TechPrufer
Get started
ScyllaDB
ScyllaDB logoScyllaDB · Feature trail

Medium Feature Store Scalability

1 post · Aug 2026

What this trail covers

Medium has evolved its feature store architecture to handle massive scale, moving from a relational feature model to a list-based model using ScyllaDB. This change was driven by the need to support 1M operations per second for real-time recommendations, addressing performance bottlenecks and complexity associated with relational features at scale. The new architecture leverages ScyllaDB's capabilities to store and retrieve features efficiently, enabling more personalized and timely user experiences.

All ScyllaDB trails
  1. How Medium Powers Real-Time Recommendations at 1M OPSThis post details Medium's transition from a relational feature model to a list-based feature store architecture, powered by ScyllaDB, to achieve 1M operations per second for real-time recommendations. It explains the limitations of the previous relational model, particularly for features with multiple values and cross-entity relationships, which led to performance degradation and complexity at scale. The new architecture utilizes ScyllaDB to store and retrieve features, focusing on efficient data modeling and retrieval for high-throughput, low-latency recommendation systems. The post highlights the importance of data modeling choices in database systems for scaling critical features.