CNI 插件与边缘网络边界说明

This page clarifies the boundary between KubeEdge and third-party CNI plugins for edge networking. Understanding which component is responsible for each part of the network stack helps users troubleshoot problems faster and choose the right tool for their use case.

Overview

KubeEdge is an edge computing framework that extends Kubernetes orchestration to edge nodes. It does not replace the container networking layer. Pod networking still requires a CNI plugin, just as it does in any standard Kubernetes cluster.

The table below summarizes which component owns each responsibility:

Responsibility	Owner
Pod IP assignment	CNI plugin (Flannel, Calico, Cilium, etc.)
Pod-to-pod connectivity on the same node	CNI plugin
Cross-node pod networking (within the same LAN)	CNI plugin
Cross-LAN / cross-region service discovery	EdgeMesh
Cloud-to-edge control plane tunnel	KubeEdge (CloudHub / EdgeHub)
Device management (DeviceTwin, Mapper)	KubeEdge
Kubernetes API proxy for edge agents	KubeEdge MetaManager
Node registration and lifecycle	KubeEdge EdgeCore

What KubeEdge Does in Networking

KubeEdge's networking involvement is limited to the control plane tunnel between CloudCore and EdgeCore. This tunnel runs over WebSocket or QUIC and carries:

Pod scheduling decisions (spec delivery)
Device twin state synchronization
ConfigMap and Secret delivery
Node heartbeat and status reports

This tunnel does not carry application pod traffic. Pod-to-pod data plane traffic is always handled by the CNI plugin installed on the node.

Cloud Node
  └─ CloudCore (EdgeHub) ──────────────────────┐
                              WebSocket / QUIC  │ (control plane only)
Edge Node                                       │
  └─ EdgeCore (EdgeHub) ◄──────────────────────┘
       ├─ MetaManager   (API proxy for CNI agent calls)
       ├─ DeviceTwin    (device state management)
       └─ (CNI plugin runs independently alongside EdgeCore)

What the CNI Plugin Does

The CNI plugin is responsible for all pod data plane networking:

Allocating an IP address to each pod
Attaching a network interface to the pod's network namespace
Programming routing rules and iptables entries so pods can reach each other
Enforcing NetworkPolicy rules

KubeEdge does not intercept or modify pod traffic after the CNI plugin has set up the interface. If a pod can reach another pod on the same node, that is entirely the CNI plugin working correctly.

What EdgeMesh Does

EdgeMesh is an optional data plane component developed by the KubeEdge project. It is not a replacement for a CNI plugin. It sits on top of the network provided by the CNI and adds:

Service discovery for pods running on edge nodes that are behind NAT or on different LANs
Traffic proxying across network boundaries using LibP2P tunnels
Cross-region pod-to-service communication without requiring a flat, routable network

You would deploy EdgeMesh in addition to your CNI plugin when edge nodes are on isolated networks that cannot directly route to each other.

Edge Node Configuration Considerations

Because edge nodes are often managed separately from the cloud, there are a few CNI-specific things users need to handle manually on edge nodes that are typically automatic on cloud nodes:

CNI binaries: Ensure that CNI binaries are present at /opt/cni/bin (or the path configured in your container runtime, e.g., cni_bin_dir in containerd) on every edge node. Some CNI DaemonSets do not schedule onto edge nodes by default and need a toleration or nodeSelector to do so.
Preventing CNI DaemonSets from scheduling on edge nodes unintentionally: If your CNI DaemonSet is not designed for edge nodes, use a nodeSelector or affinity rule to prevent it from scheduling there. For example, when using Cilium with KubeEdge, the official configure_cilium.sh script automatically patches the Cilium DaemonSet to exclude edge nodes.
MetaManager as API proxy: CNI agents (such as cilium-agent) need to call the Kubernetes API server to configure themselves. On edge nodes, KubeEdge's MetaManager acts as a local proxy for these API calls so they work reliably even when the edge-to-cloud connection is intermittent. This requires KubeEdge v1.16 or later. See Enable Cilium with KubeEdge for a working setup guide.
IPv6 and dual-stack: KubeEdge supports IPv6. See IPv6 Support for details.

Troubleshooting Guide

Use the following decision tree to identify whether a networking problem is a CNI issue or a KubeEdge issue:

Symptom	Likely Owner	Where to Look
Pod on edge node has no IP address	CNI plugin	CNI agent logs on the edge node, `/opt/cni/bin` presence
Pods on the same edge node cannot ping each other	CNI plugin	CNI agent logs, iptables rules on the edge node
Pod on edge node cannot reach a pod on a different edge node (same LAN)	CNI plugin or routing	CNI overlay config, node-to-node routes
Pod on edge node cannot reach a pod on a different LAN	EdgeMesh	EdgeMesh installation and configuration
EdgeCore is not connecting to CloudCore	KubeEdge	CloudCore and EdgeCore logs, firewall rules on port 10000
Device twin state is not syncing	KubeEdge	EdgeCore DeviceTwin logs, Mapper logs
CNI agent (e.g., cilium-agent) crashes on edge node	KubeEdge MetaManager	Requires KubeEdge >= v1.16, check if `metaServer` is enabled in `edgecore.yaml`

Supported CNI Plugins

KubeEdge has been tested with the following CNI plugins in edge deployments:

CNI Plugin	Notes
Cilium	Fully supported with KubeEdge v1.16+. Requires `configure_cilium.sh`. See Enable Cilium with KubeEdge.
Flannel	Works for basic intra-LAN scenarios. No special edge-specific configuration required.
Calico	Works in routed mode. Edge nodes may require manual BGP peer configuration depending on topology.

Overview​

What KubeEdge Does in Networking​

What the CNI Plugin Does​

What EdgeMesh Does​

Edge Node Configuration Considerations​

Troubleshooting Guide​

Supported CNI Plugins​

Related Documentation​