How to Improve Your Kubernetes Experience with Service Mesh and MLOps
Talkat KubeCon + CloudNativeCon North America 2021 · on Wednesday, 2021-10-13· at 11:00· Concourse Hall 152 + Online·
summary
Private cloud runs 50+ on‑prem clusters (500+ compute nodes, 10+ Istio meshes) and collects Istio service‑mesh telemetry into Prometheus/Timeseries storage to build directed call graphs weighted by latency. Service grouping by heavy edges collocates chatty services via nodeSelectors/pod‑affinity/topologySpread to cut network hops; anomaly detection uses DBSCAN clustering and personalized PageRank to point to root causes (example: reviews v2 → container memory working set → heap overflow). Predictive autoscaling blends CatBoost + moving average on time‑series features to provision replicas proactively; models run in recommendation mode with plans for realtime rollout.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 2:25Monitoring Issues
- 4:51HighLevel Concepts
- 7:54Conceptual Approach
- 8:59Service Mesh Example
- 10:33Simple Data Array
- 12:38Grouping Mechanism
- 14:31Normally Detection
- 15:00Booking for Application
- 15:19Anomaly Detection
- 17:24Directed Application Graph
- 17:54Anomaly Subgraph
- 18:41Personal Pagerank Algorithm
- 21:19Predictive Autoscaling
- 22:13HPA and Predictive Autoscaling
- 25:39Conclusion
- 26:11Future plans
- 28:30Thank you
related talks
24mDemo of Kubernetes at the Edge for AI Model Deployments
26mRunning Machine Learning Workloads on a Service Mesh
39m3+1 Ways to Do MLOps with Kubernetes
03mSponsored Keynote: Redefining MLOps at Scale on Top of Kubernetes
36mMLOps from the edge to on demand visualization
38mBeyond Kubernetes: Adapting To Spe...
33mFrom spreadsheet scheduling to Kubernetes: building an on-premise ML platform
07mScaling Predictive Manufacturing Analytics: Kubernetes-Powered CRM