'Storage' to Replace Traditional Filesystems?

JigSaw writes "OSNews is reporting on Storage, an innovative project which aims to replace the traditional hierarchical filesystems with a new document store which is database-based (PostgreSQL). The current implementation, built under Gnome 2.x for now, offers natural language access, network transparency, and a number of other features. The project is currently in alpha (screenshots already available), and it is part of the next major generation of Gnome. It is currently developed by Seth Nickell, the person responsible for the enhanced Gnome usability on 2.x and its HIG, among other things."

Comments

How does the metadata get into the database?

[farnz]

My major concern with all these database type filesystems is that the gains are always shown as things like, "Find all films directed by Steven Spielberg", and yet this is not information that the computer can necessarily gather for itself.

Outside of a work environment, I've rarely encounter anyone who keeps consistent, useful filenames, let alone metadata indexes; it seems to me that people will skimp on the metadata, and thus limit the usefulness to metadata that the computer can collect automatically ("All movies that last under 90 minutes"). It's like CD collections, or books; libraries have nicely catalogued and ordered collections. Private individuals don't; they have roughly ordered collections on the shelf, and don't bother keeping them in any better order. I suspect the same will happen with these metadata systems; people won't do the work needed to make them truely useful.

Re:How does the metadata get into the database?

[Tom]

That's why we have community products. For music, CDDB works pretty good and is a working real-life example.

Other metadata is automatically inserted. When you install OpenOffice, it asks for your name and inserts that as the author into any new documents you create, for example.

Sure, the metadata on my personal machine will never be comparable to what a library could do. But it doesn't have to be - it has to be useful for me, not - like the library - for thousands of people with very different interests and approaches.

Re:How does the metadata get into the database?

[Anonymous Coward]

There are two kinds of metadata: intrinsic and extrinsic. Intrinsic metadata, as the name would imply, is information that's contained entirely within the file.

Some intrinsic metadata can be extracted with automation. For example, it's pretty trivial to examine a TIFF and tell you that it's X by Y pixels in a given color space. It's harder, but still possible, to tell you that the TIFF is predominantly red and green. It's impossible for the computer to tell you that it's a picture of a barn.

The same is true of extrinsic metadata: some of it can be extracted automatically, but not much. An example of extrinsic metadata would be a relationship. The computer can tell you that main.c and somefunction.c are both C language source code files, but it may or may not be able to tell you that they're both part of the same program. If the two files are explicitly related to each other through a makefile or some such, then the computer can know that they're related. But consider the collection of JPEGs I just copied to my home directory from my digital camera. A dozen of those pictures were taken in Fiji. The computer cannot know this unless I tell it, nor can it know which pictures were taken in Fiji and which were taken in my back yard last Tuesday. Thanks to my camera, the computer can know what apeture and focal length were used for each picture. In theory, if my camera had a GPS receiver in it, the computer might even be able to tell me there, on earth, the camera was when it snapped a given photo. But these are just automatic methods of telling the computer about the pictures. They're conceptually no different from sitting down and typing the information in. The point remains that the computer can figure out some things on its own, but it cannot know most things unless it is told.

You don't have to strain your imagination to think about this stuff. Consider your MP3 collection. Your computer can tell you that a given MP3 is 6:15 long, and that it's 192 kbps, and that it's stereo. It can't know that it's "Treefingers" by Radiohead unless somebody tells it first.

So you're basically right: automatically extracted metadata is marginally useful, but the really useful stuff has to be manually entered. And generally speaking, even in business environments, that sort of information simply never makes it into the database. It exists exclusively in people's heads.

That's the biggest challenge of digital asset management--which is, incidentally, essentially what we're talking about here. The biggest challenge is how to take information that people have in their heads and store it in some structured, persistent form. That form might be a three-ring binder or a card catalog or an Oracle database; the challenge is the same even if the technology is different.

Bottom line: this technology is really neat, and has limited applications in which it's very useful. But it is not generally useful, nor does it have widespread applications.

Re:How does the metadata get into the database?

[selderrr]

Very true and insightful !

