CloudZone Streamlines and Adapts Machine Learning Model Lifecycles in the Cloud for LSports

LSports had a large data repository in its data warehouse and wanted to create ML models from this data to improve the accuracy of its largely human-driven predictions. The company also wanted to streamline operations so it could scale up to meet increasing demand for its services. The company’s in-house approach to training ML models was slow and took large amounts of processing power that its team’s individual laptops struggled to handle, making it nearly impossible to leverage the large data sets required to improve accuracy. Much like the athletes at the heart of its business, LSports hoped to move faster, stronger, and smarter. And it planned to achieve this by moving to a cloud-powered ML platform.

The challenge was finding an ML platform that could efficiently handle the complexities of pre-processing, training, and deploying models at scale, while also simplifying the ML development lifecycle with a unified system for managing models, tracking versions, and ensuring transparency in model evolution. To address these challenges and capitalize on the opportunities in the rapidly evolving sports betting industry, LSports engaged CloudZone to take its sports betting game to the next level.

To help LSports meet its goals for efficiency and accuracy, CloudZone developed a full ML pipeline which was designed and built on top of Amazon SageMaker, a fully managed service that brings together a broad set of tools to perform various ML tasks like data processing, training, evaluation and inference on high volumes of data at low cost. With Amazon SageMaker, you can build, train and deploy ML models at scale using tools and strategies like notebooks, debuggers, profilers, pipelines, MLOps, and more—all in one platform.

CloudZone’s solution implemented MLOps methodologies, defining the responsibility of each team that has a hand in the model lifecycle, and delivered a training pipeline that controls data preparation, model training, and model deployment. All stages were implemented on Amazon SageMaker Pipeline, and the pipeline was wrapped in Amazon Web Services (AWS)-native tools to control triggers and manual approvals. The data science team also integrated Amazon SageMaker Studio as an integrated development environment (IDE) to leverage purpose-built tools for ML development, like managing experiments, explainability capabilities, data visualization, and more. Leveraging the power of AWS helped the organization’s data science teams move faster as they were no longer limited by the computational power of their laptops.

Key steps in the CloudZone’s ML pipeline include:

1. Triggering the CI pipeline: Users can initiate the pipeline using a schedule, webhook, or manual trigger, which starts the continuous integration (CI) process.
2. Lambda function execution: An AWS Lambda function is executed to trigger the Amazon SageMaker pipeline and waits for it to be complete.
3. Amazon SageMaker pipeline processing:
   1. Processing Step (Data Wrangler): Prepares and processes the data for training.
   2. Training Step (Model Training): Trains the ML model using the prepared data.
   3. Processing Step (Model Evaluation): Evaluates the trained model's performance.
   4. Condition Step (Model Accuracy): Checks if the model meets the accuracy requirements.
4. Model registration and approval:
   1. If the model meets the accuracy criteria, it proceeds to the Create Model Step and Register Model Step, where the model is packaged and registered.
   2. The process then waits for manual approval before proceeding further.
5. Model deployment:
   1. Once approved, the process triggers AWS EventBridge, which is connected to the continuous deployment (CD) pipeline.
   2. The CD pipeline deploys the model first to a staging environment using a Lambda function and AWS CloudFormation.
   3. After another manual approval, the model is deployed to the production environment.