def test_sync_job_config(shutdown_only): num_java_workers_per_process = 8 runtime_env = {"env_vars": {"key": "value"}} ray.init(job_config=ray.job_config.JobConfig( num_java_workers_per_process=num_java_workers_per_process, runtime_env=runtime_env, )) # Check that the job config is synchronized at the driver side. job_config = ray.worker.global_worker.core_worker.get_job_config() assert job_config.num_java_workers_per_process == num_java_workers_per_process job_runtime_env = RuntimeEnv.deserialize( job_config.runtime_env_info.serialized_runtime_env) assert job_runtime_env.env_vars() == runtime_env["env_vars"] @ray.remote def get_job_config(): job_config = ray.worker.global_worker.core_worker.get_job_config() return job_config.SerializeToString() # Check that the job config is synchronized at the worker side. job_config = gcs_utils.JobConfig() job_config.ParseFromString(ray.get(get_job_config.remote())) assert job_config.num_java_workers_per_process == num_java_workers_per_process job_runtime_env = RuntimeEnv.deserialize( job_config.runtime_env_info.serialized_runtime_env) assert job_runtime_env.env_vars() == runtime_env["env_vars"]
def test_serialization(self): env1 = RuntimeEnv(pip=["requests"], env_vars={ "hi1": "hi1", "hi2": "hi2" }) env2 = RuntimeEnv(env_vars={ "hi2": "hi2", "hi1": "hi1" }, pip=["requests"]) assert env1 == env2 serialized_env1 = env1.serialize() serialized_env2 = env2.serialize() # Key ordering shouldn't matter. assert serialized_env1 == serialized_env2 deserialized_env1 = RuntimeEnv.deserialize(serialized_env1) deserialized_env2 = RuntimeEnv.deserialize(serialized_env2) assert env1 == deserialized_env1 == env2 == deserialized_env2
async def get_job_info(self): """Return info for each job. Here a job is a Ray driver.""" request = gcs_service_pb2.GetAllJobInfoRequest() reply = await self._gcs_job_info_stub.GetAllJobInfo(request, timeout=5) jobs = {} for job_table_entry in reply.job_info_list: job_id = job_table_entry.job_id.hex() metadata = dict(job_table_entry.config.metadata) config = { "namespace": job_table_entry.config.ray_namespace, "metadata": metadata, "runtime_env": RuntimeEnv.deserialize(job_table_entry.config.runtime_env_info. serialized_runtime_env), } info = self._get_job_info(metadata) entry = { "status": None if info is None else info.status, "status_message": None if info is None else info.message, "is_dead": job_table_entry.is_dead, "start_time": job_table_entry.start_time, "end_time": job_table_entry.end_time, "config": config, } jobs[job_id] = entry return jobs
def runtime_env(self): """Get the runtime env used for the current driver or worker. Returns: The runtime env currently using by this worker. The type of return value is ray.runtime_env.RuntimeEnv. """ return RuntimeEnv.deserialize(self.get_runtime_env_string())
async def list_runtime_envs(self, *, option: ListApiOptions) -> List[dict]: """List all runtime env information from the cluster. Returns: A list of runtime env information in the cluster. The schema of returned "dict" is equivalent to the `RuntimeEnvState` protobuf message. We don't have id -> data mapping like other API because runtime env doesn't have unique ids. """ replies = await asyncio.gather(*[ self._client.get_runtime_envs_info(node_id, timeout=option.timeout) for node_id in self._client.get_all_registered_agent_ids() ]) result = [] for node_id, reply in zip(self._client.get_all_registered_agent_ids(), replies): states = reply.runtime_env_states for state in states: data = self._message_to_dict(message=state, fields_to_decode=[]) # Need to deseiralize this field. data["runtime_env"] = RuntimeEnv.deserialize( data["runtime_env"]).to_dict() data["node_id"] = node_id data = filter_fields(data, RuntimeEnvState) result.append(data) # Sort to make the output deterministic. def sort_func(entry): # If creation time is not there yet (runtime env is failed # to be created or not created yet, they are the highest priority. # Otherwise, "bigger" creation time is coming first. if "creation_time_ms" not in entry: return float("inf") elif entry["creation_time_ms"] is None: return float("inf") else: return float(entry["creation_time_ms"]) result.sort(key=sort_func, reverse=True) return list(islice(result, option.limit))
