def error_messages(self, job_id=None):
"""Get the error messages for all jobs or a specific job.
Args:
job_id: The specific job to get the errors for. If this is None,
then this method retrieves the errors for all jobs.
Returns:
A dictionary mapping job ID to a list of the error messages for
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(
ray.gcs_utils.TablePrefix_ERROR_INFO_string + "*")
job_ids = [
key[len(ray.gcs_utils.TablePrefix_ERROR_INFO_string):]
for key in error_table_keys
]
return {
binary_to_hex(job_id): self._error_messages(ray.DriverID(job_id))
for job_id in job_ids
}
def error_messages(self, driver_id=None):
"""Get the error messages for all drivers or a specific driver.
Args:
driver_id: The specific driver to get the errors for. If this is
None, then this method retrieves the errors for all drivers.
Returns:
A dictionary mapping driver ID to a list of the error messages for
that driver.
"""
self._check_connected()
if driver_id is not None:
assert isinstance(driver_id, ray.DriverID)
return self._error_messages(driver_id)
error_table_keys = self.redis_client.keys(
ray.gcs_utils.TablePrefix_ERROR_INFO_string + "*")
driver_ids = [
key[len(ray.gcs_utils.TablePrefix_ERROR_INFO_string):]
for key in error_table_keys
]
return {
binary_to_hex(driver_id): self._error_messages(
ray.DriverID(driver_id))
for driver_id in driver_ids
}
def fetch_and_execute_function_to_run(self, key):
"""Run on arbitrary function on the worker."""
(driver_id, serialized_function,
run_on_other_drivers) = self.redis_client.hmget(
key, ["driver_id", "function", "run_on_other_drivers"])
if (utils.decode(run_on_other_drivers) == "False"
and self.worker.mode == ray.SCRIPT_MODE
and driver_id != self.worker.task_driver_id.binary()):
return
try:
# Deserialize the function.
function = pickle.loads(serialized_function)
# Run the function.
function({"worker": self.worker})
except Exception:
# If an exception was thrown when the function was run, we record
# the traceback and notify the scheduler of the failure.
traceback_str = traceback.format_exc()
# Log the error message.
utils.push_error_to_driver(
self.worker,
ray_constants.FUNCTION_TO_RUN_PUSH_ERROR,
traceback_str,
driver_id=ray.DriverID(driver_id))
def _load_actor_class_from_gcs(self, driver_id, function_descriptor):
"""Load actor class from GCS."""
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
# 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.
(driver_id_str, class_name, module, pickled_class,
actor_method_names) = self._worker.redis_client.hmget(
key, [
"driver_id", "class_name", "module", "class",
"actor_method_names"
])
class_name = six.ensure_str(class_name)
module_name = six.ensure_str(module)
driver_id = ray.DriverID(driver_id_str)
actor_method_names = json.loads(six.ensure_str(actor_method_names))
actor_class = None
try:
with self._worker.lock:
actor_class = pickle.loads(pickled_class)
except Exception:
logger.exception(
"Failed to load actor class %s.".format(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.utils.format_error_message(
traceback.format_exc())
# Log the error message.
push_error_to_driver(
self._worker, ray_constants.REGISTER_ACTOR_PUSH_ERROR,
"Failed to unpickle actor class '{}' for actor ID {}. "
"Traceback:\n{}".format(class_name,
self._worker.actor_id.hex(),
traceback_str), driver_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
def fetch_and_register_remote_function(self, key):
"""Import a remote function."""
