feature: Support for multi variant endpoint invocation with target variant param (#1577)

aakashpydi · chuyang-deng · web-flow · commit 69d06ad9bba0 · 2020-06-11T17:40:58.000-07:00
Co-authored-by: Chuyang &lt;chuyangdeng@gmail.com&gt;
diff --git a/setup.py b/setup.py
@@ -34,7 +34,7 @@ def read_version():
 
 # Declare minimal set for installation
 required_packages = [
-    "boto3>=1.13.6",
+    "boto3>=1.13.24",
     "numpy>=1.9.0",
     "protobuf>=3.1",
     "scipy>=0.19.0",
diff --git a/src/sagemaker/local/local_session.py b/src/sagemaker/local/local_session.py
@@ -343,6 +343,7 @@ def invoke_endpoint(
         Accept=None,
         CustomAttributes=None,
         TargetModel=None,
+        TargetVariant=None,
     ):
         """
 
@@ -370,6 +371,9 @@ def invoke_endpoint(
         if TargetModel is not None:
             headers["X-Amzn-SageMaker-Target-Model"] = TargetModel
 
+        if TargetVariant is not None:
+            headers["X-Amzn-SageMaker-Target-Variant"] = TargetVariant
+
         r = self.http.request("POST", url, body=Body, preload_content=False, headers=headers)
 
         return {"Body": r, "ContentType": Accept}
diff --git a/src/sagemaker/predictor.py b/src/sagemaker/predictor.py
@@ -83,7 +83,7 @@ def __init__(
         self._endpoint_config_name = self._get_endpoint_config_name()
         self._model_names = self._get_model_names()
 
-    def predict(self, data, initial_args=None, target_model=None):
+    def predict(self, data, initial_args=None, target_model=None, target_variant=None):
         """Return the inference from the specified endpoint.
 
         Args:
@@ -98,6 +98,9 @@ def predict(self, data, initial_args=None, target_model=None):
             target_model (str): S3 model artifact path to run an inference request on,
                 in case of a multi model endpoint. Does not apply to endpoints hosting
                 single model (Default: None)
+            target_variant (str): The name of the production variant to run an inference
+            request on (Default: None). Note that the ProductionVariant identifies the model
+            you want to host and the resources you want to deploy for hosting it.
 
         Returns:
             object: Inference for the given input. If a deserializer was specified when creating
@@ -106,7 +109,7 @@ def predict(self, data, initial_args=None, target_model=None):
                 as is.
         """
 
-        request_args = self._create_request_args(data, initial_args, target_model)
+        request_args = self._create_request_args(data, initial_args, target_model, target_variant)
         response = self.sagemaker_session.sagemaker_runtime_client.invoke_endpoint(**request_args)
         return self._handle_response(response)
 
@@ -123,12 +126,13 @@ def _handle_response(self, response):
         response_body.close()
         return data
 
-    def _create_request_args(self, data, initial_args=None, target_model=None):
+    def _create_request_args(self, data, initial_args=None, target_model=None, target_variant=None):
         """
         Args:
             data:
             initial_args:
             target_model:
+            target_variant:
         """
         args = dict(initial_args) if initial_args else {}
 
@@ -144,6 +148,9 @@ def _create_request_args(self, data, initial_args=None, target_model=None):
         if target_model:
             args["TargetModel"] = target_model
 
+        if target_variant:
+            args["TargetVariant"] = target_variant
+
         if self.serializer is not None:
             data = self.serializer(data)
 
