Slashdot Mirror
Ask Slashdot: Practical Alternatives To Systemd?
First time accepted submitter systemDead (3645325) writes "I looked mostly with disinterest at Debian's decision last February to switch to
systemd as the default init system for their future operating system releases. The Debian GNU/Linux distribution is, after all, famous for allowing users greater freedom to choose what system components they want to install. This appeared to be the case with the init system, given the presence of packages such as sysvinit-core, upstart, and even openrc as alternatives to systemd.
Unfortunately, while still theoretically possible, installing an alternative init system means doing without a number of useful, even essential system programs. By design, systemd appears to be a full-blown everything-including-the-kitchen-sink solution to the relatively simple problem of starting up a Unix-like system. Systemd, for example, is a hard-coded dependency for installing Network Manager, probably the most user-friendly way for a desktop Linux system to connect to a wireless or wired network. Just this week, I woke up to find out that systemd had become a dependency for running PolicyKit, the suite of programs responsible for user privileges and permissions in a typical Linux desktop.
I was able to replace Network Manager with connman, a lightweight program originally developed for mobile devices. But with systemd infecting even the PolicyKit framework, I find myself faced with a dilemma. Should I just let systemd take over my entire system, or should I retreat to my old terminal-based computing in the hope that the horde of the systemDead don't take over the Linux kernel itself?
What are your plans for working with or working around systemd? Are there any mainstream GNU/Linux distros that haven't adopted and have no plans of migrating to systemd? Or is migrating to one of the bigger BSD systems the better and more future-proof solution?"
Unfortunately, while still theoretically possible, installing an alternative init system means doing without a number of useful, even essential system programs. By design, systemd appears to be a full-blown everything-including-the-kitchen-sink solution to the relatively simple problem of starting up a Unix-like system. Systemd, for example, is a hard-coded dependency for installing Network Manager, probably the most user-friendly way for a desktop Linux system to connect to a wireless or wired network. Just this week, I woke up to find out that systemd had become a dependency for running PolicyKit, the suite of programs responsible for user privileges and permissions in a typical Linux desktop.
I was able to replace Network Manager with connman, a lightweight program originally developed for mobile devices. But with systemd infecting even the PolicyKit framework, I find myself faced with a dilemma. Should I just let systemd take over my entire system, or should I retreat to my old terminal-based computing in the hope that the horde of the systemDead don't take over the Linux kernel itself?
What are your plans for working with or working around systemd? Are there any mainstream GNU/Linux distros that haven't adopted and have no plans of migrating to systemd? Or is migrating to one of the bigger BSD systems the better and more future-proof solution?"
PolicyKit specifically can be compiled to use consolekit instead of systemd for session tracking (this is actually the default, you have to explicitly compile policykit with systemd support).
Unfortunately this is kind of the downside to binary based package management. Either PolicyKit has to be modified to support both as configurable options, probably involving a maze of symlinks and wrapper scripts, or separate policykit-systemd and policykit-consolekit packages have to be provided.
If Debian has decided to to go with systemd, this is probably going to be a common issue on that distro, as when given the option of compiling something with it, they probably will.
Aside from joining us over on the gentoo side (open-rc is life but using something else is easier as it's just a use flag for most packages), or maintaining your own sizable collection of custom-built packages, don't know what to tell you!
I wish people who wanted windows would just stick to windows instead of infecting linux
And now I use NetBSD.
systemd also has its own NetworkManager wanna be in the making as well. I also dislike this.
For shameless plug I currently maintain dhcpcd which does your DHCP, IPv4LL, IPv6RS and DHCPv6. Other nicities like carrier detection, SSID and ARP profiles, routing preferences all come as standard. All in 155k. For kicks there is even a basic GTK+ system tray notification widget that also talks to wpa supplicant to allow wireless network selection and password entry.
It isn't just the boot. Lennart now calls it "Core OS" and he means it. NetworkManager was crap, admit it. After years it still couldn't do everything the software it replaced did but it no longer matters. Latest systemd now even nukes it and replaces it with a all new Core OS replacement that won't work. Which is part of the pattern of destruction that defines Pottering's way of working. PulseAudio is still mostly broken and that was his first project that got any widespread attention. Guy is leaving a trail of destruction wherever he goes and for some strage reason he being allowed to go everywhere.
