14 different schemes multiplied by some acting slightly different in every version. Sure you can pin it, but that fixes only their internal back and forth, is only possible via the kernel cmdline and there is no guarantee for how long the old versions will stay available, as they deprecated much more invasive things in the past (e.g., cgroupv1) I'd expect them to also drop older versions here, breaking ones naming again.

And sure, one can pin interfaces to custom names, but why should anybody have to bother with such things?!

I like systemd a lot, but this is one of the thing they fumbled big time and seemingly still aren't done.

Pinning interfaces by their MAC to a short and usable name, would e.g. have been much more stable as doing that by PCI slot, which firmware updates, new hardware, newer kernel exposing newer features, ... changes rather often. This works well for all but virtual functions, but those are sub-devices of their parent interface anyway and can just get named with a suffix added to the parent name.

I imagine they went against mac address because it is not immutable, some folks rotate mac addresses for privacy/security reasons.

The original one is still there. Systemd knows even about that, it's differentiated as MAC vs PermanentMAC.

There are, unfortunately, some older devices (like some Sun systems) which use the same MAC address for every network interface on the device.

i thought about that, but couldn't you access the hardcoded address to identify the card?

but you also want to be able to change a card in a server without the device name changing. at least that used to be an issue in the past.

> as they deprecated much more invasive things in the past (e.g., cgroupv1) I'd expect them to also drop older versions here, breaking ones naming again

Note that the naming scheme is in control of systemd, not the kernel. Even if it is passed on the kernel commandline.

Yeah, I know, I spent more than a week into looking for options to reduce impact for all of our users.

And note that cgroupv1 also still works in the kernel just fine, only the part that systemd controlled was removed from systemd. You can still boot with cgroupv1 support on, e.g., Alpine Linux and OpenRC as init 1. So not sure if that will lessen my concerns about no guarantees for older naming-scheme versions, maintaining triple digits of them sure has its cost too.

And don't understand me wrong, sunsetting cgroupv1 was reasonable, but it was a lot of churn, it at least was a one time thing. The network interface naming situation is periodic churn, guaranteed to bite you every now and then just by using the defaults.

Can you tell me why NamePolicy=keep doesn't do the trick?

Looking myself for options to keep a Debian bare metal server I admin from going deaf and mute the next time I upgrade it... It still uses an /etc/network/interfaces file that configures a bridge for VMs to use, and the bridge_ports parameter requires an interface name which, when I upgraded to Bookworm, changed.

At this rate maybe I'll write a script that runs on boot and fixes up that file with whatever interface it finds, then restarts the network.

This worked brilliantly in Debian for more than a decade, had almost zero downside, and just did what asked. I went through 3+ dist-upgrades, for the first time in my life, without a NIC change.

It was deprecated for this nonsense in systemd.

Yes, there were edge cases in the Debian scheme. Yet it did work with VMs (as most VMs kept the same MAC in config files), and it was easy to maintain if you wanted 'fresh'. Just rm the pin file in the udev dir. Done.

Again it worked wonderful on every VM, every bare metal system I worked with.

One of the biggest problems with systemd, is it seems to be developed by people that have no real world, industrial scale admin experience. It's almost like a bunch of DEVs got together, couldn't understand why things were "so confusing", and just figured "Oh, it must be a mistake".

Nope.

It's called covering edge cases, ensuring things are stable for decades, because Linux and the init system are the bottom of the stack. The top of the stack changes like the wind in spring, but the bottom of the stack must be immensely stable, consensus driven, I repeat stable change.

Systemd just doesn't "get" that.

systemd's design choices here were influenced by a lot of bugs Red Hat received where failed hardware was swapped out and interface names changed as a result. Real world enterprise users wanted this, it wasn't an arbitrary design choice.

That's quite the jump.

Some real world users asked for a fix. They did not mean they asked specifically for this fix.

There were other ways to handle this.

With Debian's system, you could wipe the state files, and for example eth0/etc would be reassigned per initialization order. Worked fine.

Even if you didn't like that, pre-Systemd udev allowed assigned by a variety of properties, including bus identifiers.

It was merely that Redhat, as usual, was so lacking in sophistication, unlike Debian.

It turns out that people do not love having to log into a machine after a network card swap to get the new network card to have the same name. Initialisation order is explicitly not guaranteed by the kernel and so absolutely does not work every time.

> systemd's design choices here were influenced by a lot of bugs Red Hat received where failed hardware was swapped out and interface names changed as a result.

Under RH-based systems the ifcfg-* files had a HWADDR variable, so if you swapped a card you could get the new MAC address and plug it in there and get the same interface name. There was also udevd rules where you map names to particular hardware, including particular MACs.

> Real world enterprise users wanted this, it wasn't an arbitrary design choice.

As a real world sysadmin, working now a few of decades in this field (starting with non-EL-RH, then BSD, then Solaris, then RHEL, Debian, and now Ubuntu), I have never wanted this.

Great. A tech swaps out a network card, now how do I log in to rewrite the ifcfg file when the interface wasn't brought up with the correct config because it has a different name?

> now how do I log in to rewrite the ifcfg file when the interface wasn't brought up with the correct config because it has a different name?

