Case Study: Discord achieves double-digit safety improvements with StreamNative

Revolutionizing Real-time Streaming Machine Learning at Discord

Discord, the real-time communication platform with more than 150 million monthly active users, needed a better way to power safety and personalization features with machine learning. Its existing rules-based system and Google Cloud Pub/Sub setup created problems with stateful feature engineering, batch and streaming data gaps, slow watermarks, and high operational overhead, making real-time ML difficult at scale.

StreamNative helped Discord move to a streaming ML platform built on Apache Pulsar, Apache Flink, and Apache Iceberg, with managed Pulsar through StreamNative and an actively maintained Pulsar-Flink connector. The new Kappa architecture enabled backfills in 3–4 hours, simpler infrastructure, and direct model inference in Flink, resulting in double-digit improvements in spam detection and account security, faster iteration, and lower maintenance for a very small engineering team.

Discord

David Christle

Staff Machine Learning Engineer