That is a case I only become aware of when I read blogs like this. Technically I could solve it the same way, but these days you have so many tasks on your desk, you don't think about the problem and implications at all and that awareness/discipline is drowned in the noise/unlearned over time.
If someone only gave me 2 minutes for this, because they think it is very simple (as always), I'd have done a count of files of a specific pattern in the directory and then picked a random index, very naive and quick and dirty programming, no sampling at all, just to avoid discussions why it takes so long with people who don't want to hear it.
This reminds me of when I did a lot of C#, Java, JS, Python in my life, filling maps of strings and objects until I started with zig and noticed how expensive and complicated strings and data structure allocations can be. It kind of blew my mind how much memory and computation we waste when we try to get stuff done as fast as possible because of budget/time constraints.
Two things that have helped me with quickly recognizing these situations:
Programming for MCUs. Less so today when multi-hundred MHz MCUs are cheap, but even several years ago there were a lot of products where you needed to use the cheapest MCU and everything it did had to be optimized to avoid stalls and edge cases.
Second is doing LeetCode problems for fun/practice. This will elicit a groan from a lot of people, but the algorithms and pathological edge cases you learn really do change your thinking. The most interesting ones are the hard problems where they’ve added some hidden test case that causes naive solutions and algorithms to blow up. You start thinking on high alert for edge cases and Big O problems. It’s more fun when you’re doing it to learn on your own than for forced interview prep.
I'd bet this bug also shows up in Vista file explorer with auto arrange disabled. Nobody will ever need more than 100 icons on their 800x600 workstation. Ship it.
How many user pictures are there to choose from, though? This is a nice, elegant way to sample from the set, but the nature, size, and frequency of the problem don't justify more than five or ten seconds' or thought and one or two minutes of coding.
Exactly this, if you count the amount of time the windows 11 context menu needs to pop up and then the second click to get to the old context menu across the entire globe for a month, you would get an insane amount of time and cycles wasted
Man, every post from Raymond Chen regarding Windows internals is like a little Xmas. I wonder whether he has to ask someone for permission before publishing this knowledge, though.
It's been many years since I talked to him, but he's an extremely by-the-book guy, so I'm sure he has a process for clearing everything he writes. His reputation does probably provide him a lot of leeway, though. He also has a queue of oldnewthing articles about a mile long and we could probably get years' worth of articles after his death if they let them run.
Interestingly, when reading Raymond Chen's article I thought "reservoir sampling would compare the random number (between 1 and n) to 1, not to n, because that extends more easily to picking more than one element" - and that's what the actual Windows code uses.
yes, you would have to allocate space for up to 100 file paths, but the article says
> it’s more efficient because it reduces the amount of calls into the file system, which is where the bottleneck is
Upon reading it a few times I think the article is alluding to a crappy two-pass solution where you don't store a filename but instead an index into a directory, which is flawed anyway due to being racy.
>Raymond has been involved in the evolution of Windows for more than 30 years. He occasionally appears on the Windows Dev Docs Twitter account to tell stories which convey no useful information.
> it’s more efficient because it reduces the amount of calls into the file system
OK, but isn't the kernel keeping the directory listing in the disk cache? Won't that prevent extra physical I/O if you do just read the directory twice?
If so, then in the second pass, it's all cache hits, and you're just paying the cost of calling into the file system. Hopefully that's pretty fast. But even if not, it's still absolutely dwarfed by the physical I/O required for the first pass. Windows XP era storage was spinning hard drives, not flash.
And if not, then I'm probably going to put my user icon coding task on the back burner and go ask the kernel team why a seemingly very common usage pattern isn't optimized.
(I realize he's not claiming the performance benefit was significant. I'm just trying to see it in the right perspective.)
This is a fun example of the cognitive switch you have to employ when first starting to program a computer. It's extremely easy for a human to pick at random one thing from a pile of things: you reach out your hand and grab it, maybe swirling them around on the table first to shuffle the order. For a computer, there's no direct analogy to that. They just can't do it. And the human process is nothing even slightly like the one the computer follows: we don't have to count the sets and iterate over them, or count the items and then generate a random number to pick the nth item, or risk picking a null item.