Unlike most desktops, basically all servers got out-of-band management (e.g. IPMI) and a NIC swap is something that needs a tech physically near the server, so even a simple serial console is easily plugged in. Or how will that new NIC work with the whole network, like any basic networking setup or firewall won't allow traffic from arbitrary MACs, so normally this needs to be coordinated already anyway in an enterprise setting, e.g. through a change management process.

And why would one optimize the whole design for network naming for the edge case and not the much more common one like simple software updates.

And the design is not even being able to guarantee it for the edge case. Plugin that NIC in a different PCI slot, or let the firmware to a blip and report it differently–all things that happened!–and you still got no network with net naming scheme. Worse, you reboot after a systemd update, and you can have no network either. Or the kernel learns that your NIC supports virtual functions, guess what, no network because the (seemingly just-in-time) predictable naming scheme now sees that information changing its previous prediction.

I never will be able to understand how one can argue for breaking the common use case, nobody argues that there isn't a real problem or that there is the One True Way™ to solve it (at least I do not intend so), but arguing for using a certainly not ideal default that optimized for an edge case feels a bit like some sunk cost fallacy to me.

Sorry for my wall of text, I would really like to care less, but at $work I am exposed to this mess directly, not only for our infra but for all users of our projects, can all be done and managed, sure, but the churn and hours I have to put in thanks to this feels unnecessary and could be used for much more useful things.

> A tech swaps out a network card, now how do I log in to rewrite the ifcfg file when the interface wasn't brought up with the correct config because it has a different name?

IPMI/iDRAC/iLO/XCC/etc.

I dislike systemd's Predictable Network Interface Names, so I disable them with this kernel command line option: net.ifnames=0

Welcome back, eth0. :)

Yup.. use this default on all my systems. Did a bookworm->trixie upgrade today on my mailserver, and everything worked, as it still just has eth0 ;)

The "stable" interface naming scheme is a scam. And I have proof. Test upgraded a VM today, from bookworm to trixie. And guess what. Everything worked, except after reboot the network interface was unconfigured? Guess what. The name changed...

That can only happen if the emulated hardware layout presented to the VM changes. I'd look at that before calling anything a scam.

Scam is probably the wrong word, and it's choice might be a bit feeling fueled, but it's really not true that this only depends on the HW.

systemd also changes behavior in what naming policies are the default and what it considered as input, it did that since ever but started to version that since v238 [0]. Due to that the HW can stay exactly the same but names still change. I see this in VMs that stay exactly the same, no software update, not change in how the QEMU cli gets generated, really nothing changed from the outside virtual HW POV, interface name still changes.

The underlying problem was a real one, the solution seems like a bit of a sunken cost fallacy, and it added more problem dimensions than there previously exist.

Besides, even if the HW would change, shouldn't a _predicatble_ naming scheme be robust to not care about that as long as the same NIC is still plugged in somewhere?

Disclaimer, as stated elsewhere: I really like systemd, I'm not one that speaks out against it lightly, but the IF naming is not something they got right, but rather made worse for the default case. Being able to easily pin interface names through .link files is great, but requiring users to do that or have no network after an upgrade, especially for simple one-NIC use cases in a completely controlled environment like a VM is just bonkers.

[0]: https://www.freedesktop.org/software/systemd/man/latest/syst...

Ah, ok, I didn't think of systemd version changes. Thanks.

Regarding your rhetorical question about "the same NIC", I think the problem is in determining whether the NIC is the same, and it is not an easy one to solve. I remember that older Suse Linux versions used to pin the interface name to the NIC's MAC address in an udev rule file that got autogenerated when a NIC with a given MAC first appeared on the system, but they stopped doing that.

Yeah, the permanent MAC address (i.e., the one the card actually reports to the system not the one dynamic one it can use) would be the safest bet, as that is the most stable thing there is, and more importantly, it is very relevant for switches and firewalls in enterprise settings, so if it changes it's often likely that network access will be broken any way, so one basically can only win with using the MAC as main identifier IMO, at least compared to the current status quo.

Sadly a NIC's permanent MAC is known to not always be unique: https://www.howtogeek.com/228286/how-is-the-uniqueness-of-ma...

As long as you only got NICs with different permanent MAC addresses installed that does not matter for getting actually long-term stable names.

And for the other case you can still fallback to the other policies, it still will be much more stable by default.

Please note that I don't say that MAC is perfect, but using something that is actually tied to a NIC itself would fare much better by default compared to the NICs position as determined by a bunch of volatile information, and what normally does not matter to me at all, as e.g., I will always use a 100G NIC as ceph private network while the 25G ones as public one, no matter where they are plugged in. That someone configures something by location is the excpection, not the norm.

I don't know if this is still the case but the last time I went without ifnames=0 adding a GPU would cause all the network interfaces to get new names. Junk.

That’s not a scam and that’s not proof. That’s an upgrade problem. Stop misusing the word and devaluing it.

I've had too many incidents to fix / work around since "stable" interface names were introduced. Sorry that you are offended. But the fact of the matter is, most of these issues would not have hit me if we'd just kept the old ethX naming scheme.

I do change ifupdown to systemd-networkd on every server I care about. Matching on MACAddress is a great imprvement. It doesn't help that Debian still insists on keeping ifupdown.

The best use of AI I've gotten so far is having it explain to me how to manage a Fedora Server's core infrastructure "the right way". Which files, commands, etc. to permanently or temporarily change network, firewall, DNS, NTP settings.