Dynamically Tuning Pods: Leveraging Time Series ML Models with KubeFlow
summary
Demonstrates using LSTM on Prometheus metrics via Kubeflow pipelines to predict pod CPU/memory trends and suggest request values, with statistical baselines using z‑scores and a two‑sigma rule for safe requests. Demo uses PyTorch notebooks, two weeks of data (test prediction runs in seconds; six months runs ~1–2 minutes), forecasts a day or two ahead, misses seasonal spikes, and recommends experimental pipelines (Argo/Tecton/OpenDataHub) rather than automatically overwriting production requests.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:43Christophers background
- 1:42Problem Statement
- 2:37Cloud Computing Myth
- 4:52Example Application
- 5:50zCore Example
- 6:47Limits
- 9:16Statistical Method
- 9:46Long Short Term Memory
- 11:58The Process
- 13:05Challenges
- 15:20Jupiter Notebook
- 18:35Results
- 22:17Questions
related talks
51mMachine Learning using Kubeflow and Kubernetes
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
31mIntroduction to Using and Use Cases of KubeFlow
14mKubeflow: Where ML Meets Kubernetes
54mScaling Machine Learning Pipelines in Cloud
31mKubernetes for Machine Learning: Managing and Scaling AI Workloads
39mDeploy your ML Workloads on Kubernetes with Kubeflow
46mKubeflow, TensorFlow on Kubernetes