diff --git a/tests/integ/test_multi_variant_endpoint.py b/tests/integ/test_multi_variant_endpoint.py
@@ -0,0 +1,318 @@
+# Copyright 2019-2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+from __future__ import absolute_import
+
+import json
+import os
+import math
+import pytest
+import scipy.stats as st
+
+from sagemaker.s3 import S3Uploader
+from sagemaker.session import production_variant
+from sagemaker.sparkml import SparkMLModel
+from sagemaker.utils import sagemaker_timestamp
+from sagemaker.content_types import CONTENT_TYPE_CSV
+from sagemaker.utils import unique_name_from_base
+from sagemaker.amazon.amazon_estimator import get_image_uri
+from sagemaker.predictor import csv_serializer, RealTimePredictor
+
+
+import tests.integ
+
+
+ROLE = "SageMakerRole"
+MODEL_NAME = "test-xgboost-model-{}".format(sagemaker_timestamp())
+DEFAULT_REGION = "us-west-2"
+DEFAULT_INSTANCE_TYPE = "ml.m5.xlarge"
+DEFAULT_INSTANCE_COUNT = 1
+XG_BOOST_MODEL_LOCAL_PATH = os.path.join(tests.integ.DATA_DIR, "xgboost_model", "xgb_model.tar.gz")
+
+TEST_VARIANT_1 = "Variant1"
+TEST_VARIANT_1_WEIGHT = 0.3
+
+TEST_VARIANT_2 = "Variant2"
+TEST_VARIANT_2_WEIGHT = 0.7
+
+VARIANT_TRAFFIC_SAMPLING_COUNT = 100
+DESIRED_CONFIDENCE_FOR_VARIANT_TRAFFIC_DISTRIBUTION = 0.999
+
+TEST_CSV_DATA = "42,42,42,42,42,42,42"
+
+SPARK_ML_MODEL_LOCAL_PATH = os.path.join(
+    tests.integ.DATA_DIR, "sparkml_model", "mleap_model.tar.gz"
+)
+SPARK_ML_DEFAULT_VARIANT_NAME = (
+    "AllTraffic"
+)  # default defined in src/sagemaker/session.py def production_variant
+SPARK_ML_WRONG_VARIANT_NAME = "WRONG_VARIANT"
+SPARK_ML_TEST_DATA = "1.0,C,38.0,71.5,1.0,female"
+SPARK_ML_MODEL_SCHEMA = json.dumps(
+    {
+        "input": [
+            {"name": "Pclass", "type": "float"},
+            {"name": "Embarked", "type": "string"},
+            {"name": "Age", "type": "float"},
+            {"name": "Fare", "type": "float"},
+            {"name": "SibSp", "type": "float"},
+            {"name": "Sex", "type": "string"},
+        ],
+        "output": {"name": "features", "struct": "vector", "type": "double"},
+    }
+)
+
+
+@pytest.fixture(scope="module")
+def multi_variant_endpoint(sagemaker_session):
+    """
+    Sets up the multi variant endpoint before the integration tests run.
+    Cleans up the multi variant endpoint after the integration tests run.
+    """
+    multi_variant_endpoint.endpoint_name = unique_name_from_base(
+        "integ-test-multi-variant-endpoint"
+    )
+    with tests.integ.timeout.timeout_and_delete_endpoint_by_name(
+        endpoint_name=multi_variant_endpoint.endpoint_name,
+        sagemaker_session=sagemaker_session,
+        hours=2,
+    ):
+
+        # Creating a model
+        bucket = sagemaker_session.default_bucket()
+        prefix = "sagemaker/DEMO-VariantTargeting"
+        model_url = S3Uploader.upload(
+            local_path=XG_BOOST_MODEL_LOCAL_PATH,
+            desired_s3_uri="s3://" + bucket + "/" + prefix,
+            session=sagemaker_session,
+        )
+
+        image_uri = get_image_uri(sagemaker_session.boto_session.region_name, "xgboost", "0.90-1")
+
+        multi_variant_endpoint_model = sagemaker_session.create_model(
+            name=MODEL_NAME,
+            role=ROLE,
+            container_defs={"Image": image_uri, "ModelDataUrl": model_url},
+        )
+
+        # Creating a multi variant endpoint
+        variant1 = production_variant(
+            model_name=MODEL_NAME,
+            instance_type=DEFAULT_INSTANCE_TYPE,
+            initial_instance_count=DEFAULT_INSTANCE_COUNT,
+            variant_name=TEST_VARIANT_1,
+            initial_weight=TEST_VARIANT_1_WEIGHT,
+        )
+        variant2 = production_variant(
+            model_name=MODEL_NAME,
+            instance_type=DEFAULT_INSTANCE_TYPE,
+            initial_instance_count=DEFAULT_INSTANCE_COUNT,
+            variant_name=TEST_VARIANT_2,
+            initial_weight=TEST_VARIANT_2_WEIGHT,
+        )
+        sagemaker_session.endpoint_from_production_variants(
