Top 5 systemd troubles - a strategic view for distros

systemd is a new way to start a Linux-system with the expressed goal of rethinking all of init. These are my top 5 gripes with it. (»skip the updates«)

Update (2016-09-28): Systemd is an exploit kit just waiting to be activated. And once it is active, only those who wrote it will be able to defuse it — and check whether it is defused. And it is starting: How to crash systemd in one tweet? Alternatives? Use OpenRC for system services. That’s simple and fast and full-featured with minimal fuss. Use runit for process supervision of user-services and system-services alike.

Update (2014-12-11): One more deconstruction of the strategies around systemd: systemd: Assumptions, Bullying, Consent. It shows that the attitude which forms the root of the dangers of systemd is even visible in its very source code.

Update (2014-11-19): The Debian General Resolution resulted in “We do not need a general resolution to decide systemd”. The vote page provides detailed results and statistics. Ian Jackson resigned from the Technical Committee: “And, speaking personally, I am exhausted.”

Update (2014-10-16): There is now a vote on a General Resolution in Debian for preserving the ability to switch init systems. It is linked under “Are there better solutions […]?” on the site Shall we fork Debian™? :^|.

Update (2014-10-07): Lennart hetzt (german) describes the rhetoric tricks used by Lennart Poettering to make people forget that he is a major part of the communication problems we’re facing at times - and to hide valid technical, practical, pragmatical, political und strategical criticism of Systemd.

Update (2014-09-24): boycott systemd calls for action with 12 reasons against systemd: “We do recognize the need for a new init system in the 21st century, but systemd is not it.”

Update (2014-04-03): And now we have Julian Assange warning about NSA control over Debian, Theodore Ts’o, maintainer of ext4, complaining about incomprehensible systemd, and Linus Torvalds (you know him, right?) rant against disrupting behavior from systemd developers, going as far as refusing to merge anything from the developers in question into Linux. Should I say “I said so”? Maybe not. After all, I came pretty late. Others saw this trend 2 years before I even knew about systemd. Can we really assume that there won’t be intentional disruption? Maybe I should look for solutions. It could be a good idea to start having community-paid developers.

Update (2014-02-18): An email to the mailing list of the technical committee of debian summarized the strategic implications of systemd-adoption for Debian and RedHat. It was called conspiracy theory right away, but the gains for RedHat are obvious: RedHat would be dumb not to try this. And only a fool trusts a company. Even the best company has to put money before ethics.

Update (2013-11-20): Further reading shows that people have been giving arguments from my list since 2011, and they got answers in the range of “anything short of systemd is dumb”, “this cannot work” (while OpenRC clearly shows that it works well), requests for implementation details without justification and insults and further insults; but the arguments stayed valid for the last 2 years. That does not look like systemd has a friendly community - or is healthy for distributions adopting it. Also an OpenRC developer wrote the best rebuttal of systemd propaganda I read so far: “Alternativlos”: Systemd propaganda (note, though, that I am biased against systemd due to problems I had in the past with udev kernel-dependencies)

  • Losing Control: systemd does so many crucial things itself that the developers of distributions lose their control over the init process: If systemd developers decide to change something, the distributions might actually have to fork systemd and keep the fork up-to-date, and this requires rare skills and lots of resources (due to the pace of systemd). See the Gentoo eudev-Project for a case where this had to happen so the distribution could keep providing features its users rely on. Systemd nowadays incorporates udev. Go reason how systemd devs will act.1 Why losing control is a bad idea: Strategy Letter V: Commodities

  • No scripts (as if you can know beforehand all the things the init system will need to do in each distribution). Nowadays any system should be user-extendable to avoid bottlenecks for development. This essentially boils down to providing a scripting language. Using the language which almost every system administrator knows is a very sane choice for that - and means making it possible to use Shell-Scripts to extend the init-system. Scripts mean that the distribution will never be in a position where it is blocked because it absolutely can’t provide a given fringe feature. And as the experiment with paludis in Gentoo shows, an implementation in C isn’t magically faster than one in a scripting language and can actually be much slower (just compare paludis to pkgcore), because the execution time of the language only very rarely is the real bottleneck - and you can easily shell out that part to a faster language with negligible time loss,2 especially in shell-scripts (pun partially intended). While systemd can be told to run a shell script, this requires a mental context switch and the script cannot tie into all the machinery inside systemd. If there’s a bug in systemd, you need to fix systemd, if you need more than systemd provides out of the box, you need either a script or you have to patch systemd, and otherwise you write in a completely different language (so most people won’t have the skills to go beyond the fences of the ground defined by the systemd developers as proper for users). Why killing scripts is a bad idea: Bloatware and the 80/20 Myth

