stable-diffusion-aws-extension/middleware_api/lambda/inference_v2/sagemaker_endpoint_api.py

492 lines
20 KiB
Python

import logging
import os
import uuid
from dataclasses import dataclass
import boto3
from datetime import datetime
from botocore.exceptions import BotoCoreError, ClientError
from common.ddb_service.client import DynamoDbUtilsService
from multi_users._types import PARTITION_KEYS, Role
from multi_users.utils import get_user_roles, check_user_permissions, get_permissions_by_username
from _types import EndpointDeploymentJob
from _enums import EndpointStatus
sagemaker_endpoint_table = os.environ.get('DDB_ENDPOINT_DEPLOYMENT_TABLE_NAME')
user_table = os.environ.get('MULTI_USER_TABLE')
S3_BUCKET_NAME = os.environ.get('S3_BUCKET_NAME')
ASYNC_SUCCESS_TOPIC = os.environ.get('SNS_INFERENCE_SUCCESS')
ASYNC_ERROR_TOPIC = os.environ.get('SNS_INFERENCE_ERROR')
INFERENCE_ECR_IMAGE_URL = os.environ.get("INFERENCE_ECR_IMAGE_URL")
# logger = Logger(service="sagemaker_endpoint_api", level="INFO")
logger = logging.getLogger('inference_v2')
logger.setLevel(logging.INFO)
sagemaker = boto3.client('sagemaker')
ddb_service = DynamoDbUtilsService(logger=logger)
# GET /endpoints?name=SageMaker_Endpoint_Name&username=&filter=key:value,key:value
def list_all_sagemaker_endpoints(event, ctx):
_filter = {}
if 'queryStringParameters' not in event:
return {
'statusCode': '500',
'error': 'query parameter status and types are needed'
}
parameters = event['queryStringParameters']
endpoint_deployment_job_id = parameters['endpointDeploymentJobId'] if 'endpointDeploymentJobId' in parameters and \
parameters[
'endpointDeploymentJobId'] else None
username = parameters['username'] if 'username' in parameters and parameters['username'] else None
if endpoint_deployment_job_id:
scan_rows = ddb_service.query_items(sagemaker_endpoint_table,
key_values={'EndpointDeploymentJobId': endpoint_deployment_job_id},
)
else:
scan_rows = ddb_service.scan(sagemaker_endpoint_table, filters=None)
results = []
user_roles = []
if username:
user_roles = get_user_roles(ddb_service=ddb_service, user_table_name=user_table, username=username)
if 'x-auth' not in event or not event['x-auth']['username']:
return {
'statusCode': '400',
'error': 'no auth user provided'
}
requestor_name = event['x-auth']['username']
requestor_permissions = get_permissions_by_username(ddb_service, user_table, requestor_name)
requestor_created_roles_rows = ddb_service.scan(table=user_table, filters={
'kind': PARTITION_KEYS.role,
'creator': requestor_name
})
for requestor_created_roles_row in requestor_created_roles_rows:
role = Role(**ddb_service.deserialize(requestor_created_roles_row))
user_roles.append(role.sort_key)
for row in scan_rows:
# Compatible with fields used in older data, must be 'deleted'
if 'status' in row and row['status']['S'] == 'deleted':
row['endpoint_status']['S'] = EndpointStatus.DELETED.value
endpoint = EndpointDeploymentJob(**(ddb_service.deserialize(row)))
if 'sagemaker_endpoint' in requestor_permissions and \
'list' in requestor_permissions['sagemaker_endpoint'] and \
endpoint.owner_group_or_role and \
username and check_user_permissions(endpoint.owner_group_or_role, user_roles, username):
results.append(endpoint.__dict__)
elif 'sagemaker_endpoint' in requestor_permissions and 'all' in requestor_permissions['sagemaker_endpoint']:
results.append(endpoint.__dict__)
# Old data may never update the count of instances
for result in results:
if 'current_instance_count' not in result:
result['current_instance_count'] = 'N/A'
return {
'statusCode': 200,
'endpoints': results
}
@dataclass
class DeleteEndpointEvent:
delete_endpoint_list: [str]
username: str
# DELETE /endpoints
def delete_sagemaker_endpoints(raw_event, ctx):
try:
# delete sagemaker endpoints in the same loop
event = DeleteEndpointEvent(**raw_event)
creator_permissions = get_permissions_by_username(ddb_service, user_table, event.username)
if 'sagemaker_endpoint' not in creator_permissions or \
('all' not in creator_permissions['sagemaker_endpoint'] and 'create' not in creator_permissions['sagemaker_endpoint']):
return {
'statusCode': 400,
'errMsg': f"User {event.username} has no permission to delete a Sagemaker endpoint",
}
for endpoint in event.delete_endpoint_list:
try:
delete_response = sagemaker.delete_endpoint(EndpointName=endpoint)
logger.info(delete_response)
short_id = endpoint.split('-')[2]
sagemaker.delete_model(ModelName=f"infer-model-{short_id}")
sagemaker.delete_endpoint_config(EndpointConfigName=f"infer-config-{short_id}")
except (BotoCoreError, ClientError) as error:
if error.response['Error']['Code'] == 'ResourceNotFound':
logger.info("Endpoint not found, no need to delete.")