(driver_id_str, function_id_str, function_name, serialized_function,
num_return_vals, module, resources,
max_calls) = self._worker.redis_client.hmget(key, [
"driver_id", "function_id", "name", "function", "num_return_vals",
"module", "resources", "max_calls"
])
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)
# This is a placeholder in case the function can't be unpickled. This
# will be overwritten if the function is successfully registered.
def f():
raise Exception("This function was not imported properly.")
self._function_execution_info[driver_id][function_id] = (
FunctionExecutionInfo(
function=f, function_name=function_name, max_calls=max_calls))
self._num_task_executions[driver_id][function_id] = 0
try:
function = pickle.loads(serialized_function)
except Exception:
# If an exception was thrown when the remote function was imported,
# we record the traceback and notify the scheduler of the failure.
traceback_str = format_error_message(traceback.format_exc())
# Log the error message.
push_error_to_driver(
self._worker,
ray_constants.REGISTER_REMOTE_FUNCTION_PUSH_ERROR,
traceback_str,
driver_id=driver_id,
data={
"function_id": function_id.binary(),
"function_name": function_name
})
else:
# 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`
function.__module__ = module
self._function_execution_info[driver_id][function_id] = (
FunctionExecutionInfo(
function=function,
function_name=function_name,
max_calls=max_calls))
# Add the function to the function table.
self._worker.redis_client.rpush(
b"FunctionTable:" + function_id.binary(),
self._worker.worker_id)
def fetch_and_register_actor(self, actor_class_key):
"""Import an actor.
This will be called by the worker's import thread when the worker
receives the actor_class export, assuming that the worker is an actor
for that class.
Args:
actor_class_key: The key in Redis to use to fetch the actor.
"""
actor_id = self._worker.actor_id
(driver_id_str, class_name, module, pickled_class,
actor_method_names) = self._worker.redis_client.hmget(
actor_class_key, [
"driver_id", "class_name", "module", "class",
"actor_method_names"
])
class_name = decode(class_name)
module = decode(module)
driver_id = ray.DriverID(driver_id_str)
actor_method_names = json.loads(decode(actor_method_names))
# In Python 2, json loads strings as unicode, so convert them back to
# strings.
if sys.version_info < (3, 0):
actor_method_names = [
method_name.encode("ascii")
for method_name in actor_method_names
]
# Create a temporary actor with some temporary methods so that if
# the actor fails to be unpickled, the temporary actor can be used
# (just to produce error messages and to prevent the driver from
# hanging).
class TemporaryActor(object):
pass
self._worker.actors[actor_id] = TemporaryActor()
def temporary_actor_method(*xs):
raise Exception(
"The actor with name {} failed to be imported, "
"and so cannot execute this method".format(class_name))
# Register the actor method executors.
for actor_method_name in actor_method_names:
function_descriptor = FunctionDescriptor(module, actor_method_name,
class_name)
function_id = function_descriptor.function_id
temporary_executor = self._make_actor_method_executor(
actor_method_name,
temporary_actor_method,
actor_imported=False)
self._function_execution_info[driver_id][function_id] = (
FunctionExecutionInfo(function=temporary_executor,
function_name=actor_method_name,
max_calls=0))
self._num_task_executions[driver_id][function_id] = 0
try:
unpickled_class = pickle.loads(pickled_class)
self._worker.actor_class = unpickled_class
except Exception:
# If an exception was thrown when the actor was imported, we record
# the traceback and notify the scheduler of the failure.
traceback_str = ray.utils.format_error_message(
traceback.format_exc())
# Log the error message.
push_error_to_driver(
self._worker, ray_constants.REGISTER_ACTOR_PUSH_ERROR,
"Failed to unpickle actor class '{}' for actor ID {}. "
"Traceback:\n{}".format(class_name, actor_id.hex(),
traceback_str), driver_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.
else:
# TODO(pcm): Why is the below line necessary?
unpickled_class.__module__ = module
self._worker.actors[actor_id] = unpickled_class.__new__(
unpickled_class)
actor_methods = inspect.getmembers(unpickled_class,
predicate=is_function_or_method)
for actor_method_name, actor_method in actor_methods:
function_descriptor = FunctionDescriptor(
module, actor_method_name, class_name)
function_id = function_descriptor.function_id
executor = self._make_actor_method_executor(
actor_method_name, actor_method, actor_imported=True)
self._function_execution_info[driver_id][function_id] = (
FunctionExecutionInfo(function=executor,
function_name=actor_method_name,
max_calls=0))
