Add CloudFormation support for SageMaker Endpoint Configs and Endpoints (#3863)

* Create SageMaker EndpointConfig with CloudFormation

Implement attributes for SM Endpoint Configs with CloudFormation

Delete SM Endpoint Configs with CloudFormation

Update SM Endpoint Configs with CloudFormation

* Fix typos in SM CF Model update test and refactor helper function for CF stack outputs

* Fixup weird commas in SM CF Test Configs from using black

* Create SageMaker Endpoints with CloudFormation

* Fix typos in SM CF update tests
This commit is contained in:
Zach Churchill 2021-04-17 08:49:46 -04:00 committed by GitHub
parent f6dda54a6c
commit 9b3e932822
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 450 additions and 54 deletions

View File

@ -140,7 +140,7 @@ class FakeTrainingJob(BaseObject):
)
class FakeEndpoint(BaseObject):
class FakeEndpoint(BaseObject, CloudFormationModel):
def __init__(
self,
region_name,
@ -184,8 +184,70 @@ class FakeEndpoint(BaseObject):
+ endpoint_name
)
@property
def physical_resource_id(self):
return self.endpoint_arn
class FakeEndpointConfig(BaseObject):
def get_cfn_attribute(self, attribute_name):
# https://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuide/aws-resource-sagemaker-endpoint.html#aws-resource-sagemaker-endpoint-return-values
from moto.cloudformation.exceptions import UnformattedGetAttTemplateException
if attribute_name == "EndpointName":
return self.endpoint_name
raise UnformattedGetAttTemplateException()
@staticmethod
def cloudformation_name_type():
return None
@staticmethod
def cloudformation_type():
# https://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuide/aws-resource-sagemaker-endpoint.html
return "AWS::SageMaker::Endpoint"
@classmethod
def create_from_cloudformation_json(
cls, resource_name, cloudformation_json, region_name
):
sagemaker_backend = sagemaker_backends[region_name]
# Get required properties from provided CloudFormation template
properties = cloudformation_json["Properties"]
endpoint_config_name = properties["EndpointConfigName"]
endpoint = sagemaker_backend.create_endpoint(
endpoint_name=resource_name,
endpoint_config_name=endpoint_config_name,
tags=properties.get("Tags", []),
)
return endpoint
@classmethod
def update_from_cloudformation_json(
cls, original_resource, new_resource_name, cloudformation_json, region_name,
):
# Changes to the Endpoint will not change resource name
cls.delete_from_cloudformation_json(
original_resource.endpoint_arn, cloudformation_json, region_name
)
new_resource = cls.create_from_cloudformation_json(
original_resource.endpoint_name, cloudformation_json, region_name
)
return new_resource
@classmethod
def delete_from_cloudformation_json(
cls, resource_name, cloudformation_json, region_name
):
# Get actual name because resource_name actually provides the ARN
# since the Physical Resource ID is the ARN despite SageMaker
# using the name for most of its operations.
endpoint_name = resource_name.split("/")[-1]
sagemaker_backends[region_name].delete_endpoint(endpoint_name)
class FakeEndpointConfig(BaseObject, CloudFormationModel):
def __init__(
self,
region_name,
@ -308,6 +370,70 @@ class FakeEndpointConfig(BaseObject):
+ model_name
)
@property
def physical_resource_id(self):
return self.endpoint_config_arn
def get_cfn_attribute(self, attribute_name):
# https://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuide/aws-resource-sagemaker-endpointconfig.html#aws-resource-sagemaker-endpointconfig-return-values
from moto.cloudformation.exceptions import UnformattedGetAttTemplateException
if attribute_name == "EndpointConfigName":
return self.endpoint_config_name
raise UnformattedGetAttTemplateException()
@staticmethod
def cloudformation_name_type():
return None
@staticmethod
def cloudformation_type():
# https://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuide/aws-resource-sagemaker-endpointconfig.html
return "AWS::SageMaker::EndpointConfig"
@classmethod
def create_from_cloudformation_json(
cls, resource_name, cloudformation_json, region_name
):
sagemaker_backend = sagemaker_backends[region_name]
# Get required properties from provided CloudFormation template
properties = cloudformation_json["Properties"]
production_variants = properties["ProductionVariants"]
endpoint_config = sagemaker_backend.create_endpoint_config(
endpoint_config_name=resource_name,
production_variants=production_variants,
data_capture_config=properties.get("DataCaptureConfig", {}),
kms_key_id=properties.get("KmsKeyId"),
tags=properties.get("Tags", []),
)
return endpoint_config
@classmethod
def update_from_cloudformation_json(
cls, original_resource, new_resource_name, cloudformation_json, region_name,
):
# Most changes to the endpoint config will change resource name for EndpointConfigs
cls.delete_from_cloudformation_json(
original_resource.endpoint_config_arn, cloudformation_json, region_name
)
new_resource = cls.create_from_cloudformation_json(
new_resource_name, cloudformation_json, region_name
)
return new_resource
@classmethod
def delete_from_cloudformation_json(
cls, resource_name, cloudformation_json, region_name
):
# Get actual name because resource_name actually provides the ARN
# since the Physical Resource ID is the ARN despite SageMaker
# using the name for most of its operations.
endpoint_config_name = resource_name.split("/")[-1]
sagemaker_backends[region_name].delete_endpoint_config(endpoint_config_name)
class Model(BaseObject, CloudFormationModel):
def __init__(

