AI
What Karmada Does and How It Is Built
Karmada is a CNCF project for Kubernetes-native multi-cluster management. Its architecture mirrors a standard Kubernetes cluster: a control plane joined to a set of member clusters. The central component is the standard kube API server, which keeps compatibility with existing Kubernetes tooling. In place of kube-scheduler, Karmada ships its own scheduler that makes placement decisions across an entire cluster fleet. Member clusters can sit on-prem, in a private cloud, or across multiple public cloud providers.
Scheduling, Failover, and Cross-Cluster Features
Karmada includes advanced policies to cover different scheduling requirements. Cross-cluster application failover handles application continuity during disaster scenarios. Unified authentication, authorization, and auditing apply across all member clusters from a single control plane. A global resource view gives operators one query entry point to inspect resources across the full fleet. Multi-cluster service discovery rounds out the set of cross-cluster capabilities built into the project.
CNCF Maturity and Production Adoption
Karmada was published as open source in 2021 and became a CNCF sandbox project that same year. In 2023 it moved to CNCF incubating status. The project applied for CNCF graduation last year and is on track to graduate in 2025. Production adopters include Huawei, Bloomberg, and Trip.com. Contributors from many different companies sustain the project, and Abellard described the community as fast-growing.
Bloomberg’s Use and Upcoming KubeCon Talks
Bloomberg runs Karmada in production and Abellard will speak at KubeCon on a zero-downtime migration from the retiring Ingress-Nginx project to a multi-cluster Kubernetes platform built on Karmada at Bloomberg. A separate talk covers decomposing and governing large LLM jobs across multiple clusters managed by Karmada. A third session addresses disaster-resilient Argo Workflows on multi-cluster Kubernetes powered by Karmada. A Karmada maintainer track talk goes deeper on the project itself.
Notable Quotes
multicluster orchestration has become a critical building block of the modern um platform engineering stack Joe Nathan Abellard · ▶ 00:33
the central component um here labeled as comma API server is just the standard cube API server which enables seamless integration with the existing Kubernetes tool chain Joe Nathan Abellard · ▶ 01:16
tomorrow I’ll be doing a talk on how we executed a zero downtime migration from the now retiring engineext project to multicluster Kubernetes platform built at top commada um at Bloomberg Joe Nathan Abellard · ▶ 03:26
Key Takeaways
- Karmada replaces kube-scheduler with its own scheduler for fleet-wide placement decisions.
- The project hit CNCF incubating in 2023 and targets graduation in 2025.
- Bloomberg and Trip.com both run Karmada in production today.
About the Speaker(s)
Joe Nathan Abellard is a Senior Software Engineer on Bloomberg’s Cloud Native Compute Services (CNCS) Platform Engineering team. He is a core owner of the Karmada project and an active open source contributor.