Project Lightning Talk: Multi-Cluster Orchestration System: Karmada Updates A... Joe Nathan Abellard

In this lightning talk, Joe, a senior software engineer at Bloomberg and core owner of the Karmada project, provides an overview of Karmada, a CNCF project designed for Kubernetes-native multi-cluster management. As organizations increasingly scale their AI and analytics workloads, Karmada facilitates seamless orchestration across various cloud environments and on-premises setups. The talk details the architecture of Karmada, which resembles a typical Kubernetes cluster setup, comprising a control plane linked to member clusters. Key features highlighted include advanced multi-cluster management capabilities, policy-driven scheduling, and cross-cluster application failover. Joe also emphasizes the project's growth since its open-source launch in 2021 and its journey towards becoming a CNCF graduate project. Additionally, he teases the upcoming Karmada presentations that delve deeper into its utilization and future developments.

Welcome everyone. I'm Joe. I work at Bloomberg as a senior software engineer and I'm also one of the core owners of the Karmada project. Um and in this lightning talk, I'll provide you with a high-level overview of the project. But before I begin, um who here has heard or currently is using Karmada? Oh, nice. Okay, so let's get started. Um so as organizations scale AI, general compute

um and analytics workflow workloads across um clusters, regions, and clouds, uh multi-cluster orchestration has become a critical building block of the modern um platform engineering stack. And this is exactly where Karmada, short for Kubernetes Armada, comes into the picture. It's a CNCF project for um Kubernetes native multi-cluster management that enables you to run your workloads across your entire fleet of clusters. Um architecturally, a Karmada setup looks

very similar to that of a Kubernetes cluster setup. It's basically a control plane that is joined to a set of um member clusters. And Karmada is cloud-agnostic, so those clusters could either be logging in your private cloud, on prem, or across multiple um cloud providers. In the control plane, the central component um here labeled as Karmada API server, is just the standard kube API server, which enables

seamless integration with the existing Kubernetes toolchain. And instead of kube scheduler, you have Karmada scheduler, responsible for making intelligent placement decisions across your entire fleet um of clusters. Um some features that are of note are obviously very powerful multi-cluster management, um very advanced um policies to meet different type of scheduling requirements and scenarios, um cross-cluster application failover um to handle failover of applications in disaster scenarios, um

unified authentication, authorization, and auditing, um global resource view, which provides a unified entry point to query resources across your entire fleet of clusters, as well as multi-cluster service discovery. Um the project was published as open source in 2021. We became a CNCF sandbox project that same year. Um in 2023, we became a CNCF incubating Um and last year we applied to become a CNCF graduate project, and

we're currently on track to be a graduate project this year. Um we have a very fast-growing and vibrant community with sustained contributions from contributors across many different companies, and the project itself has been widely adopted. Um on here I have a list of public adopters in production, and some of the companies that you may recognize on here are Huawei, Bloomberg, where I work, as well as trip.com.

Um we're actually just getting started with the Karmada party at KubeCon, so we're going to have more Karmada for you. There's going to be a talk on how to decompose and govern giant LM jobs across multiple clusters managed by Karmada. And we'll also have the Karmada maintainer track talk, which will talk about the project in a lot more detail. And when you think we're done, we're going

to have even more Karmada. So, tomorrow I'll be doing a talk on how we executed a zero-downtime migration from the now-retiring ingress-nginx project to Istio in a multi-cluster Kubernetes platform built atop Karmada um at Bloomberg. Um and then we're going to have a talk by one of the tenants of that platform um for running disaster-resilient Trino on multi-cluster Kubernetes powered by Karmada um and Trino gateway. in

closing, whether you are building distributed AI infrastructure, um designing public cloud platforms, or just exploring how multi-cluster fits into the evolving Kubernetes landscape, I can assure you that Karmada can definitely help. Um we'll be at the Karmada project booth, so to learn more about the or just multi-cluster orchestration in general, just stop by, and we'll be more than happy to talk with you. Thank you. Thank you,

sir.