feat: Telemetry metrics (#4414)

* Emit additional telemetry metrics

* Fix unit tests

* Emit endpoint failure to telemetry

* Address PR Comments

* Emit latency in telemetry

* Address PR Comments

* Addressed PR Comments

* Address PR Comments

* Fix tests

* Fix integ tests

---------

Co-authored-by: Jonathan Makunga <makung@amazon.com>
Co-authored-by: Erick Benitez-Ramos <141277478+benieric@users.noreply.github.com>

Author: 3 people

.gitignore

@@ -34,4 +34,4 @@ env/
 **/_repack_script_launcher.sh
 tests/data/**/_repack_model.py
 tests/data/experiment/sagemaker-dev-1.0.tar.gz
-src/sagemaker/serve/tmp_workspace
+src/sagemaker/serve/tmp_workspace

src/sagemaker/serve/utils/telemetry_logger.py

@@ -13,12 +13,15 @@
 """Placeholder docstring"""
 from __future__ import absolute_import
 import logging
+from time import perf_counter
+
 import requests
 
-from sagemaker import Session
+from sagemaker import Session, exceptions
 from sagemaker.serve.mode.function_pointers import Mode
 from sagemaker.serve.utils.exceptions import ModelBuilderException
 from sagemaker.serve.utils.types import ModelServer
+from sagemaker.user_agent import SDK_VERSION
 
 logger = logging.getLogger(__name__)
 
@@ -63,13 +66,21 @@ def wrapper(self, *args, **kwargs):
                 f"{func_name}"
                 f"&x-modelServer={MODEL_SERVER_TO_CODE[str(self.model_server)]}"
                 f"&x-imageTag={image_uri_tail}"
+                f"&x-sdkVersion={SDK_VERSION}"
             )
 
             if self.model_server == ModelServer.DJL_SERVING or self.model_server == ModelServer.TGI:
                 extra += f"&x-modelName={self.model}"
 
+            if self.sagemaker_session and self.sagemaker_session.endpoint_arn:
+                extra += f"&x-endpointArn={self.sagemaker_session.endpoint_arn}"
+
+            start_timer = perf_counter()
             try:
                 response = func(self, *args, **kwargs)
+                stop_timer = perf_counter()
+                elapsed = stop_timer - start_timer
+                extra += f"&x-latency={round(elapsed, 2)}"
                 if not self.serve_settings.telemetry_opt_out:
                     _send_telemetry(
                         "1",
@@ -79,7 +90,15 @@ def wrapper(self, *args, **kwargs):
                         None,
                         extra,
                     )
-            except ModelBuilderException as e:
+            except (
+                ModelBuilderException,
+                exceptions.CapacityError,
+                exceptions.UnexpectedStatusException,
+                exceptions.AsyncInferenceError,
+            ) as e:
+                stop_timer = perf_counter()
+                elapsed = stop_timer - start_timer
+                extra += f"&x-latency={round(elapsed, 2)}"
                 if not self.serve_settings.telemetry_opt_out:
                     _send_telemetry(
                         "0",

src/sagemaker/session.py

@@ -243,6 +243,7 @@ def __init__(
         # sagemaker_config is validated and initialized inside :func:`_initialize`,
         # so if default_bucket is None and the sagemaker_config has a default S3 bucket configured,
         # _default_bucket_name_override will be set again inside :func:`_initialize`.
+        self.endpoint_arn = None
         self._default_bucket = None
         self._default_bucket_name_override = default_bucket
         # this may also be set again inside :func:`_initialize` if it is None
@@ -4284,9 +4285,12 @@ def create_endpoint(self, endpoint_name, config_name, tags=None, wait=True, live
             tags, "{}.{}.{}".format(SAGEMAKER, ENDPOINT, TAGS)
         )
 
-        self.sagemaker_client.create_endpoint(
+        res = self.sagemaker_client.create_endpoint(
             EndpointName=endpoint_name, EndpointConfigName=config_name, Tags=tags
         )
+        if res:
+            self.endpoint_arn = res["EndpointArn"]
+
         if wait:
             self.wait_for_endpoint(endpoint_name, live_logging=live_logging)
         return endpoint_name
@@ -4344,9 +4348,11 @@ def update_endpoint(self, endpoint_name, endpoint_config_name, wait=True):
                 "existing endpoint name".format(endpoint_name)
             )
 
-        self.sagemaker_client.update_endpoint(
+        res = self.sagemaker_client.update_endpoint(
             EndpointName=endpoint_name, EndpointConfigName=endpoint_config_name
         )
+        if res:
+            self.endpoint_arn = res["EndpointArn"]
 
         if wait:
             self.wait_for_endpoint(endpoint_name)

tests/unit/sagemaker/serve/utils/test_telemetry_logger.py

@@ -20,6 +20,7 @@
     _construct_url,
 )
 from sagemaker.serve.utils.exceptions import ModelBuilderException, LocalModelOutOfMemoryException
+from sagemaker.user_agent import SDK_VERSION
 
