Five Years of OS Updates: An Engineer on What Actually Changed
Archi3 min read
Contents
Five Years of OS Updates: An Engineer on What Actually Changed#
We spoke with a platform engineer who manages about 9,000 endpoints across three operating systems, and who has kept the same patch dashboard running since 2021. The conversation has been edited for length and clarity.
The update cadence#
Q: What is the biggest difference between patching machines in 2021 and now?
A: The size of the update. Five years ago a monthly rollup on a typical fleet was 400 to 800 MB. Now most of ours land between 90 and 250 MB because more of the change ships as component-level deltas. That sounds like a cosmetic win. It is not. It changed how we schedule. We used to block a two-hour window per site. We now run background installs during the workday and rarely hear about it.
Q: Did reboots get better?
A: Somewhat. Restart times on our standard laptop image went from roughly 90 seconds in 2021 to about 35 seconds, mostly because services start in parallel and the update step no longer rewrites the whole system partition. The catch: more updates now require a full restart instead of a logoff. We traded time per reboot for a higher number of reboots.
Who controls the machine#
Q: What about the security side?
A: The threat model shifted more than the tools. In 2021, most of what we defended against arrived as an email attachment or a malicious download. Now the interesting incidents start with a signed driver, a browser extension, or a package pulled from a public registry. The operating system vendors responded by moving trust decisions down a layer. Driver signing was already a thing, but the enforcement is much stricter, and in some configurations you cannot load an unsigned kernel module at all.
That is good for a managed fleet. It is a headache for anyone running older lab hardware or a custom peripheral.
The hardware bill nobody mentions#
Q: Is there a cost people overlook?
A: Memory. The baseline went from 4 GB to 8 GB on desktops, and 16 GB is now what we buy. Part of that is the OS. Part of it is that the browser and the chat client each sit at 1.5 GB. Our replacement cycle used to be five years. It is closer to four now if the machine has to run the current release with reasonable headroom.
Q: How do you decide when to move a fleet to the newest version?
A: We wait for the first point release and then pilot 200 machines for three weeks. In 2021 we piloted for six weeks. The shorter window is not confidence. It is that the vendor now ships fixes on a faster cadence, so a bad build gets corrected before it reaches most of the fleet. We accept a little more churn in exchange for less time on an old branch.
The parts that got worse#
Q: Anything that regressed?
A: Settings sprawl. Every release adds a privacy panel, a permissions page, and a new notification category. An ordinary user cannot tell which toggle controls what. We now publish a one-page internal guide, updated twice a year, that tells staff the six settings that actually matter for us. Five years ago that guide was two paragraphs.
Q: And the parts that got better?
A: Recovery. A failed update in 2021 often meant a reimage. Today most failures roll back automatically after one or two restart attempts. On our numbers, machines that needed hands-on repair after a bad patch fell from about 1 in 60 to roughly 1 in 400.
Advice for anyone managing machines#
Q: What would you tell a smaller team setting up now?
A: Measure your restart times before you promise a maintenance window. Test the update on the oldest hardware you still support, not the newest. Keep one machine off the automatic channel so you have something to compare against. And read the release notes for driver and permission changes, because that is where the surprises are.
Q: One sentence on the last five years?
A: The updates got smaller and the trust model got tighter, and the work moved from scheduling downtime to managing settings.
More articles