Scaling ML workflows for real-time moderation challenges
summary
Twitch describes a real-time ML stack built to catch channel-level ban evasion and other chat abuse at ~31 million daily viewers, with models served via SageMaker and an RPC layer called turb; streaming features use Kinesis, Lambda and DynamoDB while offline work uses S3, Redshift Spectrum and Spark. They build a custom batch backend called blocks (180M records in ~40 minutes) to avoid ~$50k/month DynamoDB costs for 10M records every 3 hours, plus an inference gateway and Airflow-based conductor for CI/CD and feature logging, and report that these systems cut latency and cost while enabling fast reaction to evasion.
AI-written summary · about
chapters (9)
- 0:00Proactive Detection team intro
- 2:55Twitch real-time moderation challenges
- 8:11Scaling ML for chat moderation
- 9:15ML infrastructure for stability
- 11:03Handling growth and attacks
- 13:33Real-time reaction and reliability
- 15:21ML Ops for model maintenance
- 16:2424/7 safety systems availability
- 17:28Closing and careers at Twitch
related talks
50mThe tech behind modern AI/ML use cases in Gaming
24mEngaging new and diverse audiences with ML
23mML Platform Challenges & Applications at Freshworks
38mReal Time ML Pipelines in Multi-Tenant Environments
32mLiveMLP: ML Platform for Assisting Contact Center Agents in Real-Time
26mManaging ML Models @ Scale
29mData strategies to enable ML and analytics at scale
24mUnder Pressure: Applying ML in Real Time