I think this article highlights more the importance of understanding system limits in the 1990s versus today. No-one would care to much today if the code review for this feature had “files.count()” or whatever in it, but in the mid 90s that would have been a huge performance red flag because a user would literally hear their hard drive clicking away and see the blinkenlights.
The problem isn’t counting the files (the algorithm in the article also counts the files), but that if you determine that you want to use the ith file only after counting all files, you have to iterate over the whole directory again (or over expected half of it) to find that file.
The mechanism is interesting, but I'm not fully understanding the importance. We say it was done this way because a user would appreciate the speedup. The difference is one traversal versus expected one and one-half traversals. How slow was this traversal at the time for this difference to be significant?
Apart from the performance issues, it also prevents a bug. If some files get deleted between the first and the second pass, then you end up with unexpected behaviour and probably a crash.
It will depend on details like file system fragmentation (Windows XP could run on FAT32), but it could conceivably make a perceptible difference on a slow HDD when there are many pictures in the directory. You also have to check more error cases, for when the second iteration fails for some reason. The mindset was probably "why complicate the code with multiple iterations and make it less efficient?" when the efficient solution is straightforward and arguably simpler.
That's why now we need gigabytes of ram to do the same thing we managed in 128Mb back then. I mean sure, there's an AI chatbot in your PDF viewer and ads in your start menu, but the functionality of many daily things is still the same.
There's a better way to do this, you use inverse CDF to avoid all the RNG calls.
Generate a random number, then skip items until you reach that number:
selectRandomFromIteratorOptimized(iterator)
{
if (!iterator.moveNext()) {
return null;
}
var winner = iterator.current();
var count = 1;
while (true) {
var u = random_float_open(0.0, 1.0);
var skip = (int)Math.Floor(Math.Log(u) / Math.Log(1.0 - (1.0 / (count + 1))));
for (var i = 0; i < skip; ++i) {
if (!iterator.moveNext()) {
return winner;
}
++count;
}
if (!iterator.moveNext()) {
return winner;
}
++count;
winner = iterator.current();
}
}
Back in the day, Windows OS kernel programming avoided use of floating point numbers - certainly transcendental functions would've been frowned upon - when CPUs didn't include an FPU.
I don't know if they've relaxed this since the days of non-FPU CPUs - anyone know? If they let the Weather app use a webview, there must be some floating point usage in there.
This code is at a much higher level though - at the user shell level, explorer.exe.
Anyways, asking Google's AI to remove the above code's use of floating point results in code resembling the original version.
Either that call would have to do the same (i.e., walking the files and counting), or you'd need some additional metadata in the directory entry to store how many files there are, requiring additional storage accesses for adding and removing files. Adding to that that both FAT32 and NTFS are quite old and had to run on older hardware. Cycles and disk accesses are not free.
On top of that, how often is it necessary to efficiently know the number of files in a directory while at the same time not caring about the files enough to list or display them? This algorithm is a special case where you could use the count of using a bit simpler code that ultimately would have the same file system API calls (since you cannot tell the FS to give you file #37 from that directory, so you'd have to use FindNextFile 37 times anyway, just like the sampling algorithm).
> or you'd need some additional metadata in the directory entry to store how many files there are, requiring additional storage accesses for adding and removing files.
Most unix filesystems use inodes. inodes have the same format whether they represent a file or directory. So directory have a (usually unused) size member. NTFS doesn't use inodes, but the records in the MFT work the same way.
When adding or removing a file from a directory, you have to update the modification time of the folder, so you have to rewrite the entire inode anyway. Updating the size/file count at the same time would be free.
In my opinion the likely reason why file count isn't tracked is a lot more pedestrian: It wasn't tracked initially and we can never add it to existing file systems because the metadata would get out of sync if the FS was mounted on a kernel with no count support.
I base this assumption on the fact that many modern file systems do indeed keep track of the count.
Not that it counters any point you're making, but ZFS displays the number of contained entries of a directory in the directory's size field; mind that . and .. are included, so you usually need to subtract 2 to get the count you actually want. I do find it useful sometimes to know the count without getting the listing; the former is a very inexpensive operation (since ZFS is keeping track of metadata like you suggested), the latter is expensive, potentially extremely with hundreds of thousands or more of entries.
