Prototype named actors.#2129

Merged

robertnishihara merged 6 commits intoray-project:masterfrom

heyucongtom:master

May 24, 2018

Contributor

heyucongtom commented May 23, 2018 •

edited

Loading

What do these changes do?

Since it's a total re-factorization, I feel like it may be cleaner to open up a new commit tree for the modification. It refer to the same issue as discussed in the pull-request #2120. (Please feel free to close that one.)

This commit deal with:

Implement the get_actor and register_actor inside the experimental directory.
Add unit-test for single driver environment.
Raise appropriate errors for the operation.

Related issue number

Yucong He added 2 commits

May 23, 2018 13:13


          Clean up and move named_actor implementations to the experimental fol…

2180f0e

…der. Add tests


          clean up

1a86d3f

robertnishihara mentioned this pull request

Named actor #2120

Closed

robertnishihara reviewed

View reviewed changes

Collaborator

robertnishihara left a comment

This looks great! Please make sure the linting passes.

test/actor_test.py

+                      # Test getting f
+                      f2 = ray.experimental.get_actor("f1")
+                      self.assertEqual(f1._actor_id, f2._actor_id)

Collaborator

robertnishihara May 23, 2018

Let's modify the actor to be something like

def method(self):
    self.x += 1
    return self.x

so that here we can do

self.assertEqual(ray.get(f1.method.remote(), 1)
self.assertEqual(ray.get(f2.method.remote(), 2)
self.assertEqual(ray.get(f1.method.remote(), 3)
self.assertEqual(ray.get(f2.method.remote(), 4)

Contributor Author

heyucongtom May 23, 2018

Handled.

test/actor_test.py Outdated

+                      # Test getting an unexist actor
+                      with self.assertRaises(AssertionError):
+                          err = ray.experimental.get_actor("unexisted")

Collaborator

robertnishihara May 23, 2018

unexisted -> nonexistent

test/actor_test.py Outdated

                       # we should also test this from a different driver.
                       ray.get(new_f.method.remote())
+                  def testRegisterAndGetActorHandle(self):

Collaborator

robertnishihara May 23, 2018

can we change this to testRegisterAndGetNamedActors? It'll be easier to find by searching

python/ray/experimental/named_actors.py

+              def _calculate_key_(name):
+                  return b"Actor:" + str.encode(name)
+              def get_actor(name):

Collaborator

robertnishihara May 23, 2018

Please add docstrings for these methods with the format used e.g., in

ray/python/ray/worker.py

Lines 2443 to 2458 in 4584193

    
           def get(object_ids, worker=global_worker): 
        
               """Get a remote object or a list of remote objects from the object store. 
        
               This method blocks until the object corresponding to the object ID is 
        
               available in the local object store. If this object is not in the local 
        
               object store, it will be shipped from an object store that has it (once the 
        
               object has been created). If object_ids is a list, then the objects 
        
               corresponding to each object in the list will be returned. 
        
               Args: 
        
                   object_ids: Object ID of the object to get or a list of object IDs to 
        
                       get. 
        
               Returns: 
        
                   A Python object or a list of Python objects. 
        
               """

python/ray/experimental/named_actors.py Outdated

+                  assert not is_existed, \
+                      "Error: the actor with name={} already exists".format(name)
+                  pickled_state = pickle.dumps(actor_handle)
+                  worker.redis_client.hmset(actor_hash, {name: pickled_state})

Collaborator

robertnishihara May 23, 2018

There is a race condition with this approach. It's possible that another worker could add publish a named actor with the same name between the calls to hexists and hmset. You can avoid this with hsetnx. See https://redis.io/commands/hsetnx.

python/ray/experimental/named_actors.py

+              from __future__ import absolute_import
+              from __future__ import division
+              from __future__ import print_function

Collaborator

robertnishihara May 23, 2018

Let's reserve assertions for checking things that should never actually happen. The errors (in both register_actor and get_actor) can happen if the user passes in the wrong values, so we should raise some sort of exception

Contributor Author

heyucongtom May 23, 2018

Handled.

python/ray/experimental/named_actors.py Outdated

+              def register_actor(name, actor_handle):
+                  worker = ray.worker.get_global_worker()
+                  actor_hash = _calculate_key_(name)
+                  assert type(actor_handle) == ray.actor.ActorHandle, \

Collaborator

robertnishihara May 23, 2018

should raise TypeError

python/ray/experimental/named_actors.py Outdated

+                  assert type(actor_handle) == ray.actor.ActorHandle, \
+                      "Error: you could only store named-actors."
+                  is_existed = worker.redis_client.hexists(actor_hash, name)
+                  assert not is_existed, \

Collaborator

robertnishihara May 23, 2018

should raise ValueError

python/ray/experimental/named_actors.py Outdated

+              def get_actor(name):
+                  worker = ray.worker.get_global_worker()
+                  actor_hash = _calculate_key_(name)
+                  pickled_state = worker.redis_client.hmget(actor_hash, name)

Collaborator

robertnishihara May 23, 2018

You can just use hget and then remove the assertion below.

python/ray/experimental/named_actors.py Outdated

+                  pickled_state = worker.redis_client.hmget(actor_hash, name)
+                  assert len(pickled_state) == 1, \
+                      "Error: Multiple actors under this name."
+                  assert pickled_state[0] is not None, \

Collaborator

robertnishihara May 23, 2018

should raise ValueError

robertnishihara changed the title ~~New commit tree for the named actor~~ Prototype named actors.

