A Day in the Life of a Data Scientist Conquer Machine Learning Lifecycle on Kubernetes Brian Redmond • Cloud Architect @ Microsoft (18 years) • Azure Global Black Belt Team • Live in Pittsburgh, PA idle • Parallel training instead of sequential: huge time saver for large trainings Kubeflow • Machine Learning Toolkit for Kubernetes • To make ML workflows on Kubernetes simple, portable, and scalable

protect secrets? ● Attractive target ○ Controls access or use of sensitive resources ● Common attack vector ○ Checked into Github ○ Accessible by users who shouldn’t have access, e.g., CEO ○ Stored in public

Kubernetes repo support • Agile, flexible and convenient Node-Operator: Overview • User: SREs who can scale & offline Nodes through posting Machine CRs. • Node-Operator: difference Machine and Node state state, manage Node softwares and configure files. • Machine: the instance of Machine CRD with node basic information, which represent a node desired in the Kubernetes. • NPD(Node Problem Detector):

+ kata not easy kubernetes(dockershim) does not support to choose OCI runtime k8s + docker + kata not easy kata container network hotplug (support now) Dockershim / Docker Containerd Cri-o a.create implements Linux by way of Linux ü another approach to enhance container isolation gVisor is special Machine-level virtualization Rule-based execution gVisor Technology landscape DEMO

Processing Pipelines Data Flow Graphs for Machine Intelligence Kubernetes Contributors opensource.google.com A strong community with corporate and independent support. Independent Google Red Hat Huawei for Alpha ● Simple CLI installation ● Online and Offline installation ● Private container registry support ● Latest 3 versions of k8s ● High-availability control plane ● Auto-repair Installation and Configuration

Pod container image "gcr.io/google_containers/pause:0.8.0" alrea 20-管理容器的计算资源 85 dy present on machine Tue, 07 Jul 2015 12:53:51 -0700 Tue, 07 Jul 2015 12:53:51 -0700 1 {kubelet kubernetes-node-tf0f} created by the secret volume mount will have permission 0400 . Note that the JSON spec doesn’t support octal notation, so use the value 256 for 0400 permissions. If you use yaml instead of json for the proxy subresource查询 Heapster。 需在集群上部署Heapster并在kube-system这个namespace中运⾏。 有关使⽤REST客户端访问的详细信息，请参阅 Support for custom metrics 。 autoscaler通过 scale sub-resource访问相应的Replication Controller、Deployment或ReplicaSet。Scale是⼀个允许您

instructions ! I have only a local machine You need an easy way ! Local-machine Solutions local-up-cluster kubeadm Minikube microk8s kubeadm-dind-cluster Local-machine Solutions https://github.com/kubernetes/minikube https://microk8s.io/ local-up-cluster Minikube Kubeadm kubeadm-dind-cluster microk8s My machine can't run Kubernetes... Oops ! Online playgrounds ! https://labs.play-with-k8s.com/ Online playgrounds

sudo chown -R $(whoami) /usr/local/lib/pkgconfig ## VM 驱动程序 VirtualBox brew install docker-machine-driver-hyperkit ~/.kube/config ⽂件中查看所有可⽤的环境 启动 # 可添加参数 # --vm-driver=hyperkit minikube start

Cloud Native Way 阿里云容器服务团队 张维 KUBECON CHINA 2018 应用部署演化： Going native with cloud Virtual Machine/Bare Metal Increasing focus on business logic Decreasing concern (and control) on infrastructure

Platform (ACP)，包括：DevOps、Service Framework、Service Alauda Kubernetes 白皮书 灵雀云 5 Mesh，Machine Learning 等，构成完整的产品闭环，可以根据企业定制化的使用场景，选 择使用不同的产品。 5. 客户价值 使用 Alauda Kubernetes，有助于：