Migrate Python C extension to Cython (#3541)

This commit is contained in:
Si-Yuan
2019-01-25 01:17:14 +08:00
committed by Philipp Moritz
parent c1a52b1c86
commit 48139cf861
49 changed files with 1909 additions and 1827 deletions
+44
View File
@@ -0,0 +1,44 @@
# NOTE: These must be checked before including Cython packages to ensure we are using the right python version.
# Segfaults could happen if we are using the wrong version.
set(PYTHON_INCLUDE_DIR ${PYTHON_INCLUDE_DIRS})
# Find Cython executable
get_filename_component(_python_path ${PYTHON_EXECUTABLE} PATH)
find_program(CYTHON_EXECUTABLE
NAMES cython cython.bat cython3
HINTS ${_python_path})
include(FindPackageHandleStandardArgs)
FIND_PACKAGE_HANDLE_STANDARD_ARGS(Cython REQUIRED_VARS CYTHON_EXECUTABLE)
include(UseCython)
include_directories("${NUMPY_INCLUDE_DIR}")
include_directories("${CMAKE_CURRENT_LIST_DIR}/../src")
include_directories("${CMAKE_CURRENT_LIST_DIR}")
# Include some generated Arrow headers.
include_directories("${ARROW_SOURCE_DIR}/../arrow_ep-build/src")
# If the pyx file is a C++ file, we should specify that here.
set_source_files_properties(
${CMAKE_CURRENT_LIST_DIR}/ray/_raylet.pyx
PROPERTIES CYTHON_IS_CXX TRUE)
set(RAY_SRC_DIRECTORY "${CMAKE_CURRENT_LIST_DIR}/../src/ray")
cython_add_module(
_raylet
${RAY_SRC_DIRECTORY}/raylet/raylet_client.cc
${CMAKE_CURRENT_LIST_DIR}/ray/_raylet.pyx)
add_dependencies(_raylet ray_static)
if(APPLE)
target_link_libraries(_raylet "-undefined dynamic_lookup" ray_static)
else()
target_link_libraries(_raylet ray_static)
endif()
# Make sure that the Python extensions are built before copying the files.
add_dependencies(copy_ray _raylet)
+12 -3
View File
@@ -49,7 +49,12 @@ except ImportError as e:
modin_path = os.path.join(os.path.abspath(os.path.dirname(__file__)), "modin")
sys.path.append(modin_path)
from ray.raylet import ObjectID, _config # noqa: E402
from ray._raylet import (UniqueID, ObjectID, DriverID, ClientID, ActorID,
ActorHandleID, FunctionID, ActorClassID, TaskID,
Config as _Config) # noqa: E402
_config = _Config()
from ray.profiling import profile # noqa: E402
from ray.worker import (error_info, init, connect, disconnect, get, put, wait,
remote, get_gpu_ids, get_resource_ids, get_webui_url,
@@ -72,8 +77,12 @@ __all__ = [
"remote", "profile", "actor", "method", "get_gpu_ids", "get_resource_ids",
"get_webui_url", "register_custom_serializer", "shutdown",
"is_initialized", "SCRIPT_MODE", "WORKER_MODE", "LOCAL_MODE",
"PYTHON_MODE", "global_state", "ObjectID", "_config", "__version__",
"internal"
"PYTHON_MODE", "global_state", "_config", "__version__", "internal"
]
__all__ += [
"UniqueID", "ObjectID", "DriverID", "ClientID", "ActorID", "ActorHandleID",
"FunctionID", "ActorClassID", "TaskID"
]
import ctypes # noqa: E402
+320
View File
@@ -0,0 +1,320 @@
# cython: profile=False
# distutils: language = c++
# cython: embedsignature = True
# cython: language_level = 3
from libc.stdint cimport int32_t, int64_t
from libcpp cimport bool as c_bool
from libcpp.memory cimport unique_ptr
from libcpp.string cimport string as c_string
from libcpp.utility cimport pair
from libcpp.unordered_map cimport unordered_map
from libcpp.vector cimport vector as c_vector
from ray.includes.common cimport (
CUniqueID, CTaskID, CObjectID, CFunctionID, CActorClassID, CActorID,
CActorHandleID, CWorkerID, CDriverID, CConfigID, CClientID,
CLanguage, CRayStatus, LANGUAGE_CPP, LANGUAGE_JAVA, LANGUAGE_PYTHON)
from ray.includes.libraylet cimport (
CRayletClient, GCSProfileTableDataT, GCSProfileEventT,
ResourceMappingType, WaitResultPair)
from ray.includes.task cimport CTaskSpecification
from ray.includes.ray_config cimport RayConfig
from ray.utils import decode
from cython.operator import dereference, postincrement
cimport cpython
include "includes/unique_ids.pxi"
include "includes/ray_config.pxi"
include "includes/task.pxi"
if cpython.PY_MAJOR_VERSION >= 3:
import pickle
else:
import cPickle as pickle
import numpy
cdef int check_status(const CRayStatus& status) nogil except -1:
if status.ok():
return 0
with gil:
message = status.message().decode()
raise Exception(message)
cdef c_vector[CObjectID] ObjectIDsToVector(object_ids):
"""A helper function that converts a Python list of object IDs to a vector.
Args:
object_ids (list): The Python list of object IDs.
Returns:
The output vector.
"""
cdef:
ObjectID object_id
c_vector[CObjectID] result
for object_id in object_ids:
result.push_back(object_id.data)
return result
cdef VectorToObjectIDs(c_vector[CObjectID] object_ids):
result = []
for i in range(object_ids.size()):
result.append(ObjectID.from_native(object_ids[i]))
return result
def compute_put_id(TaskID task_id, int64_t put_index):
if put_index < 1 or put_index > kMaxTaskPuts:
raise ValueError("The range of 'put_index' should be [1, %d]" % kMaxTaskPuts)
return ObjectID.from_native(ComputePutId(task_id.data, put_index))
def compute_task_id(ObjectID object_id):
return TaskID.from_native(ComputeTaskId(object_id.data))
cdef c_bool is_simple_value(value, int *num_elements_contained):
num_elements_contained[0] += 1
if num_elements_contained[0] >= RayConfig.instance().num_elements_limit():
return False
if (cpython.PyInt_Check(value) or cpython.PyLong_Check(value) or value is False or
value is True or cpython.PyFloat_Check(value) or value is None):
return True
if cpython.PyBytes_CheckExact(value):
num_elements_contained[0] += cpython.PyBytes_Size(value)
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
if cpython.PyUnicode_CheckExact(value):
num_elements_contained[0] += cpython.PyUnicode_GET_SIZE(value)
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
if cpython.PyList_CheckExact(value) and cpython.PyList_Size(value) < RayConfig.instance().size_limit():
for item in value:
if not is_simple_value(item, num_elements_contained):
return False
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
if cpython.PyDict_CheckExact(value) and cpython.PyDict_Size(value) < RayConfig.instance().size_limit():
# TODO(suquark): Using "items" in Python2 is not very efficient.
for k, v in value.items():
if not (is_simple_value(k, num_elements_contained) and is_simple_value(v, num_elements_contained)):
return False
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
if cpython.PyTuple_CheckExact(value) and cpython.PyTuple_Size(value) < RayConfig.instance().size_limit():
for item in value:
if not is_simple_value(item, num_elements_contained):
return False
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
if isinstance(value, numpy.ndarray):
if value.dtype == "O":
return False
num_elements_contained[0] += value.nbytes
return num_elements_contained[0] < RayConfig.instance().num_elements_limit()
return False
def check_simple_value(value):
"""Check if value is simple enough to be send by value.
This method checks if a Python object is sufficiently simple that it can be
serialized and passed by value as an argument to a task (without being put in
the object store). The details of which objects are sufficiently simple are
defined by this method and are not particularly important. But for
performance reasons, it is better to place "small" objects in the task itself
and "large" objects in the object store.
Args:
value: Python object that should be checked.
Returns:
True if the value should be send by value, False otherwise.
"""
cdef int num_elements_contained = 0
return is_simple_value(value, &num_elements_contained)
cdef class Language:
cdef CLanguage lang
def __cinit__(self, int32_t lang):
self.lang = <CLanguage>lang
@staticmethod
cdef from_native(const CLanguage& lang):
return Language(<int32_t>lang)
def __eq__(self, other):
return isinstance(other, Language) and (<int32_t>self.lang) == (<int32_t>other.lang)
def __repr__(self):
if <int32_t>self.lang == <int32_t>LANGUAGE_PYTHON:
return "PYTHON"
elif <int32_t>self.lang == <int32_t>LANGUAGE_CPP:
return "CPP"
elif <int32_t>self.lang == <int32_t>LANGUAGE_JAVA:
return "JAVA"
else:
raise Exception("Unexpected error")
# Programming language enum values.
cdef Language LANG_PYTHON = Language.from_native(LANGUAGE_PYTHON)
cdef Language LANG_CPP = Language.from_native(LANGUAGE_CPP)
cdef Language LANG_JAVA = Language.from_native(LANGUAGE_JAVA)
cdef unordered_map[c_string, double] resource_map_from_python_dict(resource_map):
cdef:
unordered_map[c_string, double] out
c_string resource_name
if not isinstance(resource_map, dict):
raise TypeError("resource_map must be a dictionary")
for key, value in resource_map.items():
out[key.encode("ascii")] = float(value)
return out
cdef class RayletClient:
cdef unique_ptr[CRayletClient] client
def __cinit__(self, raylet_socket,
ClientID client_id,
c_bool is_worker,
DriverID driver_id):