View File

@ -42,6 +42,14 @@ class TestConfig:
def get_cloudformation_template(self, include_outputs=True, **kwargs):
pass
def run_setup_procedure(self, sagemaker_client):
"""Provides a method to set up resources with a SageMaker client.
Note: This procedure should be called while within a `mock_sagemaker`
context so that no actual resources are created with the sagemaker_client.
"""
pass
class NotebookInstanceTestConfig(TestConfig):
"""Test configuration for SageMaker Notebook Instances."""
@ -186,6 +194,131 @@ class ModelTestConfig(TestConfig):
if include_outputs:
template["Outputs"] = {
"Arn": {"Value": {"Ref": self.resource_name}},
"Name": {"Value": {"Fn::GetAtt": [self.resource_name, "ModelName"],}},
"Name": {"Value": {"Fn::GetAtt": [self.resource_name, "ModelName"]}},
}
return json.dumps(template)
class EndpointConfigTestConfig(TestConfig):
"""Test configuration for SageMaker Endpoint Configs."""
@property
def resource_name(self):
return "TestEndpointConfig"
@property
def describe_function_name(self):
return "describe_endpoint_config"
@property
def name_parameter(self):
return "EndpointConfigName"
@property
def arn_parameter(self):
return "EndpointConfigArn"
def get_cloudformation_template(self, include_outputs=True, **kwargs):
num_production_variants = kwargs.get("num_production_variants", 1)
production_variants = [
{
"InitialInstanceCount": 1,
"InitialVariantWeight": 1,
"InstanceType": "ml.c4.xlarge",
"ModelName": self.resource_name,
"VariantName": "variant-name-{}".format(i),
}
for i in range(num_production_variants)
]
template = {
"AWSTemplateFormatVersion": "2010-09-09",
"Resources": {
self.resource_name: {
"Type": "AWS::SageMaker::EndpointConfig",
"Properties": {"ProductionVariants": production_variants},
},
},
}
if include_outputs:
template["Outputs"] = {
"Arn": {"Value": {"Ref": self.resource_name}},
"Name": {
"Value": {"Fn::GetAtt": [self.resource_name, "EndpointConfigName"]}
},
}
return json.dumps(template)
def run_setup_procedure(self, sagemaker_client):
"""Adds Model that can be referenced in the CloudFormation template."""
sagemaker_client.create_model(
ModelName=self.resource_name,
ExecutionRoleArn="arn:aws:iam::{}:role/FakeRole".format(ACCOUNT_ID),
PrimaryContainer={
"Image": "404615174143.dkr.ecr.us-east-2.amazonaws.com/linear-learner:1",
},
)
class EndpointTestConfig(TestConfig):
"""Test configuration for SageMaker Endpoints."""
@property
def resource_name(self):
return "TestEndpoint"
@property
def describe_function_name(self):
return "describe_endpoint"
@property
def name_parameter(self):
return "EndpointName"
@property
def arn_parameter(self):
return "EndpointArn"
def get_cloudformation_template(self, include_outputs=True, **kwargs):
endpoint_config_name = kwargs.get("endpoint_config_name", self.resource_name)
template = {
"AWSTemplateFormatVersion": "2010-09-09",
"Resources": {
self.resource_name: {
"Type": "AWS::SageMaker::Endpoint",
"Properties": {"EndpointConfigName": endpoint_config_name},
},
},
}
if include_outputs:
template["Outputs"] = {
"Arn": {"Value": {"Ref": self.resource_name}},
"Name": {"Value": {"Fn::GetAtt": [self.resource_name, "EndpointName"]}},
}
return json.dumps(template)
def run_setup_procedure(self, sagemaker_client):
"""Adds Model and Endpoint Config that can be referenced in the CloudFormation template."""
sagemaker_client.create_model(
ModelName=self.resource_name,
ExecutionRoleArn="arn:aws:iam::{}:role/FakeRole".format(ACCOUNT_ID),
PrimaryContainer={
"Image": "404615174143.dkr.ecr.us-east-2.amazonaws.com/linear-learner:1",
},
)
sagemaker_client.create_endpoint_config(
EndpointConfigName=self.resource_name,
ProductionVariants=[
{
"InitialInstanceCount": 1,
"InitialVariantWeight": 1,
"InstanceType": "ml.c4.xlarge",
"ModelName": self.resource_name,
"VariantName": "variant-name-1",
},
],
)

