I should note, that I wrote this section under the pretend context that I
had coded a naive/bad restore routine that didn't change the invocation ID
at restore.  But obviously this can happen at any image based restore.

  http://brett.shirley.name/ImageBackupsBad.ppt

And the slides aren't really about all backup/restore, b/c I've removed
about 100 slides from the deck ...

Cheers,
BrettSh [msft]


On Thu, 5 May 2005, joe wrote:

> Brett, I think you would like Halo... I can visualize you getting into
> shooting other people. Plus it isn't just shooting, when playing MP, there
> is a lot of strategizing and thinking going on. :o)
> 
> On the hair issue. It makes you stick out in a crowd - literally. Keep it up
> - again literally.
> 
> 
> 
> As for the rest, great post. 
> 
> I like the way you specified the distributed system part. I knew what I
> wanted to say, basically that a DC doesn't exist in a vacuum; especially if
> it is in a forest with other DCs. You can't just pluck one out and roll it
> back without helping the other DCs understand that that has occurred. I
> couldn't think of a good way to say it. I think your method of stating it as
> a distributed system consistency issue will help folks have an understanding
> of it as well.
> 
> 
> The RID issue I hadn't even thought of in this context... Obviously a great
> concern to point out.
> 
> 
> 
> > [BrettSh] I thought I was being nice ... wow, it's going to suck, when
> someone actually annoys me. ;)
> 
> I know, that is why I said you were being extremely nice on this list... 
> 
> 
> 
> How's that blog coming..... This would be good info on it... As I have oft
> said to ~Eric - There is so much bad and mediocre info out there because so
> few of the people who really know the stuff aren't telling anyone about
> it... To that end, I have no problem saying stupid things in this forum and
> others if it incites/instigates you and other folks in the know to write it
> up and correct me; we all benefit and I don't mind looking stupid as much as
> like learning how things work. Barring that... People can just listen to my
> guesses and go with them[1]. :o)
> 
> 
> Again, great post. Thanks.
> 
> 
> 
>    joe
> 
> 
> 
> 
> [1] My guesses get better and better every day.... Trust me....
> 
>  
> 
> -----Original Message-----
> From: [EMAIL PROTECTED]
> [mailto:[EMAIL PROTECTED] On Behalf Of Brett Shirley
> Sent: Thursday, May 05, 2005 2:19 PM
> To: [email protected]
> Subject: RE: [ActiveDir] best practice?
> 
> 
> I don't really have serious time to answer this right now ...  so for now,
> you're going to have to trust me, it's not just a little bad you can recover
> from it with X, it is _really_ bad to do an image based restore, and hard to
> restore normality afterwards ...
> 
> I'll prop a portion of a slide deck later on, where I show to the backup
> vendors how the inconsistency is introduced ... but I don't know if it will
> make sense w/o my delivery.  It is also a bit simplified.  joe is close
> below, some comments inline, in joe's mail, as it's the closest so far to
> understanding why this is bad ...
> 
> BTW, clean and dirty AD DB have _nothing_ to do with this.  clean/dirty is
> an ESE / JET Blue level concept, this is an entirely AD Logical issue.
> Nothing prevents an ESE database from being imaged.  The AD has a design
> decision that prevents image based restores.
> 
> I don't play XBox or any computer games really.  I know that sounds weird,
> that a computer geek would not play video games, but I met a girl at a party
> the other day who is a huge FPS player, so I think the world somehow
> balances out in that respect.  How could that compare to the relaxing sense
> of accomplishment of working out paticularly cunning methods of compressing
> replication metadata ... I mean really?  Same goes for hair maintanence
> tasks.
> 
> On Thu, 5 May 2005, joe wrote:
> 
> > I am actually waiting for Brett or ~Eric to respond to your post as 
> > well. I am positive they could give you a bulleted list of things that 
> > you as well as the rest of us are completely unaware of that will go 
> > pear shaped both because they have seen things like that or just know 
> > it from familiarity with the code paths involved.
> > 
> > AD will not do a complete reload of the DB on its own, that was an NT4 
> > thing that occurred if the change log rolled. All gone now.
> > 
> > Do some searching on DSA IDs/GUIDs and Invocation IDs/GUIDS. A DSA ID 
> > is the GUID for the DC itself[1], it doesn't change for the life of 
> > the DC from my understanding. The invocation GUID[2] changes on 
> > restores, again to flag, hey new DB,
> 
> [BrettSh] It's not a new DB so much, as a new logical stream of changes to
> the distributed system ...
