Make more tests run on Windows (#8446)

* Remove worker Wait() call due to SIGCHLD being ignored

* Port _pid_alive to Windows

* Show PID as well as TID in glog

* Update TensorFlow version for Python 3.8 on Windows

* Handle missing Pillow on Windows

* Work around dm-tree PermissionError on Windows

* Fix some lint errors on Windows with Python 3.8

* Simplify torch requirements

* Quiet git clean

* Handle finalizer issues

* Exit with the signal number

* Get rid of wget

* Fix some Windows compatibility issues with tests

Co-authored-by: Mehrdad <noreply@github.com>

python/ray/monitor.py

@@ -53,7 +53,12 @@ class Monitor:
     def __del__(self):
         """Destruct the monitor object."""
         # We close the pubsub client to avoid leaking file descriptors.
-        self.primary_subscribe_client.close()
+        try:
+            primary_subscribe_client = self.primary_subscribe_client
+        except AttributeError:
+            primary_subscribe_client = None
+        if primary_subscribe_client is not None:
+            primary_subscribe_client.close()
 
     def subscribe(self, channel):
         """Subscribe to the given channel on the primary Redis shard.

python/ray/test_utils.py

@@ -1,4 +1,5 @@
 import asyncio
+import errno
 import json
 import fnmatch
 import os
@@ -12,6 +13,9 @@ import ray
 
 import psutil  # We must import psutil after ray because we bundle it with ray.
 
+if sys.platform == "win32":
+    import _winapi
+
 
 class RayTestTimeoutException(Exception):
     """Exception used to identify timeouts from test utilities."""
@@ -27,11 +31,24 @@ def _pid_alive(pid):
     Returns:
         This returns false if the process is dead. Otherwise, it returns true.
     """
+    no_such_process = errno.EINVAL if sys.platform == "win32" else errno.ESRCH
+    alive = True
     try:
-        os.kill(pid, 0)
-        return True
-    except OSError:
-        return False
+        if sys.platform == "win32":
+            SYNCHRONIZE = 0x00100000  # access mask defined in <winnt.h>
+            handle = _winapi.OpenProcess(SYNCHRONIZE, False, pid)
+            try:
+                alive = (_winapi.WaitForSingleObject(handle, 0) !=
+                         _winapi.WAIT_OBJECT_0)
+            finally:
+                _winapi.CloseHandle(handle)
+        else:
+            os.kill(pid, 0)
+    except OSError as ex:
+        if ex.errno != no_such_process:
+            raise
+        alive = False
+    return alive
 
 
 def wait_for_pid_to_exit(pid, timeout=20):

python/ray/tests/test_actor.py

@@ -358,7 +358,7 @@ def test_decorator_args(ray_start_regular):
     with pytest.raises(Exception):
 
         @ray.remote(invalid_kwarg=0)  # noqa: F811
-        class Actor:
+        class Actor:  # noqa: F811
             def __init__(self):
                 pass
 
@@ -366,25 +366,25 @@ def test_decorator_args(ray_start_regular):
     with pytest.raises(Exception):
 
         @ray.remote(num_cpus=0, invalid_kwarg=0)  # noqa: F811
-        class Actor:
+        class Actor:  # noqa: F811
             def __init__(self):
                 pass
 
     # This is a valid way of using the decorator.
     @ray.remote(num_cpus=1)  # noqa: F811
-    class Actor:
+    class Actor:  # noqa: F811
         def __init__(self):
             pass
 
     # This is a valid way of using the decorator.
     @ray.remote(num_gpus=1)  # noqa: F811
-    class Actor:
+    class Actor:  # noqa: F811
         def __init__(self):
             pass
 
     # This is a valid way of using the decorator.
     @ray.remote(num_cpus=1, num_gpus=1)  # noqa: F811
-    class Actor:
+    class Actor:  # noqa: F811
         def __init__(self):
             pass
 

python/ray/tests/test_actor_failures.py

@@ -15,6 +15,8 @@ from ray.test_utils import (relevant_errors, wait_for_condition,
                             wait_for_errors, wait_for_pid_to_exit,
                             generate_internal_config_map)
 
+SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
+
 
 @pytest.fixture
 def ray_checkpointable_actor_cls(request):
@@ -153,7 +155,7 @@ def test_actor_restart():
     pid = ray.get(actor.get_pid.remote())
     results = [actor.increase.remote() for _ in range(100)]
     # Kill actor process, while the above task is still being executed.
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     # Make sure that all tasks were executed in order before the actor's death.
     res = results.pop(0)
     i = 1
@@ -194,7 +196,7 @@ def test_actor_restart():
     # kill actor process one more time.
     results = [actor.increase.remote() for _ in range(100)]
     pid = ray.get(actor.get_pid.remote())
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     # The actor has exceeded max restarts, and this task should fail.
     with pytest.raises(ray.exceptions.RayActorError):
         ray.get(actor.increase.remote())
@@ -235,7 +237,7 @@ def test_actor_restart_with_retry():
     pid = ray.get(actor.get_pid.remote())
     results = [actor.increase.remote() for _ in range(100)]
     # Kill actor process, while the above task is still being executed.
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     # Check that none of the tasks failed and the actor is restarted.
     seq = list(range(1, 101))
     results = ray.get(results)
@@ -255,7 +257,7 @@ def test_actor_restart_with_retry():
     # kill actor process one more time.
     results = [actor.increase.remote() for _ in range(100)]
     pid = ray.get(actor.get_pid.remote())
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     # The actor has exceeded max restarts, and this task should fail.
     with pytest.raises(ray.exceptions.RayActorError):
         ray.get(actor.increase.remote())
@@ -351,7 +353,7 @@ def test_actor_restart_without_task(ray_start_regular):
     pid = ray.get(actor.get_pid.remote())
 
     p = probe.remote()
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     ray.get(p)
     assert wait_for_condition(lambda: not actor_resource_available())
 