View File

@ -5,25 +5,44 @@ import sure # noqa
from botocore.exceptions import ClientError
from moto import mock_cloudformation, mock_sagemaker
from moto.sts.models import ACCOUNT_ID
from .cloudformation_test_configs import (
NotebookInstanceTestConfig,
NotebookInstanceLifecycleConfigTestConfig,
ModelTestConfig,
EndpointConfigTestConfig,
EndpointTestConfig,
)
def _get_stack_outputs(cf_client, stack_name):
"""Returns the outputs for the first entry in describe_stacks."""
stack_description = cf_client.describe_stacks(StackName=stack_name)["Stacks"][0]
return {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
@mock_cloudformation
@mock_sagemaker
@pytest.mark.parametrize(
"test_config",
[
NotebookInstanceTestConfig(),
NotebookInstanceLifecycleConfigTestConfig(),
ModelTestConfig(),
EndpointConfigTestConfig(),
EndpointTestConfig(),
],
)
def test_sagemaker_cloudformation_create(test_config):
cf = boto3.client("cloudformation", region_name="us-east-1")
sm = boto3.client("sagemaker", region_name="us-east-1")
# Utilize test configuration to set-up any mock SageMaker resources
test_config.run_setup_procedure(sm)
stack_name = "{}_stack".format(test_config.resource_name)
cf.create_stack(
@ -46,22 +65,23 @@ def test_sagemaker_cloudformation_create(test_config):
NotebookInstanceTestConfig(),
NotebookInstanceLifecycleConfigTestConfig(),
ModelTestConfig(),
EndpointConfigTestConfig(),
EndpointTestConfig(),
],
)
def test_sagemaker_cloudformation_get_attr(test_config):
cf = boto3.client("cloudformation", region_name="us-east-1")
sm = boto3.client("sagemaker", region_name="us-east-1")
# Utilize test configuration to set-up any mock SageMaker resources
test_config.run_setup_procedure(sm)
# Create stack and get description for output values
stack_name = "{}_stack".format(test_config.resource_name)
cf.create_stack(
StackName=stack_name, TemplateBody=test_config.get_cloudformation_template()
)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
# Using the describe function, ensure output ARN matches resource ARN
resource_description = getattr(sm, test_config.describe_function_name)(
@ -81,22 +101,24 @@ def test_sagemaker_cloudformation_get_attr(test_config):
"Notebook Instance Lifecycle Config does not exist",
),
(ModelTestConfig(), "Could not find model"),
(EndpointConfigTestConfig(), "Could not find endpoint configuration"),
(EndpointTestConfig(), "Could not find endpoint"),
],
)
def test_sagemaker_cloudformation_notebook_instance_delete(test_config, error_message):
cf = boto3.client("cloudformation", region_name="us-east-1")
sm = boto3.client("sagemaker", region_name="us-east-1")
# Utilize test configuration to set-up any mock SageMaker resources
test_config.run_setup_procedure(sm)
# Create stack and verify existence
stack_name = "{}_stack".format(test_config.resource_name)
cf.create_stack(
StackName=stack_name, TemplateBody=test_config.get_cloudformation_template()
)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: outputs["Name"]}
)
@ -119,7 +141,7 @@ def test_sagemaker_cloudformation_notebook_instance_update():
test_config = NotebookInstanceTestConfig()
# Set up template for stack with initial and update instance types
# Set up template for stack with two different instance types
stack_name = "{}_stack".format(test_config.resource_name)
initial_instance_type = "ml.c4.xlarge"
updated_instance_type = "ml.c4.4xlarge"
@ -132,24 +154,18 @@ def test_sagemaker_cloudformation_notebook_instance_update():
# Create stack with initial template and check attributes
cf.create_stack(StackName=stack_name, TemplateBody=initial_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
initial_notebook_name = outputs["Name"]
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: initial_notebook_name}
)
initial_instance_type.should.equal(resource_description["InstanceType"])
