Linked by Eugenia Loli on Tue 26th Apr 2005 07:59 UTC, submitted by Malx
General Development Transactional NTFS, coming in Longhorn, allows developers to group filesystem operations into transactions, and make those changes atomically. Changes made by transactions are isolated from each other, so that one transaction can 'see' a different set of files compared to another transaction. Transactions can also be used to view a frozen version of a file, fixed at a point in time, while another task updates the same file.
Order by: Score:
transactional ntfs
by Peter on Tue 26th Apr 2005 08:10 UTC

"So far, you've only heard vague talk about Transactional NTFS"

Ehm,jounaling file systems aka xfs,ext3,reiserfs???

benefit?
by Renaldo on Tue 26th Apr 2005 08:14 UTC

I don't see the benefit to this beyond being able to roll back a botched software update.

RE: transactional ntfs
by TraxTech on Tue 26th Apr 2005 08:14 UTC

No, journaling file systems only ensure that the FS meta-data is never corrupted if interrupted.

Transactions at the file level is "one more step right" to prevent file corruption (soft crashing when saving...).

RE: transactional ntfs
by Edward on Tue 26th Apr 2005 08:16 UTC

NTFS has had journalling since the beginning. This is different.
With plain journalling the filesystem will roll back to its last consistent state on restarting from a crash or a power failure.
This transactional support allow the application to have control over a set of operations and roll them back explicitly if they don't complete.
Which other file systems already have this feature?

transactional ntfs
by w on Tue 26th Apr 2005 08:19 UTC

NTFS is already journaling file system like ext3 and reiserfs. The thing is about grouping stream of filesystem operations in one transaction that you can commit or abort. Have you such control over FS operations in linux when using ext3 ? No.

v @w
by raver31 on Tue 26th Apr 2005 08:23 UTC
useful
by Evert on Tue 26th Apr 2005 08:30 UTC

to be honest, ntfs was journaling long before ext3 even existed :-)

i think the examples given in the original article (did some commenters here read it?) are a good proof of the usefulness of transactional file commit.

like patches and updates, guaranteed file / directory synchronizations, preventing broken links in your website, ...

by . on Tue 26th Apr 2005 08:35 UTC

Isn't this similar to Reiser4's atomic operation?

...
by Morin on Tue 26th Apr 2005 08:47 UTC

I also seem to remember that ReiserFS had this some time ago.

RE: Renaldo
> I don't see the benefit to this beyond being able to roll
> back a botched software update.

I think this could be important for web servers which store data in files (instead of databases), since many operations can be in progress on the same data at once.

In the end it's just another step in convergence of file systems and databases.

@raver31
by gnothing on Tue 26th Apr 2005 08:52 UTC

Sorry, but NTFS is a journaling file system. Sorry. It has been journaling system before the ext3-hack and long before ReiserFS ever became stable enough to use it.

Btw, you should read the article: handling file operations in a transactional way through a standard API is pretty awesome and I foresee that people will use this extensively.

@gnothing
by raver31 on Tue 26th Apr 2005 09:03 UTC

I never said NTFS is not a journaling file system. I was simply amused by how quickly everyone came out spouting the virtues of their OS's and the filesystems on their chosen OS.

I just wonder how many of them actually understand that a journaled filesystem is not a good thing on their desktops ?

And also, no matter what you read in magazines, a journal will not prevent data loss in the event of a crash, it simply makes filechecking quicker afterwards.

Reiser4 ?
by Vide on Tue 26th Apr 2005 09:13 UTC

Seems like these feats are already implemented in reiser4, which supports also plugins to expand its functions.

@raver31
by gnothing on Tue 26th Apr 2005 09:16 UTC

Hm, okay then. Valid point.

The world doesn't run on desktops alone, you know. Anyhow, I do think that a transactional FS tightly coupled with an OS - and its standard API - has a lot of merit. For instance: a crash is _not_ the only cause of IO failure, and even then it depends what kind of crash. You could have network failures, dependency failures, locking failures; they all could be rolled back without going through too much hassle. In _most_ cases this will make sure your FS is never in an inconsistant state (not that this happens a lot nowadays).

