When your data model is the bottleneck: lessons from Medium’s feature store

▤ Summary

Medium reworked its feature store data model after the original relational approach became a scaling bottleneck. The team moved to a list-based model so it could fetch user history and recommendation signals in a single query with better performance. Medium also used ScyllaDB TTL, clustering keys, and local secondary indexes to handle high-volume read and write patterns efficiently. Benchmarking showed more predictable low-latency behavior than DynamoDB on the workloads discussed. The company now plans to extend ScyllaDB to more feature store use cases because the new model removed a major architectural constraint.

▦ Linked Companies

ScyllaDB

$1M to $5M