Does anybody have experience with both systems enough to compare them?
I’m currently using ifupdown on my Debian server as that’s the default, but it seems that the modern way of managing the local network is via systemd-networkd so I’m contemplating putting the effort in to migrate.
Would those of you who have experience with it, recommend it?
In my short investigation, I have made the following observations:
- using networkd means you can use networkctl to manually control the interfaces which is quite convenient
- networkd aims to be fully declarative
- networkd separates the creation of virtual interfaces (netdev files) from their configuration (network files)
- networkd doesn’t support all networking features (e.g. namespaces)
- networkd is systemd, but surprisingly I can’t find information on how to create other unit files that depend on the individual network files going up or down, other than networkd-dispatcher. I don’t like dispatcher because just like ifupdown it triggers all the scripts and you need if tests to exclude all interfaces you don’t need to be affected. I’d like to write unit files that can be targeted to activate and deactivate when a particular interface goes up or down.
- networkd, other than via dispatcher, does not seem to support adding arbitrary commands to run like ifupdown supports via e.g. pre-down, post-up, etc.
I’m using ifupdown2 and have services depending on the state of virtual network devices (BindsTo=sys-devices-virtual-net-.device).
I hate systemd with a passion, as the refuses to wait for networking when you haven some service specified to be started After networking, but it start when the start of networking actions has sbeen tarted (not after networking is finished bringing up everything)
The only think I let systemd do is monitor the state of the devices and set the interface name, as udev seems to mess these names up in combination with systemd (and there is to much depending on systemd now in Debian to get rid of the junk)
If you apply what is written at https://www.freedesktop.org/wiki/Software/systemd/NetworkTarget/#cutthecraphowdoimakesurethatmyservicestartsafterthenetworkisreallyonline it will work.
Systemd-networkd has a learning curve, once you learn it you’ll find it superior and more flexible than anything else.
Then I’ll need to do a global rewrite of all distrubution delivered service files and replace network.target with network-online.target. Then I don’t understand why maintainers all keep using network.target. For every service that needs some kund of netwirj active I alkready have iverrides to link then to the virtual or physical interfaces they actually use.
I still find it a solution desperately looking for a problem to solve. Yes, you can control loads of dependencies, but you could already do that with the init scripts. The main selling point on introduction was ‘it is faster’. Why would you want something faster when you use it once every kernel update? When you design your solutions correctly (redundant), you won’t even notice a reboot of 1 system in the setup.
Because most services are able to dynamically accommodate networking changes and act accordingly and it’s rare to have cases where we really need a working link with addresses and a proper connection when the service starts. For those special cases, as described, you need to enable
systemd-networkd-wait-online.service
and includeAfter=network-online.target
and/orWants=network-online.target
in your service.Programs should be designed to detect and react to networking changes and both Apache and Nginx are good examples of software that does that. There are simpler cases like stuff that needs to bind to non-existent IPs at boot (before networking) that can be dealt with
ipv4.ip_nonlocal_bind
andnet.ipv6.ip_nonlocal_bind
as described here.This may also interest you as it includes what “up” means a few other details: https://systemd.io/NETWORK_ONLINE/
Systemd does a LOT more than “you can control loads of dependencies”, it solves tons of painful issues and provides a cohesive ecosystem of tools to manage stuff like:
journalctl
as a unified interface so you can quickly browse them by specific, filters, timestamps and events.crontab
with something auditable that actually makes sense;machinectl
to manage them.Systemd is very useful and most people (including myself) don’t know about half of the things it can do. It isn’t teached because of the current poor state of thing when it comes to software development but if you use it you’ll quickly find how to have systems running very well with less processes than conventionally.