feature: add support for async inline error notifications

Author: doddaspk-amzn

doc/overview.rst

@@ -1164,7 +1164,8 @@ More information about SageMaker Asynchronous Inference can be found in the `AWS
 
 To deploy asynchronous inference endpoint, you will need to create a ``AsyncInferenceConfig`` object.
 If you create ``AsyncInferenceConfig`` without specifying its arguments, the default ``S3OutputPath`` will
-be ``s3://sagemaker-{REGION}-{ACCOUNTID}/async-endpoint-outputs/{UNIQUE-JOB-NAME}``. (example shown below):
+be ``s3://sagemaker-{REGION}-{ACCOUNTID}/async-endpoint-outputs/{UNIQUE-JOB-NAME}``, ``S3FailurePath`` will
+be ``s3://sagemaker-{REGION}-{ACCOUNTID}/async-endpoint-failures/{UNIQUE-JOB-NAME}`` (example shown below):
 
 .. code:: python
 
@@ -1174,18 +1175,21 @@ be ``s3://sagemaker-{REGION}-{ACCOUNTID}/async-endpoint-outputs/{UNIQUE-JOB-NAME
     async_config = AsyncInferenceConfig()
 
 Or you can specify configurations in ``AsyncInferenceConfig`` as you like. All of those configuration parameters
-are optional but if you don’t specify the ``output_path``, Amazon SageMaker will use the default ``S3OutputPath``
+are optional but if you don’t specify the ``output_path`` or ``failure_path``, Amazon SageMaker will use the
+default ``S3OutputPath`` or ``S3FailurePath``
 mentioned above (example shown below):
 
 .. code:: python
 
-    # Specify S3OutputPath, MaxConcurrentInvocationsPerInstance and NotificationConfig in the async config object
+    # Specify S3OutputPath, S3FailurePath, MaxConcurrentInvocationsPerInstance and NotificationConfig
+    # in the async config object
     async_config = AsyncInferenceConfig(
         output_path="s3://{s3_bucket}/{bucket_prefix}/output",
         max_concurrent_invocations_per_instance=10,
         notification_config = {
             "SuccessTopic": "arn:aws:sns:aws-region:account-id:topic-name",
             "ErrorTopic": "arn:aws:sns:aws-region:account-id:topic-name",
+            "IncludeInferenceResponseIn": ["SUCCESS_NOTIFICATION_TOPIC","ERROR_NOTIFICATION_TOPIC"],
         }
     )
 

src/sagemaker/async_inference/async_inference_config.py

@@ -31,6 +31,7 @@ def __init__(
         max_concurrent_invocations_per_instance=None,
         kms_key_id=None,
         notification_config=None,
+        failure_path=None,
     ):
         """Initialize an AsyncInferenceConfig object for async inference configuration.
 
@@ -45,6 +46,9 @@ def __init__(
             kms_key_id (str): Optional. The Amazon Web Services Key Management Service
                 (Amazon Web Services KMS) key that Amazon SageMaker uses to encrypt the
                 asynchronous inference output in Amazon S3. (Default: None)
+            failure_path (str): Optional. The Amazon S3 location that endpoints upload model
+                responses for failed requests. If no value is provided, Amazon SageMaker will
+                use default Amazon S3 Async Inference failure path. (Default: None)
             notification_config (dict): Optional. Specifies the configuration for notifications
                 of inference results for asynchronous inference. Only one notification is generated
                 per invocation request (Default: None):
@@ -54,17 +58,24 @@ def __init__(
                 * error_topic (str): Amazon SNS topic to post a notification to when inference
                 fails. If no topic is provided, no notification is sent on failure.
                 The key in notification_config is 'ErrorTopic'.
+                * include_inference_response_in (list): Optional. When provided the inference
+                response will be included in the notification topics. If not provided,
+                a notification will still be generated on success/error, but will not
+                contain the inference response.
+                Valid options are SUCCESS_NOTIFICATION_TOPIC, ERROR_NOTIFICATION_TOPIC
         """
         self.output_path = output_path
         self.max_concurrent_invocations_per_instance = max_concurrent_invocations_per_instance
         self.kms_key_id = kms_key_id
         self.notification_config = notification_config
+        self.failure_path = failure_path
 