Another argument to prove you right is the "rate this song" options in iTunes. With that feature, one can assign 1 to 5 stars to a song so that later, you can quickly select your favourites. Such system is flawed in 2 ways :

- I have yet to encounter anyone who uses it exhaustively. Most folks rate a few dozen songs. I have a library with 9000 mp3s and sure as hell I'm not going to spend a whole week rating them.
- I have yet to encounter someone who uses it consistently. Today I might consider Chris Isaak a 4star song since I'm in a depressed mood and it's raining outside. Tomorrow the weather might be beautifull and I mod him down to 2 stars cause he's a bloody negative wanker.

This is offcourse iTunes specific, but it shows that the assignment of metadata is far far far more complex than the methods to search/organize the stuff, which is what the "Storage" software above is about.

As an extra complication : consider that my metadata might not match someone elses. For instance if I were to label a mail "message from my brother", the same content would be "message from my son" for my father !

The fact that metadata based filesystems are not on our desktop is perhaps more to the fact that it's not a valid solution for data on desktop computers. Maybe, just maybe this is not due to MS squashing Be, as someone else was karmawhoring above.

Re:Obvious advantages

[dabadab]

"Integrated mime-types. No more relying on file extensions and other hacks. The mime-type (and subsequent viewer) is right there in the query"

And how does that meta data gets to the db? Oh, right, it will rely on file extensions and other hacks :)

Re:Obvious advantages

[azaroth42]

Obvious disadvantages:

SQL is slow compared to things like BerkeleyDB

We already have journaled file systems that can save metadata (though not user defined, I think)

Your database becomes corrupt, you lose everything.

Sorry, give me something that gives me back my data -fast-. If I want to do selects for files, I'll use locate and xargs.

--Azaroth

Re:filesystem is a database

[Viol8]

"What this world needs is a really big injection of orginal thought"

They are original ideas, they just don't make it into the PC world where MS dominates. MS come up with as many original ideas as McDonalds
and since all KDE & Gnome (and frankly most open source projects) are doing is playing catchup with MS then originality is never going to be
a prime concern.

Re:Windows?

[Zocalo]

Not quite, NTFS is a traditional file table with some bells and whistles, but it's not a "database" in the sense meant here(1). The next version of Windows, "Longhorn", is supposed to introduce a new file system called WinFS that will use a version of SQLServer as its backend. Whether they will actually deliver or not is another matter, since we were promised this in 1995 with Cairo and Taligent (remember them?), and now that Longhorn appears to have been pushed back...

There are also issues with gaining acceptance for the change in the way things work. This kind of thing has not really been done on a large scale in the wild before, on any OS, so whether people will be willing to accept the security and reliablity issues that may ensue is another matter. For example, what are the implications of a compromise in the database engine? MS is planning on using SQL, so if things go awry and it becomes possible to maliciously inject raw SQL to the filesystem interface... Oops. On the otherhand, the benefits for data retrival are *huge*. Imagine being able to find any audio files on your entire system by Justin Timberlake or Britney Spears and delete them all in one go by searching on the tag fields! ;)

(1) Technically, all filesystems are databases, it's just that current ones are a collection of flatfile database tables that can point to each other, generally in a heirarchial manner. When people say "database" in the same sentence as "filesystem" they usually mean "relational database". As an aside however, high end databases usually forgo the need for a file system and provide the ability to write their tables directly to disk on a dedicated partition.

"Damn, I left that on my roommate's desk"

[kfg]

"Well, where do you go?"

"Stanford."

"No problemo, I'm heading that way later and I can grab it for you. What's your room?"

"Dorm 5, Room 109. It's the desk on the left."

( We didn't bother to state earth.us because we were already inside those directories)

Yes, yes we do think heirarchically. Most of the history of human thought has been fitting everything we can lay our filthy little brain cells on into heirarcheis, whether they wish to fit into them or not. It's intuitive.

As for natural language didn't we learn about that with COBOL? Natural language only speeds the learning process slightly ( the majority of the learning still lying in the realm of understanding the basic concepts involved), but then becomes a pain in the ass forever afterward.

Looking at the screenshots it's also ugly as all sin. The physicist in me can't help but feel that a model that ugly can't possibly be correct.

I think this makes just about as much sense as using a document preperation language (XML) as the basis of a database.