# We know that we are using Python, so just skip the language parameter.
# TODO(suquark): Should we allow unicode chars in "raylet_socket"?
self.client.reset(new CRayletClient(raylet_socket.encode("ascii"), client_id.data,
is_worker, driver_id.data, LANGUAGE_PYTHON))
def disconnect(self):
check_status(self.client.get().Disconnect())
def submit_task(self, Task task_spec):
check_status(self.client.get().SubmitTask(task_spec.execution_dependencies.get()[0], task_spec.task_spec.get()[0]))
def get_task(self):
cdef:
unique_ptr[CTaskSpecification] task_spec
with nogil:
check_status(self.client.get().GetTask(&task_spec))
return Task.make(task_spec)
def task_done(self):
check_status(self.client.get().TaskDone())
def fetch_or_reconstruct(self, object_ids,
c_bool fetch_only, TaskID current_task_id=TaskID.nil()):
cdef c_vector[CObjectID] fetch_ids = ObjectIDsToVector(object_ids)
check_status(self.client.get().FetchOrReconstruct(fetch_ids, fetch_only, current_task_id.data))
def notify_unblocked(self, TaskID current_task_id):
check_status(self.client.get().NotifyUnblocked(current_task_id.data))
def wait(self, object_ids, int num_returns, int64_t timeout_milliseconds,
c_bool wait_local, TaskID current_task_id):
cdef:
WaitResultPair result
c_vector[CObjectID] wait_ids
wait_ids = ObjectIDsToVector(object_ids)
check_status(self.client.get().Wait(wait_ids, num_returns, timeout_milliseconds,
wait_local, current_task_id.data, &result))
return VectorToObjectIDs(result.first), VectorToObjectIDs(result.second)
def resource_ids(self):
cdef:
ResourceMappingType resource_mapping = self.client.get().GetResourceIDs()
unordered_map[c_string, c_vector[pair[int64_t, double]]].iterator iterator = resource_mapping.begin()
c_vector[pair[int64_t, double]] c_value
resources_dict = {}
while iterator != resource_mapping.end():
key = decode(dereference(iterator).first)
c_value = dereference(iterator).second
ids_and_fractions = []
for i in range(c_value.size()):
ids_and_fractions.append((c_value[i].first, c_value[i].second))
resources_dict[key] = ids_and_fractions
postincrement(iterator)
return resources_dict
def push_error(self, DriverID job_id, error_type, error_message,
double timestamp):
check_status(self.client.get().PushError(job_id.data,
error_type.encode("ascii"),
error_message.encode("ascii"),
timestamp))
def push_profile_events(self, component_type, UniqueID component_id,
node_ip_address, profile_data):
cdef:
GCSProfileTableDataT profile_info
GCSProfileEventT *profile_event
c_string event_type
if len(profile_data) == 0:
return # Short circuit if there are no profile events.
profile_info.component_type = component_type.encode("ascii")
profile_info.component_id = component_id.binary()
profile_info.node_ip_address = node_ip_address.encode("ascii")
for py_profile_event in profile_data:
profile_event = new GCSProfileEventT()
if not isinstance(py_profile_event, dict):
raise TypeError("Incorrect type for a profile event. Expected dict instead of '%s'" % str(type(py_profile_event)))
# TODO(rkn): If the dictionary is formatted incorrectly, that could lead
# to errors. E.g., if any of the strings are empty, that will cause
# segfaults in the node manager.
for key_string, event_data in py_profile_event.items():
if key_string == "event_type":
profile_event.event_type = event_data.encode("ascii")
if profile_event.event_type.length() == 0:
raise ValueError("'event_type' should not be a null string.")
elif key_string == "start_time":
profile_event.start_time = float(event_data)
elif key_string == "end_time":
profile_event.end_time = float(event_data)
elif key_string == "extra_data":
profile_event.extra_data = event_data.encode("ascii")
if profile_event.extra_data.length() == 0:
raise ValueError("'extra_data' should not be a null string.")
else:
raise ValueError("Unknown profile event key '%s'" % key_string)
# Note that profile_info.profile_events is a vector of unique pointers, so
# profile_event will be deallocated when profile_info goes out of scope.
# "emplace_back" of vector has not been supported by Cython
profile_info.profile_events.push_back(unique_ptr[GCSProfileEventT](profile_event))
check_status(self.client.get().PushProfileEvents(profile_info))
def free_objects(self, object_ids, c_bool local_only):
cdef c_vector[CObjectID] free_ids = ObjectIDsToVector(object_ids)
check_status(self.client.get().FreeObjects(free_ids, local_only))
@property
def language(self):
return Language.from_native(self.client.get().GetLanguage())
@property
def client_id(self):
return ClientID.from_native(self.client.get().GetClientID())
@property
def driver_id(self):
return DriverID.from_native(self.client.get().GetDriverID())
@property
def is_worker(self):
return self.client.get().IsWorker()
+38 -33
View File
@@ -12,12 +12,12 @@ import traceback
import ray.cloudpickle as pickle
from ray.function_manager import FunctionDescriptor
import ray.raylet
import ray.ray_constants as ray_constants
import ray.signature as signature
import ray.worker
from ray.utils import _random_string
from ray import ObjectID
from ray import (ObjectID, ActorID, ActorHandleID, ActorClassID, TaskID,
DriverID)
DEFAULT_ACTOR_METHOD_NUM_RETURN_VALS = 1
@@ -38,11 +38,12 @@ def compute_actor_handle_id(actor_handle_id, num_forks):
Returns:
An ID for the new actor handle.
"""
assert isinstance(actor_handle_id, ActorHandleID)
handle_id_hash = hashlib.sha1()
handle_id_hash.update(actor_handle_id.id())
handle_id_hash.update(actor_handle_id.binary())
handle_id_hash.update(str(num_forks).encode("ascii"))
handle_id = handle_id_hash.digest()
return ObjectID(handle_id)
return ActorHandleID(handle_id)
def compute_actor_handle_id_non_forked(actor_handle_id, current_task_id):
@@ -65,11 +66,13 @@ def compute_actor_handle_id_non_forked(actor_handle_id, current_task_id):
Returns:
An ID for the new actor handle.
"""
assert isinstance(actor_handle_id, ActorHandleID)
assert isinstance(current_task_id, TaskID)
handle_id_hash = hashlib.sha1()
handle_id_hash.update(actor_handle_id.id())
handle_id_hash.update(current_task_id.id())
handle_id_hash.update(actor_handle_id.binary())
handle_id_hash.update(current_task_id.binary())
handle_id = handle_id_hash.digest()
return ObjectID(handle_id)
return ActorHandleID(handle_id)
def set_actor_checkpoint(worker, actor_id, checkpoint_index, checkpoint,
@@ -83,7 +86,8 @@ def set_actor_checkpoint(worker, actor_id, checkpoint_index, checkpoint,
checkpoint: The state object to save.
frontier: The task frontier at the time of the checkpoint.
"""
actor_key = b"Actor:" + actor_id.id()
assert isinstance(actor_id, ActorID)
actor_key = b"Actor:" + actor_id.binary()
worker.redis_client.hmset(
actor_key, {
"checkpoint_index": checkpoint_index,
@@ -155,7 +159,8 @@ def get_actor_checkpoint(worker, actor_id):
exists, all objects are set to None. The checkpoint index is the .
executed on the actor before the checkpoint was made.
"""
actor_key = b"Actor:" + actor_id.id()
assert isinstance(actor_id, ActorID)
actor_key = b"Actor:" + actor_id.binary()
checkpoint_index, checkpoint, frontier = worker.redis_client.hmget(
actor_key, ["checkpoint_index", "checkpoint", "frontier"])
if checkpoint_index is not None:
@@ -370,7 +375,7 @@ class ActorClass(object):
raise Exception("Actors cannot be created before ray.init() "
"has been called.")
actor_id = ObjectID(_random_string())
actor_id = ActorID(_random_string())
# The actor cursor is a dummy object representing the most recent
# actor method invocation. For each subsequent method invocation,
# the current cursor should be added as a dependency, and then
@@ -423,6 +428,7 @@ class ActorClass(object):
num_return_vals=1,
resources=resources,
placement_resources=actor_placement_resources)
assert isinstance(actor_cursor, ObjectID)
actor_handle = ActorHandle(
actor_id, self._modified_class.__module__, self._class_name,
@@ -502,14 +508,17 @@ class ActorHandle(object):
actor_method_cpus,
actor_driver_id,
actor_handle_id=None):
assert isinstance(actor_id, ActorID)
assert isinstance(actor_driver_id, DriverID)
self._ray_actor_id = actor_id
self._ray_module_name = module_name
# False if this actor handle was created by forking or pickling. True
# if it was created by the _serialization_helper function.
self._ray_original_handle = actor_handle_id is None
if self._ray_original_handle:
self._ray_actor_handle_id = ObjectID.nil_id()
self._ray_actor_handle_id = ActorHandleID.nil()
else:
assert isinstance(actor_handle_id, ActorHandleID)
self._ray_actor_handle_id = actor_handle_id
self._ray_actor_cursor = actor_cursor
self._ray_actor_counter = 0
@@ -646,7 +655,7 @@ class ActorHandle(object):
# not just the first one.
worker = ray.worker.get_global_worker()
if (worker.mode == ray.worker.SCRIPT_MODE
and self._ray_actor_driver_id.id() != worker.worker_id):
and self._ray_actor_driver_id.binary() != worker.worker_id):
# If the worker is a driver and driver id has changed because
# Ray was shut down re-initialized, the actor is already cleaned up
# and we don't need to send `__ray_terminate__` again.
@@ -684,22 +693,22 @@ class ActorHandle(object):
else:
actor_handle_id = self._ray_actor_handle_id
# Note: _ray_actor_cursor and _ray_actor_creation_dummy_object_id
# could be None.
state = {
"actor_id": self._ray_actor_id.id(),
"actor_handle_id": actor_handle_id.id(),
"actor_id": self._ray_actor_id,
"actor_handle_id": actor_handle_id,
"module_name": self._ray_module_name,
"class_name": self._ray_class_name,
"actor_cursor": self._ray_actor_cursor.id()
if self._ray_actor_cursor is not None else None,
"actor_cursor": self._ray_actor_cursor,
"actor_method_names": self._ray_actor_method_names,
"method_signatures": self._ray_method_signatures,
"method_num_return_vals": self._ray_method_num_return_vals,
# Actors in local mode don't have dummy objects.
"actor_creation_dummy_object_id": self.
