From c6f15a00576fdafbc5cb379c890839abe2bf407b Mon Sep 17 00:00:00 2001
From: Eric Liang <ekhliang@gmail.com>
Date: Tue, 19 Mar 2019 09:58:45 -0700
Subject: [PATCH] Revert [rllib] Reserve CPUs for replay actors in apex (#4404)

* Revert "[rllib] Reserve CPUs for replay actors in apex (#4217)"

This reverts commit 2781d74680bba8dd1a490514161d4e779280c893.

* comment
---
 python/ray/rllib/agents/dqn/dqn.py               | 16 ----------------
 .../rllib/optimizers/async_replay_optimizer.py   |  5 +++--
 python/ray/rllib/tests/test_supported_spaces.py  |  2 +-
 3 files changed, 4 insertions(+), 19 deletions(-)

diff --git a/python/ray/rllib/agents/dqn/dqn.py b/python/ray/rllib/agents/dqn/dqn.py
index 9c51e5653..31f7f12cc 100644
--- a/python/ray/rllib/agents/dqn/dqn.py
+++ b/python/ray/rllib/agents/dqn/dqn.py
@@ -12,7 +12,6 @@ from ray.rllib.agents.dqn.dqn_policy_graph import DQNPolicyGraph
 from ray.rllib.evaluation.metrics import collect_metrics
 from ray.rllib.utils.annotations import override
 from ray.rllib.utils.schedules import ConstantSchedule, LinearSchedule
-from ray.tune.trial import Resources
 
 logger = logging.getLogger(__name__)
 
@@ -142,21 +141,6 @@ class DQNAgent(Agent):
     _policy_graph = DQNPolicyGraph
     _optimizer_shared_configs = OPTIMIZER_SHARED_CONFIGS
 
-    @classmethod
-    @override(Agent)
-    def default_resource_request(cls, config):
-        cf = dict(cls._default_config, **config)
-        Agent._validate_config(cf)
-        if cf["optimizer_class"] == "AsyncReplayOptimizer":
-            extra = cf["optimizer"]["num_replay_buffer_shards"]
-        else:
-            extra = 0
-        return Resources(
-            cpu=cf["num_cpus_for_driver"],
-            gpu=cf["num_gpus"],
-            extra_cpu=cf["num_cpus_per_worker"] * cf["num_workers"] + extra,
-            extra_gpu=cf["num_gpus_per_worker"] * cf["num_workers"])
-
     @override(Agent)
     def _init(self):
         self._validate_config()
diff --git a/python/ray/rllib/optimizers/async_replay_optimizer.py b/python/ray/rllib/optimizers/async_replay_optimizer.py
index 51ab68a0c..72612ed2a 100644
--- a/python/ray/rllib/optimizers/async_replay_optimizer.py
+++ b/python/ray/rllib/optimizers/async_replay_optimizer.py
@@ -230,8 +230,7 @@ class AsyncReplayOptimizer(PolicyOptimizer):
         return sample_timesteps, train_timesteps
 
 
-# reserve 1 CPU so that our method calls don't get stalled
-@ray.remote(num_cpus=1)
+@ray.remote(num_cpus=0)
 class ReplayActor(object):
     """A replay buffer shard.
 
@@ -317,6 +316,8 @@ class ReplayActor(object):
         return stat
 
 
+# note: we set num_cpus=0 to avoid failing to create replay actors when
+# resources are fragmented. This isn't ideal.
 @ray.remote(num_cpus=0)
 class BatchReplayActor(object):
     """The batch replay version of the replay actor.
diff --git a/python/ray/rllib/tests/test_supported_spaces.py b/python/ray/rllib/tests/test_supported_spaces.py
index 93a8366bf..7d59a04fb 100644
--- a/python/ray/rllib/tests/test_supported_spaces.py
+++ b/python/ray/rllib/tests/test_supported_spaces.py
@@ -105,7 +105,7 @@ def check_support_multiagent(alg, config):
 
 class ModelSupportedSpaces(unittest.TestCase):
     def setUp(self):
-        ray.init(num_cpus=10)
+        ray.init(num_cpus=4)
 
     def tearDown(self):
         ray.shutdown()