How Uber Optimize Marketplaces with Ray
summary
Claim: Ray delivers up to 47.8x speedup across Uber’s marketplace workflows by parallelizing per-city per-week ADMM subproblems. Implementation: a hybrid Spark+Ray pipeline that preps data in Spark, launches Ray tasks for surrogate model training and optimization, writes/reads Parquet via HDFS to avoid large driver transfers, and uses parent/child Ray clusters plus short image-less deploys to cut launch time to under two minutes. Results: surrogate modeling, evaluation, and allocation optimization run dramatically faster and scale.
AI-written summary · about
chapters (13)
- 0:00Introduction to talk
- 0:24Marketplace optimization problem
- 5:23Adopting Ray at Uber
- 9:39Spark-heavy application before Ray
- 13:13Hybrid Spark-Ray architecture
- 16:09Fast deployment with cloud storage
- 18:18Data transfer bottlenecks
- 20:55Ray cluster management tools
- 22:20Multi-cluster Ray control
- 23:46Hybrid mode notebooks
- 24:30Benchmark results and Ray benefits
- 25:58ADMM optimization with Ray
- 28:54Q&A session
related talks
26mUsing Ray On Large-scale Applications at Ant Group
28mDistributed XGBoost on Ray
17mScaling and Unifying SciKit Learn and Spark Pipelines using Ray
26mData Processing on Ray
03mRay: A Distributed Execution Framework for Emerging AI Applications
44mPanel: ML Platform on Ray
26mRayDP: Build Large-scale End-to-end Data Analytics and AI Pipelines Using Spark and Ray
33mScaling PyTorch Training with Ray Data on Uber's Michelangelo AI Platform