From SpellML to SageMaker: A Successful MLOps Journey
Discover How Whatnot Boosted Their Data Science with SageMaker
🏢 ABOUT THE COMPANY
Whatnot, a social marketplace where you can discover some of your favorite products, such as sports cards, sneakers, luxury handbags, and women's thrift, sought to enhance its MLOps capabilities to optimize its Recommendation Engine. The existing infrastructure, built on SpellML, was reaching its limitations in terms of scalability and flexibility.
⛰ CHALLENGES
The company faced the following challenges:
Scalability: The existing infrastructure struggled to handle increasing data volumes and model complexity.
Flexibility: The rigid nature of the SpellML framework hindered rapid experimentation and iteration.
MLOps Maturity: The company aimed to establish a robust MLOps practice to streamline the entire ML lifecycle.
✅ SOLUTION
To address these challenges, we embarked on a migration project to Amazon SageMaker, a fully managed platform for machine learning. The key steps involved:
Infrastructure Design: Designed a scalable and resilient infrastructure using Amazon SageMaker's managed services, including SageMaker Studio, SageMaker Pipelines, and SageMaker Endpoint.
Utilized Terraform to automate the provisioning of infrastructure resources.
Training Pipeline Development: Created efficient training pipelines using SageMaker Pipelines, incorporating data ingestion, feature engineering, model training, and model evaluation.
Leveraged Python and popular ML frameworks (e.g., TensorFlow, PyTorch) to build and train the Recommendation Engine models.
Model Deployment: Deployed trained models as real-time endpoints using SageMaker Endpoint.
Implemented load testing to ensure the endpoints could handle production traffic.
MLOps Integration: Integrated the SageMaker infrastructure with the company's existing MLOps tools and processes.
Implemented CI/CD pipelines using GitHub Actions to automate the entire ML workflow, from code commit to model deployment.
🏆 OUTCOMES
By migrating to Amazon SageMaker, Whatnot has not only enhanced its MLOps capabilities but also empowered its data science team, recognizing its pivotal role in delivering innovative solutions and driving business growth.
The successful migration to Amazon SageMaker has yielded many benefits, reassuring the team and stakeholders about the project's positive outcomes.
Enhanced Scalability: SageMaker's cloud-native architecture enabled seamless infrastructure scaling to accommodate growing data volumes and model complexity.
Accelerated Development: The streamlined workflows and pre-built components of SageMaker significantly reduced development time and increased iteration speed.
Improved MLOps Maturity: The adoption of SageMaker fostered a robust MLOps practice, enabling the team to focus on model development and experimentation.
Cost Optimization: Leveraging SageMaker's pay-per-use pricing model and spot instances helped optimize costs.