This is truly unhinged. I wonder if running an installer under wine in win95 mode will do this.
Granted, but at the same time it's also resolutely pragmatic.
Apparently there was already lots of software out there which expected to be able to write new versions of system components. As well as buggy software that incidentally expected to be able to write old versions, because its developers ignored Microsoft's published best practices (not to mention common sense) and and didn't bother to do a version comparison first.
The choice was to break the old software, or let it think it succeeded then clean up after the mess it made. I'd bet they considered other alternatives (e.g. sandbox each piece of software with its own set of system libraries, or intercept and override DLL calls thus ignoring written files altogether) but those introduce more complexity and redirection with arguably little benefit. (I do wonder if the cleanup still happens if something like an unexpected reboot or power loss happens at exactly the wrong time).
Could the OS have been architected in a more robust fashion from the get-go? Of course.
Could they have simply forbidden software from downgrading system components? Sure, but it'd break installers and degrade the user experience.
Since the OS historically tolerated the broken behavior, they were kind of stuck continuing to tolerate it. One thing I learned leading groups of people is if you make a rule but don't enforce it, then it isn't much of a rule (at least not one you can rely on).
I would argue the deeper mistake was not providing more suitable tooling for developers to ensure the presence of compatible versions of shared libraries. This requires a bit of game theory up front; you want to always make the incorrect path frictiony and the correct one seamless.
Why did programs overwrite system components? Because Microsoft regularly pushed updates with VC++ or Visual Studio and if you built your program with Microsoft's tools you often had to distribute the updated components for your program to work - especially the Visual C runtime and the Common Controls. This even started in the Win3.11 days when you had to update common controls to get the fancy new "3d" look. And sometimes a newer update broke older programs so installers would try to force the "correct" version to be installed... but there's no better option here. Don't do that and the program the user just installed is busted. Do it and you break something else. There was no auto-update or internet access so you had to make a guess at what the best option was and hope. Mix in general lack of knowledge, no forums or Stack Overflow to ask for help, and general incompetence and you end up with a lot of badly made installers doing absolute nonsense.
Why force everyone to share everything? Early on primarily for disk space and memory reasons. Early PCs could barely run a GUI so few hundred kilobytes to let programs have their own copy of common controls was a non-starter. There was no such thing as "just wait for everyone to upgrade" or "wait for WindowsUpdate to roll this feature out to everyone". By the early 2000s the biggest reason was because we hadn't realized that sharing is great in theory but often terrible in practice and a system to manage who gets what version of each library is critical. And we also later had the disk space and RAM to allow it.
But the biggest issue was probably Microsoft's refusal to provide a system installer. Later I assume antitrust concerns prevented them from doing more in this area. Installers did whatever because there were a bunch of little companies making installers and every developer just picked one and built all their packages with it. Often not updating their installer for years (possibly because it cost a lot of money).
Note: When I say "we" here that's doing a lot of heavy lifting. I think the Unix world understood the need for package managers and control of library versions earlier but even then the list of problems and the solutions to them in these areas varied a lot. Dependency management was far from a solved problem.
This is bog-standard boring stuff (when presented with a similar problem, Linux invented containers lol) - read some of his other posts to realize the extent Microsoft went to maintain backwards compatibility - some are insane, some no doubt led to security issues, but you have to respect the drive.
Containers are in fact redirecting writes so an installer script could not replace system libraries.
The equivalent would be a Linux distro having the assumption that installer scripts will overwrite /usr/lib/libopenssl.so.1 with its own version and just keeping a backup somewhere and copying it back after the script executes.
No OS that I know of does that because it’s unhinged and well on Linux it would probably break the system due to ABI compatibility.
If they had taken essentially the same approach as wine and functionally created a WINEPREFIX per application then it would not be unhinged.
edit: also to be clear, I respect their commitment to backwards compatibility which is what leads to these unhinged decisions. I thoroughly enjoy Raymond Chen’s dev blog because of how unhinged early windows was.
Windows may have suffered its share of bad architectural decisions, but unhinged is a word that I wouldn't apply to their work on Windows.
Just because I am saying it’s unhinged doesn’t mean I don’t think it’s cool
I’ve never read any windows source so I can still contribute to wine but I’ve read the NT kernel is really high quality
You say, oh, obviously you just should redirect writes to a shadow layer or something (and later Windows can do that), but at the time they faced the rather large problem that there is no formal concept of an installer or package in Windows. An installer is just an ordinary program and the OS has no app identity available. So, how do you know when to activate this redirection, and what is the key identifying the layer to which redirects happen, and how do you handle the case where some writes are upgrades and others are downgrades, etc, and how do you do all that in a short amount of time when shipping (meant literally in those days) will start in just a few months?
>there is no formal concept of an installer or package in Windows.
this one is on them, I think package managers already existed - doesn’t seem like there was ever a blocker for windows to have a package manager but Microsoft never bothered until very recently
As TFA says:
You even had installers that took even more extreme measures and said, “Okay, fine, I can’t overwrite the file, so I’m going to reboot the system and then overwrite the file from a batch file, see if you can stop me.”
Man, wouldn't it have been nice if everyone had enough hard drive space in those days in order to do something like that...
If an installer expects to be able to overwrite a file and fails to do so, it might crash, leaving the user with a borked installation.
Of course you can blame the installer, but resolution of the problem might take a long time, or might never happen, depending on the willingness of the vendor to fix it.
This as well. I know there are a million ways for a malicious installer to brick Win95, but a particularly funny one is hijacking the OS to perpetually rewrite its own system components back to compromised version number ∞ whenever another installer tries to clean things up.