_ray_actor_creation_dummy_object_id.id()
if self._ray_actor_creation_dummy_object_id is not None else None,
_ray_actor_creation_dummy_object_id,
"actor_method_cpus": self._ray_actor_method_cpus,
"actor_driver_id": self._ray_actor_driver_id.id(),
"actor_driver_id": self._ray_actor_driver_id,
"ray_forking": ray_forking
}
@@ -711,7 +720,7 @@ class ActorHandle(object):
# to release, since it could be unpickled and submit another
# dependent task at any time. Therefore, we notify the backend of a
# random handle ID that will never actually be used.
new_actor_handle_id = ObjectID(_random_string())
new_actor_handle_id = ActorHandleID(_random_string())
# Notify the backend to expect this new actor handle. The backend will
# not release the cursor for any new handles until the first task for
# each of the new handles is submitted.
@@ -733,7 +742,7 @@ class ActorHandle(object):
worker.check_connected()
if state["ray_forking"]:
actor_handle_id = ObjectID(state["actor_handle_id"])
actor_handle_id = state["actor_handle_id"]
else:
# Right now, if the actor handle has been pickled, we create a
# temporary actor handle id for invocations.
@@ -747,25 +756,21 @@ class ActorHandle(object):
# same actor is likely a performance bug. We should consider
# logging a warning in these cases.
actor_handle_id = compute_actor_handle_id_non_forked(
ObjectID(state["actor_handle_id"]), worker.current_task_id)
# This is the driver ID of the driver that owns the actor, not
# necessarily the driver that owns this actor handle.
actor_driver_id = ObjectID(state["actor_driver_id"])
state["actor_handle_id"], worker.current_task_id)
self.__init__(
ObjectID(state["actor_id"]),
state["actor_id"],
state["module_name"],
state["class_name"],
ObjectID(state["actor_cursor"])
if state["actor_cursor"] is not None else None,
state["actor_cursor"],
state["actor_method_names"],
state["method_signatures"],
state["method_num_return_vals"],
ObjectID(state["actor_creation_dummy_object_id"])
if state["actor_creation_dummy_object_id"] is not None else None,
state["actor_creation_dummy_object_id"],
state["actor_method_cpus"],
actor_driver_id,
# This is the driver ID of the driver that owns the actor, not
# necessarily the driver that owns this actor handle.
state["actor_driver_id"],
actor_handle_id=actor_handle_id)
def __getstate__(self):
@@ -887,7 +892,7 @@ def make_actor(cls, num_cpus, num_gpus, resources, actor_method_cpus,
Class.__module__ = cls.__module__
Class.__name__ = cls.__name__
class_id = _random_string()
class_id = ActorClassID(_random_string())
return ActorClass(Class, class_id, checkpoint_interval,
max_reconstructions, num_cpus, num_gpus, resources,
+1 -1
View File
@@ -215,7 +215,7 @@ class PlasmaEventHandler:
if not isinstance(object_id, ray.ObjectID):
raise TypeError("Input should be an ObjectID.")
plain_object_id = plasma.ObjectID(object_id.id())
plain_object_id = plasma.ObjectID(object_id.binary())
fut = PlasmaObjectFuture(loop=self._loop, object_id=plain_object_id)
if check_ready:
+1 -1
View File
@@ -91,7 +91,7 @@ def _task_table_shard(shard_index):
for key in task_table_keys:
task_id_binary = key[len(TASK_PREFIX):]
results[binary_to_hex(task_id_binary)] = ray.global_state._task_table(
ray.ObjectID(task_id_binary))
ray.TaskID(task_id_binary))
return results
+27 -26
View File
@@ -25,7 +25,7 @@ def parse_client_table(redis_client):
Returns:
A list of information about the nodes in the cluster.
"""
NIL_CLIENT_ID = ray.ObjectID.nil_id().id()
NIL_CLIENT_ID = ray.ObjectID.nil().binary()
message = redis_client.execute_command("RAY.TABLE_LOOKUP",
ray.gcs_utils.TablePrefix.CLIENT,
"", NIL_CLIENT_ID)
@@ -216,8 +216,7 @@ class GlobalState(object):
"""Fetch and parse the object table information for a single object ID.
Args:
object_id_binary: A string of bytes with the object ID to get
information about.
object_id: An object ID to get information about.
Returns:
A dictionary with information about the object ID in question.
@@ -229,7 +228,7 @@ class GlobalState(object):
# Return information about a single object ID.
message = self._execute_command(object_id, "RAY.TABLE_LOOKUP",
ray.gcs_utils.TablePrefix.OBJECT, "",
object_id.id())
object_id.binary())
gcs_entry = ray.gcs_utils.GcsTableEntry.GetRootAsGcsTableEntry(
message, 0)
@@ -284,15 +283,15 @@ class GlobalState(object):
"""Fetch and parse the task table information for a single task ID.
Args:
task_id_binary: A string of bytes with the task ID to get
information about.
task_id: A task ID to get information about.
Returns:
A dictionary with information about the task ID in question.
"""
assert isinstance(task_id, ray.TaskID)
message = self._execute_command(task_id, "RAY.TABLE_LOOKUP",
ray.gcs_utils.TablePrefix.RAYLET_TASK,
"", task_id.id())
"", task_id.binary())
gcs_entries = ray.gcs_utils.GcsTableEntry.GetRootAsGcsTableEntry(
message, 0)
@@ -303,23 +302,23 @@ class GlobalState(object):
execution_spec = task_table_message.TaskExecutionSpec()
task_spec = task_table_message.TaskSpecification()
task_spec = ray.raylet.task_from_string(task_spec)
function_descriptor_list = task_spec.function_descriptor_list()
task = ray._raylet.Task.from_string(task_spec)
function_descriptor_list = task.function_descriptor_list()
function_descriptor = FunctionDescriptor.from_bytes_list(
function_descriptor_list)
task_spec_info = {
"DriverID": task_spec.driver_id().hex(),
"TaskID": task_spec.task_id().hex(),
"ParentTaskID": task_spec.parent_task_id().hex(),
"ParentCounter": task_spec.parent_counter(),
"ActorID": (task_spec.actor_id().hex()),
"ActorCreationID": task_spec.actor_creation_id().hex(),
"DriverID": task.driver_id().hex(),
"TaskID": task.task_id().hex(),
"ParentTaskID": task.parent_task_id().hex(),
"ParentCounter": task.parent_counter(),
"ActorID": (task.actor_id().hex()),
"ActorCreationID": task.actor_creation_id().hex(),
"ActorCreationDummyObjectID": (
task_spec.actor_creation_dummy_object_id().hex()),
"ActorCounter": task_spec.actor_counter(),
"Args": task_spec.arguments(),
"ReturnObjectIDs": task_spec.returns(),
"RequiredResources": task_spec.required_resources(),
task.actor_creation_dummy_object_id().hex()),
"ActorCounter": task.actor_counter(),
"Args": task.arguments(),
"ReturnObjectIDs": task.returns(),
"RequiredResources": task.required_resources(),
"FunctionID": function_descriptor.function_id.hex(),
"FunctionHash": binary_to_hex(function_descriptor.function_hash),
"ModuleName": function_descriptor.module_name,
@@ -351,7 +350,7 @@ class GlobalState(object):
"""
self._check_connected()
if task_id is not None:
task_id = ray.ObjectID(hex_to_binary(task_id))
task_id = ray.TaskID(hex_to_binary(task_id))
return self._task_table(task_id)
else:
task_table_keys = self._keys(
@@ -364,7 +363,7 @@ class GlobalState(object):
results = {}
for task_id_binary in task_ids_binary:
results[binary_to_hex(task_id_binary)] = self._task_table(
ray.ObjectID(task_id_binary))
ray.TaskID(task_id_binary))
return results
def function_table(self, function_id=None):
@@ -439,7 +438,7 @@ class GlobalState(object):
# events and should also support returning a window of events.
message = self._execute_command(batch_id, "RAY.TABLE_LOOKUP",
ray.gcs_utils.TablePrefix.PROFILE, "",
batch_id.id())
batch_id.binary())
if message is None:
return []
@@ -877,9 +876,10 @@ class GlobalState(object):
Returns:
A list of the error messages for this job.
"""
assert isinstance(job_id, ray.DriverID)
message = self.redis_client.execute_command(
"RAY.TABLE_LOOKUP", ray.gcs_utils.TablePrefix.ERROR_INFO, "",
job_id.id())
job_id.binary())
# If there are no errors, return early.
if message is None:
@@ -891,7 +891,7 @@ class GlobalState(object):
for i in range(gcs_entries.EntriesLength()):
error_data = ray.gcs_utils.ErrorTableData.GetRootAsErrorTableData(
gcs_entries.Entries(i), 0)
assert job_id.id() == error_data.JobId()
assert job_id.binary() == error_data.JobId()
error_message = {
"type": decode(error_data.Type()),
"message": decode(error_data.ErrorMessage()),
@@ -912,6 +912,7 @@ class GlobalState(object):
that job.
"""
if job_id is not None:
assert isinstance(job_id, ray.DriverID)
return self._error_messages(job_id)
error_table_keys = self.redis_client.keys(
@@ -922,6 +923,6 @@ class GlobalState(object):
]
return {
binary_to_hex(job_id): self._error_messages(ray.ObjectID(job_id))
binary_to_hex(job_id): self._error_messages(ray.DriverID(job_id))
for job_id in job_ids
}
+18 -17
View File
@@ -223,7 +223,7 @@ class FunctionDescriptor(object):
ray.ObjectID to represent the function descriptor.
"""
if self.is_for_driver_task:
return ray.ObjectID.nil_id()
return ray.FunctionID.nil()
function_id_hash = hashlib.sha1()
# Include the function module and name in the hash.
function_id_hash.update(self.module_name.encode("ascii"))
@@ -232,7 +232,7 @@ class FunctionDescriptor(object):
function_id_hash.update(self._function_source_hash)
# Compute the function ID.
function_id = function_id_hash.digest()
return ray.ObjectID(function_id)
return ray.FunctionID(function_id)
def get_function_descriptor_list(self):
"""Return a list of bytes representing the function descriptor.
