wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-09-11 12:43:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Serve] Ensure strict traffic splitting (#5929)

Browse Source 
* [Serve] Ensure strict traffic splitting

* Fix test
This commit is contained in:
Simon Mo
2019-10-20 20:18:14 -07:00
committed by GitHub
parent bc4a0de4da
commit 6b36ef1138
2 changed files with 35 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/ray/experimental/serve/queues.py
+28 -28
View File
@@ -73,19 +73,22 @@ class CentralizedQueues:
# service_name -> traffic_policy
self.traffic = defaultdict(dict)
# backend_name -> worker queue
# backend_name -> worker request queue
self.workers = defaultdict(deque)
# backend_name -> worker payload queue
self.buffer_queues = defaultdict(deque)
def is_ready(self):
return True
def _serve_metric(self):
return {
"service_{}_queue_size".format(service_name): {
"backend_{}_queue_size".format(backend_name): {
"value": len(queue),
"type": "counter",
}
for service_name, queue in self.queues.items()
for backend_name, queue in self.buffer_queues.items()
}
def enqueue_request(self, service, request_args, request_kwargs,
@@ -129,39 +132,36 @@ class CentralizedQueues:
return list(backends_in_policy.intersection(available_workers))
def _flush(self):
# perform traffic splitting for requests
for service, queue in self.queues.items():
# while there are incoming requests and there are backends
while len(queue) and len(self.traffic[service]):
backend_names = list(self.traffic[service].keys())
backend_weights = list(self.traffic[service].values())
chosen_backend = np.random.choice(
backend_names, p=backend_weights).squeeze()
request = queue.popleft()
self.buffer_queues[chosen_backend].append(request)
# distach buffer queues to work queues
for service in self.queues.keys():
ready_backends = self._get_available_backends(service)
for backend in ready_backends:
# no work available
if len(self.buffer_queues[backend]) == 0:
continue
while len(queue) and len(ready_backends):
# Fast path, only one backend available.
if len(ready_backends) == 1:
backend = ready_backends[0]
request, work = (queue.popleft(),
self.workers[backend].popleft())
ray.worker.global_worker.put_object(
work.work_object_id, request)
# We have more than one backend available.
# We will roll a dice among the multiple backends.
else:
backend_weights = np.array([
self.traffic[service][backend_name]
for backend_name in ready_backends
])
# Normalize the weights to 1.
backend_weights /= backend_weights.sum()
chosen_backend = np.random.choice(
ready_backends, p=backend_weights).squeeze()
buffer_queue = self.buffer_queues[backend]
work_queue = self.workers[backend]
while len(buffer_queue) and len(work_queue):
request, work = (
queue.popleft(),
self.workers[chosen_backend].popleft(),
buffer_queue.popleft(),
work_queue.popleft(),
)
ray.worker.global_worker.put_object(
work.work_object_id, request)
ready_backends = self._get_available_backends(service)
@ray.remote
class CentralizedQueuesActor(CentralizedQueues):