Merge branch 'main' into python_required

Author: fege

tests/model_serving/model_server/serverless/test_concurrency_auto_scale.py

@@ -53,7 +53,8 @@ def test_auto_scale_using_concurrency(
         for pods in inference_service_pods_sampler(
             client=admin_client,
             isvc=s3_models_inference_service,
-            timeout=Timeout.TIMEOUT_1MIN,
+            timeout=Timeout.TIMEOUT_2MIN,
+            sleep=10,
         ):
             if pods:
                 if len(pods) > 1 and all([pod.status == pod.Status.RUNNING for pod in pods]):
@@ -66,6 +67,7 @@ def test_pods_scaled_down_when_no_requests(self, admin_client, s3_models_inferen
             client=admin_client,
             isvc=s3_models_inference_service,
             timeout=Timeout.TIMEOUT_4MIN,
+            sleep=10,
         ):
             if pods and len(pods) == 1:
                 return

tests/model_serving/model_server/serverless/utils.py

@@ -137,22 +137,25 @@ def verify_canary_traffic(
         )
 
 
-def inference_service_pods_sampler(client: DynamicClient, isvc: InferenceService, timeout: int) -> TimeoutSampler:
+def inference_service_pods_sampler(
+    client: DynamicClient, isvc: InferenceService, timeout: int, sleep: int = 1
+) -> TimeoutSampler:
     """
     Returns TimeoutSampler for inference service.
 
     Args:
         client (DynamicClient): DynamicClient object
         isvc (InferenceService): InferenceService object
         timeout (int): Timeout in seconds
+        sleep (int): Sleep time in seconds
 
     Returns:
         TimeoutSampler: TimeoutSampler object
 
     """
     return TimeoutSampler(
         wait_timeout=timeout,
-        sleep=1,
+        sleep=sleep,
         func=get_pods_by_isvc_label,
         client=client,
         isvc=isvc,