Re: Documentation volunteer (was: Cluster replication: multiple masters)

["Mark H. Wood" <mwood@IUPUI.Edu>]

On Tue, 15 Oct 2002, Derek Simkowiak wrote:
[snip]
> 	My first question is, "huh?".  Why would multi-master replication
> require any code whatsoever?  It seems like you'd just configure slurpd as
> per my last message in this thread.  If someone familiar with this
> 'experimental' code could comment, I would be very grateful.

Scenario:  host H updates object O on master server S.  Host I updates
object O on master server T.  When S and T exchange updates, which one
wins:  the one from H or the one from I?  S and T must have identical and
precise rules for deciding which update to bury or their subsequent query
responses will be inconsistent.

This is why directory servers tend to attach timestamps, update serial
numbers, and the like to objects' records.  This stuff is not needed in
the single-master case; the single master is always right and the slaves
strive to mirror it exactly.  Update sequencing code is only needed when
it is possible to concurrently update two or more different servers
directly.  So a product which did not initially support multimaster
service will need new code if it is to support such service, and that new
code will be less mature and probably more buggy.

-- 
Mark H. Wood, Lead System Programmer   mwood@IUPUI.Edu
MS Windows *is* user-friendly, but only for certain values of "user".