Tune Your Microservices by Learning from Traces
summary
Pipeline ingests ~700,000 Istio traces per day, moves heavy ES load into a scalable store and trains models with Kubeflow on Kubernetes, saving versions to object storage. Time-series models use LSTM/seq2seq to predict service duration and flag anomalies; evaluation uses ~1,990-ish test points. Short-window and long-window models run in parallel to avoid mistaking normal periodic swings for faults. Deployment integrates versioned Kubeflow runs with alerting so detected anomalies notify ops, reduce false positives, and support root-cause correlation.
AI-written summary · about
chapters (5)
- 0:00Start
- 1:34Architecture Overview
- 5:20Training & Modeling
- 15:09Model Serving
- 17:01Tuning microservices based on result
related talks
29mTune Your Microservices by Learning from Traces
35mAdvanced Model Inferencing Leveraging KNative, Istio & Kubeflow Serving
14mPerformance Optimization in High-Volume Transaction Systems
30mDebugging at Scale Using Elastic and Machine Learning
20mML Ops design patterns with Kubeflow Pipelines
32mScaling Kubeflow for Multi-tenancy at Spotify
29mML Engineering Kubeflow
39mDeploy your ML Workloads on Kubernetes with Kubeflow