Empowering Real-Time ML Inference and Training with GRIS
summary
GRIS delivers a real system: a feature store and gRPC inference API built to support real-time inference and offline training with Iceberg historical datasets, using Cassandra behind a Netflix TS layer and Flink stream processing. The MVP ships in one year by a team of two and exposes hybrid aggregation: tumbling-window partials plus gRPC last‑mile sums. Operational lessons include amplification up to ~4,000 partials for 30‑day/1‑min freshness, occasional gRPC latency >100ms, and production tooling: streaming jobs, on‑call alerts, and online/offline data quality checks. Cassandra + TS + Flink + gRPC
AI-written summary · about
chapters (16)
- 0:00GRIS introduction
- 4:40Motivation and challenges
- 7:56ML lifecycle with GRIS
- 11:07Lambda architecture
- 14:39Real-time storage requirements
- 16:51Time-window aggregations
- 27:39Feature management DSL
- 31:08Data quality in GRIS
- 33:54Online vs offline features
- 34:35Key takeaways and future
- 36:21Q&A: GRIS trade-offs
- 38:10Q&A: Netflix real-time ML
- 39:17Q&A: Flink pain points
- 40:20Q&A: Supervised learning features
- 41:22Q&A: Time-series adoption
- 42:26Q&A: GRIS next steps
related talks
19mAI Inference Pipelines – Building Low-Latency Systems With gRPC
16mDeep Stream Dynamic Graph Analytics with Grapharis
15mData science at scale using Apache Flink: Dynamic model serving and real-time feature generation
24mUnder Pressure: Applying ML in Real Time
33mReal-time Trade Data Processing and Machine Learning at Giga-Sca...
37mBreakdown New Data Silos in Generative AI
04mInterview
38mReal Time ML Pipelines in Multi-Tenant Environments