This is more-or-less unique to ZFS. Other file systems even on Linux and FreeBSD generally don't provide this behavior.
That's a matter of taste, but for many it was the Windows they used first and/or spend the most time on and there's a lot of love for that reason alone. I never used XP all that much, but I always changed the theme to the Windows 2000 look. I really didn't like the default, it looked unprofessional and clunky in my eyes. Upon release XP was also pretty universally mocked as a Fisher Price-like UI.
To me Windows 95/NT 4 is still the gold standard in Windows UI. It's certainly not the prettiest, that would be Windows 2000, but it was easy to use, easy to navigate and efficient with space.
Of course that's a matter of taste but Luna interface/theme made whole XP being more appealing to ordinary people - especially with these task-oriented elements. On the other hand, "classic" widgets had that strong bare-bone technical and functional look that made whole system scarier to some degree.
I had tons of "Visual Styles" back then and in the last XP days I opted for grayish Royal or Royale. It's such shame that MS has abandoned Watercolor theme - that was a middle ground: interface was updated and yet, still similar to classic design. And it was even in some elements flat before that style become a dominant. Luckily we're slowly moving away from that and I won't be missing it.
win2k was peak for me. lean, functional, just a pinch of glitter here and there (short fade-ins). it was on par with the amazing stability brought by nt5 kernel.
i kinda miss xp at a cultural level since it was a bit the end of that computing culture cycle (after that apple started to dominate and ubuiquitous computing influenced desktop ui)
When did Apple dominate? I'm not trying to hate on Apple here, but there's this weird belief that Apple have had a lead in personal computer OS market share at some point in the last few decades, and it isn't really true since about 1984.
You can argue that they should dominate, but that doesn't seem to have happened.
I remember Apple products being the realm of graphics designers, videographers, and other art professionals back in the XP days. I thought they went mainstream later, as Windows Vista coincided with the iPhone release.
Took me years to like Windows XP. Hated the new start menu. I always kept installing Windows 2000 (or ME, probably because it could use Windows 98 drivers but look like Windows 2000).
I eventually got used to it and ended up liking it.
I always ran it with the 2000 style theme. That said XP offered me nothing over 2000 so 2000 is what I ran on my main machine since it used less ram but did all the same things, often a bit faster.
For awhile I ran XP 64-bit though, that did do one thing 2000 couldn't do.
It was at this moment I realized that people are talking about liking the non-classic-look XP. I always immediately put things in classic look, no grouped windows, etc.
The first Windows I used was 98, and I sure didn't miss the blocky grey menus nor the pixelated icons. Maybe it was Fisher Price-ish but the new rounded buttons and borders were way more appealing to me.
The "Windows Classic" look on XP is exceptionally ugly, because those 3D-rendered icons just do not belong on a Windows 2000-style grey background for menus or buttons. (I have recently had to test both these OSes in VMs…)
That is true, it looked completely off, as if no one really bothered to tests it and make it look nice. It got rid of the horrible window decoration of the default theme, but the start menu looked terrible.
>During the Luna studies, that people’s reaction to Luna was often, “Wow this would be a great UI for X,” where X was “my dad” or “my employees” or “my daughter”. People didn’t look at it as the UI for themselves; rather, they thought it was a great UI for somebody else.
(Luna is the name of the default look of Windows XP)
That’s Vista. And there’s no contest whatsoever. They’ve never been able to again achieve the perfect balance of elegance and functionality that Vista’s UI had.
What is amazing is the amount of consideration given to an issue which would escape the majority of users. It is surprising how complex an apparently easy process turns out when considering certain special cases.
But the naive way of doing this also wouldn't really require two passes, right? It would just require more memory because you would first save all file names in an array (stopping at 100), then pick a random one in constant time.
How do you know how big your array has to be in a single pass? I don't think the WinXP source uses vectors or similarly ergonomic auto-growing arrays. You could preallocate an array big enough for 100 paths of length MAX_PATH, but that's a bit wasteful. And it doesn't sound like you'd actually end up with fewer lines of code (in that flavor of C++, in python it would be different)
Practically speaking, I might just allocate an array of 100 pointers. That's only 400 bytes. Then as you encounter each filename, allocate just enough memory for the actual length of the string (plus null terminator) and store the pointer in the array.
