AutoSys: The Design and Operation of Learning-Augmented Systems
summary
AutoSys is a unified framework that runs Hyperband, TPE, SMAC tuners and trial benchmarks to auto-tune five Web Search engines (selection, ranking, KV stores) for latency, CPU and relevance. Centralized training generates candidate configs, a trial manager prioritizes noisy, non‑deterministic jobs, and a rule engine blacklist validates actuations to avoid learning‑induced failures. Production use yields better or comparable performance in days to weeks versus decades of manual tuning, but training costs, transfer gaps, and human‑injected bias remain practical challenges; core components are on GitHub.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:25Learning-Augmented Systems
- 5:40Towards A Unified Framework - AutoSys
- 7:39Jobs Within AutoSys
- 11:03Overview - Learning
- 14:13Overview - Auto-Tuning Actuations
- 16:29Summary of Production Deployments
- 17:37Long-term Lessons Learned
- 22:09Conclusion
related talks
29mContinuous tuning of Microservices via Bayesian Optimization
11mSimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large...
17mSelfTune: Tuning Cluster Managers
21mRise of the Machines: Removing the Human-in-the-Loop
46mLinux Kernel Autotuning
15mAuto-Adaptive Systems: Changing the Game for Marketplace Management
27mAutomating XGBoost Hyperparameter Tuning with Agentic AI
41mAutomated Performance Tuning with Bayesian Optimization