+            name=multi_variant_endpoint.endpoint_name, production_variants=[variant1, variant2]
+        )
+
+        # Yield to run the integration tests
+        yield multi_variant_endpoint
+
+        # Cleanup resources
+        sagemaker_session.delete_model(multi_variant_endpoint_model)
+        sagemaker_session.sagemaker_client.delete_endpoint_config(
+            EndpointConfigName=multi_variant_endpoint.endpoint_name
+        )
+
+    # Validate resource cleanup
+    with pytest.raises(Exception) as exception:
+        sagemaker_session.sagemaker_client.describe_model(
+            ModelName=multi_variant_endpoint_model.name
+        )
+        assert "Could not find model" in str(exception.value)
+        sagemaker_session.sagemaker_client.describe_endpoint_config(
+            name=multi_variant_endpoint.endpoint_name
+        )
+        assert "Could not find endpoint" in str(exception.value)
+
+
+def test_target_variant_invocation(sagemaker_session, multi_variant_endpoint):
+
+    response = sagemaker_session.sagemaker_runtime_client.invoke_endpoint(
+        EndpointName=multi_variant_endpoint.endpoint_name,
+        Body=TEST_CSV_DATA,
+        ContentType=CONTENT_TYPE_CSV,
+        Accept=CONTENT_TYPE_CSV,
+        TargetVariant=TEST_VARIANT_1,
+    )
+    assert response["InvokedProductionVariant"] == TEST_VARIANT_1
+
+    response = sagemaker_session.sagemaker_runtime_client.invoke_endpoint(
+        EndpointName=multi_variant_endpoint.endpoint_name,
+        Body=TEST_CSV_DATA,
+        ContentType=CONTENT_TYPE_CSV,
+        Accept=CONTENT_TYPE_CSV,
+        TargetVariant=TEST_VARIANT_2,
+    )
+    assert response["InvokedProductionVariant"] == TEST_VARIANT_2
+
+
+def test_predict_invocation_with_target_variant(sagemaker_session, multi_variant_endpoint):
+    predictor = RealTimePredictor(
+        endpoint=multi_variant_endpoint.endpoint_name,
+        sagemaker_session=sagemaker_session,
+        serializer=csv_serializer,
+        content_type=CONTENT_TYPE_CSV,
+        accept=CONTENT_TYPE_CSV,
+    )
+
+    # Validate that no exception is raised when the target_variant is specified.
+    predictor.predict(TEST_CSV_DATA, target_variant=TEST_VARIANT_1)
+    predictor.predict(TEST_CSV_DATA, target_variant=TEST_VARIANT_2)
+
+
+def test_variant_traffic_distribution(sagemaker_session, multi_variant_endpoint):
+    variant_1_invocation_count = 0
+    variant_2_invocation_count = 0
+
+    for i in range(0, VARIANT_TRAFFIC_SAMPLING_COUNT):
+        response = sagemaker_session.sagemaker_runtime_client.invoke_endpoint(
+            EndpointName=multi_variant_endpoint.endpoint_name,
+            Body=TEST_CSV_DATA,
+            ContentType=CONTENT_TYPE_CSV,
+            Accept=CONTENT_TYPE_CSV,
+        )
+        if response["InvokedProductionVariant"] == TEST_VARIANT_1:
+            variant_1_invocation_count += 1
+        elif response["InvokedProductionVariant"] == TEST_VARIANT_2:
+            variant_2_invocation_count += 1
+
+    assert variant_1_invocation_count + variant_2_invocation_count == VARIANT_TRAFFIC_SAMPLING_COUNT
+
+    variant_1_invocation_percentage = float(variant_1_invocation_count) / float(
+        VARIANT_TRAFFIC_SAMPLING_COUNT
+    )
+    variant_1_margin_of_error = _compute_and_retrieve_margin_of_error(TEST_VARIANT_1_WEIGHT)
+    assert variant_1_invocation_percentage < TEST_VARIANT_1_WEIGHT + variant_1_margin_of_error
+    assert variant_1_invocation_percentage > TEST_VARIANT_1_WEIGHT - variant_1_margin_of_error
+
+    variant_2_invocation_percentage = float(variant_2_invocation_count) / float(
+        VARIANT_TRAFFIC_SAMPLING_COUNT
+    )
+    variant_2_margin_of_error = _compute_and_retrieve_margin_of_error(TEST_VARIANT_2_WEIGHT)
+    assert variant_2_invocation_percentage < TEST_VARIANT_2_WEIGHT + variant_2_margin_of_error
+    assert variant_2_invocation_percentage > TEST_VARIANT_2_WEIGHT - variant_2_margin_of_error
+
+
+def test_spark_ml_predict_invocation_with_target_variant(sagemaker_session):
+
+    spark_ml_model_endpoint_name = unique_name_from_base("integ-test-target-variant-sparkml")
