Data science at scale using Apache Flink: Dynamic model serving and real-time feature generation
summary
Describes a production platform called Mitra built on Apache Flink + Kafka + RocksDB that serves 500+ e‑commerce companies and performs real‑time model scoring and feature generation. Claims predict results within 200 milliseconds and generate hundreds of features on the fly during serving, implements a dynamic rule engine, Graph DB and ML model server, and uses Flink CEP, broadcast state and Async IO across ~60 operators and 40+ in‑memory states to achieve low‑latency, distributed model serving.
AI-written summary · about
chapters (9)
- 0:00Introduction to speakers
- 1:14Problem with real-time data
- 2:33Industries needing dynamic models
- 5:20Active body systems in Flink
- 7:35Features of Flink platform
- 10:18Manual mode state for evening
- 11:57Joining the screen example
- 13:12Real-time model serving explanation
- 14:17Conclusion and Q&A
related talks
21mReal-time Stream Analytics and Scoring Using Apache Flink, Druid & Cassandra
28mEnabling Machine Learning with Apache Flink
25mReal-Time Fraud Detection at Scale
20mBuilding FeatureFlow, Criteo’s feature data generation platform
29mBuilding Resilient AI Platforms: Financial Fraud Detection
25mData: Moving with the flow
21mIntroducing FlinkML
24mUnder Pressure: Applying ML in Real Time