinstance(random_seed, (np.random.Generator, np.random.BitGenerator)): + raise TypeError( + f"random_seed must be an int, a sequence of non-negative integers, " + f"or a numpy.random.SeedSequence, not a " + f"{type(random_seed).__name__}. Pass the seed the generator was " + f"built from." + ) + return np.random.SeedSequence(random_seed) + + +def _root_state_of(root): + """A root as the four picklable values a worker can rebuild it from. + + Sent to each worker instead of the object, and instead of the list of + children, so a run of a million simulations costs four values. The entropy + is copied because a sequence one is kept by reference all the way from the + caller, who could otherwise still move every child by editing their list. + """ + return ( + deepcopy(root.entropy), + tuple(root.spawn_key), + root.pool_size, + root.n_children_spawned, + ) + + +def _seed_of_simulation(root_state, sim_idx): + """The seed for one simulation index, without spawning the ones before it. + + ``spawn`` derives child ``i`` by appending ``n_children_spawned + i`` to + the parent spawn key, so rebuilding that one child directly reproduces it + and any index can be reached from the four values above alone. + """ + entropy, spawn_key, pool_size, base = root_state + return np.random.SeedSequence( + entropy=entropy, + spawn_key=(*spawn_key, base + sim_idx), + pool_size=pool_size, + ) + + def _refuse_logs_this_run_cannot_write( input_file, output_file, error_file, export_config=None ): @@ -265,6 +318,8 @@ def simulate( append=False, parallel=False, n_workers=None, + *, + random_seed=None, **kwargs, ): """ @@ -317,6 +372,11 @@ def simulate( self._export_config = kwargs self.number_of_simulations = number_of_simulations self._initial_sim_idx = self.num_of_loaded_sims if append else 0 + # Validated here, before __setup_files truncates anything, so an + # unusable seed cannot cost a previous run its results. Kept as four + # picklable values rather than as the object, since a worker rebuilds + # any index from them. + self.__root_state = _root_state_of(_root_seed_sequence(random_seed)) # Before anything is opened: __setup_files truncates for append=False. _refuse_logs_this_run_cannot_write( @@ -422,6 +482,7 @@ def __run_in_serial(self): sim_idx = sim_monitor.increment() - 1 inputs_json, outputs_json = "", "" + self.__seed_this_simulation(sim_idx) flight = self.__run_single_simulation() inputs_json = self.__evaluate_flight_inputs(sim_idx) outputs_json = self.__evaluate_flight_outputs(flight, sim_idx) @@ -473,13 +534,14 @@ def __run_in_parallel(self, n_workers=None): ) processes = [] - seeds = np.random.SeedSequence().spawn(n_workers) - for seed in seeds: + # No seed per worker any more: every simulation takes its own from + # its index, so the workers are interchangeable and how many there + # are does not reach the sampling. + for _ in range(n_workers): sim_producer = multiprocess.Process( target=self.__sim_producer, args=( - seed, sim_monitor, mutex, simulation_error_event, @@ -521,13 +583,11 @@ def __validate_number_of_workers(self, n_workers): raise ValueError("Number of workers must be at least 2 for parallel mode.") return n_workers - def __sim_producer(self, seed, sim_monitor, mutex, error_event): # pylint: disable=too-many-statements + def __sim_producer(self, sim_monitor, mutex, error_event): """Simulation producer to be used in parallel by multiprocessing. Parameters ---------- - seed : int - The seed to set the random number generator. sim_monitor : _SimMonitor The simulation monitor object to keep track of the simulations. mutex : multiprocess.Lock @@ -536,15 +596,11 @@ def __sim_producer(self, seed, sim_monitor, mutex, error_event): # pylint: disa Event signaling an error occurred during the simulation. """ try: - # Ensure Processes generate different random numbers - self.environment._set_stochastic(seed) - self.rocket._set_stochastic(seed) - self.flight._set_stochastic(seed) - while sim_monitor.keep_simulating(): sim_idx = sim_monitor.increment() - 1 inputs_json, outputs_json = "", "" + self.__seed_this_simulation(sim_idx) flight = self.__run_single_simulation() inputs_json = self.__evaluate_flight_inputs(sim_idx) outputs_json = self.__evaluate_flight_outputs(flight, sim_idx) @@ -584,6 +640,20 @@ def __sim_producer(self, seed, sim_monitor, mutex, error_event): # pylint: disa error_event.set() mutex.release() + def __seed_this_simulation(self, sim_idx): + """Reseed the three models from this index's own child of the root. + + Per index rather than per worker, which is what makes a simulation's + inputs the same however the run was split up. The child is split three + ways so the environment, rocket and flight draw independently instead + of sharing one stream. + """ + child = _seed_of_simulation(self.__root_state, sim_idx) + environment, rocket, flight = child.spawn(3) + self.environment._set_stochastic(_seed_sequence_to_int(environment)) + self.rocket._set_stochastic(_seed_sequence_to_int(rocket)) + self.flight._set_stochastic(_seed_sequence_to_int(flight)) + def __run_single_simulation(self): """Runs a single simulation and returns the inputs and outputs. diff --git a/tests/unit/simulation/test_monte_carlo_seeding.py b/tests/unit/simulation/test_monte_carlo_seeding.py new file mode 100644 index 000000000..b820c5152 --- /dev/null +++ b/tests/unit/simulation/test_monte_carlo_seeding.py @@ -0,0 +1,157 @@ +import json +import os + +import numpy as np +import pytest + +from rocketpy.simulation.monte_carlo import ( + MonteCarlo, + _root_seed_sequence, + _root_state_of, + _seed_of_simulation, +) + + +def _sampled_inputs(analysis): + with open(analysis.input_file, "r", encoding="utf-8") as written: + rows = [json.loads(line) for line in written if line.strip()] + return {row["index"]: row for row in rows} + + +def _a_run(tmp_path, stem, models, *, parallel=False, workers=None, seed=None, count=4): + environment, rocket, flight = models + analysis = MonteCarlo( + filename=str(tmp_path / stem), + environment=environment, + rocket=rocket, + flight=flight, + ) + analysis.simulate( + number_of_simulations=count, + append=False, + parallel=parallel, + n_workers=workers, + random_seed=seed, + ) + return analysis + + +@pytest.fixture(name="models") +def _models(stochastic_environment, stochastic_calisto, stochastic_flight): + return stochastic_environment, stochastic_calisto, stochastic_flight + + +# --------------------------------------------------------------- the derivation + + +@pytest.mark.parametrize("seed", [42, None, [1, 2, 3], np.random.SeedSequence(7)]) +def test_a_simulation_gets_the_child_spawn_would_have_given_it(seed): + # The whole point of deriving one index directly: it has to be the same + # child, bit for bit, as spawning every index before it would produce. + root = _root_seed_sequence(seed) + state = _root_state_of(root) + spawned = np.random.SeedSequence(**root.state).spawn(6) + + for index, expected in enumerate(spawned): + assert np.array_equal( + expected.generate_state(4), + _seed_of_simulation(state, index).generate_state(4), + ) + + +def test_deriving_an_index_costs_nothing_for_the_ones_before_it(): + state = _root_state_of(_root_seed_sequence(42)) + + far = _seed_of_simulation(state, 1_000_000) + + assert far.spawn_key == (1_000_000,) + + +def test_two_indices_do_not_share_a_stream(): + state = _root_state_of(_root_seed_sequence(42)) + + first = _seed_of_simulation(state, 0).generate_state(4) + second = _seed_of_simulation(state, 1).generate_state(4) + + assert not np.array_equal(first, second) + + +# ------------------------------------------------------------------- the root + + +def test_a_caller_seed_sequence_is_not_consumed(): + given = np.random.SeedSequence(42) + + _root_seed_sequence(given).spawn(5) + + assert given.n_children_spawned == 0 + + +def test_a_caller_cannot_move_the_run_by_editing_the_list_it_passed(): + # SeedSequence keeps a sequence entropy by reference, so without a copy of + # its own the run would follow whatever the caller did to that list next. + given = [1, 2, 3] + state = _root_state_of(_root_seed_sequence(given)) + + before = _seed_of_simulation(state, 0).generate_state(4) + given[0] = 999 + + assert