@@ -355,13 +355,13 @@ class FunctionActorManager(object):
check_oversized_pickle(pickled_function,
remote_function._function_name,
"remote function", self._worker)
key = (b"RemoteFunction:" + self._worker.task_driver_id.id() + b":" +
remote_function._function_descriptor.function_id.id())
key = (b"RemoteFunction:" + self._worker.task_driver_id.binary() + b":"
+ remote_function._function_descriptor.function_id.binary())
self._worker.redis_client.hmset(
key, {
"driver_id": self._worker.task_driver_id.id(),
"driver_id": self._worker.task_driver_id.binary(),
"function_id": remote_function._function_descriptor.
function_id.id(),
function_id.binary(),
"name": remote_function._function_name,
"module": function.__module__,
"function": pickled_function,
@@ -377,8 +377,8 @@ class FunctionActorManager(object):
"driver_id", "function_id", "name", "function", "num_return_vals",
"module", "resources", "max_calls"
])
function_id = ray.ObjectID(function_id_str)
driver_id = ray.ObjectID(driver_id_str)
function_id = ray.FunctionID(function_id_str)
driver_id = ray.DriverID(driver_id_str)
function_name = decode(function_name)
max_calls = int(max_calls)
module = decode(module)
@@ -406,7 +406,7 @@ class FunctionActorManager(object):
traceback_str,
driver_id=driver_id,
data={
"function_id": function_id.id(),
"function_id": function_id.binary(),
"function_name": function_name
})
else:
@@ -423,7 +423,8 @@ class FunctionActorManager(object):
max_calls=max_calls))
# Add the function to the function table.
self._worker.redis_client.rpush(
b"FunctionTable:" + function_id.id(), self._worker.worker_id)
b"FunctionTable:" + function_id.binary(),
self._worker.worker_id)
def get_execution_info(self, driver_id, function_descriptor):
"""Get the FunctionExecutionInfo of a remote function.
@@ -524,14 +525,14 @@ class FunctionActorManager(object):
"You might have started a background thread in a non-actor task, "
"please make sure the thread finishes before the task finishes.")
driver_id = self._worker.task_driver_id
key = (b"ActorClass:" + driver_id.id() + b":" +
function_descriptor.function_id.id())
key = (b"ActorClass:" + driver_id.binary() + b":" +
function_descriptor.function_id.binary())
actor_class_info = {
"class_name": Class.__name__,
"module": Class.__module__,
"class": pickle.dumps(Class),
"checkpoint_interval": checkpoint_interval,
"driver_id": driver_id.id(),
"driver_id": driver_id.binary(),
"actor_method_names": json.dumps(list(actor_method_names))
}
@@ -556,8 +557,8 @@ class FunctionActorManager(object):
# because of https://github.com/ray-project/ray/issues/1146.
def load_actor(self, driver_id, function_descriptor):
key = (b"ActorClass:" + driver_id.id() + b":" +
function_descriptor.function_id.id())
key = (b"ActorClass:" + driver_id.binary() + b":" +
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
@@ -588,7 +589,7 @@ class FunctionActorManager(object):
class_name = decode(class_name)
module = decode(module)
driver_id = ray.ObjectID(driver_id_str)
driver_id = ray.DriverID(driver_id_str)
checkpoint_interval = int(checkpoint_interval)
actor_method_names = json.loads(decode(actor_method_names))
@@ -645,7 +646,7 @@ class FunctionActorManager(object):
ray_constants.REGISTER_ACTOR_PUSH_ERROR,
traceback_str,
driver_id,
data={"actor_id": actor_id.id()})
data={"actor_id": actor_id.binary()})
# 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.
+1 -1
View File
@@ -58,7 +58,7 @@ def construct_error_message(driver_id, error_type, message, timestamp):
The serialized object.
"""
builder = flatbuffers.Builder(0)
driver_offset = builder.CreateString(driver_id.id())
driver_offset = builder.CreateString(driver_id.binary())
error_type_offset = builder.CreateString(error_type)
message_offset = builder.CreateString(message)
+2 -2
View File
@@ -112,7 +112,7 @@ class ImportThread(object):
if (utils.decode(run_on_other_drivers) == "False"
and self.worker.mode == ray.SCRIPT_MODE
and driver_id != self.worker.task_driver_id.id()):
and driver_id != self.worker.task_driver_id.binary()):
return
try:
@@ -131,5 +131,5 @@ class ImportThread(object):
self.worker,
ray_constants.FUNCTION_TO_RUN_PUSH_ERROR,
traceback_str,
driver_id=ray.ObjectID(driver_id),
driver_id=ray.DriverID(driver_id),
data={"name": name})
View File
+120
View File
@@ -0,0 +1,120 @@
from libcpp.string cimport string as c_string
from libcpp cimport bool as c_bool
from libc.stdint cimport int64_t
from libcpp.unordered_map cimport unordered_map
from libcpp.vector cimport vector as c_vector
from ray.includes.unique_ids cimport (
CUniqueID, TaskID as CTaskID, ObjectID as CObjectID,
FunctionID as CFunctionID, ActorClassID as CActorClassID, ActorID as CActorID,
ActorHandleID as CActorHandleID, WorkerID as CWorkerID,
DriverID as CDriverID, ConfigID as CConfigID, ClientID as CClientID)
cdef extern from "ray/status.h" namespace "ray" nogil:
cdef cppclass StatusCode:
pass
cdef cppclass CRayStatus "ray::Status":
RayStatus()
RayStatus(StatusCode code, const c_string &msg)
RayStatus(const CRayStatus &s);
@staticmethod
CRayStatus OK()
@staticmethod
CRayStatus OutOfMemory()
@staticmethod
CRayStatus KeyError()
@staticmethod
CRayStatus Invalid()
@staticmethod
CRayStatus IOError()
@staticmethod
CRayStatus TypeError()
@staticmethod
CRayStatus UnknownError()
@staticmethod
CRayStatus NotImplemented()
@staticmethod
CRayStatus RedisError()
c_bool ok()
c_bool IsOutOfMemory()
c_bool IsKeyError()
c_bool IsInvalid()
c_bool IsIOError()
c_bool IsTypeError()
c_bool IsUnknownError()
c_bool IsNotImplemented()
c_bool IsRedisError()
c_string ToString()
c_string CodeAsString()
StatusCode code()
c_string message()
# We can later add more of the common status factory methods as needed
cdef CRayStatus RayStatus_OK "Status::OK"()
cdef CRayStatus RayStatus_Invalid "Status::Invalid"()
cdef extern from "ray/status.h" namespace "ray::StatusCode" nogil:
cdef StatusCode StatusCode_OK "OK"
cdef StatusCode StatusCode_OutOfMemory "OutOfMemory"
cdef StatusCode StatusCode_KeyError "KeyError"
cdef StatusCode StatusCode_TypeError "TypeError"
cdef StatusCode StatusCode_Invalid "Invalid"
cdef StatusCode StatusCode_IOError "IOError"
cdef StatusCode StatusCode_UnknownError "UnknownError"
cdef StatusCode StatusCode_NotImplemented "NotImplemented"
cdef StatusCode StatusCode_RedisError "RedisError"
cdef extern from "ray/id.h" namespace "ray" nogil:
const CTaskID FinishTaskId(const CTaskID &task_id)
const CObjectID ComputeReturnId(const CTaskID &task_id,
int64_t return_index)
const CObjectID ComputePutId(const CTaskID &task_id, int64_t put_index)
const CTaskID ComputeTaskId(const CObjectID &object_id)
const CTaskID GenerateTaskId(const CDriverID &driver_id,
const CTaskID &parent_task_id,
int parent_task_counter)
int64_t ComputeObjectIndex(const CObjectID &object_id)
cdef extern from "ray/gcs/format/gcs_generated.h" nogil:
cdef cppclass GCSArg "Arg":
pass
cdef cppclass CLanguage "Language":
