[Serve] Hotfix: Fix actor handle hashing in metric monitoring (#5886)

python/ray/experimental/serve/global_state.py

@@ -98,6 +98,8 @@ class GlobalState:
             logger.debug((LOG_PREFIX + "Checking if HTTP server is ready."
                           "{} retries left.").format(retries))
             time.sleep(backoff_time_s)
+            # Exponential backoff
+            backoff_time_s *= 2
             retries -= 1
             if retries == 0:
                 raise Exception(

python/ray/experimental/serve/metric.py

@@ -27,11 +27,11 @@ class MetricMonitor:
         return True
 
     def add_target(self, target_handle):
-        hex_id = target_handle._ray_actor_id.hex()
+        hex_id = target_handle._ray_core_handle.actor_id().hex()
         self.actor_handles[hex_id] = target_handle
 
     def remove_target(self, target_handle):
-        hex_id = target_handle._ray_actor_id.hex()
+        hex_id = target_handle._ray_core_handle.actor_id().hex()
         self.actor_handles.pop(hex_id)
 
     def scrape(self):

python/ray/experimental/serve/tests/test_api.py

@@ -7,19 +7,21 @@ from ray.experimental import serve
 
 
 def test_e2e(serve_instance):
-    serve.create_endpoint("endpoint", "/api")
+    serve.create_endpoint("endpoint", "/api", blocking=True)
     result = ray.get(
         serve.global_state.kv_store_actor_handle.list_service.remote())
     assert result == {"/api": "endpoint"}
 
-    retry_count = 3
+    retry_count = 5
+    timeout_sleep = 0.5
     while True:
         try:
             resp = requests.get("http://127.0.0.1:8000/").json()
             assert resp == result
             break
         except Exception:
-            time.sleep(0.5)
+            time.sleep(timeout_sleep)
+            timeout_sleep *= 2
             retry_count -= 1
             if retry_count == 0:
                 assert False, "Route table hasn't been updated after 3 tries."

python/ray/experimental/serve/tests/test_metric.py

@@ -38,11 +38,10 @@ def test_metric_gc(ray_instance, start_target_actor):
     target_actor = start_target_actor
     # this means when new scrapes are invoked, the
     metric_monitor = MetricMonitor.remote(gc_window_seconds=0)
-    metric_monitor.add_target.remote(target_actor)
+    ray.get(metric_monitor.add_target.remote(target_actor))
 
     ray.get(metric_monitor.scrape.remote())
     df = ray.get(metric_monitor._get_dataframe.remote())
-    print(df)
     assert len(df) == 102
 
     # Old metric sould be cleared. So only 1 counter + 101 list values left.
@@ -56,10 +55,10 @@ def test_metric_system(ray_instance, start_target_actor):
 
     metric_monitor = MetricMonitor.remote()
 
-    metric_monitor.add_target.remote(target_actor)
+    ray.get(metric_monitor.add_target.remote(target_actor))
 
     # Scrape once
-    metric_monitor.scrape.remote()
+    ray.get(metric_monitor.scrape.remote())
 
     percentiles = [50, 90, 95]
     agg_windows_seconds = [60]