[Serve] Implement metric interface (#5852)

* Implement metric interface

* Address comment: made actor_handles a dict

* Fix iteration

* Lint

* Mark lightweight actors as num_cpus=0 to prevent resource starvation

* Be more explicit about the readiness condition

* Make task_runner non-blocking

* Lint

python/ray/experimental/serve/queues.py

@@ -76,6 +76,18 @@ class CentralizedQueues:
         # backend_name -> worker queue
         self.workers = defaultdict(deque)
 
+    def is_ready(self):
+        return True
+
+    def _serve_metric(self):
+        return {
+            "service_{}_queue_size".format(service_name): {
+                "value": len(queue),
+                "type": "counter",
+            }
+            for service_name, queue in self.queues.items()
+        }
+
     def enqueue_request(self, service, request_args, request_kwargs,
                         request_context):
         query = Query(request_args, request_kwargs, request_context)