Fix actor garbage collection by breaking cyclic references (#1064)

* Fix bug in wait_for_pid_to_exit, add test for actor deletion.

* Fix actor garbage collection by breaking cyclic references

* Add test for calling actor method immediately after actor creation.

* Fix bug, must dispatch tasks when workers are killed.

* Fix python test

* Fix cyclic reference problem by creating ActorMethod objects on the fly.

* Try simply increasing the time allowed for many_drivers_test.py.

test/actor_test.py

@@ -5,10 +5,12 @@ from __future__ import print_function
 import collections
 import random
 import numpy as np
+import os
 import time
 import unittest
 
 import ray
+import ray.test.test_utils
 
 
 class ActorAPI(unittest.TestCase):
@@ -279,6 +281,40 @@ class ActorMethods(unittest.TestCase):
 
         ray.worker.cleanup()
 
+    def testActorDeletion(self):
+        ray.init(num_workers=0)
+
+        # Make sure that when an actor handles goes out of scope, the actor
+        # destructor is called.
+
+        @ray.remote
+        class Actor(object):
+            def getpid(self):
+                return os.getpid()
+
+        a = Actor.remote()
+        pid = ray.get(a.getpid.remote())
+        a = None
+        ray.test.test_utils.wait_for_pid_to_exit(pid)
+
+        actors = [Actor.remote() for _ in range(10)]
+        pids = ray.get([a.getpid.remote() for a in actors])
+        a = None
+        actors = None
+        [ray.test.test_utils.wait_for_pid_to_exit(pid) for pid in pids]
+
+        @ray.remote
+        class Actor(object):
+            def method(self):
+                return 1
+
+        # Make sure that if we create an actor and call a method on it
+        # immediately, the actor doesn't get killed before the method is
+        # called.
+        self.assertEqual(ray.get(Actor.remote().method.remote()), 1)
+
+        ray.worker.cleanup()
+
     def testActorState(self):
         ray.init()
 

test/jenkins_tests/multi_node_tests/many_drivers_test.py

@@ -30,10 +30,10 @@ class Actor1(object):
 
 
 def driver(redis_address, driver_index):
-    """The script for driver 0.
+    """The script for all drivers.
 
-    This driver should create five actors that each use one GPU and some actors
-    that use no GPUs. After a while, it should exit.
+    This driver should create five actors that each use one GPU. After a while,
+    it should exit.
     """
     ray.init(redis_address=redis_address)
 
@@ -44,7 +44,7 @@ def driver(redis_address, driver_index):
     for i in range(driver_index - max_concurrent_drivers + 1):
         _wait_for_event("DRIVER_{}_DONE".format(i), redis_address)
 
-    def try_to_create_actor(actor_class, timeout=100):
+    def try_to_create_actor(actor_class, timeout=500):
         # Try to create an actor, but allow failures while we wait for the
         # monitor to release the resources for the removed drivers.
         start_time = time.time()