 MOCK_SESSION = Mock()
 MOCK_FUNC_NAME = "Mock.deploy"
@@ -32,6 +33,7 @@
 )
 MOCK_HUGGINGFACE_ID = "meta-llama/Llama-2-7b-hf"
 MOCK_EXCEPTION = LocalModelOutOfMemoryException("mock raise ex")
+MOCK_ENDPOINT_ARN = "arn:aws:sagemaker:us-west-2:123456789012:endpoint/test"
 
 
 class ModelBuilderMock:
@@ -72,15 +74,22 @@ def test_capture_telemetry_decorator_djl_success(self, mock_send_telemetry):
         mock_model_builder.model = MOCK_HUGGINGFACE_ID
         mock_model_builder.mode = Mode.LOCAL_CONTAINER
         mock_model_builder.model_server = ModelServer.DJL_SERVING
+        mock_model_builder.sagemaker_session.endpoint_arn = MOCK_ENDPOINT_ARN
 
         mock_model_builder.mock_deploy()
 
+        args = mock_send_telemetry.call_args.args
+        latency = str(args[5]).split("latency=")[1]
         expected_extra_str = (
             f"{MOCK_FUNC_NAME}"
             "&x-modelServer=4"
             "&x-imageTag=djl-inference:0.25.0-deepspeed0.11.0-cu118"
+            f"&x-sdkVersion={SDK_VERSION}"
             f"&x-modelName={MOCK_HUGGINGFACE_ID}"
+            f"&x-endpointArn={MOCK_ENDPOINT_ARN}"
+            f"&x-latency={latency}"
         )
+
         mock_send_telemetry.assert_called_once_with(
             "1", 2, MOCK_SESSION, None, None, expected_extra_str
         )
@@ -93,15 +102,22 @@ def test_capture_telemetry_decorator_tgi_success(self, mock_send_telemetry):
         mock_model_builder.model = MOCK_HUGGINGFACE_ID
         mock_model_builder.mode = Mode.LOCAL_CONTAINER
         mock_model_builder.model_server = ModelServer.TGI
+        mock_model_builder.sagemaker_session.endpoint_arn = MOCK_ENDPOINT_ARN
 
         mock_model_builder.mock_deploy()
 
+        args = mock_send_telemetry.call_args.args
+        latency = str(args[5]).split("latency=")[1]
         expected_extra_str = (
             f"{MOCK_FUNC_NAME}"
             "&x-modelServer=6"
             "&x-imageTag=huggingface-pytorch-inference:2.0.0-transformers4.28.1-cpu-py310-ubuntu20.04"
+            f"&x-sdkVersion={SDK_VERSION}"
             f"&x-modelName={MOCK_HUGGINGFACE_ID}"
+            f"&x-endpointArn={MOCK_ENDPOINT_ARN}"
+            f"&x-latency={latency}"
         )
+
         mock_send_telemetry.assert_called_once_with(
             "1", 2, MOCK_SESSION, None, None, expected_extra_str
         )
@@ -126,6 +142,7 @@ def test_capture_telemetry_decorator_handle_exception_success(self, mock_send_te
         mock_model_builder.model = MOCK_HUGGINGFACE_ID
         mock_model_builder.mode = Mode.LOCAL_CONTAINER
         mock_model_builder.model_server = ModelServer.DJL_SERVING
+        mock_model_builder.sagemaker_session.endpoint_arn = MOCK_ENDPOINT_ARN
 
         mock_exception = Mock()
         mock_exception_obj = MOCK_EXCEPTION
@@ -134,12 +151,18 @@ def test_capture_telemetry_decorator_handle_exception_success(self, mock_send_te
         with self.assertRaises(ModelBuilderException) as _:
             mock_model_builder.mock_deploy(mock_exception)
 
+        args = mock_send_telemetry.call_args.args
+        latency = str(args[5]).split("latency=")[1]
         expected_extra_str = (
             f"{MOCK_FUNC_NAME}"
             "&x-modelServer=4"
             "&x-imageTag=djl-inference:0.25.0-deepspeed0.11.0-cu118"
+            f"&x-sdkVersion={SDK_VERSION}"
             f"&x-modelName={MOCK_HUGGINGFACE_ID}"
+            f"&x-endpointArn={MOCK_ENDPOINT_ARN}"
+            f"&x-latency={latency}"
         )
+
         mock_send_telemetry.assert_called_once_with(
             "0",
             2,