Wow....someone asks what they can do about having a software package shoved down there throat and your response is just open wide and swallow? I thought this was supposed to be about freedom. Wait, GNU/Linux is about freedom, as long as it's what they want you to do....
On a more serious note, any software that wants UNIX compatibility will keep supporting SystemV/BSD init. I get the distinct feeling that Oracle and especially the BSD guys don't want anything to do with systemD.
I'm starting to think GNU is the problem with "GNU/Linux" these days.
There are significant numbers of people who understand it just perfectly and have valid criticisms that are not bugs.
http://ewontfix.com/14/
The systemd team has pissed of Torvalds:
https://lwn.net/Articles/59368...
Additionally, they repeatedly deny that anyone should have a text log for any reason, dismissing criticisms as 'just hook in syslog *too* as an *optional* thing'. Basically systemd discards decades of sensibilities ecosystem to 'do it better', while throwing out the baby with the bathwater (ditching modularity and portable log data and such).
It's not just that 'if you don't like it, fix it'. People don't like the very fundamental aspects of the design that the systemd did *on purpose*.
XML is like violence. If it doesn't solve the problem, use more.
Your immediate recourse is, indeed, to try and sample the *BSD offerings. Their rc.conf approach I find a lot simpler to deal with than sysv's kludgy linkfarming ever was. It works very well without imposing all sorts of requirements on the rest of the system.
But the problem is political, and so the solution isn't technical. On the political side, I'm highly annoyed by the approach that resulted in this damage, but it's actually endemic in the linux world: Identify problem, then go berserk on the over-re-design-engineering like you're deliberately aiming for a strong case of second-system effect. One (and my pet-) example is the "better replacements" to their broken ifconfig, incompatible with everyone else (and three mutually incompatible attempts down the road there's no end in sight), but there are many more. The latest batch just have taken the previous failures to new heights of technically working incompetence.
What is new-ish is that the damage is spreading, in the sense that by design systemd is linux-only yet now various programs that previously worked on Unix in general are starting to depend on it. Apparently a certain bunch of influential people in the linux-sphere want to become their own vendor-lock-in-enabled bubble, to be the next redmond. This is... not good.
There really is very little recourse other than starting your own lobby war to stop the bunch. Because the problem is mostly politican, the technical side is but a symptom, almost a sideshow.
Without political pressure, soon linux will be akin to macosx, except with poorer code quality and less unified design: Technically some Unix-heritage, in practice it's its own thing, incompatible with the world. So if you'd like a Unix, your route is to *BSD. If not, you can stay and put up with the slowly mounting pile of crap of which systemd is but one thing, if possibly a tipping point-inducing thing. The *BSD people will still have to find some sort of answer, and soon, or they'll have to decide that everything depending on systemd+friends will be a lost cause anyhow and find alternative software with similar functionality, for the current crop no longer works outside of this brave new linux.
Slackware is an alternative mainstream Linux distribution which does not use systemd. Instead of systemd, it uses a combination of custom rc scripts and sysvinit. If Slackware ever adopts systemd as the default system init, they would likely lose most of their user base.
If you really must avoid systemd, then Slackware is probably the way to go. Alternatively, FreeBSD/PC-BSD are prettly much safe from ever getting systemd. For now you could stick with Debian Stable or Ubuntu LTS, both of them will run for years on the older init systems. So, really, you are pretty safe from systemd for at least three to five years, even in the Debian/Ubuntu corner of the Linux ecosystem.
But, really, you might ask yourself why go to all the trouble? Is it a philosophy issue? Is it just hating change? Is there something technical causing problems with your computer that is caused by systemd? A lot of people claim to hate it, but rarely give any practical reasons. Sure, there are plenty of philosophical issues with systemd (and lots of personal issues where its developers are concerned), but take a good long look at why you don't like systemd before you try to avoid it.