Collaborator

robertnishihara commented May 23, 2018

To do linting, you can do (from ray/)

yapf --style .style.yapf -i  python/ray/experimental/__init__.py  python/ray/experimental/named_actors.py test/actor_test.py

also

flake8 python/ray/experimental/__init__.py  python/ray/experimental/named_actors.py test/actor_test.py

robertnishihara reviewed

View reviewed changes

python/ray/experimental/__init__.py Outdated

                   _flush_finished_tasks_unsafe_shard, _flush_evicted_objects_unsafe_shard)
+              from .named_actors import (
+                  get_actor, register_actor
+              )

Collaborator

robertnishihara May 23, 2018

looks like this can all fit on one line without the parentheses

AmplabJenkins commented May 23, 2018

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5597/
Test FAILed.

Yucong He and others added 3 commits

May 23, 2018 16:29


          Modify error type, add docstrings, fix (potential?) bug for pickling …

aefd107

…an actorhandle created by pickling.


          Update __init__.py

e72f278


          Small changes.

a29098a

robertnishihara reviewed

View reviewed changes

python/ray/actor.py

                           "actor_driver_id": self._ray_actor_driver_id.id(),
-                          "previous_actor_handle_id": self._ray_actor_handle_id.id(),
+                          "previous_actor_handle_id": self._ray_actor_handle_id.id()
+                          if self._ray_actor_handle_id else None,

Collaborator

robertnishihara May 24, 2018

Why is this change needed?

Collaborator

robertnishihara May 24, 2018

Hmm.. ok I see that the test failures without it.

robertnishihara approved these changes

View reviewed changes

Collaborator

robertnishihara left a comment

Thanks! I pushed some stylistic changes. I'll merge this once the tests pass.


          Add back check.

40d1b03

AmplabJenkins commented May 24, 2018

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5602/
Test PASSed.

AmplabJenkins commented May 24, 2018

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5606/
Test PASSed.

AmplabJenkins commented May 24, 2018

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5609/
Test PASSed.

AmplabJenkins commented May 24, 2018

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5610/
Test PASSed.

robertnishihara approved these changes

View reviewed changes

robertnishihara merged commit 3509a33 into ray-project:master

Collaborator

robertnishihara commented May 24, 2018

One thing I noticed is that calling get_actor twice on the same name will cause problems (once you try to use the actor). This is probably the same issue as #2115.

Contributor Author

heyucongtom commented May 24, 2018 via email

I just came across this bug thread earlier today. Would take a look at it later.

On Thu, May 24, 2018 at 12:40 AM Robert Nishihara ***@***.***> wrote: One thing I noticed is that calling get_actor twice on the same name will cause problems (once you try to use the actor). This is probably the same issue as #2115 <#2115>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#2129 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGbH8bpCJ3UOImrfIXH-TVnTKyaYIwWIks5t1mPVgaJpZM4ULH2Z> .

alok added a commit to alok/ray that referenced this pull request


          Merge branch 'master' into fix-a3c-torch

e865a09

* master:
  Prototype named actors. (ray-project#2129)
  Update arrow to latest master (ray-project#2100)
  [DataFrame] Speed up dtypes (ray-project#2118)
  do not fetch from dead Plasma Manager (ray-project#2116)
  [DataFrame] Refactor GroupBy Methods and Implement Reindex (ray-project#2101)
  Initial Support for Airspeed Velocity (ray-project#2113)
  Use automatic memory management in Redis modules. (ray-project#1797)
  [DataFrame] Test bugfixes (ray-project#2111)
  [DataFrame] Update initializations of IndexMetadata which use outdated APIs (ray-project#2103)

alok added a commit to alok/ray that referenced this pull request


          Merge branch 'fix-a3c-torch' into trpo

a62fa6e

* fix-a3c-torch:
  Prototype named actors. (ray-project#2129)
  Update arrow to latest master (ray-project#2100)
  [DataFrame] Speed up dtypes (ray-project#2118)
  do not fetch from dead Plasma Manager (ray-project#2116)

llan-ml mentioned this pull request

[rllib] whether support reset_args for the function reset of environments? #2969

Closed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet