This is somehow the opposite of #1768. From a separation of concerns standpoint, both will achieve the separation. However, this separation is similar to the current dependencies present in Firecracker, and might work with less fuss than #1768, which needs a new design, or at least thoughts/discussions about what can be done differently to support it. Mentioned this here to keep these issues close, and have both suggestions at hand when resolving them.

Hmm, why do you say they're opposite?

I think of them as separate issues. This is related to the VirtioNetwork interface and how we make the interface/implementation not depend on Mmds. This issue doesn't really care where the MmdsNetworkStack is defined. It is just about separating the core functionality of the virtio network device from mmds in such a way that mmds can easily become a feature for example.

Hmm, why do you say they're opposite?

I think of them as separate issues. This is related to the VirtioNetwork interface and how we make the interface/implementation not depend on Mmds. This issue doesn't really care where the MmdsNetworkStack is defined. It is just about separating the core functionality of the virtio network device from mmds in such a way that mmds can easily become a feature for example.

I said that their opposite because of the the fact that with pluggable network stack, Firecracker virtio net will still depend on the MMDS network stack, while #1768 suggests to reverse this dependency. However, reading again what you're thinking at, it looks like the changes might be very similar, and separate, as you say, or sequential maybe. I take back what I've said. :D

Related to #518

Marking this as parked, as untangling MMDS from the networking path is something we are still interesting in, yet do not have any immediate plans to work on

Hello, I'm interested in this issue. To decouple MMDS from the network stack, I think vsock is a good way to go. The basic idea is to launch an HTTP server on the host side that listens on a vsock endpoint, so that the guest can directly send HTTP requests to the corresponding socket and port. Since MMDS is used far less frequently than the general network transmission, separating the MMDS service seems very natural. It also paves the way for offloading the normal network datapath to a vhost‑user‑net backend driver without having to worry about filtering the MMDS traffic.

I’m curious to hear what the team think about this approach. Any feedback would be greatly appreciated!