[Serve] Modify max_concurrency in actor options to respect max_ongoing_requests (#47681)

python/ray/_private/ray_constants.py

@@ -445,6 +445,12 @@ def env_set_by_user(key):
 # tasks.
 DEFAULT_TASK_MAX_RETRIES = 3
 
+# Default max_concurrency option in @ray.remote for threaded actors.
+DEFAULT_MAX_CONCURRENCY_THREADED = 1
+
+# Default max_concurrency option in @ray.remote for async actors.
+DEFAULT_MAX_CONCURRENCY_ASYNC = 1000
+
 # Prefix for namespaces which are used internally by ray.
 # Jobs within these namespaces should be hidden from users
 # and should not be considered user activity.

python/ray/actor.py

@@ -965,7 +965,11 @@ def _remote(self, args=None, kwargs=None, **actor_options):
         is_asyncio = has_async_methods(meta.modified_class)
 
         if actor_options.get("max_concurrency") is None:
-            actor_options["max_concurrency"] = 1000 if is_asyncio else 1
+            actor_options["max_concurrency"] = (
+                ray_constants.DEFAULT_MAX_CONCURRENCY_ASYNC
+                if is_asyncio
+                else ray_constants.DEFAULT_MAX_CONCURRENCY_THREADED
+            )
 
         if client_mode_should_convert():
             return client_mode_convert_actor(self, args, kwargs, **actor_options)

python/ray/serve/_private/deployment_state.py

@@ -13,6 +13,7 @@
 
 import ray
 from ray import ObjectRef, cloudpickle
+from ray._private import ray_constants
 from ray.actor import ActorHandle
 from ray.exceptions import RayActorError, RayError, RayTaskError, RuntimeEnvSetupError
 from ray.serve import metrics
@@ -474,6 +475,18 @@ def start(self, deployment_info: DeploymentInfo) -> ReplicaSchedulingRequest:
         }
         actor_options.update(deployment_info.replica_config.ray_actor_options)
 
+        # A replica's default `max_concurrency` value can prevent it from
+        # respecting the configured `max_ongoing_requests`. To avoid this
+        # unintentional behavior, use `max_ongoing_requests` to override
+        # the Actor's `max_concurrency` if it is larger.
+        if (
+            deployment_info.deployment_config.max_ongoing_requests
+            > ray_constants.DEFAULT_MAX_CONCURRENCY_ASYNC
+        ):
+            actor_options[
+                "max_concurrency"
+            ] = deployment_info.deployment_config.max_ongoing_requests
+
         return ReplicaSchedulingRequest(
             replica_id=self.replica_id,
             actor_def=actor_def,

python/ray/serve/tests/unit/test_deployment_state.py

@@ -5,6 +5,7 @@
 
 import pytest
 
+from ray._private.ray_constants import DEFAULT_MAX_CONCURRENCY_ASYNC
 from ray.serve._private.autoscaling_state import AutoscalingStateManager
 from ray.serve._private.common import (
     DeploymentHandleSource,
@@ -284,6 +285,7 @@ def deployment_info(
     info = DeploymentInfo(
         version=version,
         start_time_ms=0,
+        actor_name="abc",
         deployment_config=DeploymentConfig(
             num_replicas=num_replicas, user_config=user_config, **config_opts
         ),
@@ -2914,6 +2916,23 @@ def test_default_value(self):
         assert actor_replica.health_check_period_s == DEFAULT_HEALTH_CHECK_PERIOD_S
         assert actor_replica.health_check_timeout_s == DEFAULT_HEALTH_CHECK_TIMEOUT_S
 
+    def test_max_concurrency_override(self):
+        actor_replica = ActorReplicaWrapper(
+            version=deployment_version("1"),
+            replica_id=ReplicaID(
+                "abc123",
+                deployment_id=DeploymentID(name="test_deployment", app_name="test_app"),
+            ),
+        )
+        max_ongoing_requests = DEFAULT_MAX_CONCURRENCY_ASYNC + 1
+        d_info, _ = deployment_info(max_ongoing_requests=max_ongoing_requests)
+        replica_scheduling_request = actor_replica.start(d_info)
+        assert (
+            "max_concurrency" in replica_scheduling_request.actor_options
+            and replica_scheduling_request.actor_options["max_concurrency"]
+            == max_ongoing_requests
+        )
+
 
 def test_get_active_node_ids(mock_deployment_state_manager):
     """Test get_active_node_ids() are collecting the correct node ids