np.array_equal(_seed_of_simulation(state, 0).generate_state(4), before) + + +@pytest.mark.parametrize("given", [np.random.default_rng(42), np.random.PCG64(42)]) +def test_a_generator_is_refused_rather_than_read(given): + # Using a consume-on-use object as an immutable seed cannot mean what it + # says, so it is refused instead of quietly meaning something else. + with pytest.raises(TypeError, match="random_seed must be"): + _root_seed_sequence(given) + + +def test_an_unusable_seed_costs_the_previous_run_nothing(models, tmp_path): + analysis = _a_run(tmp_path, "study", models, seed=42, count=2) + kept = _sampled_inputs(analysis) + + with pytest.raises(TypeError): + analysis.simulate(2, append=False, random_seed=np.random.default_rng(1)) + + assert _sampled_inputs(analysis) == kept + + +# ------------------------------------------------------------- what a run gives + + +def test_one_seed_gives_one_set_of_inputs(models, tmp_path): + first = _a_run(tmp_path, "first", models, seed=42) + again = _a_run(tmp_path, "again", models, seed=42) + + assert _sampled_inputs(first) == _sampled_inputs(again) + + +def test_another_seed_gives_another_set(models, tmp_path): + # The control. Without this the test above passes on a run that ignores + # the seed entirely. + first = _a_run(tmp_path, "first", models, seed=42) + other = _a_run(tmp_path, "other", models, seed=7) + + assert _sampled_inputs(first) != _sampled_inputs(other) + + +@pytest.mark.skipif(os.cpu_count() < 4, reason="needs four workers to be four") +def test_an_index_gets_the_same_inputs_however_the_run_was_split(models, tmp_path): + # This is the guarantee. Splitting the work differently must not change + # what any one simulation drew. + serial = _a_run(tmp_path, "serial", models, seed=42) + two = _a_run(tmp_path, "two", models, parallel=True, workers=2, seed=42) + four = _a_run(tmp_path, "four", models, parallel=True, workers=4, seed=42) + + assert _sampled_inputs(serial) == _sampled_inputs(two) + assert _sampled_inputs(serial) == _sampled_inputs(four) + + +def test_an_appended_run_carries_the_same_stream_on(models, tmp_path): + whole = _a_run(tmp_path, "whole", models, seed=42, count=4) + + part = _a_run(tmp_path, "part", models, seed=42, count=2) + part.simulate(4, append=True, random_seed=42) + + assert _sampled_inputs(part) == _sampled_inputs(whole) From 34a0a62a5c1d079dd726c23dbc52c3fd4121469c Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Tue, 18 Aug 2026 23:36:50 +0800 Subject: [PATCH 4/6] DOC: say where a Monte Carlo run's seed goes simulate() gained random_seed with no entry in its own Parameters block, and the stochastic page hands users to the MonteCarlo class without saying that a run is fixed there rather than on the models. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- docs/user/stochastic.rst | 8 ++++++++ rocketpy/simulation/monte_carlo.py | 14 ++++++++++++++ 2 files changed, 22 insertions(+) diff --git a/docs/user/stochastic.rst b/docs/user/stochastic.rst index 6e3376236..1fd8c2f56 100644 --- a/docs/user/stochastic.rst +++ b/docs/user/stochastic.rst @@ -289,3 +289,11 @@ better reflecting the inherent uncertainties in rocketry. .. note:: See the ``MonteCarlo`` class documentation for more information on how to run \ Monte Carlo simulations with stochastic objects. + +.. note:: + A whole run is fixed by ``MonteCarlo.simulate(random_seed=...)`` rather than + by seeding these models yourself. Each simulation takes its seed from its + own index, so simulation 7 draws the same inputs whether the run was serial + or split over any number of workers, and appending with the same seed + carries the same stream on. Without it a run draws fresh entropy and + reproduces nothing. diff --git a/rocketpy/simulation/monte_carlo.py b/rocketpy/simulation/monte_carlo.py index 7c4b86a34..44f019119 100644 --- a/rocketpy/simulation/monte_carlo.py +++ b/rocketpy/simulation/monte_carlo.py @@ -339,6 +339,20 @@ def simulate( number of workers will be equal to the number of CPUs available. A minimum of 2 workers is required for parallel mode. Default is None. + random_seed : int, sequence of int or numpy.random.SeedSequence, optional + Fixes what every simulation draws. Simulation ``i`` takes the same + inputs whichever way the run was split up, so serial and parallel + results agree and the number of workers does not reach the + sampling. Keyword-only. Default is None, which draws fresh entropy + and reproduces nothing. + + Appending continues the same stream when the same seed is given + again, since an index maps to a seed and to nothing else. Nothing + here records the seed, so nothing here can tell you that a later + append was given the same one; that is #1075. + + A ``Generator`` or ``BitGenerator`` is refused rather than read. + Pass the seed it was built from. kwargs : dict Custom arguments for simulation export of the ``inputs`` file. Options are: From 7b8549950a10ed5f8e3e6c93ea77c4dbdf01fb54 Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Tue, 25 Aug 2026 22:28:50 +0800 Subject: [PATCH 5/6] TST: compare what a simulation drew, not which process wrote it RocketPyEncoder records hash(obj) beside every serialized Function, and that is the object's identity in the process that wrote the row. Two runs sharing memory agree on it and two that do not, differ, so the comparison was answering a question about the start method rather than about the seed. Windows uses spawn, and both its legs failed on exactly those fields. Reproduced on Linux with set_start_method("spawn"). Every hash is dropped at any depth before comparing now. This does not make the split-independence test pass under spawn: with the identity gone it still differs on power_off_drag and power_on_drag, which is a separate and so far unexplained difference. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- .../simulation/test_monte_carlo_seeding.py | 22 ++++++++++++++++++- 1 file changed, 21 insertions(+), 1 deletion(-) diff --git a/tests/unit/simulation/test_monte_carlo_seeding.py b/tests/unit/simulation/test_monte_carlo_seeding.py index b820c5152..b83818023 100644 --- a/tests/unit/simulation/test_monte_carlo_seeding.py +++ b/tests/unit/simulation/test_monte_carlo_seeding.py @@ -12,10 +12,30 @@ ) +def _without_object_identity(value): + """The same record with every ``hash`` dropped, at any depth. + + ``RocketPyEncoder`` records ``hash(obj)`` beside a serialized ``Function``, + and that is the object's identity in the process that wrote it, not + anything that was drawn. It agrees between two runs that share memory and + differs between two that do not, which is a property of the start method + rather than of the seed. + """ + if isinstance(value, dict): + return { + key: _without_object_identity(item) + for key, item in value.items() + if key != "hash" + } + if isinstance(value, list): + return [_without_object_identity(item) for item in value] + return value + + def _sampled_inputs(analysis): with open(analysis.input_file, "r", encoding="utf-8") as written: rows = [json.loads(line) for line in written if line.strip()] - return {row["index"]: row for row in rows} + return {row["index"]: _without_object_identity(row) for row in rows} def _a_run(tmp_path, stem, models, *, parallel=False, workers=None, seed=None, count=4): From f9a10fa2829576f878ec1ef9b44c8eb175bbbde1 Mon Sep 17 00:00:00 2001 From: thc1006 <84045975+thc1006@users.noreply.github.com> Date: Wed, 26 Aug 2026 00:13:36 +0800 Subject: [PATCH 6/6] TST: compare what a simulation drew, not what wrote the row down The Windows legs of this branch hung and then reported a mismatch on power_off_drag and power_on_drag, and I read that as the guarantee failing under spawn. It was the test. Two fields a serialized Function carries belong to the writer rather than to the draw. hash is the object's identity in that process. A callable source is its pickle, and the same callable pickles to different bytes in a spawned child: measured on one drag curve, the parent and a forked child agree and a spawned child does not. Both of those drag curves are declared None on the fixture, so nothing varies them and neither field ever carried a draw. Measured with the models seeded by hand, no per-index seeding in the way, one seed: the parent, a forked child and a spawned child all build the same rocket, mass to the last digit. So the guarantee does hold on the start method Windows uses, and there is now a test that says so rather than an assumption. Removing the per-index seeding still turns three of these red, so dropping those two fields has not made the comparison vacuous. Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com> --- .../simulation/test_monte_carlo_seeding.py | 52 +++++++++++++++---- 1 file changed, 41 insertions(+), 11 deletions(-) diff --git a/tests/unit/simulation/test_monte_carlo_seeding.py b/tests/unit/simulation/test_monte_carlo_seeding.py index b83818023..a24829789 100644 --- a/tests/unit/simulation/test_monte_carlo_seeding.py +++ b/tests/unit/simulation/test_monte_carlo_seeding.py @@ -1,6 +1,7 @@ import json import os +import multiprocess import numpy as np import pytest @@ -12,30 +13,32 @@ ) -def _without_object_identity(value): - """The same record with every ``hash`` dropped, at any depth. +def _what_was_drawn(value): + """The same record without the parts that say which process wrote it. - ``RocketPyEncoder`` records ``hash(obj)`` beside a serialized ``Function``, - and that is the object's identity in the process that wrote it, not - anything that was drawn. It agrees between two runs that share memory and - differs between two that do not, which is a property of the start method - rather than of the seed. + Two of the fields a serialized ``Function`` carries are properties of the + writer rather than of the draw. ``hash`` is the object's identity in that + process. A callable ``source`` is its pickle, and the same callable pickles + to different bytes under ``spawn``: measured on one drag curve, the parent + and a forked child agree and a spawned child does not, for a value the + fixture does not vary at all. Everything a run actually draws is numeric + and stays. """ if isinstance(value, dict): return { - key: _without_object_identity(item) + key: _what_was_drawn(item) for key, item in value.items() - if key != "hash" + if key != "hash" and not (key == "source" and isinstance(item, str)) } if isinstance(value, list): - return [_without_object_identity(item) for item in value] + return [_what_was_drawn(item) for item in value] return value def _sampled_inputs(analysis): with open(analysis.input_file, "r", encoding="utf-8") as written: rows = [json.loads(line) for line in written if line.strip()] - return {row["index"]: _without_object_identity(row) for row in rows} + return {row["index"]: _what_was_drawn(row) for row in rows} def _a_run(tmp_path, stem, models, *, parallel=False, workers=None, seed=None, count=4): @@ -175,3 +178,30 @@ def test_an_appended_run_carries_the_same_stream_on(models, tmp_path): part.simulate(4, append=True, random_seed=42) assert _sampled_inputs(part) == _sampled_inputs(whole) + + +@pytest.fixture(name="spawned_workers") +def _spawned_workers(): + """Start workers the way Windows does, wherever the test happens to run.""" + was = multiprocess.get_start_method() + multiprocess.set_start_method("spawn", force=True) + yield + multiprocess.set_start_method(was, force=True) + + +@pytest.mark.usefixtures("spawned_workers") +@pytest.mark.skipif(os.cpu_count() < 4, reason="needs four workers to be four") +def test_an_index_keeps_its_inputs_when_the_workers_are_spawned(models, tmp_path): + """The same guarantee on the start method Windows uses. + + A spawned child rebuilds the models by unpickling rather than inheriting + them, so this is a different question from the one above and was worth + asking separately: the first version of these tests compared serialized + callables, which differ between processes for a value nothing varies. + """ + serial = _a_run(tmp_path, "serial", models, seed=42) + two = _a_run(tmp_path, "two", models, parallel=True, workers=2, seed=42) + four = _a_run(tmp_path, "four", models, parallel=True, workers=4, seed=42) + + assert _sampled_inputs(serial) == _sampled_inputs(two) + assert _sampled_inputs(serial) == _sampled_inputs(four)