AI
KubeEdge Graduation: What It Took and Why It Matters
KubeEdge launched as part of the Kubernetes IoT working group in November 2018. Huawei donated it to the CNCF in March 2019. It entered incubation in 2020 and graduated in October 2024 at KubeCon Chicago. That makes it the first edge project to graduate under the CNCF umbrella. Only 20 to 30 projects out of all CNCF members have reached that status. The project now holds 8,000+ GitHub stars, 2,000+ forks, and 1,600 contributors from more than 100 organizations.
How KubeEdge Handles Cloud-to-Edge Coordination
The core design runs the Kubernetes control plane in the cloud and places a modified agent at the edge instead of a standard kubelet. A bidirectional communication channel connects the two sides. When the connection drops, the edge node restores itself to the last desired state the control plane set. This matters because edge deployments face long latency, intermittent connectivity, and data-privacy constraints that a pure data-center topology does not encounter.
Federated Learning and Joint Inference at the Edge
KubeEdge’s SIG AI, active since 2018, supports TensorFlow, PyTorch, PaddlePaddle, and MindSpore. Federated learning keeps sensitive data on the edge node. A small clinic might hold a few patient records; a large hospital might hold millions. Neither dataset leaves the local facility. Each site trains locally, and only intermediate model weights go to the cloud for aggregation. Joint inference takes the opposite approach: a compressed small model runs on the edge and handles requests that pass a confidence threshold. Requests that fall short route to the full cloud model.
Satellite, Highway, and Vehicle Deployments
Three production cases illustrate scale. A satellite operator runs KubeEdge on each satellite as an edge node, processing imagery on-orbit and sending only results to the ground station. A Chinese highway authority manages 100,000 heterogeneous toll gates, spanning x86 and ARM architectures, running 50,000 applications, with firmware upgrades triggered by a single button. A new-energy vehicle program treats every car as an edge device: the car extracts sensor features locally, the cloud trains a battery-lifetime model, and updated versions push back to the fleet continuously.
Smart Retail and Collaborative Robots
Retail stores use an edge server per location connected to cameras and MQTT devices. When a camera detects a shopper pausing in front of a shelf, the edge node processes the dwell-time signal and the cloud responds by pushing a targeted video advertisement to a nearby screen. The robot case routes natural-language commands through a cloud NLP model, which translates intent into machine instructions. The edge device then dispatches those instructions to a group of collaborative robots that plan their own routes.
Building a Community That Outlasts the Founding Companies
Graduation required open governance first. KubeEdge runs two TSC seats and separate SIGs for networking, AI, robotics, security, and device management. No single company or country controls the project. Post-graduation, the community shifted focus from developer contributions to industry and end-user outreach, including university programs, developer awards, and CNCF mentorship slots each cycle. Hardware and software vendors can earn interoperability verification against the project’s conformance program, similar to CNCF AI conformance testing.
Notable Quotes
graduation for a project is not easy. Yeah, under CSF umbrella only about 20 to 30 project got graduated. Hongbing Zhang · ▶ 00:09
the sensitive data never leave your edge. So that for the enterprise use very important. Yin Ding · ▶ 05:52
every day the senet will capture connect and produce much data but the majority of the data cannot be transferred back to ground station for further processing. Hongbing Zhang · ▶ 10:58
this project will not be controlled or dominated by one or few country or few companies it’s really you know run by the community. Hongbing Zhang · ▶ 22:52
Key Takeaways
- KubeEdge graduated CNCF in October 2024, the first edge project to do so.
- Federated learning keeps sensitive patient and enterprise data on-device, never uploading raw records.
- Joint inference routes low-confidence edge results to a larger cloud model automatically.
- One Chinese highway authority runs KubeEdge across 100,000 heterogeneous toll-gate nodes.
- Post-graduation governance shifts the project from developer-driven to industry and end-user driven.
About the Speaker(s)
Yin Ding is an Engineering Manager at VMware leading the Photon team, with more than 15 years in large-scale distributed computing. He co-founded the CNCF KubeEdge project and chairs the LF Edge Akraino TSC. He has also contributed to Kubernetes, OpenStack, and Cloud Foundry.
Yue Bao is a software engineer at Huawei Cloud working full-time on open source, focused on lightweight edge computing for KubeEdge. She maintains the project and leads the KubeEdge SIG Release and Node.
Hongbing Zhang is COO of DaoCloud and a KubeEdge TSC member. He founded the IBM China Linux team in 2011 and led contributions to the Linux kernel, OpenStack, and Hadoop before moving to cloud-native infrastructure.