Slashdot Mirror
WinFS' Spot on Back Burner Nothing New
osViews.com writes "Charles Arthur of Independant.co.uk has an interesting editorial which analyzes Microsoft's recently postponed 'WinFS,' the file system that Microsoft had been planning to implement in Longhorn. His editorial reminds us that this technology, previously referred to as the 'NT Object Filing System' was intended for a previous version of one of Microsoft's operating system's code named 'Cairo.' Microsoft first spoke of the 'NT Object Filing System' in 1992 and scheduled a beta release in 1996 and then a full release in 1997. But limitations cause it to continue being delayed."
Yeah, something like Tivo. Once you get it working and get used to it, you would feel like losing one hand without it.
Just my 2c
If programs would be read like poetry, most programmers would be Vogons.
Is it possible that NTFS's meta-data was the first foray into actually implementing this? Like WinFS, they might have started wanting to categorise everything and link it together but settled on GUIs to change ID3 tags, as well as other meta-data (like Word), planning to implement the search engine and filesystem service layer (WinFS) later on.
However, NTFS works fine as it is. Like the parent, I too question the need for WinFS when some of its features have been implemented over several iterations of Windows. Perhaps that's why they dropped it.
From what I know of WinFS, it really won't be all that important anyway. It is supposed to provide a way for all files to be treated the same by the OS (roughly) right? Thus making it easier for users to search, browse, or otherwise find these files?
Well, I don't know all of the juicy details of WinFS but I have played with the new Longhorn build. The search tool that is in the Alpha release (MSDN) is much improved over the current WinXP search. It was pretty cool, although some of it can be chalked up to eye candy. It still had a certain ease of use to it.
I doubt WinFS will ever be complete, personally. But I am sure some of the innovation and development benefits will still reach us as consumers. I know where I work, we spend time doing things the customers will never see. But they will still reap many of the benefits.
I'm working on a object file system right now, and it's really not easy.
It's a simple concept:
Store on a standard journaled b-tree (or similar) filesystem the binary data, and store in a database all sorts of meta-information about the data. Also if you want, store a reverse index of the textual info and maybe another 'index' of image features if it's an image.
Then if you want to get anything, no need to go through the filesystem's tree, you can hit the DB indexes and get info instantly.
The real problem is keeping all of this in synch, with almost flawless atomic operations. (of course it's pretty much impossible to be flawlessly atomic, but one should come as close as the current journaled filesystems are).
So if you're using 2 components, let's say, a filesystem and a SQL database, then you need to open a SQL transaction, do your inserts/updates/deletes, then do the filesystem operation, then do the SQL transaction commit. If anything fails, you can revert the SQL modifications and everything goes back to normal. But if the filesystem has problems, then you can't keep the damn DB synchronized, and at some point you'll have to resynch both.
On 100k files, no problem. On 200MM files (what I'm aiming for), you're pretty much screwed. Then you have to start thinking of a self-healing system with a constantly-running checker that must ensure that it's very resource-efficient, etc...
It's just a huge problem. Supposedly Apple is solving this by Q1 2005, but I wouldn't be surprised if we see a massive increase in filesystem corruption bugs for a while on OS X (unless the DB indexing piece is just that, an indexer that runs x times a day and isn't atomically joined to the filesystem operations).
And that's why it's taking so long. Accessing filesystems as SQL data has always been a dream of anyone who has had many files. They just never knew about it.
WinFS is the 'real' solution IMO to all things like iTunes playlist managers, and expensive Content Management Systems yadi yada.
Sure, no consumer is expected to actually use SQL statements, but that doesn't mean that user mode programs should *implement* SQL features. User mode programs should only be the 'translation' layer between the user's point and click GUI, and the OS' internal implementation of the db. Surely, anyone can see that collecting meta data from the file system, and duplicating it in usermode so that you can have search capabilities on it is wasteful.
This article wasn't news to me, I've actually been waiting for this damn WinFS since just about 1996... And by god, is it ever turning into Duke Nukem Forever, but you know what, it's such a cool feature that I still can't wait for it to come out... (figuratively speaking)
Let's put this in perspective. In '92 MS was looking at the Sybase source code and thinking about building a new filesystem around a database engine. Chicago AKA Win95 was almost out the door and it seemed reasonable to shoehorn this into Cairo (NT4). They were absolutely the dominant and fastest growing player.
I commented to a collegue in '93 (paraphrasing Robert Heinlein) that I did business with MS for the same reason I obeyed Newton's laws.
What happened around 1995? The internet became a commercial entity. Suddenly, MS needed to provide new applications (like IIS, IE, Outlook Express, an SMTP aware Exchange server, etc.) not just dork with cool OS technologies. A few years later, they are comfortable again after playing catch-up and start thinking about filesystems again, this time in "Longhorn". Again, they started talking about the capability two OS releases into the future.
However, this isn't a feature that is going to drive sales. MS needs to keep developers of home and office apps happy so they develop yet another new graphics system to replace DirectX. The perception of Windows security has never been lower and is starting to affect sales. IIS is losing ground again to Apache/Linux.
It's time to focus on revenue streams again and the revolutionary, expensive, difficult-to-build features get axed. It's probably not a bad idea. Think about the problems they've had with MS-SQL and ask yourself if you want a similar technology built into every teenager's game and grandmother's email box.
