Replacing a legacy relational model with a new approach on ScyllaDB
Medium keeps its millions of readers engaged by delivering personalized recommendations in real time. This personalization was originally implemented using an ML feature store and cross-entity relations. But the team realized that a different approach would be needed to support the constantly-increasing volume of requests with low latency.
Join Andréas Saudemont, Principal Engineer at Medium, as he shares how Medium re-architected its feature store to power real-time recommendations at scale. You will learn:
- Why Medium moved away from relational features and the limitations they faced
- How list features were designed for scalability and efficient access
- Data modeling strategies for high-throughput workloads
- Benchmark results comparing ScyllaDB and DynamoDB in production use
