def test_override_uris(self): child = {} parent = {"uris": ["a", "b"]} assert override_task_or_actor_runtime_env(child, parent) == parent child = {"uris": ["a", "b"]} parent = {"uris": ["c", "d"]} assert override_task_or_actor_runtime_env(child, parent) == child child = {"uris": ["a", "b"]} parent = {} assert override_task_or_actor_runtime_env(child, parent) == child
def test_no_mutate(self): child_env = {} parent_env = {"working_dir": "other_dir", "uris": ["a", "b"]} result_env = override_task_or_actor_runtime_env(child_env, parent_env) assert result_env == {"uris": ["a", "b"]} # The dicts passed in should not be mutated. assert child_env == {} assert parent_env == {"working_dir": "other_dir", "uris": ["a", "b"]}
def test_working_dir_override(self): child_env = {"working_dir": "uri://abc"} parent_env = {"working_dir": "uri://def"} result_env = override_task_or_actor_runtime_env(child_env, parent_env) assert result_env == {"working_dir": "uri://abc"} # The dicts passed in should not be mutated. assert child_env == {"working_dir": "uri://abc"} assert parent_env == {"working_dir": "uri://def"}
def test_uri_override(self): child_env = {"uris": ["c", "d"]} parent_env = {"working_dir": "other_dir", "uris": ["a", "b"]} result_env = override_task_or_actor_runtime_env(child_env, parent_env) assert result_env["uris"] == ["c", "d"] assert result_env.get("working_dir") is None # The dicts passed in should not be mutated. assert child_env == {"uris": ["c", "d"]} assert parent_env == {"working_dir": "other_dir", "uris": ["a", "b"]}
def test_override_env_vars(self): # (child, parent, expected) TEST_CASES = [ ({}, {}, {}), (None, None, None), ({"a": "b"}, {}, {"a": "b"}), ({"a": "b"}, None, {"a": "b"}), ({}, {"a": "b"}, {"a": "b"}), (None, {"a": "b"}, {"a": "b"}), ({"a": "b"}, {"a": "d"}, {"a": "b"}), ({"a": "b"}, {"c": "d"}, {"a": "b", "c": "d"}), ({"a": "b"}, {"a": "e", "c": "d"}, {"a": "b", "c": "d"}) ] # yapf: disable for idx, (child, parent, expected) in enumerate(TEST_CASES): child = {"env_vars": child} if child is not None else {} parent = {"env_vars": parent} if parent is not None else {} expected = {"env_vars": expected} if expected is not None else {} assert override_task_or_actor_runtime_env( child, parent) == expected, f"TEST_INDEX:{idx}"
def _remote(self, args=None, kwargs=None, num_cpus=None, num_gpus=None, memory=None, object_store_memory=None, resources=None, accelerator_type=None, max_concurrency=None, max_restarts=None, max_task_retries=None, name=None, namespace=None, lifetime=None, placement_group="default", placement_group_bundle_index=-1, placement_group_capture_child_tasks=None, runtime_env=None): """Create an actor. This method allows more flexibility than the remote method because resource requirements can be specified and override the defaults in the decorator. Args: args: The arguments to forward to the actor constructor. kwargs: The keyword arguments to forward to the actor constructor. num_cpus: The number of CPUs required by the actor creation task. num_gpus: The number of GPUs required by the actor creation task. memory: Restrict the heap memory usage of this actor. object_store_memory: Restrict the object store memory used by this actor when creating objects. resources: The custom resources required by the actor creation task. max_concurrency: The max number of concurrent calls to allow for this actor. This only works with direct actor calls. The max concurrency defaults to 1 for threaded execution, and 1000 for asyncio execution. Note that the execution order is not guaranteed when max_concurrency > 1. name: The globally unique name for the actor, which can be used to retrieve the actor via ray.get_actor(name) as long as the actor is still alive. namespace: Override the namespace to use for the actor. By default, actors are created in an anonymous namespace. The actor can be retrieved via ray.get_actor(name=name, namespace=namespace). lifetime: Either `None`, which defaults to the actor will fate share with its creator and will be deleted once its refcount drops to zero, or "detached", which means the actor will live as a global object independent of the creator. placement_group: the placement group this actor belongs to, or None if it doesn't belong to any group. Setting to "default" autodetects the placement group based on the current setting of placement_group_capture_child_tasks. placement_group_bundle_index: the index of the bundle if the actor belongs to a placement group, which may be -1 to specify any available bundle. placement_group_capture_child_tasks: Whether or not children tasks of this actor should implicitly use the same placement group as its parent. It is True by default. runtime_env (Dict[str, Any]): Specifies the runtime environment for this actor or task and its children (see :ref:`runtime-environments` for details). This API is in beta and may change before becoming stable. Returns: A handle to the newly created actor. """ if args is None: args = [] if kwargs is None: kwargs = {} meta = self.