def __init__(self):
self.my_counter = metrics.Count(
"my_counter",
description=("The number of excellent requests to this backend."),
tag_keys=("backend", ))
self.my_counter.set_default_tags({
"backend": serve.get_current_backend_tag()
})
def __init__(self, controller_name):
controller = ray.get_actor(controller_name)
self.router = Router(controller)
self.long_poll_client = LongPollAsyncClient(
controller, {
LongPollKey.ROUTE_TABLE: self._update_route_table,
})
self.request_counter = metrics.Count(
"num_http_requests",
description="The number of HTTP requests processed",
tag_keys=("route", ))
async def fetch_config_from_controller(self, controller_name):
assert ray.is_initialized()
controller = ray.get_actor(controller_name)
self.route_table = await controller.get_router_config.remote()
self.request_counter = metrics.Count(
"num_http_requests",
description="The number of HTTP requests processed",
tag_keys=("route", ))
self.router = Router(controller)
await self.router.setup_in_async_loop()
def __init__(self, controller_name):
# Set the controller name so that serve.connect() will connect to the
# controller instance this proxy is running in.
ray.serve.api._set_internal_controller_name(controller_name)
self.client = ray.serve.connect()
controller = ray.get_actor(controller_name)
self.route_table = {} # Should be updated via long polling.
self.router = Router(controller)
self.long_poll_client = LongPollAsyncClient(controller, {
LongPollKey.ROUTE_TABLE: self._update_route_table,
})
self.request_counter = metrics.Count(
"num_http_requests",
description="The number of HTTP requests processed",
tag_keys=("route", ))
def __init__(self,
router: Router,
endpoint_name,
handle_options: Optional[HandleOptions] = None):
self.router = router
self.endpoint_name = endpoint_name
self.handle_options = handle_options or HandleOptions()
self.handle_tag = f"{self.endpoint_name}#{get_random_letters()}"
self.request_counter = metrics.Count(
"serve_handle_request_counter",
description=("The number of handle.remote() calls that have been "
"made on this handle."),
tag_keys=("handle", "endpoint"))
self.request_counter.set_default_tags({
"handle": self.handle_tag,
"endpoint": self.endpoint_name
})
def __init__(self, controller_handle: ActorHandle):
"""Router process incoming queries: choose backend, and assign replica.
Args:
controller_handle(ActorHandle): The controller handle.
"""
self.controller = controller_handle
self.endpoint_policies: Dict[str, EndpointPolicy] = dict()
self.backend_replicas: Dict[str, ReplicaSet] = defaultdict(ReplicaSet)
self._pending_endpoints: Dict[str, asyncio.Future] = dict()
# -- Metrics Registration -- #
self.num_router_requests = metrics.Count(
"serve_num_router_requests",
description="The number of requests processed by the router.",
tag_keys=("endpoint", ))
def __init__(self, controller_name):
# Set the controller name so that serve.connect() will connect to the
# controller instance this proxy is running in.
ray.serve.api._set_internal_replica_context(None, None,
controller_name)
self.client = ray.serve.connect()
controller = ray.get_actor(controller_name)
self.router = starlette.routing.Router(default=self._not_found)
