[Replication] Using time offsets instead of absolute time

[Alex Karasulu <ak...@apache.org>]

Hi all,

The first thread on replication after Emmanuel's comments started taking a
turn towards the question of time synchronization across replicating
servers.  Time synchronization is very nice to have and forget about while
writing the code around replication.  And yes we do not have control over it
since administrators can fudge this up.

Another alternative would be to have all the servers in the cluster
communicate with one another on occasion to share their current time to
establish offsets.  These offsets (differences in time) between servers in
replication agreements can be used to correct for any time variance we might
encounter.  Furthermore if the time is adjusted on the hosts, the cluster
can adjust to time changes.

WDYT?

Alex

Re: [Replication] Using time offsets instead of absolute time

[Howard Chu <hy...@symas.com>]

Alex Karasulu wrote:
> Hi all,
>
> The first thread on replication after Emmanuel's comments started taking
> a turn towards the question of time synchronization across replicating
> servers. Time synchronization is very nice to have and forget about
> while writing the code around replication. And yes we do not have
> control over it since administrators can fudge this up.
>
> Another alternative would be to have all the servers in the cluster
> communicate with one another on occasion to share their current time to
> establish offsets. These offsets (differences in time) between servers
> in replication agreements can be used to correct for any time variance
> we might encounter. Furthermore if the time is adjusted on the hosts,
> the cluster can adjust to time changes.
>
> WDYT?

Doing this accurately requires the same amount of work as NTP, to factor in 
network roundtrip delays etc.

In one of my prior jobs I took this route for a Kerberos client implementation 
on Windows. (Which, if you're curious, is used by Eudora...) Since Kerberos 
would refuse to authenticate a client if its clock was more than 5 minutes 
different from the server, we simply recorded the clock offset of each KDC in 
the client. But - wtf - we had to use SNTP to maintain the offsets anyway...

-- 
   -- Howard Chu
   CTO, Symas Corp.           http://www.symas.com
   Director, Highland Sun     http://highlandsun.com/hyc/
   Chief Architect, OpenLDAP  http://www.openldap.org/project/

Re: [Replication] Using time offsets instead of absolute time

[David Boreham <da...@bozemanpass.com>]

It's 2009 not 1989. Systems need correct time anyway. Many other 
replication systems require
synchronized clocks. Don't attempt to solve this problem yourself, use NTP.

Alex Karasulu wrote:
> Hi all,
>
> The first thread on replication after Emmanuel's comments started 
> taking a turn towards the question of time synchronization across 
> replicating servers.  Time synchronization is very nice to have and 
> forget about while writing the code around replication.  And yes we do 
> not have control over it since administrators can fudge this up.
>
> Another alternative would be to have all the servers in the cluster 
> communicate with one another on occasion to share their current time 
> to establish offsets.  These offsets (differences in time) between 
> servers in replication agreements can be used to correct for any time 
> variance we might encounter.  Furthermore if the time is adjusted on 
> the hosts, the cluster can adjust to time changes.
>
> WDYT?
>
> Alex

Re: [Replication] Using time offsets instead of absolute time

[Alex Karasulu <ak...@gmail.com>]

He he okie I get the message guys.  You, Howard and David are right.
Effectively this would be like implementing NTP over LDAP.

Thanks,
Alex

On Tue, Jan 20, 2009 at 7:24 AM, Emmanuel Lecharny <el...@gmail.com>wrote:

> IMO, NTP is the way to go. Time synchronizing servers when doing
> replication is a part of the deal. Any other solution is a bit
> overworkilling, when kikking some sysadmin's bottom could do the job
> more efficiently ;)
>
> On Tue, Jan 20, 2009 at 3:57 AM, Alex Karasulu <ak...@apache.org>
> wrote:
> > Hi all,
> >
> > The first thread on replication after Emmanuel's comments started taking
> a
> > turn towards the question of time synchronization across replicating
> > servers.  Time synchronization is very nice to have and forget about
> while
> > writing the code around replication.  And yes we do not have control over
> it
> > since administrators can fudge this up.
> >
> > Another alternative would be to have all the servers in the cluster
> > communicate with one another on occasion to share their current time to
> > establish offsets.  These offsets (differences in time) between servers
> in
> > replication agreements can be used to correct for any time variance we
> might
> > encounter.  Furthermore if the time is adjusted on the hosts, the cluster
> > can adjust to time changes.
> >
> > WDYT?
> >
> > Alex
> >
>
>
>
> --
> Regards,
> Cordialement,
> Emmanuel Lécharny
> www.iktek.com
>

Re: [Replication] Using time offsets instead of absolute time

[Emmanuel Lecharny <el...@gmail.com>]

IMO, NTP is the way to go. Time synchronizing servers when doing
replication is a part of the deal. Any other solution is a bit
overworkilling, when kikking some sysadmin's bottom could do the job
more efficiently ;)

On Tue, Jan 20, 2009 at 3:57 AM, Alex Karasulu <ak...@apache.org> wrote:
> Hi all,
>
> The first thread on replication after Emmanuel's comments started taking a
> turn towards the question of time synchronization across replicating
> servers.  Time synchronization is very nice to have and forget about while
> writing the code around replication.  And yes we do not have control over it
> since administrators can fudge this up.
>
> Another alternative would be to have all the servers in the cluster
> communicate with one another on occasion to share their current time to
> establish offsets.  These offsets (differences in time) between servers in
> replication agreements can be used to correct for any time variance we might
> encounter.  Furthermore if the time is adjusted on the hosts, the cluster
> can adjust to time changes.
>
> WDYT?
>
> Alex
>



-- 
Regards,
Cordialement,
Emmanuel Lécharny
www.iktek.com