> 
> >  ... you don't know what my state is, so it can be brought into a 
> > consistent state.
> 
> [BrettSh] Don't like the term "consistent state" here.  I also don't like
> how we're talking about the DB ... I know all the AD repl docs, talked about
> it as a new database GUID, but that was poor taste ... there is a subtle but
> key difference between
> 
>       [local] database consistency, and 
>       distributed system consistency.
> 
> It's the later we're worried about.  +The later requires multiple nodes /
> DCs to have followed all the rules.+  Most of the rules are coded into the
> way AD behaves, when possible.  Thou shalt not image restore, is
> unfortunately not coded, and hard to be defensible against ... well, without
> sacraficing availability ... but lets not get into that trade-off right now.
> 
> > You should find hits on invocation id with topics of replication 
> > consistency, usn polling, AD restores, etc as it is key to all of them 
> > though it has been awhile since I went searching for that stuff.
> > Something I have read on a couple of occasions but can't say I agree 
> > with is that allegedly the DSA ID and invocation id are identical 
> > unless a restore has occurred. I don't think I have EVER seen them 
> > identical so I don't know where that info came from. I am noting it 
> > simply because I recall seeing documentation to that effect in the past.
> 
> [BrettSh] They should've been the same until the first restore ... there is
> a bug somewhere, that no one bothered to iron out.
> 
> BTW, we also change the InvocationID when we _re_-host an Application
> Directory Partition ... I'll leave the discussion of why to your
> imagination.  
> 
> Oh and since IFM is like throwing AD Restore and dcpromo into a blender for
> 30 seconds, IFM based dcpromo sort of changes the InvocationID.  
> You'll notice the invocationID of the DC you took the original backup from
> in the retired DSA signature of the newly dcpromo'd DC.
> 
> > 
> > Really try to find detailed info on how replication works. High USN is 
> > just the tip of the iceberg, there is a lot of underlying details but 
> > I understand where the misconceptions can come in, a lot of the 
> > documentation out there in the public realm simplifies the crap out of 
> > this stuff with analogies and very high level details without ever 
> > indicating that it is really quite more involved than that. This can 
> > burn you when you start making decisions based on those simplified
> examples.
> > 
> > If you really want to get into it, start fishing through the platform 
> > sdk
> > Ds* API calls. I would especially recommend the 
> > DsGetDCInfo/DsGetDcInfo2 functions and out of those the ones 
> > concerning DS_REPL_NEIGHBOR structures which gives a feeling of how 
> > much info there is involved with replication and consistency.
> > 
> > While it may be possible to force the invocationid to change after the 
> > image restore, I am not aware of a method other than doing a proper DB 
> > restore. It could be as simple as tapping that attribute in the 
> > nTDSDSA object but I certainly would NOT be willing to test that in 
> > production even if it worked great in the lab.
> 
> [BrettSh] 
> 
>       Plausible Proposal #1: (please see big warning below)
>       _Technically_, yes if you trigger an Invocation ID change after
>       you lay down the image, _AND THIS IS THE KEY_ ... before the DC
>       talks to any other DCs, and takes any new changes to the database.  
> 
>               This is one of those rules that all the nodes must follow,
>               and if you use an AD based backup/restore program, the
>               appropriate logic will be triggered, and the rules for
>               distributed consistency upheld.
> 
>       _Even_ booting the DC, may institute a change, that causes
>       distributed system inconsistency.  Obviously, tapping the object
>       from LDAP is not an option, you have to do it from DSRM.
>       Unfortunately, I've forgotten to tell you how you can trigger a
>       invocation ID change from DSRM ...
> 
> In short don't go there. These are not the droids you're looking for.
> 
> > 
> > Certainly, do not image DCs and use that as a recovery mechanism. The 
> > one way to do that, IMO, would involve snap shooting and rolling back 
> > all DCs in a forest at the same time. I don't see how this could 
> > effectively be done in the real world on real hardware. I visualize 
> > possibilities with virtualization software, but that would require a 
> > lot of testing and work to get there and some how guarantee that the 
> > snapshot was done at the exact time for all images.
> 
> [BrettSh] 
> 
>       Plausible Proposal #2: (please see the big warning below)
>       _Technically_, this will work too.  Requires all DCs to be off at
>       the same time when you take the image based back ups (I  
>       think).  Requires all the existing DCs to be turned off before
>       you restart the first restored image.  I think that is all that
>       is required ... but I'm not sure ... I don't care enough to try
>       to give anyone 
> 
>       Plausible Proposal #3: (please see the big warning below)
>       Of course a single DC forest can be image based restored as well,
>       though ... you're more likely to get SIDs reissued, and have old
>       wacky ACLs in this case, b/c IIRC we invalidate the present RID
>       pool on restore.  This can be mitigated by booting the DC, and
>       before creating any security principals, booting the next rid up,
>       can't remember how that is done off the top of my head though ...
