AWS-MLOps-module

This repo contains a terraform module with corresponding AWS resources that enable training, deploying and re-training AWS-hosted machine learning models with corresponding cloud infrastructure.

Warning: This repo is a basic template for MLOps resources on AWS. Please apply appropriate security enhancements for your project in production.

High-Level Solution Architecture

Example Usage

module "MLOps" {
 source  = "github.com/crederauk/terraform-aws-mlops-module?ref=<MODULE_VERSION>"
 resource_naming_prefix  = "your-app"
 data_s3_bucket          = "your-bucket-name"
 data_location_s3        = "/your_s3_folder/your_data.csv"
 model_target_variable   = "y"
 tuning_metric           = "AUC"
 retrain_model_bool      = true
 retraining_schedule     = "cron(0 8 1 * ? *)"
 algorithm_choice        = "classification"
 sagemaker_training_notebook_instance_type = "ml.m4.xlarge"
 inference_instance_count = 1
 preprocessing_script_path = "terraform/preprocess_data.py"
 tags = {
   my-tag-key = "my-tag-value"
 }
} 

Requirements

Name	Version
terraform	>= 1.0
archive	2.4.0
aws	>= 4.0
local	>= 2.4
random	>= 3.6

Providers

No providers.

Modules

Name	Source	Version
ecr	./modules/ecr	n/a
retraining_job	./modules/glue	n/a
s3	./modules/s3	n/a
sagemaker	./modules/sagemaker	n/a

Resources

No resources.

Inputs

Name	Description	Type	Default	Required
algorithm_choice	Machine learning problem type e.g classification, regression, clustering, anomaly, time_series	string	n/a	yes
data_location_s3	The path to a file in the data S3 bucket within which training data is located. Should be in the format //. If the file is in the root of the bucket, this should be set to / only.	string	n/a	yes
data_s3_bucket	The name of an S3 bucket within which training data is located.	string	n/a	yes
data_s3_bucket_encryption_key_arn	The ARN of the KMS key using which training data is encrypted in S3, if such a key exists.	string	""	no
inference_instance_count	The initial number of instances to serve the model endpoint	number	1	no
inference_instance_type	The instance type to be created for serving the model. Must be a valid EC2 instance type	string	"ml.t2.medium"	no
model_target_variable	The dependent variable (or 'label') that the model aims to predict. This should be a column name in the dataset.	string	n/a	yes
preprocessing_script_path	The path the user provides if they want to include their own data cleaning logic	string	null	no
resource_naming_prefix	Naming prefix to be applied to all resources created by this module	string	n/a	yes
retrain_model_bool	Boolean to indicate if the retraining pipeline shoud be added	bool	false	no
retraining_schedule	Cron expression for the model retraining frequency in the AWS format. See https://docs.aws.amazon.com/lambda/latest/dg/services-cloudwatchevents-expressions.html for details	string	""	no
sagemaker_training_notebook_instance_type	The Sagemaker notebook instance type to be created for training the model. Must be a valid EC2 instance type	string	"ml.t2.medium"	no
tags	Tags applied to your resources	map(string)	{}	no
tuning_metric	The metric user want to focus when tuning hyperparameter	string	n/a	yes

Outputs

Name	Description
config_bucket	Config S3 Bucket Terraform object
ecr	The ECR repository module outputs. Contains both 'repository' and 'encryption_key' attributes, that are the ECR repository and KMS encryption key Terraform object respectively.
ecr_repository	The ECR repository Terraform object.
glue	The Glue module outputs. Contains both 'retraining_job' and 'retraining_role' attributes, that are the Glue retraining job and IAM role Terraform objects respectively.
glue_retraining_role	The Glue retraining job IAM role Terraform object.
model_bucket	Model S3 Bucket Terraform object
s3_encryption_key	S3 encryption KMS key Terraform Object
sagemaker_endpoint_name	Sagemaker model endpoint name
sagemaker_model_name	Sagemaker model name
sagemaker_notebook_instance	Sagemaker notebook instance Terraform object

Destroying Resources

After creating the resources made using this the module, the resources:

• Sagemaker model
• Sagemaker Endpoint
• Endpoint configuration

Will not be tracked by your Terraform state file so if you decide to run "terraform destroy" these resources will not be deleted.

To destroy these resourses we recommend that you add these commands to your CI/CD pipeline:

aws sagemaker delete-model --model-name < demo-regression-model >
aws sagemaker delete-endpoint-config --endpoint-config-name < demo-regression-model-config >
aws sagemaker delete-endpoint --endpoint-name < demo-regression-model >    

But before this you will need to add your AWS credentials to the environment if you have not do already:

aws-access-key-id: < aws-access-key-id >
aws-secret-access-key: < aws-secret-access-key >
aws-region: < region >