# route -> (endpoint_tag, methods). Updated via long polling.
self.route_table: Dict[str, Tuple[EndpointTag, List[str]]] = {}
self.long_poll_client = LongPollAsyncClient(
controller, {
LongPollKey.ROUTE_TABLE: self._update_route_table,
})
self.request_counter = metrics.Count(
"serve_num_http_requests",
description="The number of HTTP requests processed.",
tag_keys=("route", ))
def __init__(self, _callable: Callable, backend_config: BackendConfig,
is_function: bool, controller_handle: ActorHandle) -> None:
self.backend_tag = ray.serve.api.get_current_backend_tag()
self.replica_tag = ray.serve.api.get_current_replica_tag()
self.callable = _callable
self.is_function = is_function
self.config = backend_config
self.batch_queue = BatchQueue(self.config.max_batch_size or 1,
self.config.batch_wait_timeout)
self.reconfigure(self.config.user_config)
self.num_ongoing_requests = 0
self.request_counter = metrics.Count(
"serve_backend_request_counter",
description=("The number of queries that have been "
"processed in this replica."),
tag_keys=("backend", ))
self.request_counter.set_default_tags({"backend": self.backend_tag})
self.long_poll_client = LongPollAsyncClient(
controller_handle, {
LongPollKey.BACKEND_CONFIGS: self._update_backend_configs,
})
self.error_counter = metrics.Count(
"serve_backend_error_counter",
description=("The number of exceptions that have "
"occurred in the backend."),
tag_keys=("backend", ))
self.error_counter.set_default_tags({"backend": self.backend_tag})
self.restart_counter = metrics.Count(
"serve_backend_replica_starts",
description=("The number of times this replica "
"has been restarted due to failure."),
tag_keys=("backend", "replica"))
self.restart_counter.set_default_tags({
"backend": self.backend_tag,
"replica": self.replica_tag
})
self.queuing_latency_tracker = metrics.Histogram(
"serve_backend_queuing_latency_ms",
description=("The latency for queries in the replica's queue "
"waiting to be processed or batched."),
boundaries=DEFAULT_LATENCY_BUCKET_MS,
tag_keys=("backend", "replica"))
self.queuing_latency_tracker.set_default_tags({
"backend":
self.backend_tag,
"replica":
self.replica_tag
})
self.processing_latency_tracker = metrics.Histogram(
"serve_backend_processing_latency_ms",
description="The latency for queries to be processed.",
boundaries=DEFAULT_LATENCY_BUCKET_MS,
tag_keys=("backend", "replica", "batch_size"))
self.processing_latency_tracker.set_default_tags({
"backend":
self.backend_tag,
"replica":
self.replica_tag
})
self.num_queued_items = metrics.Gauge(
"serve_replica_queued_queries",
description=("The current number of queries queued in "
"the backend replicas."),
tag_keys=("backend", "replica"))
self.num_queued_items.set_default_tags({
"backend": self.backend_tag,
"replica": self.replica_tag
})
self.num_processing_items = metrics.Gauge(
"serve_replica_processing_queries",
description="The current number of queries being processed.",
tag_keys=("backend", "replica"))
self.num_processing_items.set_default_tags({
"backend": self.backend_tag,
"replica": self.replica_tag
})
self.restart_counter.record(1)
asyncio.get_event_loop().create_task(self.main_loop())
def __init__(self, backend_tag: str, replica_tag: str, _callable: Callable,
backend_config: BackendConfig, is_function: bool) -> None:
self.backend_tag = backend_tag
self.replica_tag = replica_tag
self.callable = _callable
self.is_function = is_function
self.config = backend_config
self.batch_queue = BatchQueue(self.config.max_batch_size or 1,
self.config.batch_wait_timeout)
self.num_ongoing_requests = 0
self.request_counter = metrics.Count(
"backend_request_counter",
description=("Number of queries that have been "
"processed in this replica"),
tag_keys=("backend", ))
self.request_counter.set_default_tags({"backend": self.backend_tag})
self.error_counter = metrics.Count(
"backend_error_counter",
description=("Number of exceptions that have "
"occurred in the backend"),
tag_keys=("backend", ))
self.error_counter.set_default_tags({"backend": self.backend_tag})
self.restart_counter = metrics.Count(
"backend_worker_starts",
description=("The number of time this replica workers "
"has been restarted due to failure."),
tag_keys=("backend", "replica_tag"))
self.restart_counter.set_default_tags({
"backend": self.backend_tag,
"replica_tag": self.replica_tag
})
self.queuing_latency_tracker = metrics.Histogram(
"backend_queuing_latency_ms",
description=(
"The latency for queries waiting in the replica's queue "
"waiting to be processed or batched."),
boundaries=DEFAULT_LATENCY_BUCKET_MS,
tag_keys=("backend", "replica_tag"))
self.queuing_latency_tracker.set_default_tags({
"backend": self.backend_tag,
"replica_tag": self.replica_tag
})
self.processing_latency_tracker = metrics.Histogram(
"backend_processing_latency_ms",
description="The latency for queries to be processed",
boundaries=DEFAULT_LATENCY_BUCKET_MS,
tag_keys=("backend", "replica_tag", "batch_size"))
self.processing_latency_tracker.set_default_tags({
"backend": self.backend_tag,
"replica_tag": self.replica_tag
})
self.num_queued_items = metrics.Gauge(
"replica_queued_queries",
description=("Current number of queries queued in the "
"the backend replicas"),
tag_keys=("backend", "replica_tag"))
self.num_queued_items.set_default_tags({
"backend": self.backend_tag,
"replica_tag": self.replica_tag
})
self.num_processing_items = metrics.Gauge(
"replica_processing_queries",
description="Current number of queries being processed",
tag_keys=("backend", "replica_tag"))
self.num_processing_items.set_default_tags({
"backend": self.backend_tag,
"replica_tag": self.replica_tag
})
self.restart_counter.record(1)
asyncio.get_event_loop().create_task(self.main_loop())
async def setup(self, name, controller_name):