def test_convert_from_and_to_dataclass(): runtime_env = RuntimeEnv() test_plugin = TestPlugin( field1=[ ValueType(nfield1=["a", "b", "c"], nfield2=False), ValueType(nfield1=["d", "e"], nfield2=True), ], field2="abc", ) runtime_env.set("test_plugin", test_plugin) serialized_runtime_env = runtime_env.serialize() assert "test_plugin" in serialized_runtime_env runtime_env_2 = RuntimeEnv.deserialize(serialized_runtime_env) test_plugin_2 = runtime_env_2.get("test_plugin", data_class=TestPlugin) assert len(test_plugin_2.field1) == 2 assert test_plugin_2.field1[0].nfield1 == ["a", "b", "c"] assert test_plugin_2.field1[0].nfield2 is False assert test_plugin_2.field1[1].nfield1 == ["d", "e"] assert test_plugin_2.field1[1].nfield2 is True assert test_plugin_2.field2 == "abc"
def test_serialize_deserialize(option): runtime_env = dict() if option == "pip_list": runtime_env["pip"] = ["pkg1", "pkg2"] elif option == "pip_dict": runtime_env["pip"] = { "packages": ["pkg1", "pkg2"], "pip_check": False, "pip_version": "<22,>20", } elif option == "conda_name": runtime_env["conda"] = "env_name" elif option == "conda_dict": runtime_env["conda"] = {"dependencies": ["dep1", "dep2"]} elif option == "container": runtime_env["container"] = { "image": "anyscale/ray-ml:nightly-py38-cpu", "worker_path": "/root/python/ray/_private/workers/default_worker.py", "run_options": ["--cap-drop SYS_ADMIN", "--log-level=debug"], } else: raise ValueError("unexpected option " + str(option)) typed_runtime_env = RuntimeEnv(**runtime_env) serialized_runtime_env = typed_runtime_env.serialize() cls_runtime_env = RuntimeEnv.deserialize(serialized_runtime_env) cls_runtime_env_dict = cls_runtime_env.to_dict() if "pip" in typed_runtime_env and isinstance(typed_runtime_env["pip"], list): pip_config_in_cls_runtime_env = cls_runtime_env_dict.pop("pip") pip_config_in_runtime_env = typed_runtime_env.pop("pip") assert { "packages": pip_config_in_runtime_env, "pip_check": False, } == pip_config_in_cls_runtime_env assert cls_runtime_env_dict == typed_runtime_env
async def DeleteRuntimeEnvIfPossible(self, request, context): self._logger.info( f"Got request from {request.source_process} to decrease " "reference for runtime env: " f"{request.serialized_runtime_env}.") try: runtime_env = RuntimeEnv.deserialize( request.serialized_runtime_env) except Exception as e: self._logger.exception("[Decrease] Failed to parse runtime env: " f"{request.serialized_runtime_env}") return runtime_env_agent_pb2.GetOrCreateRuntimeEnvReply( status=agent_manager_pb2.AGENT_RPC_STATUS_FAILED, error_message="".join( traceback.format_exception(type(e), e, e.__traceback__)), ) self._reference_table.decrease_reference( runtime_env, request.serialized_runtime_env, request.source_process) return runtime_env_agent_pb2.DeleteRuntimeEnvIfPossibleReply( status=agent_manager_pb2.AGENT_RPC_STATUS_OK)
def runtime_env(self): """Get the runtime env of the current job/worker. If this API is called in driver or ray client, returns the job level runtime env. If this API is called in workers/actors, returns the worker level runtime env. Returns: A new ray.runtime_env.RuntimeEnv instance. To merge from the current runtime env in some specific cases, you can get the current runtime env by this API and modify it by yourself. Example: >>> # Inherit current runtime env, except `env_vars` >>> Actor.options( # doctest: +SKIP ... runtime_env=ray.get_runtime_context().runtime_env.update( ... {"env_vars": {"A": "a", "B": "b"}}) ... ) """ return RuntimeEnv.deserialize(self._get_runtime_env_string())
