async def invoke_single(self, request_item: Query) -> Any:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id))
args, kwargs = parse_request_item(request_item)
start = time.time()
method_to_call = None
try:
method_to_call = sync_to_async(
self.get_runner_method(request_item))
result = await method_to_call(*args, **kwargs)
result = await self.ensure_serializable_response(result)
self.request_counter.inc()
except Exception as e:
import os
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
function_name = "unknown"
if method_to_call is not None:
function_name = method_to_call.__name__
result = wrap_to_ray_error(function_name, e)
self.error_counter.inc()
latency_ms = (time.time() - start) * 1000
self.processing_latency_tracker.observe(latency_ms)
return result
async def invoke_single(self, request_item: Query) -> Any:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id))
args, kwargs = parse_request_item(request_item)
start = time.time()
method_to_call = None
try:
runner_method = self.get_runner_method(request_item)
method_to_call = sync_to_async(runner_method)
result = None
if len(inspect.signature(runner_method).parameters) > 0:
result = await method_to_call(*args, **kwargs)
else:
# The method doesn't take in anything, including the request
# information, so we pass nothing into it
result = await method_to_call()
result = await self.ensure_serializable_response(result)
self.request_counter.inc()
except Exception as e:
import os
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
function_name = "unknown"
if method_to_call is not None:
function_name = method_to_call.__name__
result = wrap_to_ray_error(function_name, e)
self.error_counter.inc()
latency_ms = (time.time() - start) * 1000
self.processing_latency_tracker.observe(latency_ms)
return result
async def invoke_single(self, request_item: Query) -> Any:
args, kwargs, is_web_context = parse_request_item(request_item)
serve_context.web = is_web_context
method_to_call = self.get_runner_method(request_item)
args = args if self.has_positional_args(method_to_call) else []
method_to_call = ensure_async(method_to_call)
start = time.time()
try:
result = await method_to_call(*args, **kwargs)
self.request_counter.record(1, {"backend": self.backend_tag})
except Exception as e:
result = wrap_to_ray_error(e)
self.error_counter.record(1, {"backend": self.backend_tag})
finally:
self._reset_context()
self.processing_latency_tracker.record(
(time.time() - start) * 1000, {
"backend": self.backend_tag,
"replica": self.replica_tag,
"batch_size": "1"
})
return result
async def invoke_batch(self, request_item_list: List[Query]) -> List[Any]:
args = []
call_methods = set()
batch_size = len(request_item_list)
# Construct the batch of requests
for item in request_item_list:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, item.metadata.request_id))
args.append(parse_request_item(item))
call_methods.add(self.get_runner_method(item))
timing_start = time.time()
try:
if len(call_methods) != 1:
raise RayServeException(
f"Queries contain mixed calling methods: {call_methods}. "
"Please only send the same type of requests in batching "
"mode.")
self.request_counter.inc(batch_size)
call_method = sync_to_async(call_methods.pop())
result_list = await call_method(args)
if not isinstance(result_list, Iterable) or isinstance(
result_list, (dict, set)):
error_message = ("RayServe expects an ordered iterable object "
"but the replica returned a {}".format(
type(result_list)))
raise RayServeException(error_message)
# Normalize the result into a list type. This operation is fast
# in Python because it doesn't copy anything.
result_list = list(result_list)
if (len(result_list) != batch_size):
error_message = ("Worker doesn't preserve batch size. The "
"input has length {} but the returned list "
"has length {}. Please return a list of "
"results with length equal to the batch size"
".".format(batch_size, len(result_list)))
raise RayServeException(error_message)
for i, result in enumerate(result_list):
result_list[i] = (await
self.ensure_serializable_response(result))
except Exception as e:
wrapped_exception = wrap_to_ray_error(call_method.__name__, e)
self.error_counter.inc()
result_list = [wrapped_exception for _ in range(batch_size)]
latency_ms = (time.time() - timing_start) * 1000
self.processing_latency_tracker.observe(
latency_ms, tags={"batch_size": str(batch_size)})
return result_list
async def invoke_batch(self, request_item_list: List[Query]) -> List[Any]:
args = []
call_methods = set()
batch_size = len(request_item_list)
# Construct the batch of requests
for item in request_item_list:
args.append(parse_request_item(item))
call_methods.add(self.get_runner_method(item))
timing_start = time.time()
try:
if len(call_methods) != 1:
raise RayServeException(
f"Queries contain mixed calling methods: {call_methods}. "
"Please only send the same type of requests in batching "
"mode.")