else:
logger.error(error)
return "Endpoint deleted"
except Exception as e:
logger.error(f"error deleting sagemaker endpoint with exception: {e}")
return f"error deleting sagemaker endpoint with exception: {e}"
@dataclass
class CreateEndpointEvent:
instance_type: str
initial_instance_count: str
autoscaling_enabled: bool
assign_to_roles: [str]
creator: str
endpoint_name: str = None
# POST /endpoints
def sagemaker_endpoint_create_api(raw_event, ctx):
logger.info(f"Received event: {raw_event}")
logger.info(f"Received ctx: {ctx}")
event = CreateEndpointEvent(**raw_event)
try:
endpoint_deployment_id = str(uuid.uuid4())
short_id = endpoint_deployment_id[:7]
sagemaker_model_name = f"infer-model-{short_id}"
sagemaker_endpoint_config = f"infer-config-{short_id}"
sagemaker_endpoint_name = f"infer-endpoint-{short_id}"
if event.endpoint_name:
sagemaker_endpoint_name = f'infer-endpoint-{event.endpoint_name}'
image_url = INFERENCE_ECR_IMAGE_URL
model_data_url = f"s3://{S3_BUCKET_NAME}/data/model.tar.gz"
s3_output_path = f"s3://{S3_BUCKET_NAME}/sagemaker_output/"
initial_instance_count = int(event.initial_instance_count) if event.initial_instance_count else 1
instance_type = event.instance_type
# check if roles have already linked to an endpoint?
creator_permissions = get_permissions_by_username(ddb_service, user_table, event.creator)
if 'sagemaker_endpoint' not in creator_permissions or \
('all' not in creator_permissions['sagemaker_endpoint'] and 'create' not in creator_permissions['sagemaker_endpoint']):
return {
'statusCode': 400,
'message': f"Creator {event.creator} has no permission to create Sagemaker",
}
endpoint_rows = ddb_service.scan(sagemaker_endpoint_table, filters=None)
for endpoint_row in endpoint_rows:
endpoint = EndpointDeploymentJob(**(ddb_service.deserialize(endpoint_row)))
# Compatible with fields used in older data, endpoint.status must be 'deleted'
if endpoint.endpoint_status != EndpointStatus.DELETED.value and endpoint.status != 'deleted':
for role in event.assign_to_roles:
if role in endpoint.owner_group_or_role:
return {
'statusCode': 400,
'message': f"role [{role}] has a valid endpoint already, not allow to have another one",
}
_create_sagemaker_model(sagemaker_model_name, image_url, model_data_url)
_create_sagemaker_endpoint_config(sagemaker_endpoint_config, s3_output_path, sagemaker_model_name,
initial_instance_count, instance_type)
logger.info('Creating new model endpoint...')
response = sagemaker.create_endpoint(
EndpointName=sagemaker_endpoint_name,
EndpointConfigName=sagemaker_endpoint_config
)
logger.info(f"Successfully created endpoint: {response}")
current_time = str(datetime.now())
raw = EndpointDeploymentJob(
EndpointDeploymentJobId=endpoint_deployment_id,
endpoint_name=sagemaker_endpoint_name,
startTime=current_time,
endpoint_status=EndpointStatus.CREATING.value,
max_instance_number=event.initial_instance_count,
autoscaling=event.autoscaling_enabled,
owner_group_or_role=event.assign_to_roles,
current_instance_count="0",
)
ddb_service.put_items(table=sagemaker_endpoint_table, entries=raw.__dict__)
logger.info(f"Successfully created endpoint deployment: {raw.__dict__}")
return {
'statusCode': 200,
'message': f"Endpoint deployment started: {sagemaker_endpoint_name}",
'data': raw.__dict__
}
except Exception as e:
logger.error(e)
return {
'statusCode': 200,
'message': str(e),
}
# lambda: handle sagemaker events
def sagemaker_endpoint_events(event, context):
logger.info(event)
endpoint_name = event['detail']['EndpointName']
endpoint_status = event['detail']['EndpointStatus']
endpoint = get_endpoint_with_endpoint_name(endpoint_name)
if not endpoint:
# maybe the endpoint is not created by sde or already deleted
logger.error(f"No matching DynamoDB record found for endpoint: {endpoint_name}")
return {'statusCode': 200}
endpoint_deployment_job_id = endpoint['EndpointDeploymentJobId']
business_status = get_business_status(endpoint_status)
update_endpoint_field(endpoint_deployment_job_id, 'endpoint_status', business_status)
# update the instance count if the endpoint is not deleting or deleted
if business_status not in [EndpointStatus.DELETING.value, EndpointStatus.DELETED.value]:
status = sagemaker.describe_endpoint(EndpointName=endpoint_name)
logger.info(f"Endpoint status: {status}")