@@ -507,7 +509,7 @@ def test_multiple_actor_restart(ray_start_cluster_head):
 def kill_actor(actor):
     """A helper function that kills an actor process."""
     pid = ray.get(actor.get_pid.remote())
-    os.kill(pid, signal.SIGKILL)
+    os.kill(pid, SIGKILL)
     wait_for_pid_to_exit(pid)
 
 

python/ray/tests/test_advanced.py

@@ -334,19 +334,19 @@ def test_identical_function_names(ray_start_regular):
         return 1
 
     @ray.remote  # noqa: F811
-    def g():
+    def g():  # noqa: F811
         return 2
 
     @ray.remote  # noqa: F811
-    def g():
+    def g():  # noqa: F811
         return 3
 
     @ray.remote  # noqa: F811
-    def g():
+    def g():  # noqa: F811
         return 4
 
     @ray.remote  # noqa: F811
-    def g():
+    def g():  # noqa: F811
         return 5
 
     result_values = ray.get([g.remote() for _ in range(num_calls)])

python/ray/tests/test_autoscaler.py

@@ -1,3 +1,4 @@
+import os
 import shutil
 import tempfile
 import threading
@@ -165,7 +166,7 @@ SMALL_CLUSTER = {
     },
     "auth": {
         "ssh_user": "ubuntu",
-        "ssh_private_key": "/dev/null",
+        "ssh_private_key": os.devnull,
     },
     "head_node": {
         "TestProp": 1,
@@ -310,7 +311,7 @@ class AutoscalingTest(unittest.TestCase):
         return path
 
     def testInvalidConfig(self):
-        invalid_config = "/dev/null"
+        invalid_config = os.devnull
         with pytest.raises(ValueError):
             StandardAutoscaler(
                 invalid_config, LoadMetrics(), update_interval_s=0)

python/ray/tests/test_component_failures.py

@@ -1,5 +1,5 @@
 import os
-from signal import SIGKILL
+import signal
 import sys
 import time
 
@@ -8,6 +8,8 @@ import pytest
 import ray
 from ray.test_utils import run_string_as_driver_nonblocking, SignalActor
 
+SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
+
 
 # This test checks that when a worker dies in the middle of a get, the plasma
 # store and raylet will not die.

python/ray/tests/test_component_failures_2.py

@@ -11,6 +11,8 @@ import ray.ray_constants as ray_constants
 from ray.cluster_utils import Cluster
 from ray.test_utils import RayTestTimeoutException
 
+SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
+
 
 @pytest.fixture(params=[(1, 4), (4, 4)])
 def ray_start_workers_separate_multinode(request):
@@ -62,7 +64,7 @@ def test_worker_failed(ray_start_workers_separate_multinode):
     time.sleep(0.1)
     # Kill the workers as the tasks execute.
     for pid in pids:
-        os.kill(pid, signal.SIGKILL)
+        os.kill(pid, SIGKILL)
         time.sleep(0.1)
     # Make sure that we either get the object or we get an appropriate
     # exception.

python/ray/tests/test_multinode_failures.py

@@ -11,6 +11,8 @@ import ray.ray_constants as ray_constants
 from ray.cluster_utils import Cluster
 from ray.test_utils import RayTestTimeoutException
 
+SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
+
 
 @pytest.fixture(params=[(1, 4), (4, 4)])
 def ray_start_workers_separate_multinode(request):
@@ -62,7 +64,7 @@ def test_worker_failed(ray_start_workers_separate_multinode):
     time.sleep(0.1)
     # Kill the workers as the tasks execute.
     for pid in pids:
-        os.kill(pid, signal.SIGKILL)
+        os.kill(pid, SIGKILL)
         time.sleep(0.1)
     # Make sure that we either get the object or we get an appropriate
     # exception.

python/ray/tests/test_reference_counting_2.py

@@ -3,6 +3,7 @@ import json
 import logging
 import os
 import signal
+import sys
 
 import numpy as np
 
@@ -12,6 +13,8 @@ import ray
 import ray.cluster_utils
 from ray.test_utils import SignalActor, put_object, wait_for_condition
 
+SIGKILL = signal.SIGKILL if sys.platform != "win32" else signal.SIGTERM
+
 logger = logging.getLogger(__name__)
 
 
@@ -269,7 +272,7 @@ def test_recursively_return_borrowed_object_id(one_worker_100MiB, use_ray_put,
     _fill_object_store_and_get(final_oid_bytes)
 
     if failure:
-        os.kill(owner_pid, signal.SIGKILL)
+        os.kill(owner_pid, SIGKILL)
     else:
         # Remove all references.
         del head_oid

python/ray/worker.py

@@ -9,7 +9,6 @@ import json
 import logging
 import os
 import redis
-import signal
 from six.moves import queue
 import sys
 import threading
@@ -900,7 +899,7 @@ atexit.register(shutdown, True)
 
 # TODO(edoakes): this should only be set in the driver.
 def sigterm_handler(signum, frame):
-    sys.exit(signal.SIGTERM)
+    sys.exit(signum)
 
 
 try:
@@ -1366,7 +1365,12 @@ def disconnect(exiting_interpreter=False):
     worker.node = None  # Disconnect the worker from the node.
     worker.cached_functions_to_run = []
     worker.serialization_context_map.clear()
-    ray.actor.ActorClassMethodMetadata.reset_cache()
+    try:
+        ray_actor = ray.actor
+    except AttributeError:
+        ray_actor = None  # This can occur during program termination
+    if ray_actor is not None:
+        ray_actor.ActorClassMethodMetadata.reset_cache()
 
 
 @contextmanager