IIRC "NT Object Filing System != WinFS"
WinFS is supposed to be based on SQL Server, when NTOFS was announced, MicroSoft hadn't yet acquired SQL Server.
I thought NTOFS was what morphed into the fast-find thingie that shipped with Office.
I don't need no instructions to know how to rock!!!!
Exchange4Linux does the same thing, but with Python and PostgreSQL. At first I hated the idea but man it does work remarkably well.
One correction - filesystems (at least most UNIX filesystems) are not constrained to tree structure; the leaf nodes may have any number of parents, i.e. a file may be in any number of directories simultaneously. (Use the "ln" command). And using ln -s you can practically place a directory in any number of parent directories.
I use this to organize my music collection alphabetically by artist, by genre, and by the date I got the music simultaneously. (I tend to be most interested in music I got recently, because I'm not tired of it yet).
I know people tend to organize files and directories in a tree structure anyways. If you ask me that's because people are happy to maintain the analogy of a physical item that can only be in one place at a time - so what does that mean for WinFS?
Every filesystem is a database at heart. They already contain other attributes like permissions, create and modify date etc. The place to store this stuff is in the FS because the database is already there. All you need to do is add some more stuff like extended description, a few topic reference fields, and and slap of a query engine on it. The query engine does not need to be real complex either. You can get away with little or no formating/sorting/grouping support as the user space app which performs the query should take care of that. All you need is basic bool logic and string comparision. Most of this code already exists out there under a free license, I am not saying it would be a copy past job but there are examples of required algorithms which developers can look at safely, without running afowl of and IP.
The one tough thing WINFS aims to do that would be simple in user space is it hopes to be able to look in files and gleen some atributes form them. This is great if you can hook into some of the libraries form office or adobe et al, it saves you from having to implement parseing for all that stuff. I am not quite sure how you solve that one at the FS level. I just fear a user space system will get real crufty real fast and break when major changes occur to the files and their real attribes on disk that the DB can't know about. Like if a mount point gets moved or everything is resotored form a tarball and the dates get changed/permissions change a little because someone was careless. I think overall getting the neccecary info form the user when new files are created would be a fair compromise, the only issues is rule one of DATA "crap in crap out".
Then there are all the problems that you mostly have to deal with wether you do it in the FS or as some user space hack/bloatware thing:
Note that file creation would constitute just that you would want/need for efficency archives to contain all that info for the file in them, so the user does not have to enter it. Makefiles and the like would have to be update to do magic and fill in that data for the output files. Then you naturally have to fix all the gui tool kits so their fileIO dialogs support that info, any apps with custom dialogs will need to be patched as will console apps. Some sort of default values would be need for apps that just can't resonably support collecting that info as well. I don't want to have to fill in values everytime I "cat" somethig, I mean to unlink moments later.
I think its clear there are lots of differcult usability problems to solve. Some could probably extend and of the major OSS filesystems to include some extra attributes and add a crude query system, its all a question of what do you really do with it once you have it. I am sure R&D at Microsoft is just as perplexed on that point as I am. I feel sory for them since the marketing dept has been pushing this as the next big thing for almost a decade now, the pressure must be intense.
Repeal the 17th Amendment TODAY! Also Please Read http://www.gnu.org/philosophy/right-to-read.html
Except it is relevant because Reiser4 has metadata built-in. WinFS is supposed to be built on top of NTFS but its (NTFS+WinFS) purpose is similar to that of Reiser4.
Time makes more converts than reason
Except it is relevant because Reiser4 has metadata built-in. WinFS is supposed to be built on top of NTFS but its (NTFS+WinFS) purpose is similar to that of Reiser4.
NTFS has always had metadata built in. That's not what WinFS provides.
Coming soon - pyrogyra
The key here is this: I am not at all interested in a system that fundamentally assumes I am stupid. I will be utterly devoted to a system that fundamentally assumes I am lazy.
WinFS and masses of metadata assumes the stupid and not the lazy. The reason I don't want to have complicated trees of directories is that i am too damn lazy to do so and maintain it. Requiring me to add masses of metadata instead of a directory heirarchy does not address the problem: I am lazy!
Such a system will work well for limited uses - anything that has self populating metadata (such as music collections where files will either come with suitable metadata attached, or if I rip a CD I'll automatically attach suitable metadata via FreeDB or what have you. Similarly for a certain amount of video etc.
Such a system will work passingly well when you have a reasonable amount of attached metadata automatically, for instance email.
It won't work well for general user created documents and the like.
In the end a lot of data is purely user created - from speadsheets and letters to photos downloaded off digital cameras.
Find a way for me to be lazy and still have quick and easy access to all of those, and then you'll have my interest.
Jedidiah.
Craft Beer Programming T-shirts
I wrote the first design document for ReiserFS in 1984....
The nice thing about being slow in solving a hard problem is that others are also slow....
Just to use your example for one type of problem that NTFS/FAT16/FAT32 users have just now (although there are several types of problems if you think about it for a while).
You have some mp3s for a band called "Green Day" Do they go under emo, punk, rock (or even 'pop'). You may have strong feelings one way or the other as to which category they fall under, and therefore be able to save these files in one place and find them again at a later time. But will other people who use your computer/network have the same feelings about what kind of music green day play? How will other users now find those Green Day mp3s if they dont know which directory to look under?
This is at the heart of the arguments behind metadata and multiple inheritance, The reiserFS home page has lots of good information on the issues involved with file systems