RE: AW: Partitioning questions II

[Dan Shriver <dshriver@sharemedia.com>]

----- Original Message -----
From: Tiefnig Daniel <daniel.tiefnig@infonova.at>
To: openldap-software@OpenLDAP.org
Sent: 6/18/01 11:56 AM
Subject: AW: Partitioning questions II    

> > Is a multi-master set up necessary for partitioning?
> 
> no.

great, thanks for this info.
 
> > Does partitioning follow referrals
> 
> what do you mean by "partitioning follows refferals"..?
> 
> > (to determine where to
> > add/delete/search/modify) if not what should I use?
> 
> if you perform a modify/search on a server, that holds a
referral for the
> subtree you want to modify, the server will return the
referral and that's
> it. ok?

My understanding is that if you want to add/remove/modify
entries you have to go through the master?  Am I wrong here?

In any case, the design I'd like would have one "master" who
deals with all such requests and if it doesn't own the entry (a
delegated subtree ("partitioning")) it adds the entry to the
appropriate slave.  I understand that in a normal case it adds
the entry to itself and then replicates down to the slave, but
shouldn't it be using a similar mechanism (following the
referral and doing an action similar to replication on the
appropriate slave)?

I'm stuck in a "chicken and egg" problem here.  If I want to set
up a partitioned directory how do I initially populate it?  I
can't add entries directly to the slave containing the right
partition (also I wouldn't want to make the client need to know
where to go- I'd prefer them to go to a central point that then
does the appropriate work in the right spot... because then all
I'd have to do is to keep referrals up to date (for
repartitioning... which I see as very likely in the future)).  I
can't seem to add them to the master (and you're saying that I
am conceptually and/or syntactically not doing the right thing
by trying that).  I do not see what the right way to look at
this problem is.  It can't be that I set up N+1 boxes (for the N
partitions and the one master) each as a master containing its
own partition, populate them, then turn N of them into slaves
that point to the one master.....  This would mean each time I'd
want to make a change somewhere I'd have to break up the system,
make the change, and reconnect it.
 
> > Why (after setting up referrals (which I am assuming allow
the
> > master to find/modify info on the slaves
> 
> wrong. a referral has nothing to do with master - slave
setup..

Ok, a referral simply tells where to go for
subbordinate/superior info (this isn't in my tree you want to go
to...).  But in a single master setup doesn't this have an
implied master-slave mapping (that if you want to add/modify you
percolate up to the top of your tree (or just before the top, if
you point to a global directory...) where the master lives and
tell him to do the work and he percolates it down to the
appropriate slave?

My understanding is that in "normal replication" the master
binds to all the slaves with the binddn and tells them to add
the appropriate entry.  Wouldn't partitioning work the same
(except that the master only binds to the slave(s) corresponding
to the subtree where the entry lives).
  
> you're mixing up distribution / replication of
ldapsubdirectories again.

Obviously I am very confused, but I don't have another
conceptual model to use here.  How are things supposed to work? 
Is there a good reference for how I should be looking at this
problem?