bridge-marker-dp

Device plugin compatible implementation for Kubevirt's Bridge Marker.

Why?

The original bridge-marker project relies on Kubernetes' built in ability to advertise extended resources, it does so by directly patching the node/status subresource and doesn't check for device health (Kubelet is completely unaware of the extended resource).

The device plugin implementation aims to address several of the original's issue:

• Remove API permissions completely - As a device plugin, bridge-marker only requires access to the host network and Kubelet's device plugins directory, Kubelet takes care of updating the node status instead.
• Allow for proper health monitoring - The previous iteration of bridge-marker doesn't check whether the bridge is "healthy" (e.g. up) and therefore doesn't update the allocatable section of node/status for the device which may allow for improper scheduling of consuming Pods. See manifests/helm for deployment.

How?

bridge-marker-dp follows the basic implementation of a Kubernetes device plugin but it is unique in the sense that bridges aren't considered proper devices. The only implemented part (and the required one) is ListAndWatch to serve Kubelet with device updates. Allocation setup is taken care of by bridge CNI.

bridge-marker-dp leverages https://github.com/vishvananda/netlink to subscribe for link updates and act upon them to accurately measure bridge health as well as allow for adding new bridge device plugins at runtime.

Deploy?

To build and deploy locally using Kubevirt's cluster-up, make sure you have the following installed binaries:

• helm
• kubectl

And run:

make cluster-sync

To build locally:

make build