# Update stack with new instance type and check attributes
# Update stack and check attributes
cf.update_stack(StackName=stack_name, TemplateBody=updated_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
updated_notebook_name = outputs["Name"]
updated_notebook_name.should.equal(initial_notebook_name)
@ -167,7 +183,7 @@ def test_sagemaker_cloudformation_notebook_instance_lifecycle_config_update():
test_config = NotebookInstanceLifecycleConfigTestConfig()
# Set up template for stack with initial and update instance types
# Set up template for stack with two different OnCreate scripts
stack_name = "{}_stack".format(test_config.resource_name)
initial_on_create_script = "echo Hello World"
updated_on_create_script = "echo Goodbye World"
@ -180,11 +196,8 @@ def test_sagemaker_cloudformation_notebook_instance_lifecycle_config_update():
# Create stack with initial template and check attributes
cf.create_stack(StackName=stack_name, TemplateBody=initial_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
initial_config_name = outputs["Name"]
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: initial_config_name}
@ -194,13 +207,10 @@ def test_sagemaker_cloudformation_notebook_instance_lifecycle_config_update():
resource_description["OnCreate"][0]["Content"]
)
# Update stack with new instance type and check attributes
# Update stack and check attributes
cf.update_stack(StackName=stack_name, TemplateBody=updated_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
outputs = _get_stack_outputs(cf, stack_name)
updated_config_name = outputs["Name"]
updated_config_name.should.equal(initial_config_name)
@ -221,7 +231,7 @@ def test_sagemaker_cloudformation_model_update():
test_config = ModelTestConfig()
# Set up template for stack with initial and update instance types
# Set up template for stack with two different image versions
stack_name = "{}_stack".format(test_config.resource_name)
image = "404615174143.dkr.ecr.us-east-2.amazonaws.com/kmeans:{}"
initial_image_version = 1
@ -235,32 +245,159 @@ def test_sagemaker_cloudformation_model_update():
# Create stack with initial template and check attributes
cf.create_stack(StackName=stack_name, TemplateBody=initial_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
inital_model_name = outputs["Name"]
outputs = _get_stack_outputs(cf, stack_name)
initial_model_name = outputs["Name"]
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: inital_model_name}
**{test_config.name_parameter: initial_model_name}
)
resource_description["PrimaryContainer"]["Image"].should.equal(
image.format(initial_image_version)
)
# Update stack with new instance type and check attributes
# Update stack and check attributes
cf.update_stack(StackName=stack_name, TemplateBody=updated_template_json)
stack_description = cf.describe_stacks(StackName=stack_name)["Stacks"][0]
outputs = {
output["OutputKey"]: output["OutputValue"]
for output in stack_description["Outputs"]
}
updated_notebook_name = outputs["Name"]
updated_notebook_name.should_not.equal(inital_model_name)
outputs = _get_stack_outputs(cf, stack_name)
updated_model_name = outputs["Name"]
updated_model_name.should_not.equal(initial_model_name)
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: updated_notebook_name}
**{test_config.name_parameter: updated_model_name}
)
resource_description["PrimaryContainer"]["Image"].should.equal(
image.format(updated_image_version)
)
@mock_cloudformation
@mock_sagemaker
def test_sagemaker_cloudformation_endpoint_config_update():
cf = boto3.client("cloudformation", region_name="us-east-1")
sm = boto3.client("sagemaker", region_name="us-east-1")
test_config = EndpointConfigTestConfig()
# Utilize test configuration to set-up any mock SageMaker resources
test_config.run_setup_procedure(sm)
# Set up template for stack with two different production variant counts
stack_name = "{}_stack".format(test_config.resource_name)
