aiacademy: AI environment, clkao

• clkao

• What yo will learn today

Problems

• Setup machine learning enviroment is gard

• Reproducible researches require lots of tooling

• It may take lots of time(months) to apply AI/ML reserach result in pruduction

ModelOps

Paradigm: Interactivity & Agile

Jupyter Architecture

• jupyter-notebook

• message-jupyter

---

---

---

• 願望清單

  • CLKAO 的同事，每年玩一種語言

---

---

---

JupyterHub (on K8s) Architecture

Container & Orchestration

• Containers: Isolated process space, filesystem
• Orchestrator: Decide where to run things
  • Kubernetes: declare desired state
• Container Networking: connect (or disconnect) w/ other pod
• Persisten Storage: when container requires persistent fs

---

---

---

• docker

  • 從 linux 的 LSC 建置起來的

---

---

---

• Docker and Kubernetes Orchestration

  • link

kubernetes

• Pod

  • pod

• Nodes

  • node

• Ingress 最外層(入口)

• kubernetes-101-and-fun

  • reference

• hello-minikube

  • example

---

---

---

GPU

• openCL 公有
• NVIDIA 私有

---

---

---

aiacademy: Jupyterhub

• keycloak
  • link

• Spawner

• Juyterhub Routing

Components

• JupyterHub + JypyterLab

  • kubespawner
  • pre-project storage, gpu quota

• Keycloask: SSO

• Gitlab: courses material and data management

  • Image Bilding & Registry

• Custom DaemenSet

  • git-sync

What can possibly go wrong?

• What can cause cascade failures?

• SPoF!
  • storage: needs to be distributed & HA
    • ex: ceph
  • hub fafilures
  • 死不透的 issues
• hardware failure - memory, gpu card, power

---

---

---

• clkao 建議學習項目

  • PostgreSQL

  • Berkeley: data-8

---

---

---

The littlest JupyterHub

• TLJH (The Littlest JupyterHub) vs. Z2JH (from zero to JupyterHub on Kubernetes)

PrimeHub

• clkao 的公司

  • InfuseAi

• Communtiy Edition:

  • Github

KubeCon

• link

• KubeCon EU 2018 slides:
  • reading
• KubeCon NA 2018 slides:
  • reading

Futther readings

• Six easy ways to run your Jupyter Notebook in the cloud