Yes, you could allocate it on the stack. I think back then (still?) a filename could be at most 260 characters, each encoded with 16 bits, so about 52k of stack allocation.
Because it would always return the first match in that case.
You still need to see all of the items once.
Imagine you have 2 items.
First one has 100% chance to be selected. So it does. Then the second has 50% chance to be selected. If it isn’t you effectively chosen the first one and have 50/50 chance to return either.
Now you add a third item. There is 50/50 chance of having either selected. And 1/3 chance of replacing the selection with the new one. Resulting in a 1/3 chance of selecting any of the three. (Because 1/2-1/6 = 1/3) 1/6 because there is 50% chance you will “steal” the selection.
Thank you for writing this out, I didn't quite get what was going on at first. But then, to formalize the recursion from your example: let's assume we're at item n in the iterator, and at that point we've selected a winner from the previous n-1 items with equal probability, i.e. each item had a 1/(n-1) chance of being selected. The probability that item n will override it is 1/n. The probability that the old winner will remain selected is thus (n-1)/n. That means that the old winner remains selected with probability 1/(n-1) * (n-1)/n, which cancels out to 1/n, so each item is indeed selected with equal probability in the end.
If you are at picture 1, you have 100% chance of selecting it as the current winner.
If you are at picture 2, you have 1/2 chance of selecting it as the current winner, or 1/2 chance of keeping the previous fairly selected winner.
At picture 3, 1/3 chance of picking it, or 2/3 chance of retaining the previous fairly-selected winner. There are two of them, so 1/3 chance of each.
At picture n, you have a 1/n chance of picking it, or an (n-1)/n chance of retaining the previous fairly-selected winner. There are n-1 previous pictures, so all of them have had 1/n chance of being picked.
At every single step, there is the invariant of all pictures being considered that far having had an equal chance of being selected, and the next step always retains the invariant.
Oh, I understand, should've examined more carefully, the count starts at 0 and increments, so random is not from the total but from the elements counted so far.
No there is not. First element is defacto winner, but you still have to loop through the rest with 1/n chance of being selected to fully give each element a chance of winner selection
The len(pics) can be O(n), especially if iterators are used like here. Also, an O(1) lookup would require a previous O(n) pass over the data anyway.
The picture selection algorithm's kind of single-pass iterator usage might have been more performant back in the XP days, as it avoids possibly expensive operations.
Modern CPU/other optimizations might make a multi-pass approach more performant due to better memory locality or other factors.
On count == 1, the winner gets set to the first element, true. But the function does not return yet! So the value might get overwritten during the remainder of the for-loop.
I don't understand Microsoft. These guys solve the most mundane problems with most elegant solutions and with sound edge-case handling scenarios, then they destroy all the effort with subpar programming where it matters and with user hostile behavior where they can't botch it.
That's simple. The Windows Kernel is a collection of mostly elegant solutions, with a strong peppering of backwards-compatibility cruft, all the way from NT3 to Windows 11. The Windows userland received a lot of effort until about Windows XP, and since then is a collection of subpar programming, half-finished projects and user-hostile patterns (with some notable exceptions)
Different teams with different goals and different management
> Different teams with different goals and different management
Not to mention the stack ranking policies that caused engineers to sabotage other people's work and avoid collaboration in order to keep themselves high on the stack in order to avoid being fired.
We had bigger problems back then, and the function ran considerably rarely when compared the other parts of the OS, so it was a valid assumption at that age.
However, I still remember Wine laughing at Windows for WMF exploit and end up being affected from the same exploit. Now, that was a good laugh.
Early in my career, I have read a blog post from someone senior about boring/entertaining work. In short it stated to following:
Even if you work on exciting things, the project will always have boring parts.
You'll need to develop these parts as well, so be prepared and show the same respect to all parts.
So, I'd personally rather work on all parts equally well rather than winging boring parts and create something subpar.
There's no way my memory of this is reliable anymore, but I also remember my administrator account being chess pieces (and my user account being an orange fish).
On initial install, sure. But user accounts can also be created at arbitrary times. The user may have changed the set of photos in the intervening time and might even be editing the directory during profile creation.
I thought the same thing, and then wondered if this must be an enterprise problem, because what home user adds loads of icons - so many that IO becomes a concern - and also wants them to be randomly assigned to other family members?
