Files
ray-project--ray/python/ray/tests/test_reference_counting_2.py
2026-07-13 13:17:40 +08:00

525 lines
16 KiB
Python

"""All tests in this file use a module-scoped fixture to reduce runtime.
If you need a customized Ray instance (e.g., to change system config or env vars),
put the test in `test_reference_counting_standalone.py`.
"""
# coding: utf-8
import copy
import logging
import os
import pickle
import signal
import sys
import time
from typing import Union
import numpy as np
import pytest
import ray
import ray._private.gcs_utils as gcs_utils
from ray._common.test_utils import SignalActor, wait_for_condition
from ray._private.internal_api import memory_summary
from ray._private.test_utils import (
put_object,
wait_for_num_actors,
)
SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
logger = logging.getLogger(__name__)
@pytest.fixture(scope="module")
def one_cpu_100MiB_shared():
config = {
"task_retry_delay_ms": 0,
"object_timeout_milliseconds": 1000,
"automatic_object_spilling_enabled": False,
}
yield ray.init(
num_cpus=1, object_store_memory=100 * 1024 * 1024, _system_config=config
)
ray.shutdown()
def _fill_object_store_and_get(
obj: Union[ray.ObjectRef, bytes],
*,
succeed: bool = True,
object_MiB: float = 20,
num_objects: int = 5,
timeout_s: float = 10.0,
):
for _ in range(num_objects):
ray.put(np.zeros(object_MiB * 1024 * 1024, dtype=np.uint8))
if type(obj) is bytes:
obj = ray.ObjectRef(obj)
if succeed:
wait_for_condition(
lambda: ray._private.worker.global_worker.core_worker.object_exists(obj),
timeout=timeout_s,
)
else:
wait_for_condition(
lambda: not ray._private.worker.global_worker.core_worker.object_exists(
obj
),
timeout=timeout_s,
)
# Test that an object containing object refs within it pins the inner IDs
# recursively and for submitted tasks.
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
@pytest.mark.parametrize(
"use_ray_put,failure", [(False, False), (False, True), (True, False), (True, True)]
)
def test_recursively_nest_ids(one_cpu_100MiB_shared, use_ray_put, failure):
@ray.remote(max_retries=1)
def recursive(ref, signal, max_depth, depth=0):
unwrapped = ray.get(ref[0])
if depth == max_depth:
ray.get(signal.wait.remote())
if failure:
os._exit(0)
return
else:
return recursive.remote(unwrapped, signal, max_depth, depth + 1)
signal = SignalActor.remote()
max_depth = 5
array_oid = put_object(np.zeros(20 * 1024 * 1024, dtype=np.uint8), use_ray_put)
nested_oid = array_oid
for _ in range(max_depth):
nested_oid = ray.put([nested_oid])
head_oid = recursive.remote([nested_oid], signal, max_depth)
# Remove the local reference.
array_oid_bytes = array_oid.binary()
del array_oid, nested_oid
tail_oid = head_oid
for _ in range(max_depth):
tail_oid = ray.get(tail_oid)
# Check that the remote reference pins the object.
_fill_object_store_and_get(array_oid_bytes)
# Fulfill the dependency, causing the tail task to finish.
ray.get(signal.send.remote())
if not failure:
ray.get(tail_oid)
else:
# There is only 1 core, so the same worker will execute all `recursive`
# tasks. Therefore, if we kill the worker during the last task, its
# owner (the worker that executed the second-to-last task) will also
# have died.
with pytest.raises(ray.exceptions.OwnerDiedError):
ray.get(tail_oid)
# Reference should be gone, check that array gets evicted.
_fill_object_store_and_get(array_oid_bytes, succeed=False)
# Test that serialized ObjectRefs returned from remote tasks are pinned until
# they go out of scope on the caller side.
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
@pytest.mark.parametrize(
"use_ray_put,failure", [(False, False), (False, True), (True, False), (True, True)]
)
def test_return_object_ref(one_cpu_100MiB_shared, use_ray_put, failure):
@ray.remote
def return_an_id():
return [put_object(np.zeros(20 * 1024 * 1024, dtype=np.uint8), use_ray_put)]
@ray.remote(max_retries=1)
def exit():
os._exit(0)
outer_oid = return_an_id.remote()
inner_oid_binary = ray.get(outer_oid)[0].binary()
# Check that taking a reference to the inner ID and removing the outer ID
# doesn't unpin the object.
inner_oid = ray.get(outer_oid)[0] # noqa: F841
del outer_oid
_fill_object_store_and_get(inner_oid_binary)
if failure:
# Check that the owner dying unpins the object. This should execute on
# the same worker because there is only one started and the other tasks
# have finished.
with pytest.raises(ray.exceptions.WorkerCrashedError):
ray.get(exit.remote())
else:
# Check that removing the inner ID unpins the object.
del inner_oid
_fill_object_store_and_get(inner_oid_binary, succeed=False)