initial_num_production_variants = 1
updated_num_production_variants = 2
initial_template_json = test_config.get_cloudformation_template(
num_production_variants=initial_num_production_variants
)
updated_template_json = test_config.get_cloudformation_template(
num_production_variants=updated_num_production_variants
)
# Create stack with initial template and check attributes
cf.create_stack(StackName=stack_name, TemplateBody=initial_template_json)
outputs = _get_stack_outputs(cf, stack_name)
initial_endpoint_config_name = outputs["Name"]
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: initial_endpoint_config_name}
)
len(resource_description["ProductionVariants"]).should.equal(
initial_num_production_variants
)
# Update stack and check attributes
cf.update_stack(StackName=stack_name, TemplateBody=updated_template_json)
outputs = _get_stack_outputs(cf, stack_name)
updated_endpoint_config_name = outputs["Name"]
updated_endpoint_config_name.should_not.equal(initial_endpoint_config_name)
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: updated_endpoint_config_name}
)
len(resource_description["ProductionVariants"]).should.equal(
updated_num_production_variants
)
@mock_cloudformation
@mock_sagemaker
def test_sagemaker_cloudformation_endpoint_update():
cf = boto3.client("cloudformation", region_name="us-east-1")
sm = boto3.client("sagemaker", region_name="us-east-1")
test_config = EndpointTestConfig()
# Set up template for stack with two different endpoint config names
stack_name = "{}_stack".format(test_config.resource_name)
initial_endpoint_config_name = test_config.resource_name
updated_endpoint_config_name = "updated-endpoint-config-name"
initial_template_json = test_config.get_cloudformation_template(
endpoint_config_name=initial_endpoint_config_name
)
updated_template_json = test_config.get_cloudformation_template(
endpoint_config_name=updated_endpoint_config_name
)
# Create SM resources and stack with initial template and check attributes
sm.create_model(
ModelName=initial_endpoint_config_name,
ExecutionRoleArn="arn:aws:iam::{}:role/FakeRole".format(ACCOUNT_ID),
PrimaryContainer={
"Image": "404615174143.dkr.ecr.us-east-2.amazonaws.com/linear-learner:1",
},
)
sm.create_endpoint_config(
EndpointConfigName=initial_endpoint_config_name,
ProductionVariants=[
{
"InitialInstanceCount": 1,
"InitialVariantWeight": 1,
"InstanceType": "ml.c4.xlarge",
"ModelName": initial_endpoint_config_name,
"VariantName": "variant-name-1",
},
],
)
cf.create_stack(StackName=stack_name, TemplateBody=initial_template_json)
outputs = _get_stack_outputs(cf, stack_name)
initial_endpoint_name = outputs["Name"]
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: initial_endpoint_name}
)
resource_description["EndpointConfigName"].should.match(
initial_endpoint_config_name
)
# Create additional SM resources and update stack
sm.create_model(
ModelName=updated_endpoint_config_name,
ExecutionRoleArn="arn:aws:iam::{}:role/FakeRole".format(ACCOUNT_ID),
PrimaryContainer={
"Image": "404615174143.dkr.ecr.us-east-2.amazonaws.com/linear-learner:1",
},
)
sm.create_endpoint_config(
EndpointConfigName=updated_endpoint_config_name,
ProductionVariants=[
{
"InitialInstanceCount": 1,
"InitialVariantWeight": 1,
"InstanceType": "ml.c4.xlarge",
"ModelName": updated_endpoint_config_name,
"VariantName": "variant-name-1",
},
],
)
cf.update_stack(StackName=stack_name, TemplateBody=updated_template_json)
outputs = _get_stack_outputs(cf, stack_name)
updated_endpoint_name = outputs["Name"]
updated_endpoint_name.should.equal(initial_endpoint_name)
resource_description = getattr(sm, test_config.describe_function_name)(
**{test_config.name_parameter: updated_endpoint_name}
)
resource_description["EndpointConfigName"].should.match(
updated_endpoint_config_name
)