Im currently running Gentoo. it offers systemd as a package and ive even run it a few times with success. What it offers, along with uefi, is a chance to drastically speed up the boot process but at a cost to the Linux ethos of 'do one thing and do it well.' Im just as conflicted, and seeing as i work in a RedHat shop i fear ill have to start using it eventually. TFA from sporkbox in the summary highlights the major pain points of systemd quite nicely but the other problem it poses is the homogenization of linux and what that means to numerous Linux community members personally. Linux used to be about choice, but so many distros are systemd/gnome/networkmangler now that its almost horrifying. I get that a unified platform is the key to a 'year of the linux desktop' but the sense of alienation and loss that systemd imparts is very palpable for many of us.
Back on topic though, Gentoos commitment to choice means you can run OpenRC. Its a fine time-tested alternative to SystemDoEverything and while your coworkers might be confused by it, at least you wont have to hack through binlogs for ages to fix a problem in it. You're best not trying to hack out systemd or any of its dependencies in distros like Fedora or Ubuntu as theyre basically so intrinsic to the OS as to render it useless if removed.
Sorry i cant offer more closure for the issue, I hope someone in the thread can though. For me i worry in another ten years ill be deploying machines that are exclusively systemd, quietly muttering the free software lyric, 'You'll be free, hackers, you'll be free.'
Good people go to bed earlier.
And I wish people that want Linux to stay frozen would just stop upgrading or move to a system that sure to stay in 1970.
depinit. written by richard lightman because he too did not trust the overcomplexity of sysv initscripts and wanted parallelism, it was adopted by linux from scratch and seriously considered for adoption in gentoo at the time. richard is extremely reclusive and his web site is now offline: you can get a copy of depinit however using archive.org.
using depinit in 2006 i had a boot to X11 on a 1ghz pentium in 17 seconds, and a shutdown time of under three. depinit has two types of services: one is the "legacy" service (supporting old style /etc/init.d/backgrounddaemon) and the other relied on stdin and stdout redirection. in depinit you can not only chain services together for their dependencies but also chain their *stdin and stout* _and_ stderr together.
that has some very interesting implications. for example: rather than have some stupid system which monitors /var/log/apache2/logfile for security alerts or /var/log/auth.log for sshd attacks, what you do is run sshd or apache2 as a *foreground* service outputting log messages to stderr, chained to a "security analysis" service which then chains to a log file service.
the "security analysis" service could then *immediately* check the output looking for unauthorised logins and *immediately* ban repeat offenders by blocking their IP address, rather than having to either poll the files (with associated delays and/or CPU untilisation) or have some insane complex monitoring of inodes which _still_ has associated delays.
also depinit catches *all* signals - not just a few - and allows services to be activated based on those signals. richard also had a break-in on one system, and they deployed the usual fork-and-continue trick, so he wrote some code which allowed the service-stopping code to up the agressiveness on hunting down and killing child processes. this also turned out to be very useful in cases where services went a bit awry.
basically the list of innovations that richard added to depinit is very very long, in what is actually an extremely small amount of code. i simply haven't the space to list them all, and no, richard was not a fan of network-manager either.
btw you might also want to look at the replacement for /bin/login that richard wrote. it was f****g awesome. basically what he did was use gpg key passphrases as the login credentials.... and ran gpg-agent automatically as part of the *login*. i have never even seen a PAM module which does this trick. it would be awesome to do the same trick for ssh as well.
it's fascinating what someone can get up to when they have the programming skill and the logical reasoning abilities to analyse existing systems that everyone else takes for granted, work out that those sytems are actually not up to scratch and can write their *own* replacements. it's just such a pity that nobody seems to have noticed what he achieved.
I've pinned systemd in apt to -1 (so it won't ever install on my machines). So far i didn't have any problem. Debian will continue to support sysv for years and years, and in that timeframe this silly systemd fad will have passed away, and people eventually regain their minds and (hopefully) balls.
This "inevitability" horse shit is that: horse shit. Linux is equally useful without systemd, provided you have a mininum of experience.