+
+    model_data = sagemaker_session.upload_data(
+        path=SPARK_ML_MODEL_LOCAL_PATH, key_prefix="integ-test-data/sparkml/model"
+    )
+
+    with tests.integ.timeout.timeout_and_delete_endpoint_by_name(
+        spark_ml_model_endpoint_name, sagemaker_session
+    ):
+        spark_ml_model = SparkMLModel(
+            model_data=model_data,
+            role=ROLE,
+            sagemaker_session=sagemaker_session,
+            env={"SAGEMAKER_SPARKML_SCHEMA": SPARK_ML_MODEL_SCHEMA},
+        )
+
+        predictor = spark_ml_model.deploy(
+            DEFAULT_INSTANCE_COUNT,
+            DEFAULT_INSTANCE_TYPE,
+            endpoint_name=spark_ml_model_endpoint_name,
+        )
+
+        # Validate that no exception is raised when the target_variant is specified.
+        predictor.predict(SPARK_ML_TEST_DATA, target_variant=SPARK_ML_DEFAULT_VARIANT_NAME)
+
+        with pytest.raises(Exception) as exception_info:
+            predictor.predict(SPARK_ML_TEST_DATA, target_variant=SPARK_ML_WRONG_VARIANT_NAME)
+
+        assert "ValidationError" in str(exception_info.value)
+        assert SPARK_ML_WRONG_VARIANT_NAME in str(exception_info.value)
+
+        # cleanup resources
+        spark_ml_model.delete_model()
+        sagemaker_session.sagemaker_client.delete_endpoint_config(
+            EndpointConfigName=spark_ml_model_endpoint_name
+        )
+
+    # Validate resource cleanup
+    with pytest.raises(Exception) as exception:
+        sagemaker_session.sagemaker_client.describe_model(ModelName=spark_ml_model.name)
+        assert "Could not find model" in str(exception.value)
+        sagemaker_session.sagemaker_client.describe_endpoint_config(
+            name=spark_ml_model_endpoint_name
+        )
+        assert "Could not find endpoint" in str(exception.value)
+
+
+@pytest.mark.local_mode
+def test_target_variant_invocation_local_mode(sagemaker_session, multi_variant_endpoint):
+
+    if sagemaker_session._region_name is None:
+        sagemaker_session._region_name = DEFAULT_REGION
+
+    response = sagemaker_session.sagemaker_runtime_client.invoke_endpoint(
+        EndpointName=multi_variant_endpoint.endpoint_name,
+        Body=TEST_CSV_DATA,
+        ContentType=CONTENT_TYPE_CSV,
+        Accept=CONTENT_TYPE_CSV,
+        TargetVariant=TEST_VARIANT_1,
+    )
+    assert response["InvokedProductionVariant"] == TEST_VARIANT_1
+
+    response = sagemaker_session.sagemaker_runtime_client.invoke_endpoint(
+        EndpointName=multi_variant_endpoint.endpoint_name,
+        Body=TEST_CSV_DATA,
+        ContentType=CONTENT_TYPE_CSV,
+        Accept=CONTENT_TYPE_CSV,
+        TargetVariant=TEST_VARIANT_2,
+    )
+    assert response["InvokedProductionVariant"] == TEST_VARIANT_2
+
+
+@pytest.mark.local_mode
+def test_predict_invocation_with_target_variant_local_mode(
+    sagemaker_session, multi_variant_endpoint
+):
+
+    if sagemaker_session._region_name is None:
+        sagemaker_session._region_name = DEFAULT_REGION
+
+    predictor = RealTimePredictor(
+        endpoint=multi_variant_endpoint.endpoint_name,
+        sagemaker_session=sagemaker_session,
+        serializer=csv_serializer,
+        content_type=CONTENT_TYPE_CSV,
+        accept=CONTENT_TYPE_CSV,
+    )
+
+    # Validate that no exception is raised when the target_variant is specified.
+    predictor.predict(TEST_CSV_DATA, target_variant=TEST_VARIANT_1)
+    predictor.predict(TEST_CSV_DATA, target_variant=TEST_VARIANT_2)
+
+
+def _compute_and_retrieve_margin_of_error(variant_weight):
+    """
+    Computes the margin of error using the Wald method for computing the confidence
+    intervals of a binomial distribution.
+    """
+    z_value = st.norm.ppf(DESIRED_CONFIDENCE_FOR_VARIANT_TRAFFIC_DISTRIBUTION)
+    margin_of_error = (variant_weight * (1 - variant_weight)) / VARIANT_TRAFFIC_SAMPLING_COUNT
+    margin_of_error = z_value * math.sqrt(margin_of_error)
+    return margin_of_error
diff --git a/tests/unit/test_predictor.py b/tests/unit/test_predictor.py