Btw, I don't think the transactional part is of that much importance here, it's the generalisation and virtualisation of the concept by embedding it in an easy-to-use API.

@gnothing
by Renaldo on Tue 26th Apr 2005 09:37 UTC

"Btw, I don't think the transactional part is of that much importance here, it's the generalisation and virtualisation of the concept by embedding it in an easy-to-use API."

Microsoft makes easy-to-use API's? I'm shocked!

@Renaldo
by gnothing on Tue 26th Apr 2005 09:47 UTC

Easy-to-use API's? Yes, really, they do make 'em, except for the MFC-debacle of course. All in all, Win32 or ATL aren't that hard to use and IMHO the .Net framework is a thing of beauty (although it's far from perfect, I know). I wonder if they'll incorporate the transactional functionality into the .NET IO namespace (they should).



Easy to use API's
by Hettar on Tue 26th Apr 2005 09:58 UTC

Sure MS makes easy to use API's. If you can master the Windows security API's anything is easy. More to the point if you don't commit suicide after spending weeks trying to figure out what some api function with pathetic documentation does, then You can master anything. Not that I'm bitter or anything :-)

@gnothing
by Renaldo on Tue 26th Apr 2005 10:00 UTC

well, MFC was just Microsoft doing what Microsoft does in trying to squash Borland's OWL.

I never really liked pure Win32 programming either... of course I've always had a love for Delphi.

Transactional FS update is nice but ...
by Sch on Tue 26th Apr 2005 10:11 UTC

remember that this "blog" is a technical PR marketing front-end.

You can do transactional file operation yourself on any file system. See http://users.auriga.wearlab.de/~alb/libjio/ for such an example for POSIX file system. The metadata file system journalisation is the step that help having a full transaction protected access is you want to.

Meanwhile, having the transaction mechanism inside the file system layer is better from performance point of view.

I wonder how transactional NTFS will manage conflicting transactional access to a file. This is were headaches start. Files do not spend all their life being accessed by only one program like "blog" example show us.

Easy to use api are a myth
by Sch on Tue 26th Apr 2005 10:17 UTC

Cannot remember where is saw those lines, buts the advice was "always look for where the API abstraction fails". This is a wise advice. It helps me a lot to known what were the programmers/designers intent. Usefull also for knowning how my own API are shaped.

Good API are when it took time to hit the wall of failed abstraction.

@Hettar
by gnothing on Tue 26th Apr 2005 10:23 UTC

Okay, maybe I'll just write a 10 page essay about which Windows API's are - IN FACT - easy to use, and which are - IN FACT - not. This was not my point. My point was: embed the functionality we're currently discussing (remember?) into an API that is widely accessible and embed it in such a way that one doesn't require extraordinary programming skills to implement it.

Some API's are a bloody mess, that's true, and some aren't, no matter what OS. And I have committed suicide, only two years ago. Thank God for that rollback-functionality.

storage
by bonjour on Tue 26th Apr 2005 11:34 UTC

Speaking of journaling, I was downloading stuff off the msdn site yesterday through our company's universal subscription service and I came across a package under server tools called "Volume Shadow Copy Service" which are supposedly a set of services that allow a developer to more easily take snapshots of file systems for better backup/restores.

Then I stumbled across this page:

http://www.microsoft.com/windowsserversystem/storage/default.mspx

Storage at MS? That's funny. I see a lot of ideas being taken from NetApp and EMC.

This transactional file system thing makes sense, it converges the theory of databases and filesystems, why not. Filesystems should be able to join soon too on that token then.

Will this break compatibility.
by Aaron Cohen on Tue 26th Apr 2005 11:36 UTC

The NTFS section of the kernel is becomming more and moremature. Will this break compatibility?

A related but different technology...
by Anonymous on Tue 26th Apr 2005 11:48 UTC

Logical Volume Management (e.g. LVM2 in Linux) can allow two views of an entire filesystem to co-exist, so long as they don't diverge by more than the amount of available space (a restriction which I'm sure also applies to this NTFS feature)

The traditional example is a long-running backup. You don't want to put your DBMS or similar stateful services into a quiescent state for the entire period of the backup (perhaps several hours). But you daren't run the backup while these services are accessing their data because there's no guarantee that the copy you backup is consistent.

