directory-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Martin Alderson" <Martin.Alder...@salfordsoftware.co.uk>
Subject Re: [Mitosis] Interface vs classes
Date Fri, 09 Jan 2009 16:14:27 GMT

Yeah, I noticed that you were looking at mitosis again.  That's good
news.

Before I started working on mitosis I thought a multi-master
replication system would be pretty easy - the complexity  was
surprising.  The biggest problems (both with MM replication in general
and in mitosis) are the cases that aren't so obvious at first (to me
anyway) like adding new replicas, servers that are down for a while,
concurrent overlapping modifications, large quantities of data, etc.

For my use of mitosis everything had to work or at least fail
gracefully.  Data loss was unacceptable (I was replicating account
balances) and having the server fall over or stop responding wouldn't be
great, since this was a minor component of our system and not one that
our users would want to watch - they would barely even know about it.

I think mitosis is a decent enough implementation that almost works. 
The design seems reasonable - it almost has a configurable log store
already for example.  The transport seems overly confusing to me and I
would love to see that replaced with an LDIF based format compatible
with openLDAP, ideally with encryption support.  There are some concepts
which don't even seem to be used at the moment, like the tombstone
entries and entry UUIDs.  The entry UUIDs seem to be generated and
stored in each entry, but never used.  If we can be assured that all
change log entries will be applied in order (as I believe they are at
the moment) then I'm not convinced we even need UUID's.

I'll keep monitoring this list and try to help out where I can.

Martin


>>> Emmanuel Lecharny <elecharny@gmail.com> 09/01/2009 13:57 >>>
Martin Alderson wrote:
> I agree, given that the CSN interface implies that all implementors
> have timestamp, replicaID and operationSequence it really doesn't
seem
> likely that anyone would want to make a custom CSN class.  CSNFactory
is
> just as bad.
>   
Cool, thanks Martin !

Btw, we are going to review the complete Mitosis code, and your
insingth 
might be very valuable, as you worked on it.

DO you think there are some areas we will have to dig ?

Can you give us a feedback about the Good, the Bad and the Ugly ?

Many thanks !
 

--
cordialement, regards,
Emmanuel L├ęcharny
www.iktek.com 
directory.apache.org



Mime
View raw message