Beispiel #1
0
    def _load_actor_class_from_gcs(self, job_id,
                                   actor_creation_function_descriptor):
        """Load actor class from GCS."""
        key = (b"ActorClass:" + job_id.binary() + b":" +
               actor_creation_function_descriptor.function_id.binary())
        # Wait for the actor class key to have been imported by the
        # import thread. TODO(rkn): It shouldn't be possible to end
        # up in an infinite loop here, but we should push an error to
        # the driver if too much time is spent here.
        while key not in self.imported_actor_classes:
            try:
                # If we're in the process of deserializing an ActorHandle
                # and we hold the function_manager lock, we may be blocking
                # the import_thread from loading the actor class. Use cv.wait
                # to temporarily yield control to the import thread.
                self.cv.wait()
            except RuntimeError:
                # We don't hold the function_manager lock, just sleep regularly
                time.sleep(0.001)

        # Fetch raw data from GCS.
        vals = self._worker.gcs_client.internal_kv_get(
            key, KV_NAMESPACE_FUNCTION_TABLE)
        fields = [
            "job_id", "class_name", "module", "class", "actor_method_names"
        ]
        if vals is None:
            vals = {}
        else:
            vals = pickle.loads(vals)
        (job_id_str, class_name, module, pickled_class,
         actor_method_names) = (vals.get(field) for field in fields)

        class_name = ensure_str(class_name)
        module_name = ensure_str(module)
        job_id = ray.JobID(job_id_str)
        actor_method_names = json.loads(ensure_str(actor_method_names))

        actor_class = None
        try:
            with self.lock:
                actor_class = pickle.loads(pickled_class)
        except Exception:
            logger.debug("Failed to load actor class %s.", class_name)
            # If an exception was thrown when the actor was imported, we record
            # the traceback and notify the scheduler of the failure.
            traceback_str = format_error_message(traceback.format_exc())
            # The actor class failed to be unpickled, create a fake actor
            # class instead (just to produce error messages and to prevent
            # the driver from hanging).
            actor_class = self._create_fake_actor_class(
                class_name, actor_method_names, traceback_str)

        # The below line is necessary. Because in the driver process,
        # if the function is defined in the file where the python script
        # was started from, its module is `__main__`.
        # However in the worker process, the `__main__` module is a
        # different module, which is `default_worker.py`
        actor_class.__module__ = module_name
        return actor_class
Beispiel #2
0
    def _load_actor_class_from_gcs(self, job_id,
                                   actor_creation_function_descriptor):
        """Load actor class from GCS."""
        key = (b"ActorClass:" + job_id.binary() + b":" +
               actor_creation_function_descriptor.function_id.binary())
        # Wait for the actor class key to have been imported by the
        # import thread. TODO(rkn): It shouldn't be possible to end
        # up in an infinite loop here, but we should push an error to
        # the driver if too much time is spent here.
        while key not in self.imported_actor_classes:
            time.sleep(0.001)

        # Fetch raw data from GCS.
        (job_id_str, class_name, module, pickled_class,
         actor_method_names) = self._worker.redis_client.hmget(
             key,
             ["job_id", "class_name", "module", "class", "actor_method_names"])

        class_name = ensure_str(class_name)
        module_name = ensure_str(module)
        job_id = ray.JobID(job_id_str)
        actor_method_names = json.loads(ensure_str(actor_method_names))

        actor_class = None
        try:
            with self.lock:
                actor_class = pickle.loads(pickled_class)
        except Exception:
            logger.exception("Failed to load actor class %s.", class_name)
            # The actor class failed to be unpickled, create a fake actor
            # class instead (just to produce error messages and to prevent
            # the driver from hanging).
            actor_class = self._create_fake_actor_class(
                class_name, actor_method_names)
            # If an exception was thrown when the actor was imported, we record
            # the traceback and notify the scheduler of the failure.
            traceback_str = ray._private.utils.format_error_message(
                traceback.format_exc())
            # Log the error message.
            push_error_to_driver(
                self._worker,
                ray_constants.REGISTER_ACTOR_PUSH_ERROR,
                f"Failed to unpickle actor class '{class_name}' "
                f"for actor ID {self._worker.actor_id.hex()}. "
                f"Traceback:\n{traceback_str}",
                job_id=job_id)
            # TODO(rkn): In the future, it might make sense to have the worker
            # exit here. However, currently that would lead to hanging if
            # someone calls ray.get on a method invoked on the actor.

        # The below line is necessary. Because in the driver process,
        # if the function is defined in the file where the python script
        # was started from, its module is `__main__`.
        # However in the worker process, the `__main__` module is a
        # different module, which is `default_worker.py`
        actor_class.__module__ = module_name
        return actor_class
Beispiel #3
0
    def _load_actor_class_from_gcs(self, job_id,
                                   actor_creation_function_descriptor):
        """Load actor class from GCS."""
        key = make_function_table_key(
            b"ActorClass",
            job_id,
            actor_creation_function_descriptor.function_id.binary(),
        )

        # Fetch raw data from GCS.
        vals = self._worker.gcs_client.internal_kv_get(
            key, KV_NAMESPACE_FUNCTION_TABLE)
        fields = [
            "job_id", "class_name", "module", "class", "actor_method_names"
        ]
        if vals is None:
            vals = {}
        else:
            vals = pickle.loads(vals)
        (job_id_str, class_name, module, pickled_class,
         actor_method_names) = (vals.get(field) for field in fields)

        class_name = ensure_str(class_name)
        module_name = ensure_str(module)
        job_id = ray.JobID(job_id_str)
        actor_method_names = json.loads(ensure_str(actor_method_names))

        actor_class = None
        try:
            with self.lock:
                actor_class = pickle.loads(pickled_class)
        except Exception:
            logger.debug("Failed to load actor class %s.", class_name)
            # If an exception was thrown when the actor was imported, we record
            # the traceback and notify the scheduler of the failure.
            traceback_str = format_error_message(traceback.format_exc())
            # The actor class failed to be unpickled, create a fake actor
            # class instead (just to produce error messages and to prevent
            # the driver from hanging).
            actor_class = self._create_fake_actor_class(
                class_name, actor_method_names, traceback_str)

        # The below line is necessary. Because in the driver process,
        # if the function is defined in the file where the python script
        # was started from, its module is `__main__`.
        # However in the worker process, the `__main__` module is a
        # different module, which is `default_worker.py`
        actor_class.__module__ = module_name
        return actor_class