Solaris has it's own abomination called SMF. Good luck debugging network problems on Solaris without a GUI unless you're experienced with SMF. I use Solaris only to maintain my open source projects (I value portable code), and I _hate_ dealing with the system. So convoluted.
The system with the most straightforward configuration and init system, IMO, is OpenBSD. It's soooooooo nice. The only major change in nearly 15 years has been the move to an rc.d/ (init.d-style) startup script directory. Contrast that with number of convoluted changes in Linux administration over those past 15 years, and it seems like a miracle.
If SMF and launchd (OS X) are any indicator, I'm definitely going to hate working with systemd.
(NOTE: I haven't used Slackware since the 1990s, so maybe it's remained stable all these years, too. For Linux I tend to only use Debian and Ubuntu.)
To be fair, LILO is very primitive and sensitive. It doesn't read filesystems; it has an installed map (the result of running lilo) that lists the exact blocks to load for a given entry. You cannot load anything that's not in its map. Touch any of those blocks and it can fall apart. GRUB was a vast improvement, but also adds a great deal of complexity. (GRUB2 even more so.)
That Oracle dislikes something isn't a condemnation. It's more nearly a recommendation.
That said, I'm dubious about systemd. I almost understand how to use init. OTOH, I prefer the interface of the pre-grub2 grub to the current one. I assume that there must be SOME benefits to the change, but I haven't found any. I expect to end up feeling the same way about systemd.
I think we've pushed this "anyone can grow up to be president" thing too far.
Hate to break it to you, but when you install a distribution, you have a lot of software "shoved down your throat." It is what a distribution is, after all, and has been the case since forever. The maintainers decide what functionality is in the base system, what packages are installed in meta packages, what versions, what optional features to compile in. The only way around it is to use a source distribution like gentoo.
Never used SMF, but systemd is quite a bit better than launchd. The configuration files are all plain text. The major difference from a configuration point of view is that instead of writing a script, you just specify executable information, dependencies, sockets, etc, in a config file. That's it. Doesn't seem like such a big deal to me and in many ways seems quite a bit better than sysV.
the systemd init may be brilliant, if it would be isolated. But its mixed up with udev, syslog and even gnome to some extent. This cannot be an good idea, because stuff like init needs to KISS.
http://wizardofbits.tumblr.com...
The configuration system for daemons that systemd has is an enormous leap forward over the old shell scripts. If systemd would stick to be an init system, it would not be such a problem.
When it takes over file system mounting, including hiding most mount points from /etc/fstab and breaking silently if there are perfectly valid mounts in there which it happens not to like, people complain.
When it takes over system logging, previously one of the major advantages of Unix-based systems over Windows, people complain.
And so on and so forth.
Finally! A year of moderation! Ready for 2019?
It's a theme on the Internet. If you don't qualify every minor nuance of your statement, or carve out an exception for every conceivable corner case, someone calls you out on it. Nothing can be left as an excercise for the reader, because too many readers are pedantic or intellectually dishonest.
Usually it's intentional equivocation masked as an attempt to sound intelligent or continue the argument when they no longer have a real point. Often they boil down to syntactic or semantic arguments, belaboring point after point until those with solid points are swarmed by nits. Unfortunately, that makes it very difficult to tell when someone is being obtuse versus when they are being curious or have a legitimate point.
The road to tyranny has always been paved with claims of necessity.
When it takes over file system mounting, including hiding most mount points
I can see how this is annoying, especially when you don't know what is going on, but mounting filesystems is an integral part of the startup process and therefore should be managed (in part) by systemd. It was manged using the old sysV scripts too btw, so it is consistent as far as init systems go. The difference with systemd is it actually knows what a filesystem is and that it is different from a service, so it can manage and monitor them accordingly. What this means is that filesystems associated with booting (root, swap, dev, ...) are now systemd entries instead of /etc/fstab entries. Once you realize this, it is not that hard to manage. And /etc/fstab does still work, of course, for filesystems you want to manage yourself. There are reasons you might want to create systemd entries for those too, though. Automounting, for example, is handled much better with systemd than the old autofs route that we had to use before.
and breaking silently if there are perfectly valid mounts in there which it happens not to like
That is either a bug, or possibly a conflict. Again, I can see how it is annoying, but if you have an /etc/fstab entry that wants to steamroll a systemd entry, it is understandable that systemd will try to stop that from happening. The correct fix in that case would be to edit the systemd entry to match the changes you are trying to make with the /etc/fstab entry.
When it takes over system logging
So, systemd doesn't just start/stop services, it also monitors them and can be configured to take certain actions depending on what happens to a particular process. So, needless to say, logging is kind of inherent to the whole thing. Correct me if I'm wrong, but I don't think it "taking over logging" is your real objection, but rather the way in which it does its logging. Instead of splitting things into separate text files that are managed by their respective daemons, systemd collects all this information and stores it in a standardized, indexed way within its own file. This is a design decision, and with all design decisions there are tradeoffs. In this case you are sacrificing the ability to just cat/grep individual text files for the ability to filter and have other processes able to monitor the log files, as well as some benefits for auditing and security. I definitely prefer text files because I don't manage complex scenarios, but I can also see how journald is critical for certain enterprise infrastructure. If you don't like journald, you can install syslog-ng. You just have to make sure it doesn't trample on journald (ie: it has to listen on the journald socket instead of /dev/log). I believe CentOS 6 is configured this way (journald+syslog-ng), so it is not that unusual or hard to do.
"First, I have to ask, what is wrong with systemd?"
It's a massive, complicated, and very poorly behaved substitute for a simple, robust, and well behaved program. And it's not just a regular program, it is (if used as intended) a critical system component that will take your entire system down when it goes wrong.
If it were just a bad program, that's no big deal in and of itself, there are plenty, you just avoid them. But these people are not hackers, they have a marketing engine and are aggressively attempting to push themselves into a position where it WILL become impossible to avoid them and still use many new programs. That's beyond offensive. That's an attack on the Free Software ecosystem itself.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Friends don't let friends enable ecmascript.
I like how people automatically assume change == good. Maybe I'm getting old, but it seems to be a young person thing (as is the rewrite everything from scratch mentality).
Change is change. It can be good, it can be bad. I'm not an expert on such things, but from everything I've read, the change to systemd is bad. And it seems to be a bad change in much the same ways the examples of change you gave (Metro, Unity, etc.) have turned out to be bad.
The Unix philosophy has always been to do big things by using little pieces. To violate this philosophy is not necessarily bad, but it would seem like trying to fit a round peg into a square hole. Sure, if you hammer it in hard enough, the thing will fit. But your square hole might have trouble fitting square pegs through afterwards, and your wooden board might crack after you fit more things through the hole irrespective of shape.
I'd have used a car analogy, but the best I could come up with is using the wrong kind of motor oil, which when put that way, doesn't seem quite as severe as the systemd problem.
"If a nation expects to be ignorant and free in a state of civilization, it expects what never was and never will be."
Systemd COULD be a good thing if it would stick to starting up the system. It should START udevd, not BE udevd. It should START dbus, not BE dbus. It should be trivially easy to do any of:
Switch from systemd to SysV, switch from SysV to systemd, use systemd but honor the SysV init scripts. With a bit of work fron the systemd folks, it should even be fairly easy to use SysV and have it start systemd to monitor select daemons.
Do that and every single objection would go away immediately.
What's wrong with it? Here's my starting list and I'm sure I'll think of more....
- Binary Logs: Sorry, but there is no advantage to not being able to easily look at a log file.
- Failure to Log to the Console: There is nothing more frustrating than watching 5 screens of "Failed, use journalctl to blah, blah, blah..." come by when you know that your root filesystem isn't mounted read/write. There went *ALL* your debug information.
- Failure to Drop to a Shell When It Breaks: If my boot is broken,I want a shell. Not a hang. There's a way to force it to go to a shell, but that's before it does *anything* so you don't get to debug the failure, you get to guess what the failure might be and see if you can debug *that*.
- No way to see WTF it's doing: There's supposedly a command to make it tell you what order (and presumably what'll happen in parallel) things are going to start in. However, if you use that command as root, it tells you not to run it as root. If you do it as a normal user, it doesn't have permission to read all the files to tell you what it's doing.