Which is to say, none.

KFG

Re:Obvious advantages

[Anonymous Coward]

XFS limits user-defined extended attributes to 32 KB. Big, but not unlimited.

Also, extended attributes are fundamentally broken because they're stored in the inode. They do not survive, for example, a copy operation. Worse, they do not survive an open/save cycle in most cases, because most programs do not write to open files. Instead, they open a new file under a temporary name, write the data into it, close it, unlink the original file, then rename the temporary file to the original file's name. That way the data is safe if the program or computer fails during the save operation. This creates a new inode for the file data, however, which means extended attributes go bye-bye.

Extended attributes are not the answer. I'm not sure exactly what the question is, but I'm sure extattr are not the answer.

Nope

[varjag]

> SQL is slow compared to things like BerkeleyDB

BerkeleyDB is a hierarchial database. SQL is godzillion times faster on complex searches.

> Your database becomes corrupt, you lose everything.

Your filesystem becomes corrupt, you lose everything.

And yeah, I know about journaling, so don't bother :) But modern RDBMSes have integrity control facilities as well.

Re:so is everyone copying BeOS

[Twylite]

Summary of developments:

• BeOS has a good idea
• Microsoft announces a breakthrough in file system technology (around 1996), nothing happens
• newdocms announced on Slashdot in January 2003. Integrates with KDE, so no-one cares
• Microsoft announces WinFS plans for Longhorn. Slashdot decides that Microsoft sucks.
• Initial release of Haystack from MIT. Screenshot has XP interface so no-one gives a toss
• WinFS is reviewed, Slashdot has a flame war about file system layout, and concludes that MS sucks and a database file system is a stupid idea anyway and no-one wants one
• YEDFS (Yet Another Database File System) announced calling itself "Storage". Integrates with GNOME. FLOSS community bows and worships the superiority, leadership and sheer innovativeness of the application.

Re:i think

[glwtta]

and users weren't forced to learn the artificial concept of a "file"

Um, artificial as they may be, these so called "files" have been around for some time, in fact long before computers. Users can quite intuitively understand the concepts of "file" and "folder." I really think you are trying to make the difference seem greater than it actually is. (on the user side, that is)

Re:Windows' filesystem

[Zocalo]

Yeah, but as I mentioned in an earlier post, *all* filesystems are databases of some type, it's just a matter of context. Generally, when someone says a "database filesystem" today, what they actually mean is "a relational database driven, virtual filesystem providing an infinite variety of views onto a soup of metadata". I think I prefer the former and leaving the rest up to inference, but I'm sure that when these new products finally ship the marketroids are going to think otherwise.

I do deserve my wrists slapping though... I'd completely forgotten about BeOS! For shame!

as long as it stays in user space

[penguin7of9]

Of course, databases are very useful for organizing user data. People already keep PIM info, images, and lots of other stuff in databases. Lotus Notes is built entirely on databases.

But "replacing the traditional file system" carries with it the notion of ripping ext3 out of the kernel and putting a relational database there. That's a very bad idea. Databases don't belong into the kernel. They are far too inefficient to handle most storage needs, they are far too complex to go into the kernel, and they just don't need to be in the kernel. Operating system kernels need simple, fast storage systems. Something like ext3. ReiserFS is pushing the limits. PostgreSQL would be going too far.

As an aside, this is an idea that just about every nerd has when they learn about databases and retrieval. It's been tried various times since the 1960's. There are probably good reasons why interfaces don't use them. Perhaps most importantly, keep in mind that the vast majority of files on your system are not user files, they are bits and pieces of the operating system. And for the files that actually are used by users (mail, PIM info, images, text, etc.), they usually already have special-purpose database interfaces available to them as part of the applications that users use to access them.

Re:Obvious advantages

[smallpaul]

The file should be self-describing. It should have a header saying its type. You can never trust intermediary software to properly keep data and metadata together. The problem isn't just other operating systems. It is file formats like ZIP and prototols like FTP. Plus there is a problem that the file type a user gives a file on their computer may be just a means of triggering a bit of software (e.g. change a JSP file to HTML so it launches your HTML editor). But the intrinsic type of the file should not be corrupted by these user preferences.