pass
# This is a workaround for C++ enum class since Cython has no corresponding representation.
cdef extern from "ray/gcs/format/gcs_generated.h" namespace "Language" nogil:
cdef CLanguage LANGUAGE_PYTHON "Language::PYTHON"
cdef CLanguage LANGUAGE_CPP "Language::CPP"
cdef CLanguage LANGUAGE_JAVA "Language::JAVA"
cdef extern from "ray/raylet/scheduling_resources.h" namespace "ray::raylet" nogil:
cdef cppclass ResourceSet "ResourceSet":
ResourceSet()
ResourceSet(const unordered_map[c_string, double] &resource_map)
ResourceSet(const c_vector[c_string] &resource_labels, const c_vector[double] resource_capacity)
c_bool operator==(const ResourceSet &rhs) const
c_bool IsEqual(const ResourceSet &other) const
c_bool IsSubset(const ResourceSet &other) const
c_bool IsSuperset(const ResourceSet &other) const
c_bool AddResource(const c_string &resource_name, double capacity)
c_bool RemoveResource(const c_string &resource_name)
c_bool AddResourcesStrict(const ResourceSet &other)
void AddResources(const ResourceSet &other)
c_bool SubtractResourcesStrict(const ResourceSet &other)
c_bool GetResource(const c_string &resource_name, double *value) const
double GetNumCpus() const
c_bool IsEmpty() const
const unordered_map[c_string, double] &GetResourceMap() const
const c_string ToString() const
+64
View File
@@ -0,0 +1,64 @@
from libc.stdint cimport int64_t
from libcpp cimport bool as c_bool
from libcpp.memory cimport unique_ptr
from libcpp.string cimport string as c_string
from libcpp.utility cimport pair
from libcpp.unordered_map cimport unordered_map
from libcpp.vector cimport vector as c_vector
from ray.includes.common cimport (
CUniqueID, CTaskID, CObjectID, CFunctionID, CActorClassID, CActorID,
CActorHandleID, CWorkerID, CDriverID, CConfigID, CClientID,
CLanguage, CRayStatus)
from ray.includes.task cimport CTaskSpecification
cdef extern from "ray/gcs/format/gcs_generated.h" nogil:
cdef cppclass GCSProfileEventT "ProfileEventT":
c_string event_type
double start_time
double end_time
c_string extra_data
GCSProfileEventT()
cdef cppclass GCSProfileTableDataT "ProfileTableDataT":
c_string component_type
c_string component_id
c_string node_ip_address
c_vector[unique_ptr[GCSProfileEventT]] profile_events
GCSProfileTableDataT()
ctypedef unordered_map[c_string, c_vector[pair[int64_t, double]]] ResourceMappingType
ctypedef pair[c_vector[CObjectID], c_vector[CObjectID]] WaitResultPair
cdef extern from "ray/raylet/raylet_client.h" nogil:
cdef cppclass CRayletClient "RayletClient":
CRayletClient(const c_string &raylet_socket,
const CClientID &client_id,
c_bool is_worker, const CDriverID &driver_id,
const CLanguage &language)
CRayStatus Disconnect()
CRayStatus SubmitTask(const c_vector[CObjectID] &execution_dependencies,
const CTaskSpecification &task_spec)
CRayStatus GetTask(unique_ptr[CTaskSpecification] *task_spec)
CRayStatus TaskDone()
CRayStatus FetchOrReconstruct(c_vector[CObjectID] &object_ids,
c_bool fetch_only,
const CTaskID &current_task_id)
CRayStatus NotifyUnblocked(const CTaskID &current_task_id)
CRayStatus Wait(const c_vector[CObjectID] &object_ids, int num_returns,
int64_t timeout_milliseconds, c_bool wait_local,
const CTaskID &current_task_id, WaitResultPair *result)
CRayStatus PushError(const CDriverID &job_id, const c_string &type,
const c_string &error_message, double timestamp)
CRayStatus PushProfileEvents(const GCSProfileTableDataT &profile_events)
CRayStatus FreeObjects(const c_vector[CObjectID] &object_ids,
c_bool local_only)
CLanguage GetLanguage() const
CClientID GetClientID() const
CDriverID GetDriverID() const
c_bool IsWorker() const
const ResourceMappingType &GetResourceIDs() const
+83
View File
@@ -0,0 +1,83 @@
from libc.stdint cimport int64_t, uint64_t
from libcpp.string cimport string as c_string
from libcpp.unordered_map cimport unordered_map
cdef extern from "ray/ray_config.h" nogil:
cdef cppclass RayConfig "RayConfig":
@staticmethod
RayConfig &instance()
int64_t ray_protocol_version() const
int64_t handler_warning_timeout_ms() const
int64_t heartbeat_timeout_milliseconds() const
int64_t debug_dump_period_milliseconds() const
int64_t num_heartbeats_timeout() const
uint64_t num_heartbeats_warning() const
int64_t initial_reconstruction_timeout_milliseconds() const
int64_t get_timeout_milliseconds() const
uint64_t max_lineage_size() const
int64_t worker_get_request_size() const
int64_t worker_fetch_request_size() const
int64_t actor_max_dummy_objects() const
int64_t num_connect_attempts() const
int64_t connect_timeout_milliseconds() const
int64_t local_scheduler_fetch_timeout_milliseconds() const
int64_t local_scheduler_reconstruction_timeout_milliseconds() const
int64_t max_num_to_reconstruct() const
int64_t local_scheduler_fetch_request_size() const
int64_t kill_worker_timeout_milliseconds() const
int64_t max_time_for_handler_milliseconds() const
int64_t size_limit() const
int64_t num_elements_limit() const
int64_t max_time_for_loop() const
int64_t redis_db_connect_retries()
int64_t redis_db_connect_wait_milliseconds() const
int64_t plasma_default_release_delay() const
int64_t L3_cache_size_bytes() const
int64_t max_tasks_to_spillback() const
int64_t actor_creation_num_spillbacks_warning() const
int node_manager_forward_task_retry_timeout_milliseconds() const
int object_manager_pull_timeout_ms() const
int object_manager_push_timeout_ms() const
int object_manager_repeated_push_delay_ms() const
uint64_t object_manager_default_chunk_size() const
int num_workers_per_process() const
int64_t max_task_lease_timeout_ms() const
void initialize(const unordered_map[c_string, int] &config_map)
+146
View File
@@ -0,0 +1,146 @@
from ray.includes.ray_config cimport RayConfig
cdef class Config:
@staticmethod
def ray_protocol_version():
return RayConfig.instance().ray_protocol_version()
@staticmethod
def handler_warning_timeout_ms():
return RayConfig.instance().handler_warning_timeout_ms()
@staticmethod
def heartbeat_timeout_milliseconds():
return RayConfig.instance().heartbeat_timeout_milliseconds()
@staticmethod
def debug_dump_period_milliseconds():
return RayConfig.instance().debug_dump_period_milliseconds()
@staticmethod
def num_heartbeats_timeout():
return RayConfig.instance().num_heartbeats_timeout()
@staticmethod
def num_heartbeats_warning():
return RayConfig.instance().num_heartbeats_warning()
@staticmethod
def initial_reconstruction_timeout_milliseconds():
return RayConfig.instance().initial_reconstruction_timeout_milliseconds()
@staticmethod
def get_timeout_milliseconds():
return RayConfig.instance().get_timeout_milliseconds()
@staticmethod
def max_lineage_size():
return RayConfig.instance().max_lineage_size()
@staticmethod
def worker_get_request_size():
return RayConfig.instance().worker_get_request_size()
@staticmethod
def worker_fetch_request_size():
return RayConfig.instance().worker_fetch_request_size()
@staticmethod
def actor_max_dummy_objects():
return RayConfig.instance().actor_max_dummy_objects()
@staticmethod
def num_connect_attempts():
return RayConfig.instance().num_connect_attempts()
@staticmethod
def connect_timeout_milliseconds():
return RayConfig.instance().connect_timeout_milliseconds()
@staticmethod
def local_scheduler_fetch_timeout_milliseconds():
return RayConfig.instance().local_scheduler_fetch_timeout_milliseconds()
@staticmethod
def local_scheduler_reconstruction_timeout_milliseconds():
return RayConfig.instance().local_scheduler_reconstruction_timeout_milliseconds()
@staticmethod
def max_num_to_reconstruct():
return RayConfig.instance().max_num_to_reconstruct()
@staticmethod
def local_scheduler_fetch_request_size():
return RayConfig.instance().local_scheduler_fetch_request_size()
@staticmethod
def kill_worker_timeout_milliseconds():
return RayConfig.instance().kill_worker_timeout_milliseconds()
@staticmethod
def max_time_for_handler_milliseconds():
return RayConfig.instance().max_time_for_handler_milliseconds()
@staticmethod
def size_limit():
return RayConfig.instance().size_limit()
@staticmethod
def num_elements_limit():
return RayConfig.instance().num_elements_limit()
@staticmethod
def max_time_for_loop():
return RayConfig.instance().max_time_for_loop()
@staticmethod
def redis_db_connect_retries():
return RayConfig.instance().redis_db_connect_retries()
@staticmethod
def redis_db_connect_wait_milliseconds():
return RayConfig.instance().redis_db_connect_wait_milliseconds()
@staticmethod
def plasma_default_release_delay():
return RayConfig.instance().plasma_default_release_delay()
@staticmethod
def L3_cache_size_bytes():
return RayConfig.instance().L3_cache_size_bytes()
@staticmethod
def max_tasks_to_spillback():
return RayConfig.instance().max_tasks_to_spillback()
@staticmethod
def actor_creation_num_spillbacks_warning():
return RayConfig.instance().actor_creation_num_spillbacks_warning()
@staticmethod
def node_manager_forward_task_retry_timeout_milliseconds():
return RayConfig.instance().node_manager_forward_task_retry_timeout_milliseconds()
@staticmethod
def object_manager_pull_timeout_ms():
return RayConfig.instance().object_manager_pull_timeout_ms()
@staticmethod
def object_manager_push_timeout_ms():
return RayConfig.instance().object_manager_push_timeout_ms()
@staticmethod
def object_manager_repeated_push_delay_ms():
return RayConfig.instance().object_manager_repeated_push_delay_ms()
@staticmethod
def object_manager_default_chunk_size():
return RayConfig.instance().object_manager_default_chunk_size()
@staticmethod
def num_workers_per_process():
return RayConfig.instance().num_workers_per_process()
@staticmethod
def max_task_lease_timeout_ms():
return RayConfig.instance().max_task_lease_timeout_ms()
+102
View File
@@ -0,0 +1,102 @@
from libc.stdint cimport int64_t, uint8_t
from libcpp cimport bool as c_bool
from libcpp.memory cimport unique_ptr, shared_ptr
from libcpp.string cimport string as c_string
from libcpp.unordered_map cimport unordered_map
from libcpp.vector cimport vector as c_vector
from ray.includes.common cimport (
CUniqueID, CTaskID, CObjectID, CFunctionID, CActorClassID, CActorID,
CActorHandleID, CWorkerID, CDriverID, CConfigID, CClientID,
CLanguage, ResourceSet)
cdef extern from "ray/raylet/task_execution_spec.h" namespace "ray::raylet" nogil:
cdef cppclass CTaskExecutionSpecification "ray::raylet::TaskExecutionSpecification":
CTaskExecutionSpecification(const c_vector[CObjectID] &&dependencies)
CTaskExecutionSpecification(const c_vector[CObjectID] &&dependencies, int num_forwards)
c_vector[CObjectID] ExecutionDependencies() const
void SetExecutionDependencies(const c_vector[CObjectID] &dependencies)
int NumForwards() const
void IncrementNumForwards()
int64_t LastTimestamp() const
void SetLastTimestamp(int64_t new_timestamp)
cdef extern from "ray/raylet/task_spec.h" namespace "ray::raylet" nogil:
cdef cppclass CTaskArgument "ray::raylet::TaskArgument":
pass
cdef cppclass CTaskArgumentByReference "ray::raylet::TaskArgumentByReference":