async def _setup_runtime_env( serialized_runtime_env, serialized_allocated_resource_instances ): runtime_env = RuntimeEnv.deserialize(serialized_runtime_env) allocated_resource: dict = json.loads( serialized_allocated_resource_instances or "{}" ) # Use a separate logger for each job. per_job_logger = self.get_or_create_logger(request.job_id) # TODO(chenk008): Add log about allocated_resource to # avoid lint error. That will be moved to cgroup plugin. per_job_logger.debug(f"Worker has resource :" f"{allocated_resource}") context = RuntimeEnvContext(env_vars=runtime_env.env_vars()) await self._container_manager.setup( runtime_env, context, logger=per_job_logger ) for (manager, uri_cache) in [ (self._working_dir_manager, self._working_dir_uri_cache), (self._conda_manager, self._conda_uri_cache), (self._pip_manager, self._pip_uri_cache), ]: uri = manager.get_uri(runtime_env) if uri is not None: if uri not in uri_cache: per_job_logger.debug(f"Cache miss for URI {uri}.") size_bytes = await manager.create( uri, runtime_env, context, logger=per_job_logger ) uri_cache.add(uri, size_bytes, logger=per_job_logger) else: per_job_logger.debug(f"Cache hit for URI {uri}.") uri_cache.mark_used(uri, logger=per_job_logger) manager.modify_context(uri, runtime_env, context) # Set up py_modules. For now, py_modules uses multiple URIs so # the logic is slightly different from working_dir, conda, and # pip above. py_modules_uris = self._py_modules_manager.get_uris(runtime_env) if py_modules_uris is not None: for uri in py_modules_uris: if uri not in self._py_modules_uri_cache: per_job_logger.debug(f"Cache miss for URI {uri}.") size_bytes = await self._py_modules_manager.create( uri, runtime_env, context, logger=per_job_logger ) self._py_modules_uri_cache.add( uri, size_bytes, logger=per_job_logger ) else: per_job_logger.debug(f"Cache hit for URI {uri}.") self._py_modules_uri_cache.mark_used(uri, logger=per_job_logger) self._py_modules_manager.modify_context( py_modules_uris, runtime_env, context ) # Add the mapping of URIs -> the serialized environment to be # used for cache invalidation. if runtime_env.working_dir_uri(): uri = runtime_env.working_dir_uri() self._uris_to_envs[uri].add(serialized_runtime_env) if runtime_env.py_modules_uris(): for uri in runtime_env.py_modules_uris(): self._uris_to_envs[uri].add(serialized_runtime_env) if runtime_env.conda_uri(): uri = runtime_env.conda_uri() self._uris_to_envs[uri].add(serialized_runtime_env) if runtime_env.pip_uri(): uri = runtime_env.pip_uri() self._uris_to_envs[uri].add(serialized_runtime_env) if runtime_env.plugin_uris(): for uri in runtime_env.plugin_uris(): self._uris_to_envs[uri].add(serialized_runtime_env) def setup_plugins(): # Run setup function from all the plugins for plugin_class_path, config in runtime_env.plugins(): per_job_logger.debug( f"Setting up runtime env plugin {plugin_class_path}" ) plugin_class = import_attr(plugin_class_path) # TODO(simon): implement uri support plugin_class.create( "uri not implemented", json.loads(config), context ) plugin_class.modify_context( "uri not implemented", json.loads(config), context ) loop = asyncio.get_event_loop() # Plugins setup method is sync process, running in other threads # is to avoid blocks asyncio loop await loop.run_in_executor(None, setup_plugins) return context
async def list_runtime_envs(self, *, option: ListApiOptions) -> ListApiResponse: """List all runtime env information from the cluster. Returns: A list of runtime env information in the cluster. The schema of returned "dict" is equivalent to the `RuntimeEnvState` protobuf message. We don't have id -> data mapping like other API because runtime env doesn't have unique ids. """ agent_ids = self._client.get_all_registered_agent_ids() replies = await asyncio.gather( *[ self._client.get_runtime_envs_info(node_id, timeout=option.timeout) for node_id in agent_ids ], return_exceptions=True, ) result = [] unresponsive_nodes = 0 for node_id, reply in zip(self._client.get_all_registered_agent_ids(), replies): if isinstance(reply, DataSourceUnavailable): unresponsive_nodes += 1 continue elif isinstance(reply, Exception): raise reply states = reply.runtime_env_states for state in states: data = self._message_to_dict(message=state, fields_to_decode=[]) # Need to deseiralize this field. data["runtime_env"] = RuntimeEnv.deserialize( data["runtime_env"] ).to_dict() data["node_id"] = node_id result.append(data) partial_failure_warning = None if len(agent_ids) > 0 and unresponsive_nodes > 0: warning_msg = NODE_QUERY_FAILURE_WARNING.format( type="agent", total=len(agent_ids), network_failures=unresponsive_nodes, log_command="dashboard_agent.log", ) if unresponsive_nodes == len(agent_ids): raise DataSourceUnavailable(warning_msg) partial_failure_warning = ( f"The returned data may contain incomplete result. {warning_msg}" ) result = self._filter(result, option.filters, RuntimeEnvState) # Sort to make the output deterministic. def sort_func(entry): # If creation time is not there yet (runtime env is failed # to be created or not created yet, they are the highest priority. # Otherwise, "bigger" creation time is coming first. if "creation_time_ms" not in entry: return float("inf") elif entry["creation_time_ms"] is None: return float("inf") else: return float(entry["creation_time_ms"]) result.sort(key=sort_func, reverse=True) return ListApiResponse( result=list(islice(result, option.limit)), partial_failure_warning=partial_failure_warning, )
async def GetOrCreateRuntimeEnv(self, request, context): self._logger.debug( f"Got request from {request.source_process} to increase " "reference for runtime env: " f"{request.serialized_runtime_env}.") async def _setup_runtime_env(runtime_env, serialized_runtime_env, serialized_allocated_resource_instances): allocated_resource: dict = json.loads( serialized_allocated_resource_instances or "{}") # Use a separate logger for each job. per_job_logger = self.get_or_create_logger(request.job_id) # TODO(chenk008): Add log about allocated_resource to # avoid lint error. That will be moved to cgroup plugin. per_job_logger.debug(f"Worker has resource :" f"{allocated_resource}") context = RuntimeEnvContext(env_vars=runtime_env.env_vars()) await self._container_manager.setup(runtime_env, context, logger=per_job_logger) for manager in self._base_plugin_cache_managers.values(): await manager.create_if_needed(runtime_env, context, logger=per_job_logger) def setup_plugins(): # Run setup function from all the plugins for name, config in runtime_env.plugins(): per_job_logger.debug( f"Setting up runtime env plugin {name}") plugin = self._runtime_env_plugin_manager.get_plugin(name) if plugin is None: raise RuntimeError( f"runtime env plugin {name} not found.") # TODO(architkulkarni): implement uri support plugin.validate(runtime_env) plugin.create("uri not implemented", json.loads(config), context) plugin.modify_context( "uri not implemented", json.loads(config), context, per_job_logger, ) loop = asyncio.get_event_loop() # Plugins setup method is sync process, running in other threads # is to avoid blocking asyncio loop await loop.run_in_executor(None, setup_plugins) return context async def _create_runtime_env_with_retry( runtime_env, serialized_runtime_env, serialized_allocated_resource_instances, setup_timeout_seconds, ) -> Tuple[bool, str, str]: """ Create runtime env with retry times. This function won't raise exceptions. Args: runtime_env(RuntimeEnv): The instance of RuntimeEnv class. serialized_runtime_env(str): The serialized runtime env. serialized_allocated_resource_instances(str): The serialized allocated resource instances. setup_timeout_seconds(int): The timeout of runtime environment creation. Returns: a tuple which contains result(bool), runtime env context(str), error message(str). """ self._logger.info( f"Creating runtime env: {serialized_env} with timeout " f"{setup_timeout_seconds} seconds.") serialized_context = None error_message = None for _ in range(runtime_env_consts.RUNTIME_ENV_RETRY_TIMES): try: # python 3.6 requires the type of input is `Future`, # python 3.7+ only requires the type of input is `Awaitable` # TODO(Catch-Bull): remove create_task when ray drop python 3.6 runtime_env_setup_task = create_task( _setup_runtime_env( runtime_env, serialized_env, request.serialized_allocated_resource_instances, )) runtime_env_context = await asyncio.wait_for( runtime_env_setup_task, timeout=setup_timeout_seconds) serialized_context = runtime_env_context.serialize() error_message = None break except Exception as e: err_msg = f"Failed to create runtime env {serialized_env}." self._logger.exception(err_msg) error_message = "".join( traceback.format_exception(type(e), e, e.__traceback__)) await asyncio.sleep( runtime_env_consts.RUNTIME_ENV_RETRY_INTERVAL_MS / 1000 ) if error_message: self._logger.error( "Runtime env creation failed for %d times, " "don't retry any more.", runtime_env_consts.RUNTIME_ENV_RETRY_TIMES, ) return False, None, error_message else: self._logger.info( "Successfully created runtime env: %s, the context: %s", serialized_env, serialized_context, ) return True, serialized_context, None try: serialized_env = request.serialized_runtime_env runtime_env = RuntimeEnv.deserialize(serialized_env) except Exception as e: self._logger.exception("[Increase] Failed to parse runtime env: " f"{serialized_env}") return runtime_env_agent_pb2.GetOrCreateRuntimeEnvReply( status=agent_manager_pb2.AGENT_RPC_STATUS_FAILED, error_message="".join( traceback.format_exception(type(e), e, e.__traceback__)), ) # Increase reference self._reference_table.increase_reference(runtime_env, serialized_env, request.source_process) if serialized_env not in self._env_locks: # async lock to prevent the same env being concurrently installed self._env_locks[serialized_env] = asyncio.Lock() async with self._env_locks[serialized_env]: if serialized_env in self._env_cache: serialized_context = self._env_cache[serialized_env] result = self._env_cache[serialized_env] if result.success: context = result.result self._logger.info("Runtime env already created " f"successfully. Env: {serialized_env}, " f"context: {context}") return runtime_env_agent_pb2.GetOrCreateRuntimeEnvReply( status=agent_manager_pb2.AGENT_RPC_STATUS_OK, serialized_runtime_env_context=context, ) else: error_message = result.result self._logger.info("Runtime env already failed. " f"Env: {serialized_env}, " f"err: {error_message}") # Recover the reference. self._reference_table.decrease_reference( runtime_env, serialized_env, request.source_process) return runtime_env_agent_pb2.GetOrCreateRuntimeEnvReply( status=agent_manager_pb2.AGENT_RPC_STATUS_FAILED, error_message=error_message, ) if SLEEP_FOR_TESTING_S: self._logger.info(f"Sleeping for {SLEEP_FOR_TESTING_S}s.") time.sleep(int(SLEEP_FOR_TESTING_S)) runtime_env_config = RuntimeEnvConfig.from_proto( request.runtime_env_config) # accroding to the document of `asyncio.wait_for`, # None means disable timeout logic setup_timeout_seconds = ( None if runtime_env_config["setup_timeout_seconds"] == -1 else runtime_env_config["setup_timeout_seconds"]) start = time.perf_counter() ( successful, serialized_context, error_message, ) = await _create_runtime_env_with_retry( runtime_env, serialized_env, request.serialized_allocated_resource_instances, setup_timeout_seconds, ) creation_time_ms = int( round((time.perf_counter() - start) * 1000, 0)) if not successful: # Recover the reference. self._reference_table.decrease_reference( runtime_env, serialized_env, request.source_process) # Add the result to env cache. self._env_cache[serialized_env] = CreatedEnvResult( successful, serialized_context if successful else error_message, creation_time_ms, ) # Reply the RPC return runtime_env_agent_pb2.GetOrCreateRuntimeEnvReply( status=agent_manager_pb2.AGENT_RPC_STATUS_OK if successful else agent_manager_pb2.AGENT_RPC_STATUS_FAILED, serialized_runtime_env_context=serialized_context, error_message=error_message, )