The most successful operating system ever written is the one that kept the fewest of the rules it is famous for keeping. Both halves of that sentence are true, and the whole of this piece is an attempt to hold them at once.
For three Mondays this family has been drawn by a single measure: how faithfully a system keeps the Unix discipline, whatever its parentage. The certified members hold the paper; the BSDs keep the line without it; QNX, a fortnight ago, kept the discipline with none of the blood at all. This Monday closes the family on the member hardest to place, the one largest by a distance and least disciplined by the same distance, and both of those things at once.
Picture the family as a stand of trees grown from one seed, and off to the side a single tree, larger than all of them together, grown from a seed of its own. It is not a branch of the others. It was written from nothing in 1991 to look and feel like them, and it succeeded at the looking and feeling so thoroughly that most of the computing world now takes it for the original. It is Unix-like by resemblance, not by descent, and over thirty years the resemblance has drifted.
The Success, Plainly
Before the drift, the success, because the success is not in doubt and pretending otherwise would be a poor start. Linux runs every one of the five hundred fastest supercomputers on earth, and has done since 2017, when the last two holdouts fell off the list. It is the kernel inside Android, which is to say inside most of the telephones on the planet. It is the near-universal default of the cloud, the thing a workload lands on whether anyone asked for it or not, and it is in the cars, the routers, the televisions and the light switches besides. No operating system has ever been deployed on this scale, and the achievement is entirely real.
Where It Drifted, and Why
The discipline it drifted from is the one the earlier Mondays kept describing: small pieces, each doing one thing, composed through plain interfaces, in a system one team can hold whole. Linux keeps the surface of that idea and has, over three decades, quietly relaxed most of the interior. Five departures, each with a reason that was sound at the time.
The kernel is the first. Unix orthodoxy, by the 1980s, leaned toward keeping the kernel small and pushing what it could into user space; QNX, a fortnight ago, showed how far that can be taken. Linux went the other way, a large monolithic kernel with drivers built in or loaded as modules, and more recently with eBPF, which lets verified programs run inside the kernel itself. The reason was never ideology. A monolithic kernel is quicker on the hardware of the day and far easier to make support an enormous range of devices, and supporting an enormous range of devices is exactly how Linux came to be everywhere.
Then there is systemd, the departure people argue about at parties, if they go to the right sort of party. The old Unix init did one small job, starting processes in order at boot. systemd, first released by Lennart Poettering in 2010 and adopted as Debian's default in 2014, starts the processes and also keeps the logs, manages the network, resolves the names, tracks the logins and mounts the disks, one integrated complex where there had been a dozen small tools. It is the most deliberate break with doing one thing anywhere in the system, and it was not made from spite: a laptop that boots in two seconds and a fleet that reports its state in a single format are things people genuinely wanted, and the small tools were not supplying them.
The third drift the second Monday set up in advance, and it is the one an administrator feels on a Tuesday. The BSDs are four projects, but each is a single tree, a base system written and released as one whole. Linux is a kernel, and only a kernel; everything around it is assembled by a distribution, and the distributions do not agree, nor are they always on speaking terms with their own past selves. Debian, Red Hat, Arch and Alpine share the kernel and remarkably little else, not the package manager, not the init, not the default shell, not even the C library, since Alpine swapped the usual one for a leaner musl. There is a genuine freedom in this, and it would be less than honest to pretend otherwise: the same looseness that lets Alpine fold down to five megabytes for a container lets Debian sit unmoved for a decade, each distribution free to fit a niche no single vendor would have troubled to serve. The freedom keeps a bill, and the administrator settles it. Tools get swapped release to release, ifconfig for ip, one init for another, and the same command can quietly do two different things on two machines that both answer to the name Linux. A fix to anything shared, meanwhile, has to travel through every distribution's packagers, each on their own schedule, each rebuilding and re-testing, so the patch one team wrote in an afternoon reaches the fleet some weeks later, assuming the backporting queue has not lost interest along the way. Learn Linux, it turns out, means learn one of several families that happen to share an engine; patch Linux means wait your turn behind a dozen teams that have never met.
The licence is the fourth, and it shaped the other three more than it appears to. Linux is under the GPL, version 2 and no later, a copyleft licence obliging anyone who ships a modified kernel to publish their changes; Torvalds declined to move to version 3 in 2007, and the kernel has stayed on 2 since. Copyleft made it rational for rival companies to pool their work in one tree rather than hoard private forks, and so most of the kernel is now written by engineers paid by companies to write it. That is a genuine strength, and it is also a drift: the thing is tended less like a craft tradition and more like a shared industrial asset, which is a different sort of institution from the one that produced the discipline.
The last drift is the world it now optimises for, and it is where the family tells on itself most plainly. FreeBSD had been running containers since 2000: jails, an isolation built whole into the base system, in production for thirteen years before Linux had a settled answer of its own. That answer was Docker, in 2013, which assembled one out of separate kernel primitives, control groups and namespaces begun at Google in 2006 and merged in 2008, and made them usable; Kubernetes, released by Google in 2014 out of its Borg experience, made that the ordinary way large fleets are run. Solaris had shipped zones in the same integrated spirit in 2005. Docker's real contribution was ergonomics and reach, since the isolation itself had been production kit on FreeBSD since the turn of the century. And it bent Linux toward packing a data centre, a question Unix never asked.
Not a Verdict
None of that is a charge sheet, and it would be a small dishonesty to let it read as one. Every drift here is an adaptation to a constraint the original Unix never met, made by people solving the problem in front of them, and the evidence that they solved it is the scale two sections above. Drift is not decay. A system that stayed perfectly disciplined and ran on nothing would have kept the paper and mislaid the point.
The fair way to hold it is the two-axis way this family has used throughout. Membership is one axis, the code and the certificate; discipline is another, the ongoing practice. Linux sits low on the first, being a resemblance rather than a descendant, and it has loosened the second over time. FreeBSD, the tree these Mondays keep returning to, sits differently: one base system, a narrower range of hardware, a permissive licence, a slower and more coherent centre. That is not a scoreboard with a winner on it. It is two answers to two different questions, and the enormous one is enormous precisely because it answered the question the market was actually asking.
The Point
Four Mondays, four ways of belonging to a family that turns out not to be a single thing. The certified members showed that the paper is a niche pass and not a bloodline. The BSDs showed that a lineage can fork and stay whole. QNX showed that the discipline travels without the descent. And Linux shows the last and least comfortable thing: that success and discipline were never plotted on the same axis, and that a system can drift a very long way from the one while running clean away with the other.
Which is why the family marker standing over this piece is a quiet joke at its own expense. The most Unix-like thing in the world, measured by reach, is the least Unix-like thing in the family, measured by discipline. Both are true. Hold them together, and the family finally makes sense.
Success and discipline were never the same axis. A system can drift a very long way from the one while running clean away with the other.