- Races: I no longer have any idea what order things are starting in. I've had a cluster where everything worked fine. Until the a week and a few reboots later and then it occasionally failed. Don't even start to tell me that "I must have my dependencies wrong". I *KNOW* they're wrong. But I have no tools to help me figure out what "right" is. Plus, have you looked at how many unit files systemd starts on a normal system? I can't hold that much of a graph in my head. With SysV init, unless I turned on some weird parallel mode, everything starts in the same order every time.
- Complexity: I'm not a professional sysadmin. I'm a developer who has to maintain development systems (as well as personal systems) part time. If I worked with systemd every day, I'd probably be able to figure out ways to make it work for me. But I don't. SysV is just shell scripts. I *DO* deal with *those* every day so it's pretty easy to debug.
- Complexity, Part 2: The previous version of init essentially had no bugs. Ok, I'm sure that's not really true but they sure didn't surface very often. Since the results of your Process #1 dumping core are catastrophic (ie, a kernel panic), ideally that process should do as little as possible. That is *CLEARLY* not the design philosophy of systemd. Further, it consumes a decent amount of RAM and the more RAM you consume, the more likely (statistically) you are of hitting a memory error.
- YACL (Yet Another Config Language): Ok, so this is really a minor complaint but I get to learn yet another way of writing config files.
- Filesystems: SysV init tended to mount local filesystems *very* early in boot (some of that broke when udev got involved, but you could usually hack around that) and network filesystems not long after. I'm not entirely sure where systemd mounts filesystems, but it breaks *HORRIBLY* if you move some of the files needed by a service onto a filesystem that's not a "normal" filesystem. I'm sure there's some way to set all the dependencies to make that work, too, but see above, I have no f'ing way of figuring out what should depend on what.
From all outward appearances, the developers have *no* interest in fixing much of any of those complaints. The whole "debug on the kernel command line" fiasco is a pretty clear indication that they "don't play well with others". In the end, I'll see what Slackware has or maybe move (back) to the BSDs.
Go Badgers! -- #include "std/disclaimer.h"
> How long until all of the software packages that BSD wants to use require so much work to retrofit to use a different init mechanism that they just throw in the towl and accept defeat?
Keep in mind that *BSD is not alone. There are other GNU/Linux distributions that avoid it. Gentoo are among the distributions working on things like eudev (so you can keep on using udev without systemd).
9/11: Never forget it was a false-flag operation
The problem is: A *tiny* init process won't be able to offer the *exactly same* functionality. The functionality has to come from somewhere, it does not fall from the sky: Some code needs to implement it.
If you want to keep PID 1 tiny then you can implement the actual functionality in separate processes. You now have two or more process and now you need code in the tiny init process that makes sure the controlling processes are getting started (and restarted). Remember: Those daemons provide the actual functionality, so PID1 can not depend on that to start those daemons in the first place.
You need code that facilitates a communication channel between the processes. You need code to lock out processes that are not meant to talk to your tiny init process. You need a protocol that the init process speaks and that allows it to be remote-controlled. All of sudden that tiny process is no longer tiny and your architecture is much more complex than it would be otherwise.
That complexity requires you to add more code to mitigate communication failures, to synchronize data structures between all the different processes that need access to them, you need to be careful not to introduce race conditions between those processes. In the end you end up with a pretty big init process and a bunch of big and nearly equally critical daemons surrounding it. I do understand where the systemd guys come from: Keep the architecture simple, and put absolute minimum amount of code into PID1 to provide the functionality they want. That makes the overall system less complex and easier to reason about, which is good for security and robustness.
Read the code, it is actually pretty ok.
Regards, Tobias
Do one thing well. Build more complex actions by putting smaller parts together. Swiss army knife system utilities need not apply. That is the Unix way.
Mount -a is the perfect way to mount filesystems needed to init the machine. The rest can be mounted by a daemon as they become available. My / filesystem resides on an HDD that is bolted in to the system, if it's not there, there is nothing to boot, so why does it need to be 'monitored'?