With no special support from the DBMS you can use LVM2 to run this backup safely with only a few moments downtime. Simply shut down the DBMS safely, then instruct the LVM to provide a read-only fork of the filesystem and restart the DBMS. The backup process runs on the read-only view, which contains a consistent, stopped database, and once it has finished that view can be destroyed and the space made available for the next time.

v misread..
by memory on Tue 26th Apr 2005 12:15 UTC
Transactions
by Luke McCarthy on Tue 26th Apr 2005 12:30 UTC

Btw, you should read the article: handling file operations in a transactional way through a standard API is pretty awesome and I foresee that people will use this extensively.

Oh, their API is wonderful! *pukes*

On a more serious note, a good idea in principal (for some applications). One more step towards becoming a database, no?

Re: Will this break compatibility
by Anonymous on Tue 26th Apr 2005 12:31 UTC

> The NTFS section of the kernel is becomming more and more
> mature. Will this break compatibility?

From the Linux-NTFS project web site (uppercasing the Linux kernel driver part below is by me):
http://linux-ntfs.sourceforge.net/info/ntfsresize.html#longhorn

Question: Does ntfsresize support Windows Longhorn NTFS?

Answer: Yes, it supports. There have been reports that ntfsresize AND THE LINUX KERNEL DRIVER can recognize and handle flawlessly the NTFS version of the beta releases of Microsoft's next generation, code-named Longhorn, operation system. During startup the softwares of the Linux-NTFS project check the NTFS version and report unsupported NTFS if the version is unknown. So far they reported the NTFS version Longhorn uses is 3.1, exactly the same what Windows XP and Windows Server 2003 use. It will be seen if this remains true. If you'd ever find otherwise please let us know.

comment:
by somebody fine on Tue 26th Apr 2005 12:36 UTC

> Which other file systems already have this feature?

I think this is what DragonFly (BSD yes folks) is doing with their journaling layer.

http://leaf.dragonflybsd.org/mailarchive/kernel/2004-12/msg00105.ht...

I don't count on M$ doing it right, they never did...

Reiser4
by Anonymous on Tue 26th Apr 2005 13:06 UTC

Reiser4 is a step into this direction, but it has far weaker guaranties.

It only ensures, that one "write" operation either takes place or not. This means that if all writes in a program perserve the file format invariant, then the file can not become corrupt.

Traditional journaling filesystems do not journal the data written to the files, just the metadata part. So the filesystem will allways be consistent, but the data in the files can easyly become corrupted.

Full data journaling (when done in the traditional way) makes all write operations twice as slow (data is once written to the journal, and then to it's final location later), this is why no linux filesystem does it by default.

Reiser4 has some clever algorithms to reduce this factor 2 in the write times when doing full data journaling by keeping the file data itself if B* a tree.

One question about hard drives
by Anonymous on Tue 26th Apr 2005 13:11 UTC

How does the OS know that a given block has really been written to permanent storage (for example on a (S)ATA hard drive)?

I've heard that server disk controllers used to have non volatile caches on them, so when the data made it to the controller, it was permanent.

What is the current situation with common PC parts? Does anybody know?

Hard drive cache
by Zan Lynx on Tue 26th Apr 2005 14:40 UTC

Consumer and server hard drives with write cache are definitely volatile. If the power goes out before it finishes writing, kiss that data goodbye.

Server drives do support barrier operations which mean that the OS can ask the drive to finish all writes before the barrier. Then once that barrier operation is complete, the OS knows everything is on disk. These ops let the OS use drives with write cache and not be unreliable.

Lots of consumer drives don't have reliable barrier implementations and they will also lie about write completions. On those drives you're pretty much required to turn off write cache for reliability.

A UPS makes it less likely to lose data in write cache but there's still a chance the power supply will go dead or the drive electronics will die.

Server RAID cards with lots of cache often have built-in battery backup. It won't keep the data alive forever, but long enough to get the power back on so that they can finish updating the data on disk. These cards often disable the write cache on disk and just use their own.