# Test that serialized ObjectRefs returned from remote tasks are pinned if
# passed into another remote task by the caller.
@pytest.mark.parametrize(
"use_ray_put,failure", [(False, False), (False, True), (True, False), (True, True)]
)
def test_pass_returned_object_ref(one_cpu_100MiB_shared, use_ray_put, failure):
@ray.remote
def return_an_id():
return [put_object(np.zeros(20 * 1024 * 1024, dtype=np.uint8), use_ray_put)]
# TODO(edoakes): this fails with an ActorError with max_retries=1.
@ray.remote(max_retries=0)
def pending(ref, signal):
ray.get(signal.wait.remote())
ray.get(ref[0])
if failure:
os._exit(0)
signal = SignalActor.remote()
outer_oid = return_an_id.remote()
inner_oid_binary = ray.get(outer_oid)[0].binary()
pending_oid = pending.remote([outer_oid], signal)
# Remove the local reference to the returned ID.
del outer_oid
# Check that the inner ID is pinned by the remote task ID and finishing
# the task unpins the object.
ray.get(signal.send.remote())
try:
# Should succeed because inner_oid is pinned if no failure.
ray.get(pending_oid)
assert not failure
except ray.exceptions.WorkerCrashedError:
assert failure
def ref_not_exists():
worker = ray._private.worker.global_worker
inner_oid = ray.ObjectRef(inner_oid_binary)
return not worker.core_worker.object_exists(inner_oid)
wait_for_condition(ref_not_exists)
# Call a recursive chain of tasks that pass a serialized reference that was
# returned by another task to the end of the chain. The reference should still
# exist while the final task in the chain is running and should be removed once
# it finishes.
@pytest.mark.skipif(sys.platform == "win32", reason="Flaky on Windows.")
@pytest.mark.parametrize(
"use_ray_put,failure", [(False, False), (False, True), (True, False), (True, True)]
)
def test_recursively_pass_returned_object_ref(
one_cpu_100MiB_shared, use_ray_put, failure
):
@ray.remote
def return_an_id():
return put_object(np.zeros(20 * 1024 * 1024, dtype=np.uint8), use_ray_put)
@ray.remote(max_retries=1)
def recursive(ref, signal, max_depth, depth=0):
inner_id = ray.get(ref[0])
if depth == max_depth:
ray.get(signal.wait.remote())
if failure:
os._exit(0)
return inner_id
else:
return inner_id, recursive.remote(ref, signal, max_depth, depth + 1)
max_depth = 5
outer_oid = return_an_id.remote()
signal = SignalActor.remote()
head_oid = recursive.remote([outer_oid], signal, max_depth)
# Remove the local reference.
inner_oid = None
outer_oid = head_oid
for i in range(max_depth):
inner_oid, outer_oid = ray.get(outer_oid)
# Check that the remote reference pins the object.
_fill_object_store_and_get(outer_oid, succeed=False)
# Fulfill the dependency, causing the tail task to finish.
ray.get(signal.send.remote())
try:
# Check that the remote reference pins the object.
ray.get(outer_oid)
_fill_object_store_and_get(inner_oid)
assert not failure
except ray.exceptions.OwnerDiedError:
# There is only 1 core, so the same worker will execute all `recursive`
# tasks. Therefore, if we kill the worker during the last task, its
# owner (the worker that executed the second-to-last task) will also
# have died.
assert failure
inner_oid_bytes = inner_oid.binary()
del inner_oid
del head_oid
del outer_oid
# Reference should be gone, check that returned ID gets evicted.
_fill_object_store_and_get(inner_oid_bytes, succeed=False, timeout_s=20)
# Call a recursive chain of tasks. The final task in the chain returns an
# ObjectRef returned by a task that it submitted. Every other task in the chain
# returns the same ObjectRef by calling ray.get() on its submitted task and
# returning the result. The reference should still exist while the driver has a
# reference to the final task's ObjectRef.
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
@pytest.mark.parametrize(
"use_ray_put,failure", [(False, False), (False, True), (True, False), (True, True)]
)
def test_recursively_return_borrowed_object_ref(
one_cpu_100MiB_shared, use_ray_put, failure
):
@ray.remote
def recursive(num_tasks_left):
if num_tasks_left == 0:
return (
put_object(np.zeros(20 * 1024 * 1024, dtype=np.uint8), use_ray_put),
os.getpid(),
)
return ray.get(recursive.remote(num_tasks_left - 1))
max_depth = 5
head_oid = recursive.remote(max_depth)
final_oid, owner_pid = ray.get(head_oid)
final_oid_bytes = final_oid.binary()
# Check that the driver's reference pins the object.
_fill_object_store_and_get(final_oid_bytes)
# Remove the local reference and try it again.
_fill_object_store_and_get(final_oid_bytes)
if failure:
os.kill(owner_pid, SIGKILL)
else:
# Remove all references.
del head_oid
del final_oid
# Reference should be gone, check that returned ID gets evicted.
_fill_object_store_and_get(final_oid_bytes, succeed=False)
if failure:
with pytest.raises(ray.exceptions.OwnerDiedError):
ray.get(final_oid)
@pytest.mark.parametrize("failure", [False, True])
def test_borrowed_id_failure(one_cpu_100MiB_shared, failure):
@ray.remote
class Parent:
def __init__(self):
pass
def pass_ref(self, ref, borrower):
self.ref = ref[0]
ray.get(borrower.receive_ref.remote(ref))
if failure:
sys.exit(-1)
def ping(self):
return
@ray.remote
class Borrower:
def __init__(self):
self.ref = None
def receive_ref(self, ref):
self.ref = ref[0]
def resolve_ref(self):
assert self.ref is not None
if failure:
with pytest.raises(ray.exceptions.ReferenceCountingAssertionError):
ray.get(self.ref)
else:
ray.get(self.ref)
def ping(self):
return
parent = Parent.remote()
borrower = Borrower.remote()
ray.get(borrower.ping.remote())
obj = ray.put(np.zeros(20 * 1024 * 1024, dtype=np.uint8))
if failure:
with pytest.raises(ray.exceptions.RayActorError):
ray.get(parent.pass_ref.remote([obj], borrower))
else:
ray.get(parent.pass_ref.remote([obj], borrower))
obj_bytes = obj.binary()
del obj
_fill_object_store_and_get(obj_bytes, succeed=not failure)
# The borrower should not hang when trying to get the object's value.
ray.get(borrower.resolve_ref.remote())
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
def test_actor_constructor_borrowed_refs(one_cpu_100MiB_shared):
@ray.remote
class Borrower:
def __init__(self, borrowed_refs):
self.borrowed_refs = borrowed_refs
def test(self):
ray.get(self.borrowed_refs)
# Actor is the only one with a ref.
ref = ray.put(np.random.random(1024 * 1024))
b = Borrower.remote([ref])
del ref
# Check that the actor's ref is usable.
for _ in range(3):
ray.get(b.test.remote())
time.sleep(1)
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
def test_deep_nested_refs(one_cpu_100MiB_shared):
@ray.remote
def f(x):
print(f"=> step {x}")
if x > 25:
return x
return f.remote(x + 1)
r = f.remote(1)
i = 0
while isinstance(r, ray.ObjectRef):
print(i, r)
i += 1
r = ray.get(r)
@pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows.")
def test_forward_nested_ref(one_cpu_100MiB_shared):
@ray.remote
def nested_ref():
return ray.put(1)
@ray.remote
def nested_nested_ref():
return nested_ref.remote()
@ray.remote
class Borrower:
def __init__(self):
return
def pass_ref(self, middle_ref):
self.inner_ref = ray.get(middle_ref)
def check_ref(self):
ray.get(self.inner_ref)
@ray.remote
def pass_nested_ref(borrower, outer_ref):
ray.get(borrower.pass_ref.remote(outer_ref[0]))
b = Borrower.remote()
outer_ref = nested_nested_ref.remote()
x = pass_nested_ref.remote(b, [outer_ref])
del outer_ref
ray.get(x)
for _ in range(3):
ray.get(b.check_ref.remote())
time.sleep(1)
def test_out_of_band_actor_handle_deserialization(one_cpu_100MiB_shared):
@ray.remote
class Actor:
def ping(self):
return 1
actor = Actor.remote()
@ray.remote
def func(config):
# deep copy will pickle and unpickle the actor handle.
config = copy.deepcopy(config)
return ray.get(config["actor"].ping.remote())
assert ray.get(func.remote({"actor": actor})) == 1
def test_out_of_band_actor_handle_bypass_reference_counting(one_cpu_100MiB_shared):
@ray.remote
class Actor:
def ping(self):
return 1
actor = Actor.remote()
serialized = pickle.dumps({"actor": actor})
del actor
wait_for_num_actors(1, gcs_utils.ActorTableData.DEAD)
config = pickle.loads(serialized)
with pytest.raises(ray.exceptions.RayActorError):
ray.get(config["actor"].ping.remote())
def test_generators(one_cpu_100MiB_shared):
@ray.remote(num_returns="dynamic")
def remote_generator():
for _ in range(3):
yield np.zeros(10 * 1024 * 1024, dtype=np.uint8)
gen = ray.get(remote_generator.remote())
refs = list(gen)
for r in refs:
_fill_object_store_and_get(r)
# Outer ID out of scope, we should still be able to get the dynamic
# objects.
del gen
for r in refs:
_fill_object_store_and_get(r)
# Inner IDs out of scope.
refs_oids = [r.binary() for r in refs]
del r
del refs
for r_oid in refs_oids:
_fill_object_store_and_get(r_oid, succeed=False)
def test_lineage_leak(one_cpu_100MiB_shared):
@ray.remote
def process(data):
return b"\0" * 100_000_000
data = ray.put(b"\0" * 100_000_000)
ref = process.remote(data)
ray.get(ref)
del data
del ref
def check_usage():
return "Plasma memory usage 0 MiB" in memory_summary(stats_only=True)
wait_for_condition(check_usage)
if __name__ == "__main__":
sys.exit(pytest.main(["-sv", __file__]))