     def _to_request_dict(self):
         """Generates a request dictionary using the parameters provided to the class."""
         request_dict = {
             "OutputConfig": {
                 "S3OutputPath": self.output_path,
+                "S3FailurePath": self.failure_path,
             },
         }
 

src/sagemaker/async_inference/async_inference_response.py

@@ -17,7 +17,11 @@
 from botocore.exceptions import ClientError
 from sagemaker.s3 import parse_s3_url
 from sagemaker.async_inference import WaiterConfig
-from sagemaker.exceptions import ObjectNotExistedError, UnexpectedClientError
+from sagemaker.exceptions import (
+    ObjectNotExistedError,
+    UnexpectedClientError,
+    AsyncInferenceModelError,
+)
 
 
 class AsyncInferenceResponse(object):
@@ -32,6 +36,7 @@ def __init__(
         self,
         predictor_async,
         output_path,
+        failure_path,
     ):
         """Initialize an AsyncInferenceResponse object.
 
@@ -43,10 +48,13 @@ def __init__(
                 that return this response.
             output_path (str): The Amazon S3 location that endpoints upload inference responses
                 to.
+            failure_path (str): The Amazon S3 location that endpoints upload model errors
+                for failed requests.
         """
         self.predictor_async = predictor_async
         self.output_path = output_path
         self._result = None
+        self.failure_path = failure_path
 
     def get_result(
         self,
@@ -71,28 +79,36 @@ def get_result(
 
         if self._result is None:
             if waiter_config is None:
-                self._result = self._get_result_from_s3(self.output_path)
+                self._result = self._get_result_from_s3(self.output_path, self.failure_path)
             else:
                 self._result = self.predictor_async._wait_for_output(
-                    self.output_path, waiter_config
+                    self.output_path, self.failure_path, waiter_config
                 )
         return self._result
 
-    def _get_result_from_s3(
-        self,
-        output_path,
-    ):
+    def _get_result_from_s3(self, output_path, failure_path):
         """Get inference result from the output Amazon S3 path"""
         bucket, key = parse_s3_url(output_path)
         try:
             response = self.predictor_async.s3_client.get_object(Bucket=bucket, Key=key)
             return self.predictor_async.predictor._handle_response(response)
-        except ClientError as ex:
-            if ex.response["Error"]["Code"] == "NoSuchKey":
-                raise ObjectNotExistedError(
-                    message="Inference could still be running",
-                    output_path=output_path,
-                )
-            raise UnexpectedClientError(
-                message=ex.response["Error"]["Message"],
-            )
+        except ClientError as e:
+            if e.response["Error"]["Code"] == "NoSuchKey":
+                try:
+                    failure_bucket, failure_key = parse_s3_url(failure_path)
+                    failure_response = self.predictor_async.s3_client.get_object(
+                        Bucket=failure_bucket, Key=failure_key
+                    )
+                    failure_response = self.predictor_async.predictor._handle_response(
+                        failure_response
+                    )
+                    raise AsyncInferenceModelError(message=failure_response)
+                except ClientError as ex:
+                    if ex.response["Error"]["Code"] == "NoSuchKey":
+                        raise ObjectNotExistedError(
+                            message="Inference could still be running", output_path=output_path
+                        )
+                    else:
+                        raise UnexpectedClientError(message=ex.response["Error"]["Message"])
+            else:
+                raise UnexpectedClientError(message=e.response["Error"]["Message"])

src/sagemaker/exceptions.py

@@ -77,3 +77,12 @@ def __init__(self, **kwargs):
         msg = self.fmt.format(**kwargs)
         Exception.__init__(self, msg)
         self.kwargs = kwargs
+
+
+class AsyncInferenceModelError(AsyncInferenceError):
+    """Raised when model returns errors for failed requests"""
+
+    fmt = "Model returned error: {message} "
+
+    def __init__(self, message):
+        super().__init__(message=message)

