AWARE: Automate Workload Autoscaling with Reinforcement Learning in Production...
summary
AWARE is an extensible framework that deploys RL autoscalers on Kubernetes with an OpenAI Gym-like interface, combines offline replay-buffer training with online actuation via bootstrapping, and conditions policies on per-workload embeddings learned by meta-learning. It uses a DDPG actor‑critic agent, adapts to new workloads with 5.5× faster convergence (using ~32 trajectories), keeps online reward degradation under 3.6%, raises CPU/memory utilization ~47%/39%, and cuts SLO violations up to 16.9×.
AI-written summary · about
chapters (9)
- 0:00Introduction to RL in Cloud
- 2:40RL Framework for Systems
- 4:42Challenges in Production
- 7:50The Aware Framework
- 9:42MPA Implementation Details
- 12:06Bootstrapping and Monitoring
- 13:34Meta-Learning for Adaptation
- 17:35Experimental Evaluation
- 20:20Conclusion and Future Work
related talks
17mScaling Reinforcement Learning
25mIntelligent Kubernetes Workload Optimization
19mDynaRL: Flexible and Dynamic Scheduling of Large-Scale Reinforcement Learning Training
30mPut the ‘Auto’ in Autoscaling — Erwin Daria
17mReinforcement Learning with Human Feedback in Distributed Systems
31mAccelerate AI adoption: automated cloud efficiency
28mScaling AI Infrastructure
23mAlphaDow: Reinforcement Learning for Industrial Production Scheduling