But then it occurred to me that XP was of an era when you would absolutely spend an afternoon downloading some megapack of 10000 cool user icons.
> As a final safety check, the code stops after sampling 100 pictures. This avoids pathological behavior if somebody puts a million files in the Default Pictures directory
If someone only gave me 2 minutes for this, because they think it is very simple (as always), I'd have done a count of files of a specific pattern in the directory and then picked a random index, very naive and quick and dirty programming, no sampling at all, just to avoid discussions why it takes so long with people who don't want to hear it.
This reminds me of when I did a lot of C#, Java, JS, Python in my life, filling maps of strings and objects until I started with zig and noticed how expensive and complicated strings and data structure allocations can be. It kind of blew my mind how much memory and computation we waste when we try to get stuff done as fast as possible because of budget/time constraints.
Programming for MCUs. Less so today when multi-hundred MHz MCUs are cheap, but even several years ago there were a lot of products where you needed to use the cheapest MCU and everything it did had to be optimized to avoid stalls and edge cases.
Second is doing LeetCode problems for fun/practice. This will elicit a groan from a lot of people, but the algorithms and pathological edge cases you learn really do change your thinking. The most interesting ones are the hard problems where they’ve added some hidden test case that causes naive solutions and algorithms to blow up. You start thinking on high alert for edge cases and Big O problems. It’s more fun when you’re doing it to learn on your own than for forced interview prep.
Its not like windows is the pinnacle of software craftsmanship.
Being tenured and one of the top engineers in your company with a very proven track record gives you quite a bit more freedom.
If it were the Windows 11 source, it'd get nuked immediately
Most modern Windows code was written in 1995. Don't assume for one moment that it isn't in production Win11 today.
Interestingly, when reading Raymond Chen's article I thought "reservoir sampling would compare the random number (between 1 and n) to 1, not to n, because that extends more easily to picking more than one element" - and that's what the actual Windows code uses.
> it’s more efficient because it reduces the amount of calls into the file system, which is where the bottleneck is
Upon reading it a few times I think the article is alluding to a crappy two-pass solution where you don't store a filename but instead an index into a directory, which is flawed anyway due to being racy.
OK, but isn't the kernel keeping the directory listing in the disk cache? Won't that prevent extra physical I/O if you do just read the directory twice?
If so, then in the second pass, it's all cache hits, and you're just paying the cost of calling into the file system. Hopefully that's pretty fast. But even if not, it's still absolutely dwarfed by the physical I/O required for the first pass. Windows XP era storage was spinning hard drives, not flash.
And if not, then I'm probably going to put my user icon coding task on the back burner and go ask the kernel team why a seemingly very common usage pattern isn't optimized.
(I realize he's not claiming the performance benefit was significant. I'm just trying to see it in the right perspective.)
selectRandomFromIteratorOptimized(iterator) { if (!iterator.moveNext()) { return null; }
I don't know if they've relaxed this since the days of non-FPU CPUs - anyone know? If they let the Weather app use a webview, there must be some floating point usage in there.
This code is at a much higher level though - at the user shell level, explorer.exe.
Anyways, asking Google's AI to remove the above code's use of floating point results in code resembling the original version.
On top of that, how often is it necessary to efficiently know the number of files in a directory while at the same time not caring about the files enough to list or display them? This algorithm is a special case where you could use the count of using a bit simpler code that ultimately would have the same file system API calls (since you cannot tell the FS to give you file #37 from that directory, so you'd have to use FindNextFile 37 times anyway, just like the sampling algorithm).
Most unix filesystems use inodes. inodes have the same format whether they represent a file or directory. So directory have a (usually unused) size member. NTFS doesn't use inodes, but the records in the MFT work the same way.
When adding or removing a file from a directory, you have to update the modification time of the folder, so you have to rewrite the entire inode anyway. Updating the size/file count at the same time would be free.
In my opinion the likely reason why file count isn't tracked is a lot more pedestrian: It wasn't tracked initially and we can never add it to existing file systems because the metadata would get out of sync if the FS was mounted on a kernel with no count support.
I base this assumption on the fact that many modern file systems do indeed keep track of the count.
