hi

Signed-off-by: Judy Ng <[email protected]>

tests/integration-tests/configs/scaling_stress_test.yaml

@@ -2,7 +2,7 @@ test-suites:
   performance_tests:
     test_scaling.py::test_scaling_stress_test:
       dimensions:
-        - regions: [ "us-east-1" ]
+        - regions: [ "use1-az6" ]
           instances: [ "c5.large" ]
           oss: [ "alinux2" ]
           schedulers: [ "slurm" ]

tests/integration-tests/tests/common/scaling_common.py

@@ -71,16 +71,19 @@ def get_bootstrap_errors(remote_command_executor: RemoteCommandExecutor, cluster
     ip_addresses_with_bootstrap_errors = remote_command_executor.run_remote_command(command=f"cat $HOME/bootstrap_errors.txt").stdout
 
     os.makedirs("bootstrap_errors", exist_ok=True)
-
     client = boto3.client("ec2", region_name=region)
     for ip_address in ip_addresses_with_bootstrap_errors.splitlines():
-        instance_id = client.describe_instances(
-            Filters=[{"Name": "private-ip-address", "Values": [ip_address]}]
-        )["Reservations"][0]["Instances"][0]["InstanceId"]
-        logging.info(f"Instance {instance_id} had bootstrap errors. Check the logs for details.")
-        compute_node_log = client.get_console_output(InstanceId=instance_id, Latest=True)["Output"]
-        with open(f"bootstrap_errors/{region}-{cluster_name}-{instance_id}-bootstrap-error.txt", "w") as file:
-            file.write(compute_node_log)
+        try:
+            instance_id = client.describe_instances(
+                Filters=[{"Name": "private-ip-address", "Values": [ip_address]}]
+            )["Reservations"][0]["Instances"][0]["InstanceId"]
+            logging.warning(f"Instance {instance_id} had bootstrap errors. Check the logs for details.")
+            compute_node_log = client.get_console_output(InstanceId=instance_id, Latest=True)["Output"]
+            with open(f"bootstrap_errors/{region}-{cluster_name}-{instance_id}-bootstrap-error.txt", "w") as file:
+                file.write(compute_node_log)
+        except Exception:
+            logging.error("Error when retrieving the compute node logs for instance with ip address %s", ip_address)
+            raise
 
 
 def get_scaling_metrics(

tests/integration-tests/tests/performance_tests/test_scaling.py

@@ -82,7 +82,7 @@ def _get_scaling_time(capacity_time_series: list, timestamps: list, scaling_targ
 
 
 @pytest.mark.usefixtures("scheduler")
-@pytest.mark.parametrize("scaling_strategy", ["all-or-nothing", "best-effort"])
+@pytest.mark.parametrize("scaling_strategy", ["all-or-nothing"])
 def test_scaling_stress_test(
     test_datadir,
     instance,