Panel: Cracking the Data Locality Puzzle
Panelat Data on Kubernetes Day NA 2025 · on Monday, 2025-11-10· at 16:00· Building B | Level 2 | B211-212·
summary
Panel summarizes concrete experiments and gaps: Kubernetes schedulers lack topology and bandwidth visibility so moving hundreds of gigabyte to terabyte model shards is slow; teams build tools like RunAI model streamer to stream weights directly to GPUs and a Kai-based topology-aware scheduling prototype that improves model-parallel throughput. CNCF Batch Subproject tests policies—splitting jobs, adding location metadata, scheduling on historical latency and price—and non-Kubernetes systems use worker-pull and multi-dimensional knapsack math to optimize placement; checkpointing, write consistency, and cross-region bandwidth remain unsolved in practice.
AI-written summary · about
chapters (10)
- 0:00Data locality challenges
- 3:57Panelists' definitions
- 8:33AI workload gaps
- 12:59Kubernetes scheduling limits
- 18:35Improving data locality initiatives
- 27:27Modeling and community efforts
- 29:35Kubernetes topology awareness?
- 31:04Q&A: writes, reads, GDPR
- 37:30Future networks and directions
- 39:14Closing remarks
related talks
53mPanel Discussion: Fragmentation of the Batch Ecosystem in Kubernetes, Challenges and Solutions
34mPanel Discussion: Training and Inference at Planet Scale
27mKeynote Panel: Future of Cluster Management
12mHow the Data Center Shapes the Workload
21mBeyond the Theory: What Actually Breaks When You Scale Your Disaggregat...
31mAn Adventure in Multi Cloud with Kubernetes and Distributed Data
43mSystems @Scale Summer Q&A Week 2
35mBeyond Centralized Models