__ray_metadata__ actor_has_async_methods = len( inspect.getmembers(meta.modified_class, predicate=inspect.iscoroutinefunction)) > 0 is_asyncio = actor_has_async_methods if max_concurrency is None: if is_asyncio: max_concurrency = 1000 else: max_concurrency = 1 if max_concurrency < 1: raise ValueError("max_concurrency must be >= 1") if client_mode_should_convert(auto_init=True): return client_mode_convert_actor( self, args, kwargs, num_cpus=num_cpus, num_gpus=num_gpus, memory=memory, object_store_memory=object_store_memory, resources=resources, accelerator_type=accelerator_type, max_concurrency=max_concurrency, max_restarts=max_restarts, max_task_retries=max_task_retries, name=name, namespace=namespace, lifetime=lifetime, placement_group=placement_group, placement_group_bundle_index=placement_group_bundle_index, placement_group_capture_child_tasks=( placement_group_capture_child_tasks), runtime_env=runtime_env) worker = ray.worker.global_worker worker.check_connected() if name is not None: if not isinstance(name, str): raise TypeError( f"name must be None or a string, got: '{type(name)}'.") elif name == "": raise ValueError("Actor name cannot be an empty string.") if namespace is not None: ray._private.utils.validate_namespace(namespace) # Check whether the name is already taken. # TODO(edoakes): this check has a race condition because two drivers # could pass the check and then create the same named actor. We should # instead check this when we create the actor, but that's currently an # async call. if name is not None: try: ray.get_actor(name, namespace=namespace) except ValueError: # Name is not taken. pass else: raise ValueError( f"The name {name} (namespace={namespace}) is already " "taken. Please use " "a different name or get the existing actor using " f"ray.get_actor('{name}', namespace='{namespace}')") if lifetime is None: detached = False elif lifetime == "detached": detached = True else: raise ValueError( "actor `lifetime` argument must be either `None` or 'detached'" ) # Set the actor's default resources if not already set. First three # conditions are to check that no resources were specified in the # decorator. Last three conditions are to check that no resources were # specified when _remote() was called. if (meta.num_cpus is None and meta.num_gpus is None and meta.resources is None and meta.accelerator_type is None and num_cpus is None and num_gpus is None and resources is None and accelerator_type is None): # In the default case, actors acquire no resources for # their lifetime, and actor methods will require 1 CPU. cpus_to_use = ray_constants.DEFAULT_ACTOR_CREATION_CPU_SIMPLE actor_method_cpu = ray_constants.DEFAULT_ACTOR_METHOD_CPU_SIMPLE else: # If any resources are specified (here or in decorator), then # all resources are acquired for the actor's lifetime and no # resources are associated with methods. cpus_to_use = (ray_constants.DEFAULT_ACTOR_CREATION_CPU_SPECIFIED if meta.num_cpus is None else meta.num_cpus) actor_method_cpu = ray_constants.DEFAULT_ACTOR_METHOD_CPU_SPECIFIED # LOCAL_MODE cannot handle cross_language if worker.mode == ray.LOCAL_MODE: assert not meta.is_cross_language, \ "Cross language ActorClass cannot be executed locally." # Export the actor. if not meta.is_cross_language and (meta.last_export_session_and_job != worker.current_session_and_job): # If this actor class was not exported in this session and job, # we need to export this function again, because current GCS # doesn't have it. meta.last_export_session_and_job = (worker.current_session_and_job) # After serialize / deserialize modified class, the __module__ # of modified class will be ray.cloudpickle.cloudpickle. # So, here pass actor_creation_function_descriptor