src/sagemaker/model.py

@@ -1277,7 +1277,10 @@ def deploy(
 
         async_inference_config_dict = None
         if is_async:
-            if async_inference_config.output_path is None:
+            if (
+                async_inference_config.output_path is None
+                or async_inference_config.failure_path is None
+            ):
                 async_inference_config = self._build_default_async_inference_config(
                     async_inference_config
                 )
@@ -1316,11 +1319,19 @@ def deploy(
 
     def _build_default_async_inference_config(self, async_inference_config):
         """Build default async inference config and return ``AsyncInferenceConfig``"""
-        async_output_folder = unique_name_from_base(self.name)
-        async_output_s3uri = "s3://{}/async-endpoint-outputs/{}".format(
-            self.sagemaker_session.default_bucket(), async_output_folder
-        )
-        async_inference_config.output_path = async_output_s3uri
+        unique_folder = unique_name_from_base(self.name)
+        if async_inference_config.output_path is None:
+            async_output_s3uri = "s3://{}/async-endpoint-outputs/{}".format(
+                self.sagemaker_session.default_bucket(), unique_folder
+            )
+            async_inference_config.output_path = async_output_s3uri
+
+        if async_inference_config.failure_path is None:
+            async_failure_s3uri = "s3://{}/async-endpoint-failures/{}".format(
+                self.sagemaker_session.default_bucket(), unique_folder
+            )
+            async_inference_config.failure_path = async_failure_s3uri
+
         return async_inference_config
 
     def transformer(

src/sagemaker/predictor_async.py

@@ -12,10 +12,11 @@
 # language governing permissions and limitations under the License.
 """Placeholder docstring"""
 from __future__ import absolute_import
-
+import threading
+import time
 import uuid
 from botocore.exceptions import WaiterError
-from sagemaker.exceptions import PollingTimeoutError
+from sagemaker.exceptions import PollingTimeoutError, AsyncInferenceModelError
 from sagemaker.async_inference import WaiterConfig, AsyncInferenceResponse
 from sagemaker.s3 import parse_s3_url
 from sagemaker.session import Session
@@ -98,7 +99,10 @@ def predict(
         self._input_path = input_path
         response = self._submit_async_request(input_path, initial_args, inference_id)
         output_location = response["OutputLocation"]
-        result = self._wait_for_output(output_path=output_location, waiter_config=waiter_config)
+        failure_location = response["FailureLocation"]
+        result = self._wait_for_output(
+            output_path=output_location, failure_path=failure_location, waiter_config=waiter_config
+        )
 
         return result
 
@@ -141,9 +145,11 @@ def predict_async(
         self._input_path = input_path
         response = self._submit_async_request(input_path, initial_args, inference_id)
         output_location = response["OutputLocation"]
+        failure_location = response["FailureLocation"]
         response_async = AsyncInferenceResponse(
             predictor_async=self,
             output_path=output_location,
+            failure_path=failure_location,
         )
 
         return response_async
@@ -209,30 +215,81 @@ def _submit_async_request(
 
         return response
 
-    def _wait_for_output(
-        self,
-        output_path,
-        waiter_config,
-    ):
+    def _wait_for_output(self, output_path, failure_path, waiter_config):
         """Check the Amazon S3 output path for the output.
 
-        Periodically check Amazon S3 output path for async inference result.
-        Timeout automatically after max attempts reached
-        """
-        bucket, key = parse_s3_url(output_path)
-        s3_waiter = self.s3_client.get_waiter("object_exists")
-        try:
-            s3_waiter.wait(Bucket=bucket, Key=key, WaiterConfig=waiter_config._to_request_dict())
-        except WaiterError:
-            raise PollingTimeoutError(
-                message="Inference could still be running",
-                output_path=output_path,
-                seconds=waiter_config.delay * waiter_config.max_attempts,
-            )
+        This method waits for either the output file or the failure file to be found on the
+        specified S3 output path. Whichever file is found first, its corresponding event is
+        triggered, and the method executes the appropriate action based on the event.
 