@gnothing
by Rayiner Hashem on Tue 26th Apr 2005 15:47 UTC

Win32 is truely awful, ATL isn't that bad. I'm sorry, but POSIX is a lot older than Win32, and it hasn't had to create a __fork_ex_2() yet...

@Anonymous
by Rayiner Hashem on Tue 26th Apr 2005 15:59 UTC

That's a wee-bit misleading. Between write() and the metadata operations, all system calls are atomic (read() is idempotent, of course). Now, Reiser4 doesn't just guarantee atomicity for a single write(). It tracks modifications in 'atoms', which contain the results of several system calls, and commits those units atomically. Currently, I don't think there is an API for delineating your own atoms (effectively defining your own transactions), instead the system flushes atoms at regular intervals. However, the infrastructure for it is already there, the work is now in the API.

Re: @Anonymous
by Rayiner Hashem on Tue 26th Apr 2005 16:02 UTC

Much better reference. This is an e-mail from Hans Reiser to the reiser4 mailing list, dated Feb-2005:

"I think it is essential that you use the atomic transactions functionality to get good performance on top of reiser4 for what you are doing. Unfortunately, that API is still kernel internal at this time, and we have not yet coded the export of it to outside the kernel. We will though, we just need some time.

Hans"

Like Zope
by Ben on Tue 26th Apr 2005 17:31 UTC

I may be shot down in flames (literally) for being wrong, but is this similar to the ZODB (Zope object database)???

Ben

NTFS Format
by Malx on Tue 26th Apr 2005 19:37 UTC

Just to clarify: the NTFS on-disk format will not change to support this feature; the NTFS on-disk format has not changed since Windows 2000, and will not change in Longhorn. Existing NTFS tools should continue to work.

NTFS ... bah !
by Burbzman on Tue 26th Apr 2005 20:14 UTC

Ripped off from OS/2's HPFS.

MS developer behavior
by Marcelo on Tue 26th Apr 2005 23:28 UTC

"I went to install an OS X update to my iMac. "

A M$ developer using Mac ? Hum....


"I started here at Microsoft a little over three months ago. It's been an amazing transition from self-confessed "GNU hippy" to now, and has been more than a mild culture shock. The good news is that Seattle has far better weather than it's reputed to have."


He sold your soul to devil...

Re: NTFS Format
by Anonymous on Wed 27th Apr 2005 05:34 UTC

Hello Malx, thank you for the interesting article and comments on the NTFS formats. Unfortunately you're wrong considering the latter one. The NTFS on-disk format did change between Windows 2000 and XP. The former is version 3.0 and the later is 3.1. Please see the

1) NTFS FAQ: http://linux-ntfs.sourceforge.net/info/ntfs.html#1.4

2) NTFS technical documentation: http://linux-ntfs.sourceforge.net/ntfs/

3) Linux NTFS source code dealing with the changes: http://linux-ntfs.sourceforge.net/downloads.html

However the on-disk format indeed didn't change between XP/Win2003 and Longhorn, all of them are version 3.1: http://linux-ntfs.sourceforge.net/info/ntfsresize.html#longhorn

Glad to hear also that no changes are planned.

Re: Anonymous from Ericsson
by Malx on Wed 27th Apr 2005 23:52 UTC

Yes, you're right about those version numbers. Thing to remember here is that if a major version number changes, older implementations won't attempt to mount that volume. A minor version increment is designed to be backwards compatible. So Win2k should be able to mount a LH volume. I haven't tested this, but if you have any problems, let me know =)

Re: NTFS Format
by Anonymous on Thu 28th Apr 2005 12:44 UTC

Third party partition handling tools had problems with the newer format. The most famous is probably this one: http://support.microsoft.com/default.aspx?scid=kb;en-us;308322

Interestingly the open source Linux NTFS code wasn't hit by the 3.0->3.1 changes, contrary to common belief (that was the 1.2->3.0 transition). The NTFS 3.0+ write suppport was already disabled when XP came out in the original, NT4 driver used by the 2.4 and previous kernels and the rewritten NTFS code, supporting the 3.0+ formats, takes into account the different NTFS variants.