to make # sure export actor class correct. worker.function_actor_manager.export_actor_class( meta.modified_class, meta.actor_creation_function_descriptor, meta.method_meta.methods.keys()) resources = ray._private.utils.resources_from_resource_arguments( cpus_to_use, meta.num_gpus, meta.memory, meta.object_store_memory, meta.resources, meta.accelerator_type, num_cpus, num_gpus, memory, object_store_memory, resources, accelerator_type) # If the actor methods require CPU resources, then set the required # placement resources. If actor_placement_resources is empty, then # the required placement resources will be the same as resources. actor_placement_resources = {} assert actor_method_cpu in [0, 1] if actor_method_cpu == 1: actor_placement_resources = resources.copy() actor_placement_resources["CPU"] += 1 if meta.is_cross_language: creation_args = cross_language.format_args(worker, args, kwargs) else: function_signature = meta.method_meta.signatures["__init__"] creation_args = signature.flatten_args(function_signature, args, kwargs) if placement_group_capture_child_tasks is None: placement_group_capture_child_tasks = ( worker.should_capture_child_tasks_in_placement_group) placement_group = configure_placement_group_based_on_context( placement_group_capture_child_tasks, placement_group_bundle_index, resources, actor_placement_resources, meta.class_name, placement_group=placement_group) if runtime_env and not isinstance(runtime_env, ParsedRuntimeEnv): runtime_env = ParsedRuntimeEnv(runtime_env) elif isinstance(runtime_env, ParsedRuntimeEnv): pass else: runtime_env = meta.runtime_env parent_runtime_env = worker.core_worker.get_current_runtime_env() parsed_runtime_env = override_task_or_actor_runtime_env( runtime_env, parent_runtime_env) concurrency_groups_dict = {} for cg_name in meta.concurrency_groups: concurrency_groups_dict[cg_name] = { "name": cg_name, "max_concurrency": meta.concurrency_groups[cg_name], "function_descriptors": [], } # Update methods for method_name in meta.method_meta.concurrency_group_for_methods: cg_name = meta.method_meta.concurrency_group_for_methods[ method_name] assert cg_name in concurrency_groups_dict module_name = meta.actor_creation_function_descriptor.module_name class_name = meta.actor_creation_function_descriptor.class_name concurrency_groups_dict[cg_name]["function_descriptors"].append( PythonFunctionDescriptor(module_name, method_name, class_name)) actor_id = worker.core_worker.create_actor( meta.language, meta.actor_creation_function_descriptor, creation_args, max_restarts or meta.max_restarts, max_task_retries or meta.max_task_retries, resources, actor_placement_resources, max_concurrency, detached, name if name is not None else "", namespace if namespace is not None else "", is_asyncio, placement_group.id, placement_group_bundle_index, placement_group_capture_child_tasks, # Store actor_method_cpu in actor handle's extension data. extension_data=str(actor_method_cpu), serialized_runtime_env=parsed_runtime_env.serialize(), runtime_env_uris=parsed_runtime_env.get_uris(), concurrency_groups_dict=concurrency_groups_dict or dict()) actor_handle = ActorHandle(meta.language, actor_id, meta.method_meta.decorators, meta.method_meta.signatures, meta.method_meta.num_returns, actor_method_cpu, meta.actor_creation_function_descriptor, worker.current_session_and_job, original_handle=True) return actor_handle
def test_inherit_pip(self): child_env = {"uris": ["a"]} parent_env = {"pip": ["pkg-name"], "uris": ["a", "b"]} result_env = override_task_or_actor_runtime_env(child_env, parent_env) assert result_env == {"uris": ["a"], "pip": ["pkg-name"]}
def test_inherit_conda(self): child_env = {"uris": ["a"]} parent_env = {"conda": "my-env-name", "uris": ["a", "b"]} result_env = override_task_or_actor_runtime_env(child_env, parent_env) assert result_env == {"uris": ["a"], "conda": "my-env-name"}
def test_working_dir_in_child_invalid(self): child_env = {"working_dir": "some_dir"} parent_env = {"working_dir": "other_dir", "uris": ["a", "b"]} with pytest.raises(NotImplementedError): override_task_or_actor_runtime_env(child_env, parent_env)
