Shortcuts on this page
r m x toggle line displays
j k next/prev highlighted chunk
0 (zero) top of page
1 (one) first highlighted chunk
1# This file is part of Hypothesis, which may be found at
2# https://github.com/HypothesisWorks/hypothesis/
3#
4# Copyright the Hypothesis Authors.
5# Individual contributors are listed in AUTHORS.rst and the git log.
6#
7# This Source Code Form is subject to the terms of the Mozilla Public License,
8# v. 2.0. If a copy of the MPL was not distributed with this file, You can
9# obtain one at https://mozilla.org/MPL/2.0/.
11import contextlib
12import gc
13import random
14import sys
15import warnings
16from collections.abc import Generator, Hashable
17from itertools import count
18from random import Random
19from typing import TYPE_CHECKING, Any, Callable, Optional
20from weakref import WeakValueDictionary
22import hypothesis.core
23from hypothesis.errors import HypothesisWarning, InvalidArgument
24from hypothesis.internal.compat import FREE_THREADED_CPYTHON, GRAALPY, PYPY
26if TYPE_CHECKING:
27 from typing import Protocol
29 # we can't use this at runtime until from_type supports
30 # protocols -- breaks ghostwriter tests
31 class RandomLike(Protocol):
32 def seed(self, *args: Any, **kwargs: Any) -> Any: ...
33 def getstate(self, *args: Any, **kwargs: Any) -> Any: ...
34 def setstate(self, *args: Any, **kwargs: Any) -> Any: ...
36else: # pragma: no cover
37 RandomLike = random.Random
39_RKEY = count()
40_global_random_rkey = next(_RKEY)
41# This is effectively a WeakSet, which allows us to associate the saved states
42# with their respective Random instances even as new ones are registered and old
43# ones go out of scope and get garbage collected. Keys are ascending integers.
44RANDOMS_TO_MANAGE: WeakValueDictionary[int, RandomLike] = WeakValueDictionary(
45 {_global_random_rkey: random}
46)
49class NumpyRandomWrapper:
50 def __init__(self) -> None:
51 assert "numpy" in sys.modules
52 # This class provides a shim that matches the numpy to stdlib random,
53 # and lets us avoid importing Numpy until it's already in use.
54 import numpy.random
56 self.seed = numpy.random.seed
57 self.getstate = numpy.random.get_state
58 self.setstate = numpy.random.set_state
61NP_RANDOM: Optional[RandomLike] = None
64if not (PYPY or GRAALPY):
66 def _get_platform_base_refcount(r: Any) -> int:
67 return sys.getrefcount(r)
69 # Determine the number of refcounts created by function scope for
70 # the given platform / version of Python.
71 _PLATFORM_REF_COUNT = _get_platform_base_refcount(object())
72else: # pragma: no cover
73 # PYPY and GRAALPY don't have `sys.getrefcount`
74 _PLATFORM_REF_COUNT = -1
77def register_random(r: RandomLike) -> None:
78 """Register (a weakref to) the given Random-like instance for management by
79 Hypothesis.
81 You can pass instances of structural subtypes of ``random.Random``
82 (i.e., objects with seed, getstate, and setstate methods) to
83 ``register_random(r)`` to have their states seeded and restored in the same
84 way as the global PRNGs from the ``random`` and ``numpy.random`` modules.
86 All global PRNGs, from e.g. simulation or scheduling frameworks, should
87 be registered to prevent flaky tests. Hypothesis will ensure that the
88 PRNG state is consistent for all test runs, always seeding them to zero and
89 restoring the previous state after the test, or, reproducibly varied if you
90 choose to use the :func:`~hypothesis.strategies.random_module` strategy.
92 ``register_random`` only makes `weakrefs
93 <https://docs.python.org/3/library/weakref.html#module-weakref>`_ to ``r``,
94 thus ``r`` will only be managed by Hypothesis as long as it has active
95 references elsewhere at runtime. The pattern ``register_random(MyRandom())``
96 will raise a ``ReferenceError`` to help protect users from this issue.
97 This check does not occur for the PyPy interpreter. See the following example for
98 an illustration of this issue
100 .. code-block:: python
103 def my_BROKEN_hook():
104 r = MyRandomLike()
106 # `r` will be garbage collected after the hook resolved
107 # and Hypothesis will 'forget' that it was registered
108 register_random(r) # Hypothesis will emit a warning
111 rng = MyRandomLike()
114 def my_WORKING_hook():
115 register_random(rng)
116 """
117 if not (hasattr(r, "seed") and hasattr(r, "getstate") and hasattr(r, "setstate")):
118 raise InvalidArgument(f"{r=} does not have all the required methods")
120 if r in [
121 random
122 for ref in RANDOMS_TO_MANAGE.data.copy().values() # type: ignore
123 if (random := ref()) is not None
124 ]:
125 return
127 if not (PYPY or GRAALPY): # pragma: no branch
128 # PYPY and GRAALPY do not have `sys.getrefcount`.
129 gc.collect()
130 if not gc.get_referrers(r):
131 if sys.getrefcount(r) <= _PLATFORM_REF_COUNT:
132 raise ReferenceError(
133 f"`register_random` was passed `r={r}` which will be "
134 "garbage collected immediately after `register_random` creates a "
135 "weakref to it. This will prevent Hypothesis from managing this "
136 "PRNG. See the docs for `register_random` for more "
137 "details."