CTaskArgumentByReference(const c_vector[CObjectID] &references);
cdef cppclass CTaskArgumentByValue "ray::raylet::TaskArgumentByValue":
CTaskArgumentByValue(const uint8_t *value, size_t length);
cdef cppclass CTaskSpecification "ray::raylet::TaskSpecification":
CTaskSpecification(const CDriverID &driver_id, const CTaskID &parent_task_id,
int64_t parent_counter,
const c_vector[shared_ptr[CTaskArgument]] &task_arguments,
int64_t num_returns,
const unordered_map[c_string, double] &required_resources,
const CLanguage &language,
const c_vector[c_string] &function_descriptor)
CTaskSpecification(
const CDriverID &driver_id, const CTaskID &parent_task_id, int64_t parent_counter,
const CActorID &actor_creation_id, const CObjectID &actor_creation_dummy_object_id,
int64_t max_actor_reconstructions, const CActorID &actor_id,
const CActorHandleID &actor_handle_id, int64_t actor_counter,
const c_vector[CActorHandleID] &new_actor_handles,
const c_vector[shared_ptr[CTaskArgument]] &task_arguments,
int64_t num_returns,
const unordered_map[c_string, double] &required_resources,
const unordered_map[c_string, double] &required_placement_resources,
const CLanguage &language, const c_vector[c_string] &function_descriptor)
CTaskSpecification(const c_string &string)
c_string SerializeAsString() const
CTaskID TaskId() const
CDriverID DriverId() const
CTaskID ParentTaskId() const
int64_t ParentCounter() const
c_vector[c_string] FunctionDescriptor() const
c_string FunctionDescriptorString() const
int64_t NumArgs() const
int64_t NumReturns() const
c_bool ArgByRef(int64_t arg_index) const
int ArgIdCount(int64_t arg_index) const
CObjectID ArgId(int64_t arg_index, int64_t id_index) const
CObjectID ReturnId(int64_t return_index) const
const uint8_t *ArgVal(int64_t arg_index) const
size_t ArgValLength(int64_t arg_index) const
double GetRequiredResource(const c_string &resource_name) const
const ResourceSet GetRequiredResources() const
const ResourceSet GetRequiredPlacementResources() const
c_bool IsDriverTask() const
CLanguage GetLanguage() const
c_bool IsActorCreationTask() const
c_bool IsActorTask() const
CActorID ActorCreationId() const
CObjectID ActorCreationDummyObjectId() const
int64_t MaxActorReconstructions() const
CActorID ActorId() const
CActorHandleID ActorHandleId() const
int64_t ActorCounter() const
CObjectID ActorDummyObject() const
c_vector[CActorHandleID] NewActorHandles() const
cdef extern from "ray/raylet/task.h" namespace "ray::raylet" nogil:
cdef cppclass CTask "ray::raylet::Task":
CTask(const CTaskExecutionSpecification &execution_spec,
const CTaskSpecification &task_spec)
const CTaskExecutionSpecification &GetTaskExecutionSpec() const
const CTaskSpecification &GetTaskSpecification() const
void SetExecutionDependencies(const c_vector[CObjectID] &dependencies)
void IncrementNumForwards()
const c_vector[CObjectID] &GetDependencies() const
void CopyTaskExecutionSpec(const CTask &task)
cdef c_string SerializeTaskAsString(const c_vector[CObjectID] *dependencies,
const CTaskSpecification *task_spec)
+185
View File
@@ -0,0 +1,185 @@
from libc.stdint cimport uint8_t
from libcpp.memory cimport shared_ptr, make_shared, static_pointer_cast
from ray.includes.task cimport CTaskSpecification, CTaskArgument, CTaskArgumentByValue, CTaskArgumentByReference, SerializeTaskAsString
from ray.utils import _random_string
cdef class Task:
cdef:
unique_ptr[CTaskSpecification] task_spec
unique_ptr[c_vector[CObjectID]] execution_dependencies
def __init__(self, DriverID driver_id, function_descriptor, arguments,
int num_returns, TaskID parent_task_id, int parent_counter,
ActorID actor_creation_id,
ObjectID actor_creation_dummy_object_id,
int32_t max_actor_reconstructions, ActorID actor_id,
ActorHandleID actor_handle_id, int actor_counter,
new_actor_handles, execution_arguments, resource_map,
placement_resource_map):
cdef:
unordered_map[c_string, double] required_resources
unordered_map[c_string, double] required_placement_resources
c_vector[shared_ptr[CTaskArgument]] task_args
c_vector[CActorHandleID] task_new_actor_handles
c_vector[c_string] c_function_descriptor
c_string pickled_str
c_vector[CObjectID] references
for item in function_descriptor:
if not isinstance(item, bytes):
raise TypeError("'function_descriptor' takes a list of byte strings.")
c_function_descriptor.push_back(item)
# Parse the resource map.
if resource_map is not None:
required_resources = resource_map_from_python_dict(resource_map)
if required_resources.count(b"CPU") == 0:
required_resources[b"CPU"] = 1.0
if placement_resource_map is not None:
required_placement_resources = resource_map_from_python_dict(placement_resource_map)
# Parse the arguments from the list.
for arg in arguments:
if isinstance(arg, ObjectID):
references = c_vector[CObjectID]()
references.push_back((<ObjectID>arg).data)
task_args.push_back(static_pointer_cast[CTaskArgument, CTaskArgumentByReference](make_shared[CTaskArgumentByReference](references)))
else:
pickled_str = pickle.dumps(arg, protocol=pickle.HIGHEST_PROTOCOL)
task_args.push_back(static_pointer_cast[CTaskArgument, CTaskArgumentByValue](make_shared[CTaskArgumentByValue](<uint8_t *>pickled_str.c_str(), pickled_str.size())))
for new_actor_handle in new_actor_handles:
task_new_actor_handles.push_back((<ActorHandleID?>new_actor_handle).data)
self.task_spec.reset(new CTaskSpecification(
CUniqueID(driver_id.data), parent_task_id.data, parent_counter, actor_creation_id.data,
actor_creation_dummy_object_id.data, max_actor_reconstructions, CUniqueID(actor_id.data),
CUniqueID(actor_handle_id.data), actor_counter, task_new_actor_handles, task_args, num_returns,
required_resources, required_placement_resources, LANGUAGE_PYTHON,
c_function_descriptor))
# Set the task's execution dependencies.
self.execution_dependencies.reset(new c_vector[CObjectID]())
if execution_arguments is not None:
for execution_arg in execution_arguments:
self.execution_dependencies.get().push_back((<ObjectID?>execution_arg).data)
@staticmethod
cdef make(unique_ptr[CTaskSpecification]& task_spec):
cdef Task self = Task.__new__(Task)
self.task_spec.reset(task_spec.release())
# The created task does not include any execution dependencies.
self.execution_dependencies.reset(new c_vector[CObjectID]())
return self
@staticmethod
def from_string(const c_string& task_spec_str):
"""Convert a string to a Ray task specification Python object.
Args:
task_spec_str: String representation of the task specification.
Returns:
Python task specification object.
"""
cdef Task self = Task.__new__(Task)
# TODO(pcm): Use flatbuffers validation here.
self.task_spec.reset(new CTaskSpecification(task_spec_str))
# The created task does not include any execution dependencies.
self.execution_dependencies.reset(new c_vector[CObjectID]())
return self
def to_string(self):
"""Convert a Ray task specification Python object to a string.
Returns:
String representing the task specification.
"""
return self.task_spec.get().SerializeAsString()
def _serialized_raylet_task(self):
return SerializeTaskAsString(self.execution_dependencies.get(), self.task_spec.get())
def driver_id(self):
"""Return the driver ID for this task."""
return DriverID.from_native(self.task_spec.get().DriverId())
def task_id(self):
"""Return the task ID for this task."""
return TaskID.from_native(self.task_spec.get().TaskId())
def parent_task_id(self):
"""Return the task ID of the parent task."""
return TaskID.from_native(self.task_spec.get().ParentTaskId())
def parent_counter(self):
"""Return the parent counter of this task."""
return self.task_spec.get().ParentCounter()
def function_descriptor_list(self):
"""Return the function descriptor for this task."""
cdef c_vector[c_string] function_descriptor = self.task_spec.get().FunctionDescriptor()
results = []
for i in range(function_descriptor.size()):
results.append(function_descriptor[i])
return results
def arguments(self):
"""Return the arguments for the task."""
cdef:
CTaskSpecification *task_spec = self.task_spec.get()
int64_t num_args = task_spec.NumArgs()
int count
arg_list = []
for i in range(num_args):
count = task_spec.ArgIdCount(i)
if count > 0:
assert count == 1
arg_list.append(ObjectID.from_native(task_spec.ArgId(i, 0)))
else:
serialized_str = task_spec.ArgVal(i)[:task_spec.ArgValLength(i)]
obj = pickle.loads(serialized_str)
arg_list.append(obj)
return arg_list
def returns(self):
"""Return the object IDs for the return values of the task."""
cdef CTaskSpecification *task_spec = self.task_spec.get()
return_id_list = []
for i in range(task_spec.NumReturns()):
return_id_list.append(ObjectID.from_native(task_spec.ReturnId(i)))
return return_id_list
def required_resources(self):
"""Return the resource dictionary of the task."""
cdef:
unordered_map[c_string, double] resource_map = self.task_spec.get().GetRequiredResources().GetResourceMap()
c_string resource_name
double resource_value
unordered_map[c_string, double].iterator iterator = resource_map.begin()
required_resources = {}
while iterator != resource_map.end():
resource_name = dereference(iterator).first
py_resource_name = str(resource_name) # bytes for Py2, unicode for Py3
resource_value = dereference(iterator).second
required_resources[py_resource_name] = resource_value
postincrement(iterator)
return required_resources
def actor_creation_id(self):
"""Return the actor creation ID for the task."""
return ActorID.from_native(self.task_spec.get().ActorCreationId())
def actor_creation_dummy_object_id(self):
"""Return the actor creation dummy object ID for the task."""
return ObjectID.from_native(self.task_spec.get().ActorCreationDummyObjectId())
def actor_id(self):
"""Return the actor ID for this task."""
return ActorID.from_native(self.task_spec.get().ActorId())
def actor_counter(self):
"""Return the actor counter for this task."""
return self.task_spec.get().ActorCounter()
+34
View File
@@ -0,0 +1,34 @@
from libcpp cimport bool as c_bool
from libcpp.string cimport string as c_string
from libc.stdint cimport uint8_t
cdef extern from "ray/id.h" namespace "ray" nogil:
cdef cppclass CUniqueID "ray::UniqueID":
CUniqueID()
CUniqueID(const CUniqueID &from_id)
@staticmethod
CUniqueID from_random()
@staticmethod
CUniqueID from_binary(const c_string & binary)
@staticmethod
const CUniqueID nil()
size_t hash() const
c_bool is_nil() const
c_bool operator==(const CUniqueID& rhs) const
c_bool operator!=(const CUniqueID& rhs) const
const uint8_t *data() const
uint8_t *mutable_data();
size_t size() const;
c_string binary() const;
c_string hex() const;
ctypedef CUniqueID TaskID
ctypedef CUniqueID ObjectID
ctypedef CUniqueID FunctionID
ctypedef CUniqueID ActorID
ctypedef CUniqueID ActorClassID
ctypedef CUniqueID ActorHandleID
ctypedef CUniqueID WorkerID
ctypedef CUniqueID DriverID
ctypedef CUniqueID ConfigID
ctypedef CUniqueID ClientID
+280
View File
@@ -0,0 +1,280 @@
"""This is a module for unique IDs in Ray.