  • Linux-specific3 (are you serious??). This makes the distribution an add-on to the kernel instead of the distribution being a focus point of many different development efforts. This is a second point where distributions become commodities, and as for systemd itself, this is against the interest of the distributions. On the other hand, enabling the use of many different kernels strengthens the Distribution - even if currently only few people are using them. Why being Linux-only is a bad idea for distributions: Strategy Letter V: Commodities

  • Requiring an up-to-date kernel. This problem already gives me lots of headaches for my OLPC due to udev (from the same people as systemd… which is one of the reasons why I hope that Gentoo-devs will succeed with eudev), since it is not always easy to go to a newer kernel when you’re on a fringe platform (I’m currently fighting with that). An init system should not require some special kernel version just to boot… Why those hard dependencies are a bad idea: Bloatware and the 80/20 Myth AND Strategy Letter V: Commodities

  • Requiring D-Bus. D-Bus was already broken a few times for me, and losing not just some KDE functionality but instead making my system unbootable is unacceptable. It’s bad enough that so much stuff relies on udev.4

In my understanding, we need more services which can survive without the others, so the system gets resilient against failures in a given part. As the system gets more and more complex, this constantly gets more important: Less interdependencies, and the services which are crucial to get my system in a debuggable state should be small and simple - and should not require many changes to implement new features.

Having multiple tools to solve the same problem looks like wasted resources, but actually this extends the range of problems which can be solved with our systems and avoids bottlenecks and single points of failure (either tools or communities), so it makes us resilient. Also it encourages standard-formats to minimize the cost of maintaining several systems side-by-side.

You can see how systemd manages to violate all these principles…

This does not mean, that the features provided by systemd are useless. It says that the way they are embedded in systemd with its heavy dependencies is detrimental to a healthy distribution.

Note: I am neither a developer of systemd, nor of upstart, sysvinit or OpenRC. I am just a humble user of distributions, but I can recognize impending horrible fallout when I see it.


I’ll finish this with a quote from 30 myths about systemd:

We try to get rid of many of the more pointless differences of the various distributions in various areas of the core OS. As part of that we sometimes adopt schemes that were previously used by only one of the distributions and push it to a level where it's the default of systemd, trying to gently push everybody towards the same set of basic configuration.
— Lennart Poettering, main developer of systemd

I could not show much clearer why distributions should be very wary about systemd than Lennart Poettering does here in the post where he tries to refute myths about systemd.

PS: I’m definitely biased against systemd, after having some horrifying experiences with kernel-dependencies in udev. Resilience looks different. And I already modified some init scripts to adjust my systems behavior so it better fits my usecase. Now go and call me part of a fringe group which wants to add “pointless differences” to the system. If you force Gentoo devs to issue a warning in the style of “you MUST activate feature X in your kernel, else your system will become unbootable”, this should be a big red flag to you that you’re doing something wrong. If you do that twice, this is a big red flag to users not to trust your software. And regaining that trust requires reestablishing a long record of solid work. Which I do not see at the moment. Also do read Bloatware and the 80/20 Myth (if you didn’t do that by now): It might be true that 80% of the users only use 20% of the features, but they do not use the same 20%.

  1. Update 2014: Actually there is no need to guess how the systemd developers will act: They showed (again) that they will keep breaking systems of their users: “udev now silently fails to do anything useful if devtmpfs is missing, almost as if resilience was a disease” — bonsaikitten, Gentoo developer, 2014-01, long after udev was subsumed into systemd. 

  2. Running a program in a subshell increases the runtime by just six milliseconds. I measured that when testing ways to run GNU Guile modules as scripts. So you have to start almost 100 subshells during bootup to lose half a second of runtime. Note that OpenRC can boot a system and power down again in under 0.7 seconds and the minimal boot-to-login just takes 250 ms. There is no need for systemd to get a faster boot. 

  3. The systemd proponents in the debian initsystem discussion explicitly stated that they don’t want to port systemd to other kernels. 

  4. And D-Bus is slow, slow, slow when your system is under heavy memory and IO-pressure, as my systems tend to be (I’m a Gentoo user. I often compile a new version of all KDE-components or of Firefox while I do regular work on the computer). From dbus I’m used to reaction times up to several seconds… 

Inhalt abgleichen
Willkommen im Weltenwald!

Beliebte Inhalte news