MLExray: Observability for Machine Learning on the Edge
Talkat Kubernetes on Edge Day Europe 2022 · on Tuesday, 2022-05-17· at 09:40· Pavilion 4, Room C | Level 2 | Central Forum·
summary
Describes MLExray’s end to end debug pipeline that logs per-layer inputs and outputs, end-to-end and per-layer latency, and resource and sensor metadata, then diffing those logs against a cloud reference pipeline to find where accuracy drops or latency spikes. Shows concrete failures: accuracy falling from ~80–90% to ~30% or >10% drop after edge deployment, and a case where non-max suppression slows because it receives ~2,000 bounding boxes. Demonstrates Pixie using eBPF uprobes to add telemetry without redeploys and plots p50 p90 p99 latencies.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:35Why Is this Important
- 1:04Amazon Echo
- 2:13Security Privacy Benefits
- 8:22Accuracy Validation
- 14:30Kubernetes Object Detection
- 22:10Resources
related talks
27mDebugging Machine Learning on the Edge with MLExray
10mML-based Log Analysis for Faster Debugging
17mLLM Observability, Evaluation, Experimentation Platform
26mML Monitoring & Observability
10mLightning Talks: Detecting Data Exfiltration on the Edge with Pixie
19mMinimize Risks and Accelerate MLOps with Model Performance Monitoring and Explainability
30mFrom Black Box To Insight: Observability for AI Agents in Production
31mRunning Machine Learning Analytics On Traces