Automated container-deployment-on-kubernetes
Download as PPTX, PDF1 like273 views
The document provides an introduction to Kubernetes, an open-source system that automates the deployment, scaling, and management of containerized applications. It covers Kubernetes architecture, cluster management, application deployment, resource request handling, and advantages of cloud versus bare metal resources. Key topics include health checks, auto-scaling, load balancing, and job scheduling in a Kubernetes environment.
Automated container-deployment-on-kubernetes
- 1. Automated Container Deployment on Kubernetes An intro to kubernetes
- 3. Outline • What is Kubernetes • Deploy a containerized app • Deploy a app to Kubernetes. Pros & Cons • Real world cases
- 4. What is Kubernetes Kubernetes is an open-source system for automating deployment, scaling, and management of containerized applications https://kubernetes.io/
- 5. Why containerize Docker - Build, Ship, and Run Any App, Anywhere Docker is an open platform for developers and sysadmins to build, ship, and run distributed applications, whether on laptops, data center VMs, or the cloud. https://www.docker.com/
- 6. Kubernetes Cluster Let’s use Kubernetes Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Pod 1 (App) We want’s to deploy our app
- 7. Kubernetes Cluster Deploy an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Pod 1 (App) We want our app. We don’t really care where it is.
- 8. Kubernetes Cluster Access an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Service (App Endpoint) Pod 1 (App) Cluster-wide endpoint by service
- 9. Kubernetes Cluster Scale an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Service (App Endpoint) Pod 1 (App) Pod 2 (App) K8s find a Node to run second pod
- 10. Kubernetes Cluster Health check an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Service (App Endpoint) Pod 1 (App) Pod 2 (App) Your service still online with zero downtime
- 11. Kubernetes Cluster Self-heal an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Service (App Endpoint) Pod (App) Pod 2 (App) Your service still online with zero downtime Pod 3 (App)
- 12. Kubernetes Cluster Self-heal an app Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Service (App Endpoint) Pod 4 (App) Pod 2 (App) Your service still online with zero downtime
- 13. Keywords • Clustering, distributed • Automation deployment • Auto scaling, Load balancing • Health check, self-healing, zero downtime
- 14. Kubernetes Cluster Resource Management Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 GPU GPU Storage Job 1 (GPU) Job 2 (GPU) One time job. Deploy, execute, save results. GPU Storage
- 15. Kubernetes Cluster Request Resource Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 GPU Storage Job 1 (GPU) Job 2 (GPU) Share loading between nodes GPU GPU Storage
- 16. Kubernetes Cluster Request Resource Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 GPU Storage Job 1 (GPU) Job 2 (GPU, Heavy disk io) Share loading between nodes GPU GPUStorage
- 17. Kubernetes Cluster Schedule Policy Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Storage Job 1 (GPU) Job 2 (GPU) GPU GPU Storage GPU Node 1 is optimized (high CPU, high disk IO…)
- 18. Kubernetes Cluster Queue Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 GPU Storage Job 1 (GPU) Job 2 (GPU) GPU GPU Storage Job 3 (GPU) Job 4 (GPU) K8s try to meet resource request
- 19. Kubernetes Cluster Queue Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Storage Job 1 (GPU) Job 2 (GPU) GPU GPU Storage Job 3 (GPU) Job 4 (GPU) K8s try to meet resource request GPU
- 20. Kubernetes Cluster Resources Node 2 10.1.15.2 Node 1 10.1.15.1 Node 3 10.1.15.3 Job 2 (GPU) GPU Storage Job 6 (GPU) Job 5 GPU CPU Memory CPUCPU Storage GPU
- 21. Keywords • Resource request, availability check, auto scheduling • Schedule policy, affinity, anti-affinity • Job queue • Label selector
- 22. Data Processing Job • Tons of data, heavy IO • GPU/CPU loading • Storage elasticity
- 23. Kubernetes Cluster Storage System Storage System GPU Computing GPU Computing CPU serverCPU server GPU Servers CPU servers Storage System GPUGPUCPUCPUCPU GPU Network Storage System Endpoint Storage Data Processing Job Pre-processing Model-training Inference
- 24. Public Cloud • Google GKE, AWS EKS, Azure AKS… • Auto scale servers (node pool) • High integration with cloud services • It won’t easily die • Expensive GPU
- 25. Bare metal servers • Cheap (compare to cloud platform) • Highly customizable, optimizable • Embrace many system / networking / infrastructure issues
- 26. Thank you • Automating deployment, scaling, containerize • Build, deploy, distributed apps, anywhere • Resource management, job queue, scheduling policy
Editor's Notes
- #4: Share great tools for DevOps, deployment Sounds like a sale I saw tremendous change within my develop team. Hard to maintain but really good when using it There are less ‘how’ in this presentation
- #5: If your issue fit those key words, try k8s 2014 released by google, open source, maintained by community, lead by google Easy to achieve automation deployment management Containerize
- #6: Docker released 2013 Containerize: build, ship, run any app, anywhere Widely used by public cloud/services like GCP Uni-interface upon OS -> App like a tree root on your server. Your engineer don’t want to move it.(bind port, bind dir, bind library, bind kernel, bind os distribution) App as replaceable part Easy to build, deploy, scale, monitor, heal, recover
- #7: Learning by using Distributed system Let’s say we already have a k8s cluster for some reason We want’s to deploy our app / training job / api server / db…
- #8: User make deploy of an app We want our app. We don’t really care where it is. Kubernetes handle schedule for you.
- #9: K8s find a node to deploy Can be any suitable node. Then, how to access it? K8s build a cluster-wide endpoint. K8s networking dns
- #10: User scale up app deployment K8s find a Node to run second pod Pod/container is easy to ship, deploy, scale
- #11: K8s found out node 3 down for some reason High Availability Auto health check, error discover Your service still online with zero downtime
- #12: Your service still online with zero downtime K8s find another suitable node for desired app number: which is 2
- #13: When node became available again (issue fix) K8s automatically bring pods back Your service back to normal
- #14: Kubernetes is a distributed sys
- #15: Let’s talk aboutOne time job. Deploy, execute, save results.
- #16: Node 3 won’t accept GPU Job
- #17: The request is very dynamic
- #18: Node 1 is optimized to run GPU Jobs (high CPU, high disk IO, already have data) Node affinity Node selector
- #19: K8s try to meet resource request
- #20: K8s try to meet resource request Resource release when job complete Your data scientists should focus on submitting jobs, not waiting
- #21: Many pre-defined resource Utilize resource for minimum costing Label your node for more specific need like bandwidth, IO, cost
- #23: It’s all about data Storage is another very long story Access big data chunk from everywhere. You don’t move thousands of GB Backups? Replicas? Model cumulative feedback (reinforcement learning)
- #24: Scale your cluster (data center) Containerize your training
- #25: Google’s kubernetes on GCP Registry, network setting, load balancer Global access
- #26: Cheap GPU Storage system