def workers(self):
"""Get a dictionary mapping worker ID to worker information."""
self._check_connected()
# Get all data in worker table
worker_table = self.global_state_accessor.get_worker_table()
workers_data = {}
for i in range(len(worker_table)):
worker_table_data = gcs_utils.WorkerTableData.FromString(
worker_table[i])
if worker_table_data.is_alive and \
worker_table_data.worker_type == gcs_utils.WORKER:
worker_id = binary_to_hex(
worker_table_data.worker_address.worker_id)
worker_info = worker_table_data.worker_info
workers_data[worker_id] = {
"node_ip_address":
decode(worker_info[b"node_ip_address"]),
"plasma_store_socket":
decode(worker_info[b"plasma_store_socket"])
}
if b"stderr_file" in worker_info:
workers_data[worker_id]["stderr_file"] = decode(
worker_info[b"stderr_file"])
if b"stdout_file" in worker_info:
workers_data[worker_id]["stdout_file"] = decode(
worker_info[b"stdout_file"])
return workers_data
def fetch_and_register_remote_function(self, key):
"""Import a remote function."""
(job_id_str, function_id_str, function_name, serialized_function,
module,
max_calls) = self._worker.redis_client.hmget(key, [
"job_id", "function_id", "function_name", "function", "module",
"max_calls"
])
function_id = ray.FunctionID(function_id_str)
job_id = ray.JobID(job_id_str)
function_name = decode(function_name)
max_calls = int(max_calls)
module = decode(module)
# This function is called by ImportThread. This operation needs to be
# atomic. Otherwise, there is race condition. Another thread may use
# the temporary function above before the real function is ready.
with self.lock:
self._num_task_executions[job_id][function_id] = 0
try:
function = pickle.loads(serialized_function)
except Exception:
def f(*args, **kwargs):
raise RuntimeError(
"This function was not imported properly.")
# Use a placeholder method when function pickled failed
self._function_execution_info[job_id][function_id] = (
FunctionExecutionInfo(function=f,
function_name=function_name,
max_calls=max_calls))
# 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.
logger.debug("Failed to unpickle the remote function "
f"'{function_name}' with "
f"function ID {function_id.hex()}. "
f"Job ID:{job_id}."
f"Traceback:\n{traceback_str}. ")
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[job_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_remote_function(self, key):
"""Import a remote function."""
(job_id_str, function_id_str, function_name, serialized_function,
module, max_calls) = self._worker.redis_client.hmget(
key, [
"job_id", "function_id", "function_name", "function",
"module", "max_calls"
])
if ray_constants.ISOLATE_EXPORTS and \
job_id_str != self._worker.current_job_id.binary():
# A worker only executes tasks from the assigned job.
# TODO(jjyao): If fetching unrelated remote functions
# becomes a perf issue, we can also consider having export
# queue per job.
return
function_id = ray.FunctionID(function_id_str)
job_id = ray.JobID(job_id_str)
function_name = decode(function_name)
max_calls = int(max_calls)
module = decode(module)
# This function is called by ImportThread. This operation needs to be
# atomic. Otherwise, there is race condition. Another thread may use
# the temporary function above before the real function is ready.
with self.lock:
self._num_task_executions[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())
def f(*args, **kwargs):
raise RuntimeError(
"The remote function failed to import on the "
"worker. This may be because needed library "
"dependencies are not installed in the worker "
"environment:\n\n{}".format(traceback_str))
# Use a placeholder method when function pickled failed
self._function_execution_info[function_id] = (
FunctionExecutionInfo(
function=f,
function_name=function_name,
max_calls=max_calls))
# Log the error message. Log at DEBUG level to avoid overly
# spamming the log on import failure. The user gets the error
# via the RuntimeError message above.
logger.debug("Failed to unpickle the remote function "
f"'{function_name}' with "
f"function ID {function_id.hex()}. "
f"Job ID:{job_id}."
f"Traceback:\n{traceback_str}. ")
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[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)