def _remote(self, args=None, kwargs=None, num_returns=None, num_cpus=None, num_gpus=None, memory=None, object_store_memory=None, accelerator_type=None, resources=None, max_retries=None, retry_exceptions=None, placement_group="default", placement_group_bundle_index=-1, placement_group_capture_child_tasks=None, runtime_env=None, name=""): """Submit the remote function for execution.""" if client_mode_should_convert(auto_init=True): return client_mode_convert_function( self, args, kwargs, num_returns=num_returns, num_cpus=num_cpus, num_gpus=num_gpus, memory=memory, object_store_memory=object_store_memory, accelerator_type=accelerator_type, resources=resources, max_retries=max_retries, retry_exceptions=retry_exceptions, placement_group=placement_group, placement_group_bundle_index=placement_group_bundle_index, placement_group_capture_child_tasks=( placement_group_capture_child_tasks), runtime_env=runtime_env, name=name) worker = ray.worker.global_worker worker.check_connected() # If this function was not exported in this session and job, we need to # export this function again, because the current GCS doesn't have it. if not self._is_cross_language and \ self._last_export_session_and_job != \ worker.current_session_and_job: # There is an interesting question here. If the remote function is # used by a subsequent driver (in the same script), should the # second driver pickle the function again? If yes, then the remote # function definition can differ in the second driver (e.g., if # variables in its closure have changed). We probably want the # behavior of the remote function in the second driver to be # independent of whether or not the function was invoked by the # first driver. This is an argument for repickling the function, # which we do here. self._pickled_function = pickle.dumps(self._function) self._function_descriptor = PythonFunctionDescriptor.from_function( self._function, self._uuid) self._last_export_session_and_job = worker.current_session_and_job worker.function_actor_manager.export(self) kwargs = {} if kwargs is None else kwargs args = [] if args is None else args if num_returns is None: num_returns = self._num_returns if max_retries is None: max_retries = self._max_retries if retry_exceptions is None: retry_exceptions = self._retry_exceptions if placement_group_capture_child_tasks is None: placement_group_capture_child_tasks = ( worker.should_capture_child_tasks_in_placement_group) if self._placement_group != "default": if self._placement_group: placement_group = self._placement_group else: placement_group = PlacementGroup.empty() elif placement_group == "default": if placement_group_capture_child_tasks: placement_group = get_current_placement_group() else: placement_group = PlacementGroup.empty() if not placement_group: placement_group = PlacementGroup.empty() check_placement_group_index(placement_group, placement_group_bundle_index) resources = ray._private.utils.resources_from_resource_arguments( self._num_cpus, self._num_gpus, self._memory, self._object_store_memory, self._resources, self._accelerator_type, num_cpus, num_gpus, memory, object_store_memory, resources, accelerator_type) if runtime_env and not isinstance(runtime_env, ParsedRuntimeEnv): runtime_env = ParsedRuntimeEnv(runtime_env) elif isinstance(runtime_env, ParsedRuntimeEnv): pass else: runtime_env = self._runtime_env parent_runtime_env = worker.core_worker.get_current_runtime_env() parsed_runtime_env = override_task_or_actor_runtime_env( runtime_env, parent_runtime_env) def invocation(args, kwargs): if self._is_cross_language: list_args = cross_language.format_args(worker, args, kwargs) elif not args and not kwargs and not self._function_signature: list_args = [] else: list_args = ray._private.signature.flatten_args( self._function_signature, args, kwargs) if worker.mode == ray.worker.LOCAL_MODE: assert not self._is_cross_language, \ "Cross language remote function " \ "cannot be executed locally." object_refs = worker.core_worker.submit_task( self._language, self._function_descriptor, list_args, name, num_returns, resources, max_retries, retry_exceptions, placement_group.id, placement_group_bundle_index, placement_group_capture_child_tasks, worker.debugger_breakpoint, parsed_runtime_env.serialize(), parsed_runtime_env.get("uris") or []) # Reset worker's debug context from the last "remote" command # (which applies only to this .remote call). worker.debugger_breakpoint = b"" if len(object_refs) == 1: return object_refs[0] elif len(object_refs) > 1: return object_refs if self._decorator is not None: invocation = self._decorator(invocation) return invocation(args, kwargs)