MoCE: A Mixture-of-Context Aware Experts Framework for Troubleshooting...
summary
MoCE implements a mixture-of-context-aware experts system where each expert is expressed as a dataflow DAG of SQL, stateful, and statistical operators; a library of ~30 operators and ~15 experts runs on a Python prototype with a lazy DAG runtime (built on Ray) to merge DAGs and avoid duplicate work. On injected and real incidents across four services MoCE reaches >95% RCA accuracy (versus ~34% for the best single expert), shows 6x speedup at 15 experts, and yields 9/10 verified findings in a week-long pilot.
AI-written summary · about
chapters (15)
- 0:00MoCE framework intro
- 1:01Internet failure costs
- 2:49RCA challenges in large systems
- 4:34Systems approach to RCA
- 5:16MoCE core ideas
- 6:39Context-aware experts
- 9:06Building experts with libraries
- 11:58Evolving operator library
- 12:18Lazy DAG runtime
- 13:00MoCE end-to-end architecture
- 14:24Implementation status
- 14:46Evaluation on failures
- 15:49Pilot study results
- 16:34Scalability comparison
- 17:16Conclusion and future
related talks
53mFrom Mixture of Experts to Mixture of Agents with Super Fast Inference
12mSmartMoE: Efficiently Training Sparsely-Activated Models through Combiningand...
24mBuilding Queryable O11y with the Model Context Protocol
27mStageML: Partial Evaluation for Multi-Tenant MoE-LoRA Inference
37mUnstructured Sparsity Meets Tensor Cores: Lessons from Sparse Attention and MoE
46mFlexOlmo: Open Language Models for Flexible Data Use
20mAccelerating Distributed MoE Training and Inference
31mVeRL: Extreme Optimization Practices for Ultra-Large-Scale MoE Models