-        s3_object = self.s3_client.get_object(Bucket=bucket, Key=key)
-        result = self.predictor._handle_response(response=s3_object)
-        return result
+        Args:
+            output_path (str): The S3 path where the output file is expected to be found.
+            failure_path (str): The S3 path where the failure file is expected to be found.
+            waiter_config (boto3.waiter.WaiterConfig): The configuration for the S3 waiter.
+
+        Returns:
+            Any: The deserialized result from the output file, if the output file is found first.
+            Otherwise, raises an exception.
+
+        Raises:
+            AsyncInferenceModelError: If the failure file is found before the output file.
+            PollingTimeoutError: If both files are not found and the S3 waiter
+             has thrown a WaiterError.
+        """
+        output_bucket, output_key = parse_s3_url(output_path)
+        failure_bucket, failure_key = parse_s3_url(failure_path)
+
+        output_file_found = threading.Event()
+        failure_file_found = threading.Event()
+
+        def check_output_file():
+            try:
+                output_file_waiter = self.s3_client.get_waiter("object_exists")
+                output_file_waiter.wait(
+                    Bucket=output_bucket,
+                    Key=output_key,
+                    WaiterConfig=waiter_config._to_request_dict(),
+                )
+                output_file_found.set()
+            except WaiterError:
+                pass
+
+        def check_failure_file():
+            try:
+                failure_file_waiter = self.s3_client.get_waiter("object_exists")
+                failure_file_waiter.wait(
+                    Bucket=failure_bucket,
+                    Key=failure_key,
+                    WaiterConfig=waiter_config._to_request_dict(),
+                )
+                failure_file_found.set()
+            except WaiterError:
+                pass
+
+        output_thread = threading.Thread(target=check_output_file)
+        failure_thread = threading.Thread(target=check_failure_file)
+
+        output_thread.start()
+        failure_thread.start()
+
+        while not output_file_found.is_set() and not failure_file_found.is_set():
+            time.sleep(1)
+
+        if output_file_found.is_set():
+            s3_object = self.s3_client.get_object(Bucket=output_bucket, Key=output_key)
+            result = self.predictor._handle_response(response=s3_object)
+            return result
+
+        failure_object = self.s3_client.get_object(Bucket=failure_bucket, Key=failure_key)
+        failure_response = self.predictor._handle_response(response=failure_object)
+
+        raise AsyncInferenceModelError(
+            message=failure_response
+        ) if failure_file_found.is_set() else PollingTimeoutError(
+            message="Inference could still be running",
+            output_path=output_path,
+            seconds=waiter_config.delay * waiter_config.max_attempts,
+        )
 
     def update_endpoint(
         self,

tests/integ/test_async_inference.py

@@ -63,6 +63,9 @@ def test_async_walkthrough(sagemaker_session, cpu_instance_type, training_set):
         assert result_no_wait_with_data.output_path.startswith(
             "s3://" + sagemaker_session.default_bucket()
         )
+        assert result_no_wait_with_data.failure_path.startswith(
+            "s3://" + sagemaker_session.default_bucket() + "/async-endpoint-failures/"
+        )
         time.sleep(5)
         result_no_wait_with_data = result_no_wait_with_data.get_result()
         assert len(result_no_wait_with_data) == 5
@@ -99,6 +102,9 @@ def test_async_walkthrough(sagemaker_session, cpu_instance_type, training_set):
         assert result_not_wait.output_path.startswith("s3://" + sagemaker_session.default_bucket())
         time.sleep(5)
         result_not_wait = result_not_wait.get_result()
+        assert result_not_wait.failure_path.startswith(
+            "s3://" + sagemaker_session.default_bucket() + "/async-endpoint-failures/"
+        )
         assert len(result_not_wait) == 5
         for record in result_not_wait:
             assert record.label["projection"] is not None