if 'ProductionVariants' in status:
instance_count = status['ProductionVariants'][0]['CurrentInstanceCount']
update_endpoint_field(endpoint_deployment_job_id, 'current_instance_count', instance_count)
else:
# sometime sagemaker don't send deleted event, so just use deleted status when deleting
update_endpoint_field(endpoint_deployment_job_id, 'endpoint_status', EndpointStatus.DELETED.value)
update_endpoint_field(endpoint_deployment_job_id, 'current_instance_count', 0)
# if endpoint is deleted, update the instance count to 0 and delete the config and model
if business_status == EndpointStatus.DELETED.value:
try:
endpoint_config_name = event['detail']['EndpointConfigName']
model_name = event['detail']['ModelName']
sagemaker.delete_endpoint_config(EndpointConfigName=endpoint_config_name)
sagemaker.delete_model(ModelName=model_name)
except Exception as e:
logger.error(f"error deleting endpoint config and model with exception: {e}")
if business_status == EndpointStatus.IN_SERVICE.value:
# if it is the first time in service
if 'endTime' not in endpoint:
check_and_enable_autoscaling(endpoint, 'prod')
current_time = str(datetime.now())
update_endpoint_field(endpoint_deployment_job_id, 'endTime', current_time)
if business_status == EndpointStatus.FAILED.value:
update_endpoint_field(endpoint_deployment_job_id, 'error', event['FailureReason'])
return {'statusCode': 200}
def _create_sagemaker_model(name, image_url, model_data_url):
primary_container = {
'Image': image_url,
'ModelDataUrl': model_data_url,
'Environment': {
'EndpointID': 'OUR_ID'
},
}
logger.info(f"Creating model resource PrimaryContainer: {primary_container}")
response = sagemaker.create_model(
ModelName=name,
PrimaryContainer=primary_container,
ExecutionRoleArn=os.environ.get("EXECUTION_ROLE_ARN"),
)
logger.info(f"Successfully created model resource: {response}")
def _create_sagemaker_endpoint_config(endpoint_config_name, s3_output_path, model_name, initial_instance_count, instance_type):
async_inference_config = {
"OutputConfig": {
"S3OutputPath": s3_output_path,
"NotificationConfig": {
"SuccessTopic": ASYNC_SUCCESS_TOPIC,
"ErrorTopic": ASYNC_ERROR_TOPIC
}
},
"ClientConfig": {
# (Optional) Specify the max number of inflight invocations per instance
# If no value is provided, Amazon SageMaker will choose an optimal value for you
"MaxConcurrentInvocationsPerInstance": 1
}
}
production_variants = [
{
'VariantName': 'prod',
'ModelName': model_name,
'InitialInstanceCount': initial_instance_count,
'InstanceType': instance_type
}
]
logger.info(f"Creating endpoint configuration AsyncInferenceConfig: {async_inference_config}")
logger.info(f"Creating endpoint configuration ProductionVariants: {production_variants}")
response = sagemaker.create_endpoint_config(
EndpointConfigName=endpoint_config_name,
AsyncInferenceConfig=async_inference_config,
ProductionVariants=production_variants
)
logger.info(f"Successfully created endpoint configuration: {response}")
def check_and_enable_autoscaling(item, variant_name):
autoscaling = item['autoscaling']['BOOL']
endpoint_name = item['endpoint_name']['S']
max_instance_number = item['max_instance_number']['S']
logger.info(f"autoscaling: {autoscaling}")
logger.info(f"endpoint_name: {endpoint_name}")
logger.info(f"max_instance_number: {max_instance_number}")
if str(autoscaling) == 'True':
if max_instance_number.isdigit():
enable_autoscaling(endpoint_name, variant_name, 0, int(max_instance_number))
else:
logger.info(f"the max_number field is not digit, just fallback to 1")
enable_autoscaling(endpoint_name, variant_name, 0, 1)
else:
logger.info(f'autoscaling_enabled is {autoscaling}, no need to enable autoscaling')
def enable_autoscaling(endpoint_name, variant_name, low_value, high_value):
client = boto3.client('application-autoscaling')
# Register scalable target
response = client.register_scalable_target(
ServiceNamespace='sagemaker',
ResourceId='endpoint/' + endpoint_name + '/variant/' + variant_name,
ScalableDimension='sagemaker:variant:DesiredInstanceCount',
MinCapacity=low_value,
MaxCapacity=high_value,
)
# Define scaling policy
response = client.put_scaling_policy(
PolicyName="Invocations-ScalingPolicy",
ServiceNamespace="sagemaker", # The namespace of the AWS service that provides the resource.