self.request_counter.record(batch_size,
{"backend": self.backend_tag})
call_method = ensure_async(call_methods.pop())
result_list = await call_method(args)
if not isinstance(result_list, Iterable) or isinstance(
result_list, (dict, set)):
error_message = ("RayServe expects an ordered iterable object "
"but the worker returned a {}".format(
type(result_list)))
raise RayServeException(error_message)
# Normalize the result into a list type. This operation is fast
# in Python because it doesn't copy anything.
result_list = list(result_list)
if (len(result_list) != batch_size):
error_message = ("Worker doesn't preserve batch size. The "
"input has length {} but the returned list "
"has length {}. Please return a list of "
"results with length equal to the batch size"
".".format(batch_size, len(result_list)))
raise RayServeException(error_message)
except Exception as e:
wrapped_exception = wrap_to_ray_error(e)
self.error_counter.record(1, {"backend": self.backend_tag})
result_list = [wrapped_exception for _ in range(batch_size)]
self.processing_latency_tracker.record(
(time.time() - timing_start) * 1000, {
"backend": self.backend_tag,
"replica_tag": self.replica_tag,
"batch_size": str(batch_size)
})
return result_list
def invoke_single(self, request_item):
args, kwargs, is_web_context = parse_request_item(request_item)
serve_context.web = is_web_context
start_timestamp = time.time()
try:
result = self.__call__(*args, **kwargs)
except Exception as e:
result = wrap_to_ray_error(e)
self._serve_metric_error_counter += 1
self._serve_metric_latency_list.append(time.time() - start_timestamp)
return result
async def invoke_single(self, request_item: Query) -> Any:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id))
arg = parse_request_item(request_item)
start = time.time()
method_to_call = None
try:
# TODO(simon): Split this section out when invoke_batch is removed.
if self.config.internal_metadata.is_asgi_app:
request: Request = arg
scope = request.scope
root_path = self.config.internal_metadata.path_prefix
# The incoming scope["path"] contains prefixed path and it
# won't be stripped by FastAPI.
request.scope["path"] = scope["path"].replace(root_path, "", 1)
# root_path is used such that the reverse look up and
# redirection works.
request.scope["root_path"] = root_path
sender = ASGIHTTPSender()
await self.callable._serve_asgi_app(
request.scope,
request._receive,
sender,
)
result = sender.build_starlette_response()
else:
method_to_call = sync_to_async(
self.get_runner_method(request_item))
result = await method_to_call(arg)
result = await self.ensure_serializable_response(result)
self.request_counter.inc()
except Exception as e:
import os
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
function_name = "unknown"
if method_to_call is not None:
function_name = method_to_call.__name__
result = wrap_to_ray_error(function_name, e)
self.error_counter.inc()
latency_ms = (time.time() - start) * 1000
self.processing_latency_tracker.observe(latency_ms,
tags={"batch_size": "1"})
return result
async def invoke_single(self, request_item: Query) -> Tuple[Any, bool]:
"""Executes the provided request on this replica.
Returns the user-provided output and a boolean indicating if the
request succeeded (user code didn't raise an exception).
"""
logger.debug(
"Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id
)
)
args, kwargs = parse_request_item(request_item)
method_to_call = None
success = True
try:
runner_method = self.get_runner_method(request_item)
method_to_call = sync_to_async(runner_method)
result = None
if len(inspect.signature(runner_method).parameters) > 0:
result = await method_to_call(*args, **kwargs)
else:
# When access via http http_arg_is_pickled with no args:
# args = (<starlette.requests.Request object at 0x7fe900694cc0>,)
# When access via python with no args:
# args = ()
if len(args) == 1 and isinstance(args[0], starlette.requests.Request):
# The method doesn't take in anything, including the request
# information, so we pass nothing into it
result = await method_to_call()
else:
# Will throw due to signature mismatch if user attempts to
# call with non-empty args
result = await method_to_call(*args, **kwargs)
result = await self.ensure_serializable_response(result)
self.request_counter.inc()
except Exception as e:
logger.exception(f"Request failed due to {type(e).__name__}:")
success = False
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
function_name = "unknown"
if method_to_call is not None:
function_name = method_to_call.__name__
result = wrap_to_ray_error(function_name, e)
self.error_counter.inc()
return result, success
async def invoke_single(self, request_item):
args, kwargs, is_web_context = parse_request_item(request_item)
serve_context.web = is_web_context
method_to_call = self.get_runner_method(request_item)
args = args if self.has_positional_args(method_to_call) else []
method_to_call = ensure_async(method_to_call)
try:
result = await method_to_call(*args, **kwargs)
self.request_counter.add()
except Exception as e:
result = wrap_to_ray_error(e)
self.error_counter.add()
return result