We define different types for different IDs for type safety.
See https://github.com/ray-project/ray/issues/3721.
"""
from ray.includes.common cimport (
CUniqueID, CTaskID, CObjectID, CFunctionID, CActorClassID, CActorID,
CActorHandleID, CWorkerID, CDriverID, CConfigID, CClientID,
ComputePutId, ComputeTaskId)
from ray.utils import decode
def check_id(b):
if not isinstance(b, bytes):
raise TypeError("Unsupported type: " + str(type(b)))
if len(b) != kUniqueIDSize:
raise ValueError("ID string needs to have length " + str(kUniqueIDSize))
cdef extern from "ray/constants.h" nogil:
cdef int64_t kUniqueIDSize
cdef int64_t kMaxTaskPuts
cdef class UniqueID:
cdef CUniqueID data
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CUniqueID& cpp_id):
cdef UniqueID self = UniqueID.__new__(UniqueID)
self.data = cpp_id
return self
@classmethod
def from_binary(cls, id_bytes):
if not isinstance(id_bytes, bytes):
raise TypeError("Expect bytes, got " + str(type(id_bytes)))
return cls(id_bytes)
@staticmethod
def nil():
return UniqueID.from_native(CUniqueID.nil())
def __hash__(self):
return self.data.hash()
def is_nil(self):
return self.data.is_nil()
def __eq__(self, other):
return self.binary() == other.binary()
def __ne__(self, other):
return self.binary() != other.binary()
def size(self):
return self.data.size()
def __len__(self):
return self.size()
def binary(self):
return self.data.binary()
def __bytes__(self):
return self.binary()
def hex(self):
return decode(self.data.hex())
def __hex__(self):
return self.hex()
def __repr__(self):
return "UniqueID(" + self.hex() + ")"
def __str__(self):
return self.__repr__()
def __reduce__(self):
return type(self), (self.binary(),)
def redis_shard_hash(self):
# NOTE: The hash function used here must match the one in GetRedisContext in
# src/ray/gcs/tables.h. Changes to the hash function should only be made
# through std::hash in src/common/common.h
return self.data.hash()
cdef class ObjectID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CObjectID& cpp_id):
cdef ObjectID self = ObjectID.__new__(ObjectID)
self.data = cpp_id
return self
@staticmethod
def nil():
return ObjectID.from_native(CObjectID.nil())
def __repr__(self):
return "ObjectID(" + self.hex() + ")"
cdef class TaskID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CTaskID& cpp_id):
cdef TaskID self = TaskID.__new__(TaskID)
self.data = cpp_id
return self
@staticmethod
def nil():
return TaskID.from_native(CTaskID.nil())
def __repr__(self):
return "TaskID(" + self.hex() + ")"
cdef class ClientID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CClientID& cpp_id):
cdef ClientID self = ClientID.__new__(ClientID)
self.data = cpp_id
return self
@staticmethod
def nil():
return ClientID.from_native(CClientID.nil())
def __repr__(self):
return "ClientID(" + self.hex() + ")"
cdef class DriverID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CDriverID& cpp_id):
cdef DriverID self = DriverID.__new__(DriverID)
self.data = cpp_id
return self
@staticmethod
def nil():
return DriverID.from_native(CDriverID.nil())
def __repr__(self):
return "DriverID(" + self.hex() + ")"
cdef class ActorID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CActorID& cpp_id):
cdef ActorID self = ActorID.__new__(ActorID)
self.data = cpp_id
return self
@staticmethod
def nil():
return ActorID.from_native(CActorID.nil())
def __repr__(self):
return "ActorID(" + self.hex() + ")"
cdef class ActorHandleID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CActorHandleID& cpp_id):
cdef ActorHandleID self = ActorHandleID.__new__(ActorHandleID)
self.data = cpp_id
return self
@staticmethod
def nil():
return ActorHandleID.from_native(CActorHandleID.nil())
def __repr__(self):
return "ActorHandleID(" + self.hex() + ")"
cdef class FunctionID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CFunctionID& cpp_id):
cdef FunctionID self = FunctionID.__new__(FunctionID)
self.data = cpp_id
return self
@staticmethod
def nil():
return FunctionID.from_native(CFunctionID.nil())
def __repr__(self):
return "FunctionID(" + self.hex() + ")"
cdef class ActorClassID(UniqueID):
def __init__(self, id):
if not id:
self.data = CUniqueID()
else:
check_id(id)
self.data = CUniqueID.from_binary(id)
@staticmethod
cdef from_native(const CActorClassID& cpp_id):
cdef ActorClassID self = ActorClassID.__new__(ActorClassID)
self.data = cpp_id
return self
@staticmethod
def nil():
return ActorClassID.from_native(CActorClassID.nil())
def __repr__(self):
return "ActorClassID(" + self.hex() + ")"
-1
View File
@@ -2,7 +2,6 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import ray.raylet
import ray.worker
from ray import profiling
+2 -2
View File
@@ -160,9 +160,9 @@ class Monitor(object):
object_table_objects = self.state.object_table()
driver_object_id_bins = set()
for object_id, _ in object_table_objects.items():
task_id_bin = ray.raylet.compute_task_id(object_id).id()
task_id_bin = ray._raylet.compute_task_id(object_id).binary()
if task_id_bin in driver_task_id_bins:
driver_object_id_bins.add(object_id.id())
driver_object_id_bins.add(object_id.binary())
def to_shard_index(id_bin):
return binary_to_object_id(id_bin).redis_shard_hash() % len(
+5 -1
View File
@@ -97,6 +97,10 @@ class Profiler(object):
time.sleep(1)
self.flush_profile_data()
except AttributeError:
# TODO(suquark): It is a bad idea to ignore "AttributeError".
# It has caused some very unexpected behaviors when implementing
# new features (related to AttributeError).
# This is to suppress errors that occur at shutdown.
pass
@@ -120,7 +124,7 @@ class Profiler(object):
component_type = "driver"
self.worker.raylet_client.push_profile_events(
component_type, ray.ObjectID(self.worker.worker_id),
component_type, ray.UniqueID(self.worker.worker_id),
self.worker.node_ip_address, events)
def add_event(self, event):
-13
View File
@@ -1,13 +0,0 @@
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from ray.core.src.ray.raylet.libraylet_library_python import (
Task, RayletClient, ObjectID, check_simple_value, compute_task_id,
task_from_string, task_to_string, _config, RayCommonError)
__all__ = [
"Task", "RayletClient", "ObjectID", "check_simple_value",
"compute_task_id", "task_from_string", "task_to_string",
"start_local_scheduler", "_config", "RayCommonError"
]
+2 -2
View File
@@ -39,14 +39,14 @@ class TaskPool(object):
Assumes obj_id only is one id."""
for worker, obj_id in self.completed():
plasma_id = ray.pyarrow.plasma.ObjectID(obj_id.id())
plasma_id = ray.pyarrow.plasma.ObjectID(obj_id.binary())
(ray.worker.global_worker.raylet_client.fetch_or_reconstruct(
[obj_id], True))
self._fetching.append((worker, obj_id))
remaining = []
for worker, obj_id in self._fetching:
plasma_id = ray.pyarrow.plasma.ObjectID(obj_id.id())
plasma_id = ray.pyarrow.plasma.ObjectID(obj_id.binary())
if ray.worker.global_worker.plasma_client.contains(plasma_id):
yield (worker, obj_id)
else:
+2 -2
View File
@@ -23,13 +23,13 @@ def pin_in_object_store(obj):
obj_id = ray.put(_to_pinnable(obj))
_pinned_objects.append(ray.get(obj_id))
return "{}{}".format(PINNED_OBJECT_PREFIX,
base64.b64encode(obj_id.id()).decode("utf-8"))
base64.b64encode(obj_id.binary()).decode("utf-8"))
def get_pinned_object(pinned_id):
"""Retrieve a pinned object from the object store."""
from ray.raylet import ObjectID
from ray import ObjectID
return _from_pinnable(
ray.get(
+3 -4
View File
@@ -15,7 +15,6 @@ import time
import uuid
import ray.gcs_utils
import ray.raylet
import ray.ray_constants as ray_constants
@@ -67,7 +66,7 @@ def push_error_to_driver(worker,
will be serialized with json and stored in Redis.
"""
if driver_id is None:
driver_id = ray.ObjectID.nil_id()
driver_id = ray.DriverID.nil()
data = {} if data is None else data
worker.raylet_client.push_error(driver_id, error_type, message,
time.time())
@@ -96,7 +95,7 @@ def push_error_to_driver_through_redis(redis_client,
will be serialized with json and stored in Redis.