> 
> > 
> > If you have done this in production already, I would recommend going 
> > back to what Brett said and doing a verification of your DB on all of your
> DCs.
> 
> [BrettSh] Jeez, I really hope no one is in this state, it can be quite
> disturbing to iron out.
> 
> > Again, Brett is someone who knows about the AD DB. Don't let his 
> > sometimes grouchy demeanor throw you off. He may get difficult at 
> > times but he is almost always trying to help, he just has interesting 
> > ways of expressing it on occasion. He has actually been extremely nice 
> > on this list compared to some other notes I have seen from him.
> 
> [BrettSh] I thought I was being nice ... wow, it's going to suck, when
> someone actually annoys me. ;)
> 
> >  Basically I say the same about him
> > that I have often said about myself; don't mistake the quality of the 
> > delivery for the quality of the information. :o)
> 
> [BrettSh]
> 
> So first let me divulge, that I am not in fact the Garage Door Operator for
> building 7, in fact I am a developer/programmer (we're call Software
> Development Engineers at Microsoft) in Windows, ON Active Directory.  
> Before my recent move to the ESE development, I worked on the AD Replication
> development for ~5.5 years, spending time working on AD Replication, AD
> backup/restore, a small bit in AD Schema/Database stuff, AD tools, and even
> dabbling in DcPromo off and on when required for those years.  Quite frankly
> I'm the one who has dealt with almost all the areas affected by a bad image
> based backup/restore, and the parts that make a good backup/restore
> possible.  I'm uniquely qualified to say:
>       Image based backup/restores are not supported for AD.
> 
> So we had this customer who wanted to use SAN based hot split on Win2k AD
> (which is even more unsupported, as they didn't shutdown all the DCs, like
> Plausible Proposal #2 above), after explaining that they'd have to shutdown
> all DCs, and them agreeing (though I doubted they'd actually do that, it's
> amazing what customers will do when they think they understand better than
> you) and then they agreed for restore, they'd take ALL the DCs back to the
> same backup time, at the same time, and working out this complicated set of
> steps they would need, I pointed out this:
> 
> --- begin quote ---
> I can't confirm if you will fail ..., but that set of steps if correctly
> followed will not cause forest corruption due to USN rollback.  Honestly, it
> isn.t worrying about this once PSS guided transition that worries me,
> following those types of steps once isn't hard . it is someone not
> understanding why each of the parts of the technique were required, and
> later trying it again, and not getting it right.  In general customers may
> not truly understand the system's requirements, EVEN after they say they do
> (b/c they believe they do, no one intentionally hoses their domain, but
> somehow it happens) so it's just easier to say "no mirror splits on
> unsupported SANs"
> --- end quote ---
> 
> So ....
> 
>  Warning!  Warning!  Danger Will Robinson!  Danger!
> 
> So the same goes for all 3 proposals above ... while technically you could
> work out the exact set of steps required, it is likely to be an error prone
> manual process ... will the next guy who maintains the corp infrastructure
> understand it all ... will you miss a step ... if you have lots of DCs in
> branches, how do you know one won't be missed ... you're playing with fire
> ... and the slightest tweaks can change the answer substantially, for
> instance auth restore for proposal #1 must be done after triggering the
> invocation ID to change, which would require a reboot ... even me with all
> my knowledge, wouldn't implement such a mechanism in a live corporate
> deployment ... it's subtle, and it is not worth the risk.
> 
> Friends don't let friends use image based backups of AD.
> 
> Cheers,
> -Brett [msft]
> I'm just kidding, I just made all the above up, I really am just the
> Building 7 Garage Door Operator ...
> 
> 
> > 
> >    joe
> > 
> > 
> > 
> > [1] It is the objectGUID attribute of the ntdsdsa object(aka NTDS 
> > Settings object).
> > [2] It is the invocationID attribute of the ntdsdsa object.
> > 
> >  
> > 
> > -----Original Message-----
> > From: [EMAIL PROTECTED]
> > [mailto:[EMAIL PROTECTED] On Behalf Of Bahta 
> > Nathaniel V Contr NASIC/SCNA
> > Sent: Thursday, May 05, 2005 10:22 AM
> > To: [email protected]
> > Subject: RE: [ActiveDir] best practice?