async def invoke_single(self, request_item: Query) -> Any:
method_to_call = ensure_async(self.get_runner_method(request_item))
arg = parse_request_item(request_item)
start = time.time()
try:
result = await method_to_call(arg)
self.request_counter.record(1)
except Exception as e:
result = wrap_to_ray_error(e)
self.error_counter.record(1)
self.processing_latency_tracker.record(
(time.time() - start) * 1000, tags={"batch_size": "1"})
return result
async def invoke_single(self, request_item):
args, kwargs, is_web_context = parse_request_item(request_item)
serve_context.web = is_web_context
start_timestamp = time.time()
method_to_call = self.get_runner_method(request_item)
args = args if self.has_positional_args(method_to_call) else []
method_to_call = ensure_async(method_to_call)
try:
result = await method_to_call(*args, **kwargs)
except Exception as e:
result = wrap_to_ray_error(e)
self.error_counter += 1
self.latency_list.append(time.time() - start_timestamp)
return result
async def invoke_single(self, request_item: Query) -> Any:
args, kwargs, is_web_context = parse_request_item(request_item)
serve_context.web = is_web_context
method_to_call = self.get_runner_method(request_item)
args = args if self.has_positional_args(method_to_call) else []
method_to_call = ensure_async(method_to_call)
try:
result = await method_to_call(*args, **kwargs)
self.request_counter.record(1, {"backend": self.backend_tag})
except Exception as e:
result = wrap_to_ray_error(e)
self.error_counter.record(1, {"backend": self.backend_tag})
finally:
self._reset_context()
return result
async def invoke_single(self, request_item: Query) -> Any:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id))
method_to_call = ensure_async(self.get_runner_method(request_item))
arg = parse_request_item(request_item)
start = time.time()
try:
result = await method_to_call(arg)
self.request_counter.record(1)
except Exception as e:
import os
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
result = wrap_to_ray_error(e)
self.error_counter.record(1)
latency_ms = (time.time() - start) * 1000
self.processing_latency_tracker.record(
latency_ms, tags={"batch_size": "1"})
return result
async def invoke_single(self, request_item: Query) -> Any:
logger.debug("Replica {} started executing request {}".format(
self.replica_tag, request_item.metadata.request_id))
args, kwargs = parse_request_item(request_item)
start = time.time()
method_to_call = None
try:
# TODO(simon): Split this section out when invoke_batch is removed.
if self.config.internal_metadata.is_asgi_app:
request: Request = args[0]
sender = ASGIHTTPSender()
await self.callable._serve_asgi_app(
request.scope,
request._receive,
sender,
)
result = sender.build_starlette_response()
else:
method_to_call = sync_to_async(
self.get_runner_method(request_item))
result = await method_to_call(*args, **kwargs)
result = await self.ensure_serializable_response(result)
self.request_counter.inc()
except Exception as e:
import os
if "RAY_PDB" in os.environ:
ray.util.pdb.post_mortem()
function_name = "unknown"
if method_to_call is not None:
function_name = method_to_call.__name__
result = wrap_to_ray_error(function_name, e)
self.error_counter.inc()
latency_ms = (time.time() - start) * 1000
self.processing_latency_tracker.observe(latency_ms)
return result
def invoke_batch(self, request_item_list):
# TODO(alind) : create no-http services. The enqueues
# from such services will always be TaskContext.Python.
# Assumption : all the requests in a bacth
# have same serve context.
# For batching kwargs are modified as follows -
# kwargs [Python Context] : key,val
# kwargs_list : key, [val1,val2, ... , valn]
# or
# args[Web Context] : val
# args_list : [val1,val2, ...... , valn]
# where n (current batch size) <= max_batch_size of a backend
arg_list = []
kwargs_list = defaultdict(list)
context_flags = set()
batch_size = len(request_item_list)
for item in request_item_list:
args, kwargs, is_web_context = parse_request_item(item)
context_flags.add(is_web_context)
if is_web_context:
# Python context only have kwargs
flask_request = args[0]
arg_list.append(flask_request)
else:
# Web context only have one positional argument
for k, v in kwargs.items():
kwargs_list[k].append(v)
# Set the flask request as a list to conform
# with batching semantics: when in batching
# mode, each argument it turned into list.
arg_list.append(FakeFlaskRequest())
try:
# check mixing of query context
# unified context needed
if len(context_flags) != 1:
raise RayServeException(
"Batched queries contain mixed context. Please only send "
"the same type of requests in batching mode.")
serve_context.web = context_flags.pop()
serve_context.batch_size = batch_size
# Flask requests are passed to __call__ as a list
arg_list = [arg_list]
start_timestamp = time.time()
result_list = self.__call__(*arg_list, **kwargs_list)
self._serve_metric_latency_list.append(time.time() -
start_timestamp)
if (not isinstance(result_list,
list)) or (len(result_list) != batch_size):
raise RayServeException("__call__ function "
"doesn't preserve batch-size. "
"Please return a list of result "
"with length equals to the batch "
"size.")
