Repository navigation
Replies: 1 comment
|
Thanks for laying out the options so clearly. I am not going to take this into my own roadmap. I do not use Docker, and NetPulse is deliberately a single self-contained binary with the agent running on the router. I do not want to add a runtime integration that I cannot test or maintain myself. The two immediate problems behind the NAS case are already covered by #1534, which keeps a portless wired client under its switch or NAS, and #1540, which derives local traffic from switch-port counters. For my own use that is enough, so I am closing this discussion. A larger container model or runtime collector is not on my roadmap. That does not block community work. If someone who runs Docker or LXC wants to prototype a small, optional slice and is willing to validate and maintain it, I am open to reviewing a PR with tests. A container runtime must never become a requirement for running NetPulse. |
Uh oh!
There was an error while loading. Please reload this page.
Motivation
Containers are everywhere (Docker on a Synology/QNAP NAS, on any Linux box, Kubernetes pods, LXC), but NetPulse has no first-class notion of a container. Today a container is just a generic
band:"cable"client attributed to its host viarouterId(or worse, force-anchored to the gateway). That makes both display and metrics weak:02:42:…), so the "device" is ephemeral across recreates.Existing precedent to build on
Device.Infraalready seals roles server-side:hypervisor | ct | vm | managed-switch | ap, and there is a genuine Proxmox integration (reads the PVE API, hangs CT/VM guests under the host; #561, #936/#940, #1446). So "a workload under its host" is an established shape — Docker/LXC/k8s are just a much more common, lighter-weight instance of it that currently isn't wired up.Options to discuss
A. A dedicated
containerinfra + a host-side collector. The host agent (or a small collector) enumerates the runtime (docker ps/ containerd //sys/fs/cgroup/*/net_*, or veth counters) and reports each container as a child of the host with real rx/tx and a stable id (image/name/container-id instead of the churny MAC). Topology renders them under the host via the device-hub path (no switch port needed).B. A generic "attached children" mechanism (no runtime API). Any device can be a hub; children attach via
attachToand are laid out around the parent regardless of port. Cheap, fixes display for the NAS case, but gives no real per-container traffic or lifecycle.C. Keep containers as plain clients and just fix the portless-child placement (my current frontend option). Cosmetic only — doesn't address semantics or traffic.
Questions for maintainers / community
Device, or a nested view under its host? (There's alreadyInfra=ct/hypervisor— is that the right home, or a newcontainerrole?)I hit all of this wiring up a NAS's Docker container and it's currently painful enough that a dedicated type seems worth doing. Happy to prototype a slice (e.g. a
containerinfra + host-agent enumeration + placement under the host) if there's appetite.All reactions