> > 
> > Joe,
> > 
> > I appreciate you indulging me in detail.  I was just curious on what 
> > the consequences may be of imaging and restoring DC's.  We are always 
> > evaluating and re-evaluating DR methods and techniques, and this was 
> > the latest hot topic.  I thought AD pushed changes up to a 
> > pre-determined amount and then it would just replicate the whole 
> > database if the number of changes were too great.  I am not sure of 
> > the in-depth implications of restoring imaged DC's but I know the 
> > difference between a clean and dirty AD DB and it sounds as though the 
> > metadata cleanup and synchronization is not meant to happen with an AD 
> > unaware application such as ghost.  Perhaps an application that could 
> > stabilize an old DC with the new AD DB would be something that would 
> > have to be looked at.  Or maybe an image of a member server and a 
> > dcpromo is the easiest way to recover a DC.  I have intentions on working
> smarter, not harder, but that does not forgo my lust for understanding right
> from wrong.
> > 
> > Thanks again for the rebuttal.  It always helps to hear things from 
> > all perspectives to get a better look at the big picture.
> > 
> > Nathaniel
> > 
> > 
> > -----Original Message-----
> > From: [EMAIL PROTECTED]
> > [mailto:[EMAIL PROTECTED] On Behalf Of joe
> > Sent: Wednesday, May 04, 2005 2:36 PM
> > To: [email protected]
> > Subject: RE: [ActiveDir] best practice?
> > 
> > I'm not Brett[1] but wanted to just say something really quick here. 
> > 
> > Well a couple of things actually.
> > 
> > 1. When it comes to AD Database consistency and replication. Brett is 
> > someone I would tend to listen to very carefully. I may not understand 
> > what he is trying to say but I will try like heck to understand it. 
> > Rough around the edges though he may be, he knows a lot about the guts 
> > of the AD DB and Replication. Keep in mind he wrote some of the most 
> > "brilliant" parts of repadmin[2].
> > 
> > 2. When you image and recover the image you are bypassing any and all 
> > logic associated with a directory DB recovery. I.E. You aren't 
> > restoring the database through the very specific DS Backup/Restore API 
> > so you don't get the cool things that it does like renaming the 
> > Database GUID aka invocation ID which effectively tells all of the other
> partners there is a "different"
> > database out here that needs to be fully updated. 
> > 
> > I haven't fully thought out the implications of that but one thing 
> > right off the bat is the thought that all DCs maintain high water 
> > vectors for all databases so they know where they are at for 
> > replication. This isn't just kept on the DC in question, this is kept 
> > all over so I could see serious possibilities of issues there. 
> > Additionally think of a change that mastered on that database and 
> > replicated out. How do you get it back if the DB is rolled back and 
> > all of the other DCs already think that DB has that info since it was
> mastered there?
> > 
> > You get ~Eric, Dean, and Brett thinking about it and I expect you 
> > could find all sorts of horrible things that this can do to you.
> > 
> > I think the idea that a DC can be restored from an image like that 
> > because it is "sort" of like restoring the DB is flawed at the very 
> > best. You don't have a full comprehension of what is being done in the 
> > backend to support that restore. If it were that simple, why do you need a
> backup api at all?
> > Mirror the DIT and zip it and there is your backup... It doesn't work 
> > that way.
> > 
> > As Brett indicated... Bad mojo... Heck I will go further, positively evil.
> > You could damage your AD in ways that you (and it) has no clue about 
> > and only later run into it when you are trying to figure out niggling 
> > consistency issues in applications that act odd some of the time.
> > 
> > 
> >    joe
> > 
> > 
> > 
> > [1] And I couldn't play him on TV either, Brett stores a good portion 
> > of his height in his hair and I store mine in my legs.
> > 
> > [2] His words when I met him in person at an MVP summit. He was quite 
> > excited to talk about that portion of the code...
> > 
> > 
> > 
> > -----Original Message-----
> > From: [EMAIL PROTECTED]
> > [mailto:[EMAIL PROTECTED] On Behalf Of Bahta 
> > Nathaniel V Contr NASIC/SCNA
> > Sent: Wednesday, May 04, 2005 1:59 PM
> > To: [email protected]
> > Subject: RE: [ActiveDir] best practice?
> > 
> > Brett,
> > 
> > What is your basis for not being able to restore a DC from a image?  
