Vulcan: Automatic Query Planning for Live ML Analytics
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Vulcan automatically generates full query plans for live ML analytics by building pipelines, deciding filter order, and jointly optimizing operator placement and configuration across heterogeneous edge tiers. It profiles each configuration once then reuses intermediate output sizes and timings to compute alternative placements with Bayesian optimization per-placement and early pruning, seeding online re-profiling with top-k/worst-k configs. The system claims 4.1×–30.1× lower search cost and up to 3.3× better query latency compared to prior ML analytics systems.
AI-written summary · about
chapters (11)
- 0:00Introduction to Vulcan and live ML analytics
- 1:00Live ML query challenges
- 3:08Pipeline construction
- 5:16Design goals
- 7:47Generating query plans
- 9:14Evaluating plans
- 10:17Bayesian optimization
- 11:21Low-cost placement
- 13:29Online adaptation
- 14:55Evaluation results
- 15:59Conclusion
related talks
27mManufacturing optimisation & planning using ML platform
41mHydra's Heads – Vladimir Ozerov
10mML-based Log Analysis for Faster Debugging
39mMachine Learning with Scala on Spark
27mLeverage ML to automatically tune K8s apps
39mAccelerating Open-Source RL and Agentic Inference with vLLM
14mPerformance Optimization in High-Volume Transaction Systems
46mShorten the Path to Production with In-Database Machine Learning