The Synthetic Data Flywheel
summary
Pinterest builds a Ray Data pipeline that turns seed queries into millions of synthetic preference pairs by streaming 256 upfront partitions through GPU generation, concurrent QPS‑limited network reactors and GPU scoring, using an operator library so teams swap models, prompts and pool sizes. Optimizations—continuous batching, dynamic GPU reallocation between stages, and one-time repartitioning—cut one job from 334 hours to 2h41m (a 227× speedup) and feeds an iterative DPO loop that retrains and reuses the model.
AI-written summary · about
chapters (16)
- 0:00Introduction to talk
- 1:11Pinterest ML stack and product
- 3:58Need for synthetic data
- 6:28Data generation bottleneck
- 7:31Modern GPU-heavy pipelines
- 9:39System challenges overview
- 10:21Eric introduces system problems
- 12:27Map batches and streaming
- 13:53Dynamic GPU reallocation
- 14:59Single upfront repartitioning
- 16:03Operator framework design
- 17:28Two Pinterest use cases
- 19:14Performance improvements
- 20:17Iterative DPO pipeline
- 21:45Acknowledgements and thanks
- 22:08Q&A session
related talks
27mPinterest's ML Evolution: Distributed Training with Ray
32mAnyscale's Ray Data: Revolutionizing Batch Inference
30m🤯 No data? No problem! Synthetic data to the rescue
18mUnleash your Private Datasets to Power up AI
35mHow Bytedance Builds Large-Scale Data Processing Pipelines for Multimodal Models with Ray
38mNetflix's Ray Platform: From Deep Learning to GenAI
55mGPU-accelerated virtual drug screening
33mNext-generation tools for synthetic data and AI training