ResourceId='endpoint/' + endpoint_name + '/variant/' + variant_name, # Endpoint name
ScalableDimension="sagemaker:variant:DesiredInstanceCount", # SageMaker supports only Instance Count
PolicyType="TargetTrackingScaling", # 'StepScaling'|'TargetTrackingScaling'
TargetTrackingScalingPolicyConfiguration={
"TargetValue": 5.0,
# The target value for the metric. - here the metric is - SageMakerVariantInvocationsPerInstance
"CustomizedMetricSpecification": {
"MetricName": "ApproximateBacklogSizePerInstance",
"Namespace": "AWS/SageMaker",
"Dimensions": [{"Name": "EndpointName", "Value": endpoint_name}],
"Statistic": "Average",
},
"ScaleInCooldown": 180,
# The cooldown period helps you prevent your Auto Scaling group from launching or terminating
"ScaleOutCooldown": 60
# ScaleOutCooldown - The amount of time, in seconds, after a scale out activity completes before another
# scale out activity can start.
},
)
step_policy_response = client.put_scaling_policy(
PolicyName="HasBacklogWithoutCapacity-ScalingPolicy",
ServiceNamespace="sagemaker", # The namespace of the service that provides the resource.
ResourceId='endpoint/' + endpoint_name + '/variant/' + variant_name,
ScalableDimension="sagemaker:variant:DesiredInstanceCount", # SageMaker supports only Instance Count
PolicyType="StepScaling", # 'StepScaling' or 'TargetTrackingScaling'
StepScalingPolicyConfiguration={
"AdjustmentType": "ChangeInCapacity",
# Specifies whether the ScalingAdjustment value in the StepAdjustment property is an absolute number or a
# percentage of the current capacity.
"MetricAggregationType": "Average", # The aggregation type for the CloudWatch metrics.
"Cooldown": 180, # The amount of time, in seconds, to wait for a previous scaling activity to take effect.
"StepAdjustments": # A set of adjustments that enable you to scale based on the size of the alarm breach.
[
{
"MetricIntervalLowerBound": 0,
"ScalingAdjustment": 1
}
]
},
)
cw_client = boto3.client('cloudwatch')
cw_client.put_metric_alarm(
AlarmName='stable-diffusion-hasbacklogwithoutcapacity-alarm',
MetricName='HasBacklogWithoutCapacity',
Namespace='AWS/SageMaker',
Statistic='Average',
EvaluationPeriods=2,
DatapointsToAlarm=2,
Threshold=1,
ComparisonOperator='GreaterThanOrEqualToThreshold',
TreatMissingData='missing',
Dimensions=[
{'Name': 'EndpointName', 'Value': endpoint_name},
],
Period=60,
AlarmActions=[step_policy_response['PolicyARN']]
)
print(f"Autoscaling has been enabled for the endpoint: {endpoint_name}")
def update_endpoint_field(endpoint_deployment_job_id, field_name, field_value):
logger.info(f"Updating DynamoDB {field_name} to {field_value} for: {endpoint_deployment_job_id}")
ddb_service.update_item(
table=sagemaker_endpoint_table,
key={'EndpointDeploymentJobId': endpoint_deployment_job_id['S']},
field_name=field_name,
value=field_value
)
def get_endpoint_with_endpoint_name(endpoint_name):
try:
record_list = ddb_service.scan(table=sagemaker_endpoint_table, filters={
'endpoint_name': endpoint_name,
})
# not include deleted endpoint anywhere
record_list = [item for item in record_list if item['endpoint_status']['S'] != EndpointStatus.DELETED.value]
if len(record_list) == 0:
logger.error("There is no endpoint deployment job info item with endpoint name:" + endpoint_name)
return {}
logger.info(record_list[0])
return record_list[0]
except Exception as e:
logger.error(e)
return {}
def get_business_status(status):
"""
Convert SageMaker endpoint status to business status
:param status: EventBridge event status(upper case)
:return: business status
"""
switcher = {
"IN_SERVICE": EndpointStatus.IN_SERVICE.value,
"CREATING": EndpointStatus.CREATING.value,
"DELETED": EndpointStatus.DELETED.value,
"FAILED": EndpointStatus.FAILED.value,
"UPDATING": EndpointStatus.UPDATING.value,
"DELETING": EndpointStatus.DELETING.value,
}
return switcher.get(status, status)