return result_list
except Exception as e:
wrapped_exception = wrap_to_ray_error(e)
self._serve_metric_error_counter += batch_size
return [wrapped_exception for _ in range(batch_size)]
async def invoke_batch(self, request_item_list):
arg_list = []
kwargs_list = defaultdict(list)
context_flags = set()
batch_size = len(request_item_list)
call_methods = set()
for item in request_item_list:
args, kwargs, is_web_context = parse_request_item(item)
context_flags.add(is_web_context)
call_method = self.get_runner_method(item)
call_methods.add(call_method)
if is_web_context:
# Python context only have kwargs
flask_request = args[0]
arg_list.append(flask_request)
else:
# Web context only have one positional argument
for k, v in kwargs.items():
kwargs_list[k].append(v)
# Set the flask request as a list to conform
# with batching semantics: when in batching
# mode, each argument is turned into list.
if self.has_positional_args(call_method):
arg_list.append(FakeFlaskRequest())
try:
# Check mixing of query context (unified context needed).
if len(context_flags) != 1:
raise RayServeException(
"Batched queries contain mixed context. Please only send "
"the same type of requests in batching mode.")
serve_context.web = context_flags.pop()
if len(call_methods) != 1:
raise RayServeException(
"Queries contain mixed calling methods. Please only send "
"the same type of requests in batching mode.")
call_method = ensure_async(call_methods.pop())
serve_context.batch_size = batch_size
# Flask requests are passed to __call__ as a list
arg_list = [arg_list]
self.request_counter.add(batch_size)
result_list = await call_method(*arg_list, **kwargs_list)
if not isinstance(result_list, Iterable) or isinstance(
result_list, (dict, set)):
error_message = ("RayServe expects an ordered iterable object "
"but the worker returned a {}".format(
type(result_list)))
raise RayServeException(error_message)
# Normalize the result into a list type. This operation is fast
# in Python because it doesn't copy anything.
result_list = list(result_list)
if (len(result_list) != batch_size):
error_message = ("Worker doesn't preserve batch size. The "
"input has length {} but the returned list "
"has length {}. Please return a list of "
"results with length equal to the batch size"
".".format(batch_size, len(result_list)))
raise RayServeException(error_message)
self._reset_context()
return result_list
except Exception as e:
wrapped_exception = wrap_to_ray_error(e)
self.error_counter.add()
self._reset_context()
return [wrapped_exception for _ in range(batch_size)]
async def invoke_batch(self, request_item_list):
# TODO(alind) : create no-http services. The enqueues
# from such services will always be TaskContext.Python.
# Assumption : all the requests in a bacth
# have same serve context.
# For batching kwargs are modified as follows -
# kwargs [Python Context] : key,val
# kwargs_list : key, [val1,val2, ... , valn]
# or
# args[Web Context] : val
# args_list : [val1,val2, ...... , valn]
# where n (current batch size) <= max_batch_size of a backend
arg_list = []
kwargs_list = defaultdict(list)
context_flags = set()
batch_size = len(request_item_list)
call_methods = set()
for item in request_item_list:
args, kwargs, is_web_context = parse_request_item(item)
context_flags.add(is_web_context)
call_method = self.get_runner_method(item)
call_methods.add(call_method)
if is_web_context:
# Python context only have kwargs
flask_request = args[0]
arg_list.append(flask_request)
else:
# Web context only have one positional argument
for k, v in kwargs.items():
kwargs_list[k].append(v)
# Set the flask request as a list to conform
# with batching semantics: when in batching
# mode, each argument is turned into list.
if self.has_positional_args(call_method):
arg_list.append(FakeFlaskRequest())
try:
# Check mixing of query context (unified context needed).
if len(context_flags) != 1:
raise RayServeException(
"Batched queries contain mixed context. Please only send "
"the same type of requests in batching mode.")
serve_context.web = context_flags.pop()
if len(call_methods) != 1:
raise RayServeException(
"Queries contain mixed calling methods. Please only send "
"the same type of requests in batching mode.")
call_method = ensure_async(call_methods.pop())
serve_context.batch_size = batch_size
# Flask requests are passed to __call__ as a list
arg_list = [arg_list]
start_timestamp = time.time()
result_list = await call_method(*arg_list, **kwargs_list)
self.latency_list.append(time.time() - start_timestamp)
if (not isinstance(result_list,
list)) or (len(result_list) != batch_size):
error_message = ("Worker doesn't preserve batch size. The "
"input has length {} but the returned list "
"has length {}. Please return a list of "
"results with length equal to the batch size"
".".format(batch_size, len(result_list)))
raise RayServeException(error_message)
return result_list
except Exception as e:
wrapped_exception = wrap_to_ray_error(e)
self.error_counter += batch_size
return [wrapped_exception for _ in range(batch_size)]