138 )
139 elif not FREE_THREADED_CPYTHON: # pragma: no branch
140 # On CPython, check for the free-threaded build because
141 # gc.get_referrers() ignores objects with immortal refcounts
142 # and objects are immortalized in the Python 3.13
143 # free-threading implementation at runtime.
145 warnings.warn(
146 "It looks like `register_random` was passed an object that could "
147 "be garbage collected immediately after `register_random` creates "
148 "a weakref to it. This will prevent Hypothesis from managing this "
149 "PRNG. See the docs for `register_random` for more details.",
150 HypothesisWarning,
151 stacklevel=2,
152 )
154 RANDOMS_TO_MANAGE[next(_RKEY)] = r
157# Used to make the warning issued by `deprecate_random_in_strategy` thread-safe,
158# as well as to avoid warning on uses of st.randoms().
159# Store just the hash to reduce memory consumption. This is an underapproximation
160# of membership (distinct items might have the same hash), which is fine for the
161# warning, as it results in missed alarms, not false alarms.
162_known_random_state_hashes: set[Any] = set()
165def get_seeder_and_restorer(
166 seed: Hashable = 0,
167) -> tuple[Callable[[], None], Callable[[], None]]:
168 """Return a pair of functions which respectively seed all and restore
169 the state of all registered PRNGs.
171 This is used by the core engine via `deterministic_PRNG`, and by users
172 via `register_random`. We support registration of additional random.Random
173 instances (or other objects with seed, getstate, and setstate methods)
174 to force determinism on simulation or scheduling frameworks which avoid
175 using the global random state. See e.g. #1709.
176 """
177 assert isinstance(seed, int)
178 assert 0 <= seed < 2**32
179 states: dict[int, object] = {}
181 if "numpy" in sys.modules:
182 global NP_RANDOM
183 if NP_RANDOM is None:
184 # Protect this from garbage-collection by adding it to global scope
185 NP_RANDOM = RANDOMS_TO_MANAGE[next(_RKEY)] = NumpyRandomWrapper()
187 def seed_all() -> None:
188 assert not states
189 # access .data.copy().items() instead of .items() to avoid a "dictionary
190 # changed size during iteration" error under multithreading.
191 #
192 # I initially expected this to be fixed by
193 # https://github.com/python/cpython/commit/96d37dbcd23e65a7a57819aeced9034296ef747e,
194 # but I believe that is addressing the size change from weakrefs expiring
195 # during gc, not from the user adding new elements to the dict.
196 #
197 # Since we're accessing .data, we have to manually handle checking for
198 # expired ref instances during iteration. Normally WeakValueDictionary
199 # handles this for us.
200 #
201 # This command reproduces at time of writing:
202 # pytest hypothesis-python/tests/ -k test_intervals_are_equivalent_to_their_lists
203 # --parallel-threads 2
204 for k, ref in RANDOMS_TO_MANAGE.data.copy().items(): # type: ignore
205 r = ref()
206 if r is None:
207 # ie the random instance has been gc'd
208 continue # pragma: no cover
209 states[k] = r.getstate()
210 if k == _global_random_rkey:
211 # r.seed sets the random's state. We want to add that state to
212 # _known_random_states before calling r.seed, in case a thread
213 # switch occurs between the two. To figure out the seed -> state
214 # mapping, set the seed on a dummy random and add that state to
215 # _known_random_state.
216 #
217 # we could use a global dummy random here, but then we'd have to
218 # put a lock around it, and it's not clear to me if that's more
219 # efficient than constructing a new instance each time.
220 dummy_random = Random()
221 dummy_random.seed(seed)
222 _known_random_state_hashes.add(hash(dummy_random.getstate()))
223 # we expect `assert r.getstate() == dummy_random.getstate()` to
224 # hold here, but thread switches means it might not.
226 r.seed(seed)
228 def restore_all() -> None:
229 for k, state in states.items():
230 r = RANDOMS_TO_MANAGE.get(k)
231 if r is None: # i.e., has been garbage-collected
232 continue
234 if k == _global_random_rkey:
235 _known_random_state_hashes.add(hash(state))
236 r.setstate(state)
238 states.clear()
240 return seed_all, restore_all
243@contextlib.contextmanager
244def deterministic_PRNG(seed: int = 0) -> Generator[None, None, None]:
245 """Context manager that handles random.seed without polluting global state.
247 See issue #1255 and PR #1295 for details and motivation - in short,
248 leaving the global pseudo-random number generator (PRNG) seeded is a very
249 bad idea in principle, and breaks all kinds of independence assumptions
250 in practice.
251 """
252 if (
253 hypothesis.core.threadlocal._hypothesis_global_random is None
254 ): # pragma: no cover
255 hypothesis.core.threadlocal._hypothesis_global_random = Random()
256 register_random(hypothesis.core.threadlocal._hypothesis_global_random)
258 seed_all, restore_all = get_seeder_and_restorer(seed)
259 seed_all()
260 try:
261 yield
262 finally:
263 restore_all()
264 # TODO it would be nice to clean up _known_random_state_hashes when no
265 # active deterministic_PRNG contexts remain, to free memory (see similar
266 # logic in StackframeLimiter). But it's a bit annoying to get right, and
267 # likely not a big deal.