"""
if driver_id is None:
driver_id = ray.ObjectID.nil_id()
driver_id = ray.DriverID.nil()
data = {} if data is None else data
# Do everything in Python and through the Python Redis client instead
# of through the raylet.
@@ -105,7 +104,7 @@ def push_error_to_driver_through_redis(redis_client,
redis_client.execute_command("RAY.TABLE_APPEND",
ray.gcs_utils.TablePrefix.ERROR_INFO,
ray.gcs_utils.TablePubsub.ERROR_INFO,
driver_id.id(), error_data)
driver_id.binary(), error_data)
def is_cython(obj):
+42 -41
View File
@@ -32,10 +32,9 @@ import ray.serialization as serialization
import ray.services as services
import ray.signature
import ray.tempfile_services as tempfile_services
import ray.raylet
import ray.ray_constants as ray_constants
from ray import import_thread
from ray import ObjectID
from ray import ObjectID, DriverID, ActorID, ActorHandleID, ClientID, TaskID
from ray import profiling
from ray.function_manager import (FunctionActorManager, FunctionDescriptor)
import ray.parameter
@@ -161,7 +160,8 @@ class Worker(object):
self.serialization_context_map = {}
self.function_actor_manager = FunctionActorManager(self)
# Identity of the driver that this worker is processing.
self.task_driver_id = ObjectID.nil_id()
# It is a DriverID.
self.task_driver_id = DriverID.nil()
self._task_context = threading.local()
@property
@@ -182,13 +182,13 @@ class Worker(object):
# If this is running on the main thread, initialize it to
# NIL. The actual value will set when the worker receives
# a task from raylet backend.
self._task_context.current_task_id = ObjectID.nil_id()
self._task_context.current_task_id = TaskID.nil()
else:
# If this is running on a separate thread, then the mapping
# to the current task ID may not be correct. Generate a
# random task ID so that the backend can differentiate
# between different threads.
self._task_context.current_task_id = ObjectID(random_string())
self._task_context.current_task_id = TaskID(random_string())
if getattr(self, '_multithreading_warned', False) is not True:
logger.warning(
"Calling ray.get or ray.wait in a separate thread "
@@ -286,7 +286,7 @@ class Worker(object):
try:
self.plasma_client.put(
value,
object_id=pyarrow.plasma.ObjectID(object_id.id()),
object_id=pyarrow.plasma.ObjectID(object_id.binary()),
memcopy_threads=self.memcopy_threads,
serialization_context=self.get_serialization_context(
self.task_driver_id))
@@ -450,7 +450,7 @@ class Worker(object):
# smaller fetches so as to not block the manager for a prolonged period
# of time in a single call.
plain_object_ids = [
plasma.ObjectID(object_id.id()) for object_id in object_ids
plasma.ObjectID(object_id.binary()) for object_id in object_ids
]
for i in range(0, len(object_ids),
ray._config.worker_fetch_request_size()):
@@ -567,16 +567,16 @@ class Worker(object):
with profiling.profile("submit_task", worker=self):
if actor_id is None:
assert actor_handle_id is None
actor_id = ObjectID.nil_id()
actor_handle_id = ObjectID.nil_id()
actor_id = ActorID.nil()
actor_handle_id = ActorHandleID.nil()
else:
assert actor_handle_id is not None
if actor_creation_id is None:
actor_creation_id = ObjectID.nil_id()
actor_creation_id = ActorID.nil()
if actor_creation_dummy_object_id is None:
actor_creation_dummy_object_id = ObjectID.nil_id()
actor_creation_dummy_object_id = ObjectID.nil()
# Put large or complex arguments that are passed by value in the
# object store first.
@@ -584,7 +584,7 @@ class Worker(object):
for arg in args:
if isinstance(arg, ObjectID):
args_for_local_scheduler.append(arg)
elif ray.raylet.check_simple_value(arg):
elif ray._raylet.check_simple_value(arg):
args_for_local_scheduler.append(arg)
else:
args_for_local_scheduler.append(put(arg))
@@ -625,7 +625,8 @@ class Worker(object):
# Submit the task to local scheduler.
function_descriptor_list = (
function_descriptor.get_function_descriptor_list())
task = ray.raylet.Task(
assert isinstance(driver_id, DriverID)
task = ray._raylet.Task(
driver_id,
function_descriptor_list,
args_for_local_scheduler,
@@ -693,7 +694,7 @@ class Worker(object):
# Run the function on all workers.
self.redis_client.hmset(
key, {
"driver_id": self.task_driver_id.id(),
"driver_id": self.task_driver_id.binary(),
"function_id": function_to_run_id,
"function": pickled_function,
"run_on_other_drivers": str(run_on_other_drivers)
@@ -880,7 +881,7 @@ class Worker(object):
str(failure_object),
driver_id=self.task_driver_id,
data={
"function_id": function_id.id(),
"function_id": function_id.binary(),
"function_name": function_name,
"module_name": function_descriptor.module_name,
"class_name": function_descriptor.class_name
@@ -939,14 +940,14 @@ class Worker(object):
with _changeproctitle(title, next_title):
self._process_task(task, execution_info)
# Reset the state fields so the next task can run.
self.task_context.current_task_id = ObjectID.nil_id()
self.task_context.current_task_id = TaskID.nil()
self.task_context.task_index = 0
self.task_context.put_index = 1
if self.actor_id.is_nil():
# Don't need to reset task_driver_id if the worker is an
# actor. Because the following tasks should all have the
# same driver id.
self.task_driver_id = ObjectID.nil_id()
self.task_driver_id = DriverID.nil()
# Increase the task execution counter.
self.function_actor_manager.increase_task_counter(
@@ -1100,17 +1101,16 @@ def error_applies_to_driver(error_key, worker=global_worker):
+ ray_constants.ID_SIZE), error_key
# If the driver ID in the error message is a sequence of all zeros, then
# the message is intended for all drivers.
driver_id = ObjectID(error_key[len(ERROR_KEY_PREFIX):(
driver_id = DriverID(error_key[len(ERROR_KEY_PREFIX):(
len(ERROR_KEY_PREFIX) + ray_constants.ID_SIZE)])
return (driver_id == worker.task_driver_id
or driver_id == ObjectID.nil_id())
return (driver_id == worker.task_driver_id or driver_id == DriverID.nil())
def error_info(worker=global_worker):
"""Return information about failed tasks."""
worker.check_connected()
return (global_state.error_messages(job_id=worker.task_driver_id) +
global_state.error_messages(job_id=ObjectID.nil_id()))
global_state.error_messages(job_id=DriverID.nil()))
def _initialize_serialization(driver_id, worker=global_worker):
@@ -1127,7 +1127,7 @@ def _initialize_serialization(driver_id, worker=global_worker):
# Define a custom serializer and deserializer for handling Object IDs.
def object_id_custom_serializer(obj):
return obj.id()
return obj.binary()
def object_id_custom_deserializer(serialized_obj):
return ObjectID(serialized_obj)
@@ -1656,8 +1656,8 @@ def listen_error_messages_raylet(worker, task_error_queue):
gcs_entry.Entries(0), 0)
job_id = error_data.JobId()
if job_id not in [
worker.task_driver_id.id(),
ObjectID.nil_id().id()
worker.task_driver_id.binary(),
DriverID.nil().binary()
]:
continue
@@ -1768,23 +1768,23 @@ def connect(info,
else:
# This is the code path of driver mode.
if driver_id is None:
driver_id = ObjectID(random_string())
driver_id = DriverID(random_string())
if not isinstance(driver_id, ObjectID):
raise Exception("The type of given driver id must be ObjectID.")
if not isinstance(driver_id, DriverID):
raise Exception("The type of given driver id must be DriverID.")
worker.worker_id = driver_id.id()
worker.worker_id = driver_id.binary()
# When tasks are executed on remote workers in the context of multiple
# drivers, the task driver ID is used to keep track of which driver is
# responsible for the task so that error messages will be propagated to
# the correct driver.
if mode != WORKER_MODE:
worker.task_driver_id = ObjectID(worker.worker_id)
worker.task_driver_id = DriverID(worker.worker_id)
# All workers start out as non-actors. A worker can be turned into an actor
# after it is created.
worker.actor_id = ObjectID.nil_id()
worker.actor_id = ActorID.nil()
worker.connected = True
worker.set_mode(mode)
@@ -1910,18 +1910,18 @@ def connect(info,
nil_actor_counter = 0
function_descriptor = FunctionDescriptor.for_driver_task()
driver_task = ray.raylet.Task(
driver_task = ray._raylet.Task(
worker.task_driver_id,
function_descriptor.get_function_descriptor_list(),
[], # arguments.
0, # num_returns.
ObjectID(random_string()), # parent_task_id.
TaskID(random_string()), # parent_task_id.
0, # parent_counter.
ObjectID.nil_id(), # actor_creation_id.
ObjectID.nil_id(), # actor_creation_dummy_object_id.
ActorID.nil(), # actor_creation_id.
ObjectID.nil(), # actor_creation_dummy_object_id.
0, # max_actor_reconstructions.
ObjectID.nil_id(), # actor_id.
ObjectID.nil_id(), # actor_handle_id.
ActorID.nil(), # actor_id.
ActorHandleID.nil(), # actor_handle_id.
nil_actor_counter, # actor_counter.
[], # new_actor_handles.
[], # execution_dependencies.
@@ -1933,18 +1933,18 @@ def connect(info,
global_state._execute_command(driver_task.task_id(), "RAY.TABLE_ADD",
ray.gcs_utils.TablePrefix.RAYLET_TASK,
ray.gcs_utils.TablePubsub.RAYLET_TASK,
driver_task.task_id().id(),
driver_task.task_id().binary(),
driver_task._serialized_raylet_task())
# Set the driver's current task ID to the task ID assigned to the
# driver task.
worker.task_context.current_task_id = driver_task.task_id()
worker.raylet_client = ray.raylet.RayletClient(
worker.raylet_client = ray._raylet.RayletClient(
raylet_socket,
worker.worker_id,
ClientID(worker.worker_id),
is_worker,
worker.current_task_id,
DriverID(worker.current_task_id.binary()),
)
# Start the import thread
@@ -2144,6 +2144,7 @@ def register_custom_serializer(cls,
if driver_id is None:
driver_id = worker.task_driver_id
assert isinstance(driver_id, DriverID)
def register_class_for_serialization(worker_info):
# TODO(rkn): We need to be more thoughtful about what to do if custom
@@ -2228,7 +2229,7 @@ def put(value, worker=global_worker):
if worker.mode == LOCAL_MODE:
# In LOCAL_MODE, ray.put is the identity operation.
return value
object_id = worker.raylet_client.compute_put_id(
object_id = ray._raylet.compute_put_id(
worker.current_task_id,
worker.task_context.put_index,
)
+1 -2
View File
@@ -22,8 +22,7 @@ import setuptools.command.build_ext as _build_ext
ray_files = [
"ray/core/src/ray/thirdparty/redis/src/redis-server",
"ray/core/src/ray/gcs/redis_module/libray_redis_module.so",
"ray/core/src/plasma/plasma_store_server",
"ray/core/src/ray/raylet/libraylet_library_python.so",
"ray/core/src/plasma/plasma_store_server", "ray/_raylet.so",
"ray/core/src/ray/raylet/raylet_monitor", "ray/core/src/ray/raylet/raylet",
"ray/WebUI.ipynb"
]