> > If the DC has an old copy of the directory data, it will check its 
> > USN's and update its copy.  What could cause havok if anything?  We 
> > are about to institute this very same concept here to turn DR into a 
> > 10 minute process when it comes to operating system recovery.  We will 
> > image the servers monthly and restore from said image whenever one 
> > crashes.  What could cause a problem by restoring a DC, it will be 
> > timestamped to be old and AD will synchronize it with the rest of the
> domain.
> > 
> > Please elaborate on your basis for comment.
> > 
> > Nathaniel Bahta
> > 
> > -----Original Message-----
> > From: [EMAIL PROTECTED]
> > [mailto:[EMAIL PROTECTED] On Behalf Of Brett Shirley
> > Sent: Wednesday, May 04, 2005 11:47 AM
> > To: [email protected]
> > Subject: RE: [ActiveDir] best practice?
> > 
> > jlc,
> > 
> > You can't restore a single DC via an image based backup, either.  It 
> > is not supported, it is not allowed ... it is bad mojo.
> > 
> > Well, it wouldn't cause issues if the forest had ONLY that one DC 
> > (seems unlikely the case), or for a multi-DC forest, you'd have to 
> > shutdown all the DCs in the forest at the same time, when you took your
> backup images.
> > And then on restore, restore them all at the same time.  Basically a 
> > pretty infeasible suggestion.
> > 
> > Cheers,
> > -Brett Shirley [msft]
> > 
> > This posting is provided "AS IS" with no warranties, and confers no
> rights. 
> > 
> > 
> > On Wed, 4 May 2005, Joseph L. Casale wrote:
> > 
> > > Exactly, I do it for DR purposes, the old one dies - I reimage it 
> > > and put it back out there.
> > > No poblem...
> > > jlc
> > > 
> > > -----Original Message-----
> > > From: [EMAIL PROTECTED]
> > > [mailto:[EMAIL PROTECTED] On Behalf Of Phil Renouf
> > > Sent: Wednesday, May 04, 2005 7:01 AM
> > > To: [email protected]
> > > Subject: Re: [ActiveDir] best practice?
> > > 
> > > On 5/4/05, John Shukovsky Jr <[EMAIL PROTECTED]> wrote:
> > > > BUT....as for DC's. I do "image" dc's using Symantec Livestate 
> > > > Recovery ( formerly PowerQuest V2i ). It works wonderfully. I 
> > > > primarily use for backups. I have not had to recover a server in 
> > > > production ( and hope I do not have to ) but I have in lab 10+ 
> > > > times
> > > and servers are as clean as ever.
> > > > You should take a look.
> > > 
> > > When Brett mentioned imaging DCs being a bad idea and to never ever 
> > > do it I believe that he was meaning don't Image a DC and try to use 
> > > that Image to build other new DCs and just trying to change the SID 
> > > like you would for a desktop. Bad idea!
> > > 
> > > Phil
> > > List info   : http://www.activedir.org/List.aspx
> > > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > > List archive:
> > > http://www.mail-archive.com/activedir%40mail.activedir.org/
> > > List info   : http://www.activedir.org/List.aspx
> > > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > > List archive: 
> > > http://www.mail-archive.com/activedir%40mail.activedir.org/
> > > 
> > 
> > List info   : http://www.activedir.org/List.aspx
> > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > List archive: http://www.mail-archive.com/activedir%40mail.activedir.org/
> > List info   : http://www.activedir.org/List.aspx
> > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > List archive: 
> > http://www.mail-archive.com/activedir%40mail.activedir.org/
> > 
> > List info   : http://www.activedir.org/List.aspx
> > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > List archive: http://www.mail-archive.com/activedir%40mail.activedir.org/
> > List info   : http://www.activedir.org/List.aspx
> > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > List archive: 
> > http://www.mail-archive.com/activedir%40mail.activedir.org/
> > 
> > List info   : http://www.activedir.org/List.aspx
> > List FAQ    : http://www.activedir.org/ListFAQ.aspx
> > List archive: 
> > http://www.mail-archive.com/activedir%40mail.activedir.org/
> > 
> 
> List info   : http://www.activedir.org/List.aspx
> List FAQ    : http://www.activedir.org/ListFAQ.aspx
> List archive: http://www.mail-archive.com/activedir%40mail.activedir.org/
> 
> List info   : http://www.activedir.org/List.aspx
> List FAQ    : http://www.activedir.org/ListFAQ.aspx
> List archive: http://www.mail-archive.com/activedir%40mail.activedir.org/
> 

List info   : http://www.activedir.org/List.aspx
List FAQ    : http://www.activedir.org/ListFAQ.aspx
List archive: http://www.mail-archive.com/activedir%40mail.activedir.org/

Reply via email to