This is more-or-less unique to ZFS. Other file systems even on Linux and FreeBSD generally don't provide this behavior.
https://jakeludington.com/2003/12/17/create_your_own_windows...
To me Windows 95/NT 4 is still the gold standard in Windows UI. It's certainly not the prettiest, that would be Windows 2000, but it was easy to use, easy to navigate and efficient with space.
I had tons of "Visual Styles" back then and in the last XP days I opted for grayish Royal or Royale. It's such shame that MS has abandoned Watercolor theme - that was a middle ground: interface was updated and yet, still similar to classic design. And it was even in some elements flat before that style become a dominant. Luckily we're slowly moving away from that and I won't be missing it.
i kinda miss xp at a cultural level since it was a bit the end of that computing culture cycle (after that apple started to dominate and ubuiquitous computing influenced desktop ui)
When did Apple dominate? I'm not trying to hate on Apple here, but there's this weird belief that Apple have had a lead in personal computer OS market share at some point in the last few decades, and it isn't really true since about 1984.
You can argue that they should dominate, but that doesn't seem to have happened.
I eventually got used to it and ended up liking it.
For awhile I ran XP 64-bit though, that did do one thing 2000 couldn't do.
>During the Luna studies, that people’s reaction to Luna was often, “Wow this would be a great UI for X,” where X was “my dad” or “my employees” or “my daughter”. People didn’t look at it as the UI for themselves; rather, they thought it was a great UI for somebody else.
(Luna is the name of the default look of Windows XP)
It's all been downhill from there as far as UX goes.
Its Vista. Like it or hate it. It’s Vista
nobody cares
Practically speaking, I might just allocate an array of 100 pointers. That's only 400 bytes. Then as you encounter each filename, allocate just enough memory for the actual length of the string (plus null terminator) and store the pointer in the array.
You still need to see all of the items once.
Imagine you have 2 items.
First one has 100% chance to be selected. So it does. Then the second has 50% chance to be selected. If it isn’t you effectively chosen the first one and have 50/50 chance to return either.
Now you add a third item. There is 50/50 chance of having either selected. And 1/3 chance of replacing the selection with the new one. Resulting in a 1/3 chance of selecting any of the three. (Because 1/2-1/6 = 1/3) 1/6 because there is 50% chance you will “steal” the selection.
If you are at picture 1, you have 100% chance of selecting it as the current winner.
If you are at picture 2, you have 1/2 chance of selecting it as the current winner, or 1/2 chance of keeping the previous fairly selected winner.
At picture 3, 1/3 chance of picking it, or 2/3 chance of retaining the previous fairly-selected winner. There are two of them, so 1/3 chance of each.
At picture n, you have a 1/n chance of picking it, or an (n-1)/n chance of retaining the previous fairly-selected winner. There are n-1 previous pictures, so all of them have had 1/n chance of being picked.
At every single step, there is the invariant of all pictures being considered that far having had an equal chance of being selected, and the next step always retains the invariant.
The picture selection algorithm's kind of single-pass iterator usage might have been more performant back in the XP days, as it avoids possibly expensive operations.
Modern CPU/other optimizations might make a multi-pass approach more performant due to better memory locality or other factors.
I don't understand Microsoft. These guys solve the most mundane problems with most elegant solutions and with sound edge-case handling scenarios, then they destroy all the effort with subpar programming where it matters and with user hostile behavior where they can't botch it.
Different teams with different goals and different management
Not to mention the stack ranking policies that caused engineers to sabotage other people's work and avoid collaboration in order to keep themselves high on the stack in order to avoid being fired.
I get what you mean, but it's really hard for any organization this size to drive consistent quality across the board.
Yep..
And image files were, and continue to be, a huge exploit attack vector
However, I still remember Wine laughing at Windows for WMF exploit and end up being affected from the same exploit. Now, that was a good laugh.
Just like the rest of us.
I think it's true, but not sure if I'm just falling victim to false memories... help?
I configured an account for someone with an Asian last name and it chose the fortune cookie.
Probably not voodoo, but it never seemed 100% random. More like some correlation was being done.
A simple rand/mod based on first character of username should be sufficient?
But then it occurred to me that XP was of an era when you would absolutely spend an afternoon downloading some megapack of 10000 cool user icons.