# Note: Several queues are used in the router
# - When a request come in, it's placed inside its corresponding
# endpoint_queue.
# - The endpoint_queue is dequeued during flush operation, which moves
# the queries to backend buffer_queue. Here we match a request
# for an endpoint to a backend given some policy.
# - The worker_queue is used to collect idle actor handle. These
# handles are dequed during the second stage of flush operation,
# which assign queries in buffer_queue to actor handle.
self.name = name
# -- Queues -- #
# endpoint_name -> request queue
# We use FIFO (left to right) ordering. The new items should be added
# using appendleft. Old items should be removed via pop().
self.endpoint_queues: DefaultDict[deque[Query]] = defaultdict(deque)
# backend_name -> worker replica tag queue
self.worker_queues: DefaultDict[deque[str]] = defaultdict(deque)
# backend_name -> worker payload queue
self.backend_queues = defaultdict(deque)
# -- Metadata -- #
# endpoint_name -> traffic_policy
self.traffic = dict()
# backend_name -> backend_config
self.backend_info = dict()
# replica tag -> worker_handle
self.replicas = dict()
# backend_name -> replica_tag -> concurrent queries counter
self.queries_counter = defaultdict(lambda: defaultdict(int))
# -- Synchronization -- #
# This lock guarantee that only one flush operation can happen at a
# time. Without the lock, multiple flush operation can pop from the
# same buffer_queue and worker_queue and create deadlock. For example,
# an operation holding the only query and the other flush operation
# holding the only idle replica. Additionally, allowing only one flush
# operation at a time simplifies design overhead for custom queuing and
# batching policies.
self.flush_lock = asyncio.Lock()
# -- State Restoration -- #
# Fetch the worker handles, traffic policies, and backend configs from
# the controller. We use a "pull-based" approach instead of pushing
# them from the controller so that the router can transparently recover
# from failure.
self.controller = ray.get_actor(controller_name)
traffic_policies = ray.get(
self.controller.get_traffic_policies.remote())
for endpoint, traffic_policy in traffic_policies.items():
await self.set_traffic(endpoint, traffic_policy)
backend_dict = ray.get(self.controller.get_all_worker_handles.remote())
for backend_tag, replica_dict in backend_dict.items():
for replica_tag, worker in replica_dict.items():
await self.add_new_worker(backend_tag, replica_tag, worker)
backend_configs = ray.get(self.controller.get_backend_configs.remote())
for backend, backend_config in backend_configs.items():
await self.set_backend_config(backend, backend_config)
# -- Metrics Registration -- #
self.num_router_requests = metrics.Count(
"num_router_requests",
description="Number of requests processed by the router.",
tag_keys=("endpoint", ))
self.num_error_endpoint_requests = metrics.Count(
"num_error_endpoint_requests",
description=(
"Number of requests that errored when getting results "
"for the endpoint."),
tag_keys=("endpoint", ))
self.num_error_backend_requests = metrics.Count(
"num_error_backend_requests",
description=("Number of requests that errored when getting result "
"from the backend."),
tag_keys=("backend", ))
self.backend_queue_size = metrics.Gauge(
"backend_queued_queries",
description=("Current number of queries queued "
"in the router for a backend"),
tag_keys=("backend", ))
asyncio.get_event_loop().create_task(self.report_queue_lengths())
