What's next in Kubernetes Bob Killen Program Manager, Google Cloud

Bob Killen Program Manager, Google Cloud .

Building the Foundation

June Kubernetes announced and open sourced at Dockercon 2014 2015 2019 2016 2018 2020 2021 July Kubernetes 1.0 Released & donated to the CNCF August GKE launched as first commercial Kubernetes offering January Core Workloads go GA in 1.9 Release March Kubernetes is the first project to graduate to stable in the CNCF July Google Cloud Services Platform launched April Google Anthos Launched September Custom Resources go GA in the 1.16 Release August Kubernetes support window extended to 1 year in the 1.19 release

Why multi-cluster? Security Maximize security by restricting access at the cluster level and only advertise select services across clusters. Resiliency Ensure application and service availability by spreading the load across multiple clusters. Scalability Scale components or burst to the cloud to meet the demands of your application. Latency Deploy your apps globally to minimize latency and improving the user experience.

Clusters are the new Pod

Current state us-west us-east API Gateway API Gateway Business Logic Business Logic Cache Database Load Balancer Load Balancer Load Balancer Global Load Balancer Multi-Cluster systems are complex.

Multi-cluster services us-west Multi-Cluster Gateway us-east API Gateway API Gateway Business Logic Business Logic Cache Database New API that addresses the complexities of cross-cluster networking and service discovery.

Gateway API us-west Multi-Cluster Gateway us-east Service Next generation of Ingress management designed to support both multi-tenant and multi-cluster use cases. - matches: - path: type: Prefix value: /store - matches: - path: type: Prefix value: /store - matches: - path: type: Prefix value: /admin Web Server Service Web Server Web Server Service

One GKE: Two modes Standard ● Configuration flexibility ● No security restrictions ● Pay by the node Autopilot ● Managed node configuration ● Greater default security posture ● Pay by the pod GKE (us-east) GKE (us-west) Standard Cluster GKE Standard Cluster GKE Autopilot Cluster GKE Autopilot Cluster GKE Multi-Cluster Gateway Multi-Cluster Services

What’s next for AI/ML and batch

Why is it so hard? Scheduling Batch and AI/ML workloads have complex scheduling requirements that clash with the default Kubernetes scheduler and resource types. Scalability Batch and AI/ML workloads can have hundreds of thousands of tasks and require both large singular clusters and the capability to burst. Performance Requires a highly optimized underlying system, access to specialty hardware and a tuned backend to handle the high throughput requests.

Work queues data-0 data-1 data-2 data-N Job - dproc Pod dproc-5rwd7 GCS Bucket Pub/Sub External system required to keep track of work queue Pod dproc-z3e25 Pod dproc-q26l4 Pod dproc-l6d9a

Reduced complexity: Indexed Job data-0 data-1 data-2 data-N Job - dproc Pod dproc-0 GCS Bucket Pod dproc-1 Pod dproc-2 Pod dproc-N New Job completion mode: Indexed ● Built in method to partition work ● Consistent Pod Hostname ● Job Index exposed to Pod for easy integration

● Scheduler framework ● Suspended jobs ● API priority and fairness Other improvements for batch workloads

Kubernetes is Foundational Kubernetes is Extensible Kubernetes is Mature Kubernetes is... ✔ ✔ ✔

✔ ✔ ✔ Easy? Kubernetes is Foundational Kubernetes is Extensible Kubernetes is Mature Kubernetes is...

✔ ✔ ✔ Getting easier every day Kubernetes is Foundational Kubernetes is Extensible Kubernetes is Mature Kubernetes is...

● Kubernetes Essentials from Google Cloud ● Learn Kubernetes with Google ● Google Open Source Live ● Hybrid and multi-cloud: Anthos and Google Kubernetes Engine Learn more

Thank you