Ah, now I understand your setup. To answer the title question, I’ll have to be a bit verbose with how I think Incus behaves, so that the Docker behavior can be put into context. Bear with me.

br0 has the same MAC as the eth0 interface

This behavior stood out to me, since it’s not a fundamental part of Linux bridging. And it turns out that this might be a systemd-specific thing, since creating a bridge is functionally equivalent to a software switch, where every port of the switch has its own MAC, and all “clients” to that switch also have their own MAC. If I had to guess, systemd does this so that traffic from the physical interface (eth0) that passes directly through to br0 will have the MAC from the physical network, thus making it easier to understand traffic flows in Wireshark, for example. I personally can’t agree with this design choice, since it obfuscates what Linux is really doing vis-a-vis a software switch. But reusing this MAC here is merely a weird side-effect and doesn’t influence what Incus is doing.

Instead, the reason Incus needs the bridge interface is precisely because a physical interface like eth0 will not automatically forward frames to subordinate interfaces. Whereas for a virtual switch, that’s the default. To that end, the bridge interface is combined with virtual ethernet (veth) interfaces – another networking primitive in Linux – to each container that Incus manages. The behavior of a veth is akin to a point-to-point network cable, plus the NICs on both ends. That means a veth always consists of a pair of interfaces, where traffic into one end comes out the other, and each interface has its own MAC address. Functionally, this is the networking equivalent of a bidirectional pipe.

By combining a bridge (ie a virtual switch) with veth (ie virtual cables), we have a full Layer 2 network topology that behaves identically as if it were a physical bridge with physical cables. Thus, your DHCP server is none the wiser when it sends and receives BOOTP traffic for assigning an IP address. This is the most flexible way of constructing a virtual network within Linux, since it has feature-parity with physical networks: there is no Macvlan or Ipvlan or tunneling or whatever needed to make this work. Linux is just operating as a switch, with all the attendant flexibility. This architecture is what Calico – a network framework for Kubernetes – uses, in order to achieve scalable, Layer 3 connectivity to containers; by default, Kubernetes does not depend on Layer 2 to function.

OK, so we now understand why Incus does things the way it does. For Docker, when using the Macvlan driver, the benefits of the bridge+veth model are not achieved, because Macvlan – although being a feature of Linux networking – is something which is implemented against an individual interface on the host. Compare this to a bridge, which is a standalone concept and thus can exist with or without any interfaces to the host: when Linux is actually used as a switch – like on many home routers – the host itself can choose to have zero interfaces attached to the switch, meaning that traffic flows through the box, rather than to the box as a destination.

So when creating subordinate interfaces using Macvlan, we get most of the bridging behavior as bridge+veth but the Macvlan implementation in the kernel means that outbound traffic from a subordinate interface always get put onto the outbound queue of the parent interface. This makes it impossible for a subordinate interface to exchange traffic with the host itself, by design. Had they chosen to go the extra mile, they would have just reinvented a version of bridge+veth that is excessively niche.

We also need to discuss the behavior of Docker networks. Similar to Kubernetes, containers managed by Docker mandate having IP connectivity (Layer 3). But whereas Kubernetes will not start a container unless an IPAM (IP Address Management) plugin explicitly provides an IP address, Docker’s legacy behavior is to always generate a random IP address from a default range, unless given an IP explicitly. So even though bridge+veth or Macvlan will imbue Layer 2 connectivity to a DHCP server to obtain an IP address, Docker is eager to provide an IP, just so the container has one from the very start. The distinction between Docker and Kubernetes+Calico is thus one of actual utility: by getting an address from Calico’s IPAM, Kubernetes knows that the address will actual work for networking, because Calico also creates/manages a network. Whereas Docker has no problem assigning an IP but not actually checking if this IP can be used on that network; it’s almost a pro-forma exercise.

I will say this about early Docker: although they led the charge for making containers useful, how they implemented networking was very strange and led to a whole class of engineers who now have a deep misunderstanding of how real networks operate, and that only causes confusion when scaling up to orchestrated container frameworks like Kubernetes that depend on rigorous understanding of networking and Linux implementations. But all the same, Docker was more interested in getting things working without external dependencies like DHCP servers, so there’s some sense in mandating an IP locally, perhaps because they didn’t yet envision that containers would talk to the physical network.

The plugin that you mentioned operates by requesting a DHCP-assigned address for each container, but within the Docker runtime. And once it obtains that address, it then statically assigns it to the container. So from the container’s perspective, it’s just getting an IP assigned to it, not aware that DHCP has happened at all. The plugin is thus responsible for renewing that IP periodically. It’s a kludge to satisfy Docker’s networking requirements while still using DHCP-assigned addresses. But Docker just doesn’t play well with Layer 2 physical networks, because otherwise the responsibility for running the DHCP client would fall to the containers; some containers might not even have a DHCP client to run.

If I’m missing something about MACVLAN that makes DHCP work for Docker, let me know!

Sadly, there just isn’t a really good way to do this within Docker, and it’s not the kernel’s fault. Other container runtimes like containerd – which relies wholly on the standard CNI plugins and thus doesn’t have Docker’s networking footguns – have no problem with containers running their own DHCP client on a bridged network. But for any container manager to handle DHCP assignment without the container’s cooperation always leads to the same kludge as what Docker did. And that’s probably why no major container manager does that natively; it’s hard to solve.

I do wish there could be something like Incus’ hassle-free solution for Docker or Podman.

Since your containers were able to get their own DHCP addresses from a bridged network in Incus, can you still run the DHCP client on those containers to override Docker’s randomly-assigned local IP address? You’d have to use the bridge network driver in Docker, since you also want host-container traffic to work and we know Macvlan won’t do that. But even this is a delicate solution, since if DHCP fails to assign an address, then your container still has the Docker-assigned address but it won’t be usable on the bridged network.

The best solution I’ve seen for containers on DHCP-assigned networks is to not use DHCP assignment at all. Instead, part of the IP subnet is carved out, a region which is dedicated only for containers. So in a home IPv4 network like 192.168.69.0/24, the DHCP server would be restricted to only assigning 192.168.69.2 through 192.168.69.127, and then Docker would be allowed to allocate the addresses from 192.168.69.128 to 192.168.69.254 however it wants, with a subnet mask of 255.255.255.0. This mask allows containers to speak directly to addresses in the entire 192.168.69.0/24 range, which includes the rest of the network. The other physical hosts do the same, allowing them to connect to containers.

This neatly avoids interacting with the DHCP server, but at a loss of central management and it splits the allocatable addresses into smaller parts, potentially causing exhaustion in one side while the other has spare addresses. Yet another reason to adopt IPv6 as the standard for containers, but I digress. For Kubernetes and similar orchestration frameworks, DHCP isn’t even considered since the orchestrator must have full internal authority to assign addresses with its chosen IPAM plugin.

TL;DR: if your containers are like mini VMs, DHCP assignment is doable. But if they’re pre-packaged appliances, then only sadness results when trying to use DHCP.

What about using the default docker bridge networking instead of macvlan? You can access docker containers from the host, they can talk to each other if on the same bridge network, and there’s nothing hardcoded into the docker compose files.